TCSLib

109 Learning Theory — Convex Minimax Core

109.1 Overview

This module develops the core ingredients for the convex-compact minimax theorem, following Cesa-Bianchi and Lugosi Theorem 7.1 specialized to subsets of \(\mathbb {R}\). It provides the exact minimax value for finite matrix games (derived from the Hedge approximate saddle-point theorem), Jensen-type inequalities for mixed-strategy convex combinations, a sampled-game construction that bridges finite and continuous settings, and the compactness machinery (sublevel sets, finite-intersection property) needed for the infinite-dimensional direction.

109.2 Declarations

Definition 109.1 Arbitrary mixed strategy
#

Given \(n \geq 1\), OnlineLearning.arbitraryMixedStrategy is a concrete element of \(\mathrm{MixedStrategy}(n)\) that places all probability mass on a single arbitrarily chosen action. It is used solely to witness that the type of mixed strategies is nonempty.

Definition 109.2 Lower value of a finite game

For a finite zero-sum game \(G\) with \(M\) row actions and \(N\) column actions, the lower value is

\[ v^- (G) \; =\; \sup _{p \in \Delta _M}\; \inf _{j \in [N]}\; \mathrm{payoffVsPure}(G,p,j), \]

where the row player maximises over mixed strategies and the column player minimises over pure responses.

Definition 109.3 Upper value of a finite game

The upper value of \(G\) is

\[ v^+(G) \; =\; \inf _{q \in \Delta _N}\; \sup _{i \in [M]}\; \mathrm{pureVsPayoff}(G,i,q), \]

where the column player minimises over mixed strategies and the row player maximises over pure responses.

Lemma 109.4 Payoff vs. pure column is nonnegative

For any game \(G\), mixed row strategy \(p\), and pure column action \(j\), \(\mathrm{payoffVsPure}(G,p,j) \geq 0\). This follows because every entry of \(G\) is nonnegative and the mixing weights are nonnegative.

Lemma 109.5 Payoff vs. pure column is at most one

For any game \(G\), mixed row strategy \(p\), and pure column action \(j\), \(\mathrm{payoffVsPure}(G,p,j) \leq 1\). This follows because every entry of \(G\) is at most one and the mixing weights sum to one.

Lemma 109.6 Pure row payoff vs. mixed column is nonnegative

For any game \(G\), pure row action \(i\), and mixed column strategy \(q\), \(\mathrm{pureVsPayoff}(G,i,q) \geq 0\).

Lemma 109.7 Pure row payoff vs. mixed column is at most one

For any game \(G\), pure row action \(i\), and mixed column strategy \(q\), \(\mathrm{pureVsPayoff}(G,i,q) \leq 1\).

Lemma 109.8 Row-guaranteed payoffs bounded above

When \(N \geq 1\), the set \(\bigl\{ \inf _{j}\, \mathrm{payoffVsPure}(G,p,j) : p \in \Delta _M\bigr\} \) is bounded above by \(1\).

Lemma 109.9 Column-induced upper values bounded below

When \(M \geq 1\), the set \(\bigl\{ \sup _{i}\, \mathrm{pureVsPayoff}(G,i,q) : q \in \Delta _N\bigr\} \) is bounded below by \(0\).

Lemma 109.10 Weak duality for finite games

For any finite zero-sum game \(G\) (with \(M,N \geq 1\)), \(v^-(G) \leq v^+(G)\).

For a finite zero-sum game \(G\) with \(M \gt 1\) rows and \(N \geq 1\) columns, the lower and upper values coincide:

\[ v^-(G) \; =\; v^+(G). \]

The proof extracts arbitrarily good approximate saddle points from the Hedge algorithm, then passes to the limit in \(\varepsilon \).

Definition 109.12 Convex-compact minimax statement
#

For sets \(X, Y \subseteq \mathbb {R}\) and a payoff function \(f : \mathbb {R}\to \mathbb {R}\to \mathbb {R}\), OnlineLearning.ConvexCompactMinimaxStatement is the proposition

\[ \inf _{x \in X}\, \sup _{y \in Y}\, f(x,y) \; =\; \sup _{y \in Y}\, \inf _{x \in X}\, f(x,y). \]

It is packaged as a named Prop so that intermediate lemmas can refer to it while the proof is built up incrementally.

Definition 109.13 Convex-compact minimax hypotheses

A structure bundling the assumptions of Cesa-Bianchi–Lugosi Theorem 7.1 for \(f : \mathbb {R}\to \mathbb {R}\to \mathbb {R}\) on sets \(X, Y \subseteq \mathbb {R}\): both sets are nonempty, \(X\) is compact and convex, \(Y\) is convex, \(f(\cdot ,y)\) is continuous and convex on \(X\) for every \(y \in Y\), \(f(x,\cdot )\) is concave on \(Y\) for every \(x \in X\), and \(f\) is bounded above and below on \(X \times Y\).

