Download - Numerical Solution of a Non-Smooth Eigenvalue Problem

Transcript
Page 1: Numerical Solution of a Non-Smooth Eigenvalue Problem

Numerical Solution of a Non-Smooth Eigenvalue Problem

An Operator-Splitting Approach

A. Caboussat & R. Glowinski

Page 2: Numerical Solution of a Non-Smooth Eigenvalue Problem

1. Formulation. Motivation

Our main objective is the numerical solution of the following problem from Calculus of Variations

Compute

γ = inf v Σ ∫Ω|v|dx (NSEVP)

where: Ω is a bounded domain of R2 and

Σ = {v| v H01(Ω), ∫Ω|v|2dx = 1}.

Page 3: Numerical Solution of a Non-Smooth Eigenvalue Problem

Actually, γ = 2√ π , independently of the shape and size

of Ω (holds even for non-simply connected Ω and in

fact for unbounded Ω) (G. Talenti).

A natural question is then:

Why solve numerically a problem whose exact solution is known ?

(i) If I claim that it is a new method to compute π nobody will believe me.

(ii) (NSEVP) is a fun problem to test solution methods for non-smooth & non-convex optimization problems.

Page 4: Numerical Solution of a Non-Smooth Eigenvalue Problem

(iii) ∫Ω|v|dx arises in a variety of problems from Image

Processing and Plasticity.

Actually, our motivation for investigating (NSEVP) arises from the following problem from visco-plasticity :

u L2(0,T; H01(Ω)) C0([0,T ]; L2(Ω)); u(0) = u0,

(BFP) ρ∫Ω(∂u/∂t)(t)(v – u(t))dx + μ∫Ωu(t).(v – u(t))dx +

g[ ∫Ω|v|dx – ∫Ω|u(t)|dx ] ≥ C(t)∫Ω(v – u(t))dx,

v H01(Ω), a.e. t (0, T),

with ρ > 0, μ > 0, g > 0, Ω a bounded domain of R2 and u0

L2(Ω).

Page 5: Numerical Solution of a Non-Smooth Eigenvalue Problem

(BFP) models the flow of a Bingham visco-plastic fluid in an infinitely long cylinder of cross section Ω, C being the pressure drop per unit length. Suppose that C = 0 and that T = +∞; we can show that

(C-O.PR) u(t) = 0, t ≥ Tc,

with

Tc = (ρ/μλ0)ln[1 + (μλ0/γg)||u0||L2(Ω)],

λ0 being the smallest eigenvalue of – 2 in H01(Ω).

A similar cut-off property holds if after space discretization we usethe backward Euler scheme for the time discretization of (BFP),

with λ0 and γ replaced by their discrete analogues λ0h and γh.

Page 6: Numerical Solution of a Non-Smooth Eigenvalue Problem

Suppose that the space discretization is achieved via C0-piecewise

linear finite element approximations, we have then

|λ0h – λ0| = O(h2).

But what can we say about |γh – γ| ?

The main goal of this lecture is to look for answers to the

above question !

Page 7: Numerical Solution of a Non-Smooth Eigenvalue Problem

2. Some regularization procedures There are several ways to approximate (NSEVP) – at the

continuous level – by a better posed and/or smoother

variational problem. The most obvious candidate is clearly

γε = inf v Σ ∫Ω(|v|2 + ε2)½dx, (NSEVP.1)ε

a regularization quite popular in Image Processing.

Assuming that the above problem has a minimizer uε, this

minimizer verifies the following Euler-Lagrange equation

(reminiscent of the mean curvature equation):

Page 8: Numerical Solution of a Non-Smooth Eigenvalue Problem

First regularized problem:

.1d||

,on 0,in ||

2

22

xu

uuu

u

(RP.1)

.

Page 9: Numerical Solution of a Non-Smooth Eigenvalue Problem

(RP.1) is clearly a nonlinear eigenvalue problem for a

close variant of the mean curvature operator, the eigenva

lue being γε.

