Theory of Ordinary Differential Equations EARL A. CODDINGTON Assistant Professor of Mathematics University of California, Los Angeles NORMAN LEVINSON Professor of Mathematics Massachusetts Institute of Technology TATA McGRAW-HILL PUBLISHING CO, LTD. New DolhiCopyright © 1955 by the McGraw-Hill Book Company, Inc. Alt rights reserved except those granted to the United States Covern- ment, Otherwise, this book, or parts thereaf, may not be reproduced in any form without permission of the publishers. TMH EDITION 1972 9th Reprint 1987 Reprinted in India by arrangement with McGraw-Hill Inc., New York This edition can be exported from India only by the publis ters. Tata McGraw-Hill Publishing Company Limited. PUBLISHED BY TATA MOGRAW-HILL, PUBLISHING COMPANY LIMITED AND PRINTED AT PEARL OFFSET PRESS PRIVATE LIMITED 5/33 KIRTI NAGAR INDUSTRIAL AREA, NEW-DELHE-I10015.To Faar anp SurPREFACE This book has developed from courses given by the authors and probably contains more materia! than will ordinarily be covered in a one-year course. The selection of material is partly conditioned by the interests of the authors. It is hoped that the book will be a useful text in the application of differential equations as well as for the pure mathematician. Pre- requisite for this book is a knowledge of matrices and of the essentials of functions of a complex variable. The notion of the Lebesgue integral is used in Chaps. 2, 7, 9, and 10. However, Chap. 2 is needed only for certain parts of Chap. 15, which, so far a8 applications go, are ade- quately covered by Chap. 13. The Lebesgue integral can easily be avoided in Chap. 7, as is indicated there. However, a rigorous study of Chaps. .9 and 10 requires a mathematical sophistication that would certainly include the ability to understand the atatements of the theorems required from integration theory. An alternative approach is to apply the theory of Chaps. 9 and 10 to a restricted class of functions as is done in the proof of Theorem 3.1 of Chap. 9. This approach requires a knowledge of the Riemann-Stieltjes integral only. Chapters 3 through 12 are on linear equations. For linear theory, it is not necessary to cover the existence theory of Chap. 1. For Chap. 3, the necessary theorem is sketched in Prob. 1 at the end of that chapter. The discussion in Sec. 7 of Chap. 3 suffices for Chaps. 4 and 5. For Chaps. 7 through 12, Prob. 7 of Chap. 1 provides the additional existence theory needed. Chapters 4, 5, and 6 are not needed for any later chapters. Chapter 8 is not required for any later chapter, nor are Chaps. 9and 10, Chapter 8 does not depend on Chap. 7. Chapter 12 requires only Chap. 7 and, for See. 5, also Chap. 11. Chapters 1 and 3 only are required for Chaps. 13 and 14. Chapter 1 will suffice for most of Chap. 15 and for Chaps. 16 and 17. No attempt has been made to give the historical origin of the theory, and only a limited number of references are given at the end of the book. In keeping with this approach, the authors make no mention in the text where they present new results. vitviii PREFACE The problems, in some cases, give additional material not considered in the text. ‘The preparation of this book was greatly facilitated by a gran: from the Office of Naval Research. The authors are indebted to a number of colleagues and stucents who read portions of the manuscript, in particular, F. G. Brauer, Prof. A. Horn, and Dr. J. J. Levin. Eant A. Copn:noron Norman LevinsonCONTENTS PREFACR. 6 ee ee ee we Cuaprer i. Existence anp Uniqueness or Souutions. 2. 1 6 4 1 1. Existence of Solutions © 2 2 2 2 ee ee 2, Uniqueness of Solutions. . te ee ee 8 3. The Method of Suecessive Approximations - 5 50. 0. 0.0.0.0. UL 4, Continuation of Solutions 2 2 - . we eee 13 5, Systems of Differential Equations 2... ee 8. The nth-order Equation . Po ee BL 7. Depondenee of Solutions on Initial Conditions and Parameters. |. =. (22 8. Complex Systems... Poe ee ee BB Problems... eee 8 Cuabten 2. Existence axp UNIQUENESS OF SOLUTIONS (continued) . . . 42 4. Extension of the Idea of a Solution, Maximum and Minimum Solutions . 42 2. Further Uniqueness Results. . .. soe ae wee AB 3. Uniqueness and Successive ‘Approximations woe » 83 4. Variation of Solutions with Respect to Initial Conditions and Parameters . 87 Problems . 2. 1 wee toe ee eee see 60 Cuaprer 3, Lixear Dirrenentiat Equations: soe ee ee ee BR 1, Preliminary Definitions and Notations, . 2 6 1 6 2 1 6 ee 82 2, Linear Homogencous Systems... . ee ee ee OT 3, Nonhomogencous Linear Systems . Coke ee eee 4, Linear Systems with Constant Coofticients woe . . » 6 5, Linear Systems with Periodic Cocfficients . . . 7 6, Linear Differential Equations of Ordern 2 2 0. 0. - / we ee 8 7. Linear Equations with Analytic Coefficienta . . soe ek BE 8. Asymptotic Behavior of the Solutions of Certain Lincar ‘Systems soe Problems ©. 0. ee ee Cuarren 4. Linear Systems witi Isouateo Sincunantrins: SINGULARITIES ortue First Kind 6 6 ee ee ee ee eee 108 1, Introduction. . | 2. Classification of Singularities bo ee ee ee ee ee 3, Formal Solutions . Soe ee ee ee eee UB 4, Structure of Fundamental Matrices 1 11. eee eee HB 5, The Equation of the nth Order . soe . soe ee 1 6. Singularities at Infinity . rn . . 4237 7. An Example: the Second-order Equation see . os. 10 8. The Frobeniua Method . . . Fok ee ee 132 Problema ©. 0: 6 6 6 ee ee ee 135x CONTENTS Cuarter 5. Lixzan Systems wirt IsouaTep Srxquanitina: SINGUL \RITIES or THE SEconp Kinny . 1. Introduction 2, Formal Solutions . 3. Asymptotic Series. 4, Existence of Solutions Which Mave the Formal Solutions ¢ ag Asymptotic Expansions—the Real Case 5. The Asymptotic Nature of the Formal Solutions in the Complex Guse 6. The Case Where Ay Has Multiple Characteristic Roots, soe 7. Irregular Singular Points of an ath-order Equation . 8. The Laplace Tntegea and | Asymptotic § Series . Problems . . eo Cuarter6. Asymerotic Benavior or Linear Sysress Contaixine « Lance PARAMETER, 6, ee 1. Introduction © 2. ee 2, Formal Solutions. . ee 3. Asymptotic Behavior of Solutions Be 4, The Case of Equal Characteristic Roots . . 2 1 6 1 6 5, The nth-order Equation toe Soe ke ee Problems... a Crarrer 7. Serr-anjoint Eicenvatve Pros.ems on A Finite INTERVAL - 1. Introduction : 2, Sclf-adjoint Eigenvalue Problems | |.) 3. The Existence of Eigenvalues... Co 4, ‘The Expansion and Completeness Theorems». 1. 1 se Problems . Cuapren 8. Osciutattion aNp Costrarnison THeorems ron SEcoND-onDER Lingant EQUATIONS AND APPLICATIONS . 1. Comparison Theorems 2 2 2. 1 ee 2. Existence of Eigenvalues. 3. Periodic Boundary Conditions . 4. Stability Regions of Second-order Equations with Periodic Coefficients Problems . So Cuarrer 9. Sixcurar Sevy-apsornr BounpaRy-valve PRop.ems ror Seconp-onper Equations. 1, Introduction 2. The Limit-point and Limit. circle Cases . 3, The Completeness and Expansion Theorems in ‘the Lintpaint. ¢ Cas at Infinity . 4. The Limit-circle Chae. at Infinity eT 5. Singular Behavior at Both Ends of an Interv: al soe ee Problems... es Cuarter 10. Stncuiar SeLy-apioint Bounpary-vaLue Paos.ems ror nH oRvER Equations . erry 1, Introduction . er) 2. The Expansion Theorem ‘and Parseval ‘Equality sone te 138 138 141 M8 151 161 167 169 170 173 174 wa 175 178 182 182 184 186 188 193 197 201 208 ait 213 218. 220 222 222 225 231 242 236 254 261 262CONTENTS 3. The Inverse-transform Theorem and the Uniquenes of the Spectral Matiix 4. Green's Function . tee 5, Representation of the Spectral Matrix iy Green's Funston | 5. Problems... Se Charter 11, Auoesnirc Proverrtes oF Lincan Bounpany-vaLue PROBLEMS ona Fine Invenvan. © 2. . - 1 ee eee ee 1, Introduetion . Ce ee 2. The Boundary-form Formula . tt 8. Homogencous Boundary-value Problema.end Adjoint Problems | | 4. Nonhomogeneous Boundary-value Problems and Green's Funetion Problems... eee Cuapter 12, Non-seL¥-apsoinT BOUNDARY-VALUE PROBLEMS... S288 S28 388828 ESSE x 1, Introduction .. so 2, Green's Function and’ the Expansion Theorem for the Case. Lr = wz" 3, Green’s Function and the ° Expansion Theorem for the Case Lr = +z" + gOz. .. . . . . toe 4. The nth-order Case Be 5. The Form of the Expansion. © 1... we ee ee Problems 2 2 6 ee Cuarrer 13, Asymerotic Benavior or Nonuingar Systems: Sraniity . 1. Asymptotic Stabi rn 2. First Variation: Orbital Stability pon ee ee ee 8. Asyniptotic Behavior ofaSystem. 2. 2 1 1. ee ee 4. Conditional Stability. . coe ee ee 8, Behavior of Solutions off the Stable Manifold. | 2) le Problems . . . . So ee Cuarter 14, Pesrunsation or Systems Having 4 Peniopic Sonution . 1, Nonautonomous Systems 2 2. Autonomous Systems. . 3. Perturbation of a Linear System with a Poriodie Solution in the Non- autonomous Case. Le 4. Perturbation of an Autonomous System with a Vanishing Jacobian | Problema . . . . Be Cuarter 15. Pertrursation Tizony or Two-prsmenstonaL Rea AvTono~ Mous Sysreas . . Foe ee ee lL ‘Two-dimensional Linear Systems . soe eee 2, Perturbations of Two-dimensional Linear Systems ry 3. Proper Nodes and Proper Spiral Ponts 2... ee 4, Centers. . . ee ee ee ee 5. Improper Nees 6. Baddic Points. 2 2) 1 ee ee eh Problems © 0. 1 we ee ee ee Craprea 16, Tue Posncann-Benpixson Tuzory or Two-pistenstonaL Auronomous SYSTEMS. 2 6 1. Limit Sets of an Orbit soe 2, The Poincaré-Bendixaon Theorem . 2 = 314 321 327 329 340 344 348 348 352 366 364 370 a7 371 375 377 381 384 387 388 389 389 391xi CONTENTS 3, Limit Sets with Critical Pointe. 6 2 6 ww ee ee BOE 4, The Index of an Isolated Critical Point ©. 2 2. we ee. 898 5. The Index of Simple ¢ Critieal Point. © 2 1 1 1. we ee A Problems . . hoe eee eee, 402 Cuarter 17, Dirrerentta, Equations on a Torus. 1. Introduction. . ee ee ee 404 2. The Rotation Number ee err . 405 3. The Cluster Set 2. soe eee soe ee soe 408 4, The Ergodic Case. . Do : 409 5. Characterization of Solutions i in ‘the Ergodie Case eee ee ee AES 6. A System of Two Equations, 2 6 6 1, ee ee ee AES REFERENCES, © 6, 6 ee ee . aIy UNDEX, 6 ee ee ABTHEORY OF ORDINARY DIFFERENTIAL EQUATIONSCHAPTER 1 EXISTENCE AND UNIQUENESS OF SOLUTIONS 1. Existence of Solutions Let 7 denote an open interval on the real line —-~@ << ©, thatis, the set of all real ¢ satisfying a < { 0) about the point (1,2), it is bounded there. Let M = max |f(t2)| (2) B) @=min (« 4) M Theorem 1.1. Let fe Con ihe rectangle R. Given any e > 0, there exists an e-approximate aolution ¢ of (E) on lt — +] S a auch that-e(r) = &. Proof. Let «> 0 be given. An e-approximate solution will be con- structed for the interval [r,t + a]; a similar construction will define it for [ — a, 7]. This approximate solution will consist of a polygonal path starting at (r,é), that is, a finite number of straight-line segments joined end to end. A function g is said to have a simple discontinuity at o point ¢ if the right and left Imits of g at c exist but are not equal. In caso « = 0, it will be understood that the sat 3 is empty. and4 ORDINARY DIFFERENTIAL EQUATIONS [Cuar. 1 Bince fe on R, it is uniformly continuous there, and hence, for the given ¢, there exiats a 5, > 0 such that it l4Ga) — f@a)| S (Lp i (Gz) eR, Ge R, and |t-]ss le-a ss Now divide the interval [r, + + a] into n parta Teh 0, there exists a 5, > 0, independent of f ¢ ¥ and also 4, fe Z such that W@ —fO| 0 and rationa] number r, ¢ 7, there exists an integer N’.(r.) such that Vntre) — fnlre)| <€ (nym > Nefre)) For the given ¢ there exists a 5,, independent of ¢,f and fe F such that WO-fOl max (N.(71),..., Na(fp)). This proves the uniform convergence of the sequence {f,} on J. Theorem 1,2 (Cauchy-Peano Existence Theorem). If fe on the rectangle R, then there exists a solution ye C' of (E) on |t — 1] S a for which e(r) = & Proof. Let fen},n = 1,2, ... , be a monotone decreasing sequence of positive real numbers tending to zeroasn — ©, By Theorem 1.1, for each ¢, there exists an ¢,-approximate solution, ¢,, of (E) on |t ~ 7| 3 « such that ya(r) = & Choose one such solution yp for each ¢. From (1.4) it follows that len() — en) S Mle — (1.5) Applying (1.5) to 7 = 1, one readily sees, since |i ~ r| $ b/M, that the sequence {ya} is uniformly bounded by |&| + 4. Moreover, (1.5) implies that {ya} is an equicontinuous set. By the Ascoli lemma, there exists a subsequence {gn}, k = 1, 2,..., of (yal, converging uniformly on [r — a, r + @] toa limit function y, which must be continuous aince each On is — [Indeed, it follows from (1.5) that lot) -— e| s Mit — 2.) This limit function ¢ is @ solution of (E) which meets the required specifications, To see this, one writes the relation defining y, a8 an «approximate solution in an integral form, as follows: onl) = + fl lapel) + bale) a (18) where 4.@)-= e,(!) — f(t,e.()) at those points where y, exista, and 4,(t) = O otherwise. Because ¢, is an ¢,-approximate solution, |4,(0)| S éx Since f is uniformly continuous on R, and g,,—> ¢ uniformly onSexe. 1) EXJSTENCE AND UNIQUENESS OF SOLUTIONS 7 fr -a,4 + a], as k— ©, it follows that f(t,en(Q) > f(,e(®) uniformly onfr —a,r+a],ask— ©, Replacing by nz in (1.6) one obtains, in letting kK > », of = + f' He,e(e) as an But from (1.7), o(r) = £ and, upon differentiation, y'(t) = f(t,e(®), for F(,e(@)) is a continuous function. It is clear from this that ¢ is a solution of (E) on |f — +| 3 @ of class C. In general, the choice of a subsequence of {g,} in the above proof is necessary, for there exist polygonal paths {¢.} which diverge everywhere on a whole interval about ¢ = r as e, > 0; see Prob, 12. ‘If it is assumed that a solution of (E) through (r,£) (if it exists) is unique, then every sequence of polygonal paths {y,} for which «,—+ 0 must converge on |t — | $ , and hence uniformly, to a solution, for {pn} is an equicontinuous set on |{~ r| S a. Suppose this were false. Then there would exist a sequence of polygogal paths (y,} divergent at some point on | —+| So. This implies the existence of at least two sub- sequences of {y,} tending to different limit functions. Both will be solu- tions, and this gives a contradiction. Therefore, if uniqueness is assured, the choice of a subsequence in Theorem 1.2 is unnecessary. It can happen that the choice of a subsequence is unnecessary even though uniqueness is not satisfied, The example a’ =a (1.8) illustrates this. There are an infinite number of solutions starting at (0,0) which exist on [0,1]. For any c,0 S$ ¢ & 1, the function ¢. defined by ef) =0 OsStsSe) — i ot = (2G9) exesn 19) is a solution of (1.8) on [0,1]. If the construction of Theorem L.1 is applied to Eq. (1.8), one finds that the only polygonal path starting at the point (0,0) is y:. This shows that this method cannot, in general, give all solutions of (E). Theorem 1.3. Let feC ona domain D in the (£2) plane, and suppose (7,8) is any point in D. Then there exisis a solution y of (E) on some ¢ interval containing + in its inlerior. Proof. Since D is open, there exists an r > 0 such that all points, whose distance from (7,£) is less than r, are contained in D. Let R be any closed rectangle containing (7,£), and contained in this open circle of radiusr. Then Theorem 1.2 applied to (E) on F gives the required result.8 ORDINARY DIFFERENTIAL EQUATIONS [Cuap. 1 2. Uniqueness of Solutions The example (1.8), with the solutions given in (1.9), shows that some- thing more than the continuity of f in () is required in order to guarantee that a solution passing through a given point be unique. A simple condi- tion which permits one to imply uniqueness is the Lipschitz condition. Suppose f is defined in a domain D of the (#,) plane. If there exists a constant k > 0 such that for every (é,2:) and (é,22) in D as) — fG20| S$ klar — 2] then f is said to satisfy a Lipschitz condition (with respect to z) in D, and this fact will be denoted by fe Lip in D. The constant & is called the Lipschitz constant. If, in addition fe C in D, one writes fe (C,Lip) in D. If fe Lip in D, then fis uniformly continuous in z for each fixed ¢, although nothing is implied concerning the continuity of f with respect to 4. If D is convex (that is, D contains the line segment connecting any two points in D), then an application of the mean-value theorem of differential calculus shows that the existence and boundedness of f, (= af/az) in D are sufficient for fe Lip in D. Before proceeding to the uniqueness proof, an important inequality will be deduced. In the following, D is a domain in the (i,z) plane. Theorem 2.1. Suppose fe(C,Lip) in D, with Lipschitz constani k. Let gigs be es- and «approximate solutions of (E) of class C) on some interval (a,b), satisfying for some 1,0 <7 0 the approximate solution tends to the actual solution. The inequality (2.2) is the best possible, in the sense that equality can be attained for nontrivial y, and yz. For example, let i, ¢, ¢2 be any real constants, and let P;: (0,£:), Ps: (0,€) be two points in the (¢,x) plane. Let 9:(0) = &, and gs(0) = & and let g: and ys be solutions of the equations a =ke- ag a! = ke pes10 ORDINARY DIFFERENTIAL EQUATIONS [Cuar. 1 respectively, on {0,1]. Then 1 and yg: are clearly e- and erapproximate solutions of a! = ke there. A simple calculation shows that for y: and y: the equality sign in (2.2) must hold if & 2 &. “Note that, roughly speaking, the inequality says that, if 6 and e are smail, then so is g:(¢) — g(t). In fact, if 8 = « = 0, then yg: = ys, and there is at most one solution of (E) going through any given point (7,t) in D. This proves the following uniqueness result: Theorem 2.2, Let fe (C,Lip) in D, and (7,f)2¢ D. If g, and y: are any two solutions of (B) on (a,b), a <+ 0) M = max |f(42)] (2) ¢ R) =min (a, 2 a = min (a, 57 Then there exists a (unique) solution ge C! of (E) on |t — 7| S a for which o(7) = & Proof. Let {en} be a monotone decreasing sequence of positive real numbers tending to zeroasn—» %. Choose for each ¢, an ¢n-approximate solution g,. These functions satisfy the relation onl) = + [! SGenls)) + Anlo)) ds 26) and let and where A,() = ga(t) — f(é,ea(2)) at those points where ¢/, exists, and An(t) = Q otherwise. Now A,(¢)—+ 0,a31-— ©, uniformly on |t -- 7} S a. by the very definition of ¢,. From (2.2) applied to y, and g, one obtains for |t — z| S a, len) en(O| s Set soles ~ D where & is the Lipschitz constant. Thus the sequence {¢,] is uniformly convergent on |é — 7| S a, and therefore there exists a continuous limitSec. 3] EXISTENCE AND UNIQUENESS OF SOLUTIONS YM function g on this interval such that ga({) > g(t) asn— © uniformly on \t~7r] Sa. This fact, plus the uniform continuity of f on R, implies that L(bea)) > fhe) (a @) uniformly on |t ~ 7| S a. Hence oft a tirn_f (A(s,e0(8)) + An(e)) do = f'(s,0()) ds and from (2.6) one gets finally, by letting n— ©, ol = E+ [ feos) ds which proves the existence of a solution ge C! of (E) on |é—7| Sa. It is unique by Theorem 2.2. Clearly le®) ~ oat 5 2 (2.7) etltrl — 1) k 3. The Method of Successive Approximations ‘The existence proof given in Theorem 1.2 is unsatisfactory in one respect in that there is no constructive method given for obtaining a solu- tion of (E). However, as was pointed out after that proof, if the solution through the given point is known to be unique, then the original polygonal approximate solutions can be used to obtain the solution; no subsequence need be chosen. In particular, if f satisfies a Lipschitz condition, the inequality (2.7) gives » bound for the error in using an ¢,-approximate solution in place of the actual solution. In the following a very useful method, known as the methed of successive approximations, will be con- sidered, and the existence of a solution will be deduced with its aid. Here again one can conveniently compute an upper bound on the error involved in stopping the process after a finite number of steps. The results will be deduced for the case of the rectangle 2 defined by Ry lt-7]sa |z-t sb where (7,¢) is some point in the (iz) plane, anda > 0,b>0. It will be clear that the analogue of Theorem 1.3 also holds. If feC on R, then f is bounded there; let max |f| = Jf on R, and, as before, a = min (a,0/Mf). It is clear that a solution » of (E) on |t — 7| S a for which ¢(r) = & must satisfy the integral equation ot) = + [feel ds (e-rl S a) (a) and conversely, if ¢ satisfies (3.1), it satisfies (E) and g(r) = & The12 ORDINARY DIFFERENTIAL EQUATIONS (Cuar, 1 successive approximations for (E) are defined to be the functions yo, yi, +.» , given recursively by the formulas vol!) = & vrei) = &+ [seo ds (k= 0,1,2,... 5 {t-7] Sa) (8.2) It is shown below that these functions actually exist on |f - 7] S a. Theorem 3.1 (Picard-Lindeléf). If fe (C,Lip) on R, then the successive approximations y, exist on |t — 7| S$ a as continuous functions, and con- verge uniformly on this interval to the unique solution @ of (E) such that o(r) = & Proof. Consider the interval {r, 7 + a]; similar arguments hold for {fr — a, 7). It will be shown that every y; exists on [7,7 + a], ge eC! there, and la@® -sM@-7) @elr tal) (3.3) Obviously yo, being the constant &, satisfies these conditions. Assume gr does the same; then f(t,¢s(¢)) is defined and continucus on [r, r + a]. From (8.2) this implies gy: exists on [r, r + a], yayre C? there, and leeai(t) — &| S.Af(t — 7). Therefore these properties are shared by all the g: by induction. Geometrigally, this means that all the ¢. start at (r,£) and stay within triangular region T between the lines 2—-E= +M(t—7) andt=7 +a. It rémains to prove the convergence of the gy. Let A, be defined by A = len) — Ol @ele, + + al) Then from (3.2) by subtraction and the fact that fe Lip on R with some constant ¢ > 0, a0 Se f' assls) do (3.4) But (9.3) gives fork = 1, Aolt) = ler) — eo(t)] S Mt — +) and an easy induction on (3.4) implies that ay s BEE tele tad This shows that the terms of the series y A.(!) are majorized by those kno of the power series for (M/c)e, and therefore the series ). A(t) is uni- andSec, 4] EXISTENCE AND UNIQUENESS OF BOLUTIONS 13 formly convergent on [r,7 + a]. ‘Thus the series volt) +) Corsilt) — ontt)) 20 is absolutely and uniformly convergent on [r, r + a); consequently the partial sum n= oolt) + YD, (ensll) ~ geld) = oald) eo tends uniformly on [r, 7 + aj to a continuous limit function ¢. It will be shown that the function ¢ satisfies (3.1), and is hence a solu- tion of (E) on [7, + + a] for which g(r) = & Since all the gy are within the region 7, s0is yg. Therefore f(s,o(s)) exists forse [r,7 + a]. Clearly | [ Ue — seertolde| sf" 1f6.e(2)) ~ Kewuls))1 de ef lols) — els) ds the latter inequality being duc to the fact that fe Lip on R. Now le(s) — ve(s)| + Onsk > & uniformly on [7,7 + a], and thus the above inequalities show that (3.2) yields (3.1) as k-» ©, The solution y is unique by Theorem 2.2, and this completes the proof. An upper bound for the error in approximating the solution y by the nth approximation g, is easily computed. It is given by let — exidl SY local ~ eh gE EE Eon kent) eM SP (eal! Af (cant! SY (ca)! AT (oa) #4 c BR Scar 2, Me mri” kn keatl 4, Continuation of Solutions Suppose that fe C in some domain D of the (t,x) plane and that (E) has a solution g which exists on a finite interval (a,b) and passes through some point (r,£)¢ D,a <1 0. If now is defined by aD = 6) (te (a,b) a=W eld, b +8) then @ is a solution of (I) of class C! on (a, b + 6], and g(r) = & The only point to check is the existence and continuity of the derivative 2’ at b. It will be shown that a=tt [Heads telab+s)) (4.1) This is obvious fora <¢ Sb. For ¢ > 6 it follows from the definition of @ that a = ob — 0) + f' Hs,000)) ae But b old — 0) = E+ f H60,0(6)) do which proves (4.1) for! > 6. The continuity of @ in (4.1) implies that of £{s,8(s)), and by differentiation of this integral equation for @, one obtainsSec. 5} EXISTENCE AND UNIQUENESS OF SOLUTIONS 16 the fact that ¢’(t) = f(¢,¢(¢)) for te (a,b + A]. Naturally, @ is called a continuation of the solution ¢ to (a,b + 8]. There are just as many con- tinuations of g to (a, b + A] as there are solutions of (E) issuing from (,e(6 — 0)) which exist on [b, b +- 8). If it is known that there exists at most one solution through (b,¢(6 — 0)) (for example, if fe Lip on D), then one can speak of the continuation of ¢ to (a,b + J. In general, if a con- tinuation of a solution ¢ on (a,b) exists on some interval containing (a,b), then one says y can be continued, or haa a continuation. The above remarks are summarized in the following theorem: Theorem 4,1. Let feC ina domain D of the (t,x) plane, and suppose f 4s bounded on D. If ¢ is a solution of (E) on an interval (a,b) then the limite p(a + 0) and y(b — 0) exist. If (a,e(a + 0)) [or (b,p(b — 0))] #s in D, then the solution g may be continued to the left of a (or lo the right of b). A more general analysis of the continuation problem appears in Chap. 2, See, 1. 6. Systems of Differential Equations Suppose n is a positive integer and f,, ... , fa are n realcontinuous functions defined on some domain D of the real (é,21, . . . , Zn) space. Just as in the case where n = 1, this is abbreviated f;eC in D, ¢ = 1, . ,®. One can then formulate the following problem: ‘Problem. To find n differentiable functions yi, . . . , on defined ona seal t interval I such that a) hei, ..-,en)eD (te Gi) eid) = Filer, - ~~ 1 eal) (te ils 4, fa... 2) This problem is called a system oj n ordinary differential equations of the first order, and is denoted by «) a= fet, 2, te) Hl... ny Correspondingly, if such an interval J and functions (1, . . . , gn) exist, then the set of functions (yi, . . . , @n) is called a solution of the system (E) on I. Let (7,1... &s)e¢ D. The initial-value problem consists of finding a solution (v1, « » @x) of (E) on on interval J containing + such that w(t) = & It turns out that the results so far obtained for the case n = 1 can be carried over successfully to the aystem (E). Let X denote the Euclidean dimensional space with points x having coordinates (x, . - . , Za) ‘Then the functions /; defined on the (é,2;, . . . , 7) space give rise to16 ORDINARY DIFFERENTIAL EQUATIONB [Caar..1 functions f; on the (t,x) space defined by Fiz) = filgay, .. . tn) Also associated with any point 2 in the z space is the one-column matrix ay In, called the vector associated with a; x; is called the ith componcnt of 4. Clearly j; defines a function f, of ¢ and the vector # by Silt) = Klt,2) Associated with the point (1,2) in the (é,z) space one has the vec or and this gives rise to a vector f(¢,4) defined by S(t) Jie) = fa(b2), If the z, are differentiable functions of é, then #’, the derivative cf 4, is defined by , zy x Then the equation (E) may be written simply as # = fla) In actuality, there is seldom any chance of confusing the point x with the vector 2, and so the same notation will be used for both; the circu nflex will be dropped. This has the effect of identifying the space X wit. theSec. 5] EXISTENCE AND UNIQUENESS OF SOLUTIONS 17 space X of all n-rowed one-column matrices, considered as a vector space. Similarly, no confusion results by identifying the functions fi, f;, and f., and this will be done in all that follows. With this understanding, Eq. (E) may be written as @) a! = f(t,2) where f may be thought of as a vector function of real ¢ andthe point ze X, or as a function of ¢ and the one-column matrix zy ns Tn, Here Silt,x) Slt2) = n(l,2), A solution of (E) on an interval J becomes a vector function g with com- ponents y1, . . . , ¢n defined on / satisfying (i) (9) = (hel), -.. ene D el) Gi) e'() = fe) (el) ‘The magnilude (or norm) |2| of n vector ze X with components a, . . . jn is defined byt lel = } tad 1 ‘The Euclidean length {|x| of a veetor x X is defined by tell = (3 ir) ‘The distance between two vectors 2,y eX is defined to bely ~ z|. It is + Other definitions for the magnitude of a veetor can be used, such ag Fe] = max (Iz) G=l...,m or the Euclidean length {x4 the various norms are readil ki skisals} pbs tell s nlfef Hel Steps nbtell aN] s fet ss Il As a matter of fact, the following inequalities relating ‘seen to hold: and18 ORDINARY DIFFERENTIAL EQUATIONS [Cuar, 1 obvious that this distnnee satisfies the ordinary rules for a distance function (a) ly— 2] = |e - al (b) ly-—a}20 andly—al|=0 — if and only ify =: (c) w-asly-adtk—-al This distance function allows one to consider X as 8 metric space; a sequence of vectors {2} is said to be convergent if it is convergent with respect to this distance function. Note that here the 2, are vectors and not components of a vector. Clearly {z,} is convergent if anc only if each of the component sequences (2,} (z, having compon2nts 2, t= 1,...,n) is convergent, If g is a differentiable vector function on some f interval (a,b). that is, g' exists on (a,b), and r = |g is the function defined by r() = |p} (te (a,b)) then, if agn a = a/lal for a = 0, 1 = J, a0 sen at) int Tf none of the components 9;(¢) of g(6) vanishes, clearly ri) = 25 (0) sen gs(t) im exists and I'l s le’) 6.1) In any case (whether a component of g{é) is zero or not), it is alvays true that if g‘() exists, then the right- and left-hand derivatives 7'.() and r_(t) existt and satisfy rel S ly'@l (5.2) For if t is an isolated zero of any component g; of g, a straigh forward calculation shows that the right- and left-hand derivatives of gi| exist at ¢ and do not exceed |g,|; and if ¢ is a zero of y; which is not isolated (that is, fis a limit point of zeras of gi), then an approach tod through a : equence of these zeros shows that g;(t) = 0, and hence |g,|'(t) = 0. In cither caso, (5.2) is valid, t Notice that if g ix absolutely continuous, then so is r, since be(ta) — r(tsa Ss Let) — a(t)Sze. 5] EXISTENCE AND UNIQUENESS OF SOLUTIONS 19 If [lo()|| 0, and {gil is the function defined by |[gll() = llp(e)||, then Igll'(é) exists, and it is easily verified that loi’! s te'@ll G3) For integration purposes the magnitude | | is more useful than the length |] |. If g and r = |9| are integrable over an interval (a,b), then | {oma s [roa (4) where by f . gt) dt, of course, is meant the vector whose 7th component is [oxo a. Suppose a vector f (which may have any finite number of components, not necessarily ) is defined on a domain D of the (1,r) space. If there exists a constant k > 0 such that for every (t,z) and (t,2) in D \sz) — f4)| S ble — 2 (5.5) then the vector f is said to satisfy a Lipschitz condition (with respect to x) in D, and one writes fe Lip in D, Suppose fz C on a domain D in the (¢,z) space. An capproximate solu- tion of (E) on an interval J is a vector function ye C on I such that (@) (4e@)eD (te Z) (6) geCton] except for a finite set of points S on J o le) ~fGeO)| Se (te -S) Tn terms of tHe definitions introduced above, all the theorema in Secs. 1-4 are valid for the vector equation (E) if, in their statements and proofs, 2,f are replaced by the vectors z,f and the magnitude is understood in the sense defined above for vectors. (The Ascoli lemma is valid for vectors also.) Therefore it will be assumed from now on that these theorems have been proved for the more general vector equation (E). A particularly interesting system is the linear system () a= Yau, G=1...,”) gat where the a,; are continuous functions on some closed bounded ¢ interval [a,b]. If f is the vector with components f, defined by Hba) = Faults, GA. sm)20 ORDINARY DIFFERENTIAL EQUATIONS (Crap. 1 then clearly f satisfies a Lipschitz condition on the (mn + 1)-dimensional region D: astsb \z| << (Here D is not a domain since it is not open.) In fact, if (¢,2) and (1,2) are in D, 1S@,2) — f44)| S ble — al where a k= max) ax] (eledij=d.. 4m) st Theorem 6.1, For the linear system (L), where the functions a2 C on [a,b], there exists one and only one solution y of (I.) on [a,b] passing through any point (r,£) ¢ D, that is, yz) = & Proof. Since the vector f satisfies a Lipschitz condition on D, the existence and uniqueness of a solution y through (r,£) over some interval [c,d] & {a,b] are guaranteed. It remains to show that ¥ can be continued to a unique solution y on the whole interval [a,b]. If # is any solution of (L) through (r,£) existing on any subinterval of {a,b], then applying the inequality of Theorem 2.1 to g, = Pand g: = 0, one obtains WO] S lelero (5.8) for tin the domain of definition of ¥. Now suppose y does not have a continuation to [a,b], and for definiteness assume ¥ has a contit:uation ¢ existing up tol < 6, but cannot be continued past But, from 5.6), the path (¢,¥(t)) remains inside a closed bounded subset of D, where fe C and satisfies a Lipschitz condition. Therefore, by Theorem 4.1 (interpreted for systems), ¥ may be continued beyond 7. This results in a contradic- tion, thus proving a continuation ¢ of ¢ exists on (a,b). Itis unique, for f satisfies a Lipschitz condition on D, Theorem 5.2. Let the functions aj (i,j = 1, . . . , n) be cantinuous on an open interval I, which may be unbounded. Then there exists on / one and only one solution y of (L) satisfying er) =§ Gel, lth < &) (5.7) Proof. By Theorem 5.1 there exists a solution of (L) satisfying (5.7) on every closed subinterval of / containing r; using the same argument as in that proof, any such solution may be continued to the whole of I uniquely. A more detailed trentment of the linear system will be the subject of Chap. 3.§xc. 6] EXISTENCE AND UNIQUENESS OF SOLUTIONS 21 6. The nth-order Equation Suppose f is a real continuous function defined in a domain D of the real (tm, ...,2n) space. Then the nth-order equation associated with f, ge (E.) ad = fljzz,... , t-Y) ( = ie is defined to be the following problem: Problem. To find a function ¢ defined on a real t interval I possessing n derivatives there such that @ e.g, ..,e%O)sD — (ts 2) Gi) oD) = flo. -.., eH) ta If such an interval J and function ¢ exist, then g is said to be a solution of (E,) on I. If yg is a solution, clearly ge C* on J. Notethat 2, f, and g are not vectors here, Let (rf, ..., &)¢D. Then the initial-value problem consists of finding a solution. ¢ of (E,) on an interval 7 containing ¢ such that or) = bo) = &, .. . oG) = be The theory of the equation (E,) can be reduced to the theory of a system of n first-order differential equations. Indeed, associated with Eq. (E,) ia the syatem of equations zy = Dy ty =a (&) . Tua = Zn zm = f(bty... te) which is defined for (¢,2) = (4,41, ..., %)¢D. If the vector ¢, with ¢.), is a solution of (,) on JZ, then since rene gf", components (e102, tae Pr = Py Os = Fe = Fy Fer@), . - 5 en) = Mh, ©. IO) = oP O and the first component g: of g is a solution of (E,) on J. Conversely, if eis solution of (E,) on J, then the vector ¢ with components 91,91, , g-» ig o solution of the system (B,) on I. The system (f,) is called the system (or vector equation) associated with the nth-order equa- tion (Ea). If gilt) = &, . . . eC) = En, then the vector ¢ satisfies (7) = §, where t = (f1, . . . , &), and conversely, It is thus clear that all statements proved about the system (i,) carry over directly to statements about the nth-order equation (E,). In par-22 ORDINARY DIFFERENTIAL EQUATIONS (Cuar. 1 ticular, if fs C on a domain D of the (¢,a1, . . . , Za) space, and P is a point of D, there exists a solution ye C* of (Ey) on some ¢ interval and passing through P. If, in addition, fe Lip in D, that is, if (Wea, 2) = Seti ASD bee — 4 i=l for some constant k > 0, then the solution through P is unique. %. Dependence of Solutions on Initial Conditions and Parameters A solution of a differential equation on an interval I can be considered asa function, not only of te Z, but of the coordinates of a point through which the solution passes. Tor example, the first-order equation in one dimension 2’ = x has the solution g(t) = fe'~7 through the point (7,2). This determines a function of (¢,7,t), which is also calledt y, given by o(t,r,£) = ke". In the general situation, it is important to know how g behaves as a function of (t,r,t) together, and, in particular, under what circumstances ¢ is continuous in (t,7,). In the following the behavior of the solutions as functions of the initial conditions will be investigated for the general case of a system. Let D be a domain in the (n +- 1)-dimensional real (¢,2) space and sup- pose fe(C,Lip) in D. Let y be a solution of the equation ) a! = f(z) on some interval J. Thus (4,y())eD for teZ. It follows from the existence theorem that (E) has a unique solution through any point (r,£) close enough to the given solution. However, the existence theorem assures the existence of the solution only over some short é interval con- taining r, Actually, it can be shown that the solution exists over the whole interval J, and is a continuous function of (¢,7,). The following theorem will be proved. Theorem 7.1. Let fe (C,Lip) in a domain D of the (n + 1)-dimensional (4,2) space, and suppose y is a solution of (E) on an interval Iza St Sb. There exists a 5 > 0 such that for any (r,£) « U, where U: a 0 so that the (é,2) region Ui, given by Ui: tel lz-v¥@l 5 & isin D. Then let 8 be chosen so that 5 < e~**-)3,, where k is the Lipschits constant. With this 8, define U as in the statement of the theorem; see Fig. 2 for the casen = 1. If {r,¢) e U, there exists a solution @ through (7,£) locally, and this satisfiea . ott.) = €+ f' fleeler.8) ds 7.1) as far as it exists. Moreover, for ¢e I, WO = 6) + [Neve as (7.2) + A topological mapping 7’ of a set S onto s set T(S) is a one-to-one mapping such that 7 and 7~! are continuous.a ORDINARY DIFFRRENTIAL EQUATIONS [Cuar. 1 Thus, using the fundamental inequality, (2.2) with « = 0, there results lo(tsr.2) — HOLS 1E ~ vet < ay This proves ¢ cannot leave U;, and can therefore, by Theorem 4.1, be continued to the whole interval J. The continuity of » on V will be proved by showing that ¢ is the uni- form limit of continuous functiehs on V. Note that ¢ aatisfies (7.1) on T. Define the successive approximations {y;| for (7.1) by oult) =) + E~ ¥G) elt = 8+ [Semend) de 4 =0,1,2,. Then for (7,t)¢ U lettin) ~ WO = 1 — ve) <8 which shows that (t,o(t,7,€)) e Ui for¢eZ. Clearly geCon V. From (7.3) for j = 0, and (7.2), it follows that .) (7.3) lorena) ~ volt.) =| {° Utevatene) — sep) def | ['leslor.t) — vol do| = He — VOI — a1 and hence horl7.8) — HOLS + le — ale — ¥)I < delle — yer)| ay, by tr A. The symbol exp u denotes e. i=l Theorem 7.2; Let the hypothesis of Theorem 7.1 be satisfied, and suppose Sa exists and f.eCon D. Then pe C! on V, and moreover det gi(t,7,£) = exp f tr fuls,o(syr,8)) de (7.5) Remarks: The fact that f,eC on D actually makes the explicit Lip- schitz hypothesis for f superfluous. Notice that det ye(t7,8) is just the Jacobian of the transformation, taking into ¢(é,7,¢), which was considered in the remarks following Theorem 7.1. For the case where f is an analytic function, Theorem 7.2 is easily obtained from Theorem 7.1, as is shown in Sec. 8. The reader interested mainly in this important case can therefore omit Theorem 7.2. Proof of Theorem 7.2. In order to prove the existence of ge, it is suffi- cient to consider the case of d¢/8f,, where & = (t1, ..., &.). Let A= (uy, 0,...,9), = & +h, and let (7,8) and (7,£) be in U. If x is the function defined by xtn&h) = v(t7,6) = ol(t,r,t) for (7,8) e V, then what has to be proved is that lim x(67,8,4) (7.8) 40 exists. It will be shown that the limit in (7.6) exists uniformly on V and that the limit function is continuous on V. This will prove dy/a£, exists and is continuous on V. The motivation behind the proof is very simple: The solution ¢ satisfies (E), and so e'(tn.8) = Stel 8)) Thus, if g and f are sufficiently differentiable, ae\' a0 (22) Gna = peeetirien 22 tre26 ORDINARY DIFFERENTIAL EQUATIONS. (Cuar. 1 where the latter product is an ordinary matrix product. ‘Therefore ap/dk, is a solution of a lincar differential equation. All the following proof does is to justify this procedure. Let (67,84) = ft7,6) ~ eltsr,£) Using the inequality (2.2), there results 18¢67.E,A)| S [OCsr,BA)leHe ss Yhylete-> (aD Thus as hk, + 0, 0-+ 0 uniformly for (4,7,€) ¢ V. Since ¢ is a solution of (E) Lr BA) = Sehr.) — fer.) (7.8) Using the theorem of the mean on the right side of (7.8), and recalling that f,¢C on D, there exists a matrix f = (Lj) such that O'(Er EA) = (fa(tvoltsr,€)) + P)O(,7,8,A) (7.9) where, given any e, > 0, there exists a 6,, which depends on e,, such that Il =). Wal 0 wiat ag hy ~~» 0 uniformly for (¢,7,£) « V. Since x = 6/h, (7.9) yields xin Eh) = Seoltinn xr, &sh) +y (7.10) where y = I'°0/h; so that by (7.7) Ins [Tle Thus y— 0 as hy 0 uniformly on V. In particular, given any € > 0, there exists a 6, > O such that |y| < eif |hil < 5,. Thus (7.10: states that x as a function of ¢ is an «approximate solution of the linear equation w = Sltoltr.O)y (2.1L) provided that |A,| < 6,. The initial value x(z,7,%,h) is e,, where e; is the vector with components (1, 0,. , 0). Consider now for fixed (r,t) ¢ U the solution € of (7.11) which assumes the initial value eat £= +r, That this solution exists on /:¢ S¢Sb follows from Theorem 5.1. The fact that x is an -approximate solution of (7.11) for |A,| < 5, implies by Theorem 2.1 that Ixttr.&h) ~ Arb S Eee — D + Here use ia made of the fact that for (fr,g) ¢ V the pointa (t,y(¢,7,)) ¢ U1, a closed bounded set. Thus f, is uniformly continuous on U.Sec. 7] EXISTENCE AND UNIQUENESS OF SOLUTIONS 27 for (t,7,£) on V. Clearly this implies that lim xXtr.A) = A(t,7,8) uniformly on ¥. This proves the existence of dy/d£, and also proves that it ig the solution of (7.11) which assumes the initial value e, at ¢ = +. The uniformity of the convergence of x as h—+ 0, and the continuity of x on V, imply the continuity of dy/dt, on V. An entirely similar argument proves the existence and continuity of dp/8, j = 2,..., 7, on V. Also if e is the vector with all com- ponents zero except the jth, which is 1, 2 (7,2) = j= seers Gabe. yn) (7.32) and dp/dé; is a solution of (7.11). The columns of the matrix g¢ are pre- cisely the vectors dp/d£;. Therefore the following matrix equation is valid: lr.) = Seller £)eeltr.€) (7.18) where of = dy:/a. The relation (7.12) may be written as edlrrg) = E (7.14) where E ig the n-by-n unit matrix, 10--: 0 o1- B=|[ 0° .. . - 0 Or-- a1 The relation (7.5) is a consequence of a general fact concerning matrix solutions of linear systems. Since this relation is of importance in itself, it will be proved in the next theorem. One obtains (7.5) from (7,18) below using (7.13) and (7.14) and the fact that det Z = 4, It is but a repetition of the previous arguments to show that d¢/dr also satisfies the linear equation (7.11), once it is observed that it has the initial value given by 88 (er,8) = —s0e.8) (7.18) This is shown by a direct calculation as follows: (77,6) — elt.) 8) = 0(r,7,6) — & = ort) ~ orb) = ff Keeler.) ds2 ORDINARY DIFFERENTIAL EQUATIONS {Cuap. 1 Thus eer eon) oh L getas8) de #-1 Since the integrand is continuous for (s,7,¢) € V, it follows that the limit as 7— 7 exists for (r,é) ¢ U and gives (7.15). Theorem 7.3. Let A be an n-by-n malriz with continuous elements on an interval I:a St S b, and suppose is a matriz of functions on J satisfying ¥() = A(Yd(2) (te I) (7.18) Then det ® satisfies on I the first-order equation (det &)’ = (tr A)(det +) (7.17) and thua for 7, te I det (2) = det #(r) exp f tr A(s) ds (7.18) Proof. Let ij, az be the elements in the ith row and jth column of & and A, respectively. Then (7.16) says 6) =) aulen) d= 1. ym) (7.19) kat The derivative of det is a sum of n determinants Ca Zo) en ei em (det ay = a an am + Pa rs Pim Pat Pat “* Pan Prt Put "nw Pu Pit oe pO FE , 7, - Pas Fan Pan Using (7.19) in the first determinant on the right, one gets ¥ engn ¥ ewer nt D auves ® re ’ en ps2 nee pin ent Pre rte nn and this determinant is unchanged if one subtracts from the first row a): times the second row plus ais times the third row up to ai, times the nth row. This givesSkc. 7] EXISTENCE AND UNIQUENESS OF SOLUTIONS 20 Gigi Giugpie Agia en vr Pin Par Par Pan which is just a,, det. Carrying out a similar procedure with the remaining determinants, one obtains finally (7.17). The eqnation (7.17) is of the form u’ — a({)u = 0 from which follows wexp [- f a(s) ae] = constant which gives (7.18). The case where the right member f of (E) contains a parameter vector ucan be readily dealt with. Suppose » space has & real dimensions, and let I, be the domain of y space, |x — uo] < ¢, where yo is fixed and ¢ > 0. As above, D is a domain of (4,2) space. Let D, be the domain of (¢,2,n) space D,: (4z)e D wel, and let fC on D, and satisfy a Lipschitz condition in x uniformly on Dy. The differential equation {E,) x = filszu) will be considered here. For a fixed given » = uo, let ¥ be a solution of (E,) on an interval a S$! 6. Then the fallowing theorem, which includes Theorem 7.1 as a special case, will be proved: Theorem 7.4. Let ¥ be the solution of (E,) described above. There exisis a > 0 such that for any (1,£,4) ¢ U,, where Ur acr 0 so that the (¢,2,«) region Uy given by Uni aStSb le—vO| + le — mol So isin D,. Define the approximations {¢;} by30 ORDINARY DIFFERENTIAL EQUATIONS [Cnar. 1 eolt.Bu) = vi) + &— 9G) onrlerin) = £4 f SlovestssrsBa)oa) ds Clearly lpo(t,rs 8m) — YO] = 18 — ¥6)] and Jer(,7,8n) —~ eoltr Eu)| = [ 1Hleeolo.r, 8.0 a) — floeCe),ne)} ds | (7.20) The uniform continuity of f in U,, implies that, given any ¢ > 0, there exists a 5, > O such that \f(s,0(8,7,8u),#) — S6(s),H0)| < € provided that a Ss S 8, (7,f,u) ¢ Ui, and Ie = vr) + le — aol < & (7.21) Thus (7.20) implies ler(t,r, £m) — poltz.Eu)] < elf — a] provided (7.21) is valid. Proceeding as before, there now results Sf — ppt ipe leni@rén) -- e(GrEm)) S “oer where k is the Lipschitz constant. [ct ¢ be chosen so that free — 1) <4 and let § = 8 < 8,/2 be chosen as above for thise. Then it follows easily by induction that, for all j. (4v,(/7,80)) is in the region Uy, for all (én) ¢U,. The continuity and the uniform convergence of the ¢ on V, lead to the result of the theorem. The generalization of Theorem 7.2 to (E,) is valid. [n fact, it follows directly from Theorem 7.2 itself. Theorem 7.5. Let the hypothesis of Theorem 7.4 -be satisfied and stppose thal feeC, fyeC, on Dz. Then the solution p defincd in Theorem 7.4 is of elass C' on V,. Proof. Consider the (n + k)-dimensional u space consisting of points with coordinates Ij... ,7) 1,...,4k) and define the vector function # = (Fy, .. . , Fase) on Dy by Pitu) = flttn) G=1,...,0) Pealtsu) = 0 Gel... wan ieneSec. 7] EXISTENCE AND UNIQUENESS OF SOLUTIONS 31 Then, by Theorem 7.1, the system of equations ul = Flu) (7.22) hos for a solution the vector x = (xy i. . » Xa) given by x = olhrén) (= 1... nm) xennlt) = i @=1,...,%) since x has the initial value given by xr) = & (i= 1,. ey) xen(t) =m G1 Thus yi, ... , # may be thought of os forming part of the components of an initial-value vector for the system (7.22), and the F in (7.22) satisfies the conditions in Theorem 7.2. Therefore the first partial derivatives of x with respect to 7, &, and y; exist and are continuous on V,, and from the definition of x this implies the theorem. From elér.bu) = E+ [Seelontadan) ds it follows that ! $2 tent) = f | tortorsta.ad $2 tort + ZL tow(os8u)a) | ds This shows that dy/dy; is the solution of the initial-value problem ¥f = fAlotttudady + 2E yeltrtadan) — ye) = 0 Hypotheses under which the existence of higher derivatives of ¢ with respect tor, £;, or n: can be shown to exist are readily ascertained from the fact that the first-order derivatives are solutions of a linear equation. For,example, 4/9€; is the solution 6; of y = flhelunduuly (7.23) with the initial value ¢. Clearly o%)/d#,06; is 08,/8€;, if it exists. But £ enters (7.23) as a parameter. If 7 and y are held fixed in (7.23), then & in (7.23) plays the role of » in Theorem 7.5. Thus, if f-(¢,o(¢,7,£,4),#) has & continuous derivative with respect to é, then 98/06; exists. If f has continuous partial derivatives of the second order with respect to the components of x, then f.(t,p(t,7,6,4),4) will have continuous first-order partial derivatives with respect to &. In much the same way, if f has continuous partial derivatives of the second order with respect to the components of (x,u), then d%/duduy32 ORDINARY DIFFERENTIAL EQUATIONS [Cuap. 1 exists as do the mixed derivatives d%y/dy,0¢;, ‘The case where the partial derivatives ofp are taken with respect to the components of (7,£,) is left to the reader as an exercise. 8. Complex Systems So far it has been assumed in the equation (E) that ¢,z,f were all real, If f is a continuous complex-valued function on an open connected set D in the (é,w) space, where ¢ is real and w is complex n-dimensional (real 2n-dimensional), then the equation :) w! = f(t,w) is defined to be the préblem of finding an interval J on the real f line and a (complex) differentiable function y on / such that ) (ie) eD = (te 2) i) #4 =Aueto) (te, = 4) It is an easy task to see that all the existence, uniqueness, continuation, and dependence theorems proved in Sces. 1 to 7 are valid for (E;) as well, if one defines the norm |u| of a complex vector w= (wi, . ~~, Wn) formally as before, namely, fel = y wil ied Here, of course, |t,| = (tw)? + (Sw,)*)!, where Nw; and Jw, are the real and imaginary parts of w. Morcover, the Theorems 7.4 and 7.5 con- cerning the equation E,) z= fir) can be extended in an obviows way tothe case where « is a complex parameter vector, if fis defined for complex z and », Linear systems are an important case where the above remarks apply. Usually a function defined on a set of complex numbers that cecurs in n differential equation is analytic. Let / be a vector function defined on a domain (open connected set) D of the complex n-dimensional w space. Then F is said to be analytic at a point weD if in some neighbourhood lw — wl 0, cach component F; of F is continuous in wes (ui... 5 tn) and is analytic in cach w, when all other w, ¢ # k, are held fixed. An equivalent definition is that each F; is representable by a convergent power seriesSec. 8] EXISTENCE AND UNIQUENESS OF SOLUTIONS 33 Foy oes stn) = SD Anges sey + + (oy ~ te) m0 mad in some neighborhood Jw — «| 0. The An,...m, are complex constants. A function F is said to be analytic in a domain D if it is analytic at each point of D. It will be recalled that an analytic function in a domain D possesses derivatives of all orders on D. A basie property of analytic functions is that, if a sequence of analytic functions converges uniformly on a domain D, then the limit function is analytic in D. It is evident that since an analytic function F in D is represented locally by a power series it is locally single-valued, that is, for every point we D thereisa p > Osuch that F is single-valued on |w — ul <.p. How- ever, in the large, it need not be single-valued. For example, the function F given by F(w) = w!, where w has one complex dimension, is analytic in thé ring 1 < || < 2 but is double-valued there. If w' is taken og posi- tive and real on the interval 1 < {tw < 2 and w is followed around o closed path (|w| = 4, for example), then w! assumes negative real values when w ngain reaches the positive real axis. The function F(w) = w*, a real and irrational, assumes infinitely many values in the ring. An important extension of the problem (E) is to the case where 1 may be complex. Suppose that f is an analytic complex-valued vector func- tion defined on a domain D in the complex (z,w) space, where the z space has one complex dimension, and the w space is complex n-dimensional. Phen the equation (Es) w' = f(z,w) is defined to be the problem of finding a domain #/ in the complex z plane and a (complex) differentiable locally single-valued function » [a solution of (E2)] on # such that 0) (ze(z))eD = (2e H) i @asevi (een, = 4) The existence and uniqueness of solutions of (Es) can be inferred from the method of successive approximations. Indeed, suppose f has components fy oss Sn) and w = (wy, . . . , Wn), and fis analytic on the domain Ra: jz-—al 0) which will be called a rectangle, although it ism + 1 complex dimensional. Note that wo is a vector here and not a component.34 ORDINARY DIFFERENTIAL EQUATIONS [Caar. 1 Theorem 8.1, Suppose f ia analytic and bounded on the open rectangle Ra, and let . b Me sip, Went a= min (0 3) Then there exists on |z ~ 2o| < a a unique analytic function y which is a solution of (E12) satisfying g(20) = wo. Proof. Since the matrix f. = (df/dw,;) is bounded on any closed rectangle R, C Rz, it follows that f satisfies a Lipschitz condition on Rs. Therefore one can construct the successive approximations vol2) = wo en) = wet [Gm d &=012,..) BD where the integrals can be taken along a straight line joining 2 to z. Applying the argument in Theorem 3.1, one obtains the existence of a unique solution g on the circle jz — zo| < @ which satisfies (zo) = wo. Clearly go is analytic in z on [2 — zo] < a, and thus the function f, defined by fo(z) = f(z,po(z)), being an analytic function of an analytic function, is analytic on |z — zo] < a. From (8.1) it follows that g; is analytic on lz — 2o| < a, and an ensy induction proves that all the approximations ys are analytic on jz — zol < a. Since the solution ¢ is the uniform limit of the sequence {ys} of analytic functions, it is itself analytic on |z — zo| < a. This completes the proof. Remank: Unless other restrictive assumptions are made on f, the circle of analyticity |2 — zol < a cannot be improved. For a s 6/M, this is illustrated by the case where f is independent of w, and has singularities on the circle jz — zo| =a. Fora > b/M the example w! = fw) = at iE ¢ + “yr where w is one dimensional, illustrates this. The solution ¢ of this equation for which g(0) = 0 (here 29 = wz = 0), is wo =if04sy =I = (many b om = Nin i) Clearly fis analytic and bounded in the circle |w| b/M, the solution y has a singularity in the region b u< kl 0 such that for any (t,w) e U, where Ur peH lw - vO <8 there exists a unique solution p = y(z,t,0) of (Es) on H with y(t,t,0) = o. Morcover, p ia analytic on the n + 2 complex dimensional domain Vi zeH § (Sa)eU Remanx: Actually H need not be convex, It is sufficient if 11 is simply connected and if there is some constant ¢ > 0 such that any two points of H may be joined by a polygonal are of length less than c. Proof of Theorem 8.2. The proof follows that part of the proof of Theorem 7.1 that deals with the successive approximations. The path of integration from ¢ to 2 in the successive approximations can betaken as a straight line if H is convex. In any case, the path can be taken as a polygonal path of length less than c. The argument of Theorem 7.1 carries over with the obvious modifications necessary to meet the require~ ments that the variables are complex. The approximations 9; are all analytic on V. Thus the limit function, to which the approximations converge uniformly on V, must be analytic on ¥. Since » has all derivatives with respect to z,t,0 on V, the equation o'(2,f,0) = fleol2,t,0)) can be differentiated with respect to «;, giving 5 Gade) = felewlesa)) 2 Thus dy/du; is the solution of the linear equation u! = folz,e(z,tw))y (8.2) with initial condition (0y/du)(?,¢,0) = ¢;. Thus the analogue of the main result of Theorem 7.2 is proved. The result analogous to (7.5) follows in much the same way as (7.5). The result here is36 ORDINARY DIFFERENTIAL EQUATIONS [Caap. 1 det ealetin) = exp fi tr fuleo(2,tye)) ds where the path of integration of the integral is along an are in H. - Since dy/df exists, it follows easily that it is the solution of (8.2) with initial value ~f(f,w) at t. The case (E2,) w! = f(z,w,n) where f ia analytic in (z,0,x) and » is & complex dimensional can also be dealt with. Let I, be the domain of » space given by |y — pol <¢, where yp is fixed and ¢ > 0, and let D be a domain in the n + 1 complex dimensional (z,w} space. Let D, be the set of all (z,w,u) such that (zw) e Dand peZ,. The analogue of Theorems 7.4 and 7.5 for (Ex,) is the following result: Theorem 8.3, Let f be analytic in the domain D, and suppose ¥ is a solution of (Ex) for 4 = uo which exists for ze H, where H is a closed convex domain in the z plane. There exists a 8 > 0 such thal for any (hea) e Un, where Uys fe jo dO + lu — wl <8 there exisis a unique solution p = o(z,¢,w,u) of (Es) on H with eC.ten) = 0 Moreover, is analytic in the n + k + 2 complex dimensional domata Vu: ee HH (Syn) @ Uy The proof can be obtained using either the method of proof of Theorem 7A or that of Theorem 7.5. The remark following Theorem 8.2 applies here also. A theorem which is a mixture of the results of Sees. 7 and 8 is obtained when ¢ is assumed to be real and w, a, f complex. Let D be a domain of (t,w) space, where ¢ is real and w complex n-dimensional. Let J, be the set of all p satisfying lz — ol < c for some ¢ > 0, where x is k complex dimensional. Finally let D, denote the set of all (¢,w,u) satisfying (t,w) e D and ye J, Theorem 8.4. Let feC on the domain D,, and for each fixed t suppose f is analytic in (w,u). For up = yo let y be a solution of w' = f(t,w,no) on some interval I: a St = b [thus (4,y(t)) e D for te I) satisfying yr) = wo, wherereI. There exists a & > 0 such that for any (w,x) e U, where Uy: lo = wal + |e — pol <8 there exists @ unique solution p = o(tw,u) of w' = f(tu,uj on I with olrwn) = @Suc. 8) EXISTENCE AND UNIQUENESS OF SOLUTIONS 37 Morcover, p is continuous in (tw,n) fora St S b, (w,n) 2 U,, and for each fixed Le I is an analytic function of (w,x) for (w,n) e U,. The proof is very much like that of ‘Theorem 7.4 and is left to the reader. The uniform convergence of the successive approximations, each of which is analytic in (wu) in U,, leads to the analyticity result for ¢ asa function of (w,s). . An important application of this result is to the case of a linear system involving a one-dimensional parameter « linearly. For example, let A, Bbe continuous complex-valued n-by-n matrices defined on some open real't interval J, and consider the system w' = (AQ) + aB))o Then the vector f defined by f(t,w.x) = (A() + uBQ@))w is continuous on the demain D, given by Da ted — wl + |x| < and for each fixed ¢e J it is analytic in (2,u) for |w] + |x| < ©. Apply- ing Theorems 5.2 and 8.4, it follows that the solution y passing through the point (7,2) (r2J, |v] < ©) exists for all ¢¢ J, is continuous in (t,0,4) for ted, jw] + |u| < ©, and for each fixed te J is analytic in (w,u) for lo} + |u| < o. In particular, for fixed (r,w) y is entire in x. (See also Prob, 7.) A case of great importance where the above is applied occurs in the aludy of boundary-value problems involving a parameter; see Chaps. 7-12, PROBLEMS 1. Let v, ¥, x be renl-valued continuous (or piceewise continuous) functions on a real {interval J:a St $b. Let x(t) > Oon /, and suppose for fe / that vt) 5 v0 + [i xiero(e) as Prove that on J vo 3900 + fi xto)vy exp ( fxtu) au) as Hawn: Let R(Q =f x(e)o(0) do and show that ft! — x2 5 xv. 2. A-function f defined on a domain PD of the real (t,z) space is said to be of class Lip ()-on D if there exists an integrable function & of t such that for all (t,z) and (2) in D Iz) — $6,2)] Ss kOlz — 2 Let feLip (#) on D, Let y1, 2 be two continuous functions on I: St Sb auch that (,es(4)) ¢ D for te I, and f(¢,es(1)) is integrable over J fori = 1,2. Let33 ORDINARY DIFFERENTIAL EQUATIONS (Crap. 1 et) = oe) + [Flee da + BD where re I, and suppose |gi(r) — ya(r)| & 4. Prove, if E(t) = [£1(¢)| + [Be(¢)|, that forr St 5 > tout — esta s dexn [ f* aio da] + 2m + f Bion) exp | fou) du] do and a similar result fora St $1. Huwr: Use Prob. 1. 