Operations Research

Motivation

Reinforcement learning formalizes a scenario supervised learning cannot: an agent that actively interacts with an environment, choosing actions that change both the state it observes next and the reward it receives, rather than passively receiving an i.i.d. labeled sample. Every practical treatment of this scenario — from classical dynamic programming to modern deep reinforcement learning — is built on the Markov decision process (MDP), a model in which the effect of an action depends only on the current state, not on the full history that led to it. Two questions define the theory this mission covers: given a fixed way of acting (a policy), what value does it obtain, and how is that value actually computed rather than merely characterized as the solution of a fixed-point equation? Mohri, Rostamizadeh and Talwalkar's chapter 17 answers both for the stationary, infinite-horizon discounted case, and this mission targets its two central results: that a fixed policy's value is not just characterized but uniquely determined by a linear system with an explicit closed-form solution (Theorem 17.10), and that the optimal value function — obtained instead by choosing the best action at every state — can be computed by an iterative algorithm guaranteed to converge regardless of where it starts (Theorem 17.11).

Setting

A (finite) Markov decision process consists of a finite set of states $S$, a finite set of actions $A$, a transition kernel $P[s'\mid s,a]$ giving the distribution over the next state $s'$ after taking action $a$ at state $s$, and an expected reward $\mathbb E[r(s,a)]$ for that transition. A (stationary) policy $\pi:S\to\Delta(A)$ assigns each state a distribution over actions — possibly, but not necessarily, a point mass on a single action. Fixing $\pi$ turns the MDP into an ordinary Markov chain on $S$: at each step the agent is at some state $s$, draws $a\sim\pi(s)$, receives (expected) reward $\mathbb E[r(s,a)]$, and moves to a state drawn from $P[\cdot\mid s,a]$. For a discount factor $\gamma\in[0,1)$, the value of $\pi$ at $s$ is the expected discounted sum of future rewards starting from $s$,

$$V_\pi(s) = \mathbb E_{a_t\sim\pi(s_t)}\Big[\sum_{t=0}^{+\infty}\gamma^t r(s_t,a_t) \;\Big|\; s_0=s\Big],$$

and the state-action value function $Q_\pi(s,a)$ is the analogous quantity for taking $a$ at $s$ and then following $\pi$. Marginalizing the raw kernel and reward over the mixed action $\pi(s)$ gives the induced transition matrix $P_{s,s'}=P[s'\mid s,\pi(s)]=\sum_a \pi(s)(a) P[s'\mid s,a]$ and induced reward vector $R_s=\mathbb E[r(s,\pi(s))]=\sum_a\pi(s)(a)\,\mathbb E[r(s,a)]$ — the objects that turn $\pi$'s value into a genuinely linear-algebraic quantity. A policy $\pi^*$ is optimal if $V_{\pi^*}(s)\ge V_\pi(s)$ for every policy $\pi$ and every state $s$; write $V^*$ for its value function.

Formalization targets

Theorem 17.10 (goal). For a finite MDP and a fixed policy $\pi$, the matrix $I-\gamma P$ (with $P$ the policy-induced transition matrix) is invertible, and $\pi$'s value function is the unique solution of the Bellman equations, given in closed form by

$$V_\pi = (I-\gamma P)^{-1} R.$$

Proposition 17.9 (milestone). The value function itself satisfies the linear system that Theorem 17.10 solves:

$$\forall s\in S,\quad V_\pi(s) = \mathbb E_{a\sim\pi(s)}[r(s,a)] + \gamma\sum_{s'} P[s'\mid s,\pi(s)]\,V_\pi(s').$$

Theorem 17.7 (milestone). A policy $\pi$ is optimal if and only if it places probability only on $Q_\pi$-maximizing actions: for every $(s,a)$ with $\pi(s)(a)>0$, $a\in \operatorname{argmax}_{a'} Q_\pi(s,a')$.

Theorem 17.11 (milestone). The Bellman optimality operator $\Phi$, $[\Phi(V)](s)=\max_{a} \{\mathbb E[r(s,a)]+\gamma\sum_{s'}P[s'\mid s,a]V(s')\}$, is a $\gamma$-contraction for $\lVert\cdot\rVert_\infty$; consequently, for any starting vector $V_0$, the value-iteration sequence $V_{n+1}=\Phi(V_n)$ converges to a fixed point of $\Phi$.

Significance

Theorem 17.10 is what makes policy evaluation on a finite MDP an exact, finite computation rather than an infinite limit: instead of summing an infinite discounted series or solving an implicit fixed-point equation numerically, a single $|S|\times|S|$ matrix inversion gives the policy's value at every state simultaneously. It is also the base case every planning algorithm in the chapter builds on: policy iteration alternates optimizing a policy with exactly this evaluation step. Theorem 17.11 gives the complementary guarantee for the harder problem of finding the optimal value function directly, without fixing a policy first: value iteration converges from any starting point, with a convergence rate ($O(\log(1/\epsilon))$ iterations for $\epsilon$-accuracy) that follows from the same contraction argument. Together, the two results are the mathematical content behind why dynamic-programming planning for finite MDPs is tractable at all — the discount factor $\gamma<1$, not any structural assumption on rewards or transitions, is what buys both the uniqueness in Theorem 17.10 and the convergence in Theorem 17.11. Formalizing them requires reproducing this linear-algebraic and metric content precisely, not just asserting the conclusions: an invertibility claim asserted without the operator-norm argument, or a convergence claim without the contraction property, would state something true by fiat rather than the book's actual result. No faithful prior art exists on the platform for this exact model (see Formalization scope).

Difficulty

The obvious shortcut for Theorem 17.10 is to assert $I-\gamma P$ is invertible without proof — true, but not what the book does, and not informative about why it holds. The genuine content is that $P$, being row-stochastic (every row of $P$ sums to exactly $1$, since $\pi(s)$ and $P[\cdot\mid s,a]$ are both proper distributions), has operator norm $\lVert P\rVert_\infty=1$ exactly, so $\lVert\gamma P\rVert_\infty=\gamma<1$ strictly; this rules out $1$ as an eigenvalue of $\gamma P$, which is exactly what invertibility of $I-\gamma P$ requires. The same $\gamma<1$ fact, applied differently, drives Theorem 17.11: showing $\Phi$ is $\gamma$-Lipschitz requires bounding $\Phi(V)(s)-\Phi(U)(s)$ by comparing the maximizing action for $V$ against the same action's value under $U$ (not $U$'s own maximizer), since the two suprema need not be attained at the same action — a step easy to state incorrectly as a direct comparison of two maxima. Both theorems fail if $\gamma=1$ is allowed: the discounted setting's central asset, a strict contraction, disappears exactly at that boundary.

Formalization scope

States and actions are modeled as finite types (Fintype S, Fintype A); the raw kernel and reward P : S → A → S → ℝ, Er : S → A → ℝ are unconstrained functions, with IsTransitionKernel asserting the required distribution property explicitly rather than assuming it silently. A policy is π : S → A → ℝ with IsPolicy π asserting π s is a distribution over A for every s — deliberately not π : S → A or a PMF-valued function, since Theorem 17.7's own quantifier ("for any pair (s,a) with π(s)(a) > 0") requires treating π(s) as a genuine mixture. PolicyValue is defined as the actual infinite discounted expectation (via an explicit state-occupation-distribution recursion), not as the Bellman fixed point — so that Proposition 17.9 (the value function satisfies the linear system) and Theorem 17.10 (that system has a unique, invertible-matrix solution) are both non-vacuous claims about the same object, rather than one being definitionally true of the other. The trivializing formalization this rules out is asserting IsUnit (1 - γ • P) as a bare hypothesis, or defining V_π as (1-γP)⁻¹R and calling the resulting identity a theorem; both would erase the mission's actual content. Two platform modules model related MDPs (BertsekasSSPModel, a stochastic-shortest-path model with a termination-probability deficit rather than exact row-stochasticity, and FoundationsRL.RLBasics, a finite-horizon episodic model indexed by layer) — neither specializes exactly to this chapter's stationary, always-continuing, infinite-horizon discounted convention, so every definition here is drafted fresh rather than imported. This chunk covers §17.2–17.4.2 (the MDP model, policy value, Bellman's equations, value and policy iteration); §17.4.3 (the linear-programming formulation) and §17.5 (stochastic-approximation learning algorithms — TD(0), Q-learning, SARSA) are out of scope, since they require a stochastic-approximation convergence substrate this mission does not build.

Selected references

• Mohri, M., Rostamizadeh, A., and Talwalkar, A. Foundations of Machine Learning, 2nd ed., chapter 17. MIT Press, 2018.
• Bellman, R. Dynamic Programming. Princeton University Press, 1957.
• Puterman, M. L. Markov Decision Processes: Discrete Stochastic Dynamic Programming. Wiley, 1994.

Comparing location and spread together

Chapter I's usual stochastic order compares "how large" two random variables tend to be; Chapter III's convex order compares "how spread out" they are, holding the mean fixed. Chapter IV's increasing convex and increasing concave orders combine the two: $X \le_{icx} Y$ says $X$ is both smaller and less variable than $Y$ in a single comparison, without forcing equal means. These are the orders a decision-maker with risk-averse (concave-utility) or risk-loving (convex-cost) preferences actually uses to rank random outcomes, since expected utility is exactly an expectation of an increasing concave or convex function. This mission formalizes both orders and their coupling characterization: the submartingale/supermartingale analogue, one chapter over, of Chapter III's Strassen martingale coupling for the plain convex order.

