The Moreau-Yosida Regularization: Difference between revisions
(added portmanteau) |
|||
Line 1: | Line 1: | ||
The '''Moreau-Yosida regularization''' is a technique used to approximate lower semicontinuous functions by Lipschitz functions. The main application of this result is to prove Portmanteau's Theorem, which states that integration against a lower semicontinuous and bounded below function is lower semicontinuous with respect to the narrow topology on the space of probability measures. | |||
==Definitions== | ==Definitions== | ||
Line 11: | Line 7: | ||
<math>g_k(x) := \inf\limits_{y \in X} \left[ g(y) + k d(x,y) \right].</math> | <math>g_k(x) := \inf\limits_{y \in X} \left[ g(y) + k d(x,y) \right].</math> | ||
The distance term <math>d(x,y)</math> may often be raised to a positive exponent. For example, when <math>X</math> is a Hilbert space <ref name="BC"/> <ref name="AGS"/>, <math>g_k</math> is taken to be | |||
<math>g_k(x) := \inf\limits_{y \in X} \left[ g(y) + k \| x - y \|^2 \right].</math> | |||
Note that | Note that | ||
Line 27: | Line 27: | ||
[[File:Ex 1.png|300px|thumb|Plot of <math>g(x) = x^2</math> and <math>g_k(x)</math> for <math>k = 0, 1, 2, 3</math>.]] | [[File:Ex 1.png|300px|thumb|Plot of <math>g(x) = x^2</math> and <math>g_k(x)</math> for <math>k = 0, 1, 2, 3</math>.]] | ||
== | ==Approximating Lower Semicontinuous Functions by Lipschitz Functions== | ||
'''Proposition.''' <ref name="OT"/><ref name="S"/> | '''Proposition.''' <ref name="OT"/><ref name="S"/> | ||
* If <math>g</math> is proper and bounded below, so is <math>g_k</math>. Furthermore, <math>g_k</math> is continuous for all <math>k \geq 0</math>. | * If <math>g</math> is proper and bounded below, so is <math>g_k</math>. Furthermore, <math>g_k</math> is continuous for all <math>k \geq 0</math>. | ||
Line 53: | Line 53: | ||
* By definition, <math>g_k \wedge k \in C_b(X)</math>. Since <math>g_k(x) \nearrow g(x)</math> for all <math>x \in X</math>, <math>g_k(x) \wedge k \nearrow g(x)</math> for all <math>x \in X</math>. | * By definition, <math>g_k \wedge k \in C_b(X)</math>. Since <math>g_k(x) \nearrow g(x)</math> for all <math>x \in X</math>, <math>g_k(x) \wedge k \nearrow g(x)</math> for all <math>x \in X</math>. | ||
==Portmanteau Theorem== | |||
'''Theorem (Portmanteau).''' Let <math>g : X \to (-\infty,+\infty]</math> be lower semicontinuous and bounded below. Then the functional <math>\mu \mapsto \int_X g \, \mathrm{d}\mu</math> is lower semicontinuous with respect to narrow convergence in <math>\mathcal{P}(X)</math>, that is | |||
:<math> \mu_n \to \mu \quad \text{narrowly} \Longrightarrow \liminf\limits_{n \to \infty} \int_X g_n \, \mathrm{d}\mu \geq \int_X g \, \mathrm{d}\mu </math>. | |||
'''Proof.''' By the Moreau-Yosida approximation, for all <math>k \geq 0</math>, | |||
:<math>\liminf\limits_{n \to \infty} \int_X g \, \mathrm{d} \mu_n \geq \liminf\limits_{n \to \infty} \int_X g_k \wedge k \, \mathrm{d}\mu_n = \int_X g_k \wedge k \, \mathrm{d}\mu </math>. | |||
Taking <math>k \to \infty</math>, Fatou's lemma ensures that | |||
:<math>\liminf\limits_{n \to \infty} \int_X g \, \mathrm{d} \mu_n \geq \liminf\limits_{k \to \infty} \int_X g_k \wedge k \, \mathrm{d} \geq \int_X g \, \mathrm{d}\mu \mu</math>. | |||
==The Mysterious Etymology of Portmanteau== | |||
(spurious stuff, will fill in later) | |||
==References== | ==References== | ||
Line 58: | Line 74: | ||
<!--Bauschke-Combette Ch 12.<ref name="BC" />; Santambrogio (6)<ref name="S" />; Ambrosio-Gigli-Savare (59-61)<ref name="AGS" />--> | <!--Bauschke-Combette Ch 12.<ref name="BC" />; Santambrogio (6)<ref name="S" />; Ambrosio-Gigli-Savare (59-61)<ref name="AGS" />--> | ||
<references> | <references> | ||
<ref name="AGS">Ambrosio, Luigi, Nicola Gigli, and Giuseppe Savaré. ''Gradient Flows in Metric Spaces and in the Space of Probability Measures.'' Ch. 3.1. Birkhäuser, 2005.</ref> | |||
<ref name="OT">Craig, Katy C. Lower Semicontinuity in the Narrow Topology. Math 260J. Univ. of Ca. at Santa Barbara. Winter 2022.</ref> | <ref name="OT">Craig, Katy C. Lower Semicontinuity in the Narrow Topology. Math 260J. Univ. of Ca. at Santa Barbara. Winter 2022.</ref> | ||
<ref name="BC">Bauschke, Heinz H. and Patrick L. Combettes. ''Convex Analysis and Monotone Operator Theory in Hilbert Spaces, 2nd Ed.'' Ch. 12. Springer, 2017.</ref> | |||
<ref name="S">Santambrogio, Filippo. ''Optimal Transport for Applied Mathematicians: Calculus of Variations, PDEs, and Modeling'' Ch. 1.1. Birkhäuser, 2015.</ref> | <ref name="S">Santambrogio, Filippo. ''Optimal Transport for Applied Mathematicians: Calculus of Variations, PDEs, and Modeling'' Ch. 1.1. Birkhäuser, 2015.</ref> | ||
</references> | </references> |
Revision as of 19:48, 10 February 2022
The Moreau-Yosida regularization is a technique used to approximate lower semicontinuous functions by Lipschitz functions. The main application of this result is to prove Portmanteau's Theorem, which states that integration against a lower semicontinuous and bounded below function is lower semicontinuous with respect to the narrow topology on the space of probability measures.