3, Let the functions gs, yx presented in Prab. 2 be of class Con J, and in addition let leat) ~ Se) S c(O, ef) = alt) + aff). Prove that testy ~ to] 5 Boxp [ fx) a] + fas) exp [fen du] de Hive: 60 feo) as Ke f * (a) ds, then the above inequality yields tent — esto} 5 (0 + fate de) o Clearly the above inequalities can be used to prove uniqueness of solutions of 2 = f(z) iffeLip (t) on D. 4. In the hypothesis of Thoorem 3.1 let the condition (C,Lip) in # be replaced by IFGOL Ss KOO + Ish) Wa) — G2) S ble — 3] for (f,z) and (1,2) in . Assuming f is euch that the integral of f(t,y-(¢)) is defined for any continuous function y, show that there existe an intervals $f Sr +a; (a; > 0) on which the aucceasive approximations converge uniformly to a solution. Hote: Let K@) = f "k(a) de, TEL + (e)(eX — 1) = bfor some fo in the intorval (70), bet ty 7 = ai. Otherwise let ay =a. Show that all the successive ‘approximations ; stay in |x — S$ (1 + |e)(e® — 1) for Le fe, 7 -tan). Show Joa) ~ oat] 5 OAL OY and the condition Lip (1), Reatans: If the above hypothesis on f is true for all z and and for all ¢e (a,5}, and if re {a,6], then the sucecssive approximations converge uniformly on ja,b], ‘This is the case when f is linear in x. 6. Let fe C* on the (t,2,y) set given by 0 SS 1, and all zy. Let ¢ be asolution of the second-order equation 2” = f(t,z,2") on [0,1], and let (0) = a, p{1) = b. Suppose af/ar > 0 for ¢£ [0,1] and for all z,y. Prove that if @ ia near & then there exists a solution y of 2” = f(f,2,2') auch Uhat ¥(0) = a, yl) = B, Hint: Consider the solution 9 (aa a function of (¢,a)) with initial values 6(0,a) = a, 80a) a, Lat e'(0) Sa, Then for ja — as small, @ exists for te [0,1]. LetProns.] EXISTENCE AND UNIQUENESS OF SOLUTIONS 39 ul) = 2 Gas) Then uw ZL gotoeow -2 Got,Ou 0 where u(0) = 0, u'(0) = 1. Because df/dz > 0, u is monotone nondecreasing and thus u(l) = (88/da)(1,a0) > 0. Thus the equation @(1,2) — 8 = 0 ean be solved for a ns a function of f for (a,8) in a neighborhood of (a0,). 6. The following problem shows that analyticity with respect to initial conditions may prevail for the solution of a diflcrential equation with a right member which is discontinuous, (This situation arises in practices when a curve may be replaced by a polygonal line to obtain linearization in each of several regions.) Let F be a real-valued analytic function defined on the n + 1 real dimensional space Rr |isa_ [zt sd Here ¢ has one real dimension, and z is n real dimensional, aud the analyticity of F meang that at each point of 2 it can be represented by a power series convergent ia some domain containing the point, Let the surface Sdefined by F(l,z) = 0, (tz) ¢R, divide R into Ry and Ry such that Fiz) <0 (z)e Ri Fc) > 0 x) e Rs Fez) =0 @z)eS Suppose f is a real-valued vector which is analytic for (1,2) ¢ RU S, and g isa real- valued vector analytic for (t,z)¢ 2. 8, Consider the differential equation (ly x = f(tz) (42) eR, (2) Bes g(x) ize Rr A continuous function g defined over some t interval J contained in |t] $ a igasolution of this differential equation if (to(t)) € R for (2 1, and ¢ satisfies (1) for (t,p(t)) ¢ Ri nd (2) for (I,p(t))€ 2, and if g has only a finite number of points in Sfor¢eZ, (This definition can be generalized considerably by allowing (f,¢(t)) ¢S on one or several # intervals contained in /, but this necessitates some additional hypotheses involving Jiand J; defined below.] Let (t1,21) € Ry and (é,22) ¢ 22, and suppose y is a solution an (f1,t2] such that eh) =m ols) as Suppose (o(t)}¢S fort ry... , tm, Where 0 there exists a unique analytic solution t = +{o,n), where y(t,21) > 7. ‘The solution of (2) is then considered with initial value t = y(o,n),z = ¥(r(o,n),0,)- This initial vatue is analytic in (,n). Because Ja{ri,e(r1)) > 0 this solution will not again intersect S near (r1,¢(r1)). Binee it stays close to (t,e(t)), it can be continued to ts. The solution is analytic in ¢ and in ita initial conditions. The initial conditions are analytic in (en). Thus y is analytic in (¢,,7) for (¢,,n) sufficiently near (t1,f1,21). 7, Let f be @ continuous function defined on a real f interval a S ¢ S band for all complex wand y, where w is complex n-dimensional and , is complex k-dimensional. For each fixed let f be analytic in (w,«) for Jw} + ln] < ©. For all w, i, and ¢¢ [a,b] let, Uwe) ~ £,0,0)| S MUlal)ho — a] HCO S MGal) < © where Af is a monotone Increasing function. (This hypothesis implica f is a linear function of w.) Prove that the solution » of the initial-value problem w’ = f(t,w,»), w{a) = «, is continuous in (0,4) for ¢¢ [a,b) and Jo] + |p| < . Thus ¢ for fixed ¢ is an entire function of (cu). Resane: If the hypothesis is valid for 1 D, where D is a domain in » space, instead of for |u| < ©, then the result holds for we D. ‘Hint: The successive approximations »,, where ge(t,o,u) = «, satisfy letailteen) — eiGon)| S a+ la) Se =a and cach y; is an entire function of (wu) for any fixed ¢, The result follows from the uniform convergence of gj. (The result also follows from Theorem 8.4.} 8. Lot F be a real continuous function of (t,2,y) in a real domain D containing the " point (to,20,ye). Let af/az and oF /ay exiat and be continuous in D, and suppose F(to,20,yo) = 0, (OF /y) (lo,t0,yo) #0. Prove that there exists a unique function 9 [asolution of F(,x,2’) = 0] on some interval containing to satisfying F(,e(t),¢'()) = 0, elle) = 20 oll) = yo. Rewank: The above theorem may fail where F(,2,y) = 0 and (aF/dy)(4,z,y) = 0. A solution of ¥(¢,2,2') = 0 may satisfy both these equations but uniqueness may fail. An example is (2')? — 22’ + 4z = 4f — 1 at (0,0,1) with solutions ¢ and ¢ — &, 8. In Thcorem 7.2 it was shown that 49/2E;, j = 1, . . . , # were solutions of the linear equation y’ = f-(,o(¢,r,£))y with initial values egatr. Prove thatevery solution of this linear system is a linear combination with complex coefficients of these n solutions.” Bince’ ay/dr is the solution with initial value —f(¢,£) at x,,prove that A 2 ane) + Jistea BE (ora) = 0 ‘Hint: If gis any solution and 0(r) = a, then a = Za,e, for some complex constanta ay. Prove that 0 = Ya,(d¢/at,). 10. The following problem illustrates tho abstract idea behind the Picard theorem. Consider a Banach space B (a complete normed linear apace) with the norm of an element y denoted by [yi]. Let 7 be a transformation defined on the set of all ¥Props.} EXISTENCE AND UNIQUENESS OF SOLUTIONS 4) satisfying [lyl| $ & ( > 0) which is such that [7¥l| $4, and 7 satisfics » Lipschits sondition Tv — TH s ky - vil with constant k <1. Prove that there exists an clement » such that Te = ¢, that is, 7 has a fixed point. Moreover, prove that ¢ is unique. Hixt: Define the successive approximations yo, o1, .. . by vo = 0, gir = Tei and using the Lipschitz condition show that flgi+t — vill S 4b, and hence llesam — orl SG eb kien Since ki + +++ 4 két*-1 ig the Cauchy difference for the convergent geometric series Zk, it follows that the sequence {y;] is convergent in B, and hence has a limit veB. Since lies <8, cleacly iivll Sb, and thus Ty is defined. Also [IT — vil SIT — Teil + hes — ell S Ele — esl + less — ell + 0 a8 f+, Unique-” ness follows from the Lipschita condition. 11. Let fe C on the n + | dimensional real rectangle # given by |t| Sa, |zl S 4, and assume f aatisfies the Lipschitz condition Wie) — sa) 6 Ble - al for aconstant & <1in R, Further, suppose [f(t.2)| 5 0/a for (t7}¢ R, The initial- vaine problem 2! = f(é,z), (0) = 0 is equivalent to the integral equation ¢ at) = [Keaton as Let B denote the space of all continuous vector functions y on |f] S$ @ with a norm given by {yi| = mnx |¥()| for |f| Sa, Show that B isa Banach spaco. Let T be the transformation defined for ¥ 2 # satisfying llyi] Sb by Tew = fi 166s) ds Prove that |ITyl| $6 if [ll] $ 4, and that 7 satisfies the condition lity — TYI s ly ~ vil Apply Prob. 10 to obtain the existence and uniqueness of a solution of the initial- value problem 2’ = f(t,z), 2(0) = 0, on | Sa, 12, Let 2" = |af-tz +f sin (e/t), 2(0) = 0. Show that if polygonal approximate solutions are set up as in Theorem 1,1 they need not converge as «~+ 0. Hin: Consider t % 0 and let ty = ks, k = 0,1,2,..., where 6 = (n + 3)°! for some large n. If n is even, show that, the polygonal solution ya(t) satisfies ¢n(5) = 0, en(28) = 54, pa(38) > $64. Once oa(t) A/G, it stays there as long as ¢ < 1/2,000. Indeed, for { 2 43 and as long 23 ga (0) & 41/6, ya(t) > yal(t — 8) 0 > gt — alt > A/10, Since Yett > (d/di)(/6), the reault follows. If n is odd, gn(l) < —U/6 for 38 <¢ < 1/2,000.CHAPTER 2 EXISTENCE AND UNIQUENESS OF SOLUTIONS (CONTINUED) 1. Extension of the Idea of a Solution, Maximum and Minimum Solutions It has been seen that if f is a continuous function in some (é,2) domain D, then the differential equation «) z! = f(z) together with an initial condition xr) = £ (1) is equivalent to the integral equation a) = + [ fle2(e)) da (12) That is to say, if gis. a solution of (E) on some interval J for which g(r) = £, then x = o(¢) will satisfy (1.2) on J, and conversely. Clearly the integral in (1.2) makes sense for many functions f which are not continuous. Recall that the continuity of f guaranteed that a solu- tion of (E) was of class C'. Thus, if a continuously differentiable solution of (E) is not demanded, the continuity restriction on f can be relaxed. Suppose f is a real-valued (not necessarily continuous) funetion defined in some set S of the (¢,2) space. ‘Then one can extend the notion of the differential equation (E) by defining (E) to be the following problem: Problem. To find an absolutely continuous function p defined on a real t interval I such that @ GeO)eS Gel Gi) pO =SeQ) forall Le I, except on a set of Lebesgue-measure zero. If such an interval J and function ¢ exist, then y is said to be a solution of (E) in the extended sense on I. Notice that the absolute continuity of a solution guarantees the existence of y’ almost everywhere on J (that is, except on a set of Lebesgue-measure zero), so that (ii) makes sense. Wf fe C on S, and ¢ is a solution of (E) in the above sense, then from (ii) y’e Con Z, and therefore the more general nation of the equation (E), and of solution y, reduces to the ordinary definition of (E) when feC on S. a28xc. 1] EXISTENCE AND UNIQUENESS OF SOLUTIONS 43 It will usually be clear from the context as to the meaning attached to (E) and the solution ¢, and hence it will rarely be necessary to add the phrase “tin the extended gense.” As regards the existence of a solution of (Z), Caratheodory has proved the following quite general theorem under the assumption that f be bounded by a Lebesgue-integrable function of ¢. The proof will be earried out for the case n = 1 only; it will be clear what modifications are required in the case of a system (E). 2 will denote the rectangle Rk: lt-*sSa [zx-G Sd where (7,€) is a fixed point in the (¢,2) plane, and a and b are positive reat numbers. Theorem 1.1 (Carathéodory). Let f be defined on R, and suppose tt is measurable in t for each fixed x, continuous in x for each fixed t. If there exista a Lebesgue-integrable function m on the inlerval lt — +| S a auch that Gz) Sm@ (ta) R) (1.8) then there exists a solution @ of (E) tn the extended sense on some interval \t ~ 7] SB, (A > 0), satisfying or) = & Proof. The case t 2 + will be considered; the situation is similar when tS. If M is defined by M®)=0 (<1) MO = ['me)d @Sts7 +0) a) then it is clear that M is continuous nondecreasing [m 2 0 by (1.3)], and M(r) = 0. Therefore (¢,¢ + M()) eR for some interval r St § 7 + 8 S++ a, where @ is some positive constant. Choose any 8 > 0 for which this js true, and define the approximations g (j = 1,2,...)} by a=t (rstsrt4) BE; B (1.5) ett f semtords (148 ~, From (1.3), (1.8) fie, Sm@) @ StSr+A) and since f is continuous in x for fixed t, fein) > fe) > ©) for every fixed ¢ in [r, 7 +f]. Therefore the dominated convergence theorem due to Lebesgue may be applied to give im [' sleents)) de =f flo,0(0) de as) a for any tin [7,7 +6]. But eal) = + ['tlevsley ds fi, fowls) de where it is clear that the latter integral tends to zero ask-—» oo There- fore, letting k—> «, and using (1.9), it follows that #0 = E+ [ Kay(o)) de from which the theorem follows at once.Bec. 1} EXISTENCE AND UNIQUENESS OF SOLUTIONS 45 It is interesting to remark that the original approximations (1.5) must converge to a solution in the case where a unique solution isknown. This situation does not obtain for the ordinary successive approximations; see the example in Sec. 3. For the case n = 1 it can be shown that all solutions of (I) issuing from an initial point (7,¢) can be bracketed between two specia solutions, the mazimum and minimum solutions. Let f be defined on the rectangle 2, as in'Theorem 1.1. If yw is a solution of (E) passing through (r,£), existing on some interval I containing r, with the property that every other solution ¢ of (E) passing through (r,f) and existing on J is such that ot) Soult) (te DT) then yw is called a maximum solution of (E) on Z passing through (7,£), Similarly, if g. is a solution of (E) on an interval / for which g(r) = &, and such that of) = volt) te 1) holds for every other solution of (2) on 7 for which p(r) = & then gn is called a minimum solution of (E) on 7 passing through (7,£). Clearly, the functions yx and ¢, if they exist, must be unique. The existence of yy and g, will now be demonstrated under the Carathéodory assumptions. Theorem 1.2. Let the hypothesis of Theorem 1.1 be satisfied. Then there exials a maximum solution gy and a minimum solution ya of (E) on \t — +] S 8 passing through (1,). Proof. The existence of guy on [r, 7 + 8] will be proved. Now any solution g of (2) passing through (7,¢) must satisfy ol = E+ ['Feels) as (1.10) a8 for as it exists, and from (1.10) it follows that les) — of) Sh) — MO) (1.11) for any two points f1, fs where y exists. Recall that Mf is defined by (1.4). Since Af is continuous, (1.11) implies, by the Cauchy criterion for con- vergence, that the solution g can be continued, if necessary, to the entire interval [7, r + 6}, making use of the Carathéodory existence theorem. The details of this argument are entirely similar to those given in Theorem 4.1, Chap. 1. Therefore, all solutions of (E) passing through (r,£) exist on [r, + + @]J, and ali must satisfy (1.11) there. From the uniform con- tinuity of, M on [r, + +- 8], it follows from (1.11) that the set of all solutions {te} of CE) on [7,7 + 4] is an equicontinuous set, that is, given any « > 0, there exists a 6, > 0, independent of t and the solution ¢, such that46 ORDINARY DIFFERENTIAL EQUATIONS [Cuar. 2 1e® — ¢@| 0, a 5, > 0 exists such that, not only is (1.12) true for this 3. but also for 2, fin [r, r + Al, 1Q —@| 0, choose 6, so that (1.12) “and (1.13) hold. Subdivide the interval [r, r -+ A] into n intervals by the points = to n2(te-1), define y, to the left of ts a8 ga: up to the point rps (if it exists) in (t,2,4,-1) nearest 4, such that and fort 21 ultra) = Gr—i(te-s) = Pnna(Te-2) Tf tana exiate,.define 9.() = ga-a(t) for tr St < res. If ro: does not exist, define y. on [tp-fa-1) 88 gor Tf goallet) = Paa(fe-1), define At) = go2(f) on [t-2,4s-1). Continuing in this way, one can define solution y. of (E) on [r, r + 8] passing through (r,¢), obtained by patching together solutiona of (E), and having the property 0530) et) 0. A similar result holde ata. Thus the solution p can be continued up to the boundary of D. Moreover, the same continuation ts valid for a maximum solution pu or a minimum solution ga. The proof is very similar to that of Theorem 4.1, Chap. 1, Corollary to Theorem 1.8. Let the hypothests of Theorem 1.1 be satiafied and let vy and gm, the maximum and minimum solutions through (7,8), extst over (1, r + 8], where 8 S a. Then for any c satisfying ga(r + 8) < € < pu(r + 8) there is at least one solution p through (+,) fors St 2+ Band with o(r + 8) = c. Proof. Start with the solution through (7 + 4, ¢) and continue it to the left. It need not leave the region yn(t) S x S gulf},7 StS7r +8, since it can always be continued back along one of these extreme solutiona if it meets one of them. Thus it can be continued back to (r,£). Theorem 1.4, Let the hypothests of Theorem 1.3 be valid, and suppose the maximum solution oxy of (E) through (r,t) exists over an interval [r,7 +a]. Then there exists a 8 > 0 such that (E) has a maximum solution pu, for each n, § Sy < & + 6 on [7,7 + a] with guglt) = 9. Moreover, eua— eur as 7— E+ 0, uniformly over [2,7 + a]. Proof. By Theorem 1.2, gy, certainly exists over some interval with the left end point 7, ify — ia amall enough. From the definition of the maximum solution, it follows readily that, for 7 > 4 > &, euslt) S parol) & eue(t) Thus yx, ig monotone nondecreasing in 7 and is bounded from below. ‘Therefore, for each’? on some interval {r, r + 8], there exists BY) = gmroll) & eurtt) (1.16)48 ORDINARY DIFFERENTIAL EQUATIONS [Caar. 2 Bince gu, satisfies (1.11), [@(h) ~ &42)| S 184(4) - af Gs)| 80 that ® is continuous. From oul = 2+ f Heeuds) de it follows on letting »—> & + 0 that a) = E+ f' 7le,a(0) de But this implies # is a solution of (E) through (7,2). Thus, by (1.16), (t) = gun(t) over [r, 7 +8]. The uniformity of the convergence of Pate 60 guy follows from the equicontinuity of par, in f, as proved by (1.11), The above argument is clearly valid over the range of existence of on {r, 7+ a]. Suppose that for some to S++ a and for every small Ah > O, © exists over {r, to — A] but not over (r, fo + AJ. Then for any given e > O there exists a 8, > 0 such that lpue(to — ©) — purl — | Se (Lay #OSn-§< 8%. Let the region H be the set. of points (é,z) which satisfy the inequalities Wi-tl Sy fe — euelo-— YS 7+ MW — Mie - y) By choosing y small enough, H CD. Any solution ¢ of (E) which starts on the left vertical side ¢ = ta — y of H [that is, |p(to — 7) — gurr(te — 7)} S | will, by (1.11), remain in H astinereases. Thus any such solution ean be continued to fo + 7. By choosing ¢ in (1.17) so that « = +, it follows that for0 + + a and therefore & exists over [7, + + al. 2. Further Uniqueness Results Considerable research has been done on the problem ‘of uniqueness of solutions of the system (E). The following theorem gives a criterion for ‘uniqueness which is sufficient for many practical cases, and includes as special cases many known criteria. Theorem 2.1. Let ¥ = ¥(tr) be a continuous nonnegative function defined on O0) (2.1) and nondecreasing in r for fixed t there, Suppose that for eacha,0 ) and satisfy there, for t # +, lf.) — $@,2)| S ¥(lt — xh lé — 2) (24) Then there exists at most one solution ¢ « C! of (E) in R on |t — c| S a for which g(r) = & Nore: If ¥ is the function defined on (2.1) by Or) = kr where * is a positive constant, then y satisfies the conditions of the y in Theorem 2.1. Itis an easy exercise to see that for exch a,0 < a 0. Suppose that for each a, 0 < a < a, the identically zero function ta the only absolutely continuous function on 0 S t < a which satiafies and : a'(t) = oto) (2.8) almost everywhere on 0 0. From Theorem 1.1 it follows that through the point (c¢,p(c)) there exists an absolutely continuous function p satisfying the equation PO = HG) on some interval to the left of «. As far to the left of o as p exists, it satisfies the inequality a(t) S p(t) (2.9) for if this were not the case there would exist a point to the left of «, say f, where p(t) = p(g), and p(é) > p(?) for ¢ < ¢, and sufficiently near ¢° (& = o is not excluded). Now, since g, and y: are both solutions of (E} satisfying (2.8), pit) =| f een) — s,ext0)1 at | and for small enough A > 0, vt — 4) =] Puen) — s0,9x(0)1 ae | Since |u| — [of S {u — vl, it follows by subtraction that 20) - pG — 2) S| f°, UGe0) — seen a| [i leno) — Heesto)| at s [i ,vGa a (2.10) using (2.4), From the definition of p one has, since p(t) = p(t), pt) — 0 — = fF veep) at (2.11) where / is new assumed go small that p exists on (t — h, $]. Since ¢ is nondecreasing in r, ¥G2) S ¥(Le) G-kstsy snd this, together with (2.10) and (2.11), implies that p(t — h) SrSec. 2} EXISTENCE AND UNIQUENESS OF SOLUTIONS 51 ptt — h), which contradicts the definition of ¢. This establishes the inequality (2.9). Now p(t) > Oon0 <¢ So, as for as it exists. Otherwise p(¢) = 0 for aome #, 0 < ¢ +0, exists, and by (2.12) lim p(t) = 0 tO Define p(0) to be 0. From (2.12) it follows further that a 0 for any «, 0 0 (2.14) A procedure by which (2.9) ean be proved for the minimum solution pm of (2.14), and which avoids the monotonicity restriction on ¥, is as follows: Consider the problem of finding a solution to r= vbr) te O p(t) for ¢ < ¢ and sufficiently near ¢. At such a ¢, the left-hand derivative p/(¢) exists, and PLO) S lei) — eS ¥GPOD) < Ua) + € = vid) Therefore, for h > 0 sufficiently small, pS — h) < py — hy which contradicts the definition of ¢. This proves (2.15). In the same way, it follows that lim sup pl!) 3 pall) (2.16) Now / plo) = alt) = f" Weed) dt + eo — 9 and hence on « — a S¢ Sg, (a > O), the set {a} is an equicontinuous uniformly bounded set of functions. Therefore there exists a subsequenee {ea} such that p,, tends uniformly {as & — 0) toa function pong —a S # So, where p satisfies (2.14). But by (2.16), this p must be p,, and from (2.15) pul) SP) @ — a St So) The remainder of the proof is the same as the corresponding part of the proof of Theorem 2.1, only replacing p by pn throughout.Sere, 3) EXISTENCE AND UNIQUENESS OF SOLUTIONS 53 3. Uniqueness and Successive Approximations Let 2 denote the (n + 1)-dimensional region R: {f{-rlsa@ jz-e sb (a,b > 0) Let fe C on R, and suppose M = max {fl on R. In See. 3, Chap, 1, it was shown that the successive approximations go, ¢1, ¢2, - - - defined by oot) = & oni) = E+ ['flaen(s)) de (m= 01,2649 converge to a solution » of (E) on the interval pease («=nin(o,4)) and g(r) = & provided that fe Lip on R. The following example (1 = 1) illustrates the fact that the continuity of f alone is not sufficient for the convergence of the successive approxima- tions. Let f be defined by 8.1) 0 (t= 0,-0 <2 < +0) 2t ( ©. The subsequence [ym4i}, which satisfies | omit) = [i feen(e)) de (lt $a) is uniformly convergent on |éj S @ to the function y* defined by vO = [fev ds (I Sa)Sec. 3} EXISTENCE AND UNIQUENESS OF SOLUTIONS 55 for f is uniformly continuous on || S$ a, |z| Sb. It will be shown below that on || Sa, pari) ~ Pal) +0 (m— @) (3.4) Assuming thig, it follows that emi) — em) 0 (k-> &) and this implies g* = y, that is, » is a solution of (E). Because of uniqueness, every subsequence of {ym} which is convergent will tend to the same solution, and this proves that the original sequence {ym} is con- vergent on |t| S « to the solution y. This convergence is uniform on {tl S «, for the set {¢»} is equicontinuous and convergent there. In order to prove (8.4), let w,, be the difference walt) = paul) — eal) (|| S @) and v be defined by v(Q) = lim sup wa (lel S @) me Now (0) = 0, and p is continuous on || $ a, for it is the upper limit of an equicontinuous uniformly bounded sequence of functions. To prove wnt) > 0, (m—+ ©), on |e] S a, is equivalent to showing »(¢) = 0 on \ Sa. This will be done for 0 S$ ¢ S a; the proof for —a $¢ $0 is similar. Asa matter of notation, for any At > 0, and function g defined at and t + At, let Ag be defined by 4g (t) = oft + Ad — g(t) Then from (3.1) it follows that for any ¢ and ¢ + At in the interval [0,«] [avons OLS f'?™ [fleemeals)) ~ Feyem(e))] ds and because of (2.4) [arnt OLS f** ylo,lenn(s))) de (85) Given any 6 > 0 there exists an integer Ni, independent of s and m, such that tbat [wn(s)| S 9(0) +8 (mm > Mi) (3.6) for all ¢ in the interval ¢ Ss ¢#+ At. To see this, note that v is uni- formly continuous, and the set {w,,} is equicontinuous, ont S s S t + At. Therefore corresponding to any 8 > 0, there exists an 4s > 0 such that [o(e) = oO] Me) @8) Define N; to be max Ni, (¢ = 0,1, ... ,k). Then the inequality (3.8) holds form > Na. For a fixed ain (¢, 4 + Ad] there exists an 4; such that Js — s| <<. Applying (8.7) to this s and & = s,, and combining with (8.8), the inequality (3.6) results, Since ¥ is nondecreasing in r, it follows from (3.6) that ¥(2,|2m(a)|) & ¥Ca,e(2) + 8) (mt > Ns) 6.9) and consequently, using (3.5), vm OLS f" vla(e) +8) de (m>M) B10) From the definition of », it is easy to see that [4 (1 lim sup |Au. (| and this with (3.10) shows that lao OL sf" vena) + 8) de (.11) On account of the continuity of y in 7, ¥(a,0(2) + 8) —> ¥(2,0(8)) (as 5 — 0) and from (2.5), and the dominated convergence theorem of Lebesgue, it follows that tim f°" weo(e) + 2) ds =f" yloa(e) de The last relation, along with (8.11), yields [ao | Ss rf ¥(a,v(@)) da (8.12) The inequality (3.12) implies that v is absolutely continuous over any interval in [0,2], and consequently v’ exists almost everywhere on [0,q]. From (3.12) this derivative satisfies Ol S ¥62O) @.13) almost everywhere on [0,a). .Sc. 4] HXISTENCE AND UNIQUENESS OF SOLUTIONS 87 Suppose for somec, 0 <¢ S a, thatu(c) #0. Because of the hypothe- sis on y, there exiats a function p on some intervalo — y S$ ¢ Go, (y > 0), satiafying e'() = 440) ale) = 0) > 0 ag shown in Theorem 1.1. Now on this interval a) S 0) @.14) and from this it follows that p can be continued to the entire interval O 0, there exists an », > 0 such that wal <¢ O8tSn) Hence o() = lim sup |w,.())| S ¢ m2 provided 0 S$ ¢ S$ »,,.or v(t)/t-+0 as 40+. From (3.15) it now fol- lows that p(t)/t-+0 as > 0+, or since p(0) = 0, p,(0) = 0. This contradicts the hypothesis of the theorem, for p is an absolutely con- tinuous function satisfying PO =o) O 0, and D, the set of all (t,z,u) satisfying (4,2) 2 D, wel, Suppose J is a continuous function on D, bounded by a constant M there. For = po let a= firu) 2s) = E (4) have a unique solution go on the interval [a,b], where ze [a,b]. Then there extsta a 8 > O such that, for any fized y satisfying |u — uol < 4, every solu- tion gp of (4.1) exists over [a,b] and as p— yo Pu Oo uniformly over (a,b). Nore: Though (4.1) need not have a unique solution for » # yo, never~ theless its solutions are continuous in y at po. Proof of Theorem 4.1. The proof will be carried out for the case +e(a,b). The result will first be proved over |¢ — | S a for some a > 0. Choose a small enough so that the region R: |t ~ 7| S a, [z — t] S Ma ig in D. All solutions of (4.1) with we J, exist over [ry — a, 7 + a] and remain in R, Let ¢, denote a solution. Then the set of functions {¢,), uel,, ia a uniformly bounded and equicontinuous set in [¢ — 7] S$ a. ‘This follows from the integral equation el = E+ [Holen) ds (tA Ss 2) (4.2) and the inequality |f| S 4. Suppose ¢,(2) does not tend to go(f) for some Ze [r — «, r+]. Then there exists a sequence {ui}, k = 1,2, ... , for which wz — yo, and corre- sponding solutions ,, such that y,, converges uniformly over [r — a, t+a]ask— © toa limit function » but ¥@) golf). From the fact that feC on D,, that yeC on [r — a, 7 + a], and that ¢,, converges uniformly to ¥, (4.2) for the solutions ¢,, yields WO = E+ f'Hevleud ds (t— a1 Sa) Thus y is a solution of (4.1) with » = uo. By the uniqueness hypothesis, it follows that (i) = go(t) on jt —7| Sa. Thusy(%) = ¢o(f). Thusall solutions g, on |t — 7| S @ tend to %% as n> wo. Because of the equi- continuity, the convergence is uniform.Sec. 4] EXISTENCE AND UNIQUENESS OF SOLUTIONS 58 To prove the result over [a,b], a region H similar to that used in Theorem 1.4 will be introduced. The interval [r,) will be treated. Sup- pose that ty [r,b) and that the result is valid for every small h > 0 over [r, to — A] but not over [r, fo +h]. It is clear thatto 27+ a, By the above assumption, for any small ¢ > 0, there exists a 6, > 0 such that lea(to — €) — galt — | go which has been given for |¢ — 7| < a, and is based on (4.2), also applies over [r, f + «]. Thus the assumption about the existence of ty < bisfalse. The case ty = 0 is treated in similar fashion on 4: — y S¢ S$ ts. A similar argument applies to the left of 7 and therefore the theorem is valid over [a,b]. The same reault is true if f satiafies a Carath¢odory type of hypothesis. Theorem 4.2. The conclusion of Theorem 4.1 remaina valid if the hypothesis fe C in D, is replaced by the aseumptions that on D, f is measur- able in t for each fixed y and x; f is continuous in x for each fixed t and p; for fixed t, f is continuous in (z,u) at wp = po; and zm) S mQ where m is Lebesgue integrable over {a,b}. The proof is similar to that of Theorem 4.1 with the usual changes Niecessitated by the Carathéodory type of hypothesis and is left as an exercise for the reader. Theorem 4.3, Let the hypothesis of Theorem 4.1 be satiafied. Then there exists a 8 > 0 auch that for any fixed (o,n,n) satisfying los] +l — ht le — ml <8 all solutions @ = p(t,o,,n) of a= fru) (0) = 9 exiat over [a,b]. Moreover, as {o,n,2) — (1,10), elt,osn,u) > poll) = elt7,E,H0) uniformly over [a,b].60 ORDINARY DIFFERENTIAL EQUATIONS {Cuar. 2 Proof. A proof can be constructed with minor changes from that of Theorem 4,1. An important result in connection with the continuity of solutions with reapect to initial conditions is contained in the following theorem. Theorem 4.4. Let ro, ri(7a < 71) be fixed real numbers, and (ro) a Aized point in the (n + 1)-dimensional (t,x) space. Denote by U's the set of all poinis Po: (70,£) such that IE = ol < bo (bo > 0) Suppose that through each point (t,x) in the region v: nStan |t -— fl 0, R dr/F(r) = ©, If for email ¢ and |x|, [MG2) — SE) S MWR G2 ~ 2), show that if [satisfies the hypothesis of Theorem 1.1 then the solution » of (E) satisty- ing y(0) = is unique. Hint: If ¥(r) = #(0F(), then p'(®) © vital), 0(0) = 0, implies that p() = 0. 2, Show that Theorem 2.1 is valid with y(Gr) = r/t, 3. Let feC (n = 1) on therectangle 0 $1 3 a, jz| & 5, where a,b > 0, and assume Mar) $ {ze if mS zy and f(0) ZO for 0 St Sa. Prove that the suscessivePross.] EXISTENCE AND UNIQUENESS OF SOLUTIONS 61 approximations (3.1) converge to # solution of 2’ = f(z), 2(0) = 0, on 0$¢Sa=min @b/M) where Mf = max [j| on the rectangle. 4. Let fe € on the {n + 1)-dimensional (4,2) region 0 S$ ¢ $a (a > 0), [zl < ~, and suppose ¢ is a solution of the aystem z’ = f(é,z) starting at (0,¢) and existing for OSt<2 0, there exists a positive integer N, such that |4g—4sl N, The sequence {A,,} is said to have a limit matrix A if, given any € > 0, there exists a positive integer N, such that |An — Al N, Clearly {A,,} is convergent if and only if each of the component sequences is, and this implies that {A,,.| is convergent if and only if there exists a limit matrix to which it tends. The infinite series 2 Sa mot 4s said to be convergent if the sequence of partial sums is convergent, and the sum of the series ia defined to be the limit matrix of the partial sums. A particular seriea which is of great importance for the study of linear equations is the one defining the exponential of a matrix A, namely, amet) 4 (1.2) melSec. 1) LINEAR DIFFERENTIAL EQUATIONS 685 where A™ represents the mth power of A. The series defining e4 is con- vergent for all A, since for any positive integers p,q, pte ‘ pta A™ [> mls) math maptl and the latter represents the Cauchy difference for the series el4! which is convergent for all finite |.A|. Also le] S @—1) + ell (1.3) For matrices, it is not in general true that e4+? = e4e%, but this relation is valid if A and B commute. It will be seen in Theorem 4.1 that det e4 = e4 (1.4) and hence e4 is nonsingular for all A. Since —A commutes with A, ef & (e4)-}, Every matrix A satisfies its characteristic equation det (KE — A) = 0, and this remark is sometimes useful for the actual calculation of e+. As a simple example, if 01 A= ¢ 0 then det (AB — A) =? =0, and therefore A? = 0, which implies aA" =0,m> 1. Hence, 1. aaEta=(5 ) If B is a nonsingular matrix, then it will be shown that there exists a matrix A (called a logarithm of B) such that e4 = B. Indeed, if B isin the-canonical form J of Theorem 1.1, it is evident that A can be taken as 4o 0 0 --- 0 ae(O A Orr? o 0 O-::: A, provided that e = Jj,j7 = 0,1,...,8. It ig also easily verified that a suitable Ao is given by log da 0 eee oO Agw{ 9 ber crs 0 Qo O +++ loga66 ORDINARY DIFFERENTIAL EQUATIONS [Cuar. 3 Clearly Jr deus (Bu + x 2) od where Z, is the nilpotent matrix defined after Theorem 1.1. Since large powers of Z; all vanish, the series Y irae !Z} eet has only a finite number of terms, and is thus convergent. Define log (2. + = 2) Neti to be this series, which is, of course, a polynomial in \7},Z;. Thus FOUL) = exp [log (Br, + GHZ, is a polynomial in 4¢},Z;. On the other hand, from Lf = else (ite) =il+ e-pats +h lage. Jae. ja] <1 2 ai 2 ' it follows that, when the right member is rearranged, the coefficients of a*, k = 2, are all zero, while the coefficient of z is 1. This implies the same result for F, and proves that exp [log (Ey, + Xg42s)] = Ey + G43; From this follows readily that a suitable A;,j = 1, ... , 8, is given by 1 Ay = (log dees) Br, + log | Bey + +4) Using the fact that for any matrix M, (PMP) = PM*P") (k= 1,2,...) one readily sees that PeMP-1 = ghr"! From this it follows that the result just sketched for s canonical matrix B is valid for any nonsingular matrix B. Indeed, if J = e4 and B = PJP-, then B = e4, where A = PAP-', Naturally A ig not unique. For example, et mm etetete m gitirie (e = 0, 1, £2,...)Sec. 2] LINEAR DIFFERENTIAL EQUATIONS 67 If # is an n-by-n matrix of functions defined on a real ¢ interval I (the functions may be real or complex), # is said to be continuous, differenti- able, or analytic on / if every clement of 4 is continuous, differentiable, or analytic on J. If @ is differentiable on J, then &’ denotes the matrix of derivatives. Note that if @, ¥ are differentiable (OY)! = BY + dy’ (1.5) and that #°% = ¥#' in general. If #’(t) exists and # is nonsingular at é, then #~* is differentiable at 4. This fellows from the fact that 1) ee ® det & where $ = (¢,), and ¢, is the cofactor of y, in From (1.5) and the fact that S6-? = E, it follows that (oy = -e-H'p- det @ #0 (1.6) Recall that in Sec. 7, Chap. 1, it was shown that, if A is a continuous matrix on a Linterval 7, and & satisfies ’(2) = A(é)#(2) on J, then (det 4)’ = (tr A)(det 4) a7 or, in integral form, det 20) = det) exp [tr At) de re D) (1.8) 2. Linear Homogeneous Systems Let A be a continuous n-by-n matrix of complex functions on a real tinterval J. The linear system (LH) a= A(z (tel) is called a linear homogeneous system of the nth order. It was shown in Sec. 5, Chap. 1, that given any §, and 7 J, there exists a unique solution g of (LH) on J such that g(r) = &t In the following it will be assumed at least (*) holds for A. ‘The zero vector function on J is always a solution of (LH). This will be called the trivial solution of (LH). If a solution of (LH) is zero for any re J, thon, by uniqueness, it must be zero throughout 7. + Moro generally, if each element of A is measurable on [and (*) A@ism®Q el where m is Lebesgue integrable on J, then an application of Prob. 4, Chap. 1, yields the existence and uniqueness of a solution » of (LH) estisfying @(r) = ¢ Bee also Prob, 1 at the ond of this chapter.68 ORDINARY DIFFERENTIAL EQUATIONS (Caar. 3 Theorem 2.1. The set of all solutions of (LH) on I form an n-dimen- sional vector space over the complex field. Proof. If g1, ¢2 are solutions of (LH) and cy, cz are two complex num- bers, then c.g, + ery: is again a solution of (LH). Thia shows that the solutions form a vector space, To show that the space is n-dimensional, a set of n linearly independent solutions yi, . . . , gn must be exhibited such that every other solution of (LH) is a linear combination (with complex coefficients) of these yg Let&,t=1,... ,n, be linearly independent points in the n-dimen- sional x space. For example, each £; may be taken as a vector with all components zero except the ith, which is 1. Then, by the existence theorem, if re J, there exist n solutions gy, 7 = 1, .. . ,, of (LH) such that yi(r) = &. It will be shown that these solutions satisfy the required conditions. If the g; are linearly dependent, there must exist n complex numbers ¢, not all zero, such that > awit) =0 el) This implies that n d capi(t) = 5 ck = 0 is i= and this contradicts the assumption that the £, are linearly independent, This shows that the y; are linearly independent. If v is any solution of (LH) on /, such that y(r) = &, then for some (unique) constants 6 a f= > oiks 1 for the & form a basis for the n-dimensional x space. Hence the function n > cig iow ig a solution of (LH) on J which assumes the value & at 7, and, by unique- ness, this must be y, that is, {A reader unfamiliar with the terminology of the statement of the abave theorem will find that by reading the proof be can readily restate the result in more familiar terms. Sce P. R. Halmos, Finite dimensional rector spaces, Princeton, for » discus- sion of vector spaces,Sec. 2} LINEAR DIFFERENTIAL EQUATIONS 69 z e= y CPi iwi Therefore every solution ¢ is a (unique) linear combination of the g;, and this proves Theorem 2,1. If gi... , en are a set of 2 linearly independent solutions of (LH), they are said to form a basis or a fundamental set of solutions of (LH). If + is a matrix whose n columns are 2 linearly independent solutions of (LH) on J, then + is called a fundamental mairiz for (LH). Evidently # satisfies the matrix equation eH = ANP (de D Qu By the matrie differential equation associated with (LH) on I is meant the problem of finding an n-by-n matrix & whose columns are solutions of (LH) on 7. This problem is denoted by X= AX (tet) (2.2) The matrix 4 is called a solution of (2.2) on J, and & satisfies (2.1). From Theorem 2.1 it is now evident that a complete knowledge of the set of solutions of (LH) can be obtained if one knows a fundamental mairiz for (LH), which is, of course, a particular solution of (2.2). Theorem 2.2. A necessary and sufficient condition that a solution mairiz of (2.2) be a fundamental matrix is that det (4) # 0, for te I. Remark: If det b(t) ~ 0, for some te /, then by (1.8) det b(t) = 0 for all te I. Proof of Theorem 2.2. Let & be o fundamental matrix with column vectors ¢;, and suppose ¢ is any nontrivial solution of (LH). By Theorem 2.1, there exist unique constants c1, . . . , ¢n, not all zero, such that es ¥ Cp; a or, in terms of %, e = bc where cis the column vector with components, . . . ,¢n. This relation is a system of n linear equations in the n unknowns cq, . . . , cn at any re Z, and has a unique solution for any ehoice of g(r). Hence det b(r) ¥ 0, and by the remark above, det ’(t) # Ofor any ¢e 7. Notice that this proves that the column vectors of a fundamental matrix are linearly independent at every ¢¢ I. Conversely, let be a solution matrix of (2.2) and suppose det (2) # 0 for te 7. Thus the column vectors of © are linearly independent at every te J.70 ORDINARY DIFFERENTIAL EQUATIONS [Crrap. 3 A matrix of column vectors may have a determinant identically zero on an interval J, although the, vectors may be linearly independent. For example, let be defined as #0) = (, ) for any real interval J. The content of Theorem 2.2 is that this cannot occur for vectors which are solutions of (LH). Theorem 2.3. if @ is a fundamental matriz of (LH) and Ca (complex) constant nonsingular matriz, then &C is again a fundamental matriz of (LH). Every fundamental matrix of (LH) is of this type for some non- singular C. Proof. From (2.1), if # is a fundamental matrix, PWC = AOHYC el or (@CY = A(#0) and hence C is a solution matrix of (2.2). Since det (PC) = (det #)(det C) = 0 #C is a fundamental matrix. , Conversely, if ¢, and 4, are fundamental matrices, then @, = 4,C for some constant nonsingular matrix C. To show this, let éy'@, = ¥, Then &, = $,¥, Differentiating this equation gives @, = Oy¥' + &%, Using (2.1), this gives Ad, = &,’ + Adv or dye’ = 0. Thus ¥’ = 0 and therefore ¥ = Cisaconstant, It is nonsingular since 4, and #; are. Remarks: If it is only required that ©. be a solution, then C may be singular. Observe that, if # is a fundamental matrix of (LH) and C is a constant nonsingular matrix, then C& is not in general a fundamental matrix. Two different homogeneous systems cannot have the same funda- mental matrix, for in (LH), A(t) = ®(¢)@-(). Hence & determines A uniquely, although the converse is not true. Adjoint Systems. If is a fundamental matrix for (LH), then (Y = -e8'e-1 = -o-14 or, taking the conjugate transpose, (ot) = —Ateet Therefore *-" is a fundamental matrix for the system a= —Ati2 (tel) (2.3) The systen. (2.3) is called the adjoint to (LH), and the matrix equationxc. 2] LINEAR DIFFERENTIAL EQUATIONS val Xt a —A*OX (ted) (2.4) is called the adjoint to (2.2). The relationship is symmetric, for (LH) and (2.2) are the adjoints to (2.3) and (2.4), respectively. Theorem 2.4. If S is a fundamental matriz for (LH), then ¥ ia a funda- mental matrix for its adjoint (2.3) if and only if We=C (2.5) where C is a constant nonsingular mairiz. Proof. Vf & is a fundamental matrix for (LH) and ¥ one for (2.3), then since @*-' is a particular fundamental matrix for (2.3), v=o 1D for some constant nonsingular matrix D (Theorem 2.3). Hence : wd = D* and let C = D*. Conversely, if is a fundamental matrix for (LH) and satisfies (2.5), one has ¥* = C'~! or ¥ = *-'C*, and hence, by Theorem 2.3, ¥ isa fundamental matrix of the adjoint system (2.3). If A = —A*, then since #*~' ig a fundamental matrix of (2.3) it is alsa one for (LH). Hence by Theorem 2.3 6 = *"! C, or oe =C (2.6) . where C is a constant nonsingular matrix. Equation (2.6) implies, in particular, that the Euclidean length of any solution vector ¢ of (LH) is constant, Reduction of the Order of a Homogeneous System. If m oO ) Since AP = PJ, pi, . . . , pp satisfy the relations Ap. = dipy s+) AD = Meo Apes = henPott ADgis = Pest + rersPets APgin = Petrict + dettPetr.78 ORDINARY DIFFERENTIAL EQUATIONS [Crap. 3 APnartt = gg Danrth AD nares = Prone + hossPnmestt ABa = Pa-t + detsPn The solutions ¥; are expressed in terms of the independent vectors p,, pa, .. . , Poin the preceding set of equations. (For another derivation, see Prob. 13.) The variation-of-constants formula (3.1) applied to the nonhomo- geneous system w= Az+b) (tel) (4.11) where A is a constant matrix, gives for the solution ¢ of (4.11) satisfying e(r) = 0, re Z, the following formula elt) = et flere) de = fl etry) ds a The solution gy of (4.11) satisfying g(r) = & where re J, \t| < ©, is given by od = etre fleas ds ed 6. Linear Systems with Periodic Coefficients Consider the linear homogeneous system. si= Az (~2 0 (t+ +) exponentially fast. From (5.6), (w) = (O)e*", and hence the d; may be thought of as the characteristic roots of the matrix ’-1(0)6(w). In particular, if (0) = E, then e*® = (w), and the d; are the characteristic roots of b(w), Since det (0) = Ar + de = exp [tr ACs) da (6.11) it follows that, ifn ~ 1 of the ; are known, the remaining one is deter- mined from (5.11). A real nonsingular matrix C need not have a real Jogarithm; that is, there need exist no real B such that e® = C. Indeed, the matrix of one row and coluunn € = —1 is an example. However, it is the case that if C is real then there is always a real matrix B such that C* = e*; see Prob. 41, The above used in the proof of Theorem 5.1 yields readily that if A(t) ts real in (5.1) and of period w, then corresponding to any real fundamental matrix & there exists a real matrix P of period 2u and a real constant matriz R such that et) = Pe 6, Linear Differential Equations of Order n Suppose ao, a, ... , @, are n+ 1 continneus (complex) functions defined on a real ¢ interval 7, and let L, denote the formal differential operator dn @ Ly = oo ta ae that is, if g is any function possessing n derivatives on J, Lag = ag + age + +++ + ag Further suppose ao(t) ¥ 0 for any é¢ J. Then the-equation82 ORDINARY DIFFERENTIAL EQUATIONS {Cuar. 3 Inp =O (te) [written out as aot) 4+ a(x) + +++ pane = 0, (te Z)] is defined to be the differential equation a(t) ant) a(t) aa(t) * and is called a linear homogencous differential equation of order n, ‘The system associated with this equation (see Sec. 6, Chap. 1) is then the veetor equation 2 2D goog Oped We # = AWE (6.1) where 0 1 0 0 0 0 0 1 0 0 4 0 0 Qo Qo seek 62) ian net Ang Gee || ao ao ao Qo ao Since (6.1) is a linear system with a continuous coefficient matrix A on Z, there exists n unique vector solution ¢ of (6.1) on / satisfying ar) = where re J, [f] < ©. Thus ¢1, the first component of ¢, satisfies er(t) = By, elt) = Ey 2. eM) = be (6.3) Since ¢; is a solution of Lar = 0, it is the solution satisfying (6.3). The remainder of the results so far obtained for linear systems will be interpreted for Lt = 0. If gi, . . . , gn are n solutions of L,2 = 0, then the matrix fod ns? wel Sm 64) ePrD gh PL. ge is a solution matrix for (6.1). The determinant of this matrix is called the Wronskian of L.x = 6 with respect to gi, . ~~ , es, and is denoted by Wei, . - . ga). It isa function of (on J for fixed yy, . 2. , gap its value at {is denoted by W(e1, . . . , gn)(t). From the fact that for a linear system, such as (6.1), dot H(0) = det a(-) exp fl tr Ads @eD one obtains, noting from (6.2) that tr 4 = ~ai/ao,Sega. 6} LINEAR DIFFERENTIAL EQUATIONS 83 1 Gn) (7) ox [ = aus) ds (te 7) (6.5) Theorem 6.1. A necessary and sufficient condition that n solutions ey «+ +, On Of Lot = 0 on an interval I be linearly independent there is that Wey, .-- end) = Wer - - Wen... ed #0 (el) Every solution of La = 0 ta a linear combination with complex coefficients of any n linearly independent solutions. Proof. If gi, ... , @n are linearly dependent on J, there exist con- stants ¢, . . . , ¢, not all zero such that x > cy, = 0 i Thia implies that A Yat =0 (=0,1,...,2-1) it and hence the veetors @; with components y;, yi, .-., gi", (i = 1, « , 2) are linearly dependent on J. Conversely, if the vectors ¢; are linearly dependent, so are the solutions y,, . . . , ga of Lat = 0. From ‘Theorem 2.2 a necessary and sufficient condition that the vectors ¢1, . , Qn be linearly independent is that det #() » 0 on J, where # is the matrix (6.4). But this is just the condition W(g1, . . . , en)(f) #0 on T. By (6.5), if Wipu, . . - , pn)(z) # Oforsomerel, Wy, . . . , ent) # O for any te I. Since every solution vector of (6.1) and (6.2) is a linear combination of x linearly independent vector solutions, every solution of Laz = Oisa linear combination of x linearly independent solutions of Lt = 0. This proves the theorem. Because of the properties exhibited in Theorem 6.1, a set of ” linearly independent solutions of Lat = 0 is called a basis, or a fundamental set, for Lax = 0. Theorem 6.2. Suppose y:,..., ¥n are n functions which possess continuous nth-order derivatives on a real t interval I,and Wi, . . ~ ,a){t) 0 on I. Then there exists a unique homogeneous differential equation of order n (with coefficient of t™ one) for which these functions form a fundamental set, namely — pe WG ou ees sen) (I Wee) 6.0),84 ORDINARY DIFFERENTIAL EQUATIONS |Cuap. 3 Norte: The Wronskian W(z, v1, . . . , ¢n)'is the determinant of the matrix with the firat row consisting of the elements z, gi, . . . , gn and the other rows being the derivatives of the first row up to the order n for the last row. Proof of Theorem 6.2. Clearly Wy, oy... @o) = 0, @ = 1, . ,”), for two columns of thia determinant are equal: An‘ expansion of the numerator, W(x, g1, . . . » es), of (6.6) by the first eclumn shows that (6.6) is a differential equation of order n, and the coefficient of 2 in W(e, ¢1, - - - » @») is just (—1)"W (gi, . . . , on), Which proves the coefficient of x‘) in (6.6) is one. Since W(g1, . . . , on) ¥ 0, it follows from Theorem 6.1 that y;, . . . , gs form a fundamental set for (6.6). The uniquentss of (6.6) follows from the fact that the corresponding vectors ¢, with components ¢;, %, ..., 9%" determine the coefficient matrix (6.2) of the associated syatem (6.1) uniquely. Since there is a one-to-one correspondence between linear equations of order n and linear systems of the type (6.1), (6.2), the proof is complete. If one or more solutions of Z,2 = 0 are known, then using the asso- ciated system (6.1) it follows that a reduction of order can be effected. A more direct procedure is suggested by the following process, which is the variation of constants adapted to L,z = 0. Let Lay, = 0, and set «= ye. Then L,t = 0 yields a linear differential equation of the nth order in y which has y = 1 as a solution, since gy, is a solution of L,x = 0. Thus the coefficient of y in the new equation must vanish. Considered ae an equation in u = y’, itis of ordern — 1. If oe is independent of y; and Lay: = 0, then (y2/¢1)' is a solution of the (n ~ 1)st-order equation in u, which oan, by a repetition of the above, be reduced to an equation of order n — 2, ete, Adjoint Equations. Intimately connected with the formal operator Ly ia another linear operator Lt of order n, ealled the adjoint of L,, given by ts = (1 (S) Ge) + (9 (SS) ay to tae that is, if g is any funetion on J which is such that ag (k = 0,1, . . . ,) has n — k derivatives on J, then Litg = (—1)"(Gog) 4 (—1*-aig)@- + + + ag ‘The equation Liz=0 (eD [written out as (= 1)"Aa() 2) + (— 1) z)Om-M + + + Gal = 0] called the adjoint equation to L.t = 0 on J, is defined to be the problemSec. 6] LINEAR DIFFERENTIAL EQUATIONS 85 of finding a function ¢ (a solution) on / such that dy (k = 0,1, . . . 7) has n — k derivatives on / and satisfying (=1)Mdop) (HL) $s dye = 0 on I. Hfa,e C*-* on 7 and ¢ is a solution of Ltz = 0 with n derivatives on J, then by using the product rule of differentiation Lie = (—Irdop + + and by dividing by (— 1)"do one sees ¢ is a solution of a differential equa- tion of order n of the type considered previously. Consider the special case of an LZ, where go = 1. For the system (6.1), (6.2) associated with the equation Lt = 2 base + sss tae = 0 (6.7) the adjoint system is a= —A*(£ (tel) (6.8) where, from (6.2), Qs+s 0 & ai Oc: O Gar At = -1+-- 0 Gane (6.9) . sone &s o OQsr+ -1 & In terms of components, (6.8) and (6.9) give Ty = Gaty p= Tat Gp tyate (kK=2,...,n) (6.10) “Thus if gi, . . . , gn is a solution of (6.10) for which gf? and (Gnarsrgen) OY exist, one obtains, by differentiating the kth relation in (6.10) (& — 1) times and svlving for gy, OP davn)mY + oo + + (= 1)"dngn) = 0 ‘Therefore y, satisfies the equation Liz = (—1)r2 + (1) ae) ps page = 0 which is just, the adjoint equation to (6.7). ' The importance of Lt is due to an interesting relation connecting La and L3, which is indispensable for the study of boundary-value problems (see Chaps. 7-12).86 ORDINARY DIFFERENTIAL EQUATIONS {Cnap. 3 Theorem 6.3 (Lagrange Identity). Jn L, suppose a, e C** on I (k = 0, J,...,n). If up are any two (complex) functions on I possessing n derivatives there, then Shyu — algo = [uw ( = £) (6.11) where {uv] is a form in (u, vw, ..., u™) and , v’ +g BRD) given by . : tur) = YD van ns (6.12) Titkemal FRORRO Proof. Using the product rule for differentiation, Bul) = (H1)BOr, eying yl DB EE (1m tyBe-ny? form =0,1,...,n. Thus one obtains BLyu = 6 GammUu™ HP fiat al i Miss (Gaomb)U™ + Dante DCD anand? wb Bag mat at + [ 5 (1940 (4-08) ]! proving the result, Corollary (Green’s Formula). Jf the a, in La, and uv are the same as in Theorem 6.3, then for any ty, tre I, ["* (algun — ube) dt = (wollte) — [wo)(s) (6.13) where (wv](t) is the value at t of (wo. Proof. Integrate the Lagrange identity (6.11) from f; to (2. Hf ¥ is a known solution of Liz = 0 on J, the solution of the equation L,e = 0 is reduced by (6.11) to finding a function ¢ on J satisfying an equation of order x — 1, namely, y ¥ (—Hiz®(a,_0)? = constant aL itke mtSec. 6] LINEAR DIFFERENTIAL EQUATIONS 87 The Nonhomogeneous Lincar Equation of Order n. Ona real ¢ interval T, suppose ay #0, a, ..., a, and b are continuous functions, and eonsider the equation Ene = a(x + arte"? + +++ +aQe= db) del) which is defined to be the same as (mn aly =. LO ale) 8 FF ale) = att) ‘This is called (in case b x 0) a nonhomogeneous linear equation of order n. The system associated with this equation (see Chap. 1, Sec. 6) is given by f= AWt+ 6) (tel) (6.14) where A is the matrix (6.2), and 6 is the column vector with all elements zero except the last which is b/ao. Thus the system (6.14) associated with Lz = b(¢) is a linear nonhomogeneous system, and the existence and uniqueness of solutions of (6.14) can be interpreted, as usual, as existence and uniqueness results for Lar = b(t). It is of interest to determine the explicit form that the variation-of- constants formula (3.2) takes for the special system (6.14). Only the first component ¥ = ¥1 of any vector solution } of (6.14) is of interest, since this component ig a solution of the equation Lat = b(t). Theorem 6.4. If 1, . . . , @ 8a fundamental set for the homogeneous equation Eye = 2 + ayer + +++ + ag = 0 (are C on J) then the solution y of the nonhomogencous equation Lye = b(t) (be on I) satisfying Br) =E Geb li< ©) is given by WO = WO + Sno ff (Hien 200) oy ae (618) where fo is the solution of Lax = Ofor which falr) = Eand Wiles, . . - , en) is the determinant obtained from We, . ~~, @n) by replacing the kth column by (0, ... ,0, 1). Proof. By (3.1) the first component y = y1 of the vector solution ¢ of (6.14) for which $(r) = 0 is given by WO = fi runlt,)6(0) de88 ORDINARY DIFFERENTIAL EQUATIONS (Cuap. 3 where ‘yin(¢,8) is the element in the first row and nth column of the matrix (é-"(s). Recall that the element in the tth row and jth column of (2) ia yi", and det 6() = (yr, . . . , on)(). Now the element in the ith row and nth column of &-! is given by Gin Wen -- + 1 Gn) where gin is the cofactor of g{"-” in. Therefore, a Wher, « - - » ea)(8)vinl8) = > er) Wiles, » - «5 en) (s) get where Wi(pi, . . . , @n)(s) is defined as in the statement of the theorem. Thus the solution y of L,t = b(t) satisfying P(r) = 0 is given by wi) = Yoo [ Wee and obviously (6.15) gives the solution satisfying $(r) = &, if fa(r) = & The Linear Equation of Order n with Constant Coefficients. Consider the ease where in LZ, the functions a> = 1, a1, ..., 4, are all constants. Then J may be assumed to be the entire real ¢ axis. In this case, Ina = 20 $ aye") os age = 0 (6.16) has as its associated system : ee} b(8) ds Wey = ede) = At (6.17) where A is the constant matrix 0 1 0 0 0 0 1 tee 0 Aw . : (6.18) 0 0 0 1 —G, On. Gna °° Oy Ag is to be expected, a fundamental set of solutions of (6.16) can be exhibited, and the precise form of these functions depends on the char- acteristic polynomial f(A) = det (AB — A) of the constant matrix A in (6.18). Lemma. The characteristic polynomial for A in (6,18) ts given by SO) =a + adr + oe) tay (6.19) Note that f(A) can be obtained from L,r by formally changing 2 tom.Sec, 6} LINEAR DIFFERENTIAL EQUATIONS, 89 Preof. The proof proceeds by induction. Fora = 1, A = —a;, and hence det (AE, — A) = 4 + a), and therefore (6.19) is true for 2 == 1, Agsume the result forn — 1. Then expand A ~1 0 0 0 0 » -1 . Oo Q det AZ, — A) =| + . . see : 0 0 0 tok oa Go Gat, Ang Oe Ata by the first column, and notice that the coefficient of \ is a determinant of order n — 1 equal to det (AZ,-1 — Ai), where 0 1 O ses 0 0 0 1 ) A= . : : wees 0 0 Oorsr dL ww Ont —On-2 —Gn-a 6° * th, Hence det AZ wa — Ai) = A" + at +e +++ bak. The only other nonzero element in the first column is a,, and the contribution to det QE — A) due to a, is a, itself, since the cofactor of a. is 1. Hence det QE — A) =A" + al +--+ + Gnd -+ as, which was to be proved. Theorem 6.5. Let ds, . . - , d. be the diatinet roots of the characteristic JQ) =e tad 4+ 65+ +a, =O and suppose d; has muliiplicity m; (i = 1, ... , 8). Then a fundamental set for (6.16) is given by the n functions . Bes (k=0,1,...,m—-1ji=1,...,8) (6.20) Proof. The proof can be based on the corresponding result for linear systems with constant coefficients. However, a direct proof will be given hera, which depends upon the fact that, if 4; is a root of f(4) = 0 with multiplicity m,, then it is also a root of the equations f’(k) = 0, . . f™-DQ) = 0, Now clearly Le) = faye and in general Latter) = 143 e) = A tater) = ROY = [709 + ape-naye + KAD pormaye 4 ++ + soe} e90 ORDINARY DIFFERENTIAL EQUATIONS [CHar. 3 From this it is now obvious that, for any fixed i, LA) =0 (k=0,1,...,m-1) thus proving that the functions (6.20) are solutions of Laz = 0. Suppose the functions (6.20) are not linearly independent. Then there exist constants ¢ not all zero such that s omat F eater = 0 fmt k=O or 5 Pde = 0 int where the P,() are polynomials and o S sis chosen so that P, # 0 while P,,4) = 0,7 2 1. Divide the above expression by e™ and differentiate enough times so that the polynomin! P(é) becomes zero. Note that the degrees and tho nonidentically vanishing nature of the polynomials multiplying e-, ¢ > 1, do not change under this operation. Thus there results 2 OdHe™ = 0 where Q;(/) has the same degreo as P,(i) fori 2 2. Repeating the pro- cedure results finally in a polynomial F(t) of a degree equal to that of P.{t) such that F(é) = 0 forall 4. This is impossible, since a polynomial ean vanish only at isolated points. Thus the solutions are linearly independent. 7. Linear Equations with Analytic Coefficients Suppose A is an n-by-n matrix and 6 an n-dimensional vector defined and analytic on a simply connected domain D of the z plane, and let zoe D, Using the method of successive approximations, it is readily shown that the linear system # = AG@)d + @) (7.1) has a unique analytic solution ¢ on D such that (20) =o where || < «. Indeed, let ze D and let C be an arc from zo to 2; whieh lies in D, has a continuously turning tangent, and is of length L. Let the arc lengthSze. 8] LINEAR DIFFERENTIAL EQUATIONS. 91 along G starting from zo be denoted by s. Let the constant K be large enough so that |A(z)| < K and |6(2)| < K forzon@. Let go(2) = & and one) = 0 + [* Aaa ae + [7 BG) ar where the integration is carried out along C so that ¢, is defined on C. It follows readily that l@: — 0] S K(jol + 1)s S KL(lA +0, ..-, ee eel S Kal + Ss AE al +) Clearly these appraisals are valid at all points z in D which can be reached from zo on an arc of length L on which |A(z)| and [6(z)| are bounded by X. This implies that they are valid in any fixed closed region R contained in D. Since each ¢, is analytic in 2, it follows from the uni- form convergence of ¢, that the limiting function ¢ is also analytic in 2. Tt also follows that o@) =o + f' a@em ar tf" bq) ar This proves the result in # and therefore in D. Moreover, all the theorems proved in Secs. 2 and 8, being essentially algebraic in nature, are valid for the system (7.1). Correspondingly, if a;, . . . , @a, b are n analytic functions on D, then the linear equation of order n, awl fp ar(zhwer? + ++ + + ay(z)w = b(z) (7.2) has a unique analytic solution y on D satisfying w(zo) = wi, w'(z0) = wa... , wh (20) = Wn where a, w2, . . . ,@, are any given n complex numbers. In addition, all the results of Sec. 6 carry over to the case (7.2) in an obvious way. 8. Asymptotic Behavior of the Solutions of Certain Linear Systems If the coefficients of a linear system of differential equations tend to constants as ¢—> ©, it is sometimes possible to characterize the behavior of the solutions. In the analytic case, this problem is treated in Secs. 4 and 5, Chap. 5. A real variable problem will be considered here. Simpler cases are treated in Probs. 