The increasing convex and increasing concave orders

Let $X$ be a real-valued random variable on a probability space $(\Omega,\mu)$, and let $Y$ be a real-valued random variable on a (possibly different) probability space $(\Omega',\nu)$. $X$ is smaller than $Y$ in the increasing convex order, written $X \le_{icx} Y$, if

$$E[\varphi(X)] \le E[\varphi(Y)] \quad \text{for every increasing convex } \varphi:\mathbb{R}\to\mathbb{R} \text{ for which the two expectations exist,}$$

and smaller than $Y$ in the increasing concave order, $X \le_{icv} Y$, if the same holds for every increasing concave $\varphi$. Taking $\varphi(x)=x$ (increasing and both convex and concave) gives $E[X]\le E[Y]$ under either order — unlike the convex order, no equality of means is forced. Two equivalent tail-integral characterizations (Theorem 4.A.2) make the orders tractable: $X\le_{icx}Y$ iff $\int_x^\infty \bar F(u)\,du \le \int_x^\infty \bar G(u)\,du$ for every $x$, and $X\le_{icv}Y$ iff $\int_{-\infty}^x F(u)\,du \ge \int_{-\infty}^x G(u)\,du$ for every $x$, where $\bar F,\bar G$ and $F,G$ are the survival and distribution functions.

Formalization targets

Goal: the submartingale-coupling characterization (Theorem 4.A.5, increasing convex case)

$$X \le_{icx} Y \iff \exists\,(\Omega'',\rho),\ \hat X,\hat Y:\Omega''\to\mathbb{R}\text{ with } \hat X=_{st}X,\ \hat Y=_{st}Y,\ E[\hat Y\mid\hat X]\ge\hat X\text{ a.s.}$$

Furthermore, $\hat X,\hat Y$ can be chosen so that $[\hat Y\mid\hat X=x]$ stochastically increases with $x$ (in $\le_{st}$). This is the increasing-convex analogue of Chapter III's Theorem 3.A.4: instead of a martingale, $\{\hat X,\hat Y\}$ need only be a submartingale — the copy of $Y$ is, conditionally on the copy of $X$, at least a fair randomization of it. The book states its proof is "similar to the proof of Theorem 3.A.4" and calls the constructive direction "not easy to prove"; this mission states the theorem faithfully, including the "Furthermore" strengthening, without attempting a proof.

Companion: the supermartingale-coupling characterization (Theorem 4.A.5, increasing concave case)

The same theorem's other bracketed case: $X \le_{icv} Y$ iff there exist $\hat X,\hat Y$ on a common space with $\hat X=_{st}X$, $\hat Y=_{st}Y$, and $\{\hat Y,\hat X\}$ a supermartingale, $E[\hat X\mid\hat Y]\le\hat Y$ a.s. — note the swapped roles of $\hat X$ and $\hat Y$ in the conditioning relative to the increasing-convex case, not merely a flipped inequality. Drafted as a separate Lean theorem from the goal (see Formalization scope), since the two conditioning structures are genuinely different predicates, not sign-flipped rewrites of one another.

Supporting milestones

• Theorem 4.A.1, the duality relation: $X\le_{icx}Y \iff -X\ge_{icv}-Y$, and $X\le_{icv}Y\iff -X\ge_{icx}-Y$ — the increasing-order analogue of Theorem 3.A.12(a)'s duality for the plain convex order.
• Theorem 4.A.2, the tail-integral characterizations above, for integrable $X,Y$.
• Theorem 4.A.8(d), closure under convolution: independent $X_i\le_{icx}Y_i$ (resp. $\le_{icv}$) for $i=1,\dots,m$ gives $\sum_i X_i \le_{icx} \sum_i Y_i$ (resp. $\le_{icv}$) — the increasing-order analogue of Chapter III's Theorem 3.A.12(d), which the book itself says "can be proven as in Theorem 3.A.12."

Significance

The increasing convex and concave orders are the natural language of expected-utility comparisons: a risk-averse agent with concave utility $u$ prefers $Y$ to $X$ exactly when $X\le_{icv}Y$ implies $E[u(X)]\le E[u(Y)]$ for every increasing concave $u$ — the order is defined precisely so that "every risk-averse agent with increasing utility agrees" collapses to one comparison. In operations research this underlies stochastic dominance of the second kind in portfolio and inventory models, and the increasing-convex order similarly formalizes "second-order stochastic dominance for costs" used to compare random cost/loss distributions under risk-loving or regret-averse preferences. The submartingale-coupling characterization is what turns the intractable "for every increasing convex $\varphi$" quantifier into a single explicit construction, exactly as Strassen's theorem does for the plain convex order — and, being the direct chapter-IV successor to Chapter III's coupling theorem in this series, it fixes the second data point for what a "coupling-characterization" mission in this book's series looks like when the order being characterized is not symmetric between the two variables' roles.

No platform prior art exists: GET /theorems?q=increasing+convex, q=increasing+concave, and q=submartingale return zero genuinely matching hits (the one submartingale hit, a bandit subgaussian maximal inequality, uses Doob's inequality for a concentration bound, not this order). This mission restates the increasing convex/concave orders and their coupling characterization as a foundational, self-contained pair of definitions, parallel in structure to Chunk 03's convex-order mission but independently drafted (drafts cannot import each other's Lean).

Difficulty

The chief formalization risk is the one this chapter's own brief flags explicitly: the book states Theorem 4.A.5 as a single statement with bracketed alternatives ("$X\le_{icx}Y$ [$X\le_{icv}Y$] iff … $\{\hat X,\hat Y\}$ is a submartingale [$\{\hat Y,\hat X\}$ is a supermartingale] …"), and the two cases are not symmetric rewrites of each other — the conditioning variable in E[Ŷ|X̂]≥X̂ swaps to E[X̂|Ŷ]≤Ŷ in the concave case, not just a flipped inequality on the same conditioning. A Lean statement that tried to unify both cases with a single Or or a naive sign-flip would either conflate two different predicates or silently state the wrong condition for one of the two cases. The duality theorem (4.A.1) compounds this risk in the other direction: its content is exactly that $\le_{icx}$ and $\ge_{icv}$ are related (by negating the variables), so a formalization that makes this equivalence provable by unfolding definitions (rather than as a genuine ↔ between two independently-stated predicates) would trivialize the one theorem whose entire content is that relationship.

Formalization scope

Random variables are measurable functions into $\mathbb{R}$ from arbitrary measurable spaces, matching this series' convention; IcxOrder μ ν X Y and IcvOrder μ ν X Y are drafted as two separate definitions (not one order parametrized by an Or of function classes), each quantifying over Monotone φ ∧ ConvexOn ℝ Set.univ φ (resp. ConcaveOn) with the two expectations' existence stated as explicit Integrable hypotheses inside the ∀, exactly as Chunk 03's ConvexOrder. Equality in law is ProbabilityTheory.IdentDistrib. The goal's submartingale condition uses Mathlib's conditional-expectation notation X̂ ≤ᵐ[ρ] ρ[Ŷ | m] with m the σ-algebra generated by X̂; its companion's supermartingale condition uses ρ[X̂ | m'] ≤ᵐ[ρ] Ŷ with m' generated by Ŷ instead — the conditioning variable is genuinely swapped between the two theorems, matching the book's own bracket ordering {X̂,Ŷ} vs. {Ŷ,X̂}. The "Furthermore" clause in both is kept (not dropped), via Mathlib's ProbabilityTheory.condDistrib, restated as the relevant conditional-distribution kernel's survival function being monotone at every threshold — the same shape Chapter I's UsualOrder and Chunk 03's goal theorem use, necessarily restated locally since drafts cannot import another mission's definitions.

Theorem 4.A.1 (duality) and Theorem 4.A.8(d) (convolution closure) are each drafted as a single Lean theorem with two independent conjuncts (an ∧ of two ↔s, or of two implications), one per bracketed case, since the book states both cases as the two halves of one theorem sharing every hypothesis — this is not the same shape as the goal/companion split, where the two cases have genuinely different internal structure (the swapped conditioning) rather than a shared statement instantiated at two function classes. A trivializing formalization this mission rules out: stating Theorem 4.A.1's duality by relabeling IcvOrder as IcxOrder applied to negated arguments (making the equivalence a rfl or single simp unfolding) rather than keeping the two orders as independently-defined predicates whose relationship is the theorem's actual content.

This mission draws on no platform prior art (searches for "increasing convex", "increasing concave", and "submartingale" as of 2026-09-18 return zero genuine matches — the sole submartingale hit is an unrelated bandit concentration inequality via Doob's inequality). Reusable beyond this mission: the IcxOrder/IcvOrder definition pattern and the submartingale/supermartingale coupling shape parallel Chunk 03's ConvexOrder martingale-coupling pattern closely enough that a future chapter needing "coupling characterization of a location-and-spread order" (none of the remaining chapters currently in this series' first wave) could restate the same shape with minimal adaptation.

Selected references

• M. Shaked and J. G. Shanthikumar, Stochastic Orders, Springer Series in Statistics, Springer, 2007, Chapter 4 (Univariate Monotone Convex and Related Orders), §4.A. https://doi.org/10.1007/978-0-387-34675-5
• This series' Chunk 03 (StochasticOrders.Convex), for the plain convex order and Strassen's martingale coupling this chapter's submartingale/supermartingale coupling directly generalizes.

