Sunteți pe pagina 1din 18

Numer. Math.

50, 27-44 (1986)

Numerische MathemaUk 9 Springer-Verlag1986

Bounds on Nonlinear Operators in Finite-dimensional Banach Spaces


Gustaf Stiderlind Department of Numerical Analysis and Computing Science, Royal Institute of Technology, S-10044 Stockholm, Sweden

Summary. We consider Lipschitz-continuous nonlinear maps in finite-dimensional Banach and Hilbert spaces. Boundedness and monotonicity of the operator are characterized quantitatively in terms of certain functionals. These functionals are used to assess qualitative properties such as invertibility, and also enable a generalization of some well-known matrix results directly to nonlinear operators. Closely related to the numerical range of a matrix, the Gerschgorin domain is introduced for nonlinear operators. This point set in the complex plane is always convex and contains the spectrum of the operator's Jacobian matrices. Finally, we focus on nonlinear operators in Hilbert space and hint at some generalizations of the yon N e u m a n n spectral theory.

Subject Classifications: AMS(MOS) 47H05, 47H12, 65H10, 65L07; CR:


G1.5, G1.7.

1. Introduction
Let f: D c C" ~ C" be a Lipschitz-continuous map on the path-connected set D, [14, p. 99]. Let II" II be a strictly homogeneous norm on C", i.e. II~xlt=I~l [Ixtl for all complex constants ~. The lub Lipschitz constant of f on D, subordinate to the given vector norm is defined by

(1.1)

L [ f ] = s u p Nf(u)-f(v)[I ,*~ Ilu-vll

u, reD.

Similarly, the glb Lipschitz constant is defined by

(1.2)

l [ f ] = inf IIf(u)-f(v)H .,v Llu-vll

u, reD.

Let f, g and h be Lipschitz-continuous maps defined on suitable sets. Then f + g denotes the operator that maps x onto f(x)+g(x), and fog denotes the

28

G. S6derlind

composition, i.e. f(g(x)). Note that composition is right distributive only, i.e. (f+g)oh=foh+goh, but ho(f+g)*hof+hog unless h is a linear operator. One easily shows that the functional L [ - ] has the following properties: L1) L [ f ] > O L E f ] = 0 ~ f=const., L2) L[~f]=]~[L[f], L3) L [ f ] - L [ g ] __<L[f+g] <=L[f] + L [ g ] , L4) L [ f o g ] <=L[f]L[g]. Thus it is a semi-norm on the space of Lipschitz-continuous functions. However, any two operators with f - g = c o n s t a n t on D may be interpreted as elements in the same equivalence class of operators. L [ . ] is therefore an operator norm on the quotient space formed by such equivalence classes. For a matrix A, we have LEA]= [lAtE and lEA] =glb[A]. Moreover, if A is nonsingular, then glb[A]=llA-1ll-k As a direct generalization of this, we proved the following result in [191, cf. [14, p. 106]: (l.3) Lemma. Assume that the Lipschitz-continuous map f satisfies l [ f ] > 6 > O on the path-connected set D. Then f - 1 exists and is Lipschitz-continuous on f(D), and L [ f - a ] = l [ f ] -'. Two new functionals were introduced in [191. Let I + e f denote the operator that maps x6D onto x+ef(x). The lub Dahlquist constant of f on D is then defined by L[I+ef]-I (1.4) M [ f ] = lim
~0+

Likewise, the glb Dahlquist constant of f on D is defined by (1.5) m[f]= lim


E~O+

l[I+ef]-I
E

The existence of these limits was proved in [19]. (The existence of M [ f ] follows directly from the fact that every norm is Gateaux differentiable, cf. [6, p. 49], whereas the existence of m [ f ] is slightly more complicated.) Moreover, the following analogue of Lemma (1.3) was established: (1.6)

m[f] = -M[-f].

Thus M [ f ] and m [ f ] are, respectively, the right and left-hand Gateaux differentials with increment f, of the lub Lipschitz functional at the identity operator I. These limits do, in general, not coincide, and we always have that m I f ] __< [ f ] . M The functional M [ . ] obeys the following rules: M1) - l [ f ] < = M [ f ] ,<L [ f ] , M2) M [ f + z l ] = M [ f ] + R e z , M3) M [ ~ f ] = o ~ M [ f ] a e R +, M4) M I f ] + m [g] __< I f + g] <=M I f ] + M [g]. M

Bounds on Nonlinear Operators in Finite-dimensional Banach Spaces

29

The corresponding rules for m [ ' ] are easily derived from (1.6) and the rules above. For a matrix A, we have M [A] =/~ [A], where
/~ [ A ] =

lim
e~O+

III+eAN-1

is the logarithmic norm of the matrix, introduced by Dahlquist [3], see also [10] and [18]. The relation between the logarithmic norm and the numerical range (or Bauer field of values) is also discussed in [10] and [1]. While L [ f ] is related to contractivity, M [ f ] relates to monotonicity. Thus f is monotone with respect to the given norm if either M [ f ] <=0 or ruff] >0. If the Euclidean norm 11"112 is used, then m 2 [ A ] > 0 (M2[A]<O) if and only if A is positive (negative) definite. For this reason we shall refer to f as positive monotone, strictly positive monotone or uniformly positive monotone if, respectively, r u f f ] > 0 , m [ f ] > 0 or m [ f ] > 8 > 0 on the set D, cf. [14, p. 1411 for definitions when the norm is induced by an inner product. 1 Analogous definitions, using corresponding conditions on M [ f ] , apply to negative monotonicity. A generalization of the Uniform Monotonicity Theorem of Browder and Minty, [14, p. 167], [2, 8], the following result was proved in [19]. (1.7) Lemma. Let f: DcC'~--*C m be Lipschitz-continuous

and assume that M [ f ] < = - 8 < 0 on the path-connected set D. Then f - 1 exists and is Lipschitzcontinuous on f (D), and L [ f -1] < - M [ f ] - k
(1.8) Remark. a) Note that this result is equivalent to the left inequality in Rule M1 above. b) The corresponding result with the assumption m [ f ] > 8 > O is clearly