29 and 35 at the end of this chapter. First consider the example z+ {1 +0) + rQlz = 0 where v ig ao real-valued differentiable function with lim o(¢) = 0, r is92 ORDINARY DIFFERENTIAL EQUATIONS [Cuar. 3 integrable, and fC wola<@ f° pwla has similar behavior with 7 above replaced by —7. ‘This result indicates that r does not affect the gross asymptotic behavior at all, However, the case of) = re (-K Gzhz0Sec, 8} LINEAR DIFFERENTIAL EQUATIONS. 93 jeh ff [Dia 0 it is possible to choose 80 that Bek f." |2RUr)| dr <« Thus, denoting the left side of (8.17) by J(¢), VO s «+ exp[ — f[/ m(9 do] NnOl f" eH REel)| ar Asi , it follows from (8.19) that lim sup W@| Se Since cis arbitrary, (8.17) is proved. Thusif ysis taken as y, the theorem ig proved for the case A + V(é) = A(t). The proof of Thearem 8.1 is a consequence of the following lemma. Lemma. Suppose A and V satiafy the requirements of Theorem 8.1. Then there exists a matriz S(t), which ag t+ © tends to a constant non- singular matriz 1’, such that S(A + V) = AS (8.20) where A(t) is a diagonal matrix with diagonal elements dj(t), j = 1, 2, vee ym. Ast» @, dQ) — j, where the yj are the characteristic roots of A. Moreover, for some to, fC ola < © (8.21) ‘The proof of the lemma will follow that of Theorem 8.1. Proof of Theorem 8.1. Since SQ) > T ast—+ ©, and T is nonsingular, SQ) is nonsingular for all sufficiontly large &. Choose to so large that nut96 ORDINARY DIFFERENTIAL EQUATIONS. [Cuar.3 only (8.21) is valid but. S~'(¢) exista for ¢ 2 to. Then, letting y = S(d)z in (8.1), wo = Ay + (SRS! + S'S-)y (2S te) (8.22) Let & = SRS"! + S'S“, Then, by (8.3) and (8.21), it follows that |A] is integrable, Thus the proof for the special ease of ‘Theorem 8.1 given above is valid for (8.22) so that (8.22) has as a solution 4@,, where sim a4) exp [~ f'r4(s) do] = ox Thus (8.1) has as a solution S'& = yx. Since, ast— ©, S(t) > To, it follows that t ex oxp[~ f'aiads]p. > @) where py is the kth column of 7-', Since AT-! = T-1A(@), it follows that Ap, = usps. This completes the proof of Theorem 8.!. Proof of Lemma. There exists a constant matrix 7 such that TAT = B where B is a diagonal matrixwith diagonal clements 4. Let S = ST. Then it is required that S(A + VS = STA + YTS SB + TVT-YS* S(B+ VS =A where 7 = TVT-', Because P is linear in the elemonts of V it satisfies [CO lPola<@ woaeou and P(«) = 0. Consider the matrix MOQ = B+ VQ - 8 Thus det M(Q,t) = 0 has roots d,(t), where X4(%) = yu; Denote the cofactor of the element mu(d,t) of ALQ.t) by CuQa,é). Let Sut’) = Croll) f) (8.23) Tl G4) inn where the prime on the product denotes that j = iis omitted. Because Cu:Q(),t) tends to the cofactor of the clement in the kth row and tth column of (B — yE) asi— , it follows that 5u(o) = daSxc. 8] LINEAR DIFFERENTIAL EQUATIONS 87 which is 1 if i = k, and 0 otherwise, Let the matrix with elements Su(é) be S(t). Then clearly Also S(o) = B 2» CuOAD,OlOn + a) — W(OBa] = 0 (8.28) a1 fork=1,...,2. Fort + k this is true because a determinant with two columna identical is zero. Fort = & it is true because A,(2) is a char- acteristic root of B + V {and thus also of A + V). Thus (8.23) and (8.24) imply that 3B + 0) = a8 Because S{«) = E, clearly $~* exists for large ¢ and SB + V)E 1 A Finally [CO wos @ (8.25) This follows from the fact that ¢, ia linear homogeneous in the elements % and Xj. The former are absolutely integrable. Thus it remains only to show that the a/ are absolutely integrable. Let F(A,t) = det M(a,2). Then since FQ,(),t) = 0, oF 4 ar KOKOMO + F One) = 0 Because the characterisic roots of B are distinct, (aF/0d)(A.(1),) tends to & nonvanishing limit as i> ©. The term (aF/dt)(A,(1),¢) is linear homogeneous in 9% and so is absolutely integrable. Thus A; ia absolutely integrable and the proof of (8.25) is completed, Clearly S = ST satisfies the lemma. PROBLEMS 1. Let the matrix A and the vector 5 be integrable functions of ¢ over [a,b]. Let Meals ke [oo] s sO where [wowace Lat +6 [a5] and consider the initial-value problem Fm AME +O] 2h) =e Prove that there is unique solution » over (a,b) in the sense that ye C and98 ORDINARY DIFFERENTIAL EQUATIONS [Cuar. 3 oO t+ [Auer det fords on [a,b]. Hint: Use successive approximations. Let yo(!) = E and end E+ f' A@eiedst fold G0 Provo that, iff k(4) da = K(), thon tor ~ es-rf0l (1 + ep HEE 60 that {yj} converges uniformly over [a,b]. If the nbove holds for all b < 6, thon the solution exists over (a,b). The case 6 = o is allowed. A similar situation prevails at the left end point, For uniquonesa, use Prob. 1, Chap. 1. 2, Tn this problom let the norm of a matrix A be defined by [Al = max > lasil int Then [4 + Bl S$ |Aj + (Bl and |4B] § |A{ |B]. Let A be of class C' on [a,b]. Product integration of 2’ = A(t)z is defined as followa: Divide [a,b] into m parts a = to <4 < +++ 0, m can be chosen large enough ao that |Ja(t)| $« Thus (a) = E and 4, is an «approximate solution of z’ = A(i)r. Use this to prove the existence of the fundamental solution #, (2) = EB. 9. Let the matrix A be continuous over (0, =]. A fundamental solution & of z' = A(x is uniformly bounded over (0, ©] and rt tim int # f' te AG) de > — 2 Prove that #~! is uniformly bounded over [0, «]. Moreover, prove that no solution not identically zero can satisfy y(t) Oast—+ 0. Hint: Use (1.8). 4 Consider the differential equation of Prob. 3 and also the differential equation 2 — B()z (Be C on (0,00), 8 solution of which will be designated by ¥. SuppoveProzs.} LINEAR DIFFERENTIAL EQUATIONS 99 i |A@ = BWI dt < © Prova that y is hounded over (0,00). (Ilera y is a solution of x’ = Az.) Have: Use ¥() = (0 + f* 9(e-H(9)(B@) ~ AG))v(e) de and Prob. 1, Chap. 1. &. Show in Prob. 4 that corresponding to any given ¢ there exists a unique ¥ such that o( — ¥() > Oast— a, Hint: Uso ¥() = off) — f * @e-")(B(e) ~ ACV) de. 6. 1f f,” Wolat < «, then any olution of 2' = B()z not identically zero tends ton it different from zero asé—» =, Moreover, given any constant vector c, there is a unique solution y which tends to cas t—+ «, Hint: Use A = O and o(t) = B. 7. Lot aya be continuous and periodic of period w. Lot g and ys be solutions of 2” + ar()2' + aa()z = 0, where (0) = 1, (0) = 0, 9s(0) = 0, 0} = 1. Use the system formulation and show that the multiptiers (or characteristic roota) are solu- tions of Mt — AX 4 B= 0, where A= exlo) + vi(s)and B= exp[— [" aitr ai]. 8 Let a and b bo real constants and p a real continuous funetion of # of period w. Consider 2” + [a + dp()|z = 0. Let wo: and gs be defined as in Prob. 7. Let F(ab) = ei(w) + (wu). Show that F is an entire function of (a,b). Show that if 2 < F(a,b) < 2 then the multipliers are complex conjugate and of magnitude 1 and that all solutions are uniformly bounded, together with their first derivatives, on (=, ©). Bhow that if (a,b) > 2 or F(a,b) < —2 then no solution is uniformly bounded on (- =»). 9. If in the previous problem F(a,b) = 2, show that there ig at least one solution of period w, and that if F(a,b) = —2 there is at least one solution of period 2. 10. Ifin Prob. 8, a # n?for any integer n, a Z 0,4 = 0, and w = », then show that ~2 < F(a,0) <2. From the continuity of (a,b) show that if a » n® and b is euf- ficiently small, all solutions arc uniformly bounded on (— 2, ~). 21, In Prob. 8 let p(@) = cos 2 and consider the case where @ is near 4nt and 6 is amall. This may be formulated as a" + [dn® + ye + 4 cos 2i]z = 0 where is real and y is a small real parameter. Determine the behavior of the curves on which F(ab) = F(4n® + 74, 2) = 2 in the neighborhood of (x = 0, y = 0). Hint: In vector form, # = (A + wPW)E, where As on A) POe (_, con 2 3 The fundamental eolution ¢ which is F att = 0 isan entire function of «and therefore BL) = ett pit) ob aMiald) bo where F(in? + yu, 2) = tr ®(x,4), Show that 410 = ff cternp(eyes ate) de100 ORDINARY DIFFERENTIAL EQUATIONS [Cuap. 3 where ao» ott [i ereptayeteds Attar tule) = [F omPtadet da and thorefore . tr dite) = £ tr [o~4*P(e)e4"] ds = [fw roa ©0 Thus Plant + yay u) = 2 + yt tr tale) + ut tr ale) oo and from the behavior of tr s(x) obtain the result for small y. 12. Give a direct proof of (6.5) by showing that J’ = (—a;/ae)¥ with the use of (6.4) and faz = 0. 13, Lot A bea constant square matrix. Give the analogue of the proof of Theorem 6.5 for the system 2’ = Az. Hint: Let p be a constant vector. Then (e4 - 4) (tp) = OER — Alp ‘Paling the partial derivative with respect tod, iG - 4) tp = tM(EX — Alp + ep Using the above relations, show that if Ap. = ip and Ap: = hips + Pr then oitp: and eMtps + tevtp, are solutions, Goneralize the procedure to include the general result of Sec. 4. 34 Let Lax = x + cycle“) + - + + ays, where the ay aré periodic functions of period won (— s,s), Find the form of the solutions over (~ ©, =). 16. If 1 and y are solutions of 2 + ax((}2' + az(}x = 0, show that ele — este) = coxp[ — fase) de] where ¢ is a constant, If g; is a solution, show that an ds wo foal [sea] ig an independent solution on an interval where w(t} # 0. 16. In 2” 4 ay(t)z’ + as(t)z = 0, make the change of variable s = F(t), where FW) = oxp[ ~ faut) de] and tet ¢= Gio). Show that this leads to PF 4 o(a)e = where g(a) is eNO eos evaluated at ¢ = G(s).Pross.] LINRAR DIFFERENTIAL EQUATIONS 101 11, tat 2 = yoxn (—$ fax(e)dr) in tho equation ia line 1 of Prob. 18. Bhow that it becomes oy aa) e at a -$-$)y ° 18 Let af?) > O and let ae Ct, Consider 2” + at()z = 0. Let s = Fi), where #'® = af), and let f= G{s). Then the equation becomes Powe rea0 a) = ZO Let 2 = y exp (—# [*01(0) de) and show the above becomes £4 41 +b) = 0 where b(e) is ’: JQ ice ~ a5 path= oe) df b is of class C?, the above may be repeated. If a*(¢) is a polynomial) in 4, note that bG@) + 0 na ¢-+ « and, indeed, that b is of bounded variation so that the reault of ‘Theorem 8.1 applies. 19. Show that the conjugates of Wiley, . . . , n)/Wen . . «+ en), as defined in ‘Theorem 6.4, are solutions of the adjoint equation Ltz = 0. Hint: Use the aystem formulation. 20. Obtain the result (6.15) by using the ‘“variation-of-constanta” procedure diréctly. That is, assume Wom cigr bss + Copy where the c; are to be regarded as functions of ¢ subject to the requirement. D def ao @20,1,,..,9-2) if where of denotes the ith derivative of ». Bt. Lots = (4,2) denote the solution of L, = 0 which satisfies the initial conditions zO(e) 0,7 30, 1,... 2 — 2, and x*-(s) ce 1/ao(e), Show that [seoaroee ae i s solution of Linz = b(4) which vanishes with ita first n ~ 1 derivatives at ¢ = +, Comparing this solution with (6.15), show that ‘ Walon, « -. onde) Iba) = » 00 pro waa) 92. Lot a; s C*~‘[a,b} 00 that Lt is defined. Let K(1,2) = f(,2), aa defined in Prob. Qi for s 4. Show that K is of class C*-1[a,b] as a funotion of Gj) and that a°-1K /de™-" haa simple discontinuity, (—1)*/aa(t), ats = ¢ but is con102 ORDINARY DIFFERENTIAL EQUATIONS [Cuap. 3 tinuous fora $2 StS banda St Sas, Then for any p, ¢# Cla,b] the functions u and v defined by 10 ~ [xem da ow = [P rGage as satisfy Lau = p and Lio = q, respectively, and u'P(a) = o(6) = 0, f= 0.0.5 n—1, Thus by Green’s formula, fC co ~ up at = fue) ~ ote) Since u(q) = v'(6) = 0, both terms on the right above vanish and therefore [L102 [en - Romp a =0 Sinco this holds for all p and q, it follows that K(l,s) = H(¢,¢) and the differentiability of K with respect to s follows from that of If with respect to f. 28, If the form (6.12) is written as aml fu] = > Byawthon ike O determine the form of the matrix B = (Bjs) and provo that it is nonsingular for all ée/. In fact, compute its determinant. 24. Hf wis a solution of Lr = Gand vis asclution of Liz = 0, show that [ur](l) isa constant [uv], independent of te 7. Let ys, ..., os be a fundamental set for Laz = Oon F,and let ¥1, . . . » ¥a be asimilar set for Ltz » Oon I. Show that the matrix S = (sj) = (lee) is nonsingular on /, Let S~! = (e7') be the inverse matrixto S. Define K by a Kia) = apes) (0 Se) aee1 Provo that the funetion u given by uo = f RUae)de ten isa solution of Laz = b which vanishes with its first n ~ 1 derivatives atr. Compare this result with Probs. 2£ and 22. 26, If L, = Li, prove that [uel{t) is skew Hermitian, that is, lw]@ = —feel@) What does this imply concerning the matrix B in Prob. 23, and the matrix S$ in Prob, 247Pross.] VINEAR DIFFERENTIAL EQUATIONS 103 26. Let P; be polynomials and d, be constants and n so = Y Pane pet Let m & land X; My J xk; and let none of the Py vanish identically. If o = max (Rd,) then show that Tim sup (e-“'1f())) > 0 t* Reman: This proves the linear independence of the terms P,(t)e, Hint: Case 1, Let tho P; all be constants and the A, = ia, where tho y; are real so that, J) = Y cgetmt Prom. T fim, pf koe dt = 0 prove that lim sup [f(2)| > 0. = Case 2, Let dy = iy; as above and the highest power of ¢ in any polynomial P; be M, Then SO = OAD + PAM be + a) where the f; arc as in Case 1 above and f; docs not vanish identically. Thus, for large f, easy = fo +0(3) and by Case 1 Jim sup (¢“|/@)) > 0 me General Case, Here JO) COPD + emf) Eo + + erfplt) where o) >; > +--+ >,and the f; are as in Case 2 and /, is not identically zero. Clearly, HAY = A) + O(eerter-e for some constant Q. Thus, by Case 2, lim sup [e-*+/(| > 0 ay 27. Consider the system of linear equations wf AWD pe fA where wis an m-dimensional vector and the A; aro m-by-m constant matrices, Defino 4 fundamental matrix for this equation and compute one such. 28. Let f be integrable and let f HrOldt < «104 ORDINARY DIFFERENTIAL EQUATIONS {Cnapr. 3 Prove that 2” + f()z = 0 has a solution » such that lim yp) =1 lim y'@ +0 re ie Prove that there is a solution ¥ such that tim S81 tim yo = we me Hiwr: Use successive approximations on eo = 1+ [7G aflarete ae and vo =e fyoverds ref” reeds whereaischosen so that f-" salt <$. (Therelation of’ ~ v'y = Lean be used as an alternative way to get ¥, once y has been shown to exist.) This problem is a special case of Prob. 35. 29. Let A be o constant matrix and 2 an integrable mutrix such that [moa < » It ia assumed that the canonical matrix J similar to A is diagonal, that ie, J = Jo (In particular, this is always the cnge if the characteristic roots of A are distinet.) If \; is a characteristic root of A and p; is the characteristic vector, so that Apy = Py then prove that we Ant Re has 9 solution y; such that Jim e(Nert =p Gal... ,n) {in other words, for large / the solution acts like the corresponding one for the ease RY) 4 0] Hist: Let j be fixed, let 9; = 0, and let ef = ¥i(Q -+ ¥a(t), where the elements of ¥,(¢) are sums of terns of the form e', Rx 0 and constants AK, and Ky such that IMO! s Kev BO) [Wal s Ket @ 80) Let polt) = etn, and ' - Vaal = stp; + f ¥it — 2) R(a)ya(s) ds — f Yat — 8) R(s)y@) de where a is chosen so Inrge that K+ Ky [wid <3 Let f¥o()| § Kee, tO. Then show Want — vata] s Se BerProvs.] LINEAR DIFFERENTIAL EQUATIONS 105 thus proving the existence of a limit function which is denoted by ¢; and satisfies lostt)} & BK oer vit = Otpy + [Yate — pRO eye) de — f” Yale — sDRGbe(0) de From thia, ees = Alp & BKK, fea Gy| do + KOK: f* RE] ds 5 2KeKierl [* lacy}da + Kalk + KO f” iRoolas which gives the result as #—+ «. 80, Let 2” + (1 + r(0}e = 0, where f.” (oldt < ©. Show that the equation haa solutions ¢: and ¢: such that Tim (vill) — et) =O lim (¢,() — ie) = 0 he we : and similarly for gs with ¢ replaced by —é. Bi. Formulate and prove a remult similar to the above fot 2’ — (1+ r()z = 0. 8B. Let Lyx o 29 + fay + ri(iiz- + ++ + + [an + ra(]2 = 0, where the ay are constants and [Cinoiace @ot...yn) Let the roota of A + ai\-! + - + « + Gy © Oboe distinct and let X;bearoot, Then Iz = 0 has a colution ¢; such that lim (p)') — Mee oO &eO1....2-D saad forj1,2,...,0. Hin: Use Prob. 20, 98. Lat A bo continuous and periodic of period « v= [AO + ROK where F is asin Prob. 29. Suppose the equation y’ = A(®y has n independent solu- tions of the form ep,(), where the p, have period w. Then prove that the given equation has n solutions y; such that lim [ej(Qe"*? — pp] = 0 = Gah...,n) we Hin: The equation y' = A(t)y has as a fundamental solution P(e", where B is in dingonal form, B = Jz. Cleatly, P’+ PB = AP. Let 2 = P(e. Then the equation for z becomes: 2s Bz + PORP: Now Prob. 29 may be used since B is constant and this yields the required result. Note that u’ = Bu hos as solutions e*/e;, where e; is the constant vector with jth row Land all other rows D. 84. Formulate and prove a result similar to the above for the equation Dar = 200 + [ant) + i(Q]zO"Y + + + + ion) + Ole = 0 where the a, are poriodic of period w.106 OADINARY DIFFERENTIAL EQUATIONS [Caar. 3 865. Consider the case z' = Ax + R(J)z, where A ia constant but where now the canonical form of A has, in the terminology of Theorem 1.1, submatrices Js, # & 1, and where 7 + 1 is the maximum number of rows in any matrix Ji, k 31. Then no polynomial multiplying an exponential term in any element of e4' is of degree higher than r, Here the case r 2 1 is considered. (For r = 0, sce Prob, 20.) Assume that fj” el] at < ©, Let ; be # characteristic root of A and let y’ = Ay have & solution of the form eittte + O(etie1) where cis a vector. Clearly, 0 S& Sr. Then show that 2’ = Az + R(t)z has a solution » such that Jim [e@QeM-* — J # 0 he Hur: Let 94 =o. The elements of e4/'-» are sums of terms of the form ets", OSl+mar. Let enero ow Yiltys) + Yalta) where J¥iGa)| & Kreroe ert a BI) 1¥iGés)] & Kaerteiter Gaia) That is, Yi) has all terms for which the exponential factot e+! sctisfies Ray <<. If BA, =o, then Y; has the terms which have as factor a power of f leas thank. The proof is analogous to the caser = 0. In the final step, f ‘involving ¥1 is now written t f af ty 86, Formulate and prove the analogue of the above result for an equation of the nth order, Laz = 0. 87, Formulate and prove the analogue of Prob. 35 for 7 2 1, for the case where A is replaced by a periodic matrix A(). 98, Formulute and prove the analogue of the above result for the nth-order equation Lt = 0. 89, Let A be the n-by-n matrix A =A + 2Z, where Z = (2) and xj = L if j = i+, and 4; = Ootherwise. Show that 4 is similar ta a matrix B = AE + 7Z, where y #0. Hunt: Let P = (pi), where poy = yy, and prove B= P~1AP, 40. Let A be a real a-by-n matrix. Prove that there exists a real nonsingular matrix P such that A = P~!AP has the real canonical form consisting of real square matrices As, ..., As, By... , Bm down the main diagonal. Ench A, has the form S; 0. +++ 02 0, E: S; +++ Or Oy Aya | 0; Br +++ O: Os Or Or +++ Ey Sy where 0; is the 2-by-2 zero matrix, Ez the 2-by-2 unit matrix, and s, = (2 a) ettPross.} LINEAR DIFFERENTIAL EQUATIONS 107 The 2; have the form Ay 0 ++) 00 nr B=[0 1 - 00 O 0 +) 1 wy 41. If Cis a rea) nonsingular n-by-n matrix, prove there exista a real matrix A auch that e4 = C% Hint: Use Prob. 40, and consider the two cases 4; > 0, 4; <0. Note that log (a} + 6)! — tan! (#) S; = exp tan-! (®) log (a7 + a3)tCHAPTER 4 LINEAR SYSTEMS WITH ISOLATED SINGULARITIES: SINGULARITIES OF THE FIRST KIND 1. Introduction In this and the next chapter the linear system w’ = A(z)w (2 complex) q.1) will be analyzed, where A is an n-by-n (complex-valued) matrix with at most an isolated singularity at some point zo, but is otherwise single- valued and analytic near zo. If A.is assumed to have only a pole at zo, certain very specific results can be derived concerning the nature of a solution matrix © of (1.1) near zp. However, there is one general result which gives a qualitative picture of 6 even when A has an arbitrary isolated singularity at zo. Suppose the domain in which. A is analytic and single-valued is 0 < lz — z0| <@co This domain is simply connected. The method of successive approxima- tions, as stated in Sec. 7, Chap. 3, leads readily to the existence of an analytic fundamental matrix in D for (1.1). An alternative procedure is to set z — zo = ef. Then (1,1) becomes = BGw BG) = FAG +e) Clearly B is analytic for ¢ ¢ 5, where D is the half plane, —-0 < 9 < 108See. 1) SINGULARITIES OF THE FIRST KIND 109 log a. Since D is simply connected, there exists a fundamental matrix solution ¥ analytic for $e 5. Thus (z) = Y(log (2 — 20)) is an analytic fundamental matrix for (1.1) for 2a .D. Since log (z — 2¢) is not single- valued, 6 need not be a single-valued function of z. If M is any matrix of complex numbers, let the exponential matrix 2” be defined by gM xe gllousat (1.2) Note that for z + 0, 2¥ is nonsingular for all M, and (2¥)-! = 2. Theorem 1.1. Zf A in (1.1) ts single-valued and analytic in a punctured vicinity of 2, 0 < |z — zol 0, then A may be written as A(e) = (2 — 2¢)* A(z) (2.2) where y is an integer, A analytic for |z — zo| 0, and A(zo) » 0. When » § -1, it is clear that A is then analytic at zo and hence every fundamental matrix of (2.1) is analytic for |z — zo] 0, and A(0) #0. If # is any fundamental matrix for (2.4), it must be shown that in the representation ® = Sz (see Theorem 1.1) S is either analytic or has a pole at z = 0. ‘This will be done by showing that there exists a positive integer m such that 2"S is bounded in a neighborhood of z = 0, and, by a theorem due to Riemann, this implies the result. Let y be any nonzero vector solution of (2.4), and let (p,0) = p(se"), r= |g. Then 98 = 28 (potty gi? ie (0,8) = ae (nett)ef and thus | $2 con | = $2 oe | 5 Aceon “2 But_as was seen in Sec. 5, Chap. 1, with £ in place of p, ar oe l= [| Therefore, if || A(2)|| S$ ¢ for |e| $ 1 <4, ar dp aS O 2 O O auch that if 2 = pe? wals£ 508 2,0 1) the converse of Theorem 2.1 is not in general true, For example, let n = 2, and consider the system w' = (27°C, + Cy)w a) 01 CG =(-3 Gs ( ) (=: ) 0 9 ‘This system has at 2 = 0 0 singularity of the second kind with rank «= 1. A fundamental matrix ¢ for this system is readily seen to be given by where zt 2 If Sand # are defined by s-G i) #-Gi 4)U4 ORDINARY DIFFERENTIAL EQUATIONS [Caar. 4 it is seen that @ = S2*, and from this representation of # it follows that 2 = 0 is a regular singular point. For an equation of the nth order, however, it is possible to give a neces- sary and sufficient condition on the coefficients of the equation for a point Zo to be a regular singular point; see Sec. 5 of this chapter, especially Theorems 5.1 and 5.2. It may happen that, even though the coefficient matrix A in (2.1) has a singularity at 29, every fundamental matrix is analytic at z. In this case, zo is referred to as an apparent singularity for (2.1). For example, consider the system w= li Clearly a fundamental matrix for this is given by @ = 2 = I, which is analytic atz = 0. Notice that det (0) = 0. This is the general situa- tion under these circumstances. Theorem 2.2, In (2.1) let A be single-valued and analytic in a vicinity of zo bul have a singularity af zo. If & is any fundamental matrix, then either & haa a singularity at 29, or det (zo) = 0. Proof. Suppose ® is analytic at 29, and, if possible, det ®(2o) » 0. ‘Then #-' exists at zo, and is an analytic function of z in a neighborhood of zo. Hence ®’6-! is analytic at 2. But #’@-' = A, and this gives a contradiction. 3. Formal Solutions Although Theorem 2,1 gives a qualitative idea of the solutions for a system with a singularity of the first kind at a point zo, it does not give explicit information concerning the matrix P ~ kE in (2.3), or, for that matter, a constructive procedure for calculating the solutions, This will be done in the present section. The ease zo = 0 will be treated; the modifications necessary for any Zo will be obvious. As an example, consider the case , ‘O 1 wi) (1 (“) wh, = 0} \we 2 This leads easily to the second-order equation toy’ — w/z =0. Using the fact that, by (1.15), there is at least one solution of the form 2?(s_ + a + ++), where p, 8, . . . are constants, it follows that p(p — 1)s0e""? + (p+ Upset oo ager h — aer — ts = 0 ‘Thus p(p ~ Wacz?? + [(p + 1)psi — solert + os + +p + +B Dee — silent to = 0Sze. 3] SINGULARITIES OF THE FIRST KIND 5 From this follows az one possible solution =lealaae= aia gt: pa Laem 1, =p Bs = Opag oe = aga 1 *= Giyperiy ** Thus the series * at » CFEFT satisfies w; — w/z = 0. The question arises as to whether the series represents an actual solution or, what is equivalent, whether the series is convergent. In this case, it is obvious that the series does converge. Indeed, it is always the case that a series which satisfies (2.4) formally is an actual solution and this will be proved. [It is not always the case that a formal series satisfying a more general clasa of equations than (2.4) converges. Indeed, the divergent series > eee and ig a formal solution of the second-order equation ety” + Bz — Luo’ +w=0 where w is a scalar.) It is necessary to define the notion of a formal series in sufficient generality to include all actual solutions of (2.4). By a formal (Laurent) series f will be meant an expression of the form p= Scam where the ¢,, are complex numbers, and ali but a finite number of the tn with negutive indices are zero. If Ms dz” g= is another formal series, then f is defined to be egual to g if and only if om = dm for all m. The sum, f +g, and product, fg, of two such formal series are defined by the relations fto= DY lat dade316 ORDINARY DIFFERENTIAL EQUATIONS (Caap. 4 hin fo= Ie Iw = 2 cuts ante beta Note that the sum involved in the definition of h,, is 2 finite one, and hence fg ie defined for all formal series f and g. (If the c_., do not all vaniah for sufficiently large m, then the aum which defines h,, would not be finite and hence need not converge. Thus products fy would not be defined.) If a formal series f is such that c_, = Oform-= 1,2, .. . , then fis called a formal power seriea. The derivative f’ of a formal series f is defined to be the formal series fe » (on + Weng s2” If the fp are formal Laurent series and y; are complex numbers, the finite sum. pe Suz” (log z)* fy = 0 for j + & large bk@O is said to be a formal logarithmic sum. Let a= y ane (log 2)* theo algo be a formal logarithmic sum. The sum p + q and the product pg are defined by proceeding as though the coefficients f, and gj, were scalars. The resulting coefficients. may then be combined, and yield formal Laurent series. Thus the addition or multiplication of formal logarithmic sums results in formal logarithmic sums.{ The derivative of a formal logarithmic sum p is defined by P=) Uh + adnate! + e+ Dien le(log es) = 3.1) ak=0 which is again a formal logarithmic sum. A formal logarithmic sum is said to be reduced if none of the differences wi — u;,t ¥ j,isaninteger. Clearly a formal logarithmic sum can always be reduced. A reduced sum p ia said to be zero if and only if all the coeffi- cients, fz, are zero, A formal logarithmic sum is said to be zero if and only if its reduced sum is zero. Two formal logarithmic sums are anid to be equal if their difference is zero, t Algebraically speaking, o formal logarithmic sum is an clement of the algebra over aaa tag c.nbem Benersted by formal Laurent siris powers of r, anid inteper powers a8x. 3] SINGULARITIES OF THE FIRST KIND 17 A formal logarithmic matrix L ia defined to be a matrix with elements dy(i,j = 1, . . . ,2) which are formal logarithmicsums. The sum, prod- uct, and equality of two such matrices are defined to be the usual formal matrix sum, product, and equality. The derivative L’ of such a matrix is defined to be the matrix with elements XK. Now, to return to differential equations, consider a system having & singularity of the first kind at z = 0, w' = A(z)w 3.2) where A(z) = > "A, ig a convergent Laurent series about z = 0. mend Clearly A can be regarded as a formal logarithmic matrix. By a formal solution of (3.2) is meant a formal Iogarithmic matrix & which satisfies # = A® (3.2) considered as an equality for formal logarithmic matrices. Theorem 8.1. If © is a formal solution of (3.2), then > is an actual solution, that 18, all formal series occurring in © ara convergent in a region 0 < [el 0. Proof. There exists an actual fundamental matrix $ of (3.2) which, by Theorem 2.1, has the form : $ = SeP where P is in canonical form and where 9 is single-valued, analytic for 0 < {z{ 0. If all the 4, = 0, then the equation reduces to the system w’ = z~'Rw which has a fundamental matrix ® = z®, as can readily be checked. The essential effect of the power- series perturbation in (4.1) is to introduce a power-series term in the solu- tion, that is, a fundamental matrix for (4.1) is given by & = Pz, whereSxé. 4] SINGULARITIES OF THE FIRST KIND 119 P is a power series and f is a constant matrix (see Theorem 2.1), Ina special case, # turns out to be the R appearing in (4.1). Theorem 4.1. In the system (4.1), if R has characteristic roots which do not differ by positive integers, then (4.1) has a fundamental matriz & of the form b= P® (0 0) (4.2) where P is a power series P(e) = > P, Po=E (4.3) mao Remark: From (4.2) and (4.3) follows at once the fact that a funda- mental matrix is also given by Sz**, where Ro is the canonical form of FR, and S is a power series with S(O) nonsingular. This puts the solutions in the form (1.10) with the w, analytic in a vicinity of zo = 0. Proof of Theorem 4.1, Jt will be proved that (4.1) has a formal solution of the type (4.2), (4.3), and, by Theorem 3.1, this implies that (4.2) is an actual solution. Since Py = E, it follows that P(z) is nonsingular on |e| < ¢, for some c > 0, and this implies is nonsingular for 0 < |z| < c, and hence is a fundamental matrix in this region. Let J be the canonical form of #. Then there is a nonsingular con- stant matrix 7 such that RT = TJ. J has the form given in Theorem 1.1, Chap. 3. Let the Q,. be constant matrices and let 2) = Q@)zr = (Qo+ 20+ ° +)” (4.4) be a formal logarithmic matrix. Substituting it in (4.1), there resulta (m + V2PQnar = RQ — Ql) + YF RQnet — Qnsil) a mmo ite + 2"C,, (4.5) where Cn = ) AQ, Ds AnQnns For (4.5) to hold, it is necessary and sufficient for RQ = Qo Quill + (m+ YE] = RQni + Cn (m= 0,1,2,...) (4.6) The first equation of (4.6) is satisfied by taking Qo = 7. To satisfy the other equations, it is convenient to treat the matrix equation column by column, Let the columns of Q. be denoted by gf, j= 1,...,120 ORDINARY DIFFERENTIAL EQUATIONS (Cuap, 4 The jth column of J contains two elements which may be different from sero, d; (the jth characteristic root of 2) in the jth row of this jth column and, for j = 2, af in the (j — 1)st row of this jth column, where 3f is either 0 or 1. In what follows 3} is always zero. Taking the jth column of (4.6) yields By + m+ Ulett + afete? = Raa + G=L...,0m=0,1,2,...) (47) where cY is the jth column of C,. The equation (4.7) can be written as [Oy + m + 1E — High, = 0% — atodz? (Fale... njm=0,12...) 48) The cY depend only on q’,! $ j,k Sm. Taking m = 0, the equations (4.8) are a recursive set for 99’, 7 = 1,2, .. . , n, because 4; + 1 is not a. characteristic root of 2. Taking m = 1, (4.8) is again a recursive set for g9, j = 1, . .. , 2, and by an induction it follows that the formal solution (4.4) is determined recursively by (4.8). (It may actually be found column by column, which means that n vector solutions may be found which then comprise the matrix.) Clearly #T- is also a solution of (4.1), This may be written as QT-UT2T-") = P2* where P(z) = Q@)T' = (Tt+2at--)P t= E+ t+--. This completes the proof. The general case, where may have characteristic roots which differ by positive integers, may be reduced to Theorem 4.1 by means of the following lemma. Lemma. Let the distinct characteristic roots of R (disregarding their multiplicity) in (4.1) be a, .. ~ ,(k Sn). There exists a matriz func tion V of 2, nonsingular for z > 0, and linear in z, such thal the trane- formation w = Vw transforms (4.1) into a system for w with the same properties as (4.1), . at (ht 5 hn) 0 (4.9) m=O and where & has the characteristic roots pi ~ 1, py. - + 5 Pee Proof. It will be assumed to begin with that 7 isin canonical form and Ry, 0 Re (i a) where 2; is a p,-by-p; matrix which contains ali the terms involving the Toot p; in R,Sze. 4] SINGULARITIES OF THE FIRST KIND 121 ‘on BE 0 es O ma(@ ee orrg 00 0 «++: » 4f being either 0 or 1. The matrix U is defined by zH,, 0 Us ° x) (4.10) Clearly U is nonsingular for z # 0, and wa (“om ge.) Then tw = Ue implies, by (4.1), @ = (URU — UU + SUA (4.1) But U--RU = R, and after some calculation one obtains Ri-E, 0 = ayy? a im e1U RU U7 on( 0 a) If a (4 4") Ase Man An where A;, is the block in A» of length and width p,, then (4.11) may be written as (4.9), where Re (* ~ Ep, 4x) 0 Ra This & has the required properties. In case F is not in the assumed form, the transformation U can be replaced by 7'U, where T is chosen so that T-'RT is in the desired form. Setting V = TU, -the lemma is proved. Theorem 4.2. The system (4.1) has a fundamental matrix $ of the form = PA 0< [el 0) (4.12) where P ts a power series Pe = ) oP. (4.13) eto and R is a conatant matriz with characteristic roots which do not differ by positive integers.122 ORDINARY DIFFERENTIAL EQUATIONS. [Cuar. 4 Proof. The proof-follows directly by applying successively the above Jemma and finally Theorem 4.1, Indeed, by using sufficiently many transformations V,,i = 1, ... , l, of the type determined in the above lemma, there results finally @ = V, + - - ViPz*, where P(O) = E, and Ris derived from & in an explicit fashion. ThematrixP = V,--- ViP and is thus a power series. 6. The Equation of the nth Order Consider an equation of the nth order 5 Y arnleho =O — (ale) = 1) (6.1) aso where the a, are single-valued and onalytic in a punctured vicinity of a point zp. If any of the a, have a singularity at zp, then zo is called a singular point for (5.1); otherwise 2o is called an analytic point for (5.1). Analogous to the definition of a singular point of the first kind for a system of the first order, one says zy is a singular point of the first kind for (8.1) if 2o is a singular point for (5.1) and the coefficients in (5.1) have the form ay(z) = (2 ~ ze)*bi(2) (kK =1,...,2) (5.2) where the }, are analytic at zo. The equation (5.1) is said to have at most a singularity of the first kind at zp if zo is either an analytic point or a eingu- lar point of the first kind for (5.1). The simplest equation of the nth order having the origin as a singu- larity of the first kind is 12 be Bye tO byrylD oo sp byw = 0 where the b; are constants. This equation is equivalent, in an obvious way, to the equation 2m) sb byzr—tperD oes be bw = 0 which is called Euler's equation, It can be transformed into an equation with constant coefficients by the substitution z = e*, for if @(s) = w(e*), then o fe, de (ei enr = FE (0) Calne = $8 (a) ~ 2 (ey, ote, The transformed equation BO + DOD $+ feb = 0 with constants ¢;, has a fundamental set of solutions consisting of func- tions of the formSeo. 5} SINGULARITIES OF THE FIRST KIND 123 where is a root of the characteristic equation Atal os boy = and k is a nonnegative integer less than the multiplicity of ». ‘The original Euler equation then has a fundamental set of eolutions of the form z*(log 2)* A short calculation shows that the characteristic equation which » satisfies is given in terms of the b; by A@— De Ont YFAA-D ++ A—a24FDnee-- +b =0 It is called the indicial equation for the Euler equation. Another way of obtaining the solutions of the Euler equation ia to observe that if one puts E(w) = 20 dyer byrd fos pb baw LP) = fe where f is the indicial polynomial JQ) =AW—VY ++ A-atY+FAA—N +--+ A n+ Dy moe + ba Therefore z2* is a solution if f(u) = 0. If all the roots Ay, . . . , An Of JQ) = 0 are distinct, then 2, ..., 2* is a fundamental set for the Buler equation. If » is a root of double multiplicity, then ts) = fXn) = 0 But L (3 2) = L(e log 2) = Aue = (/'0) + Gog 2)fayle® and hence z* log z is another solution in this case, Continuing in this way, one can obtain a fundamental set for the Euler equation. This idea ean be generalized so as to yield a fundamental set for an arbitrary nth order equation having the origin as a regular singular point; see Sec. 8. It is to be observed that if zo is a singular point of the first kind for (5.1), then z) may not be a singular point of the first kind for the first- order system associated with (5.1). (See Chap. 1,Sec.6.) Indeed, only in the case where the coefficients a; have at most simple poles at 29 will this be true. However, there does exist a first-order system connected with (5.1) with the property that, if zo is a singular point of the first kind, then zy is a singular point of the first kind for the system.124 ORDINARY DIFFERENTIAL EQUATIONS (Car. 4 Suppose (5.1) has, at most, a singularity of the first kind at zo, and let y be any solution of (5.1). Define the vector ¢ with components »;, +n by Ge = (2 — ag) tp GK=1...,7) (6.8) Then clenrly @—ade = &—-VNetorn (k=1,...,2-1 x (64) (@ = aes = (n— Dyn = Dba atil2) om mel ‘Therefore the vector ¢ is a solution of the linear system wl = Alejo (5.5) where A has the structure 0 1 00 0 0 i 10 0 : 0 0 21 0 A(z) = (e-—2)"") 0 0 03 0 (5.6) 0 0 00 . 1 sb bea tt Gb Obviously (¢ — zo) A is analytic at zo and does not vanish there, and hence the system (6.5), (5.6) has at zo a singularity of the first kind. From Theorem 2.1 the point zo is a regular singular point for (5.5). Since the elements of the first row of any fundamental matrix for (5.5) constitute n linearly independent solutions of (5.1) [see (5.3), (5.4)], it follows that every solution of (5.1) near zo is a finite linear combination of terms of the form (@ — 20)" (log (2 — 20))*p(2) (5.7) where r is a constant (in general, complex), & is a nonnegative integer which cannot exceed n — 1, and p is analytic at 20, p(zo) ¥ 0. Tf every solution of (5.1) can be expressed in a vicinity of zo as a finite linear combination of terms of the form (5.7), where r and p are 1s above, then zo is said to be a regular singular point for (5.1). Thus the above argument proves the following analogue of Theorem 2.1. Theorem 6.1. Jf (5.1) has at most a singularity of the first kind at zo, then zy is @ regular singular point for (5.1). From the result of Sec. 1, it follows that in any case the solution of (6.1) will be a finite linear combination of terms of the form (5.7) but with p having a possible essential singularity at zo, 80 that it is represented by a Laurent series and not necessarily a power series. In case the » cannotBec. 5] SINGULARITIES OF THE FIRST KIND 125 all be chosen as analytic at zo, the equation (6.1) is said to have an srregu- lar singularity at zo. The converse of Theorem 5,1 also holds. Theorem 6.2. If zo is. a regular singular point for (6.1), then (5.1) has at most a singularity of the firet kind at zo. Proof. Suppose the by are related to the a; in (5.1) via (5.2). Here it is not assumed the 6; are analytic at zo, but it is true that the d, are analytic and single-valued in a punctured vicinity of zo. It is clear then that the system (5.5), (5.6) meets the requirements of Theorem 1.1. Since the element in the first row of any solution vector of (5.5) is a solu- tion of (5.1), it follows from this theorem, and (1.10), (1.11), that there exists a solution ¢, of (5.1) near zo of the form ei) = (& — 20)"p(e) where 7 is single-valued and analytic in a punctured vicinity of zp. But since 2p is a regular singular point, this solution must be of the form orl) = (@ ~ 20)'9(@) (6.8) -where ¢ is a constant and g is analytic at zo, g(zo) ~ 0. If ¢ is any solution of (5,1) near zo, and v= ow (variation of parameters), then Y must be a solution of an equation ” > cnnQ}Or = 0 (6.9) and where Comm F Oremiet + (tt + Udmmaip bo” +( aot aye? + () oo” m= 0,1,...,n) 6.10) However, from (5.10), Cn = dag + Ow tos bang? + oP which is zero, for ¢; satisfies (5.1). Hence (5.9) actually is a linear equa- tion of order n ~ 1 for ®. Letting u = 2’, and dividing (5.0) through by ¢:, there results an equation aol D, dea aleur = 0 6.1) mao where del128 ORDINARY DIFFERENTIAL EQUATIONS [Cuar, 4 =e ~ Ay... of a a at (a b+ Nae + +, 4k) (e=1,...,9-—1) (12) The proof will now proceed by induction, . Consider the case n = 1, w! + a(z)w = 0 (6.18) where a; is analytic and single-valued in a punctured vicinity of zo. If the solution g, of the form (5.8) is substituted back into (5.13), one obtains ¢@) (2 — 20)ax(2) = 8 ~ (2 — &) ae) Therefore (2 — zea; is analytic at zo, which proves the theorem forn = 1, Assume the theorem for equations of ordern — 1. Since zois a regular singular point for (5.1), it is also one for (5.11), for (5.11) has as solutions the functions (;/p:)’, @ = 2,..., m), where gy, ..., os aren linearly independent solutions of (5.1), ¢: being the function in (5.8). If the functions (¢;/¢,)’ are dependent, then there are constants c; such that 3 y e(p:/eiy’ = 0. Integrating, there follows the linear dependence of fo2 gi, i= 1,2, ..., 2”, which is impossible. Thus (g;/;)' are a funda- mental set for (5.11). These derivatives (y:/y1)’ ate, by hypothesis, sums of expressions of the type og (e — 29 22) @ — a)*Qog (@ — 29) Fs where a is a constant, b an integer, p(zo) ~ 0, p analytic at zo. By the induction assumption, therefore, the coefficients d, in (6.11) have at 29 at most a pole of order k, Putting & = 1 in (5.12), it follows that a, has at most a pole of order 1. From (5.12) it follows by an induction, and noting that yi/, has at most a pole of order k at zp, that a, must have at 29 at most a pole of order k, thus proving the theorem. [The formula (5.12) is valid for k = n if d, is defined as zero.) If zo is a regular singular point for (5.1), the actual calculation of a fundamental set may be carried out by considering the corresponding system (5.5), (5.6) and then applying Theorems 4.1 and 4.2, If (3.5) is written in the form ws [« @ — aR + 2, @- 2)" ] w where & and the A, are constant matrices, then F is the residue of A at zo. If the dy in (5.2) are of the formSec. 6] SINGULARITIES OF THH FIRST KIND 127 ble) = J ret (b= 1. ym) mao then the characteristic equation of R, det (AE — R) = 0, is calculated to be MA— B+ AH At) +dAA— ++ Q—n+2 Hi bp batcd + beg = 0 (6.14) This equation is called the tndicial equation for (5.1) relative to the regu- lar singular point z). As shown in Sec. 4, the nature of the roots of the indicial equation determines the complexity of the solutions of (5.1). Tf the roots 1, . . . , An of (5.14) ave distinct and do not differ by positive integers, then a set of 7 linearly independent solutions of (5.1) is given by w= (a — 2p = G=1,...,2) where the p, can be expanded in power series convergent in a vicinity of 20, and p,(zo) #0. In more complicated situations, where logarithms are involved in the solutions, the actual labor can be lightened by using methods such as that due to Frobenius, which is sketched in Sec. 8. 8. Singularities at Infinity A function f is said to be analytic at © if it can be represented by a power series J@ = yf which converges for |z| sufficiently large. The function f has a zero of order m at © if en # O and ¢; = 0,7 < m, and hasa pole of order mat o if e~f is analytic at © fork = m but not fork avn(ew = 0 (a(e) = 1) 6.9) amo130 ORDINARY DIFFERENTIAL EQUATIONS [Cuar. 4 have regular singular points at the distinct points 2, .. . , %, ©, and no other singularities, it is necessary and sufficient that the coefficients a, be of the form k ale) = T] @—an)a(2) (k= d,...,n) (6.10) mol where b, is @ polynomial of degree at mast h{k — 1). Proof. From Theorems 5.1, 5.2, it follows that a necessary and suffi- ecient condition for 2, ... , 2 to be zegular singular points for (6.9) is that the a, be such that the b, = [] (2 — zn)*a, are analytic for all mel finite z, From Theorem 6.2, a necessary and sufficient condition that z= © bea regular singular point is that é, = z*a, be analytic at 2 = ©, Therefore, z = © is a regular singular point for (6.9) if and only if k by = 2 [] (= en)May (6.11) meat where &, is analytic at z = ©, and by analytic in the finite part of the zplane, But (6.11) is equivalent to k by = Zhen 1] (1 - a) a and since Il (1 — 2n/z)*ds is analytic at 2 = , this can hold if and only mel if bs is @ polynomial in z of degree less than or equal to h(k ~ 1). This proves the theorem. 7. An Example; the Second-order Equation The previous material will be illustrated by the case of the second- order linear equation wo’ + f(z)w! + (zw = 0 (7.1) In order that the distinct points z;,..., 2, ©, be regular singular points for (7.1), by Theorem 6.4, it is necessary and sufficient that t f= Tl (z — 2m)~f, where f is a polynomial of degree at most i — 1 cot and g is as below. Hence, f may be expanded by partial fractions k f2) = » @ s i) (an constants) (7.2)Sze. 7} SINGULARITIZS OF THN FIRST KIND 181 Similarly, & & ben on a) = » e=ayt data 7) where ba, ¢. are constants. It is easily seen that in order for 2%g to be analytic at z = ©, it is necessary and sufficient that > em = 0. el In the cage that (7.1) has two regular singular pointe, say at z = 0, gs om, then & = 1, and (7.1) becomes ate" + ayzw! + byw = 0 where a1, b, are constants. This is an Euler equation considered in the introduction to See. 5. Suppose (7.1) has exactly three regular singular points at z,-= 0, a; = 1, and atz= «©. Then (7.1) has the form (see (7.2), (7.3)] at(z — 1)*o" + (az + dele — Iw’ + (2? +dz+e)w=0 (7.4) where a, ... ,eateconstants. It ia usual to consider (7.4) in a normal- ized form. The indicial equation for (7.4) relative to z = 0 is given by {see (5.14)] AA — 1) -— Ab +e =0 (7.5) and the indicia] equation relative to z = 1 is AAD +AMG +b) + C+d+e) =O (7.8) Let r be a root of (7.5) such that (7.4) has a solution of the form Bp ayztt es 7.7) That there is always such a solution follows from Theorem 4.1 (or more directly by the considerations at the beginning of See. 8). Let s be a similar root of (7.6). Let. = wz*(z—1)~. Then the differential equation for @ obtained from (7.4) must have the same form ns (7.4) itself, since the substitution takes all analytic solutions into analytic solu. tions, except possibly at z = 0, 1, or ©, and preserves the regular singular character of solutions at z = 0, 1, and ». Moreover, since the @ equa- tion has corresponding to (7.7) a solution 1 + ez + - - + , it follows that zero is a root of the indicial equation at z = 0, which corresponds to (7.5). Thus the constant corresponding to ¢ must be zero in the @ equa- tion. A similar result holds for the constant corresponding toe +4 +6 in (7.6). With the above substitution carried out, then the equation (7.4) will have the form132 ORDINARY DIFFERENTIAL EQUATIONS (Cuar. 4 2(z — 1)w" + (az + b)w' + ew = 0 and in terms of new constants a, 8, 7, this has the form a(1 — zw" + [y — (2 +8 + 1)2lw’ — afw =0 (7.8) ‘This is the hypergeometric equation whose theory has been investigated in oetad ay let ¢ = 6z,(t) = w(t/s). Then (7.8) is transformed into the following equation for @: rt ot [y-r- 24D] a- eo =0, (-4 79) Now (7.9) has regular singular points at ¢ = 0, 8, ©, and if B— © formally in (7.9), what resulta is $0" + (y — $)0! — ot = 0 (7.10) ‘This has { = 0 as a regular singular point, but now | {= © is anirregular singular point. There are no other singular points for (7.10). The equation (7.10) is one of the forms of an equation which for obvious reagons is called the confluent hypergeometric equation. 8. The Frobenius Method The generalization to arbitrary nth-order equations of the second method of obtaining the solutions of the Euler equation (Sec. 5) is called the Frobenius method. If the origin is taken as a regular singular point, the nth-order equation assumes the form zmwte) gIbywO-D + oes + daw = 0 (8.1) where the b, are analytic in a neighborhood of the origin. Let L(w) = 2m + ethyl) 4. + os + Baw and bz) = ) but = G=1,...,n) he ‘The indicial equation associated with (8.1) is =D Wat) +baA— N+ A—at2H4-- + bout od + dao = 0 Let f(\) denote the polynomial on the left of this equation. If for jol,...,nitis true that +E. T. Copson, An introduction to the theory of functions of a exmpler variable, New York, 1935, chap, 10,Src. 8) SINGULARITIES OF THE FIRST KIND 133 be =O (k= 1,2...) (8.2) then (8.1) becomes the Euler equation. It was seen in this case that L@) = fa)? and 2 was a solution of L(w) = Oif f(A) = 0. In the more general case (8.1), one tries to find a formal series ap f =f vl) = 2 & et (@=3) such that L(y) = fO)2 This is the basic idea behind the Frobenius method, ‘The formal series for y substituted into L yields Lg) = f)2 + 1A + Yer — git s+ +O +dq—- ale + +++ (3) where the g; are linear in c1, . . . , ¢-1 With coefficients that are poly- nomials in}. The equations fa+Da=9 G=1,2,...) (8.4) form a recursive system which can be solved for ¢1, ¢;, . . . , as functions af A, except possibly at the zeros of f(\ +7). Clearly the ¢; thus deter- mined are rational functions of 4, and (8.3) becomes L(g) = fy (85) Tf: is a root of the indicial equation f(k) = 0 and fi +3) £0,721, then from (8.5) it follows that ¢ is a formal, and therefore an actual solu- tion of L(w) = 0 which will be denoted by ¢. Consider the relation (8.5) near \, and differentiate both sides with reapect tox. This results in Ze) = 00) + (log 20))2 and if one takes into account the formal commutativity abe) _ (a on A, formally one obtains 1 (24) = (0) + dog ayonne @e)134 ORDINARY DIFFERENTIAL EQUATIONS [Cuay. 4 If A, is a double root of f(A) = 0, then f(A.) = f’(A1) = 0, and (8.6) shows that 3¢/8A evaluated at A = A: is a formal, and hence an actual, solution of L(w) = 0. This solution is (log z)e1 + v2 gaz) = ) (2 hea, If X1 is a root of multiplicity m, it is readily seen that m — 1 differentia- tions with respect to \ may be carried out to yield m solutions, In case Ag is also a root of f(A) = Gand A: — Ay = kis a positive integer, then the above argument cannot be used for the root A: since f(A: +3) vanishes when j =k. Let fA: +3) #0 for 1 Sj k. Let m be the multiplicity of A; as a root of f(4) = 0. Consider now the formal series (2) = (A — Aah + cet! + ett - ++ ‘Then the same procedure which gave (8.5) now gives Dp) = $A) — a)"2 en Moreover, the equations (8.4) now yield ¢;, ¢3, . . . , ¢x_1 with (A — As)" asafactor. However, for c, the equation is LO + bee = ge and not only is (A — As)" a factor of g, but also of f( + %). Thus ¢ is determined as a rational function of \, and it does not have \; as a pole. The terms cj, j > &, are now readily obtained and also will not have ds asa pole. ‘The series for y now has (A — As)" as a factor of its first k terms but not necessarily of the later terms. If \ ia taken as Az, then (8.7) shows that ¢ ig a solution, However, the first & terms of ¢ vanish so that y can have only 2™ as its leading term in z. Indeed, the solution found in this way is merely a multiple of ¢: found above. - ‘To find a solution really associated with the indicial root As, the mth derivative of (8.7) with respect to \ is considered. This is where L a) = mYQ)e +1 3) where J has} — A: as a factor, Letting \ = As, there follows Lents) = 0Bre. 8] SINGULARITIES OF THE FIRST KIND 135 where gay is O7~/9A" at A = Ax, The leading term of gus: is mle and thus a solution different from any associated with the root A; has been found. Note that in gn,; the powers 2, 7 = k, may occur multiplied by powers of log z of order up to m. Té f(A) has Ag a8 a multiple root, then higher derivatives of » with respect to d will clearly yield further solutions. ‘The procedure for the case of three roots \1, Aa, As differing by integers is left as an exercise, as is the general formulation of the method. PROBLEMS 1, Consider the syatem Pw + 2B wOd +... + Baw = 0 where the B; are m-by-m conatant matrices and w is an m-dimensional veetor, Caleu- late a fundamental set for thia system, @ Treat in detail the system: srw) EB wD of os of Baw 20 where the B; are analytic (near the origin) m-by-m matrices and w is an m-dimensional ‘vestor. 9. Suppose (5.1) has at most a singularity of the first kind at ze. Let z — zy = e', and then find the system associated with the transformed equation, Show that it has ‘the form @ ~ au’ = A@w Ae) = Aa + (@ = addr + @ — a te Compute the characteristic equation of Ao and show that it is the aame as the indicial equation (5.14), 4. Consider the second-order equation where O wl" + feu! + g(z)w = 0 What conditions on f and g must hold if # is to be an analytic point for (*)? Show that if f and g are not both identically zero (and are analytic throughout the wholr plane minus the origin) and « is an analytic point, then the origin must be a singular Point for (*), Discuss the possible nature of the singularity at the origin, 8. Bhow that oy ep geil, Mat OO HD Fesino ttt t Tagen Mt is a solution of the hypergeometric equation, Show from (7.8) that as a function of 3, P(a6; ¥; 2) can have singularities only at 2 > Land @, 6. For appropriately restricted ranges of 8, y, and x show that FOTO!) repivn) = [ema = grein — aye dt136 ORDINARY DIFFERENTIAL EQUATIONS [Cuap. 4 %, In (7.10) let w = phre-Ht. Show that the equation for w is of the form aw 1 k,i-m 7a +[-dets ¥ Je-0 where m = $(y — 1) andk = jy — a. (See Prob. 17, Chap. 3.) &. The Bossel equation is wtive(i-3 wad w+ (14452) a eo 8. Show that, if w = 2°v, then the Bessel equation becomes: a" + Qn + Vo +e = 0 40. Find two series solutions for the Bessel equation valid for small |s| in the case where 2 ia not an integer. 11. Classify the singular points of the Legendre equation (Ll — 2*)w" — eu! + nin + Iw 0 and tho associated Legendre equation It w = ztu, show that a ~ sur" — dev! +[ nin ++ realeeo 12. The regular singular point is related to the equation with “nearly” constant coefficients considered in Probe, 20 and 35, Chap. 3, Show this for the regular point tz = © by transforming dw ‘Ao, Ar #-(44+44---)0 to Be (dat Aue + + +e 43. Let A(e) = R/z + do + Are +--+ >, where R, As, . . . are constant equare matrices, Lot ¥ denote the formal sorica tez* + e:24t! +--+, where the g; are vectora. Show that ¢;, 3, . - . can be chosen as rational functions of \ 40 that [2% - 4] v= OE ~ Rye Asin the Frobenius treatment of the nth-order equation, show that, if 4; is a character- istic root of Rand \: +3, 2 1, is not a characteristic root, then choosing A = A: and 89 83 pi, where p, is the characteristic vector Rp ops ¥ becomes an ectual solution, yx, of w’ — A(s)w = 0. If, in the above problem, M1 is a multiple root and if Ror = Mp tp $= 2,3,...50Pross.] SINGULARITIES OF THE FIRST KIND 137 a further solution can be abtained by considering [zz - ae BY mw (AE — Rays" log 2 + 90-3 Hf, in the above, so is taken as p; and Aas ds and ay//2A is then denoted by yi, it followa that [eg -4@] n= pen (Note that ¢, contains series with log z ag a factor.) From the equation [44 - 4e] = QE = Ryne! it now follows on patting so =p: and \ = Ay and culling y, va that [et - 4@] fr = Oak — Rypah-t mw — ph Thus yi -+ Ua is a solution of 0 — A(s)w = 0, Extond the above procedure to the ease where f > 2. Let \; be a characteristic root such that dy — ds = k ia a positive integer and Ma $5, 1 Sj &, ia not a characteristic root of R. Show that, if hia a roat of multiplicity m of det (AB — 2) = 0, then replacing #a by #o(A — 2)" in ¥{2) Jeads to the determination of a solution with leading term ss,CHAPTER 5 LINEAR SYSTEMS WITH ISOLATED SINGULARITIES: SINGULARITIES OF THE SECOND KIND - 1 Introduction According to the classification of singular points for linear syatema given in Chap. 4, the point z = 0 is a singularity of rank p if the system is of the form wo! = 2B(z)w (1.1) where B is analytic at ¢ = 0, and B(0) + 0. This chapter will be con- cerned with the study of the behavior of solutions of linear systems in the neighborhood of a singularity of the second kind, that is, where p is a posi- tive integer. It will be convenient to consider this singular point at 2 = © instead of at the origin. In this case, the system to be considered is the one induced by the substitution z = 1/f (see Chap. 4, Sec. 6), which has the form, after relabeling, w! = A(z) (1.2) where r ig a nonnegative integer, and A is analytic atz = ©, A(o) #0. It turns out that the study of (1.2) with r 2 0 is much more complicated than the study of (1.2) with r = —1, the case of a singularity of the first Kind atz = ©, Although it ia not easy to prove in general (only a apecial cage will be considered here), there do exist “formal” solutions of (1.2). The real difficulty now enters because there is no analogue of Theorem 3.1, Chap. 4. This was demonstrated by a simple example in Chap. 4 which showed that a formal solution of (1.2) may actually be a divergent, series, It was apparently Poincaré who first realized that even these “formal” divergent expressions have a meaning. He showed, for the case of an nth-order equation, that corresponding to the formal solutions actual solutions of (1.2) exist which have the forma) solutions as “ asymp- totic expansions.” These facts will be made more precise in what follows. The following example will give some indication of the method to be used in this chapter. The equationt {The first-order system associated with (1.3) is given by 2’ = A(t)z, where 138Suc. 1] SINGULARITIES OF THY SECOND KIND 138 att (: - s)2 =0 (1.8) where ais a constant and tis real, behaves for large ¢ almost like the con- stant coefficient case with a = 0. This fact, combined with the results in the case of the regular singular point, suggests trying as solutions for large ¢ elt) = c(t + ett? + oft + +) (1.4) and a similar expression with 7 replaced by —¢. The use of (1.4) in (1.8) formally leads toe = 0 and to an =f 5B) 6 (k 2 0, c= 1) (1.8) Unless a = m(m + 1) for some integer m, the « form a nonterminating “sequence with |e.4:/e,|-> ©, k-+ &, Thus the series in (1.4) is diver- gent for all ¢ #0. However, since (1.4), with the ¢ given by (1.5), formally satisfies (1.3), it will be called a formal solution of (1.3). If two distinct formal solutions of (1.3) were truncated, that is, the infinite series replaced by finite sums containing the early terms, it might be expected that the second-order differential equation satisfied by these truncated functions deviates from (1.3) only in terms involving large powers uf 1/t. In this way, an equation “close” to (1.3) could be found with the help of the formal solutions. For this example, however, this refined procedure will be omitted. The equation 2” + 2 = 0 is close enough to (1.3) for the purpose of getting a representation of the actual solutions of (1.3). The equation (1.3) can be written as wera he (1.6) If (1.6) has a solution g which acts like ec as + », the variation of constants suggests that of) = of — a fain ¢ — Hols) dr 7 Indeed, if y is a continuous function which is uniformly bounded as é—> 0», and satisfies (1.7), then a direct calculation shows » must satisfy (1.6), and (1.7) shows that elt) — e# +0 (E> @) 40= (42-4 x and i thie contre for empl tthe pn ath type (1:2) fore 0,140 ORDINARY DIFFERENTIAL EQUATIONS (Cuar. 5 To show that (1.7) has a solution, the successive spproximation procedure volt) = 0 : “ (1.8) gui) =e! ~ a [sind aoaleltdr (x 2 0) ean be used. Clearly le) — go] S 1 and an induction shows that each of the integrals on the right of (1.8) exists for] Sf < ©, and lend ~ ofl s HE @eoist, 2 AQ) = > rtAy (2.3) ro . Assume Ay #0 has distinct characteristic roota \1,..., Aw Then there exists a formal solution matrix for (2.2) of the form = Prtee (2.4) where P is a formal power series in 2}, P=J rtp, det Ps x0 kn Ris a diagonal matrix of complex constants, and Q iz a matrix polynomial Qa teat +20 @8) with complex diagonal matrices Mig ree 0 a=(® 9 @=G1...,9 @6) Oo ree Qe, OS? = 2) as coefficients, Remark: The simplest case of a system with a singularity of the second kind at © is the syatem w = Aw where A is a constant matrix. A solution matrix & is given by @ = et4 Perhaps the next simplest case is the system w= 2 Aw where r is a positive integer and A is a constant matrix. It is readily verified that a solution matrix of this equation is given by hb = elt rena which indicates that the lower-order terms in Q in (2.5), R, and the formal power series P in (2.4) are completely due to the effect of the terms z“'A, + 2A, + - + + in the A of (2.3). Proof of Theorem 2.1. First, it is clear that, if P, 2, Q are matrices as144 ORDINARY DIFFERENTIAL EQUATIONS (Crar. 5 described, the product Pze° is a formal log-exponential matrix, for each of the factors P, z*, e@ is one, Actually, there are no logarithmic terms in this case aince R is disgonal. Also, it may be assumed at the outset that Ao is a diagonal matrix with elements \;, . .. , As, for a simple substitution » = Tw in (2.2) would effect this, if T is that constant nonsingular matrix such that TAoT- has the diagonal form with 4, +.» Xs as Giagonal elements. Note that when A, is assumed to be diagonal, the assertion (2.6) aays, in particular, that Qo = Ao. Suppose $ in (2.4) is a formal-solution matrix of (2.2), where P, Q, R have the properties stated in the theorem, Then differentiation yields B = Peete + PR + Pa®( Qo t 1G + + OE and using the fact that the Q; and z* are diagonal one obtains B = [Po + PR + PQ + 2G + + + Qr)lete ‘But from (2.2) there results # = 2APo0 and hence P+ PR + PleQo + 101 + ++ + + OQ) = AP Using the power-series nature of P and A, this gives F erie — ey + (FA) Ot + +O) xo =o ee ® rA) Qn) Comparing coefficients of the various powers of z~* yields PrQo — AcoPo = 0 k PiQo— 4oPs = ) (APs -PuQ) sksn & 2.7) PrsraiQe — AoPaioat =), (AiPreestes — PaiesssQi) eee + J APuensia + Pi@E — BR) (BO) tert] ‘Thus a necessary condition that $ ih (2.4) be a formal solution matrix of (1.2) is that the matrices P,, Q,, R satisfy the relations (2.7). Con- vereely, if a set of matrices Py, Q:, R exist which satisfy (2.7), then } given by (2.4), (2.5), and (2.6) will be » formal-eolution matrix of (1.2).Sze. 2] SINGULARITIES OF THE SECOND KIND 145 This follows by a retracing of steps. Thus all that remains to be proved is to show that the relations (2.7) can be solved for matrices Pi, Q., FR. Since Ao is assumed to be diagonal, a solution of the first equation in (2.7) is given by Q=Ac Pox (2.8) where £ is the identity matrix. The second equation in (2.7) for k = 1 is PQ. — AP; = APo — Pods or, using (2.8), PyAy— APi = Ai- (2.9) Since A, is diagonal, the diagonal terms of the left side of (2.9) ara zero, and thus the diagonal elements of Q, must be identical with those of Aj. This determines the diagonal matrix Q, uniquely. The nondiagonal terms of P; are determined from (2.9) by Oy — dopiP = ay? DD (2.10) where p{?, ai? are the elements in the ‘th row and jth column of the matrices P,, A;, respectively, Since 4; * d; (i » 7), Eq. (2.10) deter- mines the nondiagonal elements of P; uniquely. Let P, denote the matrix with diagonal elementa zero and p{}’ in the ith row and jth column (§ +3). Then 8 solution of (2.9) is Pi =P, + Di =P, + Pod: where D, is any diagonal matrix. Here use is made of the fact that DyAq — AcD, = 0 since Ao is diagonal. Note that P, satisfies PAg — AoP, = AiPo — PQ = Ar - Q Let 1 < k Sr, and assume the existence of diagonal matrices Qs, Q:, +» +» Qe-1 and matrices Pi, .. . , Pes of the form Po= Pet PouDi + +++ + Pods Qt) where D;,... , Ds, are arbitrary diagonal matrices, the diagonal elements of the P; are zero, and the P, satisfy BAy- AP =S GH1,...,k-D (2.12) where a & =D (AP FQ) GH 1... 