Motivation

A Markov decision process (MDP) chooses a decision in every period of a dynamic system whose state evolves stochastically in response to the decision, so as to maximize expected discounted return. Firms use this model for inventory, pricing, and advertising decisions that respond to a randomly evolving demand state; engineers use it for maintaining or replacing equipment that degrades stochastically. A recurring practical question is qualitative rather than numerical: does the optimal decision increase with the state — should a firm price higher after a period of strong sales, or replace a machine sooner the worse its observed condition — without having to solve the dynamic program numerically for every instance of the model? Topkis's Chapter 3, Section 3.9 (Topkis, Supermodularity and Complementarity, 2011, building on Topkis [1968]) answers this by isolating the lattice-theoretic structure — supermodularity of the return function and of the transition law — under which monotone optimal policies are guaranteed on structural grounds alone. A closely related but logically independent question was studied earlier by Lehmann [1955], who characterized when a family of distributions is stochastically increasing in a parameter; Topkis generalizes Lehmann's characterization to any property of the parameter dependence whose defining set of functions forms a closed convex cone, of which stochastic monotonicity, supermodularity, and convexity are three instances (Theorem 3.9.1, Corollary 3.9.1). Serfozo [1976] independently develops related conditions for partially observed Markov decision processes, and Amir [1996] and Amir, Mirman, and Perkins [1991] give analogous monotonicity results for other classes of dynamic programming models.

Setting

Fix a finite planning horizon of $k$ periods, $i = 1, \dots, k$. In period $i$ the state $t$ ranges over a set $T_i \subseteq \mathbb{R}^m$; given state $t$, the decision $x$ is restricted to a finite, nonempty set $X_{t,i} \subseteq \mathbb{R}^n$ (finiteness guarantees that an optimal decision always exists — no continuity or compactness argument is used). Write $S_i = \{(x,t) : t \in T_i,\, x \in X_{t,i}\}$ for the set of admissible (decision, state) pairs in period $i$. The (bounded) expected net return of choosing decision $x$ in state $t$, period $i$, is $r_i(x,t)$. A discount rate $\beta \in [0,1]$ gives $\gamma = 1/(1+\beta)$, the value in period $i$ of one unit of return in period $i+1$. Given decision $x$, state $t$, and period $i$, the state $w$ of period $i+1$ is drawn from a distribution $F(x,t,i,\cdot)$ on $\mathbb{R}^m$.

The optimal-value function $f_i(t)$ (present value of acting optimally from state $t$, period $i$, onward) and the decision-value function $g_i(x,t)$ (present value of choosing $x$ in state $t$, period $i$, then acting optimally thereafter) are defined by backward induction from period $k$:

$$g_k(x,t) = r_k(x,t), \qquad f_i(t) = \max_{x \in X_{t,i}} g_i(x,t), \qquad g_i(x,t) = r_i(x,t) + \gamma \int f_{i+1}(w)\, dF(x,t,i,w) \quad (i < k).$$

A subset $S$ of a Euclidean space is increasing if it is upward closed under the coordinatewise order. A family of distributions $\{F(a,\cdot) : a \in D\}$ indexed by a parameter $a$ is stochastically increasing on $D$ if the probability $\int_S dF(a,w)$ of every increasing set $S$ is a monotone (non-decreasing) function of $a$ on $D$; when $D$ is a sublattice, the family is stochastically supermodular on $D$ if that same probability is a supermodular function of $a$ on $D$. A real-valued function $\varphi$ on a lattice is supermodular on a set $D$ if $\varphi(a_1) + \varphi(a_2) \le \varphi(a_1 \vee a_2) + \varphi(a_1 \wedge a_2)$ for all $a_1, a_2 \in D$ — the mission series' shared notion, defined once in chunk 02-monotonicity and reused here as Supermodularity.Monotonicity.SupermodularOn.

Formalization targets

Goal — Theorem 3.9.2 (monotone optimal policies)

$$g_i(x,t) \text{ supermodular on } S_i, \qquad f_i(t) \text{ supermodular on } T_i, \qquad \arg\max_{x \in X_{t,i}} g_i(x,t) \text{ increasing in } t,$$

together with the existence of a greatest and a least optimal decision at every state, each increasing in the state, in every period $i$ — under the hypotheses that $S_i$ is a sublattice of $\mathbb{R}^{n+m}$, $X_{t,i}$ is expanding in $t$, $r_i$ is increasing in $t$ (on sections) and jointly supermodular in $(x,t)$, and $F(x,t,i,\cdot)$ is stochastically increasing in $t$ (on sections) and stochastically jointly supermodular in $(x,t)$, for every period $i$.

This is the weakest of the targets that is still worth stating on its own: it packages four related conclusions (parts (a)–(d) of Theorem 3.9.2) that share one hypothesis set, rather than isolating just the headline monotone-decision claim, because the book's own proof derives all four together and a solver attacking part (c) or (d) needs part (a) and (b) established first.

Supporting milestones

• Lemma 3.9.4: under the monotonicity half of the goal's hypotheses alone (no supermodularity), $f_i(t)$ is increasing in $t$ for every period $i$. This is the induction Theorem 3.9.2 reuses and strengthens.
• Corollary 3.9.1(b): on a sublattice $T$, a family of distributions is stochastically supermodular in $t$ iff $\int h(w)\,dF(t,w)$ is supermodular in $t$ for every increasing $h$. This is what lets the goal's hypothesis "$F(x,t,i,\cdot)$ stochastically supermodular in $(x,t)$" be converted into "$\int f_{i+1}(w)\,dF(x,t,i,w)$ supermodular in $(x,t)$", the step that feeds $g_i$'s supermodularity.
• Theorem 3.9.1: the general closed-convex-cone characterization of which Corollary 3.9.1(b) is the supermodularity instance (monotonicity and convexity are the other two instances, not formalized here since the goal's proof only needs the supermodularity case).

Significance

The result gives a purely structural sufficient condition — no smoothness, convexity of the decision set, or specific functional form — for monotone comparative statics in dynamic optimization: whenever the one-period return and the state transition are individually monotone and jointly supermodular in (decision, state), so is the whole multi-period value function, and so are the optimal decisions. Textbook applications include optimal advertising that decreases with prior-period sales, optimal pricing that increases with prior-period sales, and optimal maintenance of a deteriorating system, none of which need to be re-derived from scratch once the structural hypotheses are checked for the specific return and transition functions at hand.

Formalizing it contributes a genuinely new layer to the mission series and, per the paper's own triage (substrate.md), to the platform's application of Mathlib's probability-kernel infrastructure: a finite-horizon MDP with a Lebesgue–Stieltjes-style transition law and its associated stochastic-dominance vocabulary (stochastically increasing / stochastically supermodular families of distributions) did not previously exist on the platform or in this mission series, and both are reusable beyond this mission by any future formalization of dynamic programming under uncertainty. The proof itself is not open — Topkis [2011] (unchanged from the 1968/1998 original) gives a complete, elementary backward-induction argument — so what this mission produces is the formalization of a known, structurally distinctive proof technique, not a new mathematical result.

Difficulty

The naive argument — "supermodularity of $r_i$ plus supermodularity of the transition kernel obviously gives supermodularity of $g_i$" — breaks exactly at the integral: supermodularity of $(x,t) \mapsto F(x,t,i,\cdot)$ is a statement about the whole family of distributions, not about a single number, so "the transition is jointly supermodular" has to be unpacked into "the probability of every increasing set is jointly supermodular in $(x,t)$" before it says anything about $\int f_{i+1}(w)\,dF(x,t,i,w)$ for a specific function $f_{i+1}$. Corollary 3.9.1(b) is exactly the step that licenses this unpacking, and it is not free: it needs Theorem 3.9.1's closed-convex-cone argument (approximating $f_{i+1}$ from below by increasing step functions and invoking monotone convergence), not a pointwise argument on $r_i$ and $F$ separately. A second place the naive argument fails is at the constraint sets: because $X_{t,i}$ only grows with $t$ rather than being fixed, monotonicity of $f_i$ (Lemma 3.9.4) needs its own induction combining that growth with monotonicity of $g_i$ — supermodularity of $g_i$ alone does not hand you monotonicity of the arg max without it.

Formalization scope

States and decisions are represented as Fin m → ℝ and Fin n → ℝ (finite-dimensional Euclidean coordinate spaces with the coordinatewise/product order, matching the book's own restriction to $\mathbb{R}^m$ and $\mathbb{R}^n$ — no abstract lattice is used where the book itself specializes). Distributions are represented as MeasureTheory.Measure on the relevant coordinate space, with IsProbabilityMeasure supplied explicitly wherever a stochastic-dominance hypothesis is used (the probability of a set is read off as (μ S).toReal, which is only faithful to $\int_S dF$ when μ is a probability measure — an unconstrained arbitrary measure would let .toReal collapse an infinite value to $0$ and make the hypothesis trivially satisfiable, a formalization this mission rules out). Integrands h are required integrable against every measure in the family wherever an integral is asserted to lie in a set V, since the Bochner integral of a non-integrable function is definitionally 0 in Lean/Mathlib and would otherwise make Theorem 3.9.1 and Corollary 3.9.1(b) trivially true. Decision sets $X_{t,i}$ are finite (Finset, not merely a bounded or compact Set) and required nonempty exactly where the book assumes it: this finiteness, not any compactness or semicontinuity argument, is what guarantees an optimal decision exists, and dropping it would silently substitute Chapter 2's compactness-based existence machinery for the different argument this section actually uses. The optimal-value and decision-value functions $f_i, g_i$ are represented as any functions satisfying the two backward-recursion equations that define them, rather than being constructed by explicit backward recursion in Lean; since the equations determine $f_i, g_i$ uniquely from $r_i$ and $F$, this is a faithful reading of "define $f_i, g_i$ by (3.9.1) and (3.9.2)," not a weakening of the theorem. The goal's part (c) is stated using the mission series' InducedSetOrder (the Veinott/strong set order, chunk 01-lattices) and part (d) as the existence of two selection functions (greatest, least optimal decision), each monotone in the state — matching the book's "there is a greatest (least) optimal decision ... and this greatest (least) optimal decision is increasing in $t$." The infinite-horizon stationary extension that the book gives immediately after Theorem 3.9.2 (relying on an unproved citation to Blackwell [1965]) is out of scope for this mission.

