CAAM641 Spring 2007

The Eikonal equation k(x)k = n(x) : models, theory, simulation.

Chapter 0 :

http://www.tangentspace.net/cz/archives/2006/11/eikonal-equa

Where :
1

 When :

Homework - Grading ?

Outline :

Where does it come from ? What is it used for ?

Formal asymptotic solutions of Helmholtz

Eikonal and Ray tracing

Going behind caustics (Maslov theory)

Wigner Transform

GTD

Gaussian Beams
2

 Other examples of occurence of the Eikonal equation

What do we know about ?

Law of refraction : History - Fermat principle

Euler Lagrange and ray tracing (CV)

Lagrangian Eulerian : Classical solutions

Basics on Viscosity solutions

Optimal control interpretation of the viscosity solution

Explicit formula Lax-Oleinik

3

 in the (M ax, +) Algebra ...

Relationship with Hyperbolic Conservation Laws ...

How do we compute it ?

Quick recap on classic methods for ray tracing

The need for more : wavefront construction

Viscosity solutions : derivation of a simple 1-D Upwind scheme

Convergence theory

Rouy-Tourin scheme in 2-D

Fast algorithms (Fast marching, Sweeping ...)

4

 LAgrangian Schemes : Explicit formula Lax-Oleinik

Fast Legendre Transfoms

Algorithm in (M ax, +) Algebra ..

Helmholtz equation
(see also W. Symes seminar notes on time domain ... and
www.cscamm.umd.edu/programs/hfw05/runborg survey hfw05.pdf )

u(X) + k2n2(X)u(X) = 0
(local) .... to make it well posed add

a (possibly infinite) domain.

Radiation conditions or absorbing boundary conditions .

Source terms or boundary conditions (diffraction problem) .

5

Formal asymptotic solution

Debye expansion (1911) - Chap 1. Equations with Rapidly Oscillating solutions Partial Differential equations V , M.V. Fedoryuk
uk (X) = eik(X)

1
Aj (X)
j
j=0 (ik)

Apply H. operator and order in powers of k

uk + k2n2uk = k2(kk2 n2)+
ik(2 A0 + A0)+
(A0 + ....)+
...
FAS :

{ 2 + n2}(A0(X)eik(X)) = O(k1)
k
6

The Eikonal equation (EE) - Ray Tracing Solution (RTS)

k(X)k2 = n(X)2
Method of characteristics (assuming is smooth) :
d Y (s) = (Y (s)). Set P = (Y ) then
ds
dP
1
dY
= P (s),
= n2(Y (s))
ds
ds
2
The phase , can be computed as the integral of kP k2 along a
ray Y (s), since
d
dY
(Y (s)) =
(Y (s)) = kP (s)k2 = n2(Y (s))
ds
ds
Rem. 1 : All this is completely local ! BC IC ?
Rem. 2 : parameterization of paths can be changed (ex : dt =
n2(Y (s))ds is a possibility )
7

A word about Amplitudes

Setting B(s) = A(Y (s)) gives (TE) :
dY
d
B(s) =
A(Y (s)) = B(s) (Y (s))
ds
ds
... Instead remark that TE :
(A2) = 0
Integrate on a ray tube {Y (s, y0)| 0 < s < t | y0 B(y 0, )} and
use the divergence th. yields
B 2nJ |Y (s,y 0)= constant
(s,y0 )
where J = det( Y y
)
0

Ex. 1 : n = 1 plane wave solution

u(X) = AeikX
if kk = 1.
RTS :
Y (s, Y0) = Y0 + s, P (s, Y0) = , (Y (s, Y0)) = 0(Y0) + s
ICs on z = 0 : Y0 = (x, 0), P (0, Y0) = , 0(x, z = 0) = xx give
(X) = X

Physical Optics
For a diffracion pbm (u = Ceike x + uscat) us the integral formulation (u = 0 on the scatterer)
u(x) =

u 0
0
G(k|x x |) dx
n

and the geometric optic approximation u = Ceike x Ceike x to

replace by
u(x) =

Z
light

G(k|x x0|)(2Ciksin)dx0

10

Ex. 2 : Hu = x0 fundamental solution : Hankel function,

point source

u(X) = H01(k|X X0|)

As k + u(X) ' (

i 4 ik|XX0 |
2
)e
e
k|XX0 |

RTS :
Y (s, ) = X0 + s~e , P (s, ) = ~e , (Y (s, )) = 0() + s
ICs for [0, 2[ : Y (0, ) = X0, P (0, ) = ~e , 0() = 0 give
(X) = |X X0|
11

Asymptotic interpretation of the R.B.C.

The solution behave asymptotically (in kr) as
eikr
u ' A()
r
which is to say that far form source or scaterrer solution behaves
like geometric point source in homegenous space with amplitudes
modulated by .