8-4 P= tm] Since Ay is diagonal, it follows from (2.12) that the diagonal elements of each 8; are zero, Placing (2.11) into the second relation of (2.7) for k,1468 ORDINARY DIFFERENTIAL EQUATIONS (Cuar, 5 one obtains, on collecting terms, bo Prdo— AaPs = J. (APra — PriQi) + (Aa ~ Qi) + Si-aDs Tel + Sadi t ++ + 8Dir (2.18) Since the diagonal terms of PyAy — AoP, aro zero, as is the case for Sy, . . . , Sit, one sees that (2.13) determines the diagonal elements of Q, uniquely, and thus the diagonal matrix Q, uniquely. As in the passage (2.9) and (2.10), a solution F, of kn1 BiAy — Adh, = b (Bos — PraQ) + (Ae — Qs) «(2.14 ms is determined uniquely as regards the elements off the main diagonal. The elements of P, on the main diagonal are taken to be zero. Then the matrix P= By + ByaDit+ <-> + PoDe (2.15) where D, is any diagonal matrix, will be a solution of (2.13), for PyAo — AvPs = ByAo — AoPy + (PriDiAo — AcPriDi) + ++ + (PoDiAo ~ AoPoD:) = PAs — AP, + (Pride APia)Di + + °° + (Poo ~ AaPo)Ds since Ag and the D, are diagonal. Using (2.12) and (2.14), one readily sees P, given by (2,15) satisfies (2.13). By induction, corresponding to the choice Qo = Ao, Po = E, this proves the existence of diagonal matrices Qi, .. . , Q, and matrices P,, . . . , , with diagonal clements all zero, satisfying (2,12) for? = 0,1, .. . ,7, and such that the matrices P; in (2.11) satisfy the second relation in (2.7) for k = i, For & = 0 in the third relation of (2.7) one obtains : Pride — AoPrar =), (ApPruict — PresiQ) + AnssPo— PR (2.16) Tet and here is where FR enters. If the P; as given by (2.11) are put in the right side of (2.16), one geta . Popo = AePous =) (APruact ~ Prosc®id + (Aver — R) dnd +5,.Di+SaDi++--+SD, (217Sxe. 2] SINGULARITIES OF THE SECOND KIND 147 The diagonal elements of the left side of (2.17) are sero, as well as those in 8Di+ +++ +S,D,. Therefore (2.17) determines uniquely the diagonal matrix 2, and as in (2.18) one obtains a solution of (2.17) of the form Pras = Pra + BD + PoDrar (2.18) where P,.,; satisfies Fido — Aba = > (AP — ParsQ) + Ana — RB (2.19) teh and has diagonal elements zero, and D,,, is an arbitrary diagonal matrix. The last relation in (2.7) for & = 1 brings a change in that no new terma involving Q, or R enter. This equation is A Pryde ~ AoPrys =) (ArPopact — PestiQ) + AnsiPa + ArgsPo tod + PE — R) (2.20) and, using the expressions (2.11), (2.18) for the P;, this yields Purse — AoPoia =D (AiPryea — Pra riQh) + ArPa + Anes tet. + PAE ~ B)+[Y (APs ~ Presi) + Ants — 2] Ds = +SDit ++ +8:Di+ Dy (2.21) But by (2.19) the expression in the brackets{ ] has a diagonal consiating of zeros, and since the diagonal terms of the left: side of (2.21) are all zero, it follows that the diagonal matrix D, is uniquely determined by (2.21). Just as before, a solution P.4. of (2.20) can be found of the form Pras = Pry + PrasDi t 00+ + PoDeas where P,,+ is a solution of (2.20) with the P, replaced by P, everywhere and D, added to the right side, and the diagonal elements of P,,; all zero, and D,42 is an arbitrary diagonal matrix. In the next step, Ds is determined uniquely and a new diagonal matrix D,4:i8 introduced. Thus r + 1 matrices D, are always being carried in the procedure, Using another induction, it follows that all B, and D, are determined uniquely from the equations (2.7), and hence all the Px are determined uniquely, once the initial choice Qo = Ao, Po = E is made. This completes the proof of the theorem.148 ORDINARY DIFFERENTIAL EQUATIONS {Caar. 6 8. Asymptotic Series Recall that in Sec. 1 the series eH ett tot? +++) (1.4) was a formal solutiont of the equation a+ (: -) z=0 (1.3) provided the c, were determined recursively by ane §(2EEHM, Gzoe=) OH If a is not of the form n(n + 1}, the series in (1.4) diverges for all ¢ 3 0. However, it was seen that corresponding to this formal solution there existed an actual solution y of (1.3) such that ot) =e Y att +00) W @) ged and in particular & 1 elt) — et ) et*|90 (E> &) 1) [o0-a Son] a The relation (3.1), which represents the usual situation as regards singular points of the second kind, expresses the fact that the formal series (1.4) is an asymptotic series for the solution ¢ of (1.3). To be more precise, let S denote a connected sot in the complex z plane containing ©, A formal power series in 27}, p= y pe 8.2) a0 with partial sums E a= Joe &=0,1,...) is said to be an asymptotic series (or expansion) in S for a function f (as \el| > ©) which is defined in S, if for every k = 0,1,2,-.., af —%)—70 — (|z| + ©) uniformly for ze S. { Strictly speaking, the notion of a formal solution of a second-order equation has not been defined. I¢ can be defined directly in an obvious manner, or it can be taken as the firat component of any vector formal solution of the frat-ordor eyatem asso- elated with (1.3).Sue. 3] SINGULARITIES OF THE SECOND KIND 149 If p is an asymptotic series for f in S, then this relation is written f~pins Often S is a part of a sector of the z plane 8S: gi SargzSe. jel Br For example, if the formal series p converges in this S, it represents in S an analytic function f, and it is clear that f ~ pin S. If f has an asymptotic expansion p in a set S, the expansion is unique, for the coefficients p, in (8.2) are uniquely determined by the conditions LS po, 2(f — Po) Pr 2°(f — Po — Piz) — ps, ete. However, different functions may have the same asymptotic series. For example, the function g = e-*, defined for the set S: |z| >0, ~e Ss argz & a, where a < 1/2, has the identically zero formal power series as an asymptotic series in S, thatis,e*~ OinS. Hence, if fis any func- tion with an asymptotic series in S, f + e~* has the same asymptotic expansion in § as f. If f, g, h are three functions defined for ze S, h * 0, and if (f ~ grt Y pert in E=0 then this is sometimes written as S~gth > pe * in § Feo For example, it was shown for real ! > 0 (that is, S is the region |z| > 0, arg z = 0), that there exists a solution ¢ of (1.3) such that ew el(L bot + at? + + in S where the c, are defined by (1.5). Theorem 3.1. Suppose f and g are functions defined in a connected set S including », and S~p= ) pet 9g~a= ) nein’ 2 , 2», Tf a, B are any two complex numbers, then in 8, @ of + Bg ~ ap + By =), (ape + Boyz &=0 2 & (b) fo~ a= J ct & = > ras a0 i-0150 ORDINARY DIFFERENTIAL EQUATIONS (Caar. 5 @ fink- Be) 4 (SPP) rg os (Ff po #0) Po cy Po Proof. The proof follows easily from the definition of an asymptotic series, and is left to the reader. Corollary. If fiG@ =1,... , m), are m functions, f;-~ p;, 2¢ 8, and g@1, . . +» 2m) 18 @ polynomial, then F(z) = g(filz), . . - » fn(2)), has an asymptotic expansion in S, and this is calculated as tf all the expansions were convergent series. Proof, The proof follows by repeated application of (a), (b) in Theorem 3.1. Application. If Ais a matrix of functions whose components have asymptotic expansions, A ~ > 2*A, in S, then det A has an asymp- Prt totic expansion there, and the first term in this expansion is det Ao Thus if det dy * 0, (det A)-' has an asymptotic expansion in S with the first term (det Aa). Since the elements of A~! are composed of (x ~ 1)-rowed minors of A (which have asymptotic expansions) divided by det A, it follows that, if det Ay # 0, A~! has an asymptotic expansion in 8. Theorem 3.2. (a) If f~ ) pat", and f ts continuous for t & ty kao ( real), then PO = |” Ue) - pe - pe ar~ Pete I , » ke (b) If, further, f’ exists and te continuous, and f’ has an asymptotic expansion, then fim — J b — Upratt km? Proof. (a) Hf — po — pit!) px, t+, and therefore F(t) exists for ¢ > to. Also, for fixed m 2 1, m41 f- (2, pitt) = efferent where ¢({)-» 0,t-> +o. Hence i Ger)|-[ [oon ‘kL lt de SadSze. 4] SINGULARITIES OF THE BECOND KIND 151 where eu(t) = BP, le(z)|]. But ex(t) > 0, 2» +, and since . 1 {OED dp = = [ amd dy om one has = (F ~ a ee +. This proves (a), kel @ Lets ~ } gt ‘Then k=O f= [r@d +s = [Geta a + [FG ~ 99 - 267) dr +54) f= att arloge— [" U'@) — a0 ~ a0) dr te where ¢ is a constant. Since f has a unique asymptotic expansion, it follows from (a) that go = q, = Oand @ = —(k — l)pu,k 22. This proves (6). If f bas an asymptotic expansion, f’ need not have one. For example, if f = e sin e¢, then f ~ 0, but f’ = —e-' sin ef + cos ¢ does not have an expansion, for lim cos ¢, 1» + ©, does not exist. 4. Existence of Solutions Which Have the Formal Solutions as Asymp- totic Expansions—-the Real Case It will now be shown that conresponding to every formal-solution vector of (2.2) there exists an actual solution with the formal solution as an asymptotic expansion which is valid in some sector in the complex z plane, for z sufficiently large. In order to do this, certain appraisals will have to be made. It is important to distinguish, in the following, between formal solutions and actual solutions. An actual-solution matrix (or vector)‘ of the system (2.2) will be denoted by # (or y), whereas formal solutions will always be denoted by ® (or ¢). In this section, unless otherwise stated, it will always be assumed that the system under consideration is the one considered in Theorem 2.1, namely wi =2hw (r= 0) (4.1) where A, has distinct characteristic roots,152 ORDINARY DIFFERENTIAL EQUATIONS {Cuar, 5 li P is a formal power series in z~', P=) oP denote by P,,) the polynomial in z~!, n Pim = DOP, (m= 0,12...) ko If & = Pz*e® ((2.4)] is a formal-solution matrix of (4.1), denote by 4, the “truncated formal solution,” $n = Prmthe? (4.2) Clearly 4,,) can be regarded as a function of z also. A sketch of the method to be used here will now be given. Since is a formal solution of (4,1), it is clear that formally PF = 2A For the truncated formal solutions it might be expected that, if Bray is defined by the equation Bay bih = 2 Bem then the early terms in B,,) are identical with those of A. This will be shown to be the case in Lemma 4.1 below. Bince ¥im:, Pm, and bz} all exist as well-defined functions of z for all z sufficiently large, $i) is an actual- (and not only a formal-) solution matrix of the system w! = 2*Bimyw (4.3) If (4.1) is written as wo! = 27 Bint + 2(A — Bom (4.4) then, since A — Biq) is small for large z, the equation (4.4) can be recast, by treating the last term as though it were a given function of z and using the variation-of-constants formula much as in (1.7), to get an integral- equation formulation. Since a solution Jim) of the homogeneous equation (4.3) corresponding to (4.4) is known, it will be shown that the integral equation can be dealt with by using the method of successive approxima- tions to obtain a solution of (4.4) [and hence of (4.1)] which behaves like Pim for large 2. Lemma 4.1, The matrices ¥(,, 7) exist for z sufficiently large, and if Bony 18 defined by ZB = Limybich then 2A = 2B + Ey (4.5)Sec. 4} SINGULARITIES OP THE SECOND KIND 153 where Bom, 2*Ewm are analytic for all z sufficiently large (including ~), andt Ew(2) = Mel") (e| > +) (4.6) Proof. It is obvious that &,, exists. Since Pia) is a polynomial in 2}, and det Py + 0, (det Pim))~! exists, for z sufficiently large, as a con- vergent power series inz~, Therefore P;;), exists and is analytic for all z sufficiently large. From (4.2), it follows that Ba = te P 4.7) exiats for z sufficiently large. Also By = Pin + Pik + Pron Q' ere? 48) since 2, Q are diagonal. From (4.7) and (4.8) one bas BBS = (Play + Pik + Pon @)Peh (49) and since Q’ is the polynomial matrix @ = Qt e+ +O it is clear that Bix) = 2~t&{,)$7;, is analytic for 2 sufficiently large. Since det Py * 0, the formal power series P has a formal reciprocal P-1, and hence & has a formal reciprocal which is given by Bt = tap Now @ = (P’ + z'PR + PQ’)z*e®, and therefore YE = (Pi + PR + PQ’)PO (4.10) But from (4.1), &’&-' = 2A, and since A is analytic for z sufficiently large, z~-*8’$~! must be a convergent power series in z~ for z sufficiently large, and hence analytic for large z. It remains to compare the expressions (4.9) and (4.10). The formal series P may be written as P= Pom + OR, (4.2) where R,, = Ro +27; + + + is another formal power series in s~!, Tt will be useful to let J, denote any formal-matrix power series in z—! having z~* 2a a factor, and thus J, is such that ao Iyer +e - for some constant matrices Jo, J,,.... Using this notation, (4.11) may be written as P= Peay + Smet (4.12) {By (4.8) is meant [Eeu(e)] - |e|"*!-7 = 0(1), |e] + +00, where the bound depends on m,154 ORDINARY DIFFERENTIAL EQUATIONS (CHap. 5 Now det P = det Pom + Inar and this implies that (det Py“! = det Pom)? + Ima If adj P is the forma! matrix such that P adj P = (det P)E then, since the matrix adj P has as its elements the cofactors of the ele- ments of P, it follows that adj P = adj Pony + Jngt Therefore PAs Poy + Ima (4.13) From (4.12) one obtains Pl = Pog + Suis (4.44) and combining (4.12), (4.18), and (4.14) there results from (4.9) and (4.10) , Sto BB) kage (4.15) The r in the last term is due to the term PQ’P~! in (4.10). But (4.15) implies that 2A = FBiny + Iniinr (4.16) and since A and Bm are both analytic for z sufficiently large, so is eI myi-r, Denoting Jingier in (4.16) by Bey, one sees that this Bum) satisfies the conditions of the lemma, The asymptotic nature of the formal solutions will first be deduced for the case when z = / is real, Theorem 2.1 and Lemma 4.1 apply to this particular case. In order to prove Lemma 4.2 below, some notation will be required. For fixed integral m 2 0, let the column vectors of ,,) be denoted by Gon (i= 1,...,n). Then emrs = Pemiltie® (4.17) where A) =r pa ge eee pare (4.18) re PET TM; ‘ . pray is the ith column of Py), and p; is the element in the ith row and eolumn of # (see Theorem 2.1). Consider, for the following, a fired i. Since tq is a polynomial, its behavior as t+ © is determined by the term of highest power in ¢. Divide the integers j = 1, . . . , n into two classes J), I, according toSec. 4] SINGULARITIES OF THE BECOND KIND 155 the following rule: gel, if RG-— gto {t-> +”) jela if R(gi — g,) is bounded above (tt +o) Of course J,, I; depend on the i chosen. Let, further, p = max {p;. (4.19) Lemma 4.2. Lei m be any positive integer such that m—r —~ p+ Ra; > 0, foralj =1,...,2. Corresponding to any column vector Gimi of &(my, there exists an actual solution vector pom of the system w araQw (4.20) such that lecens(e)] = Oe) (E> >) Proof. The solution im; will be constructed using the method of successive approximations, combined with a version of the variation-of- constants formula, Consider the two systems w = Aw = UB ew + Faw (4.21) w! = Bey w (4.22) From the definition of Bim, the matrix $,.)(t) is a fundamental matrix for (4.22), if ¢ is sufficiently large (det Py x 0). ‘Thus if (4.21) is regarded as a nonhomogeneous system with (4.22) as the corresponding homo- geneous system, the variation-of-constanta formula ean be applied to express solutions of (4.21) in terms of a quadrature of solutions of (4.22). Tn doing this, the limits of integration must be set correctly. Let fe be so large that ;,4(t) exists for ¢ 2 to, and split $73(£) into two parts Boy = etree = HE + OO, where the jth row of #12) is identical with the jth row of £;}, or identically zero, according as je 7, or je Z44 similarly for #2, Thus the nonzero rows of &*, consist of those rows of 47}, which have as factor e~* for jel ko 1,2 The integral equation to be considered is the following: 0) = Gnld) + [i Kieriule) de +f KaltsuG) dr (et < ©) (4.28) where Kilt) = Se(QVE Em), — Kalts7) = Seal") Ber) (4.24) By direct verification, it follows that if w = y(¢) satisfies (4.23), where the integral f° converges, then satisfies (4.21).1G ORDINARY DIFFERENTIAL EQUATIONS [Crar. 3 In order to solve (1.23), define the successive approximazions for {2 toby ot) = 0 t ‘ AH = Boal) + fi KieinderGe) de + fl Kalen) de, (4.25) (k=0,1,...) It is a matter of proof that cach of the approximations exists. This will be omitted since it is entirely similar to the proof which will be given below concerning the magnitudes of the successive differences. In order to appraise these, it is necessary to appraise the kernels K, and Ks. Now JKC S mh OL Eo) (4.26) and from Lemma 4.1, [Bem(r)| = OG") (F > +@) (4.27) The 7,jth element of the matrix beny(t)Pi2)(r) is given by nO HKODs = » Peat atParinry (4)"eu0-ee (4.28) a, Because of the convergence of P;}, for large enough ¢, [Pmt Pon] = 00) > +) 4,29) From (4.26) through (4.29), and a similar consideration for K,(¢,r), it then follows that there exists a constant ¢ = c(m) > 0 and a éy sufficiently large [which can be taken to be the ¢y in (4.25)] such that |Kiltz)| Sc d [ey qr t—Bta, BEG A(t F = 1,2) (4.80) dedy Further assume that éo is so large thatT yf item ioRy ot (4.31) to RG — qld) is increasing (Le Ly, 1B ) 432) Rg) — qu(t)) is nonincreasing (Le Jn, 6 & to) " From (4.25) and that P(E) = Semrill) = Pemi(feren® and therefore le) — gO] S etre ( B be) + All the integrals in (4.31) exist because of the assumption on m,Sec. 4] SINGULARITIES OF THE SECOND KIND 157 Assume AQ — g()| S c2-E-PHeRM (2 B bo) (4.33) Then from (4.25) leHO — POL sti +hs (4.34) where Pr = fi Ge) let) — AAG) dr Ta= f° UesG0l leG) — eH) de On account of (4.30) and (4.33) one has ns erarernpessin ff tee toangntein—ecarsean-aur dr and by virtue of (4.32), R(g(r) — ar) + gilt) — gi(é)) SO for to S rSh Therefore, Ms oro neeto Y f- qttemmt Re, dp ich and using (4.31) Ty S Q-@rnyegheo A similar argument shows that I’; satisfies the same inequality as Ty. ‘Thus (4.33) ia true with & replaced by k + 1, and since (4.33) is true for k = 1, (4.33) is established for all k by induction, Therefore the series oO + DO - HH) k=1 is abselutely uniformly convergent to a vector function ¢ = gomiff) on every finite interval 44 St< T <0. Also leems(e)| S oteeRaco > Qr GD = DotveMa — (¢ & ty) (4.35) ee Using the standard argument for successive approximations, it now follows that ya): satisfies (4.23), and hence the differential system (4.21), thus proving Lemma 4.2. Lemma 4.3. For any sufficiently large m, the solution yey; of Lemma 4.2 satisfies LetmGt) — Preas(E)] = O(EMectemetm) — (Ea oe) (4.36) where ts @ positive integer independent of m and of ¢. Proof. Just how large m must be will be seen in the following. From158 ORDINARY DIFFERENTIAL EQUATIONS (Cuap. 5 (4.28), if t > 2, and m>r+p— Mp, (j = 1,...., 2), there exists a solution gm; such that locust) = Pomil)] S Ar Aa + As (4.87) where bas fi? Kater) beret] de, As = fy IK il6s 1 Nocente)| dr Aa = [7 Us] boemite)| dr Consider As first. From (4.80), (4.82), and (4.35), (4.38) fn 4:8 aetvennwn f a qr tte Rey dy Now te [Baretta dem Oem Be) = O(ermtitn) (I+ Fm where y is any positive integer exceeding r + 2p — 2 min Ro. Choose m so large that » — m <0. Then Ag = Off +B.) (t+ +) (4.39) A similar argument shows that this also holds with A; replacing Ay. Turning now to A;, apply (4.30), (4.31), (4.82), (4.85) to A,. What cesults is As 5 rornn » eles e (+412) qt)? (4.40) tel Let ¢ denote the highest power of ¢ appearing in R(q.(t) — q(t). Since Ras — %) is increasing for ¢ Z to(le 7), the coefficient By of * in SR(gAt) — gi(t)) is such that Bg > 0. The coefficient of (in Rl a) — al) + a 8) a (5) is then given by £8 — bu = ba git) <0 for oz > 0. From (4.40), therefore, the term under the summation sign is 0(e~*') for some y > 0, and hence, in particular, Ay = O(Y— mre 0,0) (ft +0) (4.41)Sec. 4] SINGULARITIES OF THE BECOND KIND 159 Combining (4.37) through (4.41), one obtains (4.36), which proves the lemma. Lemma 4.4. If m ta sufficiently large, then for any fixed integer m' > m, eens) — eeons(t)| = O(e% eta") @-> +2) (4.42) where a is a posttive constant independent of m’ and m. Proof. Choose m so large that Lemma 4.3 holds, and let m’ = m +1, la positive integer. From Lemma 4.3, Pearl) = Peovslt) + O(Matemesa) (E> +o) (4.48) where m is such that u — m < 0, and hence, if ®,.) is the matrix with the column vectors giayi, . 6 +» Pian Him {Be WOU ® m Heo (De PO1L-* + O(H™) = Pent) +00) = Pet Or) (t @) Since by Theorem 2.1 det Po = 0, this proves that det Smt) * 0 for all t sufficiently large, and hence 4;.) is a fundamental matrix for w’ = (Aw for all sufficiently large 4. But (a4 is also a fundamental matrix, and hence 2 Pomill) = > eieemsns(t) (4.44) ima where the c, are constants. Recall that Pros) = Demilt eres (4.48) where pia(t) is the ith column of Pm(t) = > t-*P,, and hence is of the a0 form Pom) = Ow (4.48) and where the pix are constant vectors. Recall also that if Ao is assumed to be in diagonal form, then Po can be chosen to be the unit matrix Z. This is clearly no loss of generality, and so this will be assumed in the follow- ing. Then pj, is a vector with I in the ith row, and 0 elsewhere. By virtue of (4.43), (4.44) is equivalent to 5 8 Arad) + OCrA-meHCO) = Y caPrarndt) +0 ( J Jeale®artom-tete 0) = . irk160 ORDINARY DIFFERENTIAL EQUATIONS [Cuap. 5 and by (4.45), (4.48), this gives etc, — Lpemsll) + Y caktetiPemy(l) = O(ER4AmetE) ae +0 (3 Jeulterrmmer0) (4.47) Let, as before, J, denote the set of all integers k, (b= 1,..., 2), such that It(q: — g.) > +, (> +), and J; the complementary set in 1,2,..., 2. From the structure of the g, as polynomials with no constant term, it follows that ke 7, if and only if cither Sig; = Ng, or RG —-G— —-, (t+ +). It will now be shown that o; = 1 and Mik xi, ke In, then en = 0 in (4.44). Let Ia, be the set of all & such that 2g; — m)—4 —«, i> +e). Suppose J}, is the set of all ke Jy, such that Rg. — qx) is bounded above asi—> +, forall ke Jy. Let k’” be any integer in /%, such that Row = Row, for alk’ Tf. Divide (4.47) by t-e% and Iet i> +e. Tf attention is confined to the k”th row, what results is curv = 0. Con- tinuing in this fashion, one shows cx, = 0 for all ke In. Let Iz2 be the set of all k = i such that Ne: = Moz, and let k’ s Taz be such that Roy 2 Rox, for all ke Igo. Divide (4.47) by tet and let t— +, By observing the k’th row, in passing to the limit, it is found that cx = 0, if Rex > No. It is true that for Roy = Sta; the k’th row shows again that ey = 0. Then it follows easily that ¢; = 1, by dividing by ex and letting > +. After that, the argument goes nicely for Rav < Rp, k's Ze. Here m must bo assumed so large that 9p; — Row + uw — m <0 for all ke Ja2 such that Roy < Roy. Therefore, from (4.44) one obtains emit) = gemeni(l) + > caspomtnill) jen From this it follows, using (4.43) and (4.45), that. : eemill) = vensni(d) + O(E%MEW) where E(t) = O(e-“*) for some constant a > 0, which does not depend on m orm’. This proves the estimate (4.42). It is now possible to prove the asymptotic nature of the formal solutions in the real case. Theorem 4.1. Let ; = p,le% be any column vector of the formal solution matric & = Pt®e® of (4.20), where A satisfies the conditions of Theorem 2.1 forz = 4. Then there exists for all sufficiently large t an actual-solution vector of this system, gi, such thatSrc. 5] SINGULARITIES OF THE SECOND KIND 161 lei) — Ptmnlt)| = OGReremteRa) — E> oe) (4.48) holds for allm = 0,1,2,.... In particular, o ~ $. Proof. From Lemma 4.3, for every m’ sufficiently large, aay m’ 2 my, there exists a solution gio; such that emit) = Bemrc(l) + O(etH—meRe) — (L—> +o) (4,49) where x is a positive integer independent of m’. The integer m can be chosen so that m,> 4. By Lemma 4.4, for m sufficiently large, say m2 m: 2m, and m' > m, prmi(t) = pont) + OfeMaHmaty (P+ 4.) (4.50) where a is a positive constant. Combining (4.49) and (4.50), there regults for m = m2, and all m' > mi, tmpilt) = Bomeys(t) + O(ERA tem’ efte,O) (t4 +) (4.51) But, by the definition of @:(O, it follows that Pm) = Pemrmnens(t} + OCR’ ette (9) (to +) (4.52) Letting m = m’ — x» — 1, and combining (4.51) and (4.52), one gets for allm > m2:—-y-1 Pemrill) = Prone) + O(a) (Par fo} (4.53) It remains to prove (4.53) form = 0,1, ...,m:—2#—1. Since Memrdl) = Pem—ni(l) + OCMmmeME) (E> + 20) it follows that (4.53) holds for m = m,— » — 1. Using an induction, it ig easy to see that (4.53) must be validform = 0,1,2, . . . , thus proving the theorem if ¢; is chosen to be the solution gas. 6. The Asymptotic Nature of the Formal Solutions in the Complex Case The result given in Theorem 4,1 holds not only for real 2 = 4, but it is also obviously valid on each radial line z = ée**, for any fixed #6, How- ever, the theorem does not relate the solutions along one radial line with the solutions along another radial line. Here it will be shown that, with the aid of certain theorems in (complex) function theory, the result of Theorem 4.1 can be used to prove that a solution ¢; with the asymptotic expansion ¢, along an appropriute radial direction actually has @; as an asymptotic expansion in a sector of the z plane. The results required from function theory are due to Phragmen and Lindeléf, and are extensions of the maximum-modulus theorem.t A } For proofs of these theorems, see E. C. Titchmarsh, The theory of functions, Oxford, 1939, pp. 177-180.162 ORDINARY DIFFERENTIAL EQUATIONS [Crar, 5 statement of the results needed here follows. With z = | tion arg z = @ will be used. Theorem A. Lei f be an analytic function for ks |e] < ©, 0, S orgz 5 0 , the nota- where k, 01, 62 are reat constants. Let J@ = Oe") (z] 4 &) uniformly in 6, S$ arg z S 62, for some constants ¢ and m such that m0. — 0) <3 If f is bounded as |z| -» © on the lines arg z = 6, and arg z = 02, then f is bounded uniformly as \z| > © in 0, S argz S 62. Theorem B. Let f be analytic and uniformly bounded for the region in Theorem A. Moreover, suppose that there exist constants a and b such that fe) > a.as || > © on arg z = b, and f(z) +b as |e| > © on ang z = Os. Then a = b and f(z) ~» a uniformly in @, S arg z S 82 a8 |zj|-> &. Recall that the system under consideration is w = 2 A(z (r 2 0) (8.1) where A(z) = HA, 5.2) & de? * 62) and the latter series converges for |z| > d,forsomed > 9. Animmediate consequence of (5.1) is the existence of a constant ¢, > 0 such that any solution g of (5.1) satisfies |’| S eilz|*fe| for large jel. Ife = a/(r + 1), this implies ee) = Of") (laf @) (6.3) uniformly in any fixed sector of the z plane bounded by two radial lines, Let the integer 7 (1 S$ i S n) be chosen and kept fixed in the discussion that follows. Since the characteristic roots }; (j= 1, ... , 2} of the matrix Ao are assumed to be distinct, it follows that the equation RA -—AWe]=0 (jf ¥# 1) (6.4) determines a finite number of directions in the z plane. These are the directions arg z = @ (mod 2r) for which - cos [arg (Ai — 4s) + (r + DOT = 0 Let S; be a sector Si a Sarg: 58 such that all the directions determined by (5.4) are exterior to S,. The following result. will be proved.Sec. 5] SINGULARITIES OF THE BECOND KIND 163 Theorem 6.1. If @; = pa%e% te any formal-solution vector of (6.1), then there exiats in the sector S;, determined above, an actual solution ¢; of {5.1) for all sufficiently large |2|, such that HH uniformly in Si. An easy corollary to the above theorem is the following result. Theorem 5.2. If a sector S in the z plane contains no direction for which Rar-—AVer]=0 Giz... alg) then there ezists a fundamental set of solutions g (§ = 1, ... , ) of (5.1) in & for all z sufficiently large such that e~e G=1,...,0) uniformly in S. Proof of Theorem 5.1, For rane el... ,n, ale) — ai(2) = OW — = = 7 tor - ayye tr OM M2 and it is clear that the behavior of 9t(q; — 9;) in Si, for 7 7 #, a {zl &, depends on the firet term Qi — Mert! | Fi ] since |A; — A #0. It follows from the definition of S, that the integers g=1,2,...,¢—1,¢+1,..., m fall into two classes 7, and J:, where Ra-a—o-e Gel) (6.5) uniformly in S; as [z| > ©, and Ra~G—a—o Gels) (6.6) uniformly in S; as |z[—+ %. Theorem 4.1 gives the existence of a solution ¢; of (5.1) on the line arg z = a for all sufficiently large [z| having the property that we (arg z = a) 67 By uniqueness, it follows that g; can be continued off the line arg z = a, and hence it may be assumed that ¢; exists for all sufficiently large |z| and satisfies (5.7) on argz = a. Now let + be chosen so that a<7vse and yra< rEi (5.8) ‘It can also be assumed that 7 is chosen so that on arg 2 = -y one has, for164 ORDINARY DIFFERENTIAL EQUATIONS (Cuar. 5 any J and j, such that | # j, either RG -—g72 (argz = y) or Ra-g>o—2 (argz = 7) From Theorem 4.1 it follows that along the line arg z = ¥ there exists for sufficiently large |z| a fundamental set of solutions ¥:, . . . , Yo of (5.1) such that : wg — (arge = 7) (6.9) Thus, for some constants ¢;, . . . , Cr, 2 o= ) oh (6.10) it for all sufficiently large [2|. Assume that there exists k « 7, such that ¢ » 0 and such that for all Te Is, with | + k and ¢; # 0, it is the case that Rae — a> 0 (ange = 7) (5.11) as(z}—> ©. If m > 0 is any fixed integer, then by (5.9) and (5.10) one has in the notation of Theorem 4.1 P= CxPemra F O(lz[Ramm“ 1AM) (arg 2 = y) and in particular, if FR) = vilelensan-a then IQ) = a+) (le] + @, argz = y) where a is a constant vector, not identically zero. On arg z = « it fol- Jows, since ke J, and (5.7) holds, that . f@) = Ole") (el > @, arg z = @) Recalling (5.3) and using Theorem A and then Theorem B3 for each com- ponent of f, it follows that a = 0, which is impossible. Hence k does not exist and the nonvanishing terms on the right of (5.10) are from j 2 J, and jmi From (5,7) it follows that 92) = gilzpenatemr ia auch that az) ¢ — (e| > ©, argz = a) where ¢ is the constant vector given by € = lim Pmilzle"*z-%Src. 5] SINGULARITIES OF THE SECOND KIND 165 From (5.9) and (5.10) one has also that g@)— ac (le] > >, arg z = 7) and by Theorem B this implies that c; = 1. Thus (5.10) gives (es — Bemieruenramt! = O(1) (le > ©, arg z = ¥) This relation is obviously valid for arg z = a by (5.7). Thus, by apply- ing Theorem A, it is valid uniformly in the region w S argz $y. This is equivalent to e~d (a Sargz Sy) If y <8, then repeating the above argument a finite number of times in sectors with angle less than x/(r + 1) yields the theorem for the sector S;, The sector S, can usually be increased in size by observing the following facts.e A curve along which 9(q: — gj) = 0 for some j + 7 serves quite as well as @ radial line in the hypothesis of Theorems A and B. More- over, the results proved in Sec. 4 for z along a radial line will generally be velid along such a curve. By minor variations of the preceding argu- ment, asymptotic relations can be proved in a sector S; bounded by two such curves, but containing none such in its interior. Generally such relationships can be extended into an adjoining sector. This will be shown by an example. The method is quite general in scope, Consider the equation (1.3) with a = ~4 and the variables complex so that it becomes as a system ae _ ° 1 ' w dz - (: + id) o with roots A; = 7 and \; = —7. Taking the first component w, of the vector w and denoting it by u, He (r+h)u-o (6.12) and the two formal solutions of (5.12), as already seen in (1.5), are ace(ia te -) Biz aaee(i1- be . ‘)166 ORDINARY DIFFERENTIAL EQUATIONS [Caar. 5 Consider now the solution of (5,12), g;, which by Theorem 6.1 satisfies for § = arg z S + — 6, where 5 > 0, ana (5.18) Let y1 and 2 be the solutions of (5.12) which on arg z = O are asymptotic to 41 and @2, reapectively. Then for some constants ¢, and cs pi = cis +b cas Multiplying the above by e’, it follows that eps(e) = cre + es +0 (4) (5.14) on arg2 = Oasz— +, By (5.13) it follows that ettay(2) = o% ( +0 (4) (6.15) on arg z = 2/2 a8 |e] + 0. Let Fe = 4 feos ao where 2 is large and positive and the integral extends in the upper half plane on the are of the circle [s| = xo until arg s = arg z and then the integral is taken along the radius args = argz. Clearly (5.14) and (5.15) imply dim F@) =o (arg 2 = 0) lim F(z) = 0 (ore = ) me Thus by Theorem A followed by Theorem, B, cz: = 0. Considering next e“g,(z) on arg z = 0 and 1/2, it follows that ¢, = i. Thus forte) ~ Gan(e)ieten*t = O(1) on argz = 0 and «/2, from which follows y; ~ 4, uniformly on 0 S atgz 2/2. The argument can now be repeated for [7/2,r]. Moreover, it can be repeated for [r, 2x — 8] and [—7 +. 6, 0] for any > 0, so that finally ene (“7 +5 Sarge S 2 — 8) (5.18) (Indeed, z—y;(z) is, except for a constant factor, a Hankel function.] Similar results hold for the solution ga, vind: (—-2n-+8 Sarge Sa — 8) (6.17)Sxc. 6] SINGULARITIES OF THE SECOND KIND 167 A solution of (5.12) is zJo(z), where Jy is the Bessel function of zero order. Since yg; and g; are independent aT o(z) = exer(2) + creole) for some constants ¢, and c, [That c, = é& = (2/a)'e—*4 is readily verifiable.] From (5.16) and (5.17) follows ahta(z) ~~ erpi(z) + e2G2(z) (—7 + 6S arge Sr - 8) Since Jy is an entire function of 2, it is single-valued. The formal series on the right above is also single-valued. Because 2! occurs on the left, the left side is multiple-valued so that the above asymptotic formula cannot be valid on —x S$ argz Sr. Thus the result obtained is in a sense the best possible. By using obvious generalizations of the method of this example, asymptotic results can often be extended into sectors larger than S; of Theorem 5.1, 6. The Case Where A, Has Multiple Characteristic Roots This case is considerably more complicated than the simple situation treated in Sec. 2. The proof of the existence of formal solutions involves essential new difficulties, This can be illustrated in the real case z = £ by the example tw” +o +0 =0 (6.1) The system associated with (6.1) is Wy = We wy = ow — Ee If w is the vector with components w;, w2, then wi = (Ao +f 'A)w (6.2) Ao = ( 0) A= (2 -') (63) Therefore Ay has a double root \ = 0, with nonsimple elementary divisor. If ¢ = s*, the equation (6.1) is transformed into where wl" + sw! + 4w = 0 ( = #) (6.4) with associated system w! = (Bo + Bi) ( - 2) (65) where Be=(_° ) a= (3 -') (6.6)168 ORDINARY DIFFERENTIAL EQUATIONS [Crar. 5 Thus the characteristic roots of By are 4 = +21, and hence Theorem 2.1 ean be applied to (6.5). An implication of Theorem 2.1 is that (6.4) has a formal solution of the form ane (1 +34 ‘) (6.7) By substituting (6.7) into (6.4), r is found to be —3. Since (6.4) is real, the complex conjugate of (6.7) must also be a formal solution, Setting 3 = #, this shows that (6.1) has formal solutions of the form th = expi(tirtett + cops(ti)rte tt (6.8) where ¢;, cz are constants, and py, ps are formal power seriesin (1. Thus it is seen that fractional powers of ¢ can enter into the exponential term and the formal series. The following theorem ean be shown to apply in the genera] case when. Ao has multiple characteristic roots, but because of the complexity of the proof, the latter will be omitted, Theorem 6.1. Coenstder the system wl = A(z (6.9) where r 18 @ nonnegative integer, A(z) = eta, 3 and the latter series converges for |2| > a for somea > 0. Then there exisis a formal-solution matrix of (6.9) of the form ® = Sé& (6.10) where Q is a diagonal matriz with diagonal elements a which are poly- nomials of the type gle) = giole/*)" + ga leet + ss bg eel land h being integers, and S is a matriz whose elements 8,; are formal expres- siona of the type my My = 2h ¥ om log® z mad Here the ty are constants and the oijm are formal series Fun = » iiaSc. 7] SINGULARITIES OF THE SECOND KIND 169 where the ccm are constants. Moreover, the formal determinant of S does not vanish for large |z| < ©. It can further be shown that there exist solutions of (6.9) which have these formal solutions as asymptotic expansions valid in certain sectors of thez plane. The proofs of Seca, 4 and 5 can be readily adapted to this more general case. 7. Irregular Singular Points of an nth-order Equation Consider the nth-order equation of the form DY sren(w- = 0 (axle) = 1) (7.1) mao where r 2 0 is an integer, and the coefficients a, are analytic in a neigh- borhood of z = «, that is, Gn{2) = y Ompz-? peo and the series are convergent for |z| > aforsomea > 0. If vis any solu- tion of (7.1), let the components ¢ of a vector be given by Pe ZOU (e=1,...,n) (7.2) ‘Then it is easy to check that a= —k—- Delete (eel... 1) (73) Ge = (0 — re ten — aan: + days ++ + + aapa) Therefore, if y is a solution of (7.1), the vector with components y, given by (7.2) is a solution of the system w= 2 Alz)w (7.4) where A@) = o 1 0 of . 0 0 =r 1 ee . : : a . eee 0 0 : : ts (nro! 1 —an(2) —an1(2) ‘ ++ ~as(e) ~(n — Yee! — a(z) (7.5) Conversely, the first component of any solution vector of (7.4), (7.5) will be a solution of (7.1). Therefore Theorems 2.1, 4.1, and 6.1 can he170 ORDINARY DIFFERENTIAL EQUATIONS (CHap. 5 applied to (7.4), (7.5) to obtain the formal solutions and their asymptotic character, and, observing the first components, one gets the corresponding information for the equation (7.1). ¥f A is written as A4@= Dota =o where the A, are constant matrices, then o 10 0 0 o1 . Ao = + bee : . . ae 0 0 “00 1 I) The characteristic equation of Aa, a8 was shown in (6.19), Chap. 3, is Mf adh! $s + ay = 0 (7.6) and this can be immediately read off by replacing w® in (7.1) by \*, and z*ra,(z) by ayo, the constant term in the expansion of a;. 8. The Laplace Integral and Asymptotic Series The Laplace integral can be made the basis for proving the existence of actual solutions of (1.2) which are represented asymptotically by formal log-exponential series.t Here the special case (doz + bo)w™ + (az + Biwr? + + + + (az +b.) = 0 (81) will be treated. The a; and 6, are constants. Let P(e) = aoa® + ays! + ++ tae Q(s) = bos" + bya) + + + bn Let F be an analytic function and let ol) = f, Fleet de where C is a path to be determined in the complex s plane. Suppose ¢ is 8 solution of (8.1). Then, since formally ome) = f, F(eyeten de {ee for example E. L. Ince, Ordinary diferential equations, London, 1027, chap. 19.Sxc. 8] SINGULARITIES OF THE BECOND KIND 171 the equation (8.1) becomes f, FeateP(s) + Q(aer de = 0 Clearly, integration by parts yields f, FPze" da = roPie| - f é (FP )e* de e je de where FPe"| | Tepresenta the variation of the function on C. Thus [Fe — FP — F’P\e= de + FPe*], = 0 Choose F so that F'P + F(P’ ~ Q) =0 Thue adi * Qo) r= peo[ [33%] e) The condition that y satisfy (8.1) now becomes simply vew[[Belej-2 es If the roots of P(s) = Oare a, ... , &m and these are simple, then [3B a= x0 + Yano £ where F(s) is a polynomial and the a; are constants. Thus V = enttaw — 3)" He-» Je and C ia chosen so that V = 0, This may require that z be restricted. In case the degree of P(s) isn, then R(s) = aa, where a is a constant and V = eT e— ar", mm If Ray > 0, let gale) = f° ever T] (@ — aor da (8.4) ga- 172 ORDINARY DIFFERENTIAL EQUATIONS {Cuap. 5 where the integral is along a line from s = 4% tos = ©, If the line of integration makes an angle y with the positive real direction in the ¢ plane, then the integral for ¢, converges fora/2 < arg z + y < 30/2, and gx is 8 solution of (8.1). The range of validity of the representation may be changed by varying +. It is rather simple to show that such integrals (which are, in fact, Laplace transforms) are represented by asymptotic series. Asan example, take the equation aw" + (y ~ 2)w! ~ aw = 0 (8.5) (considered in Chap. 4). Here P(s) = at —8 Qe) = ys -—a Thus P(g) = a-(a — Mt and ¥ sare rer], If Ra > O and R(y — @) > 0, then a solution is given by fete rete ae Another solution valid for Rz < 0, R(y — «) > Oia [0 re = rete de (86) Solutions may also be represented by closed-loop integrals that make & positive turn about ¢ = 0, a negative one about s = I, a negative one about 2 = 0, and a positive one about e = 1. Ifs =o +1, then the solution (8.6) takes the form ef Ot jmtereter da (7) For all ¢ = 0 (indeed, for all o such that jarg o| <7 — 8), (tert ata = De + SEVER oy nae +o i a a Rye where F,(c) is uniformly bounded. Thus the solution (8.7) is given asymptotically by Py —«) , @ — UMy ~« +1) [aes (a — Wy(a — 2)T(y — a + 2)SE. 8] BINGULARITIES OF THE BECOND KIND 173 By changing the direction of the path of integration in the o plane, the range of validity of the asymptotic expansion may be increased to the range —x/2 < arg 2 < 52/2. PROBLEMS 1, Suppose that A in (4.1) is analytic in some domain D of the complex z plane and that AwAgtetAy bees in D. Prove that Theorems 4.1, 5.1, and 5.2 are valid, with the added restriction that ze D. @. Carry out the analogue of the treatment (5.12), at the end of See. 5, for the cauation dy b a ga t[1+$+a]uno where a and 5 are real constants, 8. Show the relationship between the problem above and (8.5) treated at the end of See, 8. (See Prob. 7, Chap, 4.) & For largo [e|, let S(2) = ete (% + bs hee ) where g(z) = got®t"/(@ +1) + «++ + gaz is 6 polynomial of degree = +1, na constant, and the b; are constant vectors, Let A(z) be asin (2.3). Show that wi = er Ale)w + Siz) has a formal solution a veet[at+---] where k = ~ rif Szandk = —rifr > z, provided (1) no characteristic root of Ag is equal to gq if + = 2, or (2) no characteristic root of Ag is zero if r > ». 6. Let g: be defined ag in (4.18). Let S be a sector of the z plane where for each j@1,2,,..,n0ither (a) Ray - go or else (b) RG -— ge aa|:|-+ = in S. Show that the differential equation of Prob. 4 haa a solution g auch that . e~y (eS) Hint: Let ¢. bo the truncated sum consisting of the first m + 1 terms of ¥. Let Saw. Then W = 7A(2)6 + Ja(z) where ef = Offefe—t) (= max (xr) Show that thero is a solution & = x(zm), x(zm)en* = O(|e|-=-1) along any fixed radius in S by use of integral equations similar to (4.23) but with #{2) and £{%, deter. mined by (a) and (b) above rather than by (4.19). ‘Then use the dovices of Seo. 5.CHAPTER 6 ASYMPTOTIC BEHAVIOR OF LINEAR SYSTEMS CONTAINING A LARGE PARAMETER 1L. Introduction Here will be considered the system of linear differential equations aim pAte)e (a Sts d) quay where r 2 1 is an integer, A is a matrix continuous in (t,o) fora S¢ Sb and |p| large, and analytic in p for large |p| so that Alte) = Dota (12) knO for large |p| with A, continuous. Such systems arise in eigenvalue problems, as will be seen in Chap. 7. (The results of this chapter will not be required for Chap. 7.) These systems also arise in cases where the highest derivative in an nth-order linear differential equation is multiplied by a small parameter—boundary-layer theory, for example, The relationship (1.2) can be asymptotic with the series divergent without, changing the results and methods of this chapter. In some cases, the solutions of (1.1) are studied with ¢ real and » com- plex and large. In other cases, p may be real and large while / is complex, or both may be complex. The method of this chapter has much in com- mon with that of Chap. 5. The case of real ¢ and complex p will be con- sidered here. The modifications required to handle the other cases are sufficiently close to the procedures of Chap. 5 and the present chapter that they will not be dealt with, A requirement here will be that the matrix Ao(?) have distinct char- acteristic roots for ¢¢ fa,b], or at least that the number of distinct char- acteristic roots of Ao(¢) does not change as ¢ goes from a to b. This eliminates from consideration here certain problems of great intereat. Such a problem arises in the case of the second-order equation SF + lt + af tale + + Ww =o as) m4Sec. 2] MINEAR SYSTEMS WITH A LARGH PARAMETER 475 in the neighborhood of f= 0, If a system formulation were used by setting w = w, and w’ = pws, then (1.3) would have r = 1 and aon (2 Clearly Ao(t) has distinct characteristic roots except at? = 0. To treat (1.3), it is replaced by an integral equation based on the use of the variation-of-constants formula and the solutions of w” + pw = 0 which are given explicitly by certain Bessel functions. In method this is very similar to what has already been done in Chap. 5 and to what will bedone here. However, there is a more complicated asymptotic formula here because of the appearance of the Bessel functions. The equation (1.3) is said to have a fransition paint att = 0. The treatment of transi- tion points will not be given here. 2. Formal Solutions For the case that will first be treated it suffices to consider format (Laurent) series in p~! with continuous functions as coefficients, that is, a series of the type Ms pe pio* kaso where the p, are continuous functions of ¢ on a S$ ¢ S b and all but a finite number of the p: with negative indices are zeroona Sib. The series need not be convergent. If each of the p, is differentiable, then the derivative p’ of p is defined as the formal series v= Drier tate Two forma! series are said to be equal if the coefficients of like powers of p-? are equal. Sums, products, etc., of forma) series are defined in the expected manner. Let g denote the polynomial in p a= goto + art + + + ald where the qe Cona $1, Formal expressions pet will be considered. ‘Two such expressions are equal if and only if the polynomials g are equal for a $ ¢ S$ } and the formal series p are equal, If p and q are differen-176 ORDINARY DIFFERENTIAL EQUATIONS (Caar. 6 tiable in i, then by definition (pety’ = (p! + pg')er Clearly (pe*)’ is an expression of the same type as pet, Formal matrices Pe? will be considered where the elements of P are formal series and Q is a diagonal matrix whose elements are polynomials in p of the type g above. Two such matrices are said to be equal if and only if the series matrices P are equal and the diagonal matrices Q are equal, Since @ is diagonal, the derivative of & is Q’e. From (1.1) and (1.2) it is clear that p"A(t») can be considered as a formal series. The formal matrix Pe is said to be a formal solution of (1.1) if formally (Pe)! = (P! + PQ’)e? = prAltp)Pe thas is, if P+ PQ! = prA(tp)P Theorem 2.1, Let the Ay in (1.2) be infinitely differentiableona & 2 Sb, and assume the characteristic roots (1) (i = 1,..., m) of Adlt) are distinct on a St Sb, so that MO - NO #0 GH#j,a8t52) (2.1) Then (1.1) has a formal matriz solution Pe®, where P= S PW O= FO +++ +0 en0 Moreover, Po(t) is nonsingular on a St Sb, Qy(t) = Ad), where AC) és the diagonal matrix with diagonal elements d,(),i = 1, ... , 2. Proof. The proof is similar to that of Theorem 2.1 of Chap. 8. From (2.1) it follows easily that the existence of all derivatives of Ao implies the same for 4, and therefore for A, For each ¢ in g $2 b a non- singular matrix Bo(!) exists such that By*(0)Ao(t)Bo(t) = A. It ia important to observe that Ap#> = BoA implies, with the use of (2.1), that Bo can be chosen so that it has all derivatives on [a,b]. Indeed, each column of Bo is unique except for a scalar factor. On the other hand, the kth column of By can be taken as a multiple of the cofactors of a row of the matrix Ap — %E, Since the roots dy are distinct by (2.1), it follows that for any fixed ¢ the cofactors of every row cannot all vanish. If the cofactors of a particular row of Ao — AZ are not all zero at t, by con- tinuity this is true for an interval containing ¢ Using the Heine-Borel theorem, there are a finite number of intervals whose union is [a,b) such that on each of the intervals the cofactors of some row of Ao — As are not all zero. Choosing one such row for each of the finite number of intervals, it is clear that the cofactors are infinitely differentiable. BySec. 2] LINEAR BY8TEMS WITH A LARGE PARAMETER 177 patching together the cofactors of these rows by use of infinitely differen- tiable scalar factors, it ie thus possible to find a kth column of By which does not vanish over [a,b] and has all derivatives. The transformation z = Boy yields a system for y of the same type as (1.1) with Ao replaced by A. Thus, with no restriction, it can be assumed Agis diagonal. The requirement that Pe® be a formal solution of (1.1) is den + ( oP.) (FQ + + +. we (Ya) (Ze) @2) =o ‘Fo Thus, setting the coefficients of p* equal, there follows PQs = AoPo Assuming that the equation (1.1) has been transformed as indicated so that Ao is in diagonal form, it is clear that Qj = A = Ayand Po = Eisa solution. The coefficients of p™' in (2.2) yield PQ) + PoQ) = AoPi + AiPo (2.3) Since Pp = E and Ay = Q; and Qj are diagonal, the elements p{? of Pi with ¢ yj satisfy Py — dA) = af? where af? are the elements of A. By (2.1) this determines p{), ¢ = 7, uniguely. Let P, be the matrix with elements pi’, 7 + j, determined as above and with diagonal elements zero. For i = j, (2.3) yields © ay = oP Thus Qj is uniquely determined. The determination of Q;(!) from Qj{¢) may be made unique by requiring Q,(a2) = 0. Let P; = PB, + PoDo, where D,(t) is an undetermined diagonal matrix. Then clearly (2.3) is satisfied. This procedure continues unti) the coefficients of p-! are equated and the term P{ enters. That P exists is clearly a consequence of the differ- entiability of Ao and A. As the coefficients of later powers of p are con- sidered, the existence of higher derivatives of the A; will enter, From the equation resulting from the p-! term, it also follows that Dj is determined. The strong similarity with the proof of Theorem 2,1, Chap. 5, is evident and further details are omitted. Remank: In case the A, are of class C,, but not of class C,.41, then the above process is valid only up to the point in the argument where the178 ORDINARY DIFFERENTIAL EQUATIONS [Caar. 6 mth derivatives enter. Thus in this case the existence of the early terms only of the formal series can be established. 8, Asymptotic Behavior of Solutions The concept of asymptotic series has already becn introduced in Chap. 5, There asymptotic expansions in the variable z were obtained. Here the asymptotic expansions will be with respect to the parameter p. A region in the p plane bounded by two arcs each of which tends toward » and which do not intersect except at their common initial point will be denoted by S. A function f = f(t,») is said to be represented asymp- totically in S by the formal series below, $lte) ~ J. esther? sro for a S$ ¢ Sb if for every nonnegative integer m there exists a constant K,, such that [se Y, 009%| om f for all sufficiently large |p|, p being in S, andfora St $b. Letqbe con- tinuous in (¢,) for ¢ in [a,b] and p in S. A function f will be said to be represented asymptotically in S by the series below, Hp) ~ es Y eet i=0 if for each m there exists a constant K,, such that |e) — ex > otde| = eee @.) 7) for pin S, |p| sufficiently large, and ¢ in [a,b]. Similarly, f is represented asymptotically by x “ ett) > en (p77 bert » if for each m & N |s40 ~ > entin >, eutbe-| Ke » lento bed i= aSee. 3] LINEAR BYSTEMS WITH A LARGE PARAMETER 179 If the relationship (3.1) is valid for m S Af, where M is an integer, then thia is denoted by Slise) eto YF ofthort #50 Tt will be seen later that the two boundary arcs of S have the property that, as |p| > © on each of them, arg p tends to a definite limiting value. Let the element in the kth row and column of the diagonal matrix Q of Theorem 2.1 be denoted by q. Then gi(t,o) = pPalt) fo -. Hypothesis H. Let there exist a region S of the complex p plane bounded by two arcs tending to © such thal for each i and j,1 S 1,7 © n, one of the inequalities Hlai(,n) — ai(e)] = 0 (3.2) Rlgilt.e) — gto) S 0 (3.3) is satisfied for all sufficiently large |p|, p in S, anda St Sb. It is not necessarily the case that S must exist. However, if the interval [a,b] is replaced by [a,c], where ¢ is close enough to a, then S does exist. Indeed, since A,(¢) — Aj(¢) s¢ 0, t J, it follows that arg pad) — A] = 78 + gull) where arg p = @ and g(t) = arg [A(t) — A,{2)]. Clearly Reds) ~ 4(O]} 4 0 (3.4) if cos [r8 + g{)] #0. Thus if c is near enough to a, y(t) is near enough to the constant (a) so that a range of @ can be found in which cos [ré + ¢i(t)] ~ 0 for all i 7. From this follows thp existence of a sector in the p plane in which (3.4) holds for all i # 7, Since p'(A, — Aj) is the dominant term of gj — 9; for large |p|, it follows readily now that S exists, Thus, if the interval in ¢ is short enough, S always exists. The fact that the boundary arcs of S tend to definite limits in arg p = ¢ as |p| > © follows readily from the fact that 99} — 9j] is a polynomial in p and thus for larga {p| has its behavior determined mainly by the term of highest power in p that appearsinit. Let the columns of P of Theorem 2.1 be denoted by p,j = 1,...,n. Then p%e% are each formal solu- tions of (1.1) forj = 1,2,...,7. Theorem 3.1. Jf Hypothesis H holds, then for each fixed integer m > 0 and for each integer k, 1S k S n, there exists a solution py’ of (1.1) euch that of? pMen for pin Sanda St sb. Reman: In case the A; are of class C¥ for some N > 0, then the above theorem is valid for m S N.180 ORDINARY DIFFERENTIAL EQUATIONS (Crap. 6 Proof of Theorem 3.1. The proof bas very much in common with that of Theorem 4.1, Chap. 5. The truncated series P,,, as well as 2%, are defined as being identical in the terms up to p-” with those of the matrix P and the vector p™, respectively, and terminating with the terms of power p~™. Thus Paltyp) = 2 etP(t) =O Since Pe® is a formal solution of (1.1), it follows that (P! + PQ'))P> = pA The existence of the formal series P—' follows easily with the same argu- ment as in Lemma 4.1, Chap. 5. Consider now B,,(t,) defined by Pi + Pn Q’)(Pn)-! = o' Bn The identity of B,, and A for terms up to and including those of order p7 (rt) follows easily. That is, there exists a constant C; depending on m such that [Atte) ~ Battal S pein a6) for large |p| and a $¢ $b. Moreover, Pc? is actually a fundamental solution of af = ptBaltyp)e (3.6) Let k, 1S & Sn, be fixed in this proof. Let p be in S and let Pn = VO 4+ V2 where V‘ haa its column of index j equal to the same column of P,, if for pinSandasisgh Rite) — alte) ZO (3.7) The columns of index j in Vi? for which (3.7) does not hold are taken as sero, This determines V{? also. Let (1.1) be written ag 2! = P Bat + p'(A — Balt Clearly p%e% is a solution of (3.6). Hf the integral equation (tp) = Bar(esp)erir) = VRE pee fe Ae-e(Pa(ryp))(A (r9) — Balsse))o(e,0) dr + VP pet fc -W9(Pa(r,p))""(Alrp) — Balte))e(r,0) dr (8.8) has a continuous solution ¢, it follows easily that ¢ is a solution of (1.1).8xc. 3] LINEAR SYSTEMS WITH A LARGE PARAMETER 181 Because of the vanishing columns of V‘, it is the case that the only exponential terms appearing in the elements of ViPee--a.») are e%t0l-a,(.0), where 7 satisfies (3.7). A similar result holds for Viehlt.o)~Rtevod Let KP ir.) = VEN Gp)etoI-9(Pa(r,0))"(A (7,9) — Bu(r,p)) and let K® be defined similarly with V2 replaced by V?, Fort $1 3b and p in S it follows for any j satisfying (8.7) that ferttorratroraenrac | = exp {— f Salle) — evel] de} S 1 Thus for S$ 7 S$ }, pin S and |p| large there exists a constant C2, depend- ing on m, such that belrrer KO Gr plemrr-ato| Ss Cy @.9) Similarly fora S37 S$ t, alto KOE plenir-aten| ¢ Cy (8.10) for some constant Cy. The integral equation (3.8) will be shown to have a solution by use of @ successive-approximations procedure. Let yo)(é,0) = 0 and let for t20 6 purn(te) = fa (ender — pr f KD (briedew re) dr +o ft KD (re) eu (rp) dr Clearly for large |p], p ¢ S, it follows from (3.9) and (3.10) that Kearny — em)em%| S Cnet max | (gy — eu-n)e7%] where the max is taken overa St Sb. If |f(4,0)| S Co and if |p| ie so large that (6 — a)(C: + Cs} & 4|ol"*!, then it follows easily that —- C leaen (40) — ver(be)| [emo Sor From this follows the uniform convergence of {yq)| to a limit which may be denoted by ¢ and which is a solution of the integral equation. More- over, it is also clear that len,ple)| S 2Co182 ORDINARY DIFFERENTIAL EQUATIONS (Crap, 6 From the integra! equation follows readily that 2Co(b — a)(Cy + Cr) le(pensir —- BiP(Lp)| S Ter which proves the theorem. 4. The Case of Equal Characteristic Roots ‘The case where two or more characteristic roots become equal at an isolated point of [a,b] will not be considered here. Iowever, » case of great interest arises where several of the roots 4,(t) are identical over {a,b]. That is, for any given 7 or j, either \,(t) = 4,(4) over [a,b] or else A) = AG(t) for every Cin [a,b]. In this ease it can be shown that, on certain subintervals of [a,b], formal solutions exist, but now, instead of involving polynomials in p and series in I/p, the solutions involve p/, where k is some positive integer. Thus the q.