L [ f - ' ] < = m [ f ] -1.


c) Slightly stronger versions of Lemma (1.3) and (1.7) were proved in [19]. Thus, if in addition D = C', then f is a homeomorphism from C m onto C m. If f is continuously differentiable on D, one can show that sup IIf'(x) ll < L [ f ] ,
xeD

sup g [ f ' ( x ) ] =<M [ f ]


xeD

with equality if D is convex. Analogous results hold for l [ f ] and m[f]. It follows that L [ f ] and l [ f ] are upper and lower bounds for the moduli of the eigenvalues of if(x) when xeD. Similarly, M [ f ] and ruff] are upper and lower bounds, respectively, for the real parts of the eigenvalues, [3, p. 12]. These results enable us to give simple geometric characterizations in the complex plane of the assumptions we use in the following sections, see also Fig. 1 in [191. The following reformulations of the limits (1.4) and (1.5) will be used frequently in the sequel: (l.9)

M [ f l = lim L [ f +rI] - r ,

1 We use this terminologyhere, although other terms, e.g. "maximal monotone'; "'accretive"and "dissipative", are frequently used in the literature for strongly related concepts, see [6]

30

G. S6derlind

L[f]

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

~[f]

'

.................................

Fig. 1. L[f+rl]-r as a function of r in a case where f is uniformly negative monotone

[iiiiiiiiiii.iiiii.. . . .iii.. . . . . . .
m[f]= lim t [ f + r I ] - r .
r~oo

(1.10)

We will also consider L [ f + r I ] - r and l [ f + r I ] - r for finite values of r; a typical plot of the former as a function of r is given above. Note that convergence to the limit is always monotonic, [3], as shown in the figure, although L I f + r l ] - r need not be continuously differentiable for all r. In the following sections we shall focus our attention on invertibility for the operator f. Our aim is not to find the weakest possible conditions under which inverses exist, but rather to find simple sufficient conditions and establish bounds for the inverses within the framework of the formalism outlined above. The main advantages of this formalism are its simplicity, the fact that there are no differentiability assumptions (all conditions are in terms of the four basic functionals) and that monotonicity is extended to Banach space without invoking duality. Furthermore, we do not assume that the domain D is convex; it is sufficient that D is path-connected.

2. lnvertibility in Banach Space


We shall now derive bounds for inverse operators in Banach space. The results obtained have the same structure as the corresponding bounds for matrices. Throughout, we assume that the Lipschitz-continuous operators f and g are defined on the same path-connected set D. The first result is a nonlinear version of the Perturbation Lemma, [14, p. 45]. Theorem. Assume that L [ g ] < l [ f ] - c 3 < l [ f ] on D. Then ( f _ g ) - i on ( f -g)(D), and 1 (2.2) L [ ( f - g)- 1] < =l[f]-L[g]" (2.1)

exists

Proof Introduce the notation A f = f ( x O - f (x2), Ag=g(xO-g(x2) and A x = x 1 - x 2 for x 1, xz~D. Then IIA(f-g)]] > llAf Jl- ][Agl[-> ( l [ f ] - L[g])NAxlI.

Bounds on Nonlinear Operators in Finite-dimensional Banach Spaces Hence l [ f - g ] > l [ f ] - L [ g ] > 6 > O Lemma (1.3).

31

and the result follows by application of

In most applications, g is a small perturbation of the invertible operator f (N.B. l I f ] > 6 >0). If the perturbation is small enough, the theorem states that the perturbed operator is also invertible with a Lipschitz-continuous inverse. There are, however, other important applications, such as the following nonlinear analogue of the Neumann Lemma: (2.3) Corollary. Assume that L [ g ] < l - 6 < l - g)(D) and (2.4) L [(I - g)- x] < _ _ on D. Then ( l - g ) -~ exists on (I
1

1 -L[g]"

Assume that g is a small perturbation of f, set h . ' = f - g and assume that h -1 exists in accordance with the assumptions in Theorem(2.1). A bound on h- 1 - f - ~ answers the question of how well h- t approximates the inverse of f Such a bound is easily obtained for matrices: A - 1 - B - ~ = A - t ( B - A ) B -1 yields the Banach Lemma, IIA -1 - B -111 < IIA-1 I[' lIB-All-lIB -1 I1- In the nonlinear case, however, we have h - 1 _ f - 1 = (h- 1o f - I) o f - 1, whence (2,5) L[h-a-f 1]<L[h-lof-I]L[f-x].

A bound in terms of L [g] = L [ f - h ] cannot be obtained if h is nonlinear, since composition is not left distributive. On the other hand, if h is linear, h(x)= Hx, or affine, h ( x ) = H x + y , then L[h - 1o f - 13 = L [ H - t ( f _ H)]. Hence, (2.6)

L[H-X _ f - l ] <

IIH-1 HL [ f - H ]

L[f-1].