Reusable infrastructure: the StochasticallyIncreasingOn/StochasticallySupermodularOn definitions are parametric in the ambient preorder/lattice and in the measure's target dimension, so a future mission on stochastic convexity (Corollary 3.9.1(c), not formalized here) or on Topkis's §3.10 stochastic inventory model (which explicitly depends on §3.9, per the book's own reading-order note) can reuse them without modification. Contributions extending this mission to the infinite-horizon case, or completing Corollary 3.9.1's monotonicity and convexity halves, are welcome.

Selected references

• D. M. Topkis, Supermodularity and Complementarity, Princeton University Press, 2011 (unchanged from the 1998 original), Chapter 3, Section 3.9. DOI: 10.1515/9781400822539.
• D. M. Topkis, "Ordered Optimal Solutions," PhD dissertation / working paper, Stanford University, 1968 (the original source for this section's results).
• E. L. Lehmann, "Ordered Families of Distributions," Annals of Mathematical Statistics 26(3), 1955, pp. 399–419. https://doi.org/10.1214/aoms/1177728487
• R. Serfozo, "Monotone Optimal Policies for Markov Decision Processes," Mathematical Programming Study 6, 1976, pp. 202–215.
• R. Amir, "Sensitivity Analysis of Multisector Optimal Economic Dynamics," Journal of Mathematical Economics 25(1), 1996, pp. 123–141.
• R. Amir, L. J. Mirman, and W. R. Perkins, "One-Sector Nonclassical Optimal Growth: Optimality Conditions and Comparative Dynamics," International Economic Review 32(3), 1991, pp. 625–644.
• D. Blackwell, "Discounted Dynamic Programming," Annals of Mathematical Statistics 36(1), 1965, pp. 226–235. https://doi.org/10.1214/aoms/1177700285

Motivation

A device's mean residual life at age $t$ — its conditional expected remaining lifetime given that it has survived to $t$ — is one of the oldest and most interpretable summaries in reliability and survival analysis: it is what an insurer, a maintenance planner, or a hospital outcomes researcher actually wants to know about a unit still in service. Comparing two mean residual life functions pointwise gives the mean residual life order $\le_{mrl}$, a natural "the survivor of $X$ is worn less, on average, than the survivor of $Y$" comparison that is weaker than the usual stochastic order but not directly comparable to it (the book states plainly that neither implies the other in general). This mission formalizes the order's definition and its precise relationship to the stronger hazard rate order $\le_{hr}$: under an extra monotone-ratio condition the two orders coincide, and one direction of that coincidence always holds. A third milestone gives one of the chapter's closure properties, showing that "decreasing mean residual life" (DMRL) — an aging notion used throughout reliability theory to describe units that wear out, rather than improve, with age — is preserved under adding independent noise.

Setting

Fix a probability space $(\Omega,\mu)$ and a real-valued random variable $X$ with survival function $\bar F(x) = P\{X>x\}$ and finite mean. The mean residual life function of $X$ at $t$ is

$$m(t) = \begin{cases} E[X-t \mid X>t], & t < t^*; \\ 0, & \text{otherwise,} \end{cases} \qquad t^* = \sup\{t : \bar F(t) > 0\}.$$

For a second random variable $Y$ on $(\Omega',\nu)$ with mrl function $l$, $X$ is smaller than $Y$ in the mean residual life order, $X \le_{mrl} Y$, if $m(t) \le l(t)$ for every $t$. The hazard rate order, restated in this mission's own namespace (Chapter 1's version cannot be imported — see Formalization scope), is the general, absolute-continuity-free comparison $\bar F(x)\bar G(y) \ge \bar F(y)\bar G(x)$ for all $x \le y$, where $\bar G$ is $Y$'s survival function. A random variable $X$ is DMRL (decreasing mean residual life) if its mrl function $m$ is decreasing in $t$.

Formalization targets

Goal — Theorem 2.A.2

$$\left(\frac{m(t)}{l(t)}\text{ increases in }t\right)\ \text{and}\ X \le_{mrl} Y \ \implies\ X \le_{hr} Y.$$

Combined with the companion milestone below, this is a genuine conditional equivalence: under the monotone-ratio hypothesis, $\le_{mrl}$ and $\le_{hr}$ coincide, and in particular $X \le_{mrl} Y \implies X \le_{st} Y$ under that condition. Without the hypothesis, the book states explicitly (the paragraph immediately preceding Theorem 2.A.1) that neither $\le_{st}$ nor $\le_{mrl}$ implies the other.

Milestones, in attack order

• Theorem 2.A.1. $X \le_{hr} Y \implies X \le_{mrl} Y$ — the one-directional link that motivates the goal theorem: the hazard rate order, strictly stronger in general, always implies the mean residual life order.
• Theorem 2.A.11. If $X$ is DMRL and $Z$ is a nonnegative random variable independent of $X$, then $X \le_{mrl} X+Z$ — one of the chapter's closure properties (§2.A.3): adding independent nonnegative noise to a DMRL random variable can only increase it in the mean residual life order.

Each milestone is stated exactly as the book states it: no constant is hard-coded, no $O(\cdot)$ or asymptotic approximation is involved, and the goal's monotone-ratio hypothesis is the genuine ratio $m(t)/l(t)$, not two separately-monotone functions (a different, unrelated condition the book itself does not state).

Significance

The mean residual life order sits at a specific point in the book's own hierarchy of orders: strictly implied by the hazard rate order (Theorem 2.A.1), and — the goal theorem — reversible into the hazard rate order under one extra monotonicity hypothesis on the ratio of the two mrl functions. This "sandwich" structure is exactly the kind of comparison-of-orders result that makes Chapter 1's usual and hazard rate orders (already formalized in Chunk 01 of this series, restated locally here since drafts cannot import each other) into a genuinely connected theory rather than a list of unrelated definitions. The DMRL closure property (Theorem 2.A.11) is separately significant: DMRL is one of the book's standard "aging" notions, used in reliability engineering to model components that wear out over time, and its preservation under adding independent noise is a basic tool for building compound reliability models (e.g. a component with an added, uncorrelated failure mode) from simpler DMRL parts.

No prior art exists on the platform for either order: GET /theorems?q=mean+residual+life returns zero hits, and GET /theorems?q=hazard+rate returns exactly one hit (DQJSQ.theorem2_ifr), an unrelated queueing-theory IFR (increasing failure rate) lemma about patience densities in a fluid queueing model, not this order — it names a different object under a coincidentally similar keyword and is not reused. This mission is a foundational island for the mean residual life order.

Difficulty

The mrl function is a genuinely two-case object: a real conditional expectation on $\{t : \bar F(t) > 0\}$, and a hard $0$ outside that region. The goal theorem's proof (not formalized here; only the statement is a milestone) differentiates $m$ and $l$, uses the identity $r(t) = m'(t)/m(t) + 1/m(t)$ relating the mrl function to the hazard rate, and compares the two resulting hazard-rate expressions using the ratio's monotonicity — a genuinely analytic argument, not a routine unfolding of definitions. The chief formalization difficulty is keeping the shape of $\le_{mrl}$ (a pointwise comparison of a derived function) visibly distinct from the function-class shape of $\le_{st}$ used in Chapter 1, since the book explicitly warns that conflating the two orders is a live error (neither implies the other in general) — see Formalization scope below for how each shape is kept separate.

Formalization scope

All three random variables in this mission's milestones are real-valued measurable functions on a MeasureTheory.Measure space, matching this series' Chapter 1 convention (Chunk 01). The mrl function mrl μ X t is defined as if 0 < P{X>t} then (∫ ω in {X>t}, (X ω - t) ∂μ) / P{X>t} else 0, formalizing the case split on $t < t^*$ directly via positivity of the survival probability (its defining equivalent under the survival function's monotonicity) rather than through the derived quantity $t^*$ itself. MrlOrder μ ν X Y is ∀ t : ℝ, mrl μ X t ≤ mrl ν Y t — a direct pointwise comparison of two functions, deliberately kept a different shape from Chapter 1's UsualOrder (a ∀ φ ∈ 𝒞, E[φ∘X] ≤ E[φ∘Y] function-class quantifier), since the book's own warning that $\le_{st}$ and $\le_{mrl}$ neither implies the other is a warning against treating them as interchangeable comparison shapes.

The hazard rate order is restated locally in this chapter's own namespace (StochasticOrders.MeanResidualLife.HazardRateOrder) rather than imported from Chunk 01's StochasticOrders.Usual.HazardRateOrder, because each chapter's mission is drafted and reviewed as an independent Prove2Me proposal and one draft cannot import another draft's unpublished Lean; its definition is identical in shape to Chunk 01's own restatement of the general, absolute-continuity-free survival-function form of $\le_{hr}$ (not the density-ratio form, which requires absolute continuity the book does not assume at this level of generality).