Another regularization, more sophisticated in some sense,

since this time the regularized problem has minimizers, is

provided (with ε > 0) by

γε = min v Σ [ ½ ε∫Ω|v|2dx + ∫Ω|v|dx ]. (NSEVP.2)ε

An associated Euler-Lagrange (multivalued) equation

reads as follows, also of the nonlinear (in fact, non-

smooth) eigenvalue type (as above the eigenvalue is γε):

Page 10: Numerical Solution of a Non-Smooth Eigenvalue Problem

– ε2uε + ∂j(uε) γεuε in Ω,

(RP.2) uε = 0 on ∂Ω,

∫Ω|uε|2dx = 1;

in (RP.2), ∂j(uε) is the subgradient at uε of the functional

j : H01(Ω) → R defined by

j(v) = ∫Ω|v|dx.

The solution of problems such as (RP.2) is discussed in

GKM (2007); the method used in the above referenceis of the operator-splitting/inverse power method type.

Page 11: Numerical Solution of a Non-Smooth Eigenvalue Problem

In order to avoid handling simultaneously two small

parameters, namely ε and h, we will address the solution

of

γ = inf v Σ ∫Ω|v|dx

without using any regularization (unless we consider the

space approximation as a kind of regularization, that it is

indeed).

Page 12: Numerical Solution of a Non-Smooth Eigenvalue Problem

3. Finite Element Approximation(i) First, we introduce a family {Ωh}h of polygonal approxi-

mations of Ω, such that

limh→0 Ωh = Ω.

(ii) With each Ωh we associate a triangulation Th verifying

the usual assumptions of: (a) compatibility between triangles, and (b) regularity.

(iii) With each Th we associate the finite dimensional space

V0h defined (classically) as follows:

Page 13: Numerical Solution of a Non-Smooth Eigenvalue Problem

V0h = {v| v C0(Ωh∂Ωh), v|T P1, T Th,

v = 0 on ∂Ωh}.

(iv) We approximate

γ = inf v Σ ∫Ω|v|dx (NSEVP)

by

γh = min v Σh ∫Ωh |v|dx (NSEVP)h

Page 14: Numerical Solution of a Non-Smooth Eigenvalue Problem

with

Σh = {v| v V0h, ||v||L2(Ωh) = 1}. It is easy to prove that:

(i) Problem (NSEVP)h has a solution, that is there exists

uh Σh such that

∫Ωh |uh|dx = γh.

(ii) limh→0 γh = γ ( = 2√π).

We would like to investigate (computationally) the order

of the convergence of γh to γ. From the non-smoothness

of the problem, we do not expect O(h2).

Page 15: Numerical Solution of a Non-Smooth Eigenvalue Problem

4. An iterative method for the solution

of (NSEVP)h We are going to look for robustness, modularity and simplicity of programming instead of performance measured in number of elementary operations (this is not image processing and/or real time). At ADI 50 ( December

2005, at Rice University), we showed that the inverse power method for eigenvalue computations has an operator-splitting interpretation; we also showed the equivalence between some augmented Lagrangian algorithms and ADI methods such as Douglas-Rachford’s and Peaceman-Rachford’s. For problem

(NSEVP)h we think that it is simpler to take the AL approa-ch, keeping in mind that it will lead to a ‘disguised’ ADI method.

Page 16: Numerical Solution of a Non-Smooth Eigenvalue Problem

For formalism simplicity, we will use the continuous

problem notation. We observe that there is equivalence

between

γ = inf v Σ ∫Ω|v|dxand

γ = inf {v, q, z} E ∫Ω|q|dx,

where

E = {{v, q, z}| v H01(Ω), q (L2(Ω))2, z L2(Ω),

v – q = 0, v – z = 0, ||z||L2(Ω) = 1}.

Page 17: Numerical Solution of a Non-Smooth Eigenvalue Problem

The above equivalence suggests introducing the followingaugmented Lagrangian functional

Lr : (H01(Ω)×Q×L2(Ω))×(Q×L2(Ω)) → R

defined as follows, with Q = (L2(Ω))2 and r = {r1, r2}, ri > 0,

Lr(v, q, z; μ1, μ2) = ∫Ω|q|dx + ½ r1 ∫Ω|v – q|2dx

+ ½ r2 ∫Ω|v – z|2dx + ∫Ω(v – q).μ1dx

+ ∫Ω(v – z)μ2dx

Page 18: Numerical Solution of a Non-Smooth Eigenvalue Problem

We consider then, the following saddle-point problem

Find {{u, p, y}, {λ1, λ2}} (H01(Ω)×Q×S)×(Q×L2(Ω))

such that

Lr(u, p, y; μ1, μ2) ≤ Lr(u, p, y; λ1, λ 2) ≤ Lr(v, q, z; λ1, λ 2), (SDP-P) {{v, q, z}, {μ1, μ2}} (H0

1(Ω)×Q×S)×(Q×L2(Ω)),

with S = {z| z L2(Ω), ||z||L2(Ω) = 1}.