Definitions
Let Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle (X,d)} be a metric space. A function Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g : X \to (-\infty,+\infty]} is said to be proper if it is not identically equal to Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle +\infty} , that is, if there exists Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle x \in X} such that Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g(x) < +\infty} .
For a given function Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g : X \to (-\infty,+\infty]} and Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle k \geq 0} , its Moreau-Yosida regularization Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_k : X \to [-\infty,+\infty]} is given by
Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_k(x) := \inf\limits_{y \in X} \left[ g(y) + k d(x,y) \right].}
The distance term Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle d(x,y)} may often be raised to a positive exponent. For example, when Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle X} is a Hilbert space [1] [2], Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_k} is taken to be
Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_k(x) := \inf\limits_{y \in X} \left[ g(y) + k \| x - y \|^2 \right].}
Note that
- Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_k(x) = \inf\limits_{y \in X} \left[ g(y) + k d(x,y) \right] \leq g(x) + k d(x,x) = g(x)} .
Examples
- If Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle k = 0} , then by definition Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_0} is constant and Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_0 \equiv \inf\limits_{y \in X} g(y)} .
- If Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g} is not proper, then Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_k = +\infty} for all Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle k \geq 0} .
Take Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle (X,d) := (\mathbb{R},|\cdot|)} . If Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g} is finite-valued and differentiable, we can explicitly write down Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_k} . Then for a fixed Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle x \in \mathbb{R}} , the map Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_{k,x} : y \mapsto g(y) + k|x - y|} is continuous everywhere and differentiable everywhere except for when Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle y = x} , where the derivative does not exist due to the absolute value. Thus we can apply standard optimization techniques from Calculus to solve for Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_k(x)} : find the critical points of Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_{k,x}} and take the infimum of Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_{k,x}} evaluated at the critical points. One of these values will always be the original function Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g} evaluated at Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle x} , since this corresponds to the critical point Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle y = x} for Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_{k,x}} .
- Let Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g(x) := x^2} . Then
- Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_k(x) = \min \left\{ x^2 , \frac{k^2}{2} + k \left| x \pm \frac{k}{2} \right| \right\}.}

Approximating Lower Semicontinuous Functions by Lipschitz Functions
- If Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g} is proper and bounded below, so is Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_k} . Furthermore, Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_k} is continuous for all Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle k \geq 0} .
- If, in addition, Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g} is lower semicontinuous, then Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_k(x) \nearrow g(x)} for all Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle x \in X} .
- In this case, Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_k \wedge k := \min(g_k,k)} is continuous and bounded and Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle g_k(x) \wedge k \nearrow g(x)} for all Failed to parse (SVG (MathML can be enabled via browser plugin): Invalid response ("Math extension cannot connect to Restbase.") from server "https://wikimedia.org/api/rest_v1/":): {\displaystyle x \in X} .

Proof.
- Since is proper, there exists such that . Then for any
Thus is proper and bounded below. Next, for a fixed , let . Then as
- ,
the family is uniformly Lipschitz and hence equicontinuous. Thus is Lipschitz continuous.
- Suppose that is also lower semicontinuous. Note that for all , . Thus it suffices to show that . This inequality is automatically satisfied when the left hand side is infinite, so without loss of generality assume that . By definition of infimum, for each there exists such that
- .
Then
is bounded below by assumption, while the only way is finite in the limit is for to go to zero. Thus converges to in , and by lower semicontinuity of ,
- .
- By definition, . Since for all , for all .
Portmanteau Theorem
Theorem (Portmanteau). Let be lower semicontinuous and bounded below. Then the functional is lower semicontinuous with respect to narrow convergence in , that is
- .
Proof. By the Moreau-Yosida approximation, for all ,
- .
Taking , Fatou's lemma ensures that
- .
The Mysterious Etymology of Portmanteau
(spurious stuff, will fill in later)
References
- ↑ Bauschke, Heinz H. and Patrick L. Combettes. Convex Analysis and Monotone Operator Theory in Hilbert Spaces, 2nd Ed. Ch. 12. Springer, 2017.
- ↑ Ambrosio, Luigi, Nicola Gigli, and Giuseppe Savaré. Gradient Flows in Metric Spaces and in the Space of Probability Measures. Ch. 3.1. Birkhäuser, 2005.
- ↑ Craig, Katy C. Lower Semicontinuity in the Narrow Topology. Math 260J. Univ. of Ca. at Santa Barbara. Winter 2022.
- ↑ Santambrogio, Filippo. Optimal Transport for Applied Mathematicians: Calculus of Variations, PDEs, and Modeling Ch. 1.1. Birkhäuser, 2015.