(4,) are polynomials in a” and the series P(t,p) are in powers of p~ + The prooff that the formal solutions exist is much more complicated than in the case considered in Sec. 2. However, the proof that there exist netual solutions asymptotic to the formal solutions in appropriate sectors is very similar to that given in Sec. 3. A trivial example is given by B= 2, ey = pe Here r = 1 and \y = A; = 0. On the other hand, it is easily seen that actual solutions y = (1,92) are given by exlt,p) = ce + egeoh gall,p) = expe! — eapte-ott where c, and cz are constants. 5. The nth-order Equation Consider the nth-order equation wh ptar(tpuer-Y $+ + + pant p)u = 0 GD over a St S b, where a(t) = J. aldpr* goo Tf u = zx, and if xy = prds, Ty pla...) Te = pte (5.2) {HL L. Turritin, Asymptotic expansions of solutions of systema of ordinary ‘incar differential equations containing a parameter, Contributions to the theory of noniinear oscillations, vol. 2, Princeton, 1952.Src. 5] LINEAR SYSTEMS WITH A LARGE PARAMETER 183 then ul = pas, ull = pry ey UO—D = pe-iry, and (5.1) becomes a = ~ptlan(ta)zi + ++ + + arlt,p)za] (5.3) Thus (5.2) and (5.3) form a system of n first-order equations and the theory of the earlier sections can be used. ‘The A,{t) are the characteristic roots of the matrix 0 1 see 0 0 0 ne 0 . . wee . (5.4) Q a ne 1 —anolt) —Go-10(t) ~~ * —aro(t) If the A,(t) are all distinct, then Theorem 3.1 may be used. If the hypothesis of Theorem 3.1 is fulfilled, then the system, (5.2) and (5.8), has n formal solutions pe%, and for any integer m > 0 there exist n actual independent solutions of (5.1) ¥; = ¥(t,p,m),i = 1, .. . ,n, such that tem) F pP beer Gal... ,n) where the 1 denotes the first component of p. The derivatives of y; clearly satisfy Vilhosti) B erpP (eer? =I, . ym) Hi Cem) B o* Pt pen = (G= 1...) ete. As an example, the equation uw + + Oly = 0 will be considered over the interval 0 = ¢ S 1 on which g is aasumed to be C*. Setting y = and y’ = pz; and using (5.4), there follows M(t) = 4, As() = —é. Thus a formal solution etot(co(t) + ex(te™! + + + +) is considered. This leads to (co + ete b+ +) + ip(oh + cont +) + alco tb ewt + ++) =0 Taking co = 1, there follows, equating successive powers of p~' to zero, ies +g =O oy + ies + ger = 0 eto.184 ORDINARY DIFFERENTIAL EQUATIONS (Cuar. 6 The determination of ¢,(d), ex(¢), ete., may be made unique by requiring that they vanish at ¢ = 0. Remark: The formal series may also be obtained from the relationship otto) = ot + pt f sin p(t ~ r)atedolee) dr (6.5) which allows at the same time a rather easy direet proof that there is a solution which it represents asymptotically. Here S may be taken aa the upper half plane Sp 2 0. Indeed, using successive approximations with go = 0 and ety = ef pu! f sin p(t ~ r)g(r}er(r,9) dr there follows easily, if |g(!)| SM, that Meter leur — vil S a Thus ¢, converges uniformly for urge |p| in Jp 2 0,0 Sf S 1, toasolu- tion gv. Also if |p} 2 2 le(4a)| S 2le| inS,0 StS 1. Using (5.5), the above gives 2M|ei| — giet] ere Ss Te Vi this is used in (5.5), there follows wet ft Lf, (<2) = gist - OO -— jio(tr—e), on pltp) = el#t ef ar) dr a f efel-O Q(z) dr + 0 ng Integration by parts used on the second integral yields = * az) 1 ete) = eft +f ‘Vip OT +0 is) This process may be continued indefinitely, thereby proving independ- ently the asymptotic formula for y in S. PROBLEMS 1. Tet Syp) = eho) $F Aa(Yort ++ for large [pl and a $¢ Sb, where aot athe) = 0% D oitortPnons,] LINEAR SYSTEMS WITH A LARGE PARAMETER 185 The hy and g; are of class C~ and is » nonnegative integer. Let Alte) = Aalt) + Ar(t}o™! + + - where the A, are C* on [a,b], Show the differential equation 2! = prA(4e)z + S(he) has a formal solution v(t) = p-Aer[colt) + ale! + ++ + wherek = rily $ rand k = wif w & +, provided (1) no characteristic root of Ao(t) vanishes on {a,b} if» re 2. Let the hypothesis of Theorem 2.1 be satisfied and let qy(t,e) be defined as in ee. 3 above (3.2), Let S be a region in the p plane, where for cachj = 1,...,7 either Rlajttn) — o'(to)] BO or else Rigi(he) — god] SO Show that the differential equation of Prob. 1 has for any fixed m > 0 a solation @ = o(te,m), where ¢(,e,m) Gr ¥(ée) for £¢ [a,b} and p eS. Hawt: Let P(t.) be the truneated sum of the first m + 1 terms of the sum in ». Lety =2— mw, Then y' = ptAy + Frft,p), where eR aha) ™ Olam) b= max (ry) Yat. B, (1,2) be ag in the proof of Theorem 3.1 and tet Pa = UL? + UD, where a column occurs in U® or U®, according to which inequality in Prob. 2 q, antisfies, Show that there is a unique solution x = x(l,p,m) of xdhosm) =f Galbsis)Pateo) de + 0 [? Gali IBaln) — AloYIxteim) de where Galtsrye) = UL U,p)e-e-CrMP (7,9) for a Sr St and Gm satisfies a similar equation for ¢ Walt a7? kel with 1 replaced by — © for the case associated with (1.4). The equality (1.7) is known as the Parseval equality. It is (1.6) and (1.7) which will be proved under quite general conditions in this chapter. Consider a slightly more general problem than (1.4), namely, be = ie =lz 2(1) = ax(0) (1.8) where @ is a constant. Let {d,} be the eigenvalues and {x3} be the cor- teaponding eigenfunctions, which exist but which need not be computed188 ORDINARY DIFFERENTIAL EQUATIONS {Cuar. 7 explicitly for present purposes. Clearly (Lxnxe) = MOcixa) Cepia) = Ralxerce) and Also by (1.8) (xara) ~ Grains) = 4 ff die + id dt = tee] = i(a — 1)x;(0)%(0) (1.9) Qy — XedGaixe) = t(ad — 1)x;(0)X4(0) and if ad = 1 it follows, if 7 = &, that A. is real and if 7 + & that Thus (xuxa) = 0 Thus, if ad = 1, the eigenvalues are real and the eigenfunctions may be taken as an orthonormal set. However, if ad = 1, it is readily seen that the eigenvalues necd not be real and that (x;,x2) * 0. If the right side of (1.9) vanishes, then the eigenfunctions may be taken as orthonormal. Thus, if the problem (1.8) is such that any two fune- -tions u and v of class C! on [0,1} which satisfy the boundary conditions also satisfy (u,v) — (u,Lr) = 0 (1.10) then the eigenfunctions of (1.8) form an orthonormal set and the cigen- values are real. The condition (1.10) is of central importance and is known as the self-adjoininess condition. It is readily verified that if u and v are of class C? on {0,1] and satisfy the boundary conditions of (1.1), then (u,v) = (u,Lv) so that (1.1) is @ self-adjoint problem, On the other hand, it is readily verified that with the boundary conditions z(0) = x(1), 2’(0) = 2z'(1) the problem —x" = Ir is not self-adjoint. . 2. Self-adjoint Eigenvalue Problems Let L be the nth-order operator given by Le = poe) + pyr) ps tpg where the p; are complex-valued functions of class C=-/ on the closed interval ¢ S$ ¢ S b and po(t) = 0 on [a,b]. Let Ug = ¥ (Mx?) + Naz*-()) Gad... yn)Sze. 2] BELF-ADJOINT PROBLEMS ON FINITE INTERVALS 189 where the My and Ng are constants. Denote the relationships Uj = 0, j=l,...,2, by Uz =0. The problem mt ig = lt Uz =0 is called an eigenvalue problem. It is anid to be self-adjoint if (Lu,v) = (u,Le) (2.1) for all u,v eC" on [a,b] which satisfy the boundary conditiona Uu= Uv=0 Here, if f,g ¢ (a,b), Ga) = fe isl= an The number (f,g) is called the inner product of f with g, and |[fi| is the norm of f in 2°(a,b). If (f,g) = 0, then f and g are said to be orthogonal. It has already been seen from the examples of Sec. 1 that the class of self-adjoint problems is not vacuous. Further examples of such problems are given in Probs. 1, 2, and 3 at the end of the chapter. Tt was shown in (6.13), Chap. 3, that to Z there corresponds an adjoint L*, where Tete = (1) (Bot) bh (— Vy 'Brz)O-P peo Bet such that for any u,v e C* on [a,b] (Catv) — (u,Ete) = [wv]{b) — [url (a) Clearly if L+ = L and U is such that Uu = Uv = 0 implies [uv)}(b) — [ue](a) = 0 then will be self-adjoint. The condition that U have this property is given in Theorem 3.2, Chap. 11. The problem x always has the trivial solution of the identically zero function. If 2 is such that + has a nontrivial solution, then Z is called an eigenvalue of x and the nontrivial solutions of x for that are called eigen- functions. It will be shown in the next section that eigenvalues always exist for a self-adjoint problem. Theorem 2.1. Let the problem x be self-adjoint. Then the eigenvalues are real and constitute an al most enumerable set with no finite cluster point. Eiigenfunctiona corresponding to distinct eigenvalues are orthogonal. Proof. Let 1 = ) be an eigenvalue with x an eigenfunction of 7. Then because Lx = dx, the relation (2.1) gives (A — X)(x,x) = 0. Because (xx) > 0, it follows thet \ = X and thus an eigenvalue must be real.190 ORDINARY DIFFERENTIAL EQUATIONS {Cuap, 7 If A: and Aq are distinct eigenvalues with eigenfunctions x: and x3, respectively, then (xux2) — Gcadlixe) = Or Aa) Ocxa) and by (2.1) this implies that (x1,x2) = 0. Let of = 9(t.D, 7 = 1, .. .,, be solutions of Lx = Ix which satisfy the initial conditions OPM Gl = ie k= 1... ym) (2.2) for some c in the interval (a,b). By Theorem 8.4 (also Prob, 7), Chap. 1, the functions gi" are continuous in (¢,!) for £e [a,b] and all J, and for fixed ¢ are entire functions of. Since the g; are linearly independent, the problem x has /as an eigenvalue if and only if there exist constants ¢, not ® all zero, such that c = > oye; Satisfies Uz = 0. This is the ease if and ist only if a ¥ ates = 0 (k=1,...,2) a hes a nontrivial solution. ‘This system of x equations for the +; has a nontrivial solution if and only if the determinant A of the matrix with Ung; in the kth row and jth column is zero, Because the gf? ure entire functions of / for fixed ¢, in particular at ¢ = a and ¢ = 6, it follows that A ig an entire function of 1. This funetion ean have only rect zeros because x has no nonreal eigenvalues. ‘Thus A is an entire function of 2 which is not. identically zero. Its zeros, which are the eigenvalues of 7, can therefore cluster only at 1 = ©. ‘This completes the proof of the theorem. The nonhomogeneous problem Le=Iz+f Ur=0 (2.3) will now be considered, where fe C on [a,b]. ‘This problem can be solved with the aid of the variation-of-constants formula, ‘Theorem 6.4, Chap. 3 (see also Prob. 21, Chup. 3). Let. the g; again be the solutions of Lc = Iz which satisfy (2.2). Fors S ¢ let eral) tet palrl) 1 itr) st eed) Kind) = Tay We, - +. OG) (2.4) PMD + gtd eid) st galtDSrc. 2] SELF-ADJOINT PROBLEMS ON FINITE INTERVALS 191 and for t <7 let K(t,,!) = 0. The Wronskian W(yi, .. . , on) in the denominator of K in (2.4) is a function of 7 only since +. = — pails) Wey ee os ed) = oo f pols) ao] as is clear from (6.5), Chap, 3. Clearly (a/K/ab)(r + 0, 7,1) = 0 for j=0,1,...,n — 2, since the determinant in (2.4) vanishes when any two rows are identical, Thus dK/at, j = 0,1, ..., — 2, is con- tinuous in (é,7,!) for ¢,r on [a,b] and all J, and is an entire function of l for fixed (7). Moreover, for j = x — 1 and 2, it is continuous in (¢,r,1) for allland fora $1 StS band fora St¢S7r Sb. Also one ane 1 geet © 40,0) — “sper — On = Tey As a function of i, K satisfies LK = 1K if t #7. From Theorem 6.4, Chap, 3, or from the above remarks, it follows that the function u defined by atl) = [° Keds) dr = f' Kee DI) de (25) is of class C* in ¢, entire inf, and Lu = Iu + f. The function K will now be modified so that the conditions Uz = 0 are also satisfied. Let Bir) = Ker) +), oexktd 2.6) fet where the ¢; are chosen so that for fixed r on (a,b), G as a function of ¢ satisfies UG = 0. That is, VG = UK + Yq =O = 1... 5m) im or, since U,K can be extended to be continuous fora <7 Sb, Daler = -UK (kel...,aj;057s0) (2.7) ft The right member in (2.7) is continuous in (7,l) for a $ + S b and alll, and is an entire function of ! for fixed 7. Since the determinant A is an entire function of 1 with zeros at the cigenvalues of 7, it follows that if A does not vanish identically (which is the case for self-adjoint problems), (2.7) determines the ¢; as functions of (7,/) continuous for 7 on [a,b] and for all | except the eigenvalues of +. Moreover, for fixed 7, the ¢, are meromorphic functions of |. Thus @ in (2,6) is determined except at192 ORDINARY DIFFERENTIAL EQUATIONS [Cuar. 7 eigenvalues of r. From the fact that (2.5) is a solution of Lx = iz +f it is clear from (2.6) that the function w defined by a = f) GurnyG) de (2.8) is a solution of (2.3) except at the eigenvalues of x. Indeed, the following theorem is true. Note that in this theorem x is not required to be. self-adjoint. Theorem 2.2. If for at least one value of l the problem x has no solution except the trivial one (which is always true for the self-adjoint case), then there exists a unique function @ = G(t,7,1) defined for (1,7) on the square @ St,7 S band for all complez | except the cigenvalues of 7 and having the following properties: (i) aG/at* (k = 0,1, ... 2 — 2) exist and are continuous in (t,r,1) Sor (i,r) on the aquare a § t,7 Sb and I not at an eigenvalue of x. More- over, 0G /at* for k = n — L and n are continuous in (t,r,1) for (t,r) on each of the trianglesa St Sr S banda Sz St S bandlnotat aneizenvalue of x. For fixed (t,r) these functions are all meromorphic functions of b. Gi) Soe tr +0, nl) ~ FE ~ 0,2) = ts (iii) As a function of t, @ satisfies La = la tf t #1. (iv) As @ function of t, G satisfiea the boundary conditions Ux = 0 for asrsb. The solution of (2.3) is given by the function u defined by (2.8). ‘The function @ is known as Green's function for x. The theorem has already been proved above except for the uniqueness. If there were two Green’s functions, G and G, for some J, not an eigenvalue, then as a func- tion of ¢, @ — G is of class C-~! and since G — G also satisfies Lx = Ix, 4 #7, it must indeed be of class C*. However, since | is not an eigenvalue, the problem 7 has only the trivial solution and thusG — G = 0, Tt is readily verified that for (1.1) sin Mé sin A(L — 1) Ostsrs) Wain ated =) ee fin Bren Dd 9 @srstsh and for (1.4) ae @sts7s1) 7 —eé ord = {ones @Srstst)Sec. 3] SELF-ADJOINT PROBLEMS ON FINITE INTERVALS 193 Noto that the latter G above is double-valued on? = 7, This is true in general for a°—'G/at*~! since it is defined to be continuous ina Sf = rSlandosr sts. It will be assumed now that / = 0 is not an eigenvalue of the self- adjoint problem x. This is no restriction since there must, in any case, exist a real constant ¢ which is not an eigenvalue; thus if Liz = Le — ex, the problem 11:1. = lz, Ux = 0, is again a self-adjoint problem because (cuv) = (u,v). Moreover, if \ is an eigenvalue of 1, then \ + ¢ is one of x, and conversely, and the eigenfunctions are the same for x and m1. Since / = 0 is not an eigenvalue of 1, G(t,7,0) exists. In the rest of this chapter this Green’s function for 2 = 0 will be denoted by G = G@(é,r) and it will be assumed that » is self-adjoint. Corresponding to this Green's function G, let ¢ be the linear integral operator defined for all fC on [a,b] by ane) = [P Gun) ar Tf fg 2 on [a,b], then (2.1) applied to u = Gf, v = Gy yields (4,80) = (Sf9) (2.9) From (2.9) it follows easily that (G/,f) is real. A further consequence of (2.9) is that, G(r) = (r,t) which also is a sufficient condition that 7 be self-adjoint. Indeed, let uy eC" on [a,b] and satisfy Ux = 0. Lets = Lu,g = Ly, Thenu— of and » — Gg are solutions of Lz = 0, Ux = 0, and therefore are zero. Thus u = Of and v = Gg and (Lu,v) = (u,Lv) follows from (2.9). The operator § is a type of inverse to the operator L in the sense that Lof=f Glu =u are valid for all f 2 C on [a,b], and ze C* on (a,5] for which Vu = 0, 3. The Existence of Eigenvaluest With G and § defined as above, it is clear that if \ is an eigenvalue and y an eigenfunction of x corresponding to A, then = Se @.1) Conversely, if pe C on [a,b], then Ge is of class Cv and LGy = ¢ so that (3.1) implies Lp = dy. Moreover, Up = 0 since UG = 0. If there exists a nontrivial ge C on [a,b] and a complex number p such that Gp = wy, then » is snid to be an eigenvalue of § and ¢ an sigen- An alternative treatment is given in Probs. 8 and 9.194 ORDINARY DIFFERENTIAL EQUATIONS {Cuar.7 function. What was proved above is that the eigenfunctions of ¢ are identical with those of m and the eigenvalues of G are reciprocals of those of x. The equation (2.9) expresses the fact that g is self-adjoint. It will be shown that such a self-adjoint operator must possess eigenvalues and in this way the result will follow for x. Use will be made of the Schwarz inequality Ia) S Is Nett Wf + oll S Isl + llet Lemma 8.1, The set of all functions {Gu}, where ueC on [a,b] and lull & 1,43 & bounded set of equicontinuous functions.} Proof. For n 2 2, @ is uniformly continuous on the square @ S ¢, +0. Hence, given any « > 0, there exists a § > 0 auch that Gir) — Ga a. {This proves that g is a completely continuous operator.Bec. 3} SELF-ADJOINT PROBLEMS ON FINITE INTERVALS 195 Lemma 3.2. The norm of § satisfics Isl = sup gum] (we on fab) Proof. By (2.9) (Gu,u) is real. If [lul] = 1, K(u,4)| S Iisull lel] s Iisll and hence » = sup |(Gu,u}| S gil. ‘lo prove the reverse inequality, note that. (S(u + v),¢ + v) = (Guyu) + (Ge) + 2R(Gu,») S alle + ol]? and similarly (Su — v),u — v) = (Gu,e) + (Srv) — 2R(Gu,v) 2 —alle — oll? Subtracting, it follows that 4W (Guy) S 2n(ull? + [fell*) (8.4) Su is not zero for we C unless u = 0, for, if it were, LGu = u would be zero. Letting » = §u/||Gujl in (3.4), where |lx|] = 1, gives ||Gul] S 9, which completes the proof. Theorem 3.1. Either \Igl{ or —\\gl| ze an eigenvalue for g. Remark: This not only proves the existence of an cigenvalue for §, and for x, but shows that an eigenfunction y corresponding to this eigenvalue is a solution of the extremal problem of finding a function we C on [a,b] such that (Guu) = f° (fP eter ute) dr) a at attains its least upper bound among functions with |[ul] = 1 or else its greatest lower bound, depending on which of these is larger in magnitude, Proof of Theorem 3.1. Suppose ||§|| = sup (Gw,u) for [ul] = lueC on [a,b]. Then there exists a sequence of functions 4,2C on [a,d], [Jum|| = 1, such that (Gum,ttn) — [SIL Let uo = |\gll. Since {Gun} is an equicontinuous, uniformly bounded sequence, { there exists a subsequence, call it {Gun} also, which is uni- formly convergent on [a,b] to a continuous function go. It will be proved go is an eigenfunction with eigenvalue po. Since max [Sun — gol + 0 (m— @) estab it follows that Sum — gol] +0 (m— o) 3.5) t A proof of the Ascoli theorem is given in See. 1, Chap. 1.198 ORDINARY DIFFERENTIAL EQUATIONS: (Cuar. 7 Also [Gul] > |lvol]. Now [Gem ~ pottmll? = Geel? + wéllesal]? — 2p0(Getmytém) G6) and the right side tends to |{yol|* — x3 as m—» », It follows that ileoll? 2 u§ > 0, and hence go is not identically zero on [a,b). From (3.6) it also follows that sinco [|Gunll? S ui, OS [Gum — portmll® S 2uh — 2po( Gremstm) which tends to zero as m—> ©. Thus [gem — Hotn|| + 0 7) But 0 S 1Se0 — zovdll S [Geo — S(Gum)fl + [G(Gua) — HoStmll + [lx0Stm — nodolf and using ||Gul| $ [III llull, (8.5), and (3.7), this yields [Geo — zovsll = 0, which proves Gyo = pave. If —[l$ll = inf (Gu,u), the proof is similar. Let xe = go/|lvol]. Then |]xoll = 1 and xo is said to be normalized. ts Gilt,r) = Gir) — woxo()Xo(r) and define ¢, for ue € on [a,b] by giut) = f Gxltz)ule) dr Then §; has the same properties as § was shown to have in Lemmas 3.1 and 3.2. In particular, if [gull = 0, and sup |(G14,x)| = [al where we C on [a,b], |[u{] = 1, and ,, real, then ,, is an eigenvalue for ¢,, and there exists a nontrivial g; ¢ C on (a,6) satisfying Sig: = me1. Let xi = gr/ljgill. Since (G1zz,x0) = 0 for any we C on [a,b], it follows that x1 is orthogonal to xo. Therefore Sx = Sixt = ix and hence x is an eigenfunction of §, From the extremal property, Waal & [aol Letting Gilly) = Abr) — maxi) and proceeding as above, the existence of xs and yz is established with lus] S [us|, and xs orthogonal to x; and xo. In this way the existence of an orthonormalt sequence {xs}, = 0, 1,2, . . . , is established. } The sequence {xe} is orthonormal if (xj,x4) = 3), the Kronecker delta.Sec. 4] SELF-ADJOINT PROBLEMS ON FINITE IXTERVALS 197 This process can terminate only if, for some m, [G,!| = 0. But for any f of class C moi Gf =$— } whxdby Ko With [|§.]| = 0, this implies mat f= Gxix 8) Ko Sinee the x; are of class C' and f ean be taken as |f — (a + 4)|, which is not Cl, (3.8) is impossible. Thus |!Gui] > O for all m and there are therefore an infinite number of eigenvalues and cigenfunctions, 4. The Expansion and Completeness Theorems The expansion in terms of eigenfunctions of x of a function fe C" on {a,b} satisfying the boundary conditions Uz = 0 will now be proved, From this the Parseyal equality and the extensions of these results to any function fe @*{a,b) will follow easily. First, an important inequality will be deduced. Lemma. If fe (a,b) and {xx} ie an orthonormal sequence for , then the scries > Wxal? k=O ia convergent, and > IGxel? S Ili? (Bessel’s inequality) Koo Proof. For any finite m 2 0 OSM — Y Gaxdeal? = lt — Yok? iv 2, (irerel ds fixe which proves the convergence of the serics in question, and Bessel’s inequality. The number (f,x2) is called the kth Fourier coefficient of f with respect to the orthonormal set [xs]. Theorem 4.1. Let feC* on [a,b] and satisfy the boundary conditions Uf =. Then on [a,b] f= y (faxed xe (4.1) so where the serica converges uniformly on a,b].198, ORDINARY DIFFERENTIAL EQUATIONS: (Cuar. 7 By multiplying (4.1) by j and integrating, there results the following: Corollary. If f is as in Theorem 4.1, Ii? =F iGixol? — Parsevat equality) e=0 ‘This is also called the completeness relation. Proof of Theorem 4.1. From J. Gtbadx) dr = rail) it follows that the kth Fourier coefficient of the function g of 7 given by g(r) = (tr) = G(r.) for fixed ¢ is uxZe(), The Bessel inequality yields . ’ > taco s [P 1ee.nk a x= for all m. Integrating in ¢ and letting m— », Las vo-@! #2 where y = sup [G(z,!)| ona S t,7 Sb. In particular, |ux|—+Oask— @: Consider for an integer m 2 1 m-1 Galle) = Gtr) — FY mxietr) ko From the extremal property of gn there follows ||Gnil = [zm{. Thus for any we on [a,b], m=) Ugmeall = |] ge — Y sateyaadace |] S ll hel Eso or since |un| + Gas m— ©, im | gu - 5 weltxadxa | = 0 (42) For any q > n 4 s pe(Uyxe)xa = GF 6) (usd) kop kop Since |Gu| S y(b — a)'||ul, it follows that [5 astuardas| s 00 — 0) (5 loool)! kan kevSec. 4] SELF-ADJOINT PROBLEMS ON FINITE INTERVALS 199 By the Bessel inequality, the Jast sum tends to zero as p,g-—> ©. Thus x oa(U, Xe) Xx kd is uniformly convergent on [a,b], and therefore represents a continuous function there. Since Gu is also continuous, (4.2) implies that gu = > see (1,%4) x8 (4.3) eno Given any fe C* on [a,b] satisfying Uz = 0, then x = Lfe C on [a,b] and f = §u. From (4.3) the expansion result (4.1) follows, since wn(yxe) = (teynaxe) = (ExK) = (Geoxx) = Gyre) Remark: The fact that ||Gnl|—» 0 ns m—» © suggests the possibility that, since ge = 1/rs, Gltx) = » xe(OXe(7) de kad This is, in fact, correct but will not be proved here. [t wiil follow from results of Chap, 12 where much less restrictive results than Theorem 4.1 will be proved. The expansion theorem and completeness relation will now be extended to the whole space {*(a,b). Theorem 4.2. If fe (a,b), then f= 5 (S.xe)ae E50 where the equality ts meant in the sense m lim ||f- ¥ Uxoxl|=0 (4A) =e k=0 Further, Parseval’s equality holds: wit = Six keo Proof. The proof depends on the fact that the set of functions of class C? on [a,b] which satisly Ux = 0 is dense in the space 2*(a,b), that. is, given any ¢ > 0, thero exists such a function f satisfying If -fl ™) and hence (4.5), (4.6) yield |7- 5 (idu|] 1) proving the expansion result (4.4), Parseval’s equality follows directly from (4.4) since [7-3 Gane = ine - 3 tor The Parseval equality has the consequence that if (f,..) = 0 for k=0,1,2,... , then fis zero almost everywhere, and in particular, if f is continuous, it is the zero function on [a,b]. A set of functions {y} is said to be closed in 2°(a,b) if, for every fe 2*(a,b), (7,4) = 0 implies f is zero almost everywhere. Thus the set {x.} is closed in 2%(a,b). This implies that if fe &(a,b) and (f,x.) = 0 fork = 0, 1,2, ... , then fis zero almost everywhere; that is, the set (x) is closed in 2(2,b), the set of Lebesgue-integrable functions on (a,b). This can be seen in the following way. Suppose fe f(a,b) and (f,x2) = 0,k = 0,1,2,.... Then there exists a continuous function » satisfying ie=f Ug=0Sec. 4] SELF-ADJUINT PROBLEMS ON FINITE INTERVALS 201 since zero is not.an eigenvalue far x, The proof consists in verifying that, for fe 2(a,b), ¢ = Of is of class Ce fora SiS b and that g@— is absolutely continuous so that Ly = f almost everywhere. Clearly (exe) = Ne elxe) = EL) = 2"Chx) = 0 which proves ¢ is orthogonal to all cigenfunctions. But since g is con- tinuous, ¢ is the zero function, and this implies, by Ly = f, that f is zero almost everywhere. Corresponding to any f ¢ €°(a,L) there is 2 unique sequence of complex numbers ¢ = {cy}, where ce = (fix) Define the norm of ¢, denoted by Me", by ill = ( teal?)! kno Then the Parseval equality may be written as ||] = lel. It is an impor- tant fact that the correspondence f-— ¢ actually uses up all sequences ¢ of complex numbers such that |jel| < ©. ‘Chis is the Riesz-Fischer theorem, the proof of which has nothing to do with differential equations, and so will be omitted. + Riesz-Fischer Theorem. Let c = {cx} be a sequence of complex nwmbera such that |lel] < 0. Then there exists a function fe ¥?(a,b) for which ce = (fxn), and Wl = Nett PROBLEMS 1. let Lr = =(px")’ + qx, where p is of class C! and ¢ of class Con [a,0] aud p # 0 on [a,b], Let Uz = O be given by az(a) + 62a) 20 y2(b) + 82") ~ 0 Show the problom w is self-adjoint if and only if p and gare real, 78 = 54 and af! = a6, which is equivalent to requiring that a, 8, 7, and 6 all be real. If Uz = 0 shove ix replaced by x(b) — ac(a) — Ar'a) = 0 z'{b) = y2(a) ~ bra) = 0 show that the conditions for self-ndjointuess become a = ce, 6 = exc, 7 = exe 3 = cei, where ¢, and @ are real and plbrieier — eres) = pla). " 8. tet Ee = VS (nyt), whore pa, C6 F and are real on [a,b] and poll) #0 2 Fx) on (ab). Let Ux = 0 be 20a) = 20(b) = 0,7 =0,1,...,2—1. Prove x is self-adjoint, { For a proof, see W. Rudin, Principles of mathematical analysis, New York, 1953.202 ORDINARY DIFFERENTIAL EQUATIONS [(Cuar. 7 4 Let + be a self-adjoint problem with orthonormal cigenfunctions {xa}. If the norm [Fj of a function F = P(t,r) of class C for tr ¢ (a,b) ia defined by wi= (f° [rast aa) then it is possible to approximate F(f,r) in this norm by finite sums of the form Lausfilgsle), where fi and g; are of class C and satisty Uj; = Ug) = 0. Show that the Parseval equality holds for F in the sense that if caf? [P Ftnerete ade then Fit = teu ‘Show that this leads to ain ownn - 3 Sener -0 Rewane: Actually the formula Gira) = Smee k=O is valid, as will follow from considerations in Chap. 12. &. Let. [a,b] be [0,1] and let Lz = —((i — #)z’)' and Uz = 0 be x(0) = 0 and 1 =£)2'()-—> Oast—- 1-0, Show that by taking l+r lr 2 glgttt! Ostsran dog Osratan Gir) = the reasoning of the chapter can be modified to obtain the existence of a complete orthonormal family of cigenfunctions. . Hint: Instead of Lemma 3.1, show that ¢ is completely continuous in the norm of @°(0,1). This is true for any G with ft L Ganhdide < © 6, Show that if + ia self-adjoint the Green's function G(é,7,l) satisfies Gr) ~ G64) thereby generalizing the result @(¢,1,0) = G(+,t,0) already shown when / = O ia not an eigenvalue. 7. Show that if x is self-adjoint the poles of G(1,r,!) are simple poles. Mist: Consider 9(6,1) = f G(r Dflr) de for f of class C. Let g have n pole of order m > lati =A, Then near! = Xs an(t) an-60) ot = phe + peat:Prous.j SELF-ADJOINT PROBLEMS ON FINITE INTERVALS 203 Since (L -— g ™ (L - Odg +f and Ug = 0, it follows that (L — )gm = 0, (EL — Mgmnt = omy os and U(g;) = 0,j = m,m—-1,. Since Gmidm) = (mL ~ M)GmWt) = (HL — ra)aniGnai) = 0 it follows that gn = 0. Since this holds for all f, @ has at most a simple pole at As. & If + is self-adjoint and fe C on (a,b) and (j,xx) = 0 for all the eigenfunctions of x, prave that f is zero by making use of g(é,!) of Prob. 7. Hist: Using the method of Prob. 7, show that because (f,x) = 0, g has no poles and is therefore an catire function of , > a). Since Up = Oand Ly = lg +f, ind show Lay =f, Lay = a9, Lar ax, ,. , Vay = 0, Show (a)-1,04) = (ajas—1) and thus that Wi4a = (aja) depends only on j +. Show that AD = (g,a0) = Wo+ Wil + s+ is an entire function of fas is P(t) = Wo + Walt + Walt 4- ++ ~ Show Wi, = Gj-naie)? & Way-2Waiys Thus, if Ws » 0, Was < Wier Waa Wa FNM meaning Dis not entire, ‘Thus W3 = 0, and soa, = 0, Lay = a: = 0, Lao =f =O. Resang: Note that Prob. 8 gives an independent proof of the closure of the eigen- functions of x. 8. Using the result of Prob. 8, prove Theorem 4.1. Hit: Let u = Lf. ‘Then, as was shown beginning just below (4.2), « a= > Het) X) x4 o is uniformly convergent on [a,8]. Clearly f — 9 is orthogonal to all xy. ‘Thus f — g is-zero by Prob. 8, which proves Theorem 4.1. 10. If Green's function for (1-1) ia expanded in the eigenfunctions, the series S 2sin ket sin ker Gerd) = Y Ben tetoin ber kel is obtained. From the nature of G(t,r,/) as a function of ¢ and the convergence proper- ties of Fourier sine series, show the series is convergent for 0 S$ ¢ 1, for all 7, and t not ancigenvalue. Show that the sories for 6G/d! also converges for all ?, (Note that 8G /at is of bounded variation as e function of 4.) 11, Let L and U be such that « is aclf-adjoint. Consider now instead of « the problem Lz = irz, Uz = 0 on [a,b], where the function reC and r(t) > 0 on [a,b]. Show that the eigenfunctions {¥x] can be chosen so that {riyx} form an orthonormal sequence. Show that this sequence is complete. Hint: Let H(4r) = ri(Ori(r)G(4,7) and show that the operator 5C defined for alt 8 Con [a,b] by I(t) = f * HUG) dr is eelfadjoint.204 ORDINARY DIFFERENTIAL EQUATIONS [Crar. 7 12. Show that G(¢,r,l) is given by Gitrt) = ole) where A(l) = det (Ujps), Kbit) eG = ++ wally attr) = det aK Ove sos Uren UsK Une, +++ Unen and K is the function defined by (2.4). 18. Let Le = port + + + + pax on [a,b] and pe C*-sIa,b] and be real. Then L = L* if and only if nis even (n = 2r) and La can be written as Le = (gor) (gaye ee. page where 9, € C~fa,b] and is real, Hint: Show that, if L = Lt and po is real, then n = 2r and pr = rp). Thus if qo = pe then Le (ger + Lye where Ly must be of order n — 2. Show that L = L* now implies L; = Lf and thus establish the result by induction. 14. Let Lz be as above but now jet p;(t) be complex-valued. Show that if L = L+ then Liz = igo + + (golgoz)'Y «YF AIG + giao Ye Ht Fn 2(Gnet(Gn—s2)")’ + ie. jai)’ Gat where the gy € G*~! and (g)"*~/ are real and ff! = pe. Hint: Use induction, Show by direct consideration of fF sraat = + Caotqanyy «= yt that the first term of L above is self-adjoint. 18. Show that if Z = L+ there exists at least one act of boundary conditions Uz = 0 which makes the problem Lr = iz, Uz = 0, aclf-adjoiat, Hine: If n is even (x = 2r), then tako x(a) = x(b) = 2’(a) om z’(b) mm ale (a) @ 2 (b) = 0 If n is odd (n = 2r + 1), take the above conditions and the added condition Cy2t0(a) = Cy2() where iCiC1 = pf?{a), 1CxCy = pf (b). 16. Let A bea square matrix and fa vector both of which are continuous functions of ffora $t 6. Consider the problem a= AWz=f Uz = Mz(a) + Nz(b) = 0 where Af and N are constant square matrices, Let the problem with f = @ have only the null solution, Show that there existe a matrix G(ér) continuous for a $7 SProps.) SELF-ADJOINT PROBLEMS ON FINITE INTERVALS 205 + S band fora Sr StS b such that fC Gunn a ig the unique solution of the problem. Hint: Let © be a fundamental matrix for 2” = A(Dc. Let. PQS Mr) + ODT (r) (o where the matrix Je Cla,b]. To satiafy Uz = 0 Mba) (2) + N&(b)o-r) + NEQ)J (7) = 0 50 that H(t) = = (Mb(a) + N44) Nab) b-HE) 17, Lot the r-by-r matrices Poand P, be continuous fora St Sb. Moreover, let Py be continuous and det Po(t) #0. Let z be a vector with r components and let Le = Por! + Pye Let P* denote the adjoint of P, that is, the transposed conjugate, and let bee = —(Ppr) + Pit If wand v are vectors with components u,, ¢;, let wee sib bot tab Show that if uv € C'{a,b] Lu-p —u-Ltv = (Pou 0)’ Let L = Lt, thatis, Pot PZ = 0, Py = Pi ~ Pl. Let Bf and N be r-by-r constant matrices and Iet Uz = AMz(a) + Nz). Suppose af and N are such that for any usr €Cia,b] and satisfying Ux = Uv = 0 fPrurvdie Putea a [a Prove the eigenfunctions {¥[a,b} satisfying Un = Uv = 0 flrwvam Putedt Show that the self-adjoint problem lente Uz=0 has a complete orthonormal ect of eigenfunctions {vs}. Hint: Let € be the vector with nr components (z, z’,... 2-2), Then Lz—-izm0 can be replaced by a first-order equation in € with nr independent solutions with fundamental matrix =, nr-by-nr, Let the first r rows of = be denoted by $, Then Le — lb = 0 and any solution ¢ of Lz — lz = Ois given by $c, where ¢ is a constant column vector with nr rows. The Green’s function G(¢,7,l), an r-by-r matrix, ean be constructed, 19. Let L be as in Prob. 13, that is, Lz = Gor )O + iz DYOD ees ge where gj ¢ Ci[a,b], qs real, and ga(t) # Oon [a,b]. Let y be a solution of Lz = O and let ¢ be the vector with components vj, where gj = wt? Gm], .. 47h Pred CATE 999) (genset! be oe eb (gop “a, G=1,...,7). Show that ¢ satisfies the formally self-adjoint system Po + Pig =0 , Py = Pt, are the matrices mm (4 o) mG) where Pp =Prozs.) SELF-ADJOINT PROBLEMS ON FINITE INTERVALS 207 where 0,, EZ, are the zero and unit matrices of r dimensions, and A, B, C are r-by-r matrices given by “de 0 re 19 , 0 (-Iyr/q the elements not shown being zero.CHAPTER 8 OSCILLATION AND COMPARISON THEOREMS FOR SECOND-ORDER LINEAR EQUATIONS AND APPLICATIONS 1. Comparison Theorems The location of the zeros of the solutions of real second-order differential equations will be considered here. The equation will be assumed to have the form Lz = (p()z')' + 9x = 0 (a 0 and that p, p’ and g are continuous on (2,b). (The continuity requirement can be relaxed. Indeed, it suffices for g to be integrable and p absolutely continuous.) A zero of a nontrivial solution of (1.1) isisolated. Indeed, let the solu~ tion g vanish até. Then g'(lo) # 0, for otherwise o(t) = 0, This proves that é is an isolated zero. Theorem 1.1. Suppose y is a real solution on (a,b) of (pr)! + gt = 0 (1.2) and y¥ a real solution on (a,b) of (px')' + gat = 0 (1.3) Let g(t) > gilt) on (a,b). If tr and ty are successive zeros of p on (a,b), then y must vanish al some point of (t,t). Proof. Suppose y does not vanish in (¢,,42). Then with no restriction it can be assumed that ¥(/) > 0 and also g(¢) > 0 over (é,,2). Multiply- ing (1.2) by ¥ and (1.8) by ¢ and subtracting, (pet ~ (phy'e ~ (ga — gidey = 0 Integrating the above, [lore - (vy'elat > 0 208Sere. 1] OSCILLATION AND COMPANISON THEOREMS 209 Since the bracket above is the derivative of p(y’¥ — yy’) and since ¢ vanishes at ¢, and é2, Pltee’ (tele) — ple’ ¥(h) > 0 (14) Since ¢(tz:) = 0 and ¢(é) > 0 immediately to the left of 2, ¢’(t2) < 0. Similarly ¢’(4;) > 0. Thus the first term on the left is nonpositive, as is the second, which shows that (1.4) is impossible. Thus y vanishes at a point inside the open interval (t1,¢2). In case g, = gz over (a,b), then y and y are solutions of Lhe same equa- tion. Hf y and ¥ are independent, then (1-4) is valid with the inequality replaced by an equality, and the above argument shows that ¥ vanishes between successive zeros of g. Since now ¢ and y are interchangeable, it also shows that y vanishes between successive zeros of ¥. Thus the zeros of two real linearly independent solutions of a real second-order linear differential equation separate one another. The above method can be further exploited, but the following pro- cedure is simpler: Let p(é)z’ = y. ‘Thus (1.1) becomes r= a yo = ~gz (1.5) Let z=rsind y=reos0 (1.6) Differentiating the equations (1.6) with respect to ¢, replacing x’ and y by use of (1.5), and then solving for r’ and 0’, there results v(t ) . =(-- sin @ cos @ 17 t (; rsin @ cos (7) v= pout 8+ gsinto (1.8) For a solution y of (1.1) there is the solution r = p(¢) and 6 = w(t) of (1.7), (1.8), where from (1.5) and (1.6) v= (p+ eo = tant (4) Since y and y’ do not vanish simultancously, it follows that p7(2) > 0 on (a,b) and thus with no restriction it can be assumed that p(é) > 0. A consequence of this is that ¢(é) = o(t) sin w(¢) can vanish only where w(t) is an integer multiple of x. Since cos? @ and sin? 6 are uniformly bounded, the equation (1.8) has asolution over any interval on which p > 0 and p and g are piecewise con-210 ORDINARY DIFFERENTIAL EQUATIONS (Citar. 8 tinuous, Indeed, it suffices for 1/p and g to be integrable. Because the right side of (1.8) is differentiable in 9, it follows that the soiution is unique in the usual sense, From (1.6) and (1.5) it follows that 2(t) cos @ — p(i)2’() sin @ = 0 (1.9) In boundary-value problems a common condition at an end point of an interval £ = ais x(a) cos a — p{a)z’(a) sina = 0 (1.10) From (1.9) it is clear that such a condition is equivalent to the simpler condition 6(a) = a(mod =). It is easy to see that (1.10) cannot hold for a solution x = g(¢) for two different values of « unless they differ by a multiple of # or unless g*(a) + (py'(a))? = p%(a) = 0. The behavior of the solutions for two equations of the form (1.1) will now be compared, The subscripts 1 and 2 will be used to distinguish between the two equa- tions, that is, Lie = (pz’)’ + gx = 0,7 = 1,2, Theorem 1.2. Let p; and g; be piecewise continuous on [a,b], and let O< pl) Spl — grlt) 2 alt) on [a,b]. Let Ligi = O and Lig: = Oand let w(a) 2 wi(a). Thea off Bott) (@stsbdb) (11) Moreover, if gz > gx on (a,b), then w(t) > wilt) (a @ such that wrt) = o(f) (a St Se) (L.E7) Indeed, suppose this is not the case. Then by (1.11) there must exist a sequence of points [tj] with @ as a cluster point such that wa(l;) > wilt;). But if (1.16) is used with a replaced by ¢;, it follows that for é > ¢; there results w(t) > wi(f). With 4 arbitrarily near @, this implies (1.12). ‘Thus (1.17) must hold. Using (1.17), then (1.14) is possible with gz > gi only if wy = we = O(mod 7) and if p, = p2 over (a,c). However, in (1.13) the case w, = w, = O(mod 7) over (a,c) is clearly impossible. ‘his proves (1.12) if ge > gi 2, Existence of Eigenvalues Application will now be made to the equation (pz')! + Or — g)z = 0 (2.1) where \ is a real parameter and p’, 7, and q are real and continuous (or piecewise continuous) over [a,b] and p > 0,r > Oover[a,b}. [By modify- ing the proofs that follow slightly it is possible for r to vanish at a and at b as well as at isolated points itNa,b).) Given real @ and 4, the values of A for which (2.1) has a solution not identically zero and satisfying (a) cos a — p(a)zx"(a) sin a = 0 (2.2) x(b) cos 8 — p(b)x"(b) sin B = 0 (23) are called eigenvalues, Wither one of the conditions (2.2) or (2.3) deter- mines a solution of (2.1) uniquely except for a multiplicative constant.212 ONDINARY DIFFERENTIAL EQUATIONS, [Car. 8 A nontrivial solution satisfying (2.1), (2.2), and (2.3) for an eigenvalue is called an eigenfunction. Theorem 2.1. There are an infinite number of eigenvalues do, di. Az, - forming a monotone increasing sequence with, — © asn—> =. Morcover, the eigenfunction corresponding to d, has exactly n zeros on (a,b). Proof. There is no restriction in assuming that 0 S a <-# and that 0 0. This means that » is an increasing function of ¢ when w = O(mod wr). Thus if for some & on (a,b), w(h,A) = Ae, then (fA) > ke for t> & and a(t,A) < ke for é O when w = 0(mod -), it follows that the leca- tion of the kth zero of g on (a,b) att, = &(A) ¢s a continuous and monotone decreasing function of .. Indeed (tan) H+ 22 (an) = 0 It is the case that for any fixed ¢ = ¢ in (a,b], w(6d) > 9 as h—> 0 (2.5) and also w(e,\) > 0 ash—> —% (2.6) The proof of (2.5) will be given first. Since a 2 0, it follows that (44) 20 since w > 0 for w = O(mod x). Thus it suffices to show that for some io, a < ty © as A> &. Let to = (a+)/2. Let P, Q, and # be constants such that over (toc) pWSP rQOZR>O gOSQ Then the equation Pe" +0R - Or =0 (2.7)Sec. 3] OSCILLATION AND COMPARISON THEOREMS 213 with solution ¢ satisfying g{t,A} = eCto,A), Pe’(to,A) = p(to)e"(do,A) has (to) = e(fo,A) and thus by ‘Pheorem 1.2 (6h) — w(to,4) 2 G(e,4) ~ G(to,d) (2.8) The successive zeros of ¢ haye spacing #{P/(AR — Q)}. This tends to zero ash—» ©, ‘Therefore @ = 0(mod x) for arbitrarily many values of 4, and since & > 0 at @ = O(mod x), @— ©. Thus the right side of (2.8) tends to infinity as \— «©, which proves the left side must do the same. This completes the proof of (2.5). To prove (2.6) the equation (2.4) is used. Choose 5 > 0 small enough so thata 0. Since 6 > 0 and since w(b,d) is monotone increasing in \, it follows there is a vulue of A, do, for which w(b,r.) = 6. Since 0 S a 0, r > 0, on ja,b], and r, p’, and q piecewise continuous. It will be assumed that p{a) = p(b). With no restriction, it can be assumed that a = 0, b = 1, and p(Q) = p(l) = 1. The boundary condition (see Prob. 4) 2) = 20) —-2(0) = 2”(1) (8.1) will be considered, as will the condition 2(0) = —2(1) 2’) = —2'(1) @B.2)214 ORDINARY DIFFERENTIAL EQUATIONS (Cuap. 8 Theorem 3.1. The eigenvalues for (2.1) with (8.1), di, 7 2 0, and for (2.1) with (3.2), X,, i 2 1, form sequences such that ~o cd O ford odd and < 0 for t even. From (3.5) follows ¢(1,u)¥'(,u)) = 1 so that 1 1 Sa) = Vad) + Faw Since for real x > 0, z + 1/z = 2, the results (3.12) for wu; now follow. If vo is the least eigenvalue of (2.1) with 2’(0) = 2’(1) = 0, then o(t,v0) is the eigenfunction and it has no zeros in (0,1). Thus ro < uo and e(1,ro) > 0. Since y’(1,r0) = 0 it follows from (3.5) that (1, vo)¥"(1,r0) = 1 Thus Keo) = ein) + sory 2 which completes the proof of (3.12). In order to consider df/dA, where f = 2 or —2, the funetion u = 4¢/00 is considered. Clearly (0,4) = u’(0,A) = 0 and from (2.1) (pw) + Qr — gu = re Thus from the variation-of-constants formula (or as can be verified directly), / wr) = fe aHerd) — el AHUANCeD) dr Thus. Be (1) = [ leLAWOA) ~ eAWLAIWA) de (6.16) and in the same way An) = [Waar - ee AWUNKWEN & Thus, not indicating \ explicitly, L = Peery) + verelrlott) ~ HCD) ~ eK) & 1D ‘The bracket in (3.17) regarded as a quadratic form in y(r), ofr) does not change sign if (e(1) — ¥’(1))? + 4y’(I)¥(1) $0. Using (3.5), this becomes [eQ.d) + WGA S 4Sec. 3] OSCILLATION AND COMPARISON THEOREMS. 27 Thus, if —2 S f(s) S 2, the bracket in (3.17) has a fixed sign. Hf SQ) = 2 or —2 then, except possibly for a factor —1, the bracket is a perfect square and df/dd cannot vanish unless the bracket is identically zeroinr. Because ¢(r) and 9(r) are independent, the bracket is identi- cally zero if and only if all the coefficients vanish, which together with (3.5) is the condition (3.7) if f = 2 and the corresponding one if f= —2. Thus df/d\ = 0,where f = 2 or —2, if and only if the eigenvalue is not simple. IfA < poor if ps < dX < piy1, then ¥(1,A) ¥ 0 and thus, iff = 2 or ~2, the bracket in (3.17) is not identically zero. Being a perfect square, df/dd has the same sign as —¥(1,4), which proves (3.13), There remains only the proof of (3.14). AtA = say. then, f = 2 and dj/dy = 0 so that (3.7) holds. Thus WO wria) = evar) =O CE water) = e(l wea) = 1 (3.18) Using the notation a= Zany a= ay and similarly for gj, ¥i, then of mm te 8.19) From (3.5), differentiating with respect to d gives the’ + Wei ~ hie - Ven = 0 (8.20) Taking account of (3.18), there results VAC uae) = —ea(1 eater) (3.21) Differentiating (3.20) again and using (3.18) and (3.21), BWrvr + 2k — Vin en =O (A = paisa) In (3.19) this yields as . , 4 TB Chuan) = eR L wares) + drCL mains) eC waren) (3.22) Using (3.15) again, it follows from (3.16) that en(L acer) = ft WrusenoCrusae(r) dr218 ORDINARY DIFFERENTIAL EQUATIONS {Cuap. 8 and in the same way WlLanies) = ff vGuaiedr) de ; 1 (3.23) ek(Lasisa) = — ff o%eranngade(e) dr Since ¥ and y are independent, the above relations and the Schwarz inequality imply that the right member of (3.22) is negative, which proves (3.14) and completes the proof of the lemma. 4. Stability Regions of Second-order Equations with Periodic Coefficients Here the real equation (p)z'Y + [ar(t) + baz = 0 @1) will be considered with a and 6 constant and with p > 0,7 > 0, and p, p', r, and q continuous over 0 S$ ¢ S 1 and also periodic of period 1. That is, r(0) = r(1), ¢(0) = q{1) and, as can be assumed with no restric- tion, p(0) = p(i) = Lb. In Sec. 5, Chap. 3, the existence of characteristic exponents and multipliers was proved. If z = p(t,a,b) and z = ¢(¢,a,b) are solutions of (4.1) with 10,0,b) = y'(0,a,b) = 0 = ¥/(0,a,8) = 9(0,a,b) = 1 then for fixed ¢, ¥, ¥’, , and ¢’ are entire funetions of (a,b) for all a and b. To determine the multipliers, the solution z = Cig + Coy is considered which satisfies Cyo(1,a,b) + Cof(1,4,b) = oC, Cry'(h,a,b) + Cap’(1,a,b) = oCa for some ¢. For a nontrivial solution (C,,C2) to exist, the determinant of the coefficients must vanish, which gives the characteristic equation ot — oly(1,a,b) + ¥/(1,4,5)] + 1 = 0 (4.2) where use is made of (3.5). If $(a,b) = p(1a,b) + ¥'(1,0,b) then the roots ¢; and o2 of (4.2) are distinct complex conjugates of mag- nitude 1 if Fab) <4 (4.3) while the roots are real and distinct if Pad) > 4 (4.4) Because a,c; = 3, in this latter case one root is always larger than 1 in magnitude and the other less than 1.Sec. 4] OSCILLATION AND COMPARISON THEOREMS 219 If the roots are distinct, two independent solutions exist, 2 = 2(t)e%! and 2 = un(te%', where uw; and wz are periodic of period 1 and e* = a, i= 1,2. Thus, in case (4.3) prevails, all solutions of (4.1) are uniformly bounded over 0 <¢< @. If (4.4) prevails, this is certainly not the case, even over (— ©,0) or over (0,~). Therefore, in this section the values of (a,b) for which (4.3) holds will be called stable while those for which (4.4) holds will be called wnstable. From the continuity of f, it follows that the stable regions and the unstable regions of the (a,b) plane have their boundarics made up of points where f*-(@,b) = 4 or, in other words, of points where either S(a,b) = 2 (4.8) or f(a,b) = -2 (4.6) For any fixed b the equation (4.1) is of the form (2.1) with a in place of 4, Thus the conditions (4.5) and (4.6) are precisely those already con- sidered in connection with the cigenvalues of (3.1) and of (3.2). When (4.5) is satisfied, the equation (4.1) has a solution of period 1 while (4.6) corresponds to a solution which satisfies (0) = ~2(1), 2’(0) = —z2’(1) and thus has period 2. It will be designated as having half-period 1. From (3.15) it follows that for any fixed 6 the values of a, a;(b), 7 = 0, 1, 2,... ,at which (4.5) is satisfied and 4,(6),7 = 1,2, ... , where (4.6) is satisfied, are related by =~ < ao(b) < Gib) S wold) S da(b) < ai(b) S wr(b) S an(b) < Ga(b) S un(b) S Gib) < as(b) S++ (4.7) where 4,{b) are the cigenvalues of (4,1) for fixed b with the conditions 7(0) = 2(1) = 0, ‘That the u; are continuous functions of 6 for each ¢ follows from the fnet that (0¢/de)(1,0,b) # 0, where ¥(1,¢,b) = 0, since, much as in (3.23), under these conditions 1 3 (1.0.8) = o(1,a,b) f rer) dr Tt remains now to show that a, and 4; are continuous single-valued functions of b, ~ 2 <6 < =, Before showing this, the consequences of this fact will be considered. For fixed 6 and all a satisfying aai41(b) < a < ax42(b) it follows from (3.12) and (8.15) that f(a,b) > 2 and thus that (4.1) is unstable. In the same way, dai 0 such that ¢(t) > 0, (2) > Don (aa +8). Let lim p@)le'y — o¥'] BO 040 Prove that if (fy) = O forsome f: on (a,b) then there is 8 ty on (a,4,) such that y(t) = 0, 8. Consider (2.1), (2.2), and (2.3) over the interval [a,b] but now let ¢ < 0 on (a,b) and suppose r changes its sign on (a,b) while, as before, p > Oon (a,b]. Show that the cigenvalues have \ = += and 4. — o as cluster points. Hunt: For the cage A > 0 consider Ge) ¢-Dene ‘Then as 2 increases, p/A decreases and r ~ g/A increasea, Let (a,) = sin a, pa)e'(a,4) = coga, Then let @ = w(t.) satisfy w(a,) = a. Clearly (6,d) is an inereasing function of >, and «(,0) is bounded, To show w(0,h)—~+ © as A> ©, consider an interval [f1,/:] on which r(!) > H > 0, where Risa constant, Let p(t) < Pon (fyfs). ‘Then for lorge A, the solutions of { Pz" ++ 5 Rz = 0 have zetos apaced r (ey on finfs|. Thus w(0,d) > edt for somo constant c. A similar argument applies as K+ =Pross.] OSCILLATION AND COMPARISON THEOREMS 221 4. Consider (2.1) p', g, and r continuous and p > 0, r > 0 on [0,1]. Develop the conclusion analogous to that of Theorem 3.1 for the boundary conditions 2(0) = az(1) + b2"(i) 2'(0) = ex(1) + dz"(1) where a, 8, ¢, and d are real constants and where (ad — bc)p{0) = p(L). Hint: Show that the eigenvalues occur at the roots of f(\) = 2, where FO) = ag(E,d) + by'G,A) + (LA) + dy) Show that if a; are the eigenvalues of 2(0) = 0, ax(1) + br'(1) = 0, then Pui) Ba and f(u:) has alternating signs. 5, In Theorem 3.1 let Asin1 < dsc4. Show that (2.1) with ¥ = dary: has a solution vast independent of yr41 such that Pre = Pail + fone where pay iso periodic function with period t. Show that similar results hold for Dares, Reva, and Kavya. 6. Using the notation of Sec. 4, enn an &,(b) and a:(d) ever intersect? What ia the significance of Prob. 10, Chap. 3, in the terminology of Sec. 4? Sketch possible configurations of a,(b} and d;(b) showing stable and unstable domains in the (a,b) plane for (4.1) with p(Q 3 rt) = 1, 7. In (4.1) leva > Oand f* q(t) dt = 0. Show that if u isa real solution satisfying u(t +1) = dou(t), where do is a constant, then v must vanish at at lenst one point in the interval (0,1), Hint: If not, V pu’? fl [Per ate frawo Show that « vanishes at two points 4, t, where (fs ~ 41 & be 8. Let f be real and of elnss C* on [a,b], und let f(a) = f(0) = 0 and f > 0 on (@,0). Prove that ona [Olas Hint: Let f attain ita maximum atc, Then for some 7, and rz 1 1 1 [10 -f@ fhe fo =10) f@ Lo ema fey — f'n). fe vrata ee s [re ®. Let r be nonnegative, continuous, and of period 1. If ffrowss show that 2” + r(!)z = 0 has stable solutions on (— ©, #). Hint: Use Probs. 7 and 8,CHAPTER 9 SINGULAR SELF-ADJOINT BOUNDARY-VALUE PROBLEMS FOR SECOND-ORDER EQUATIONS 1, Introduction ‘The treatment of Chap. 7 fails to apply in case the finite interval (a,b) becomes infinite or in case the coellicients in the differential operator have a sufficiently singwar behavior at a orb. These cases are all regarded as singular, and the second-order singular ease will be treated in this chapter. Asa preliminary example, the problem mat lp 2(0) = 0 ( = 4) for 0 S$ ¢ < = will be considered as a limiting case of the problem on the finite interval 0 $ € S$ b with the condition x(b) = 0 added, and then letting b-> «@. The finile interval problem of course gives rise to the orthonormal system {¥], where ONE pb v(t) = (?) sin At! (h=1,2,...) Any femetion f which is continuous over 0 $ ¢ S$ cand vanishesfor¢ 2 ¢ satisfies the completeness relationship [iuora= EYL sin M509 ae in the event that > oc. If ats) = fi sin st $00 at (1a) thon the completeness relationship becomes L OP dt = ) lo () f (1.2) 222Sec. 1] SINGULAR BELF-ADJOINT BECOND-ORDER PROBLEMS 223 where the c in the left side has been replaced by © since f vanishes for t> ce, Let p be a nondecreasing step function of s which increases by 2/5 when s passes through kr/b, (k = 1, 2, . . .), and is otherwise con- stant. Assume also that has been normalized so that p,(0) = 0. Then (1.2) can be written as Lo sor at = f° latent? dostay (13) Clearly, as b-> @, p(s) ~> 28/7. Thus, proceeding without regard to rigor, (1.8) yields as b—> © [Cvora=2 [7 pores aa) This is, of course, the Plancherel equation for the Fourier sine transform for the restricted class of functions under consideration here, It is easy to give a rigorous proof of (1.4). Suppose that f(0) = 0, and that f has a continuous first derivative on [0,c]. Then from (1.1) it follows by integrating by parts that 1 f[* M wosifvola=-% @>o where M representa the integral. Thus for ¢ = 1, Mt lo)? se (1.6) Since g is continuous, it is the case that for any fixed large u li f “ joleyl* desta) = 2 f “lator 1 Him J, lo(e)l? deste) = = J, late)l? de (1.8) From (1.5) and integrating by parts, it follows that “ . am [woot dow sar [eran Bf [ora BE an 2 a Similarly ° = f Ig(o)2de = ME (18) 6 B Thus using (1.6), (1.7), and (1.8) and letting «—> ©, formula (1.4) is validated for f restricted as indicated. Such f are dense in the space #(0, 0), and, using standard theorems from Lebesgue integration, (1.4) can be proved for any fe %*(0,0). Hence the analogue of the com-224 ORDINARY DIFFERENTIAL EQUATIONS (Cua. 9 pleteness theorem of Chap. 7, Sec. 4, is valid in the ease of the simple example just considered, It might be of interest to the reader to parallel the above argument for the problem —2’=lr 20) <0 and sce how the cosine transform theorem is associated with the same differential operator but with a different boundary condition. The method of the above exampie, with necessary major elaborations, provides a means of treating the general second-order problem in the singular case. Throughout the remainder of this chapter L will denote the formally self-adjoint differentinl operator defined by Lr = —(ps')' + qx where it is assumed that p, p’, ¢ are veal wnd continuous, p(é) > 0, on any Teal ¢ interval under consideration. { Of fundamental importance in all that follows is Green's formula which states that if [41,5] is any interval over which L is defined and f and g are any tivo functions fox which L/,Lg make sense, then [E ts ~ 515) at = Yolen - olen (1 where Lal = PESO — FEA) In particular, if f and g are solutions of the same equation, Lz = Ir. where / is a complex number, then Green's formula applied to f and shows that [f9](!) is 8 constant independent of ¢, and hence can be denoted by just [/9). ‘The case of the semi-infinite interval 0 S ¢ < © will be dealt with first. followed by a treatment of the case of the interval -~2 TOT 28) Since the coefficient of mt in (2.7) is [y](b), it follows that the interior of Cy in the m plane is given by [xx](0) ward) <° 2.) By Green’s formula {(1.9)], WHI) = 2egt fP Wwlrae and xxi) = 2890 f belt dt + fxd 0) Since [xx](0) = —2i9m, (2.9) becomes f xl? ae < oF (Ql #0) (2.10) which determines the interior of C,. Points m are on C; if and only if mt > 8 [vva- (Qt = 0) (2.11)228 ORDINARY DIFFERENTIAL EQUATIONS [Cnav. 9 The radius ry in (2.8) is given for $i > 0 by me = 231 f Ivlt dt (2.12) Now let-0 0), asb— & the circles C, converge either toa cirele C,, or to a point m,. If the Cy converge to a cirele, then its radius r,, = lim ry is positive, and from (2.12) this implies ¢< @(0, ©), If o,, is ony point on C,, then *,, is inside any C, for b > 0. Henee Sh. “St and letting B—+ «© one sees that g + me 22(0,0). The same argu- ment holds if of, reduces to the point m,. Therefore, if $l = 0, there always sasolution of Le = Lr of class 7(0,), In the case Cy -> C,. all solutions are of class 2*(0, 0) for $l > 0, since both p and y + Aly are, and this identifies the limit-cirele case with existenco of the circle C,,. Correspondingly, the limit-point case is identified with the existence of the point m,. In the case Cy — m,, there results lim 7 = 0, and from (2.12) this implies that ¥ is not of class £(0,%). Therefore in this situation there is only one linearly independent solution of class &°(0, @ ) for $l #0. In the limit-cirele case m is on Cy if and only if (2.11) holds. Sinee x = ol) + mp), it follows that m is on C, if and only if St [Ixia = 3m a le + ay ds <8 Since [xx](0) = —2i%m, it follows from the formula above [(2.10)] that m is on the limit circle if and only if [xx](%) = 0. The following theorem has been proved. Theorem 2.2. Jf $i #0 and ¢, ¥ are the linearly independent solutions of La = lx satisfying (2.2), then the solution x = y + my satisfies the real boundary condition (2.4) if and only if m lies on a circle Cy in the complex plane whose equation ts [xx}@) = 0 Asb— & either Cy C,, @ limit circle, or C, > m,, a limit point. Ail solutions of Lx = lz are 9%(0,0) in the former case, and if $l = O exactly one linearly independent solution is 2(0,0) in the latter case. Moreover,Sec. 2) SINGULAR SELF-ADJOINT SECOND-ORDER PROBLEMS 229 in the limit-cirele casc, a point is on the limit cirele C_() if and only if fxal(=) = 0. Tn the limit-point case, if m is any point on Cy, then m— m,, the limit point, and this holds independent of the choice of 8 in the boundary condition (2.4). In particular, this will hold when 8 = 0, and thus the limit point is given by eb) ma(l) = ~ lim en ‘The Green’s function @ associated with the boundary-value problem (2.18) Lr = tx sin a (0) — cos a p(Q)a’(0) = 0 cos 8 x(b) -+ sin B p(O)r'(b) = 0 is clearly given by Gin) = HED Lele) + mLb,8)¥(7,)) | (87) " Wad leED + mb awed} @ > +) It follows directly from ‘Theorem 2.2 that in the limit-point case Green’s funetion tends toa unique limit as b> & given by the same formula but with m replaced by m,. In the limit-cirele case there are an infinite number of limit functions to which Green’s function may tend, depending on how @ varies as 6 inereases. In any case, the limit function to which G tends is of class (0, %) as a function of L, Theorem 2.3. Jn the limit-point case the limit point ms is an anulytic function of t for $l > 0 (and $1 <0). Sm, > Ofer $l > 0 and if m,, has zeros or poles on the real axis they are all simple, Proof, From (2.8) it follows that the center and radius of the circle Cy are continuous functions of I for $1 > 0. Thus, since C; is interior to C, for b > 1, it follows that if | is restricted to a closed bounded subset A of 3t > 0, then the points m = m(Z,b,8) on Cy are uniformly bounded as b-» «, ‘The funetions m, where m(!,8) = m(J,b,8), being meromorphic and bounded on A, are analytic there, Hence, by Cauchy’s theorem, the functions 2 constitute an equicontinuous set on A, and m, converges uniformly to m,. Being the uniform limit of analytic functions, m, itself is analytic on A, and hence on $l > 0. Since m, is inside of C,, it follows from (2.10) that Jm,, > Ofor 31 > 0. This proves that if m,, has zeros or poles on the real axis they are simple and that the poles have negative residue. ‘hese remarks apply also, of course, to the meromorphic function of 4, ma. It is important to know whether a given operator L is in the limit-point or limit-circle case. A useful sufficient condition for L to be in the limit- point case is the following: Theorem 2.4, Let M be a positive differentiable function, and ky and ky230 ORDINARY DIFFERENTIAL EQUATIONS [Cuar. 9 de positive constants such that for large t a2 ho — f @M a= POM Osi s ke (2.14) Then L is in the limit-point case at infinity. Proof. Xt will be shown that La = 0 does not have two linearly inde- pendent solutions of class 2(0,«). Suppose x is a real solution of Lz =, and assume xe £7(0,%). From (px') = ¢x follows for some e>0 : a . (px')'x dys —k p PAP tm fap de kf rat Integrating by parts and using the fact that x ¢ &(0,©) there exists a constant ks such that =prxy fete x’)? , a ~ fx XM yc hy (2.18) < M Af? Let tay Ty = {| Bix? H(t) f at Then using (2.14) and then the Schwarz inequality px’xM! afl f pbx ay