Suppose that the above saddle-point problem has a solution. We have then p = u, y = u, u being a minimizer for the original mimimization problem (the primal one).

Page 19: Numerical Solution of a Non-Smooth Eigenvalue Problem

To solve the above saddle-point problem, we advocate

the algorithm below (called ALG 2 by some practitioners

(BB)):

(1) {u –1, {λ10, λ2

0}} is given in H01(Ω)×(Q×L2(Ω));

for n ≥ 0, assuming that {un – 1, {λ1n, λ2

0}} is known,

solve:

(2) {pn, yn} = arg min{q, z} Q×S Lr(un – 1, q, z; λ1n, λ 2

n),

then

(3) un = arg minv Lr(v, pn, yn; λ1n, λ 2

n), v H01(Ω),

(4) λ1n+1 = λ1

n + r1(un – pn), λ2n+1 = λ2

n + r2(un – yn).

Page 20: Numerical Solution of a Non-Smooth Eigenvalue Problem

The above algorithm is easy to implement since:

(i) Problem (3) is equivalent to the following linear variational problem in H0

1(Ω)

un H01(Ω),

r1∫Ωun.v dx + r2 ∫Ωunv dx = ∫Ω(r1pn – λ1n ).v dx

+ ∫Ω(r2yn – λ2n )v dx, v H0

1(Ω).

The solution of the discrete analogue of the above

problem is a simple task nowadays.

Page 21: Numerical Solution of a Non-Smooth Eigenvalue Problem

(ii) Problem (2) decouples as

(a) pn = arg min q Q [½ r1 ∫Ω |q|2dx + ∫Ω|q|dx

– ∫Ω(r1un + λ1n).qdx ].

(b) yn = arg min z S [½ r2 ∫Ω |z|2dx – ∫Ω(r2un + λ2n)zdx ].

Both problems have closed form solutions; indeed, since

||z||L2(Ω) = 1, z S, one has

yn = (r2un + λ2n) / ||r2un + λ2

n ||L2(Ω).

Page 22: Numerical Solution of a Non-Smooth Eigenvalue Problem

Similarly, the minimization problem in (a) can be solved

point-wise (one such elementary problem for each triangle

of Th, in practice). We obtain then, a.e. on Ω,

pn(x) = (1/r1) (1 – 1/|Xn(x)|)+ Xn(x),

where

Xn(x) = r1un(x) + λ1n(x).

Page 23: Numerical Solution of a Non-Smooth Eigenvalue Problem

5. Numerical experimentsFirst Test Problem: Ω is the unit disk

Page 24: Numerical Solution of a Non-Smooth Eigenvalue Problem
Page 25: Numerical Solution of a Non-Smooth Eigenvalue Problem

Unit Disk Test Problem

Variation of γh versus h

Page 26: Numerical Solution of a Non-Smooth Eigenvalue Problem

Unit Disk Test Problem

Variation of γh – γ versus h

Page 27: Numerical Solution of a Non-Smooth Eigenvalue Problem

Unit Disk Test Problem

Visualisation of the coarse mesh solution

Page 28: Numerical Solution of a Non-Smooth Eigenvalue Problem

Unit Disk Test Problem

Visualisation of the fine mesh solution

Page 29: Numerical Solution of a Non-Smooth Eigenvalue Problem

Unit Disk Test Problem

Coarse mesh solution contours

Page 30: Numerical Solution of a Non-Smooth Eigenvalue Problem

Unit Disk Test Problem

Fine mesh solution contours

Page 31: Numerical Solution of a Non-Smooth Eigenvalue Problem

Unit Disk Test Problem

Fine mesh solution contours (details)