Every milestone that consumes mrl carries explicit Integrable hypotheses on the random variables involved (Integrable X μ, and Integrable Y ν or Integrable Z μ as applicable), formalizing the book's own standing "finite mean" hypothesis from §2.A.1's definition of the mrl function: without it, the Bochner integral inside mrl would return its junk value 0 for a non-integrable variable on some tail set, letting a hypothesis like MrlOrder μ ν X Y hold of a function that is not actually the book's mean residual life function. DMRL μ X is Antitone (mrl μ X), the book's own "$m(t)$ is decreasing in $t$" in the weak, non-strict monotone sense used throughout the book for "increasing"/"decreasing".

A trivializing formalization this mission rules out: stating the goal theorem with the ratio hypothesis as two separate monotonicity conditions on $m$ and $l$ individually (rather than genuine monotonicity of the ratio $m(t)/l(t)$ on the region where $l(t)>0$) would be a different, strictly stronger and easier-to-satisfy hypothesis than the book's own — the milestone here states MonotoneOn (fun t => mrl μ X t / mrl ν Y t) {t | 0 < mrl ν Y t}, the genuine ratio restricted to where the denominator does not vanish, matching Theorem 2.A.2's own "$m(t)/l(t)$ increases in $t$" verbatim.

Selected references

• M. Shaked and J. G. Shanthikumar, Stochastic Orders, Springer Series in Statistics, Springer 2007, Chapter 2 (Mean Residual Life Orders), §2.A. https://doi.org/10.1007/978-0-387-34675-5
• W. Whitt, "Uniform Conditional Stochastic Order," Journal of Applied Probability, 1980 (characterizations of IFR/DFR by the likelihood ratio order, cited by the book's remarks section as background for the chapter's aging notions).
• This series' Chunk 01 (StochasticOrders.Usual), for the usual and hazard rate orders this chapter's own restated definitions parallel.

Motivation

Every method in Chapters 2-4 of this series solves a projection or proximal subproblem at every step — a Euclidean projection, or a Bregman-divergence prox-mapping — which can itself be as hard as the original problem when $X$ is a complicated feasible set (a spectrahedron, a flow polytope, a matroid base polytope). The conditional gradient method (Frank & Wolfe, 1956) sidesteps this entirely: instead of a projection, each step calls a linear optimization (LO) oracle — minimize a linear function over $X$ — which is frequently far cheaper (over a spectrahedron, this reduces to a single eigenvector computation; over many combinatorial polytopes, to a greedy algorithm). This is the origin of the modern "projection-free" family of optimization methods widely used at the scale where projections are the bottleneck.

Setting

Fix a nonempty compact convex set $X$ in a real normed space $E$ and a convex $f:X\to \mathbb R$ with $L$-Lipschitz gradient (Eq. (7.1.4)): $\|f'(x)-f'(y)\|_*\le L\|x-y\|$. The classic conditional gradient (CndG) method, Algorithm 7.1, sets $x_0\in X$, $y_0=x_0$, and for $k=1,2,\dots$: calls the LO oracle $x_k\in\arg\min_{z\in X}\langle f'(y_{k-1}),z\rangle$, then sets $y_k=(1-\alpha_k)y_{k-1}+\alpha_kx_k$ for a stepsize $\alpha_k\in[0,1]$, either the fixed schedule $\alpha_k=2/(k+1)$ (Eq. (7.1.9)) or exact line search (Eq. (7.1.10)).

Section 7.1.1.2 extends this to bilinear saddle-point problems, where $f$ itself is the (generally nonsmooth) function $f(x)=\max_{y\in Y}\{\langle Ax,y\rangle-\hat f(y)\}$ (Eq. (7.1.5)) for a compact convex $Y$ and linear operator $A$. Since $f$ is nonsmooth, the method is applied instead to a family of smooth approximations $f_\eta$ built from a strongly convex $\omega$ on $Y$ (Eq. (7.1.21)-(7.1.23)), with the smoothing parameter $\eta_k$ allowed to vary across iterations rather than being fixed in advance.

Formalization targets

Goal — Theorem 7.1

$$f(y_k) - f^* \le \frac{2L}{k(k+1)}\sum_{i=1}^k\|x_i-y_{i-1}\|^2.$$

Supporting milestones, in attack order

• Lemma 7.1: the smoothed objective family $f_\eta$ is monotone nondecreasing in $\eta\ge0$ — the one-line fact ($V(y)-D_Y^2\le0$ pointwise) that licenses a variable, decreasing smoothing schedule $\eta_k$ rather than a schedule fixed in advance from knowledge of the target accuracy.
• Theorem 7.2: the saddle-point counterpart of the goal theorem, running the same CndG algorithm on the smoothed gradients $f_{\eta_k}'$ instead of $f'$ directly, with the explicit rate $f(y_k)-f^*\le\frac{2}{k(k+1)}\sum_{i=1}^k[i\eta_iD_Y^2+\frac{\|A\|^2}{\sigma_v\eta_i} \|x_i-y_{i-1}\|^2]$.

Every constant here is exactly the book's; the goal theorem's bound is left in terms of the actual step distances $\sum\|x_i-y_{i-1}\|^2$, not a diameter-based simplification (see Difficulty).

Significance

This mission formalizes the founding convergence result of the entire projection-free family (Frank-Wolfe methods), which has become central to large-scale machine learning precisely because its per-iteration cost can be orders of magnitude below that of a projection-based method on structured feasible sets. Theorem 7.1's specific form — a rate depending on the realized step distances rather than a fixed diameter — is also the more informative, tighter statement (the book's own remarks show it recovers the classical diameter-based $O(LD_X^2/\varepsilon)$ complexity as a corollary, but also explains why the rate can be much better in practice when the iterates settle near an extreme point).

No result matching conditional gradient / Frank-Wolfe methods exists on the platform as of 2026-09-18 (q=Frank-Wolfe and q=conditional gradient both return zero hits — see Prior art in MODERATION_NOTES.md).

Difficulty

The chief formalization difficulty is representing "with the stepsize policy in (7.1.9) or (7.1.10)" faithfully without either restricting to one policy (weaker than the book's stated theorem) or introducing an awkward disjunction of two separate algorithm definitions. The book's own proof resolves this by a single observation used for both policies at once: $f(y_k)\le f(\tilde y_k)$ for $\tilde y_k$ the point the fixed schedule $\gamma_k=2/(k+1)$ would have produced — trivially by equality under (7.1.9), or because $y_k$ is chosen to minimize $f$ over the entire line segment under (7.1.10), of which $\tilde y_k$ is one point. This mission's hyk_le hypothesis states exactly this shared consequence, which is genuinely what the proof uses and genuinely covers both policies, rather than picking one arbitrarily.

A second difficulty is not collapsing $\sum_{i=1}^k\|x_i-y_{i-1}\|^2$ into a diameter bound $kD_X^2$ inside the milestone itself — the book's own remarks perform that substitution as a separate, weaker corollary (Eq. (7.1.19)) after stating Theorem 7.1 in its sharper form; folding the substitution into the goal statement itself would silently prove a different, weaker theorem.

Formalization scope

conditional_gradient_rate and saddle_point_cndg_rate state the LO oracle's exactness (x k ∈ Argmin_{z∈X}⟨fGrad(y(k-1)),z⟩) as a pointwise hypothesis rather than deriving it from IsCompact X via an existence lemma — matching the pointwise-hypothesis convention this series uses throughout for argmin-defined algorithmic steps (chunk 03-deterministic's mirror-descent updates, chunk 04-stochastic's stochastic mirror-descent update). X compact convex is still included as a hypothesis, matching the book's own standing assumption on the problem class, even though it is not itself needed to derive the stated conclusion from the other hypotheses.

smoothed_objective_monotone and saddle_point_cndg_rate realize $f_\eta$/$f$ via sSup of the image of $Y$ under the pointwise saddle-point objective, matching the book's own max_{y∈Y}{...} definition (Eq. (7.1.5), (7.1.23)) directly rather than introducing a separate Def_ file for a "bilinear saddle-point objective" structure — no other item in this mission reuses that definition verbatim, so per this series' convention (no shared substrate bundled into a structure unless reused), it is inlined at each use.

A trivializing formalization this mission rules out: stating the LO oracle via an $\varepsilon$-approximate minimizer ((fGrad (y(k-1))) (x k) ≤ (fGrad (y(k-1))) z + ε for some ε) rather than an exact one — this is explicitly a different, weaker algorithm the book does not analyze in Theorem 7.1/7.2 (the book studies approximate LO oracles separately, later in the chapter, not selected here).

Left out of scope, for time: Theorem 7.7 (the matching lower complexity bound for LO-oracle methods, Eq. (7.1.60)) — formalizing it faithfully requires first modeling the abstract class of "LCP methods" (any algorithm restricted to LO-oracle calls) as a universally-quantified object, a substantially different and more involved formalization task than the two upper-bound convergence theorems selected here; named per Hard Rule 7 rather than approximated. The d(x)=\sum x_i\log x_i entropy-smoothing remark and the primal/primal-dual averaging CndG variants (§7.1.2, not covered by this mission's page range) are likewise not attempted.

Selected references

• G. Lan, First-Order and Stochastic Optimization Methods for Machine Learning, Springer Series in the Data Sciences, Springer 2020, Chapter 7, §7.1.1. https://doi.org/10.1007/978-3-030-39568-1
• M. Frank, P. Wolfe, "An algorithm for quadratic programming," Naval Research Logistics Quarterly, 3(1-2), 1956, pp. 95-110.
• M. Jaggi, "Revisiting Frank-Wolfe: projection-free sparse convex optimization," ICML, 2013 (the modern machine-learning revival of the method).