12

Ex. 3 : A fold caustic case

gauche x < 0
n2(x, z) = 1
droite x [0, 1 ] n2(x, z) = 1 x , x [0, 1 ]
x>1
n2(x, z) = 

uinc = eik~eX

13

Separation of variables 1D
We can set u(z, x) = u
(x)eikz sin then
u
00g (x) + k2(cos2 )
ug (x) = 0

u
g (x) = eik cos x + Reikx cos

ud(x) = 0
u
00d (x) + k2(cos2 x)

u
d(x) =

2
CAi(k 3 (x))+
2
DBi(k 3 (x))

(x) = cos2 x

As k + : On the shadow side ((x) < 0 )

2
Ai(k 3 (x))

2
3
Bi(k (x))

So we take D = 0, determine R, C with the compatibility conditions at x = 0 and

14

on the lighted side ( (x) > 0 )

3
3
ik( 2 2 )
C
2)

i
i 4 ik( 2
3
e 3
+ e 4e
} + O(k1)
u
d = 1 {e
2 4

(Stationnary phase on

2
Ai(k 3 (x))

R ik((x) 3 )
3

d)

15

Next RTS
Only work in x [0, 1 ], n2(x) = 1 x, take Y (s, z0) =
(y(s, z0), z(s, z0)) and P = (py , pz ) we get :

y = py

z = pz

1
p

y
2

pz = 0

y(0, z0) = 0
z(0, z0) = z0
py (0, z0) = cos
pz (0, z0) = sin

= 1 y (0, z0) = sin z0

2
which yields Y (s, z0) = ( 1
4 s + s cos , s sin + z0 )

16

Use z as parameterization and finally

At (x, z) = Y (s, z0) 2 phases :
3/2 + 2 cos2
(x)
(x, z) = z sin 2
3
3

Away from a Fold caustic, stationary phase theorem gives

u = aei/4 eik + a+ei/4eik + O(k1)

17

18

Global Asymptotic solutions (Maslov)

Need to know the catastrophe which ruins the standard FAS.
Simplest case is the Fold caustic :
Consider (locally) the smooth manifold
M = {(y(z, z0), z, py (z, z0), pz (z, z0)) | z0 Z0, z Z}
then TH. (H
ormander, Duistermaat, ...) : (X, ), X = (x, z)
R such that

1. M = {(X, X ) | = 0, z Z}.
2. X M , (X) such that (X, (X)) = 0 and
19

2 (X, (X)) 6= 0.
d R
k
2
3. Use generalized ansatz u(X) = ( 2 )
A(X, , k)eik(X,)d
P
1 A (X, ) then away from
and expansion A(x, , k) =
j=0 (ik)j j

Caustic, stationnary phase Th. gives

u(X) =

A0(X, (X))

i/4 eik(X, (X)) + O(k 1 )

e
1

| (X, (X))| 2

superposition of simple AS. Computable by RTS (or in

this case with 2 Eikonal Equations).

4. 2 (X, (X)) = 0 indicate the caustic where the stationnary point is degenerate and give a contirbution in O(k).

Rem. 1 : Other option is find a change of variable such that

3
(X, ) = 0(X) + (X) 3 Airy function.
Rem. 2 : Other catastrophe in 2D is the cusp which leads to
Pearcey and 3 branches.

other Ex. (from Runborg)

20

Motivation 1 : Discretisation depend on k

Finite Element Solution of the Helmholtz Equation with High
Wave Number Part II: The h-p Version of the FEM
Frank Ihlenburg; Ivo Babuska SIAM Journal on Numerical Analysis, Vol. 34, No. 1. (Feb., 1997), pp. 315-358.
For degree-p Lagrangian FE in 1D
ku uhk C1(hk)p + C2k(hk)p+1
Rule of thumbs : hk < 1 Very Large systems to solve.
http://alphard.ethz.ch/hafner/Workshop/Hiptmair.pdf

21

22

Motivation 2 : Theory and practice of Migration Operators

Mathematical Theory For Seismic Migration and spatial
resolution. G. Beylkin
check W. Symes seminar notes
http://www.trip.caam.rice.edu/txt/tripinfo/other list.html

23

The Wigner Transform Approach

v(x, y) = u(x + 2 y)
u(x 2 y), (x, y) R2
R iy
f(x, ) = Fy v(x, y) = e
v(x, y)dy

Some properties
R
2
1
ku(x)k = v(x, 0) = {F f }(x, 0) = f (x, )ei0 d
v (x, 0) = {F 1 f }(x, 0) = R if (x, )ei0 d
Re{u(x)u(x)} = y

y

24

Looking at the Wigner transform of Assymptotic Solutions gives

some insight :
3 ))
ik(x)
2
ik(y(x)+O(
u = A(x) e
v(x, y) = A (x)e
+ O(2)