Lemma 109.14 Mixed-strategy convex combination stays in convex set
#

If \(S \subseteq \mathbb {R}\) is convex, \(p \in \Delta _n\) is a mixed strategy, and \(x_1,\dots ,x_n \in S\), then \(\sum _{i} p_i\, x_i \in S\).

Lemma 109.15 Jensen’s inequality for mixed-strategy combinations (convex)
#

If \(g\) is convex on \(S \subseteq \mathbb {R}\), \(p \in \Delta _n\), and every \(x_i \in S\), then

\[ g\! \left(\sum _{i} p_i\, x_i\right) \; \leq \; \sum _{i} p_i\, g(x_i). \]
Lemma 109.16 Jensen’s inequality for mixed-strategy combinations (concave)
#

If \(g\) is concave on \(S \subseteq \mathbb {R}\), \(p \in \Delta _n\), and every \(y_j \in S\), then

\[ \sum _{j} p_j\, g(y_j) \; \leq \; g\! \left(\sum _{j} p_j\, y_j\right). \]
Definition 109.17 Sampled finite game
#

Given a payoff function \(f : \mathbb {R}\to \mathbb {R}\to \mathbb {R}\) with values in \([0,1]\), finite families of row points \(x : [M] \to \mathbb {R}\) and column points \(y : [N] \to \mathbb {R}\), the sampled game is the \(M\times N\) matrix game whose \((i,j)\) entry is \(f(x_i, y_j)\).

Theorem 109.18 Finite minimax for sampled games

Under the same size hypotheses as OnlineLearning.finite_minimax_value, the sampled game satisfies \(v^-(\mathrm{sampledGame}) = v^+(\mathrm{sampledGame})\).

Lemma 109.19 Convexity gives lower bound on sampled payoff vs. pure column

If \(f(\cdot , y_j)\) is convex on \(X\) and every \(x_i \in X\), then for any mixed row strategy \(p \in \Delta _M\),

\[ f\! \left(\sum _i p_i\, x_i,\; y_j\right) \; \leq \; \mathrm{payoffVsPure}(\mathrm{sampledGame},\, p,\, j). \]

That is, convexity in the row variable means the continuous mixed-action payoff is at most the finite-game expected payoff.

Lemma 109.20 Concavity gives upper bound on sampled payoff vs. mixed column

If \(f(x_i,\cdot )\) is concave on \(Y\) and every \(y_j \in Y\), then for any mixed column strategy \(q \in \Delta _N\),

\[ \mathrm{pureVsPayoff}(\mathrm{sampledGame},\, i,\, q) \; \leq \; f\! \left(x_i,\; \sum _j q_j\, y_j\right). \]
Lemma 109.21 Weak minimax inequality for convex-compact games

Under the hypotheses OnlineLearning.ConvexCompactMinimaxHypotheses,

\[ \sup _{y \in Y}\, \inf _{x \in X}\, f(x,y) \; \leq \; \inf _{x \in X}\, \sup _{y \in Y}\, f(x,y). \]

This is the easy direction of the minimax theorem: for every fixed \((x,y)\), \(\inf _{x'} f(x',y) \leq f(x,y) \leq \sup _{y'} f(x,y')\), and the inequality is preserved when taking \(\sup \) then \(\inf \).

Definition 109.22 Minimax sublevel set
#

For a set \(X \subseteq \mathbb {R}\), a payoff function \(f\), a column point \(y\), and a threshold \(c \in \mathbb {R}\), the minimax sublevel set is

\[ \mathrm{sublevel}(X,f,y,c) \; =\; X \cap \{ x \in \mathbb {R}: f(x,y) \leq c\} . \]
Lemma 109.23 Sublevel set contained in row set

\(\mathrm{sublevel}(X,f,y,c) \subseteq X\) for all \(f\), \(y\), \(c\).

Lemma 109.24 Sublevel set is closed

If \(X\) is closed and \(f(\cdot ,y)\) is continuous on \(X\), then \(\mathrm{sublevel}(X,f,y,c)\) is closed.

Lemma 109.25 Sublevel set is compact

If \(X\) is compact and \(f(\cdot ,y)\) is continuous on \(X\), then \(\mathrm{sublevel}(X,f,y,c)\) is compact (as a closed subset of a compact set).

Lemma 109.26 Finite-intersection gives global row minimiser

Under OnlineLearning.ConvexCompactMinimaxHypotheses, let \(c \in \mathbb {R}\). If for every finite subset \(u \subseteq Y\) there exists \(x \in X\) with \(f(x,y) \leq c\) for all \(y \in u\), then there exists \(x \in X\) with \(f(x,y) \leq c\) for all \(y \in Y\). The proof applies the finite-intersection property to the compact collection of sublevel sets.