Page 32: Numerical Solution of a Non-Smooth Eigenvalue Problem

Second Test Problem: Ω is the unit square

Coarse mesh

Page 33: Numerical Solution of a Non-Smooth Eigenvalue Problem

Unit Square Test Problem

Fine mesh

Page 34: Numerical Solution of a Non-Smooth Eigenvalue Problem

Unit Square Test Problem

Variation of γh versus h

Page 35: Numerical Solution of a Non-Smooth Eigenvalue Problem

Unit Square Test Problem

Variation of γh – γ versus h

Page 36: Numerical Solution of a Non-Smooth Eigenvalue Problem

Unit Square Test Problem Visualisation of the coarse mesh solution

Page 37: Numerical Solution of a Non-Smooth Eigenvalue Problem

Unit Square Test Problem

Visualisation of the fine mesh solution

Page 38: Numerical Solution of a Non-Smooth Eigenvalue Problem

Unit Square Test Problem

Contours of the coarse mesh solution

Page 39: Numerical Solution of a Non-Smooth Eigenvalue Problem

Unit Square Test Problem

Contours of the fine mesh solution

Page 40: Numerical Solution of a Non-Smooth Eigenvalue Problem

Unit Square Test Problem

Contours of the fine mesh solution (details)

Page 41: Numerical Solution of a Non-Smooth Eigenvalue Problem

Circular Ring Test Problem (coarse mesh)

Page 42: Numerical Solution of a Non-Smooth Eigenvalue Problem

Circular Ring Test Problem (fine mesh)

Page 43: Numerical Solution of a Non-Smooth Eigenvalue Problem

A GENERALIZATION

Compute for Ω R2

γ* = infv ∫Ω |v|dx

with

= {v| v (H10(Ω))2, ||v||(L2(Ω))2 = 1}.

Page 44: Numerical Solution of a Non-Smooth Eigenvalue Problem

Conjecture (unless it is a classical result):

...04805.33

2sin1

2

13

2

0

d*γ

Page 45: Numerical Solution of a Non-Smooth Eigenvalue Problem

Square (coarse mesh)

Page 46: Numerical Solution of a Non-Smooth Eigenvalue Problem

Square (fine mesh)

Page 47: Numerical Solution of a Non-Smooth Eigenvalue Problem

Disk (coarse mesh)

Page 48: Numerical Solution of a Non-Smooth Eigenvalue Problem

Disk (fine mesh)

Page 49: Numerical Solution of a Non-Smooth Eigenvalue Problem

The results of our numerical computations

suggest very strongly that the value we conjectu-

red for γ* is the good one.

Page 50: Numerical Solution of a Non-Smooth Eigenvalue Problem

APPLICATION to a SEDIMENTATION PROBLEM

The following problem has been considered by C. Evans &

L. Prigozhin

u/t + IK(u) f in Ω × (0, T),

(SP)

u(0) = u0,

with Ω R2 and

K = {v | v H1(Ω), |v| C, v = g on Γ0 ( Ω)}.

Page 51: Numerical Solution of a Non-Smooth Eigenvalue Problem

After time-discretization by the backward Euler scheme, we

obtain

(1) u0 = u0 ;

n ≥ 1, un – 1 → un as follows

(2) un – un – 1 + IK(un) Δt f n.

“Equation” (2) is the Euler-Lagrange equation of the following

problem from Calculus of Variations:

(MP) un = arg minv K [ ½ Ωv2 dx – Ω(un – 1 + Δt f n)vdx].

Page 52: Numerical Solution of a Non-Smooth Eigenvalue Problem

The minimization problem (MP) is equivalent to:

{un, pn } =

arg min{v, q} K [½ Ωv2 dx – Ω(un – 1 + Δt f n)vdx],

with

K = {{v, q}| v H1(Ω), v = g on Γ0, |q| C, v – q = 0}.

We can compute {un, pn } via the following augmented

Lagrangian

Lr(v, q; μ) = ½ r Ω |v – q|2 dx + Ω μ.(v – q) dx

+ ½ Ωv2 dx – Ω(un – 1 + Δt f n)vdx.

Page 53: Numerical Solution of a Non-Smooth Eigenvalue Problem

River sand pile: FE mesh

Page 54: Numerical Solution of a Non-Smooth Eigenvalue Problem

River sand pile (2)

Page 55: Numerical Solution of a Non-Smooth Eigenvalue Problem

River sand pile (3)

Page 56: Numerical Solution of a Non-Smooth Eigenvalue Problem

River sand pile (4)

Page 57: Numerical Solution of a Non-Smooth Eigenvalue Problem

Rectangular pond sand pile (1)

Page 58: Numerical Solution of a Non-Smooth Eigenvalue Problem

Rectangular pond sand pile (2)

Page 59: Numerical Solution of a Non-Smooth Eigenvalue Problem