Taking  = k1 and neglecting high order terms we get :

R iy((x))
2
f (x, ) = A (x) e
dy = A2(x)( (x))

Rem. 1 : Same with sums of AS

25

The kinetic equation for

u(x) + 2n2(x)u(x) = (x)
Notice : y xv = 2 [u(x + 2 y)
u(x 2 y) u(x + 2 y)
u(x 2 y)]
and thus we have :
i [n2 (x +  y) n2 (x  y)]v (x, y) = (x, y)
iy xv(x, y) + 2


2
2

Fourier Transform and

xf + Z(x, ) ? f = Q(x, )

26

As  0
i


Fy [n2(x y) n2(x y)]
2
2
2
So formaly (Taylor expand the n2 functions above around x) :
Z(x, ) =

1
Z(x, ) 0 xn2(x) ()
2
Can also check that
Q(x, ) = Fy [(x, y)] 0 (x)(|| = 1)

27

Finally :
1
f (x, ) + xn2(x) f (x, ) = (x x0)(|| = 1)
2
which can be reformulated as a time dependent problem f (x, ) =
R

0 f (s, x, ) ds (particles go at infty with time) :

2

sf(s, x, ) + xf(s, x, ) + 1
2 x n (x) f (s, x, ) = 0,

f(0, x, ) = (x x0)(|| = 1)
Back to Ray tracing : Set
f0(s, x, ) =

P
(x Y (s, ))( P (s, ))

28

where {Y (s, ), P (s, )} satisfy the RT equations

dP
1
dY
= P (s),
= n2(Y (s))
ds
ds
2
with point source initialisation :
Y (0, ) = x0 , P (0, ) = (cos , sin )
satify the limit kinetic equation. Use a test function g(x, ) and
R
setting < f, g >= f g dx d. Then one can check formally that
(Y, P depend on s, )
P
dY
dP
< sf0, g >=
{ ds x g(Y, P ) + ds g(Y, P )}
P
1
2
=
{P x g(Y, P ) + 2 n (Y ) g(Y, P )}
1 n2 (x) g(x, ), f >
= < xg(x, ) + 2
0

1 n2 (x) f , g >
= < x f0 , g > < 2
0
29

Rem . : Morrey Campanato semi-norms involved in Helmholtz

case (see B. Perthame et al, High Freqency limit of the Helmholtz
equation. Rev. Iber. Americana ... )
Rem. : As  0 convergence is weak in S 0
Wigner Functions versus WKB-Methods in Multivalued Geometrical
Optics, Christof Sparber ? , Peter A. Markowich. y. , Norbert
J. Mauser

30

More : Level sets, Image processing PDE (second order

terms ...)
www.intlpress.com/CMS/issue4/levelset imaging chapter.pdf
www.math.uci.edu/~zhao/publication/publication.html

More : Shape From Shading

perception.inrialpes.fr/Publications/2006/PF06a/chapter-prados-faugeras.pdf

More : Geodesics Mesh Refinement, Creeping rays, RT in

Anisotropic media
www.nada.kth.se/~olofr/Publications/article-creep2-1.pdf
math.unice.fr/~rascle/psfiles/hr01.ps
31

Symes Quian elastic waves (See TRIP WWW)

More vaguely related but more Fashionable : BlackScholes
(finance) 2nd order terms...

Distance functions
d(x, x0) = |x x0| satisfies |xd| = 1.
Anticipating : (x) = minx0 d(x, x0) is the viscosity solution
of
|| = 1 in , = 0 on .

32

Level sets image processing

d(t,)
= v((t, )) and set (t, ) = {x, (x, t) = 0} then is
dt

solution of

+ v = 0
t

Ex. 1 : Normal motion v = n ||

. Note n may depend on x.

Ex. 2 : Motion by mean curvature (Evans and Spruck .... ) :

) ||
v = .( ||
Ex. 3 : Surface reconstruction : v = d, |d| = 1, d(x) =
0 f or x S
33

34

35

Shape From Shading

Lambertian scene hypothesis (L, light and n, normal to the surL n

face vectors) I(x1, x2) = R(n(x1, x2)) = cos(L, n) = |L|
|n|

36

Orthographic SFS with a far light source

L = (~l, ) is constant and such that |L| = 1
S : x = (x1, x2) (x, u(x)) is a the surface parameterization so
that n(x) = (u(x), 1).
u(x) ~l +

I(x) = q

1 + |u(x)|2

With L = (0, 0, 1) |u(x)|2 = I 21(x) 1

37

38

Mesh refinements
Now use a Riemannian metric G. Note : length of dx is defined
1
2
as |dx|G(x) = |G dx|. If G = diag(lx, ly) and lx 6= ly anisotropy.
Again : (x) = minx0 dG(x, x0) is the viscosity solution of
1
2
|G |

= 1 in , = 0 on.

