Teoría acerca de la satisfacción

As a final example, consider the state space Rn equipped with a non- degenerate Gaussian measure instead of Lebesgue measure. Let γ be the centred Gaussian probability measure onRn with invertible covariance op- erator Γ : Rn _→ Rn; that is, γ is given in terms of its Radon–Nikod´ym derivative with respect to Lebesgue measure by

dγ dx(x) = 1 (2πdet Γ)n/2 exp −x·Γ −1_x 2 . (5.4.1)

Consider the trivial energetic potentialE(t, x) ≡0 and the dissipation po- tential Ψ(v) := 1_2|v_|2. Then PhX_i(₊₁P) _∈A X (P) i =xi i ∝ Z A exp − |∆xi+1|2 2ε∆ti+1 + x·Γ −1_x 2 dx.

Completing the square in the exponent yields that the law of X_i(₊₁P) given

X_i(P)is Gaussian with meanA−1_X(P)

i and covariance operatorεhA−1, where

A:= 1+εhΓ

−1_{. Up to leading order, this is the usual discretization of an}

Ornstein–Uhlenbeck process, and so it suggests the following (the proof of which will be omitted):

Theorem 5.4.2. Let X(P) be the Markov chain calculated with respect to the Gaussian measure γ (5.4.1), the trivial energetic potential E _≡ 0, the dissipation potential Ψ(v) := 1₂|v|2 _{and a partition P of [0, T]. Then, as} JPK_→0, X(P) converges in distribution to the solutionU of the Ornstein– Uhlenbeck equation

˙

U(t) =−εΓ−1U(t) +√εW˙ (t),

where W is a standard n-dimensional Wiener process.

As before, if the continuous stochastic interpolation ˜X(P) _{is used, then}

exists a constant C≥0 such that E " sup 0≤t≤T X˜(P)(t)−U(t) 2 # ≤CJPK.

As a concluding remark on directions for possible future research, note that theorem 5.4.2 suggests that the interior-point regularization and ther- malized gradient descent method may provide a tool for the approximation of stochastic partial differential equations. A plausible conjecture would be that the Markov chain on an infinite-dimensional separable Hilbert space

H with respect to the trivial energetic potential and the dissipation poten- tial Ψ(v) := 1_2kv_k2

H converges in a suitable sense to an infinite-dimensional

Ornstein–Uhlenbeck process, i.e. the solution to a stochastic heat equation.

5.5

Proofs and Supporting Results

Proof of lemma 5.3.4. Since the weak topology on the space of probabil- ity measures on a Polish space is metrizable, it suffices to check continuity in terms of sequential continuity. Suppose that µn is a sequence in P(X) such thatµn⇀ µ; it is required to show that Υµn⇀Υµ. Let φ: Y →Rbe bounded and continuous. Define Φ :_{X →}Rby

Φ(x) := 1 Z(x) Z Y φ(y)e−W(x,y)/εdπY(y), where, as usual, Z(x) := Z Y e−W(x,y)/εdπY(y).

Since, by assumption, the integral in the denominator is finite, Φ is bounded with_kΦ_k∞≤ kφk∞.

Since W _{is continuous, so are both (x, y) 7→ e}−W(x,y)/ε _{and (x, y) 7→}

φ(y)e−W_(x,y)/ε

; hence, these functions are measurable as well. The as- sumption in the statement of the lemma ensures that Lebesgue’s domi- nated convergence theorem in the form of theorem B.2 applies. Thus, both

Z:_{X →}(0,+_∞) and

x_7→

Z

Y

5.5. PROOFS AND SUPPORTING RESULTS 93 are continuous, and so Φ is continuous.

Hence, Z Y φd(Υµn) = Z Y φ(y) Z X e−W(x,y)/ε Z(x) dµn(x)dπY(y) by definition of Υ, = Z X Z Y φ(y)e−W(x,y)/ε Z(x) dπY(y)dµn(x) by Fubini’s theorem, = Z X Φ(x) dµn(x) by definition of Φ, −−−→_n→∞ Z X Φ(x) dµ(x) sinceµn⇀ µ, = Z Y φd(Υµ) reversing lines 1–3, and so Υµn⇀Υµ, as claimed.

Lemma 5.5.1. Let (_X, d) be a metric space and π a strictly positive and locally finite Borel measure on _X.