Motivation

Empirical-risk-minimization objectives in machine learning are finite sums: $\Psi(x) = \frac{1}{m}\sum_{i=1}^m f_i(x) + h(x)$, one smooth term $f_i$ per training example (or per worker, in a distributed setting), plus a simple nonsmooth regularizer $h$. Chapter 4's basic stochastic mirror descent handles this by sampling a single random component gradient $\nabla f_{i_t}(x)$ as an unbiased estimator of $\nabla f(x)$ — but that estimator's variance is a constant throughout the algorithm, which caps the achievable convergence rate. Variance-reduced mirror descent asks a sharper question: can an unbiased finite-sum gradient estimator be built whose variance itself vanishes as the algorithm approaches the optimum? The answer — periodic full-gradient snapshots combined with single-component corrections — is the SVRG-style idea this mission formalizes in Lan's general-norm mirror-descent framework, with an explicit, sampling-distribution-dependent constant rather than a generic $O(\cdot)$.

Setting

Fix a closed convex set $X$ in a real normed space $E$, and the finite-sum composite problem $\min_{x\in X}\{\Psi(x):=f(x)+h(x)\}$ (Eq. (5.3.1)), where $f(x)=\frac1m\sum_{i=1}^m f_i(x)$ is the average of $m$ smooth convex component functions, each with $L_i$-Lipschitz gradient $\nabla f_i$ ($\|\nabla f_i(x)-\nabla f_i(y)\|_*\le L_i\|x-y\|$), and $h$ is a simple, possibly nondifferentiable convex function. $f$ is possibly $\mu$-strongly convex, $\mu\ge0$ (Eq. (5.3.2)); this mission's goal takes $\mu=0$ (§5.3.1, "Smooth Problems Without Strong Convexity"). A fixed probability distribution $Q=\{q_1,\dots,q_m\}$ on the component indices governs the algorithm's random sampling, and

$$L_Q := \frac{1}{m}\max_{i=1,\dots,m}\frac{L_i}{q_i}$$

is the section's key aggregate smoothness constant (Eq. (5.3.4)), replacing the plain average $L$ wherever component-wise variance enters the analysis. Variance-reduced mirror descent (Algorithm 5.6) is a multi-epoch method: each epoch of length $T_s$ recomputes a full gradient $\nabla f(\tilde x)$ at a snapshot point $\tilde x$, then runs $T_s$ inner iterations using the estimator $G_t := \big(\nabla f_{i_t}(x_t)-\nabla f_{i_t}(\tilde x)\big)/(q_{i_t}m) + \nabla f(\tilde x)$ and the mirror-descent-with-composite-term update $x_{t+1}:=\arg\min_{x\in X}\{\gamma[\langle G_t,x\rangle+h(x)]+V(x_t,x)\}$, where $V$ is the Bregman divergence of a fixed distance-generating function, exactly as in Chapters 3-4.

Formalization targets

Goal — Corollary 5.8

With $\theta=1$, $\gamma=1/(16L_Q)$, and the doubling epoch schedule $T_1=7$, $T_s=2T_{s-1}$ (Eq. (5.3.17)),

$$\mathbb E[\Psi(\bar x_S)-\Psi(x^*)] \le \frac{8}{2^{S-1}}\left[\frac{11}{4}\big(\Psi(x_0)-\Psi(x^*)\big)+16L_Q\,V(x_0,x^*)\right]$$

for every epoch count $S\ge1$, where $\bar x_S$ is the weighted average of the epoch snapshots (Eq. (5.3.16)).

Supporting milestones, in attack order

• Lemma 5.12 — the per-component gradient-variation bound $\frac1m\sum_i\frac1{mq_i}\|\nabla f_i(x)-\nabla f_i(x^*)\|_*^2 \le 2L_Q[\Psi(x)-\Psi(x^*)]$, the basic smoothness consequence from which the estimator's variance bound is built.
• Lemma 5.13 — unbiasedness ($\mathbb E[\delta_t]=0$) and two variance bounds ($\mathbb E[\|\delta_t\|_*^2]\le 2L_Q[\dots]$ and $\le 4L_Q[\dots]$) for the variance-reduced estimator's error $\delta_t:=G_t-\nabla f(x_t)$.
• Lemma 5.14 — the one-step progress bound combining Lemma 5.13's variance control with the mirror-descent update's three-point inequality.
• Theorem 5.6 — the general epoch-level convergence bound (with an arbitrary epoch-length schedule $T_s$ and stepsize $\gamma$ satisfying $4L_Q\gamma\le1$) that Corollary 5.8 instantiates.

Every constant is exactly the book's: $L_Q$'s own sampling-distribution-dependent definition (never specialized to uniform $q_i=1/m$), and Corollary 5.8's explicit $8/2^{S-1}$, $11/4$, $16L_Q$ — not a generic $O(\cdot)$ — are all taken verbatim.

Significance

This is the series' first genuinely finite-sum result: unlike Chapters 3-4's single abstract objective $f$, here $f$ is structurally a named average of $m$ component functions, and the sampling distribution $\{q_i\}$ over those components is a first-class free parameter of both the algorithm and the analysis (not fixed to uniform sampling) — $L_Q$ itself depends on this choice, and a formalization that hard-codes $q_i=1/m$ would understate what Lemma 5.12's own proof needs. Getting Theorem 5.6/Corollary 5.8 right also requires keeping two nested indices straight: inner iterations $t$ within an epoch, and outer epoch counts $s$, with the convergence bound stated in terms of the epoch count $S$ alone — and keeping the two "gap" quantities $\Psi(x_0)-\Psi(x^*)$ (an objective-value gap) and $V(x_0,x^*)$ (a Bregman-divergence gap) distinct throughout, since they enter Corollary 5.8's final bound with different explicit coefficients ($11/4$ vs. $16L_Q$) and neither generically bounds the other.

No result on the platform models a finite-sum objective with $m$ named component functions sampled by a general index distribution $\{q_i\}$, a variance-reduction snapshot/anchor point, or this specific SVRG-style estimator, as of 2026-09-18 (q=finite sum, q=variance reduction, q=SVRG, q=component function, q=variance reduced gradient, q=mirror descent finite sum — see Prior art below).

Difficulty

The central difficulty is Theorem 5.6's own epoch-weight sequence $w_s$: the book defines $w_s:=(1-4L_Q\gamma)(T_{s-1}-1)-4L_Q\gamma T_s$ explicitly only for $s\ge2$ (Eq. (5.3.14)), yet the displayed sums $\sum_{s=1}^S w_s$ in (5.3.15)-(5.3.16) run from $s=1$. A 2026-09-19 revision found that this, combined with the epoch snapshot $\tilde x_s$ being constrained only by membership in $X$ and not tied to the algorithm's own dynamics, made the originally drafted statements false, not merely incomplete: an adversarial, unboundedly-large-$\Psi$, $\omega$-independent $\tilde x_1$ together with $w_1\to\infty$ violates the stated conclusion. The fix restores the connection via an auxiliary epoch-boundary sequence and the per-epoch progress inequality Theorem 5.6's own proof derives from Lemma 5.14 (see epoch_convergence_bound's hepoch hypothesis), and resolves $w_1$ by extending (5.3.14)'s domain to $s\ge1$ via a fixed "epoch 0" length $T_0$ — w_1 is no longer left free beyond positivity. finite_sum_variance_reduced_rate instantiates $T_0:=T_1/2=3.5$ concretely, reproducing the arithmetic Corollary 5.8's own proof is internally consistent with ($w_1 = 3/4(3.5-1)-1/4\cdot7 = 1/8$, matching the closed form $(1/8)T_1-3/4=1/8$) — this was previously only a documented-but-unresolved observation, not yet a stated hypothesis.

Formalization scope

All five items are stated over a general real normed space [NormedAddCommGroup E] [NormedSpace ℝ E], matching the mirror-descent chunks' general-norm convention (never specialized to Euclidean space or squared distance) — $V$ is a free two-point function throughout, and each $\nabla f_i$, $\nabla f$, $G_t$ are continuous linear functionals E →L[ℝ] ℝ, whose Mathlib operator norm supplies the dual norm $\|\cdot\|_*$ with no separate definition needed. This is the trivializing formalization this mission rules out: hard-coding $q_i=1/m$ (uniform sampling) or $V(x,y)=\frac12 \|x-y\|^2$ (Euclidean Bregman divergence) would understate both $L_Q$'s dependence on the sampling distribution (the whole point of Lemma 5.12's bound) and the general-norm apparatus the rest of this book series shares.

Ψ(x_0)-Ψ(x^*) and V(x_0,x^*) are kept as two syntactically distinct terms throughout — never conflated or bounded one by the other — matching Corollary 5.8's own two separate coefficients. Corollary 5.8's own explicit constants ($8/2^{S-1}$, $11/4$, $16L_Q$) are stated verbatim rather than left as an unspecified $O(\cdot)$, per Hard Rule 6.

Left out of scope, for time: the gradient-computation-count complexity bound (Eq. (5.3.19), an $O(\cdot)$ statement about total oracle calls, not a convergence-rate inequality on $\Psi$) and §5.3.2's strongly-convex case (Theorem 5.7, a geometric-decay bound $\Delta_s\le\rho\Delta_{s-1}$ under $\mu>0$) are natural continuations reusing this mission's variance_reduced_progress_bound milestone, not attempted here.