Now define G using a surface (x, z(x)) : G(x) = Id + z z or

Id + H(z)tH(z).

39

40

41

Law of refraction : Small experiment - History/Motivation

Minimizing travel time ( =

dist/speed) in a two layered medium.

sin 1
sin 2
=
c1
c2
History (Historia Mathematica 10 (1983) 48-62 The Mathematical Technique in Fermats deduction of the Law of refraction.
K. Andersen. )

1637 Descartes Dioptrique Assuming light goes faster in denser

media.

42

 1657 (letter to C. De la Chambre) principle of least time.

application of his method to find minima : Methodus ad
disquirendam et minimam Fermat. adequation
f (a + e) ' f (a)
then simplify, divide by e and finally equate. he obtained
expressions with an irregular and fantastic proportion after long and tedious calculations and his natural inclination
towards indolence had made him stop.

1662, completes computations. Finds Descartess law !

(Differential Calculus) Leibnitz (1684) and Newtown (1687)

43

The optical length of a Path

Fermat principle says Rays are the path Y that minimize (actually
extremize) traveltime
(Y (T )) =

Z T
0

n(Y (s))|Y (s)|ds + (Y (0))

(n = 1c )
Simplify to 1+1 D problem (time + space) : Y (s) = (t, y(t))
(paraxial assumption)

44

Fix t, x,
min

{yC , y0 R y(t)=x}

F (y, y0) =

Z t
0

L(s, y(s), y(s))ds

+ 0(y0)

with L(t, y, v) = n(t, y(t)) 1 + v 2(t).

Classical problem (L.C. Young, Lecture on the Calculus of Variations ) important hypothesis : strict convexity of v L(t, y, v).
Existence : (hints)

45

First variation vanishes

Euler-Lagrange equations :
the Frechet derivative satisfy F 0(y, y0) = 0

Rt
h(s)) L(s, y(s), y(s))ds
F (y + h, y0 + h0 ) F (y, y0 ) = 0 L(s, y + h(s), y +

+ 0 (y0 +
Rt

= 0 Lx (s, y(s), y(s))h(s)

+ Lv (s, y(s), y(s))

h(s)ds
+ O(|h|2 ) + 0x (y0 )h0 + O(h20 )

|h| = sups[0,t] {|h(s)| + |h(s)|}

=

d L (s, y(s), y(s))}h(s)ds

{L
(s,
y(s),
y(s))

+ O(|h|2 ) + ...
x
0
ds v

Rt

{Lv (0, y(0), y(0))

0x (y0 )}h0 + O(h20 ).

46

Rt
F
d L (s, y(s), y(s))}h(s)ds
< y (y0, y), h >= 0{Lx(s, y(s), y(s))

ds

v
F (y , y) h = L (0, y(0), y(0))
0 (y )

v
0
0
x 0
y0

d L (s, y(s), y(s))

Lx(s, y(s), y(s))

ds

=0
v

L (0, y(0), y(0))

0
v
x (y0 ) = 0.

47

Recap. on the Legendre Fenchel transform (Lvv > 0)

Define
H(t, x, p) = L?vp(t, x, p) = sup {p.v L(t, x, v)}
vRd

Then L(t, x, v) = L?? = suppRd {p.v H(t, x, p)}

and Hp = (Lv )1 i.e.
v = Hp(t, x, p) p = Lv (t, x, v)
and are the optimal args, and for (v, p) linked as above we have
L(t, x, v) = Lv (t, x, v).vH(t, x, Lv (t, x, v)) = p.Hp(t, x, p)H(t, x, p))
and
Lx(t, x, v) = Hx(t, x, p)
48

Examples

L(v) =

1 2
1
v H(p) = p2
2
2

(Burgers equation)
q

L(t, x, v) = n(t, x) 1 + v 2 H(t, x, p) = n2(t, x) p2

(Eikonal equation)

49

The Hamiltonian system

Set
p(s) = Lv (s, y(s), y(s)),

which automatically gives

y(s)

= Hp(s, y(s), p(s)), y(0) = y0 (by construction).

take the time derivative of p
p(s)

= Hx(s, y(s), p(s)), p(0) = 0

x (y0 ), (Euler Lagrange).
R
also remark that (y(t)) = 0t L(s, y(s), y(s))ds

+ (y0) satisfies

(s)
= Hp(t, y(s), p(s)).p H(s, y(s), p(s)), (0) = 0(y0)

50

Back to RT
In practice y, p, as functions of (t, y0)
q

Apply H(t, x, p) = n2(t, x) p2

y(s)