Here H denotes both the affine map h and its Fr6chet derivative h'. The most important application of (2.6) is when the given system is linearized. Thus, if h ( x ) = f ( X o ) + f ' ( X o ) ( X - X o ) for some xo~D, then H = f ' ( x o ) . Hence we have (2.7) L[f,(Xo)-X _ f - x ] __<ilf,(Xo)-X i l L [ f _ f , ( X o ) ] L [ f 1].

Theorem (2.1) and its corollary also have counterparts in terms of monotone operators : (2.8) Theorem. Assume that M [ g ] < m [ f ] - 3 < m [ f ] on D. Then f - g formly positive monotone and invertible on ( f -g)(D), and (2.9) L [ ( f - g)- 1] _ _ <_ m [ f ] - M [gl" is uni-

Proof Using Rule M4 and (1.6), we have m [ f - g ] > = m [ f ] + m [ - g ] = re[f] - M[g] > 3 >0. The result follows by application of Lemma (1.7). The following important special case of Theorem (2.8) was used in [t91 to derive a global error bound for the implicit Euler method applied to differential

32

G. S6derlind

sEf] ~

if" /
I t f

S Y7_

M[gl

re[f]

Fig. 2. Illustrations of a) Theorem (2.1), b) Theorem (2.8) and c) Theorem (2.14)

systems, and in [201 to derive a new bound for the error in fixed-point iterations: (2.10) Corollary. Assume that M [ g ] < l - 6 < l on D. Then I - g positive monotone and invertible on (I-g)(D), and (2.11) L[(I-g)-'] < = 1-M[g]'
1

is uniformly

(2.12) Remark. By Rule M1, M[g]<L[g]. Hence the bound (2.11) holds also under the assumption of Corollary (2.3). The converse, however, is not true, i.e. under the assumption of Corollary (2.10), (2.4) is generally not valid. In fact, the bound (2.11) holds irrespective of the size of L[g]. Also note that Theorem (2.8) remains valid if M [ g ] is replaced by L[g], thus transforming the result into a "Perturbation lemma" for a uniformly positive monotone operator f. As mentioned earlier, the conditions used in the theorems can easily be given geometric interpretations in the complex plane, since the functionals used are also spectral bounds. Thus conditions involving Lipschitz constants correspond to circular domains in the complex plane, whereas conditions in terms of Dahlquist constant correspond to half-planes, see Fig. 2 above. We shall extend Theorems (2.1) and (2.8) to a general perturbation lemma. Introduce two circular domains as follows: (2.13)

~s[f] = {2~12:12+rl1--r 1 >=l[f + rl I ] --rl} aS[g] {peC: ]p+r2l-r2 <=L[g+r2I]-r2}.

The numbers r 1 and r 2, which can be chosen arbitrarily, appear on both sides of the inequalities. This may seem superfluous, but makes the inequalities well defined also in the limiting case when ri~oe. In such a case they form conditions of the type R e 2 > m [ f ] and R e p < M [ g ] , respectively, see Fig. 2b above and (1.9-10). The radii of the circles Os[f] and ffS[g] are R = l [ f + r l I ] and r = L [g + r2I], respectively. The centers are located at r 1 and r2; hence they are offset by [rl-rz] so that i n f [ p - 2 l = R - r - l r l - r 2 ] , when # e S [ g ] and 2 ~ s [ f ] , and S[g] c~s[f] =O. (2.14) Theorem. Let f and g be defined on the same path-connected set D, and let s [ f ] and S[g] be defined by (2.13). Let 2 e s [ f ] and #eS[g], and assume that

Bounds on Nonlinear Operators in Finite-dimensional Banach Spaces infl)~--#l>6>0. Then ( f _ g ) - i exists on ( f - g ) ( D ) , and (2.15)

33

L [ ( f - g ) -1] ___sup 12-p] -1.

Proof Using the same notation as in the proof of Theorem (2.1), we have
HA(f-g)]b = ] t A f + r x A x - ( A g + r 2 A x ) - ( r l - r 2 ) A x ] l

>=]IAf + q Ax]I- [IAg + r z A x ] l - Ira --r2l IIAxll >=(l[f + r11]- L[g +r21] -Irl -rzl) ]lAx]l =(R - r - l q -r2[ ) }[Ax][ = inf[2-#1. I[Ax]l
where 2 e s [ f ] , #~S[g]. H e n c e / [ f - g ] > _ - i n f l A - p l , and the bound follows from Lemma (1.3). Theorem (2.14) contains the results of Theorem (2.1) and, by using a limiting argument, Theorem (2.8). Moreover, it can easily be modified to allow the centers of the circular domains to be arbitrary points in the complex plane. The most interesting aspect of the theorem is, however, that the circular domains associated with each operator have an interpretation as representing "spectral sets" of the operators. This idea is further developed in the following section.

3. Gerschgorin Domains
The significance of (1.9) is that the limit is evaluated in the positive real direction. We may, however, also use limits of the type (3.1) M r [ f ] = lim L [ f + re i~ I] - r = M [e-i~f].
r~oo

Interpreted geometrically in the complex plane, this limit corresponds to the half-plane (3.2)

C ~ [ f ] = {z~C: R e e - ~ z < M [ e - i ~ f ] } .