©, then (2.16) implies that for all large ¢, px'x/.Mf > H/2. ‘This means x and x’ have the same sign for all large t, which con- tradicts xe 2(0,). Thus #7 remains finite so that * hx)? 2 f M di <0 (2.17) Now suppose ¢ and ¥ are two linearly independent solutions of Lz = 0 which are of class 27(0,), that is, suppose / is in the limit-cirele case. Tt can bo assumed that these solutions are real und [ed] = pled’ — ve) = 1 This implies py’ ple’ 1 ean — Yt ~ GapBec. 3] SINGULAR SELF-ADJOINT SECOND-GRDER PROBLEMS 231 By (2.17) and the Schwarz inequality, the left side of the above equation is integrable over (c,~). By hypothesis (2.14), the right side is not. ‘Thus the limit-cirele case is ruled out. In the ease W(Q) = 1 forO St < « the following corollary results. Corollary 1. Jf g(t) 2 —k, where k ts a positive constant, and. f pdt = » then L is in the limit-point case at infinity. Many second-order differential operators of practical interest have pt) = | for OS 0 < & (in fact, o simple transformation can always effect this), and in this situation Theorem 2.4 implics the following simple criterion, Corollary 2. ff nit) = 1 for U Si < = and qs positive constant k, then L is in the limit-point case at infinity. AE for some 3. The Completeness and Expansion Theorems in the Limit-point Case at Infinity As a necessary preliminary, the results for the finite interval 0 St S b<# will be given a slightly different formulation. Consider the problem Le = -—(pr') + gr = Ie sin a r(0) — cos a p(0)27(0) = 0 (3.1) eos 6 (hb) -b sin B p(l)a'(b) = where 0 S a, 8 0. Then the completeness theorem (3.2) applied to the solution xs =e + my of La = lz yields f bweorar = 2 tral? Green’s formula applied to x, and y results in ff covers al (3.13) (= on) fxs Aan) aE = Leda H) — Luar] (0) where Yon() = ¥(LAon)- since both xe and Yon (xwenl(0) = 1. Therete ‘ 7 "a f IxelQ)[* de = I wes taking into account the definition of the monotone function p,. “The fact that m, is on C, yields, by (2.11), fr vascor a = Sea From these equations follows the important equality C pa = SH a> 0) @.14) Let? = 7 in (3.14). Since C, isin C; forb > 1 there exists a constant simple calculation shows that [xin (b) = 0 fy the ne boundary condition at b, and e (3.18) impliesSec. 3] SINGULAR SELF-ADJOINT SECOND-ORDER FROKLEMS 235 & such that $m,(7) is less than k for b > 1. Thus (3.14) yields * de) 2, * [mcs (3.18) or, for » > 0, f dps(d) < k(L + 9) This, together with p(0) = 0, gives lo) 0, Jilo@r dost) ur? foQ0P der) sf anora (3.17) where A = (—@,%) — (—x,u). Applying the completeness relationship to f itself fo vere = (ff, + f) aarr day Letting § tend to infinity through the subsequence found above, it follaws, using (3.17) and the integration theorem, that [f° yeorrae - f" tooreaoay | sar? fener ae236 ORDINARY DIFFERENTIAL EQUATIONS {Cuar. 9 Now allowing » —> , there results the Parseval equality fo ora = {7 toot deo) @.18) for any f restricted as above. Rather standard arguments now suffice to show that the Parseyal equality holds for any fe 27(0,), First suppose fe 2*(0, ©) and vanishes for large t > 0. Then there exists a sequence of functions f, ¢ 27(0,°) possessing continuous second derivatives and vanishing near ¢ = 0 and for all large ¢ such that lim "ifs — flat = 0 @.19) and from (3.18) applied to f, — fa i Vn — Seal dt = fr. lan) — ga(d)|? dor) (8.20) where a0) = [0° WHA) at (21) Since the left side of (3.20) tends to zero as n, m—> %, it follows that the sequence {g,} converges in the mean in &*{p), and since the latter space is complete there exists a ge ?(p) which is the limit in the mean of this sequence. From (3.21) it is clear that g is the continuous function given by ota) = fE stove) at Using (3.18) again fo ora I fim f° Yatolt at = tim f°, aol apo {2 wor? daar which proves the Parseval relation for any fe 29(0,%) vanishing for all large £ > 0. Suppose f is any function of class {*(0,) and define fl) =f) Ostsa) =0 (a <’) and aad) =f,” folO¥CEA) at = [FOYER at Since [7 lucta) = ge00i? doy = [Ma a © in [Coleco dee =f yore there now follows the Parseval equality (3.7) for any fe (0, «). The proof of the expansion theorem (3.12) [in the sense of (3.8)]} will now be given, Jet 4 = (u,A] and fa) = foavar dp(d) (3.22) where g is the function appearing in (3.6). ‘The relation (3.7) implies that, if fife (0, ) and gi,g2 are the corresponding transforms, thent fC iheat = gn doo) (3.23) Let Pe @(0,<) and vanish for ¢ > @ > 0, and let the transform of P be Q. Then from (3.22) follows, on multiplying by P and integrating, ff rscoPaa = [° (foun dp) PO at = foo) (2 Peovtea) at) doa) = f,0)G0) a From (3.23) for fy = f and fa = P, frp a= f" 00)80) deo) Subtracting the above and letting 4¢ = (~«,«0) ~ A [soars f, o@doreo) and using the Schwarz inequality [fg soa hs f oorrdeay [jean doa) = floeitdoa) [rar Applying this inequality to the function 7? given by: PO =fO-IsQ Ot sa) =0 @ 0, there exists a constant & such that * do) 5 [ees for b> 1 and all » 20. Letting b— & through the subsequence chosen below (3.15), it follows that the above is true with pin plice of ps. Since it holds for all y, * dela) [ efi * * From (3.15), for x > 1, there exists a constant & such that “ dpsld) 2 & <* [ w u for b > i and similarly over (—©,-2). Ff Sl #0, $lo% 0 and f. (ee - a dpa() (3.28) is considered over (— «,~), (—,x) and (4, %), it follows that, if b+ © through a chosen subsequence and if then »—» ©, (3.25) must tend to i (r = ~ row) de(d) But (3.25) is just Yun) _ Salle) St Sly which tends, as b — ©, to Sn _ In, (to) 3b Sh Therefore ! “4 ee = ia eee te (3.26) where ¢ is a constant independent of t, provided 91 > 0 (or 9! < 0).Sec. 3] SINGULAR SELF-ADJOINT SECOND-ORDER PROBLEMS 239 From Theorem 2.3 Qm.()/3/> 0. From (3.26), letting 9 = 0, Sl ©, it readily follows that c 3 0.f From (3.26) results gino — men) = 3(f" (Ay - 7) do(n)) + e8(t ~ by (3.27) Since m,, is analytic in / for 31 > 0 (or $f < 0), Sm, determines Rm, to within an additive constant. The imaginary part of the analytic function of Z (for fixed lo, ly # 0) defined by the right side of (3.10) is $0n,(2) ~ my(lo)) by virtue of (3.27). Thus, this must be m,() ~ m,,(lo), which proves (3.10). Let 4, « be points of continuity for p. Then from (3.26) lim P gma(e + indy = Jim Lf dole) adv to oF = lim fr i (: = *) - tun-? (* =. 2) | aote bo J = = m(p(d) — plu)) yielding (3.9), and thus completing the proof of Theorem 3.1, In proving Theorem 3.2 the following will be required. Lemma 3.1. Let g ¢ (0) and £0 = foQ0PGM do) where Ais a finite d-interval, Then, as A-> (~ ©,0), fa ts convergent in 20, ), and thus tends in the mean fo a function fe ©(0, 2). Proof. Let A, A; and let Pe 2*(0,%) be a funetion vanishing for large. Then if Q is the transform of P, JU Ua — Fak at = fi) 900) dao Using the Schwarz inequality much as shove (3.21), and letting P= fs— fa, forO0 $¢ sa, and P = 0 for? > a, there follows QP fu sb [900 a9) Since the right side does not depend on a, this inequality holds with Letting 4 x), the proof is completed. At this : ntness of Z in the limit-point case tan be proved, ‘This is done in Prob. 13. # As has already been remarked by appraising ps, it ean be shown that, in fact, ¢ = 0.240 ORDINARY DIFFERENTIAL EQUATIONS [Crar. 9 Proof of Theorem 3.2. In view of Lemma 3.1, it remains to show that g comes from f by (3.1), that is, at) =f" oven) ae (B.1) where the equality is meant in the mean in £%(o). From Theorem 3.1 there exists ag e £7(p) such that (3.11) holds withg replaced by g. There- fore the problem reduces to showing {lo - GO)? doe) = 0 Letr = g — §; clearly ne 2%). Using the fact that fis the limit in the mean of fy, where HO = f,Gleee) dole) A = (nad it follows that the function hs = fs — fa, ha = f raven) darn tends in the mean to zero, that is, a, f\ wsora =0 (3.28) It will be shown that ha is the zero funetion. Let [be a complex number with $f > 0, and put H(t) = f. 5 wer aor (3.29) Then, since Ly = dy, LH = Ufs + ha (3.30) From (3.29) follows easily that //, satisfies the same boundary condition at zero as dos y, namely, sin a 175(0,1) — cos « p(0) Hi) = 0 By the variation-of-constants formula, (3.30) yields AsG) = f le Dr — eG )EGD har) dr + cat.) (3.31) where cy is a constant (which may depend on 4). From the fact that. reW@(o), r/& — Delp) and hence Ha, a8 a function of é for fixed 1, $l > 0, converges in the mean in €*(0,) to a function H as A> (~~,e). Using this in (3.31), it follows from (3.28) that there exists aSec. 3] SINGULAL SELF-ADJOINT BECOND-ORDER PROBLEMS 241 constant cso that AD) = eb(t.D Since ¢ is not of class 7(0, <) for $2 > 0, the constant ¢ must be zero, and hence (1,1) = 0 for 31 > 0. Thus H, tends to the zero function in #70, ©) for $1 > 0. Let TQ) = f’ wea) at that is, P, is the transform of the function which is one for0 S$ ¢ 3 sand aerofort > s. ThusT,¢ @%(e). Integrating (3.29) with respect to ¢ and using 174 —+ 0 in £2(0,@) there results L : (140) data) = 0 (3.32) Sinee T, ¢ £7(p) the Schwarz inequality shows that the integral on the left of (8.32) is absolutely convergent. Indeed [Coroor.ayl dea < © (3.33) Formula (3.32) can he inverted much as in the proof of (3.9) for p. For this purpose it will be assumed that g, and hence r, is real. ‘This is no restriction since every g is a sum gi + tg. where g; and gz are real, Taking the imaginary part of (3.32) and integrating, one obtains Ape Jim J operror@ do(o) dv = [ . r(a)T(o) lin, [ton (: = 2) — tan (# = ‘)| dp(s) =o [ * He) e) dele) = 0 (3.34) Since for A = (u,dJ, [rere dove) = J, rte) fi otter atdpte) = f' (f.r@vtee) dpte)) ae it follows that the function *& given by ks) =f, r(e)¥fo) dove) has a continuous derivative, and from (3.34) this must vanish, Thus ha) = J r0)U(6.) dol) = 0 (3.35)242 ORDINARY DIFFERENTIAL EQUATIONS, [Caar. 9 Since ¥(0,4) = cos a, it follows from (3.35) with ¢ = 0 that if cos a #0 fro) am =4 (3.36) If cos a = 0, thon sin a ¥ 0 and, by differentiating (3.35) with respect to t first and then setting é = 0, (3.36) follows. ‘Thus (3.36) is always valid. Because of the arbitrary nature of A, (3.36) implies that for any a > 0 [2,100r0) doQ) = 0 (3.87) for any step function 7. The step functions are dense in 2*{p). and sinee reQ*(p) it follows that y can be chosen so that the left side of (3.37) is arbitrarily close to fo lrO)#dao which must therefore vanish. Since this holds for all a, the theorem is proved. 4, The Limit-circle Case at Infinity If £ is in the limit-circle case at infinity, the cireles C,(2) converge to G,,() as b > & for each 1, $f #0. Ench cirele C,(0) is traced by points m = m(1,b,8) as 8 ranges over 0 SB 0 for $2 x 0, the poles and zeros of these functions can lie on the real axis only and are simple. Let g; denote the step function ps associated with the condition 8; at b; ‘Then the following theorem will be proved, Theorem 4.1. Let (le) be a point on C. (lo) and (b;,8)) a sequence such that m(lo,bj,8;) = m,(Io) tends lo sio(lo) asj—«. Then for alll jim m(l) = kD (4.1) and'|in the sense of (3.5)] Jim pjQ) = BQ) 4.2) where wh, is a meramorphic function of 1, real for real 1, and with poles and zeras that are real and simple. Moreover, § is a step function discontinuous at the poles, | = yk = 1,2, ... , af ht, only and with a jump at dy equal lo minus the residue of teats. The functions fx, where Plt) = Y(t),Sec. 4] SINGULAR SELF-ADJOINT SECOND-ORDER PROBLEMS 243 form a complete orthogonal family in @(0,%). If Re ts the function defined by Rall) = eltle) + the(lo\W(tylo), then fet.]() = 0 (4.3) for all k. On the other hand, for yn, where Wilt) = ¥(t,)), WR) #0 Gu, k=1,2,...) The condition (4.3) is actually a boundary condition satisfied by the Yaatt = «, With each point on the limit circle C,, (lo) is associated such a boundary condition. In the course of proving Theorem 4.1, two other theorems will be proved which will make clearer the nature of this boundary condition. Proof of Theorem 4.1. Let xi) = ot2) + mOPED (4.4) Apply Green’s formula to x;(,!) and %(4,o). Since both functions satisfy the same condition at 6;, it follows that. mm8) — myfle) = = 1) PY xtra at (45) Using (4.4) in (4.5), __ mill) + Ut) f eltddraltto at : 4.6) 1=C~ 1) f' vGdxttle) ‘ m,(D) In the limit-circle case all solutions of Lx = Iz are 27(0,0). Therefore, as j— , the eniire function of { whose value at 2 is given by LF eterctatey at = PP otentetebe) + my(leritede)] de which appears in the numerator of (4.6), tends to the limit Jo" eC DEeG ta) + Malad e)] dt @n) If lis restricted to some finite part A of the J plane, then it was shown in the course of the proof of Theorem 2.1 that ¢ and ¥ have norms in 27(0, « ) which are uniformly bounded in A. Thus, by the Sehwarz inequality, the integrals in (4.7) are uniformly convergent in J over any finite part of the Z plane. This implies that (4.7) defines an entire function of i. The same is true for the integral in the denominator of (4.6). Thus, as je, the meromorphic function m, tends to a limit *, which is also a meromorphic function, and this proves (4.1). The properties of mh,244 ORDINARY DIFFERENTIAL EQUATIONS {Crar. 9 follow from Sh()/3t > 0 for $l 0. It follows from (4.6) that He) + C= Ie)" oD Rte) at the) = = 1= = be) f° ve Dect) at (4.8) As in the proof of Theorem 3.1, the Helly selection theorem shows that a subsequence of {p;} exists which converges to a limit é, Sinee (3.14) is valid, the argument of Theorem 3.1 then shows that (3.9) is valid with p replaced by 4 and m, by *,. This proves that 4 is independent of the choice of sequence and thus that (4.2) is valid. Since 2. is real on the real axis, (3.9) also proves that f is a step function discontinuous only at the A, and with a jump at \, equal Lo minus the residue of the pole of 1, there. The completeness of the set {¥x] follows from (3.6), (3.7), and (3.8) with p replaced by 4. The orthogonality will be proved after Theorems 4.2 and 4.3. From (4.8) it follows that at any pole dy of si, the denominator muat vanish, that is, Oe — 0) fi GIR) dt = 1 (4.9) By Green’s formula it is easily seen, since ¥(t,Ax) is real, that. Ore — He) PP oe rnrtelble) dl = 1 — [eeval(b) Using (4.9) and letting b + ©, it follows that (4.3) is valid. Let denote the set of all functions u such that (i) wis differentiable and x’ is absolutely continuous on 0 S f S 6 for allb < %, (i) wand Lue %(0,~), Gii) sin & u(0) — cos & p(0)u'(0) = 0, fiv) [zg.I(e) = 0 Let Wl) Ra(r4e) (ES 7) Hert) = | yeoeattn) — > 2) (4.10) and for any f ¢ 240, ©) let Sof) = i Gr bye) dr The integral is absolutely convergent since f and 2, are both of class HO, o). Theorem 4.2, For any fe8(0,~) the function u = G(bo)fe® andSec. 4] SINGULAR SELF-ADJOINT SECOND-ORDER PROBLEMS 246 (L~ tu =f. Conversely, if ue, then f = (L —~ i)ue %(0,«) and w= Goft Proof. It has been shown in Theorem 2.2 that [ReRa](e) = 0 (4.11) Let y,¥ denote the functions given by g(t) = (tte), v(t) = (tl). From (2.8) it follows easily that, if #,, is the center of the circle C(I), then levl(@) + fiLI¥](@) = 6 (4.12) and the reciprocal of the radius of C.(i) is Ivl(e}{ > 0 Clearly, since 2, = ¢ + m,¥, (4.12) yields WRal(@) = (The — Fig)vyl(o) so that WRal(e) 0 (4.18) ‘The proof of the first half of the theorem follows from the use of (4.10). Indeed, if u = §(lo)f, then WO = td) fi vets) de + Wile) f° RalrdadfCe) dr From this it follows that (i) is satisfied, and : ° 5 wren = x5(td f vetayior de + 6") |” aaletaytey ar + RalLEO Since [¥2.]() = (¥%,](0) = —1, one sees easily that Lu = lu +f. ‘That ve 2°(0, ~) follows from the fact that ¥ and %, are 27(0, «) and the use of the Schwara and Minkowski inequalities. Since Lu = lu + f, Lue (0,2), and (ii) holds. Condition (iii) is valid since ¥ satisfies {iii), and (iv) follows with the aid of (4.11), The second half of the theorem will now be proved. Let f = Lu — lou. Then f £(0,) and, from the first part of the theorem, §(lo)f is of clasa & Thus w = u— &(lo)f is of class and satisfies Lw ~ lw = 0. Hence w = cy + ¢2%,, for some constants cy, cs. Used in (iii) and (iv), it follows that cz: and 1, respectively, must be zero because %,, cannot satisfy (iii) and, by (4.23), ¥ cannot satisfy (iv), This completes the proof, This result can now be used to prove the following theorem. t The statement of Theorem 4.2 is just the statement that $(le) is the inverse of the operator L ~ ts with domain 4,246 ORDINARY DIFFERENTIAL EQUATIONS [Cear. 9 Theorem 4.3. The boundary-value problem Le = iz sin a 2(0) ~ cos a p(0)z'(0) = 0 [z2.J(-) = 0 (4.14) ia self-adjoint; that te, for all u and v of class D fo Guea = f° waa ae (4.18) Proof. From Green's formula and the fact that [uv](0) = 0, (4.15) is equivalent to fus](~) = 0 (4.18) From Theorem 4.2 there exist fg «(0,) such that u = §(/))f and o = §(b)g. Expressing [uv](b) in terms of the integrals involving the Green’s function (4.10) and f and g and letting b -» ©, (4.16) follows from (4.11), and the theorem is proved. Tt is a consequence of (4.3) that Ys is of class $. From (4.15) the orthogonality of the yy is immediate. No yr, ! # A, for all k, can satisfy [z2.](@) = 0, for if it did y, would be of class and thus would be orthogonal to all ¥. This is impossible since the yz are complete. &. Singular Behavior at Both Ends of an Interval The cases where the coefficients in L have singular behavior at both ends of an interval, or singular behavior at one end and a semi-infinite interval, or an interval extending over the whole ¢ axis, are all handled similarly. Here the case where L is defined over ~ 2 <¢< © will be treated. Recall that Le = ~(pa')’ + gx where now it is assumed that p, p', ¢ are real and continuous, p(t) > 0, on—% TimiPimk where the sum is taken over all m such that Ain = Aan. Let pals +0) = ps(A) and ;(0) be the zero matrix. Clearly ps possesses the properties: (i) pais Hermitian —(oajz = Baus) Gi) pafA) = pal) — ps(u) is positive semidefinite if \ > u (A = (#,A)) Gii) The total variation of p) is finite on every finite ) interval. Any matrix ps satisfying (ii) is said to be nondecreasing. The matrix py is the counterpart on 8 of the nondecrensing spectral function p, for the problem (3.1). Applying the Parseval equality (5.5) to any continuous function f on ~o << © which vanishes outside some interval 6, contained properly in 6, one obtains 2 fC. wor a= [7 Y 20a domo (5.8) sh248 ORDINARY DIFFERENTIAL EQUATIONS, [Cuap. 9 where wr) =f" sWeslur) at As > (— 0,0) (that is, a+ —«,b— «), it can be shown that, if L is in the limit-point ease at — © and ©, there exists a matrix » having the properties (i) through (iii) such that ps— p, and (5.6) is valid for any fel(—o,a). IfLisin the limit-circle case at one, or both, of the points — or ~, limiting matrices p still exist such that (5.6) holds, but there is the usual nonuniqueness. The key to proving the existence of a limiting matrix p is an equality for ps which replaces the equality (3.14) for ». Let x. = gi + mayabea solution of La = lx (Jl > 0) satisiying the boundary condition cos a x(a) + sin a p(a)z’(e) = 0 and similarly Jet x» = ¢: + mg: be # solution of the same equation satisfying cos 8 x(b) + sin 6 p(b)z"(b) = 0 Then, as has been shown in Theorem 2.2, m. and m; lie on circles Ca and Cy in the complex m plane whose equations are, respectively, [xaxe}(a) = 0 — [xexe](6) = 0 (5.7) It is easily seen that Green’s function G, for the problem (5.1) exists, provided St x 0, and is given by ells sle) ne — mii $7) xalr,l)xo(t,2) ing) — ms Galbr)) = The completeness relationship in the form (5.3) is now applied to the functions ny = 2% cop, jo = 3 aa 40) G,k=0,1) yielding IG, 28 40, 2 on a >) [3% Gs 0 ay han(t) at [% FG GO, fan(C) dt (6.8)Sec. 5) SINGULAR SELF-ADIOINT 8ECOND-ORDER PRODLEMS: 249 From the definition of G, it follows that oto.) = EN es 0) _ xh) ma) — my | «7 (6.9) and aa, = Mall) xelt,t) a (40) = SiG) imal — ma) @s0) a Meal) “pO im) — mj &>% (6.10) Using (5.9) and (5.10) and Green's formula, the integrals in (5.8) can be evaluated. For example, mEBt J, J@a(e0,))* dt = 25g0Lmatl) — mad { f° lxeeI* at + {P botne at} = [m.) — ma (QI {[xaxa}(O) — fxoxo](0)} = WFm(l) — ma(!)) | mall) — ma? making use of (5.7). Therefore f [Gx(t,0,0 [2 at = Seu = mar (6.11) : Similarly, (ma) — mol) = dan) f GulC.0.D Fal dt = [xoltan}(6) — [xafan}(0) + [xchan}(0) — [xahanl(a) = [xa — xz)han) (0) = (m1) — m(1))[ echo] (0) = (me(t) — malt) Fin and hence f GNC O,D and dt = Fee, (6.12) Here use has been made of the fact that [xahs.J(b) = 0 which follows since both x, and Ay, satisfy the same boundary condition at }; similarly [xehan}(a) = 0. Now (5.8), (5.11), and (5.12) yield for 7 = k = 0, * donQ) _ $Man() --h—-P st where Mur) = (m.() - m0) (5.13)250 ORDINARY DIFFERENTIAL EQUATIONS {Cuar. 9 Further similar calculation shows that iz dem SMa (5.14) where My. is given by (5.13), and MinQ) = Man) = mal) + ma()) (mal) — ma)? Min(l) = ma) m(D) (mel) — wa)“ Formula (5.14) replaces (3.14) for pp. Since BP boda = —Im BEL aD Pat = Smal) for a fixed 1, 32 + 0, m,(l) and m(J) are in opposite half planes. Suppose l= ¢in (5.14), Then points m,(z) lie on a cirele C, which is in C_, for a < ~—1, whereas points m,(7) lio on Cy which is in C, forb > 1. Thus there is a constant ¢ > 0 such that |m.(i) — m(2)| > c¢ for a <—1, b> 1, Since m,(i)and m,(7) are uniformly bounded fora < —1,6 > 1, it follows from (5.14), and the definition of the Mfa.(/), that “eMex Gann for some constant K. Since 2\renFanel S lransl® + [ranal? [Sa ©. The matrix p = (oj) possesses the same properties (i) through (iii) as ps. If Z is in the limit-point case at — © and =, p is unique since in this situation both m, and m, tend to points m_., nto, respectively, and as in the proof of Theorem 3.1 the formula m pas) ~ oats) = tim, f” gatate + t0 drSec. 5] SINGULAR SELF-ADJOINT SECOND-ORDER PROBLEMS 251 can be proved. Here My(!) is the limit (which exists) of Min() as 6 (—«, a), There is also a corresponding expansion and completeness theorem, the proof of which parallels that of Theorem 3.1, and will be omitted. Theorem 6.1. Let L be in the limit-point case ai ~~ and ». There exisis a nondecreasing Hermitian matriz p = (pj) whose elements are of bounded variation on every finite interval, and which ia essentially unique in the sense that pre(d) — pans) > pnd) — pa(n) (6 (— ©, @)) at points of continuity d,u of pj. Further > ands) ~ ale) = tim? [" gataly + 9 de 6.18) where Mu = (mQ — m)~* Miu) = Mall) = 4(m_.@ + ma())(m2() — m())-" (6.16) Mall) = mm.) (m-. — me)" Analogous results hold if Z is in the limit-circle case at either or both of the end points of the interval. In order to obtain a unique spectral matrix, boundary conditions must be added at the end point where L is in the limit-circle case, as was done in Sec. 4. If both ends are in the limit-circle case, then m, and m_,, are meromorphic and thus so are Mu, Mas, and Mus. For any limiting matrix p let £7{p) denote the set of all vectors g with components gi,g2 functions of \ such that 2 lol? =f", ¥ ada0) dad) < © Jeet ‘The nondcereasing nature of p guarantees that the integral above is nonnegative. Theorem 5.2. Let pbe any limit matriz of the act {m}. If fe@(— 0,0) the vector g = (g:,g3), where . aid) = J" edt) at sonverges in 22(p), thal ia, there exists a g ¢ @{p) such that lla - geal] 70 (C+ —=,d- ») where gen) = [00nd (-@ 0 the solution e~iv# ¢ 24(—- © 0) and hence vit) = cos Vit ailt, = cos It + m_.(t) Si vit c(cos V/1t — isin V/1 1) vi for some constant c. Putting ! = 0, there rosulls ¢ = 1, and therefore mal) = -ivt Similarly mi) =ivt t Every eigenfunction is of class 7(~ =,©), See Probs. 6 and 7,Src. 5] SINGULAR BELF-ADJOINT SECOND-ORDER PROBLEMS 253 Thus Bud = 2 aot = EY ano = tao = 0 and consequently, from (5.15), dour = 2 a>0) =0 <0) dpn(A) = x Vid (A> 0) =0 (A <0) dpu(d) = don(a) = 0 The opanon theorem becomes for any f¢ 2?7(— 0,2) = xf, hte 10) dd + f sin WX gata) ed where ata) = fp f() cos Xt dt . sin Vad 92(A) f oo hr dt and the interpretation on the integrals is the same as in ‘Theorem 5.2, This is precisely the Fourier integral formula for functions f 2#(— =, ©). (The spectrum of the problem Le = —2" = Iron —© I fort >t, Usevs’ = (¢ — dv and show ya and y, can have at most one zera for t > fs. (0) Show that there exists a monotone increasing sequence Ao, Ay, . . « With De @ ag n—» © such that y has exactly n zeros on @ <¢ << © for kx: <4 S Au, Where Anam, ‘Hin: By Prob. 1(2), Ya has a finite number of zeros on (0,2). By the proof of ‘Theorem 2,1, Chap. 8 ya and 4; are of the same sign ag \A—> ~ = for any given? > 0 and thus by Prob. 1(a), ya has no zeros for \ near ~ = fand |¥{t,d}| > © ast—+ @ 80 that L is in limit-point case]. By the same theorem, ya has zeros on (0, ») if A is large enough and the seres move continuously to the left as A inerenses and to the right a8 » decreases. Thus there exists a A, such that for \ < da, va has at most n zeros on (0,0) and for A > Xo at least n + 1 zeros on (0,0). If ¢ is small enough, then vase has exactly n + 1 zeros, for if it has n + j, then as e— +0 the lost j zeros must move toward {= ©. Thus, whon the (n + L)st zero is large enough, the method of Prob. 1(a) shows there are no further zeros and also that the rth zero is not to the right of (24). Thus Ys has at least m zeros. On the other hand, if vx has exetly Fe z0r0s, $0 does Yo, for small enough «. Thus #. has exactly 1: 2 (c) The spectral function p is a step function which can ‘be discontinuous only at Yop My eee : Consider the problem on (0,b) with z(b) = 0 and eigenvalues X.;. Then since yoy, where vail) = ¥(6s;), has exactly j +1 zeros on (0,6), it follows thatProps.] SINGULAR SELF-ADJOINT SECOND-ORDER PROBLEMS 255 dy > Ay. On the other hand, given any ¢ > Oand 7, Ma < An + efor b large enough since Passe has n +E zeros on (0,%). Use a:(A) — p(A) as b—+ oe. (@) Show that the yx are $(0, ©) and orthogonal. Hint: By Prob. 1(a), $e Oast—+ « sinee, if [ys] —» =, the number of zeros of ya on (0,«) cannot change with » small change of 4 from Xs. Since ye and yj,—+ © monotonically, show that f~ Wildl < = and therefore“ Wwaidt < = and thus ¥xe2%0,0), For orthogonality, use Grcen’s formula and ¥3¥; — viva > Ons f+ ©. (©) The spectral function is discontinuous at all tu, k 2 0. Hint: H p does not have a jump at ds, then since y, is orthogonal to all yy, j * n, the Parseval equality fails. 2. Lot p = 1 and q be real and continuous on {0, ) and Tim inf qi) = 4 ae for some », — 2