2 p2

p(s)

= n2xn 2
n p

(s)
=
n2 p2

Setting q =

y(0, y0) = y0
p(0, y0) = 0
x (y0 )
(0, z0) = 0(y0)

n2 p2 and z(s)

= q and then change parameq

terization dz = ds n2 p2 to recover our original ray tracing

equations for {Y = (y, z), P = (p, q)}.
51

Lagrangian Eulerian : Classical solutions

Consider T = {y(t, y0), (t, y0) [0, T ] Rd}.
0 ) ) 6= 0. One can invert
As long as J(t, y0) = det( y(t,y
y0

y0 x = y(t, y0)
Then define (t, x) as
(t, y(t, y0)) = (t, y0).
is smooth and
x(t, y(t, y0)) = p(t, y0).
52

The Hamilton-Jacobi equation

Time differentiate :
t(t, y(t, y0)) + Hp(t, y(t, y0), p(t, y0)) x(t, y(t, y0)) =
p(s, y0) Hp(s, y(s, y0), p(s, y0)) H(s, y(s, y0), p(s, y0)),
Write in Eulerian coordinates

y(t, y0) x

p(t, y0) = x(t, y(t, y0)) x(t, x)

53

We get
t(t, x) + H(t, x, x(t, x)) = 0, (0, x) = 0(x).
q

Rem : t > 0 and H = n2 p2 (paraxial hypothesis), we find

the original 2-D the Eikonal equation.

54

Classical solutions break down at caustics

Looking around a Caustic point : {y c = y(yoc ), pc = p(yoc )}.
Fold :

y = 0
o
2

y 6= 0
2 yo

Then (TE) :
(yo yoc )2 2y c
c
c )3 }
(y
)
+
O{(y

y
y(yo) = y + 0 +
o
o
o
2
2 yo

55

Cusp :

y = 0
yo
2y

2 =0
yo

Then :
(yo yoc )3 3y c
c )4 }
c
(y
)
+
O{(y

y
y(yo) = y + 0 + 0 +
o
o
o
6
3 yo

56

An example :
n(t, x) = 2.8 if |x 0.2 sin 0.3 t| 1 and n(t, x) = 2.8 + 0.4
2
exp10 (x0.2 sin 0.3 t) else,
o 0 as initial condition.

57

58

Basics on viscosity solutions

G. Barles, Solutions de viscosit
es des Equ. de H,-J., Springer
1994
P.L. Lions, Generalized solutions of H.-J. Equations, Pitmann
Some good ressources online , in French though ... .
www.math.psu.edu/bressan/PSPDF/simy.pdf
Has evolved into a very general theory of existence and approximation for non-linear equations.
F (x, u(x), Du(x), D2u(x)) = 0
F (x, u, P, M ) Needs :
- Uniform continuity. - (Strict) convexity in p.
- Coercivity in p. - Ellipticity in M .
59

Consider simplified Pbm

u + H(x, Du) = 0 x Rd

(same with Dirichlet/Neumann or Cauchy problems).

Defs :

u(x) C(Rd) is a viscosity sub(super)-solution if C 1

x0 is a point of max. (min.) of u
u + H(x, D) () 0
u is a viscosity solution iff sub- and super-solution.
60

First Remarks

One can restrict to s.t. u(x0) = (x0).

One can restrict to strict min/max assumption : = +

|x x0|2.

Classical solutions are viscosity solutions.

61

Exemple (1-d)
|u0| = 1, u(0) = u(1) = 0 H(u0) = |u0| 1
Has no classical solutions.
Has many generalized W 1, solutions.
Has a unique BUC viscosity solution.
Viscosity solution allows upward (but not downward) kinks :

62

Properties
maximum principle...
Prop. (Unicity) : Let u (resp. u) be a l.s.c. (resp. u.s.c) super(resp. sub-) viscosity solution of then u u.
Evanescent viscosity : > 0,


u + u + H(x, Du) = 0 x

has a unique C 2 W 1, solution u which converges to the viscosity solution ( = 0).

(Sketch of the proof).
63

Back to the Calculus of Variation and Cauchy Pbm

sider

Con-

tu + H(x, Du) = 0 t > 0, xRd

u(0, x) = u0(x)

Set L = H ? (see recap. on Legendre Transform) and

v(t, x) = inf (u0(y0) +
y0

inf
W 1,(0, t)
(t) = x
(0) = y0

Z t
0

L(t, , )dt)

Then v is the unique BUC and Lipschitz viscosity solution of the

HJ equ.

64

Sketch

Inf is reached by C 2, ||
Ct .

v super-solution (inf prop.)

v sub solution (pontryagin).

Lipschitz.