Closely related to the numerical range (or Bauer field of values) of a matrix, the Gerschgorin domain of f is introduced by (3.3)

G [ f ] = ~ C~[f].
q~

Obviously G [ f ] is a convex compact set. In fact, for a matrix A, the Gerschgorin domain coincides with the numerical range if an inner product norm is used, whereas for other norms, G [A] is the convex hull of the numerical range, cf. {-1, p. 108]. While the usual generalization to Banach space of the numerical range, using duality, causes a loss of convexity and compactness in certain cases, [13, 22], the Gerschgorin domain retains its qualitative properties regardless of the choice of norm. Therefore, we shall in the sequel only work with the Gerschgorin domain. Set z = x + iy. Then (3.4)

OC ~ [ f ] = {z~C : x cos q~+ y sin q~= m ~ [ f ]}.

34
~co[f]

G. S6derlind

Im,

Fig. 3. Proof of Theorem (3.5)

(3.5)

Theorem. For every ~o, OCt[f] is a support plane of G[f].

(3.6) Remark. Note that there may be points on OG[f] having a non-unique support plane.

Proof It suffices to show that c3Co[f]={z~C: R e z = M [ f ] } plane of G [ f ] , i.e. O C o [ f ] c~G[f] +~b.

is a support

Let - ~ < ~b< 0 < ~o< ~ and introduce z* = ~ C~ [ f ] c~ ~ C~,[f]. Now suppose that OCo[f]r~G[f]=O. Then there exist q4 ~o such that OCo[f]c~C~[f] c~ C~0[f]=~b. Therefore, the statement follows by showing that this cannot occur, i.e. Re z*_->M [ f ] for ~0- ~b< ~, see Fig. 3 above, and Re z* < M [ f ] otherwise. Now, z* satisfies

Ree-i~ z* = m[e-i'P f], Ree-i* z * = m[e-iq' f].

Thus Re z* sin (tp - ~) -- m [ e - i q ' f ] sin q~- M [ e - i ~ f ] sin ~k. However, by rules M3 and M 4 we have

M [e-iq'f] sin ~0- M[e-lq'f] sin ~k= M [ s i n q)e-lq'f] + M [ - sin ~ke-i~f] > m [(sin q~e-lq'- sin tpe-i*)f] ---M[sin (tp-qJ) f ].
Hence, by M3, R e z * > M [ f ] when sin(q~-~9)>0, Re z* < M [ f ] when ~o- ~k> g. The proof is complete. i.e. for q 3 - q J < g and

Theorem (3.5) has important consequences. The following is a well-known characterization of the numerical range which carries over to the Gerschgorin domain and the nonlinear case. (3.7) Corollary. M [ f ] = m a x { R e 2 : 2 e G [ f ] } and m [ f ] = m i n { R e 2 : 2eG[f]}.

By T h e o r e m (3.5), the boundary of G[f] is the envelope formed by the straight lines c3C,[f], 0__<tp<2~z. Thus, if z = x + i y denotes a point on c~G[f], then (cf. 3.4) x cos tO+ y sin (p = M,~[f] t x sin q~+ y cos (p = M,p [ f ]
-

Bounds on Nonlinear Operators in Finite-dimensional Banach Spaces

35

where M',p[f] denotes the derivative of M ~ [ f ] with respect to the parameter q~. Hence the boundary points satisfy

(3.8)

[;] = [cos
Lsin q~

-sin
cosq~ LM~[f]J"

Note, however, that since M ~ [ f ] may not be continuously differentiable for all ~p, (3.8) is only a "formal" envelope. Thus there may be corners on OG[f] and it may also contain straight line segments. As an example, we may consider G~[A], i.e. the Gerschgorin domain of a matrix A, subject to the max norm. Since (cf. [3, 18]) (3.9) Moo [A] = max {Re a# + ~' lajkl},
J k

where ~ ' denotes the sum over off-diagonal elements, we have (3.10)

M~[e-i~A]=max {cos (p R e a j j + s i n q~Ima~j+ ~ ' Taj~l}.


J k

There derivative is then, piecewise, (3.11)

M ' [ e - i ~ A ] = - s i n q~ Re aj~+ cos q~Im ajj.

Inserting (3.10) and (3.ll) into (3.8) shows that the envelope piecewise is given by (3.12)

z = ajj + ei~~~' lajkl.


k

This is clearly an arc of a Gerschgorin circle, and hence, for each q), the support plane OCt[A] is tangent to at least one Gerschgorin circle. At a point ~o where the maximizing index j changes in (3.10), there will be a jump discontinuity in z, see (3.12). By (3.4), however, the slope of OCt[A] changes continuously at such points. Thus, these discontinuities correspond to straight line segments of the form (3.4) joining the circular arcs (3.12). Since by (3.10) G~[A] must contain all Gerschgorin disks, it follows that G~ [A] is the convex hull of the union of the (row-wise) Gerschgorin disks of A. With respect to the 11 norm, one obtains the same result for column-wise Gerschgorin disks. We note that Nirschl and Schneider [13] have come to the same conclusions, using duality, for the convex hull of the numerical range. The Gerschgorin domain has a number of interesting properties: (3.13) Theorem. a) If f~CI(D), then for every x~D, G [ f ] contains the spectrum of f'(x). b) G [ f ] c { z e C : [zl<L[f]}, c) Affine covariance: Let T be a (complex) affine transformation. Then G[Tf] = T G [ f ] .

Proof a) Since Vx~D, p [ f ' ( x ) ] < m [ f ] , it follows that the spectrum of f'(x) for every ~o is contained in the half-plane (3.2), cf. [3] or [18]. The result is also well-known for the closure of the numerical range of a matrix, see e.g. [211.