1. For everyx_{∈ X}, there exists a sequence of probability measures µnon

X such thatµnπ and µn⇀ δx.

2. If_X is separable, thenP_{π(X)is dense in}P_{(X)in the weak topology.} Proof. 1. Letµn be the probability measure defined by

µn(A) :=

π(A∩B_1/n(x))

π(B_1/n(x)) ;

sinceπ is strictly positive, the denominator is never 0, and sinceπ is locally finite, the numerator is finite for large enoughn. Let U be any open subset of_X. Ifx_∈U, then, for all large enoughn,B_1/n(x)_⊆U, and so µn(U) = 1. On the other hand, if x 6∈ U, then, for all large enough n,B_1/n(x)_∩U =_∅, and so µn(U) = 0. That is, for all large enoughn(depending on x), µn(U) =    0, ifx_6∈U, 1, ifx∈U. Hence, lim n→∞µn(U) = lim infn→∞ µn(U) =δx(U),

and the claim follows from the portmanteau theorem (theorem D.1). 2. LetD be any countable dense subset of X. By the previous part, for

every x _∈D, δx lies in the closure of Pπ(X). By theorem D.2, the closure of _{δx | x ∈ D} is P(X). Hence, the closure of Pπ(X) is

Chapter 6

Thermalized Gradient Descent

II: 1-Homogeneous Dissipation

6.1

Introductory Remarks

This chapter makes rigorous the heuristics of chapter 5 in the case that the state space is Rn, the energetic potential E is sufficiently well-behaved and the dissipation potential Ψ is homogeneous of degree one. That is, this chapter comprises the analysis of a rate-independent system

∂Ψ( ˙z(t))_{∋ −}DE(t, z(t))

in contact with a heat bath, where the effect of the heat bath is modeled by the interior-point regularization procedure. The main ingredients of this analysis were introduced in [SKTO09]. Of particular note is an effective dual dissipation potential, _F⋆

0, that controls the dynamics of the limiting

continuous-time evolution and is determined purely by the dissipation po- tential Ψ. _F0⋆ is a convex, extended-real-valued function defined on the dual spaceh6.1i of Rn:

F0⋆(w) := log

Z

Rn

exp(₋(_hw, z_i+ Ψ(z))) dz.

h6.1i_{In fact, in full generality,F}⋆

0 is defined on the cotangent bundle of the state space,

and its dual is defined on the tangent bundle. See the remarks on manifold-type state spaces in subsection 6.6.1 for a derivation that explains this.

The limiting evolution is a deterministic ordinary differential equation of the form

˙

x=−θDF0⋆(DE(t, x)).

However, it would be closer to the “gradient descent spirit” of this thesis to describe the limiting evolution as a gradient descent with a smooth, convex, nonlinear effective dissipation potential_F0 given by the convex conjugate of

F⋆

0:

D_F0(−x˙(t)/θ) = DE(t, x), (6.1.1)

where

F0(v) := sup{hℓ, vi − F0⋆(ℓ)|ℓ∈(Rn)∗}.

In most cases, Ψ and hence F⋆

0 and F0 are even, and so DF0 is odd, and

(6.1.1) can be rearranged into the more familiar form D_F0( ˙x(t)/θ) =−DE(t, x).

Many of the results of this chapter have been published in [SKTO09], although that work considered only the cases in which the Hessian of E

was either identically zero or a constant, symmetric, and positive-definite operator. The remarks on the mean-field approximation of the Koslowski– Cuiti˜no–Ortiz phase-field model in section 6.5 owe a great deal to discussions with the other authors of the joint paper [SKTO09], and their contributions are gratefully acknowledged.

6.2

Notation and Set-Up of the Problem

The following assumptions on the energetic and dissipation potentials will be hold for the rest of this chapter unless otherwise noted:

• The energetic potentialE: [0, T]_×Rn_→Ris assumed to be bounded below, smooth in space with all derivatives uniformly bounded, and such that (t, x) 7→DE(t, x) is uniformly Lipschitz. It is also assumed that E is convex, and hence that the Hessian of E is a non-negative operator.

• The dissipation potential Ψ = χ⋆

E: Rn → R is continuous, positive-

6.3. EFFECTIVE (DUAL) DISSIPATION POTENTIAL 97