+ Du = Lv
65

Intro. to explicit formula (Lax-Oleinik

Let H(p) be convex in and depending only on p.
L = H ?, then the viscosity solution of

Remember

tu + H(ux) = 0 , u(0, x) = u0(x)

can be written as (backward parameterization)
u(t, x) = inf {u0((t)) +
v

Z t
0

H ?(v)ds}

where = v, (0) = x. Then use Jensen inequality

Z t

1 t
x (t)
?
?
?
H (v)ds tH (
vds) = tH (
)
t 0
t
0
Z

66

i.e. straight lines from x y = (t) are optimal. We can rewrite

the viscosity solution as :
u(t, x) = inf {u0(y) + tH ?(
y

xy
)ds}
t

(Lax Oleinik formula).

q

For ex. H(p) = 1 p2 H ?(v) =

funtion from the Cauchy data.

1 + v 2 gives the distance

67

An example n 1 :

68

An example n 1 :

69

Link with conservation laws in 1D. Setting formally u = x

yields
t + H(x) = 0 tu + (H(u))x = 0
|p|2
Ex. : Burgers H(p) = 2

Rem. 1 : H.-J. is similar to Homegenous GO.

Rem. 2 : Inf principle of the Viscosity solution is equivalent to
u+ +u
R.H. condition. Shock speed : s =
2

70

Back to RT (numerics)

dY
ds =

dP =
ds

P (s)
1 n2 (Y (s))
2

The phase , can be computed as the integral of kP k2 along a

ray Y (s), since
d
dY
(Y (s)) =
.(Y (s)) = kP (s)k2 = n2(Y (s))
ds
ds
In practice solve for a family of rays parameterized by Y0 or .
Generally with a RK or adapative RK method.
note : should say a word on symplectic solvers...
71

Does it work ? : RT in Marmousi

http://www-rocq.inria.fr/ benamou/testproblem.html

72

Dynamical (or Paraxial Ray tracing)

Ray methods in Seismology, Cerveny, Molotkov, Pscencik, Charles
U. Praha (77) ....
G. Lambar
e 2002 GO++ winter school notes
Idea : Linearize RTS (and change notations :(x, p) : first order
variation in (x, p) = (Y, P ))

dx
ds = p

dp = ( 1 n2(x))x
ds
2

73

Propagator matrix
Because this is linear in the initial conditions, one can factor
various computations in a propagator matrix
(x, p)(s) = (x, p)(s0)P(s, s0)
(transpose)
dP
=
ds

0
Id
2 (x)) 0
( 1
n
2

P, P(s0, s0) = Id

Rem. 1 : this can be computed very accurately along each ray.

Rem 2. : this provide a 2nd order accurate estimate of the
Lagrangian submanifold {(x(s, y0), p(s, y0)) | y0 ...}.
x(y0 + y0, p0 + p0) = x(y0, p0) + ....
74

Also give second order derivatives of travel time

ray coordinates ...

... discuss

75

Wavefront construction
Main idea : maintain ray density through interpolation (adaptative gridding). Need a good criterium.
Problem : stretching or concentration may also may also occur
in the p dimensions.

76

77

Sol 1 : Vinje ... x distance. Problem stretching or concentration

may also may also occur in the p dimensions.
Sol 2 : Sun ... use x and p distance.

Sol 3 : Lambar
e et al ... Use paraxial quantities to estimate
error.

78

79

Intro to Eulerian FD upwinding

Consider the simplest non trivial HJ equ. (c > 0)
ut + cux = 0
i.e. H(p) = cp (note L(v) = 0 for v = c and else) with explicit
solution u(t, x) = u0(x ct). Then discretize : ukj = u(k dt, j dx)
and upwind. Two solutions

ukj+1 ukj

k+1
k

= uj + dt dx
uj

ukj ukj1

k+1
k
u
= uj + dt dx
j

80

Remarks

Right upwinding does not converge (try u0 = 0 for x > 0 and

6= 0 for x < 0.
dt . consider sequences s.t.
Numerical propagation speed is dx
(k dt, j dx) (t, x) as (dt, dx) 0 and define

dx

c = lim sup
dt
Then no convergence (wrong dependance) if
c < c. This is
the sense of the CFL condition :
dt
c
1
dx
81

 Left Upwinding can be rewritten

dt k
dt k
k+1
uj
= (1 c )uj + c uj1
dx
dx
Scheme is monotone (max principle) under CFL.

Back To H.J in 1+1 D

q

Take (for instance) H(p) = 1 p2 and look a a time step

[tj , tj+1] around xi.
Assume further that the phase is piewise linear with left and
right slopes x.
Let y(t) be a ray beween (tj , x) and (tj+1, xi) such that x =
y(tj ) < xi.