36

G.S~derlind

b) This result follows immediately from (3.2) and M1, i.e. IM[f]l < L [ f ] . c) Let T z = z l z + z o for fixed Zl,Zo~C, and assume that z l + 0 (otherwise the result is trivial) in order that T -1 exist. Let z~G[Tf]. Then, the half-planes (3.2) are of the form Ree-i~~ z ~ M[e-i'pTf]. Using M2 and M3, one finds that Re e-ic~o-argzl) T - 1z < M [e-it~o- a r g z ,~f]. Hence T - l z ~ G [ f ] , i.e z ~ T G [ f ] . According to Theorem (3.13b), G[f] is always contained in a circle of radius L [ f ] centered at the origin. Note, however, that there need not be any point z ~ G [ f ] for which {zl=L[f]. One may then ask whether there exists a "sharper" norm in the sense that f has a Lipschitz constant with respect to the new norm that is smaller than the previous L [ f ] . This is obviously true for matrices. For, since the spectrum is contained in G[A] (and there can be no defective eigenvalues on ~3G[A], a consequence of Theorem (3.5) and [18], see also [131) one can use the fact that there is a norm such that tlAII is arbitrarily close to the spectral radius p[A]. It is important to note, however, that the existence of a point zeG[A] such that Iz]= IIA]t does not imply that the norm is "optimal"; for the l I and l~ norms such a point exists for every matrix A. We have not been able to prove a similar result for nonlinear operators, although there are reasons to believe that it is true also in the latter case. Therefore we formulate the following conjecture: (3.14) Conjecture. Assume that f: DcCm--*C m has a lub Lipschitz constant L [ f ] with respect to a given norm U'II on C m. Let L--maxlzl, z6G[f]. If L < L [ f ] , then there exists a new norm on C" such that f has Lipschitz constant LonD. (3.15) Remark. There are also variants of Theorem (3.13b) relating to other circular domains. Thus, for instance, we have L [ f + r l ] < R ~ G [ f ] c { z ~ C : ]z+r]<=R}. We shall conclude this section by presenting a theorem similar to Theorem (2.14) based on assumptions on the Gerschgorin domains rather than circular domains in the complex plane.

Let 2 , 2 * e G [ f ] and (3.16) Theorem. Assume that G[f]nG[g]=O. It, #* e G [g] be such that [ 2 * - # * { = m i n l 2 - / z l . Then ( f _ g ) - I exists on ( f - g)(O), and
(3.17)

L [ ( f - g ) -1 ] <12" - # ' 1 - 1 .

Proof. Setq~ = arg (2* -/~*). Then e-i~,(f_ g) is uniformly positive monotone, see Fig. 4, and the bound follows by application of Theorem (2.8).
(3.18) Remark. Two special cases of this theorem are worth noting. Using the fact that G [ 2 I ] = {2}, we find that 2 r = ~ ( g - 2 I ) -1 exists, cf. Theorem (2.8)

Bounds on Nonlinear Operators in Finite-dimensional Banach Spaces

37

Fig. 4. Proof of Theorem (3.16)

and its corollary. This has also been noted by Browder [2], although in a slightly different form and in connection with inner product norms. Specializing further, we see that 0 r ~ g - 1 exists, with a Lipschitz constant equal to the reciprocal of the minimum distance from G[g] to the origin.

4. Some Bounds in Hilbert Space


The additional structure provided by an inner product enables us to extend some of the results presented in the previous two sections. To be more precise, one may in Hilbert space use more general transformations than the affine transformations used so far. Thus, we may now use M6bius transformations in some applications (although not in all: e.g. Theorem (3.13c) does not hold for M6bius transformations, since convexity may not be preserved under such mappings). We shall begin by generalizing the well-known matrix result that a positive (negative) definite matrix has a positive (negative) definite inverse. Introduce the notation Yl = f (xO, Y2=f(x2), A x = x I - x 2, Af = f ( x l ) - f ( x2) and when applicable, A f -1 = f - l (y l ) - f - l (y 2) and A y = y 1 - Y 2. Note that in Hilbert space we may define m I f ] and M [ f ] as the infimum and supremum, respectively, of the expression
Re(Ax, Af) ]IAxLI2 xl'x2~D'

where the norm is induced by the inner product, i.e. I I A x l l 2 = ( A x , A x ) . Hence we have (4.1) (4.2)
m[f]

{]Ax]l2 ~ Re ( A x , A f ) < M [ f ]

]kJx]] 2.

Theorem. A s s u m e that M [ f ] < _ _ - 6 < 0 hold on the set f ( D ) :

on D. Then the f o l l o w i n g bounds

(4.3 a) (4.3 b)

M [ f - 1] < M [ f ] / L [ f ] 2 , m If-l] > m [f]/l[f]2.

38

G. S6derlind

Proof In the inequalities below, note that M [ f ]

and m [ f ] are negative numbers. The existence of f - 1 follows from the Uniform Monotonicity Theorem (Lemma (1.7)). By (4.1), Re ( A f -1, A y) <=M[f] IlAf -1 II2 < M [ f ] l [ f -1] z llA yll z - LMN f]] z iiAyll2, [[ which yields (4.3a). Similarly, by (4.1), Re ( A f -1, Ay) >re[f] IIA1-1 II2 > m [ f ] L [ f - ~ ] 2 HAy which gives (4.3b). (4.4) Remark. a) The bounds (4.3) are not necessarily sharp and should rather be considered as estimates of the lub and glb Dahlquist constants of the inverse operator. b) If the operator is uniformly positive monotone (i.e. m I f ] >6 > 0 on D), then (4.5 a) (4.5 b)

,2 m [ f ] = i[~

hlAybl2'

M I f - 1] < M [f]/l I f ] 2, m [ f -1 ] > m [ f ] / L I f ] 2.