Prior art

q=finite sum, q=variance reduction, q=SVRG, q=component function, q=variance reduced gradient, and q=mirror descent finite sum were all searched on 2026-09-18. The only topically-adjacent hit across all six queries is ShiOptRates.Stochastic.variance_purchase_ classical ("Classical variance reduction is cost-neutral..."), which models plain minibatch SGD on a smooth objective with an i.i.d.-noise oracle characterized by a single scalar variance $\hat\sigma^2$ and a minibatch-size trade-off — no finite-sum structure with $m$ named component functions, no sampling distribution $\{q_i\}$, no snapshot/anchor point $\tilde x$, and a different question (cost-neutrality of minibatch size vs. this mission's convergence rate for a fixed variance-reduction scheme). Not reused; every item in this mission is drafted fresh.

Selected references

• G. Lan, First-Order and Stochastic Optimization Methods for Machine Learning, Springer Series in the Data Sciences, Springer 2020, Chapter 5, §5.3. https://doi.org/10.1007/978-3-030-39568-1
• R. Johnson, T. Zhang, "Accelerating stochastic gradient descent using predictive variance reduction," Advances in Neural Information Processing Systems (NeurIPS), 2013 (the SVRG estimator this section's gradient estimator generalizes to the composite mirror-descent setting).
• A. Nemirovski, A. Juditsky, G. Lan, A. Shapiro, "Robust stochastic approximation approach to stochastic programming," SIAM Journal on Optimization, 19(4), 2009, pp. 1574-1609.

Motivation

Machine learning's canonical training objective — minimize an expected or empirical risk over a data distribution — is almost never observed exactly: at each step an algorithm sees only a noisy gradient sample (a minibatch gradient, a single-example gradient, a simulation draw). Stochastic mirror descent (Nemirovski, Juditsky, Lan & Shapiro 2009) is the modern, general-norm answer to "what happens to first-order convergence guarantees when the gradient itself is a random variable": it takes the deterministic mirror-descent scheme of the previous chapter and replaces the exact subgradient with an unbiased stochastic estimate, and asks for both an expected convergence rate and, when the noise is well-behaved, an explicit probability-of-large-deviation guarantee. This is the theoretical backbone of stochastic gradient descent as used in practice.

Setting

Fix a nonempty closed convex set $X$ in a real normed space $E$, and a convex $f:X\to\mathbb R$ with $f^*:=\min_{x\in X}f(x)$ and $x^*$ an arbitrary minimizer, exactly as in Chapter 3. A stochastic oracle $G(x,\xi)$, queried at a point $x$ with a fresh random sample $\xi$, returns an estimate of a subgradient $g(x)\in\partial f(x)$: $\mathbb E[G(x,\xi)] = g(x)$ (unbiasedness), $\|g(x)\|_*\le M$ (a dual-norm Lipschitz bound, Eq. (4.1.7)), and $\mathbb E[\|G(x,\xi)-g(x)\|_*^2]\le\sigma^2$ (a second-moment/variance bound). The stochastic mirror-descent update is exactly Chapter 3's mirror-descent update with $G_t := G(x_t,\xi_t)$ in place of the deterministic $g_t$: $x_{t+1} := \arg\min_{x\in X}\gamma_t\langle G_t,x\rangle + V(x_t,x)$ (Eq. (4.1.6)), where $V$ is the Bregman divergence of a fixed distance-generating function $\nu$.

Formalization targets

Goal — Theorem 4.1

$$\mathbb E[f(\bar x^k_s)] - f^* \le \Big(\sum_{t=s}^k\gamma_t\Big)^{-1}\Big(\mathbb E[V(x_s,x^*)] + (M^2+\sigma^2)\sum_{t=s}^k\gamma_t^2\Big).$$

Supporting milestones, in attack order

• Lemma 3.4, invoked for the stochastic update: the same three-point inequality as the deterministic mirror-descent update, restated with the stochastic gradient functional $G_t$ in place of $g_t$ — the book's own remark ("It can be easily seen that the result in Lemma 3.4 holds with $g_t$ replaced by $G_t$") is exactly what licenses treating this as the same algebraic fact for a fixed sample path.
• Lemma 4.1: the martingale-difference deviation bound, a Chernoff-type concentration inequality for a conditionally sub-Gaussian martingale-difference sequence — the chapter's general-purpose probabilistic tool, proved independently of the optimization setting.

Every constant is exactly the book's; $M^2+\sigma^2$ (not a generic $O(\cdot)$) is the goal's own noise-dependent constant, taken verbatim.

Significance

This is the first mission in the series to leave the purely deterministic, real-analytic setting of Chapters 2-3 and formalize a genuinely probabilistic convergence guarantee: an expectation taken over an entire random algorithm trajectory $\xi_1,\dots,\xi_k$, not merely over a single random variable. Getting the goal theorem's statement right requires being explicit about exactly which quantities are random (the iterates $x_t$, hence $f(\bar x_s^k)$ and $V(x_s,x^*)$) and which are deterministic constants fixed in advance ($M,\sigma,\gamma_t$), and about the precise mathematical content of "the stochastic gradient's bias vanishes after conditioning on the past" — Lemma 4.1 is included specifically because it is the general machine that makes that vanishing rigorous, independent of the optimization application.

No result matching stochastic mirror descent, Assumption 4's sub-Gaussian/light-tail condition, or this martingale-difference concentration lemma exists on the platform as of 2026-09-18 (q= stochastic gradient, q=stochastic mirror descent, q=martingale, q=sub-Gaussian — see Prior art below for what these queries actually returned).

Difficulty

The central difficulty is disentangling which facts in the chapter's proof genuinely need measure theory and which do not. The per-step algorithmic relations — $x_{t+1}$'s minimality, $f$'s subgradient inequality at $x_t$, the dual-norm bound on $g$ — hold for every sample path individually and are formalized pointwise in $\omega$, exactly as chunk 03-deterministic formalizes its deterministic analogues; only the second-moment bound and the final expectation inequality are genuine integrals. The one place this pointwise treatment cannot simply mirror the deterministic case is the noise cross-term $\mathbb E[\gamma_t\langle\delta_t,x_t-x^*\rangle]=0$: in the book's proof this vanishes because $\delta_t=G_t-g(x_t)$ is conditionally mean-zero given the past and $x_t$ is a function of the past (the martingale-difference property, via the tower property of conditional expectation) — a genuinely non-pointwise fact. Rather than thread an explicit filtration through the goal theorem's own statement (which Lemma 4.1 already does, as the chapter's dedicated home for that machinery), the goal theorem takes this post-tower-property consequence directly as a named hypothesis (hcross); see Formalization scope.

Formalization scope

stochastic_mirror_iterate_three_point and stochastic_mirror_descent_bound are stated over a general real normed space [NormedAddCommGroup E] [NormedSpace ℝ E], matching chunk 03-deterministic's general-norm milestones (mirror_iterate_three_point/mirror_descent_bound) rather than the Euclidean/inner-product specialization of that chunk's §3.1 items — Chapter 4's own stochastic mirror descent is presented directly in the general-norm framework of §3.2, with no Euclidean-only warm-up. $V$ is left a free two-point function (never hard-coded to a squared Euclidean distance), and the stochastic gradient $G_t$ and the subgradient selector $g$ are continuous linear functionals E →L[ℝ] ℝ, whose Mathlib operator norm supplies the dual norm $\|\cdot\|_*$ with no separate definition needed — the same trivializing formalization chunk 03-deterministic rules out (specializing $V$ to the Euclidean case) applies here and is ruled out the same way.

martingale_difference_deviation_bound (Lemma 4.1) is a standalone probabilistic result, formalized with Mathlib's MeasureTheory.Filtration and condExp machinery: the sequence $\xi_{[t]}$'s generated filtration, $\zeta_t$'s $\mathcal F_t$-measurability, and the two conditional-expectation hypotheses (conditional mean zero, conditional sub-Gaussian tail) are all literal translations of the book's own E|ξ[t-1] notation.

Left out of scope, for time: Assumption 4 (the light-tail/sub-Gaussian oracle assumption), Proposition 4.1 (the large-deviation bound under Assumption 4, which chains Lemma 4.1's concentration bound with the constant stepsize policy (4.1.11) and a second Markov-inequality argument on $\sum\gamma_t^2\|\delta_t\|_*^2$), Lemma 4.2 and Theorem 4.2 (the smooth-$f$ case, §4.1.2, requiring a separate recursion and averaging convention $x_t^{av}$). All four are natural continuations reusing this mission's stochastic_mirror_iterate_three_point and/or martingale_difference_deviation_bound; a later mission or an amendment to this one could add them without touching what is here. Per Hard Rule 7 (faithfulness over coverage), a genuinely faithful formalization of Proposition 4.1 in particular — which needs Assumption 4's own conditional-MGF hypothesis threaded consistently with Lemma 4.1's, plus the constant-stepsize substitution and a second concentration argument — was judged to need more time than this session's budget allowed to do without shortcuts; it is named here rather than approximated.

Selected references

• G. Lan, First-Order and Stochastic Optimization Methods for Machine Learning, Springer Series in the Data Sciences, Springer 2020, Chapter 4, §4.1. https://doi.org/10.1007/978-3-030-39568-1
• A. Nemirovski, A. Juditsky, G. Lan, A. Shapiro, "Robust stochastic approximation approach to stochastic programming," SIAM Journal on Optimization, 19(4), 2009, pp. 1574-1609.
• H. Robbins, S. Monro, "A stochastic approximation method," Annals of Mathematical Statistics, 22(3), 1951, pp. 400-407 (origin of stochastic approximation).