82

Integrate the Lagrangian phase (Riemman problem)

R tj+1
p(s)Hp(s, y(s), p(s)) H(p(s))ds
(tj+1, xi) = (tj+1) = (tj ) + tj
= (tj ) + x(xi y(tj )) dtH(x)

= (tj , xi) dtH(x)

Looks like a FD scheme for the H.-J. equ.
j+1

= i dtH(x)

83

Upwinding :

A.
B.
C.
D.

x 0,
x 0,
x < 0,
x < 0,

Minimize traveltime ...

x+ 0.
x+ < 0.
x+ 0.
x+ < 0.
84

the Godunov scheme

u (t)) = 0, (0) = 0 (x ).
ti(t) = H(t, xi, x,i
i
i

u
l
+
l

x,i(t) = max((x,i(t)) , (x,i(t)) )

(t) i(t)

l (t) = i(t) i1(t) , r (t) = i+1

.
x,i
x,i
dx
dx

(p)+ = max(p, 0) and p = min(p, 0).

85

Note on B.C. I.C CFL bound on Hp ....

86

Generalization 1. (convex H with min at p = 0)

Recall H(t, x, p) = supv {p.v L(t, x, v)} and L = H ? also convex
with min at v = 0 (Hp = (lv )1) .
Remark that (omit (t, x)) H(p) = max(H +(p), H (p)) with

H +(p) = supv0{p.v L(v)} = H(p+)

H (p) = supv0{p.v L(v)} = H(p)

H.-J. becomes (upwinding)

l (t)), H ( r (t)))
ti(t) = max(H +(x,i
x,i
87

Generalisation 2 (Godunov numerical Hamiltonian)

H G(x, x+) = Ext

pI(x ,x+ )

H(p)

where

mina<p<b if a b
ExtpI(a,b) =
max
b<p<a if a > b

88

Lax-Friedrich

(t), x,i
(t)) = H(t, xi,
H LF (t, xi, x,i

x,i
(t) + x,i
(t)

x,i
(t) x,i
(t)

dx

chosen to maintain monotonicity (derivative of H).

89

Higher Order (ENO, WENO ...), time integration ...

90

Recall ....
Recall H(t, x, p) = supv {p.v L(t, x, v)} and L = H ?.
Remark that (omit (t, x)) H(p) = max(H +(p), H (p)) with

H +(p) = supv0{p.v L(v)}

H (p) = supv0{p.v L(v)}

H.-J. becomes (upwinding)

ti(t) = max{H +(x,i

(t)), H (x,i
(t))}

91

Time discretization (.k )

ik+1 = ik dt max{H +(tk , xi, Dik ), H (tk , xi, D+ik )}

where :

k
ik i1
l
k
D i =
dx

k k
i+1
i
r
k
D i =
dx

H +(p) = supv0{p.v L(v)} H (p) = supv0{p.v L(v)}

92

Rewrite as :
ik+1 = min{U ik , U +ik }
where :

dt
dt v k
k

U ik = inf v0{dt H ?(tk , xi, v) + dx

i1 + (1 dx v) i }

dt v) k }
? (tk , x , v) dt v k
U + k = inf
{dt
H
+
(1
+
i
v0
i
i
i+1
dx
dx

Important remark : can restrain to |v| V = (kHpk). The CFL

condition can be written :
dt
V
1
dx

93

Fondamental properties of the numerical scheme

Consistence : For a smooth function , looking at a sequence

s.t. (k dt, j dx) (t, x) as (dt, dx) 0 and setting ik =
(tk , xi), we get
lim{ik+1 min(U ik , U +ik )} = t(t, x) H(t, x, x(t, x))

94

 Monotone : Check that (ik+1)i is a monotone (increasing)

function of (ik )i.
dt v > 0, CFL condition is enough for U .
Just need 1 dx

95

 Stability : (under CFL) the discrete solution is uniformly

bounded (independently of (dt, dx).)
kU ik | dtkH ?k + Ckik k
so
kik+1k dtkH ?k + Ckik k
is enough for finite time horizon.

96

Outline of convergence proof.

1. Using stability define :

(t, x) = lim(k dt,i dx)(t,x) sup ik
(t, x) = lim(k dt,i dx)(t,x) inf ik
by construction .

2. Show that (monotonicity + consistence) is an upper semi

continuous viscosity sub-solution.
Show that (monotonicity + consistence) is a lower semi
continuous viscosity super-solution.