0, the spectral function is of class C1. Indeed, show that if T = a4, ots), where ¥(0,s) = 0, ¥'(0,s) = 1, satisfics A® gin (ot — ale) + 0 vits) = as {—+ ©, where A is continuous and positive, and @ is continuoua and real. Prove that ford > 0 # oy - ax xara)256 ORDINARY DIFFERENTIAL EQUATIONS (Cuar. 9 Hinr: For s > 0 show, by successive approximations, that wea) = ee — f haa a bounded solution and hence that i(fe) — ¢—» Oast—+ «. Bhow Lyi = s%1. (Note that os exists for Je 2 0.) Show similarly there isa psrelated toe, Neither vi nor es is 20,0), Show, by Prob. 1, Chap. 1, from sin a(¢ — 1) L= 2) girder) dr veh) = By fA ye apg) a that |¥| is bounded as ¢~+ © and from this that as t—> © Ao ¥0,2) — 4O ain at - a(s)) 0 hore Ages = +f” eimaterdte) dr and the integral converges uniformly. Since ¥ = cig: + crx, show by letting 1+ © that cr = Ae~//(2ie) and cy = Gy. Henee A(a) = 0 implies that ¥(1,8) = 6 for all t. Consider the problem Lz = iz, 2(0) = 2(6) = 0. Then the cigenvalucs occur at those values of s for which (bs) «0. For b large there occur (a: ~ #:)b/e + ¢ eigenvalues in (s1,82), where [el < 4 and ift Ata) 5h Vidi “ae te where e+ 0anb—+ co. Thus for large as(ee) — salon) = EE oe where 4 isin (a,8:). Asb— © ioe) ~ ato ~ Mee for some # in (81,82). Thus . a5 2s ae ©) = SG} where (3) = p(s*). 5. In Prob. 4 let s be complex and show that the function F defined by F(a) = A(ajetato isanalytic in sfer 3s > 0, Show that F can vanish for Js > 0 only if Rs that where F(e) = 0, y(a) = O(e-"") and hence an eigenvalue occurs. 6. In the case of » boundary-value problem on 0 $¢ < « which has a spectral function » show thet if {is in the point spectrum then , where 4(¢) = ¥(t,%), is of class 03(0, ©). Hine: Let f = $fort a. Let Show Provan at = 0) fire - fonePross.} SINGULAR SELF-ADJOINT BECOND-ORDER PROBLEMS 257 Let p(X +0) — p(k — 0) = %. Show 4 i fiwea 0. From cot (a1 — a) = m, (1) follows de 1 dn, a aM "TE mQyen) dd a (aon) 10. Let Lz = ~z" + g(f)ton0 $1 < =~, where gq is real, continuous, and periodic of period i. Prove that S’ ia the union of the closure of the intervals on ~ << A < ich 2 + ( — g)x = 0 is stable; that is, prove, in the notation of Sec. 3, Chap. 8, that 8’ consists of those points for which P(A) & 4. Hist: L is in the limit-point case [by Theorem 2.4 or from the fact that on a atable interval no solution is (0, <)), Thus for 3é = 0 one characteristic multiplier y() patisfics |y{I)] <1, Multipliers are roots of ¢* — of(l} + 1 = 0, where AM = e024) +490) in the notation of Chap. 8, and (i) js one of the roots (f + +/f*—4)/2. For the problem with r(0) = 0, » + mv ¢ (0, 0) and ia therefore n solution with multiplier258 ORDINARY DIFFERENTIAL EQUATIONS (Cuar. 9 y. Thus 9(1,!) + moO) = 7 and therefore mo) = Gy — 00,))/4(1,). On the interior of stable intervals f* < 4, and (1) tends to a nonreal number of magnitude las 31-0. Thus 9m, » 0 on stable intervals, On unstable intervala Im. 0 except at points where the entire function ¥(1,2) = 0. 11, Let E be ag above but now take the interval —« =, t= fi teldvend ~ vbdels:DIfe) de + ey where o( «) must exist becaure he—» h, fo—+f. Thus Lh = th +f, and h antisfies the condition at$ = 0. It remains to prove his unique. For $i » 0 this follows from the fact that y is not £10, =) for SI #0. For 3! = 0, let? =X Clearly J, where ¥(t) = (4,5), is an eigenfunction corresponding to some ps for some condition p at b aince ¢isreal. If ¥ ¢ £9(0, ©), the jump of p, at X does not tend to zero asb—» ©, and ‘Sie in the spectrum of s. Thus, if § is not in the spectrum, ¥ is not 2*(0, ©) and so A ig unique. ea ‘® denote the class of functions satisfying (i), (ii), and (iii) below (4.9). The boundary-value problem Lz = lz, sin 2 2(0) ~ cos a p(0)z'(0) = O, ia anid to be eelf- adjoint if for any uve D (typ) = t dud = tf u(Ee) dt = (u,Z0) Show that if L is in the limit-point case the problem is self-adjoint. Fine: Let Qi #0, Lu fu =f, and Lo~lbap. Then fipe8(,~). Let of," fediaod q be simitarly related to p. Let us =f" vader). Then, by Prob. 12, ua—> win 2(0,2). Let v. be similarly defined. [ira [OES [8S [orePross.] SINGULAR SELF-ADJOINT SECOND-ORDER PROBLEMS 259 Gnd» [mae fhe But (u,,p) is also given by the integral on-the right. Thus (f,us) = (ua,p). Letting a-+ ©, (Luv) = (u,Lv). 14. It is shown in Theorem 3.1 that p is unique in the sense that all p,—» pasb—> wo. Prove the following stronger result: If (3.6) and (3.8) hold with » replaced by some #, then 2 must be ¢ given by (3.9). Hint: Let fe 240, =<). Then the unique solution k of Prob. 12 can be represented two ways, so that for 9? > 0 [o224 ao - nov = Fun Letting a> © converges in 20,0) to Oana. Let rove) = ["vliaydt, Clearly re 2G) and @*(), being of the form (3.11). ‘Taking f F. dt and letting a—> © fo FREED a — nor = 0 ‘The same holds with I replaced by J, If f = ys + 4y, then O&= = a= foe rie ‘Thus ’ LC OPO) a — 90) 2 0 B= ee Integrating with respect to » from 1 to ps and letting »— 0, [f corre. ate - 90) =0 for all p, and 4s. Differentiate the above with respect to s. Then [arrvoarrade - 0 = 0 Lat 7 = Lfor 0 s. Then g(a), is Pj) and differen tiating with respect to r [f verre ae - m0) = 0 If « 2/2, Ist @ +s = 0 and use 4(0,) = cosa. Hf « = =/2, differentiate with respect tor and sand then lety = 9 = 0. In any case, [40 - me =0 proving the result. 15, In the limit-circle case, Theorem 4.1, show that E/A;7 < «, where the prime denotes omission of a vanishing eigenvalue, Hint: va = (4 — 1s) GGo)va. Thus, by the Bessel inequality appliod to G(,x,2s) tor a ff" ortara x Savage :260 ORDINARY DIFFERENTIAL EQUATIONS [Cuar. 9 for all N, where ry is the normalizing factor. Integrating in r, the result follows. 16, At sero, Le = —2" + [27 — $)/t]z, 0 0, and r, qi, and gs real and continuous for 0 t< a. Let wand v be vectors of class C!on 0 S¢< =. Let u-v denote wi6) + uyds Show that [Pac v — 6 Lo) dt = tutte) — ol) where [uo](t) = plid(usl0)61(0) — uil02(0). Let ¢ and y be solutions of Li = iz which satiaty pA(0)¥1(0,1) = gin a pA)$1(0,) = con a ph(O)pr(0,1) = —cos a PA(O)y2(0,!) = sin a (a) Develop the theory of Sec. 2 for Lz = iz, z,(0) cos a — 23(0) sin a = 0. {®) Develop the theory of Sec. 3 for the above problem in the limit-point case (define the latter notion). In particular, show that if the vector f is 2(0, @), that iy [77-440 < «, there existe a p euch that 00) = [70-vaar ae exists and st) = f° vtrrgo doom) (©) Develop the theory of See. 4 for this aystem. {d) Formulate the reults of Sec. § for this problem,CHAPTER 10 SINGULAR SELF-ADJOINT BOUNDARY-VALUE PROBLEMS FOR nTH-ORDER EQUATIONS 1. Introduction In this chapter the theory of Chap. 9 will be extended to the nth-order case. Here the formal differential operator L is defined by La = pot + pitD fos + pet Tt is assumed that the p, are complex-valued functions with n — k con- tinuous derivatives on an open interval « <¢ ransej(t dua) (2.3) yon where the ru; ate complex constants. Using (2.3) in (2.1), the Parseva) equation can be written as [ull = Le, Bid)oee(d) doa) e4 jek where auld) =f, esteayuco at (25) and the matrix ps = (ps) consists of step functions with discontinuities at the eigenvalues. The jumps at the eigenvalues are given by punts +01 — pune ~ 0) = rangfins (2.8) where the sum is taken over all m such that Am = Age (note that severalSze. 2] SINGULAR SELY-ADJOINT MTH-ORDER PROBLEMS 263 Xin can correspond to one eigenvalue As»). Clearly the matrix has the properties: (i) 9: is Hermitian Gi) ps(A} = ps(A) — pa(u) is positive semidefinite if A > u, A = (4,A] Gii) The total variation of pj is finite on every finite d interval. It is further assumed that p;(0) is the zero matrix and ps(d + 0) = ps(A). Because of (ii), ps is said to be nondecreasing. The matrix ps is called the spectral matriz. Note that it depends not only on the problem (1.2) but also on the choice of the independent solutions of La = lx. Tn this section it will be shown that the relation (2.4) is valid with 8 replaced everywhere by (4,5). The following theorem gives the existence of at least one limiting matrix p as 5 —> (a,b). Theorem 2.1, Let [8] be a set of intervals tending to (a,b) and {Usz = Q} a corresponding set of self-adjoint boundary conditions. Then {5| contains @ sequence {8;} lending to (a,b) as j—+ © such that (dh) = lim pa,(d) 22 fe exisis on ~0 <<. Moreover, the limit matrix p satisfies (i), (ii), and (iii) above. Proof. The proof is an immediate consequence of the Helly selection theorem and the following fact: given any 4 > 0 there exisisan M(y) < ©, not depending on 5 or Us, such that for |\| Su losa)| S AL CH) @.8) To prove (2.8) it is sufficient to take j = k, since from (2.6) 2" tao) sf" deus) + f% dru(a) The funetions y{*-” are continuous in (¢,\) snd at t = ¢ are equal to gn. Thus, given y, there is an h > 0 such that WMG) — aad < gs (2.0) fore $¢ Se+Aand [Al Su. Let f be a nonnegative function of class C* on (a,b) vanishing along with its first n — 1 derivatives, outside of (¢, ¢ + k) and normalized so that . [Pioas=i (2.10) The Bessel inequality applied to (—1)"—'f"- for some fixed m, (m = 1,4 ORDINARY DIFFERENTIAL EQUATIONS [Crap. 10 « _%), gives [goede ft y Gy(A)0(A) dpuel) (2.11) where here “ aatr) = (= 18 [" axeayr-o at = [OP aeranto at By (2.9) and (2.10), last) ~ baal < gy which, with (2.11), yields eth 1 f* 1 * “ [perp aed LO donee = Bf Y, down ‘ ~* “A Summing the above from m = 1 tom = n, the result follows, and hence Theorem 2.1. For any such limiting p, the space @°(p) is defined as the set of all vector functions g = (g,),j = 1, . . . ,m, which are measurable with respect to p and such that tol? = f°, ¥ aero) dewey < © hkwd Theorem 2.2. Let p be any limit matriz given by Theorem 2.1. If fe 2(a,b) there exists a vector ge L*(p) such that if air) = falas da — (8 C (a) (2.12) then lg -gll+0 (6 (a,b)) (2.13) In terms of this g, the Parseval equality ll = llell (2.14) and expansion FO = 7) Y exltarn0d doar (2.15) deer are valid, where the latter integral converges to f in the norm of (a,b). Proof. The proof follows tho same development as that of Theorem 3.1, Chap. 9, Eqs. (3.16) to (3.24).Bsc. 3] BINGULAR BRLF-ADJOINT MTH-ORDER PROBLEMS 265 Asin Chap. 9, by fotavod Ga1,...,9) will be meant the jth component of the vector g whose existence is proved in Theorem 2.2, The analogues of the remaining parts of Theorems 3.1 and 3.2 in Chap. 9 will be proved in Secs. 3 and 4 below. . 3. The Inverse-transform Theorem and the Uniqueness of the Spectral Matrix The inverse-transform theorem states that every vector g ¢ £"(p) arises from a function fe @*(¢,b), as in Theorem 2.2. This is true under certain additional assumptions on Z, which correspond in the case n = 2 of Chap. 9 to the situation where L is in the limit-point case at b, aa in Sec. 3, or in the limit-point case at both a and b, as in Sec. §. In principle, the proof is like that of Theorem 3.2, Chap. 9. As in that proof, the following lemma is required, Lemma 3.1. Let ge 2*(p) and n 5a) =f Y esltrrn don) Abo where A is a finite \ interval. Then, as A— (—©,0), fs converges in £2(a,b) to a function f « (a,b), Proof. The proof is the same as for Lemma 3.1, Chap, 9, Another fact used in the proof is the explicit representation of the inverse of the operator L -~- I for $l * 0. Lemma 3.2. Suppose for some I, 31 4 0, that La = Ix hae no nontrivial solution in @(a,b), If fe @*(a,b) and g ts any vector in 8(p) such that sOo= [ 2 eit)0s00) doie(d) ake where the integral converges in {°(a,b) to f, then the equation (L — Dx = f has a unique solution F in C(a,b) given by PW) = [7 Y= Dexa) dnd fest where the laiter integral converges in @*(a,b) to F.266 ORDINARY DIFFERENTIAL EQUATIONS [Cuar. 10 Proof. If for any finite » interval A, fa) = ff bender) doar jeot and Fag = ff 2, 0 — DriesanesOn dent Skah then (L — DF, = fs. By Lemma 3.1 there exists an F e 2?(a,b) such that Pa — Fl 0 as 4+ (— 2,00), Also, by hypothesis, [Ifa — f| > 0 as 4—(—©,0), From the variation-of-constants: formula there exist continuous functions @ such that Fad) =) etd) [{ a0rfale) er + Y oldesttd) creat fet where the ¢;(A) are constants, anda <¢ 0 as 6 — (a,b). Proof of Theorem 3.1. The existence of f is provided by Lemma 3.1. Using this f, there exists by Theorem 2.2 a je 27(p) such that air) = [? aaltryne at and if Jal) =f, Y etaddu0a) dont) 2) iw d then ||f ~ fal| + 0 as A— (—©,00), The problem is to show le - ail = 0 If fs is defined os in (8.2) with g replaced by g, then |lf — fall > 0 as A—(-—2,0), Using Lemma 3.2, it follows that the unique solution F of (L — 2)z = f which is in 0(a,b) is given by ro = [*, 2 = Detandon(A) dan) i 1 = [OY @ = de HO) dead ike Similarly, the unique solution of (L — i)z = F which is in 2°(@,b) is FiO = [7 YO dela) don) hbo = [71% 0 = 2 eltNKO) doa) jem This may be repeated to obtain for p = 1,2,..., [2. ¥ @ a7 (4x0) dead) = 0 83) ike1 where r = 9 — §. Let T' be the transform of the function which is one on (c, ¢ + 5) and zero elsewhere, that ia, TyQa) = f°" alan) at268 ORDINARY DIFFERENTIAL EQUATIONS [Cwar. 10 Integrating (3.3) with respect to ¢ from ¢ to ¢ + 8, one obtains [2.3 0 a-70,0n0) dead) = 0 @4) jket Since T and r are in $*(), the function H of | defined by a n= [" d —D-T\Q,e)74() dowel) jbat js analytic for $1 > 0. By (3.4) H and all its derivatives vanish for l=4, Thus 4@ = 0 for $1>0. A similar result holds for the lower half plane, that ia, with 2 replaced by If2 = « + i», then 1 __t a iv YI" Y=" W=+H Thus, using H() — H() = 0, -< ry 7 Skat If \, and A; are points of continuity of p and (3.5) integrated with respect to # from A; to Az, holding » fixed, and then letting vy — 0, it follows that . LY. neon00 dead) = 0 Jem If this is differentiated with respect to s, BY, ede +4, 20) dan) = 0 ikeL and if derivatives aro taken with respect to s and evaluated at s = 0, there follows by (2.2) LY ne dma =0 Gat...yn) 3.6) ket Because of the arbitrary character of }, and Az, if y; are any step functions which vanish for large |A}, then, [2 ¥ worn) dena = 0 ket This implies ||rl| = 0, which was to be proved.Bec. 3] SINGULAR BELF-ADJOINT NTH-ORDER PROBLEMS 269 Theorem 3.2. Let neither equation (L + t)r = 0 have a nontrivial solution in @(a,b). Then the matrix p is unique in the sense that if 6 ia any other matriz for which Theorem 2.2 is valid then BO) — a(n) = p(d)-— (x) at potnts of continuity d, w of p and 5. Remanx. An immediate consequence of Theorem 3.2 is that for any , # which are points of continuity of p ard) — palu) ~> oC) — p(u) ag § — (a,b), irrespective of how U, varies with 6. Thus if neither equa- tion (L + 2)z = 0 has a sulution (0) in &*(a,b), p is called the spectral matriz of the problem Lx = tz on (a,b). The spectrum, point spectrum, and continuous spectrum for this problem are defined in terms of p, as.in See. 5, Chap. 9. Proof of Theorem 3.2. Let f(t) = lle StS +7and f(t) = 0 other- wise, and let ath) = f°" axle) at Then, much as in the proof of Theorem 3.1, F, the unique solution of La = ix + f, is given by FO = [7 YO =~ d-wslt doer) deal) dike = [7X a= ae no.r) din ike Also, if on = pa — Bay HO = [7 Y= D-T,9007) dowd) jea1 must vanish for $f > 0 and this leads to a . LY 02) daar) = 0 aol in place of (8.6). Taking the derivative with respect to r of the left side above, there results LOY ale +7) dead) =0 st270 ORDINARY DIFFERENTIAL EQUATIONS [Caar. 10 Differentiating m — 1 times with respect to 7 and setting r = 0, there results [P' dom(d) = 0 which proves the theorem. The other case where the inverse-transform theorem, and also the uniquoness of p, will be proved is one for which the open interval (a,b) is replaced by [a,b), ~ © m, and ; has all the properties that py has. If p is any limit matrix obtained by letting §— (a,b) through o sequence of intervals, then clearly @ = CpC* is a matrix with elements having row or column index exceeding m equal to zero. In terms of 4, the Parseval equality and expansion theorem become for any fe 2?(a,b) lt = {7 GOO) dino) (3.8) jeed where . ai) = f° Herre at 8.9) and f= [2 Y wearne dao (10) deed where the integral on the right of (3.9) converges in the norm of &(f) and that of (3.10) in £*(a,b). The analogue of Lemma 3.1 is now valid, that is, if g ¢ 2*(f) there exists an f¢ &*(a,b) for which (3.10) is valid. Similarly, the representation of f in terms of 6 gives the analogue of Lemma 3.2 as follows: Lemma 3.3. Suppose for some 1, Sl x 0, that the problem Le=lz UM =0 (3.41) has no nontrivial solution in @{a,b). If fe (a,b) and g ie any vector in 2(4) such that (3.10) holds, then the problem (L-z=f UM, =0 has a unique solution F e 2°(a,b) given by FO = [7 Y= DWE AAO) du) ikem where the latter integral converges in 27{a,b) to F. Proof. The proof is the same as for Lemma 3.2 if it is observed that, for any finite \ interval A, the function 4 defined by Fa =f) YO - D-AEA)HO) daa) ihn satiefies (L — Fs = fa, and UF, = 0 since the y, satisfy the condition272 ORDINARY DIFFERENTIAL EQUATIONS. [Cuap. 10 UMz = 0, Thus for some constants ¢(A) and continuous functions 4 a m ¢ Fa = Y. vt [i acre) dr + Y elanvaltn) iw j=. The analogues of Theorems 3.1 and 3.2 are now immediate. ‘Theorem 3.3. Suppose neither problem ti)2=0 UM=0 has a nontrivial solution in @(a,b). If ge 2°(p), there exists an f ¢ 2*(a,b) given by (3.10), and in ferms of this f the vector g is represented by (3.9). Moreover, the mairiz p is unique in the sense that if p is any other matrix for which (3.8) through (3.10) is valid, then B(A) — Ble) = AQ) ~ AGH) at continuity pointe d,u of f and p. Thus if A,u are continuity points of , then Bi(A) — Asx) > BO) — AQ) as } = [a,b] — (a,b), irrespective of how U{? varies with 8. The matrix pis called the spectral matrix of the problem (3.11) with respect to the set {¥s]. The two cases, that of Theorem 3.1 and that of Theorem 3.3, will be referred to as Cases I and II, respectively. The inverse-transform theorem and the uniqueness of the spectral matrix hold for other cases also ‘f boundary conditions are added in the manner shown in Chap. 9 for the limit-circle case, but this will not be considered here. Let D represent the set of all functions ue {*(a,b) which are of class C*-! on (a,b), w°—? is absolutely continuous on every closed subinterval, and Lz e 9%(a,b). If a is finite, let & represent the set of all we D such that ue C*-! on {a,b) and U“u = 0. In Case I the problem La = lz on (a,b), and in Case II the problem Lz = lz, Uz = 0 on [a,b), are self- adjoint, in the sense that in Case I (La,v) = (u,Lv) for all u,v e D, and in Case Il this is valid for all uve. The proof in each case can be made to follow that of Prob. 13, Chap. 9. 4. Green’s Function In the treatment of Chap. 9 for n = 2, the existence of Green’s function in the limit-point and limit-circle cases was an immediate by-product of the method. Here a somewhat different approach will be followed. The existence of Green’s function G, for the self-adjoint problem Lx = lx, Use = 0 on 8 was proved in Chap. 7. It will be shown that there existsSec. 4) SINGULAR SELF-ADJOINT NTH-ORDER PROBLEMS 273 a sequence of intervals 4, — (a,b) such that the corresponding Green’s functions G,, tend to a function G which is a Green's function in the Cases T and I] diseussed in Sec. 3. ‘The relation of Green's function to the spectral matrix shown in Sec. 5, Chap. 9, will be shown to hold in general in the aection that follows. Use will be made of the existence of a function K = K(t,r) defined for a <é,7 < b which serves asa kernel in the variation-of-constants formula for the solution of Lz = f. This function is such that J, kGngen ar is a solution of Lz = f on any subinterval 4 of (@,b), and K has the same differentiability properties as any Green's function associated with L. In particular, @°-'K/di7-' has the same discontinuity at ¢ = + as the Greon’s functions. There are many such K, and the existence of one such is given in (2.4) of Chap. 7 for the case La = lx + f; see also Prob. 22, Chap. 3. In the following it will be convenient to denote functions such as K, considered as a function of ¢ for fixed r, as K(_ ,r), and similarly if eon- sidered os a function of x alone it will be denoted by K(t, _ ). Lemma 4.1. The set of functions {Gi} is uniformly bounded and (for n> 1) equicontinuous on every compact (t,7,1) region where $l #0. (If n= 1,t =r ts excepled.) Proof. Let the closed interval 8) be contained properly in the closed interval 6,, which in turn is properly contained in (¢,b). Let » be any real-valued function of class C* on (a,b) such that u(t) = 1 on some open intervat 5, docdocé,, and w(t) = O for t outside 8. Then dofine J by Fer) = WOK (t7) Let 8D 6; and ze é&. The function u = Gi(_,z,l) — J (yz) is of class C* on 6 and satisfies the boundary conditions Uszu = 0, Therefore, since (L — lu = —(L; — DJ, where L, denotes L applied to J considered as a function of ¢, Gallyrt) = Tea) — fee DILT (er) — Viele A.) Applying the Schwarz inequality to (4.1), there results GiGr2)| S Weel + NOs, Pile F( x) — WC zl, (4.2) The uniform boundedness of the G, for ¢,7 ¢ do and / ranging over some compact set A with 31 s 0 will follow from (4.2) once it has been shown that [|G.(, Dll, is bounded uniformly for ¢¢ 50, Le A.274 ORDINARY DIFFERENTIAL EQUATIONS [Crar, 10 However, this follows from the fact that if u = 9.(2)jf, where DFO) = fuer Nfs) ar and fe 0(8), then [ells 1SU-*USle (4a) Indeed, u satisfies (L — Du =f and Uju = 0, Using Green's formula, fi Gaya at — fuca a =0 2gtiulld =f, Ju — fa) at and using the Schwarz inequality, thia yields (4.3). Applying (4.8) tou = Gs(_,7,1) — J(_,r) for re 69 results in N@s( Dlle SWC ally + BUM 7} YC wl, 44) ‘Thus ||G(_ ,z,)|la is uniformly bounded for 7 e 8, l¢A,andé> 8. From the symmetry relation @j(¢,7,1) = Gi(,t,J) it follows that |G(t, Dilla is also uniformly bounded, and hence by (4.2) so is {Gs} for ,r ¢ 0, le A, and 8D d. . From (4.1) follows (for 2 > 1) Be ann = % us) - f Glbeu) Fado) —U(as)\ds (4.5) and, using the Schwarz inequality on the integral, the uniform bounded- ness of the set (8G,/ar] for ir ¢ dp and le A results. (If = 1, the inte- gral in (4.1) is taken as a sum of an integral from a, to 7 and an integral from to bi, where 8; = {a;,bi].} Thesymmetry of G; implies the uniform boundedness of {dG,/at| also. The uniform boundedness of [dG,/al} follows from the analyticity of G, in 2 and the uniform boundedness of {G.}. The uniform boundedness of all first partial derivatives of G; implies the equicontinuity of the set {@,}. (If n = 1, the set {G; — J} ig equicontinuous.) This completes the proof of the lemma. This lemma, together with the Ascoli lemma, proves that there exists a sequence of intervals 5,, C (a,b) (m = 2,3, . . .), where 3, — (a,b) a8 m— ~, such that the corresponding Groen’s functions G, = Gs, tend uniformly on any fixed compact subset A; of a < t, r Ai. By taking a sequence {A;} tending to the seta ©, to the right sides in (4.9) and (1.10), it is seen that G=0,1,...,2) (4.11) uniformly on any compact {t,7,!) region, where Ql 0, and provided té7whenj =2— 1,n. The symmetry relations imply that if Wn 2G =O... ym) under the same conditions that (4.11) is valid. Returning to (4.6) through (4.10), it is casy to see that the mixed derivatives o+G/atiar G,k = 0,1, ...,2— 1) exist and satisfy (iv). Relation (v) has been proved. The proof of (vi) is based on (4.4). From that inequality there exists & constant c, (depending on 4» and 8, only) such that Ha Db S lS + 1) er (ra ba) But (Gs jz) ile SIGs yzDils for 8 C 4, ond letting first 6—> (a,b) through the sequence 8,, and then §— (a,b), it follows that, for any fixed (,D, St 0, GC 7,1) 8 8%(G,b), This also gives for fixed (4.1), St x 0, S(t, De X*(a,b).See. 4] SINGULAR SELF-ADJOINT NTH-ORDEK PROBLEMS: 27 It remains to prove (vii). If fe £*(a,b), then the integral Lounyod @ 0, uniformly for ¢ on any finite subinterval of (a,b). Thus the integral be [enantio converges uniformly for 4 on any finite subinterval of (a,b), and hence represents a continuous function on (a,b) which is easily verified to be v'. From (4.3) for 5 = 4,, letting m—> ©, it follows that loll S [Sell (4.12) which proves ve £2(a,b). Since Ly = lv +f, H woot s (1 + 4h) ta yielding Lv « £*(a,b), and completing the proof that ve D. For any fe 2°(a,b) let §(J)f denote the function given by (vii) in Theo- rem 4.1, The functions G arising in Cases I and II will now be considered. In Case II suppose G is obtained by taking for 80, 6, and all 3 the interval closed at a. Thus 6 = [a6]. The boundary conditions Usz = 0 which determine G, are assumed in this case to include Uz = 0. Thus in Case IIG, as a function of ¢ satisfies Ur = 0, All the convergence properties of Theorem 4.1 can be shown to hold uniformly over a S t, + S be for any by