Exact expressions concerning the monotonicity of inverse operators can be obtained by using circular conditions on f and M/Sbius transformations: Theorem. Assume that M[f]<__-6<0 on D. Let R* be the smallest value of R for which L [ f + RI] <=R, i.e. L [ f + R*I]=R*. Similarly, let l [ f +r*I] =r*. Then (4.6) (4.7a) (4.7b) M [ f -~] = - I/(2R*), m i f - l ] = _ 1/(2r*).

on the set f (D). Proof The condition L [ f + R * I ] = R * corresponds to the disk Iz+R*I<=R* in
the complex plane (see Fig. 5) which can also be written (4.8) Re _<0. z+2R* z

The M6bius transformation w = z-1 maps this domain onto the half-plane (4.9) Re w < - 1/(2R*).

The monotonicity condition corresponding to (4.8) and the assumption L [ f + R * I ] = R * is (4.10)

Re(Af, Af+2R*Ax)<O

xl,x2~D
is

with equality if we take the supremum of the left-hand side. Since f uniformly monotone, (4.10) is equivalent to

Re(Ay, A y + 2 R * A f - 1 ) < O

yl,y2~f(D).

Bounds on Nonlinear Operators in Finite-dimensional Banach Spaces

39

Q
w = z -I

(5)

-2R*

I 2r*

I 2R*

Fig. 5. Proof of Theorem (4.6): M6bius transformation w=z -1 maps circle-boundedset onto strip Hence, Re <Ay, A f - l > < - - - I l A y l [ = 2R*

Yl, Y2ef(D),

and (4.7a) follows from the basic inequality (4.1). The proof of (4.7b) goes along the same lines. (4.11) Remark. a) The positive number R* used in the assumption in Theorem (4.6) is given by the intersection between the r axis and the graph of L I f + r I ] - r in Fig. h b) For a positive monotone operator the assumptions would be L [ f - R * I ] = R * and l [ f - r * I ] =r*, respectively. The inverse f - 1 then has Dahlquist constants given by (4.12a) (4.12b) M If--1]
=

1/(2r*),

r e [ f - ' ] = 1/(2R*).

The following result is well-known for matrices, see e.g. [15, p. 442] and is also closely related to the spectral theory of yon Neumann [9]. (4.13) Theorem. Assume that m [ f ] < 0 (I-f)(D).

on D. Then L [ ( I + f ) o ( I - f ) - l ] < l

on

Proof We obviously have


Re <A (I + f ) + A (I - f ) , A (I + f ) or, equivalently,

A (I - f ) > < 4 M [ f ] ]1Ax I12

NA(I+f)I[ 2 - IIA(l-f)ll 2 < 4 M [ f ]

IIAx[I 2

x t, x2eD.

Since M [ f ] <0, I - f is uniformly positive monotone and has an inverse on (I - f ) ( D ) . Introducing x i = (I - f ) - 1(Yi), we have (4.14)

IIA((I+f)o(l-f)-l)H 2 - llAylk2 < 4 M [ f ]

I I A ( I - f ) -t Ih z

for all y~, y 2 e ( I - f ) ( D ) . Inserting M [ f ] < 0 in (4.14) shows that (I + f ) o ( I - f ) - t is a nonexpansive map on ( I - f ) ( D ) .

40

G. S6derlind

(4.15) Remark. a) It is of fundamental importance to note that Theorem (4.13) holds only for the operator ( I + f ) o ( I - f ) -1. The fact that f is negative monotone does not imply that ( I - f ) - l o ( I + f ) is nonexpansive. Unlike in the linear case, the two operators do not commute and are defined on different domains. b) A similar result for a positive monotone function is obtained by replacing f by - f in the theorem. Theorem (4.13) is the nonlinear functional analogue of the M6bius transformation w=(l+z)/(1-z) which maps the left half-plane R e z < 0 onto the unit disk Iwl<l. The inverse transformation is z=(w-l)/(w+l), and one may therefore expect that if g is a strict contraction, then ( g - I ) o ( g + I ) - ~ is strictly negative monotone. Quantitatively, we have: (4.16) (4.17 a) (4.17b) Theorem. Assume that L[g] < 1 on D. Then the following bounds hold on M [(g - I) o(g + I) -a ] < L [g] - 1 L [ g ] + 1'

the set (g+l)(D):

m[(g-I)o(g+I)

-1 > L [ g ] + I ] = L [ g ] _ 1"

Proof For notational convenience, set 0 = L [g]. By assumption,


IIh g IIz __<0 2 IIA x I]2. Hence

Re ( A ( g - I ) , A(g+I))<(OE-1)IMxH 2 Since 0 < 1, (g+l) -1 exists (cf. Corollary (2.3)) and

xl, x:~D.

L[(g+I)-l]<=l-O'
Thus, (4.18) Re <A ((g - I)o(g + I)- ~), Ay)

l[(g+I)-l]>=l+O"
1)IlA(g+I) -~ I]2,

~(0 2 --

where Yl,Y2e(g +I)(D). Since 0 2 - 1 < 0 we use the glb Lipschitz constant to obtain an upper bound of the right-hand side of (4.18):

Re ( d((g- I)o(g + I)-l),

Ay) < = ~

02--1

IIAyll 2,

from which (4.17a) follows by taking the supremum of the left-hand side. To prove (4.17b), note that lib (g-I)ll 2 - H A ( g + I ) H 2 < e M [ g ] IIAxll 2 < 4 0 HAxII2. Hence

[[d((g-I)o(g+I)-l)l[ 2 - Ildyll2<4OlId(g+I)-llb2<
-

4O

IldyH2.