3. Then strong uniqueness garantees : .

97

Back to full 2-D (x, y)

E. Rouy and A. Tourin. A viscosity solutions approach to shape-fr

SIAM J. Numer. Anal. 3 (1992) 867--884.
The Eikonal equation
k(X)k2 = n(X)2 X = (x, y)
can be written (optimal control formulation)
sup {(x, y).Q n(x, y)} = 0.
kQk1

so can also use

Sup is reached for q = ||

sup {(x, y).Q n(x, y)} = 0.

kQk=1
98

Lagrangian discretisation
((x, y) dt Q) (x, y)
sup {
n(x, y)} = 0
dt
kQk1
Rem. : This is first order (if the solution is smooth !)
((x, y) dt Q) (x, y) = dt (x, y).Q + O(dt2)

Rem. : this is upwinding

(x, y) =

inf {((x, y) dt Q)} + dt n(x, y)

kQk1

Recall
Y = = Qoptkk
99

Grid interpolation
Grid : (xi, yj ) = (i dx, j dy). Note (xi, yj ) = ij and n(xi, yj ) =
nij
First assume that (x, y)dt q) T = {(xi, yj ), (xi+1, yj ), (xi, yj1)}
(restrict Q to point in one of the quadrant).
using a convex linear combination of the value at grid points :
((x, y) dt q) = ij + i+1j + ij+1
where , and are such that
(

(x, y) dt q = (xi, yj ) + (xi+1, yj ) + (xi+1, yj+1))

+ + = 1.

Rem : Again first order approximation.

100

The optimization problem now depends on (, , ). It can be

worked out for the four quadrants and taking dt = dx2 dy 2
dx +dy

simplifies into the discrete Hamiltonian :

gij (ij , i+1j , ij+1, i1j , ij1) = 0, (i, j)
where
gij (ij , i+1j , ij+1, i1j , ij1) =
q

max(a+, b)2 + max(c+, d)2 n(xi, yj ),

a+ = max(0, a), b = max(0, b) and

a = Dxij =

ij i1j
dx

c = Dyij = ij dyij1

b = Dx+ij =

i+1j ij
dx

ij+1 ij
+
d = Dy ij =
dy
101

Relaxation
S. J. Osher and L. Rudin
Rapid convergence of approximate solution to shape form shading
problem.
Never Published
gij (ij , i+1j , ij+1, i1j , ij1) = 0, (i, j)
is a system of nonlinear equations ... may be difficult to solve
directly. Instead is is easy to prove that the algorithm obtained
by the following relaxation scheme converges :
n)
Compute a sequence of (ij
i,j solutions of
n , n1 , n1 , n1 , n1 ) = 0, (i, j)
gij (ij
i+1j ij+1 i1j ij1

Note this is like reintroducing time ...

102

On the positive quadrant

n (h = dx = dy)
we have to solve for t = ij

n1
n1
((i+1j
t))2 + ((ij+1
t))2 h nij = 0

Solution is explicitely given as

n+1
ij

n1
0.5 (ij+1

n1
i+1j

q
n1
n1
n1
n1 2
i+1j
| < h nij
i+1j
) if |i+1j
+ 2h2 n2ij (i+1j

n+1
n1
n1
ij
= min(i+1j
, i+1j
) + h nij

else

Extend to all four quadrant by seleecting min...

103

Slightly faster way is to revisit as P 1 interpolation on each triangle made by the 5 point stencil.
((x, y) dt (qx, qy)) = ij dt qx (Dxij ) dt qy (Dy ij )
n , D , D ) and 2 equations ((qx, qy) =
Three unknown (ij
x ij
y ij
(h, 0), (qx, qy) = (0, h))
n1
n hD )
ij+1
= ij
y ij
n1
n hD )
i+1j
= ij
x ij
then close with the Eikonal equation ..

k(X)k2 = n(X)2 X = (x, y)

+ 2
+ 2
(Dx+ij
) + (Dy+ij
) nij = 0

Extension to all four quadrant yield the scheme

104

Fondamental properties of the numerical (implicit!) scheme

Consistance : For a smooth function , looking at a sequence

s.t. (i dx, j dy) (x, y) as (dy, dx) 0 and setting ij =
(xi, yj ), we get
lim gij (ij , i+1j , ij+1, i1j , ij1) = k(x, y)k n(x, y)

105

 Monotonicity : if (uij )ij < (vij )ij then for all t and all i, j we
have
gij (t, ui+1j , uij+1, ui1j , uij1) < gij (t, vi+1j , vij+1, vi1j , vij1)
(Upwind scheme acts as a time dependent equation in the
ray direction).

106

 Stability : unconditionnaly stable (implicit !)

107

Relaxation and convergence proof

108

http://www.levelset.com/system/html/modules/sections/index.php?op=v
Interview with Stanley Osher at National University of Singapore :
Stanley Osher is an extraordinary mathematician who has ....

It turned out that I had a friend who knew the District Attorney or
was then doing video image enhancement with my colleague L. Rudin,
....
I: You could be rich. Hollywood would be paying you millions.

O: People work for salaries. There is money, ego and fun. Its a ve

109