S-ar putea să vă placă și

• The Subtle Art of Not Giving a F*ck: A Counterintuitive Approach to Living a Good Life

  

  De la Everand

  The Subtle Art of Not Giving a F*ck: A Counterintuitive Approach to Living a Good Life

  Mark Manson

  Evaluare: 4 din 5 stele

  4/5 (5794)
• The Little Book of Hygge: Danish Secrets to Happy Living

  

  De la Everand

  The Little Book of Hygge: Danish Secrets to Happy Living

  Meik Wiking

  Evaluare: 3.5 din 5 stele

  3.5/5 (399)
• An Accurate Solution For CVA

  

  Document29 pagini

  An Accurate Solution For CVA

  Adam Xu

  Încă nu există evaluări
• • Reviste
  • Podcast-uri
  • Partitură
• CH 7 Measuring Growth

  

  Document15 pagini

  CH 7 Measuring Growth

  Adam Xu

  Încă nu există evaluări
• Midterm Review F10

  

  Document2 pagini

  Midterm Review F10

  Adam Xu

  Încă nu există evaluări
• Syllabus - GSI

  

  Document2 pagini

  Syllabus - GSI

  Adam Xu

  Încă nu există evaluări
• Lecture 05 - Derivatives II

  

  Document15 pagini

  Lecture 05 - Derivatives II

  Adam Xu

  Încă nu există evaluări
• Math 202A Homework No.4: Yiting Wang - SID 22579477 Page 1 of 3

  

  Document3 pagini

  Math 202A Homework No.4: Yiting Wang - SID 22579477 Page 1 of 3

  Adam Xu

  Încă nu există evaluări
• Lecture 2

  

  Document38 pagini

  Lecture 2

  Adam Xu

  Încă nu există evaluări
• Career Launcher Finance

  

  Document177 pagini

  Career Launcher Finance

  incognito123123

  0% (1)
• Ans2 PDF

  

  Document21 pagini

  Ans2 PDF

  Defri Ahmad Lagu Lamo

  Încă nu există evaluări
• Foll and Real Analysis

  

  Document28 pagini

  Foll and Real Analysis

  Adam Xu

  Încă nu există evaluări
• BJF

  

  Document35 pagini

  BJF

  Adam Xu

  Încă nu există evaluări
• A Heartbreaking Work Of Staggering Genius: A Memoir Based on a True Story

  

  De la Everand

  A Heartbreaking Work Of Staggering Genius: A Memoir Based on a True Story

  Dave Eggers

  Evaluare: 3.5 din 5 stele

  3.5/5 (231)
• Hidden Figures: The American Dream and the Untold Story of the Black Women Mathematicians Who Helped Win the Space Race

  

  De la Everand

  Hidden Figures: The American Dream and the Untold Story of the Black Women Mathematicians Who Helped Win the Space Race

  Margot Lee Shetterly

  Evaluare: 4 din 5 stele

  4/5 (894)
• The Yellow House: A Memoir (2019 National Book Award Winner)

  

  De la Everand

  The Yellow House: A Memoir (2019 National Book Award Winner)

  Sarah M. Broom

  Evaluare: 4 din 5 stele

  4/5 (98)
• Shoe Dog: A Memoir by the Creator of Nike

  

  De la Everand

  Shoe Dog: A Memoir by the Creator of Nike

  Phil Knight

  Evaluare: 4.5 din 5 stele

  4.5/5 (537)
• Elon Musk: Tesla, SpaceX, and the Quest for a Fantastic Future

  

  De la Everand

  Elon Musk: Tesla, SpaceX, and the Quest for a Fantastic Future

  Ashlee Vance

  Evaluare: 4.5 din 5 stele

  4.5/5 (474)
• Never Split the Difference: Negotiating As If Your Life Depended On It

  

  De la Everand

  Never Split the Difference: Negotiating As If Your Life Depended On It

  Chris Voss

  Evaluare: 4.5 din 5 stele

  4.5/5 (838)
• Grit: The Power of Passion and Perseverance

  

  De la Everand

  Grit: The Power of Passion and Perseverance

  Angela Duckworth

  Evaluare: 4 din 5 stele

  4/5 (587)
• Devil in the Grove: Thurgood Marshall, the Groveland Boys, and the Dawn of a New America

  

  De la Everand

  Devil in the Grove: Thurgood Marshall, the Groveland Boys, and the Dawn of a New America

  Gilbert King

  Evaluare: 4.5 din 5 stele

  4.5/5 (265)
• Yes Please

  

  De la Everand

  Yes Please

  Amy Poehler

  Evaluare: 4 din 5 stele

  4/5 (1891)
• Angela's Ashes: A Memoir

  

  De la Everand

  Angela's Ashes: A Memoir

  Frank McCourt

  Evaluare: 4.5 din 5 stele

  4.5/5 (440)
• The Emperor of All Maladies: A Biography of Cancer

  

  De la Everand

  The Emperor of All Maladies: A Biography of Cancer

  Siddhartha Mukherjee

  Evaluare: 4.5 din 5 stele

  4.5/5 (271)
• On Fire: The (Burning) Case for a Green New Deal

  

  De la Everand

  On Fire: The (Burning) Case for a Green New Deal

  Naomi Klein

  Evaluare: 4 din 5 stele

  4/5 (73)
• The Hard Thing About Hard Things: Building a Business When There Are No Easy Answers

  

  De la Everand

  The Hard Thing About Hard Things: Building a Business When There Are No Easy Answers

  Ben Horowitz

  Evaluare: 4.5 din 5 stele

  4.5/5 (344)
• Team of Rivals: The Political Genius of Abraham Lincoln

  

  De la Everand

  Team of Rivals: The Political Genius of Abraham Lincoln

  Doris Kearns Goodwin

  Evaluare: 4.5 din 5 stele

  4.5/5 (234)
• Fear: Trump in the White House

  

  De la Everand

  Fear: Trump in the White House

  Bob Woodward

  Evaluare: 3.5 din 5 stele

  3.5/5 (738)
• The Glass Castle: A Memoir

  

  De la Everand

  The Glass Castle: A Memoir

  Jeannette Walls

  Evaluare: 4.5 din 5 stele

  4.5/5 (1712)
• Rise of ISIS: A Threat We Can't Ignore

  

  De la Everand

  Rise of ISIS: A Threat We Can't Ignore

  Jay Sekulow

  Evaluare: 3.5 din 5 stele

  3.5/5 (137)
• Principles: Life and Work

  

  De la Everand

  Principles: Life and Work

  Ray Dalio

  Evaluare: 4 din 5 stele

  4/5 (599)
• The Unwinding: An Inner History of the New America

  

  De la Everand

  The Unwinding: An Inner History of the New America

  George Packer

  Evaluare: 4 din 5 stele

  4/5 (45)
• The World Is Flat 3.0: A Brief History of the Twenty-first Century

  

  De la Everand

  The World Is Flat 3.0: A Brief History of the Twenty-first Century

  Thomas L. Friedman

  Evaluare: 3.5 din 5 stele

  3.5/5 (2219)
• Steve Jobs

  

  De la Everand

  Steve Jobs

  Walter Isaacson

  Evaluare: 4.5 din 5 stele

  4.5/5 (806)
• John Adams

  

  De la Everand

  John Adams

  David McCullough

  Evaluare: 4.5 din 5 stele

  4.5/5 (2409)
• The Gifts of Imperfection: Let Go of Who You Think You're Supposed to Be and Embrace Who You Are

  

  De la Everand

  The Gifts of Imperfection: Let Go of Who You Think You're Supposed to Be and Embrace Who You Are

  Brené Brown

  Evaluare: 4 din 5 stele

  4/5 (1090)
• Bad Feminist: Essays

  

  De la Everand

  Bad Feminist: Essays

  Roxane Gay

  Evaluare: 4 din 5 stele

  4/5 (1015)
• The Outsider: A Novel

  

  De la Everand

  The Outsider: A Novel

  Stephen King

  Evaluare: 4 din 5 stele

  4/5 (1839)
• Brooklyn: A Novel

  

  De la Everand

  Brooklyn: A Novel

  Colm Tóibín

  Evaluare: 3.5 din 5 stele

  3.5/5 (1937)
• The Sympathizer: A Novel (Pulitzer Prize for Fiction)

  

  De la Everand

  The Sympathizer: A Novel (Pulitzer Prize for Fiction)

  Viet Thanh Nguyen

  Evaluare: 4.5 din 5 stele

  4.5/5 (119)
• A Man Called Ove: A Novel

  

  De la Everand

  A Man Called Ove: A Novel

  Fredrik Backman

  Evaluare: 4.5 din 5 stele

  4.5/5 (4609)
• The Light Between Oceans: A Novel

  

  De la Everand

  The Light Between Oceans: A Novel

  M.L. Stedman

  Evaluare: 4.5 din 5 stele

  4.5/5 (789)
• The Woman in Cabin 10

  

  De la Everand

  The Woman in Cabin 10

  Ruth Ware

  Evaluare: 3.5 din 5 stele

  3.5/5 (2322)
• Manhattan Beach: A Novel

  

  De la Everand

  Manhattan Beach: A Novel

  Jennifer Egan

  Evaluare: 3.5 din 5 stele

  3.5/5 (792)
• The Perks of Being a Wallflower

  

  De la Everand

  The Perks of Being a Wallflower

  Stephen Chbosky

  Evaluare: 4.5 din 5 stele

  4.5/5 (2099)
• Wolf Hall: A Novel

  

  De la Everand

  Wolf Hall: A Novel

  Hilary Mantel

  Evaluare: 4 din 5 stele

  4/5 (3811)
• Little Women

  

  De la Everand

  Little Women

  Louisa May Alcott

  Evaluare: 4 din 5 stele

  4/5 (104)
• The Art of Racing in the Rain: A Novel

  

  De la Everand

  The Art of Racing in the Rain: A Novel

  Garth Stein

  Evaluare: 4 din 5 stele

  4/5 (4200)
• Sing, Unburied, Sing: A Novel

  

  De la Everand

  Sing, Unburied, Sing: A Novel

  Jesmyn Ward

  Evaluare: 4 din 5 stele

  4/5 (1103)
• A Tree Grows in Brooklyn

  

  De la Everand

  A Tree Grows in Brooklyn

  Betty Smith

  Evaluare: 4.5 din 5 stele

  4.5/5 (1929)
• The Constant Gardener: A Novel

  

  De la Everand

  The Constant Gardener: A Novel

  John le Carre

  Evaluare: 3.5 din 5 stele

  3.5/5 (104)
• Her Body and Other Parties: Stories

  

  De la Everand

  Her Body and Other Parties: Stories

  Carmen Maria Machado

  Evaluare: 4 din 5 stele

  4/5 (821)