Consequently,

HA((g-I)o(g+l)-X)l[Z<(l+ 4(1~00)2) [IAy 1ta,

Bounds on Nonlinear Operators in Finite-dimensional Banach Spaces

41

showing that L [ ( g - l ) o ( g + 1) -1] <=(1 + 0)/(1- 0). (4.t 7b) now follows from the general rule (cf. M 1) m I f ] __>- L I f ] . The essential conclusion in Theorem (4.16) is that we may find bounds on the lub and glb Dahlquist constants of ( g _ i ) o ( g + i ) - i by, respectively, maximizing and minimizing R e ( w - 1 ) / ( w + l ) over the set bwl<O in the complex plane. This is also true for the lub and glb Lipschitz constants, in which case one uses the extrema of I(w-1)/(w + 1)l over the same disk of radius 0. Thus, the disk Iw[<O is a spectral set in the sense of v o n N e u m a n n for the operator g and the M6bius transformation under consideration. Next, we would like to quantify Theorem (4.13) in a similar way. It is clear that circular domains corresponding to a condition on the lub Lipschitz constant or half-planes corresponding to monotonicity conditions will work as spectral sets for a particular M6bius transformation provided that its singular point is outside the spectral set. However, this is too crude for Theorem (4.13), since s u p l ( l + z ) / ( 1 - z ) l = l on the set {z6C: R e z < / ~ < 0 } irrespective of the actual value of ~. We therefore have to find a more elaborate spectral set in this case. Introduce the set (4.19)

F [ f ] = {z6C: tzl < L [ f ] , m [ f ] <- Re z < M [ f ] } ,

where, as usual, f is defined on D and the norm is induced by an inner product. We shall show that this may be used as a spectral set (in a generalized sense) for f provided that the singular point of the M6bius transformation under consideration is not contained in F [ f ] . The following is a quantitative version of Theorem (4.13). (4.20) (4.21) Theorem. Assume that M [ f ] < 1 - 3 < 1 on D. Then

L [ ( I + f ) o ( I - f ) - l ] < max l + z

on ( I - f ) ( O ) . Proof. We shall begin by considering the case 0 < M [ f ] < l . ing (4.14) and (2.11), we find
4 IlA((I + f ) o ( I - f ) - l ) l l 2< [1 q - ( l ~ )M)[ / ] 2 Then, by combin-

HAYI[2.

Simplifying the expression on the right-hand side shows that

L [ ( l + f ) o ( I _ f ) - l ] < 1+ M [ f ] =l-M[f]"
By the construction of F[f], this bound is identical to the bound given by (4.21). Next, consider the case M I f ] < - 6 < 0. Introduce the functional (4.22) 2 [ / ] = M I f ] + i(L [ / ]
2 __

m If]

2)1/2

42

G. S6derlind

Fig. 6. Proof of Theorem (4.20): |mage of spectral set F [ f ] under M6bius transformation

w=(1 + z)/(l-z) and note that I)~[f]l=L[f], R e ) ~ [ f ] = M [ f ] . Consider the circle

[z+rl2=r2-1,
which can also be written (4.23)

r>l

[zl 2 + 2r Re z + 1 = 0 .

N o t e that the M6bius t r a n s f o r m a t i o n w = (1 + z ) / ( 1 - z) maps this circle onto Iwl = 0 , where OZ=(r - 1)/(r + 1)< 1, see Fig. 6. N o w take r so that z = 2 [ f ] satisfies (4.23). Then (4.23) implies that L [ f + r l ] 2 <=r2 - 1, or (4.24)

IIAflL2+2r Re(Ax, A f ) +

IIAxH2 ~ 0 ,

xa,x26O

Evidently F [ f ] is contained in the circle (4.23) and m a p p e d into Iw[<O. Rewriting (4.24), we obtain

Re(A(I +f), A(l + f ) ) + R e ( A ( I - f ) , A ( I - f ) ) + 4 r Re (Ax, A f ) <O,


[IA(I q-f)ll 2 + [ I A ( l - f ) l [ 2 +r(l[A(I

+ f)[[ 2 - t[A(I-f) N2)<O.

Hence,

]lA(I-Ff)ll2<r~ 1 IIA(I-f)N 2
and it follows that

r-i

~a, x2~D,

NA((I+f)o(I-f)-a)ll2<O2llAylt 2

yt,yz~(I-f)(D).

Thus, L [(I + f ) o(I - f ) - ~] < 0, and by the construction of (4.23), max for z = 2[-f]. (4.25) Remark. N o t e that we have not m a d e use of m [ f ] (cf. (4.19)) in the proof. The reason for using it in the definition of F I f ] is that we m a y then apply the same technique to positive m o n o t o n e operators. Thus, if the singular l+z =0

zEF[fl 1 -- Z

Bounds on Nonlinear Operators in Finite-dimensional Banach Spaces

43

point would have been located to the left of F [ f ] , we would have defined 2 I f ] = m I f ] + i ( L [ f ] 2 - m [ f ] 2 ) 1/2. One should also note that the actual spectral set is not F [ - f ] , but rather the interior of the circle (4.23), with r chosen so that z = 2 I f ] satisfies (4.23). Theorems (4.16) and (4.20) can easily be extended to general M6bius transformations, if only some care is taken. Thus we m a y consider transformations of the form z+c~ (4.26) w = 7"-~,/3, 7~R _+z+/~ and the corresponding nonlinear m a p (4.27) g = 7" ( f + ~ I) o ( f + / 3 I ) - 1.

Provided that z=-T-~f~F[f], we m a y find bounds for L [ g ] and M [ g ] on ( f+[3I)(D) by maximizing Iwl and Re w, respectively over z~F [ f ] . The yon N e u m a n n spectral theory for linear operators is generally applicable to arbitrary rational transformations of the operator. A well-known result is that if a polynomial P satisfies I P ( z ) [ < l for ] z l < l , then ]bP(A)JI<I for all linear operators satisfying H < 1 with respect to a given inner p r o d u c t norm. Aq[ It remains an open question whether it is possible to generalize results of this type to nonlinear operators. However, it is clear that one has to impose extra conditions on the polynomials, e.g. a " c a n o n i c a l " representation. For, consider the polynomial z 2 - 1 = ( z + l ) ( z - 1 ) = ( z - 1 ) ( z + l ) and the corresponding nonlinear operators f o f - I, ( f + I) o ( f - I) and ( f - I) o ( f + I), respectively. Because of non-commutativity, these operators are different and cannot be expected to have the same bounds. M o r e advanced applications of the theory presented in this paper include the stability analysis of one-step [16] and multistep [4] time discretizations of ordinary differential equations. Strong results are already k n o w n for nonlinear systems, [4]. Recently, Nevanlinna [11, 12] has extended some earlier results [4, 5] for linear systems using operator theoretical techniques closely related to the work reported here. However, further research is still needed before our techniques can be directly used in connection with multistep discretizations.

Acknowledgements. This paper was written while the author was a visitor at the Computer Science
Department of Stanford University. The author is grateful to Professor Gene Golub for providing financial support and a stimulating research environment, and to the K. and A. Wallenberg foundation of the Royal Institute of Technology for a travelling grant.

References
0. Bauer, F.L., Stoer, J., Witzgall, C.: Absolute and monotonic norms. Numer. Math. 3, 257-264 (1961) 1. Bauer, F.L.: On the field of values subordinate to a norm. Numer. Math. 4, 103-113 (1962) 2. Browder, F.E.: The solvability of nonlinear functional equations. Duke Math. J. 30, 557-566 (1963) 3. Dahlquist, G.: Stability and error bounds in the numerical integration of ordinary differential equations. Trans. Royal Inst. of Tech., No 130, Stockholm 1959

44

G. S6derlind

4. Dahlquist, G.: G-stability is equivalent to A-stability. BIT 18, 384-40l (1978) 5. Dahlquist, G., S6derlind, G.: Some problems related to stiff nonlinear differential systems. In: Computing Methods in Applied Sciences and Engineering V. (R. Glowinski, J. Lions, eds.). Amsterdam: North-Holland 1982 6. Deimling, K.: Nonlinear functional analysis. Berlin, Heidelberg, New York: Springer 1985 7. Halmos, P.: A Hilbert space problem book. Princeton: van Nostrand 1967 8. Minty, G.J.: Monotone (non-linear) operators in Hilbert space. Duke Math. J. 29, 341-346 (1962) 9. von Neumann, J.: Eine Spektraltheorie ftir allgemeine Operatoren eines unit~iren Raumes. Math. Nachr. 4, 258-281 (1951) 10. Nevanlinna, O.: On the logarithmic norms of a matrix. Report-HTKK-MAT-A94, Helsinki Univ. of Tech., 1976 11. Nevanlinna, O.: Matrix valued versions of a result of von Neumann with an application to time discretization. Report-HTKK-MAT-A224, Helsinki Univ. of Tech., 1984 12. Nevanlinna, O.: Remarks on time discretization of contraction semigroups. In: Proc. Vlth Int'l Conf. on Trends in the Theory and Practice of Nonlinear Analysis (V. Lakshmikantham, ed.). Amsterdam: North Holland (to appear). Also available as: REPORT-HTKK-MAT-A225, Helsinki Univ. of Tech., 1984 13. Nirschl, N., Schneider, H.: The Bauer field of values of a matrix. Numer. Math. 6, 355-365 (1964) 14. Ortega, J.M., Rheinboldt, W.C.: Iterative solution of nonlinear equations in several variables. New York: Academic Press 1970 15. Riesz, F., Sz-Nagy, B.: Functional Analysis. New York: Ungar 1955 16. Schmitt, B.: Norm bounds for rational matrix functions. Numer. Math. 42, 379-389 (1983) 17. Stoer, J.: On the characterization of least upper bound norms in matrix space. Numer. Math. 6, 302-314 (1964) 18. Str6m, T.: On logarithmic norms. SIAM J. Num. Anal. 2, 741-753 (1975) 19. S/Sderlind, G.: On nonlinear difference and differential equations. BIT 24, 667-680 (1984) 20. S6derlind, G.: An error bound for fixed point iterations, ibid., 391-393 21. Zarantonello, E.H.: The closure of the numerical range contains the spectrum. Pac. J. Math. 22, 575-595 (1967) 22. Zenger, C.: On convexity properties of the Bauer field of values of a matrix. Numer. Math. 12, 96-105 (1968)

Received December 23, 1985/April 15, 1986

S-ar putea să vă placă și