Motivation

A multistage stochastic program's exact deterministic equivalent grows exponentially with the number of periods, even when each period's random data takes only a handful of values (Chapter 9's concern was the growth in the number of realizations; Chapter 10 adds growth in the number of periods). One remedy, generalizing Chapter 8's single-period Jensen bound, is to replace the exact per-period random data by a coarser, aggregated version — conditional expectations over a partition of the history space at each stage — and solve the resulting smaller deterministic equivalent instead. This is only useful if the aggregated problem's optimal value is provably a bound (here, a lower bound) on the exact problem's, and Birge & Louveaux's Chapter 10, §10.1, Theorem 1 is exactly the statement that makes this legitimate, together with a genuinely necessary extra condition the book states explicitly two paragraphs before the theorem: "if not [i.e. if the extra condition fails], then the conditional expectation form ... may not actually achieve a bound." This mission formalizes that theorem.

Setting

The book's exact multistage stochastic linear program (Eq. 1.1, p. 418) is

min c¹x¹ + E_Ω[c²x² + ⋯ + cᴴxᴴ]
s.t. W¹x¹ = h¹,  Tᵗ⁻¹xᵗ⁻¹ + Wᵗxᵗ = hᵗ (t=2,…,H, a.s.),  xᵗ ≥ 0 a.s., xᵗ nonanticipative (Σᵗ-measurable),

over the exact event space Ω = Ω₁ × ⋯ × Ω_H. Given a consistent nested partition of each Ωᵗ = Ω₁ × ⋯ × Ωₜ into finitely many blocks Sᵗ₁, …, Sᵗ_νₜ, and aggregated data (h̄ᵗᵢ, T̄ᵗᵢ) = E^{Sᵗᵢ}[(hᵗ,Tᵗ)] (the conditional expectation of the true random data over block i), the aggregated problem (Eq. 1.2, p. 419) replaces the exact recursion by a finite tree of blocks, one decision per block, linked to its parent block's decision. Both (1.1) and (1.2) are, structurally, the same kind of object — a finite-tree deterministic-equivalent recourse LP — differing only in which tree and which node data they use; this mission formalizes that shared shape once (Tree, Instance, Feasible, obj) and instantiates it twice.

Formalized as: a shared Tree H structure (a finite node type, per-node stage, anc, and a root), the same representation Chunk 06's Multistage.Tree uses for the exact scenario tree of its own (different) chapter, restated here rather than imported (a draft cannot import another chunk's draft). An Instance H n m T bundles a tree's node-varying LP data (c, W, Tmat, h, p); Feasible/obj give its feasible set and objective. The exact problem (1.1) is Instance H n m TFine for a fine/exact tree TFine; the aggregated problem (1.2) is Instance H n m TCoarse for a coarser tree TCoarse, connected to TFine by an aggregation map agg : TFine.Node → TCoarse.Node.

Formalization targets

Goal — Chapter 10, Theorem 1 (p. 419)

agg respects the tree structure (root, stage, ancestor);
W, c agree between the fine and coarse instances (up to agg);
coarse.h, coarse.Tmat are the p-weighted conditional expectations of fine.h, fine.Tmat over
  each aggregation fiber;
∀ coarse nodes i,i' at the same stage sharing a "current-period outcome",
  coarse.h i = coarse.h i' ∧ coarse.Tmat i = coarse.Tmat i'
  ⟹ zCoarse ≤ zFine

This is the mission's only formalization target: BRIEF.md records that no separately numbered lemma precedes Theorem 1's proof in this section to serve as an independent milestone (the proof is a direct LP-duality argument against the theorem's own hypotheses), and that Chapter 8's Theorem 1 — the two-period case this theorem generalizes — is a cross-chapter dependency belonging to Chunk 08's own mission, not a milestone here. milestones.yaml is accordingly empty; see STATUS.md for the explicit accounting of what else in this chapter was considered and left out (Theorem 3, the aggregation error bound of §10.2, an unrelated and substantially heavier result).

Significance

Theorem 1 is what licenses every aggregation-based approximation scheme the rest of the book's multistage material builds on: it says precisely when replacing a multistage recourse problem's random data by within-period conditional expectations preserves a valid lower bound, and precisely identifies the condition (aggregated nodes sharing a current-period outcome must carry identical aggregated data) whose failure breaks the bound — a condition the book states is not decorative ("if not, then the conditional expectation form ... may not actually achieve a bound," p. 418). Formalizing it gives Prove2Me a first structural result connecting Chapter 8's single-period Jensen bound (Chunk 08) to genuinely multistage approximation, using the same finite-scenario-tree deterministic-equivalent representation Chunk 06 uses for the exact nested Benders decomposition — the two missions' shared representation choice (documented in both STATUS.md files) means a future mission relating them formally (e.g. instantiating Chunk 06's exact tree as this mission's TFine) has a compatible object to work with, even though neither imports the other's draft.

Difficulty

The theorem's proof (p. 419-420) is a direct LP weak-duality argument: given an optimal dual solution to the aggregated problem, the book constructs a dual-feasible solution to the exact problem attaining the same value, using precisely the "common outcome ⟹ equal aggregated data" hypothesis to make the constructed dual solution well-defined across the exact tree's finer structure. This is a real argument, not a citation, but it is left as sorry: formalizing the proof would need the multistage LP duality machinery (the "multistage version of Theorem 3.13" the book's own proof invokes, itself left as Exercise 1) that no chunk of this series has built. The value of this mission is the faithful statement of the bound and its exact hypotheses.

Formalization scope

• The book's own printed typo, resolved and documented. Theorem 1's hypothesis clause reads, as printed, "such that (ωt−1,ωt) ∈ Stj if and only if there exist some (ω̂t−1,ωt) ∈ Stj" — S^t_j appears on both sides of the "if and only if," where the sentence's own subject ("S^t_i and S^t_j that have a common outcome") requires the left side to range over S^t_i. Confirmed against a direct render of PDF page 436 (uv run --with pymupdf python), not assumed from OCR: the PDF's own typesetting has this repetition, not an artefact of text extraction. This formalization reads the corrected clause as "S^t_i and S^t_j project onto the same set of period-t outcomes" and states it via an explicit label type Θ and curOutcome : TCoarse.Node → Θ, since the aggregated tree alone does not carry a literal per-period outcome space to project onto (see Setting above — Tree records only history-node structure, not the underlying product space Ω = Ω₁ × ⋯ × Ω_H).
• W, c shared exactly, not aggregated, matching the book's explicit assumption that the recourse matrix and per-stage cost are deterministic and identical across (1.1) and (1.2) ("Wt known and not random," "ct = ct," p. 418) — formalized as direct equality hypotheses (hW_agree, hc_agree) rather than folding W/c into the conditional-expectation machinery that h/Tmat go through.
• zFine/zCoarse are hypothesis-characterized, not sInf-defined, avoiding the real infimum's junk value 0 on an unbounded-below or empty feasible set (reference/FAITHFULNESS_TRAPS.md trap 5) — neither tree-LP's feasible set is shown bounded or nonempty by the hypotheses alone.
• The conditional-expectation defining equations are weighted, p·h/p·Tmat, not h/Tmat alone, matching the book's own E^{Sti}[·] = (h̄ti,T̄ti) read as "the fiber-sum of p·(h,T) equals p_i·(h̄ti,T̄ti)" — the standard definition of a conditional expectation against counting measure on a finite partition. Instance's own hp_pos (every node's probability is strictly positive) rules out the degenerate case a bare unweighted equation would need to guard separately (a coarse node of probability 0, which cannot occur, is what the read-back of this theorem flags as the one case where the weighted equation would not pin down h_coarse/ Tmat_coarse themselves — moot here since hp_pos excludes it).
• Trivialization risk (this chapter's own). A formalization that let coarse.h/coarse.Tmat be arbitrary constants unrelated to fine.h/fine.Tmat (dropping the conditional-expectation defining equations) would still typecheck a "lower bound" conclusion but assert nothing about aggregation — exactly the risk BRIEF.md flags: "a formalization that treats (h̄ti,T̄ti) as arbitrary constants rather than as conditional expectations over a partition of the scenario space at time t loses the theorem's actual content." Both hCoarse_h/hCoarse_T (the defining equations) and hCommonOutcome (the theorem's own extra hypothesis) are load-bearing and present.

Selected references

• Birge, J.R., Louveaux, F. Introduction to Stochastic Programming, 2nd ed., Springer 2011, Chapter 10, §10.1 (pp. 417-420), Theorem 1 (p. 419).
• Birge, J.R. "Decomposition and partitioning methods for multistage stochastic linear programs." Operations Research 33 (1985), 989-1007 — the source Chapter 10's aggregation bounds draw on (cited in §10.2, the neighboring section this mission does not formalize).

Motivation

Two-stage stochastic programs with recourse — choose a first-stage decision $x$ now, observe a random outcome $\xi$, then choose a second-stage recourse decision $y(\xi)$ to repair whatever $x$ left infeasible or suboptimal — are the workhorse model of the field, used for capacity planning, inventory and financial portfolio problems since the 1950s (Dantzig 1955; Beale 1955). When $\xi$ ranges over a finite set of scenarios, the recourse function $Q$ that averages the second-stage cost over scenarios is piecewise linear and convex in $x$, so the overall problem is itself a large linear program — but one whose constraint matrix has a scenario for every column block and can be far too large to hand to a general-purpose LP solver directly. Van Slyke and Wets' L-shaped method (1969), the subject of this mission, is the algorithm that made two-stage recourse problems with finite scenario sets practically solvable: it is Benders decomposition specialized to this block structure, alternating between a small master program over $x$ (and a scalar $\theta$ approximating the recourse cost) and, at each candidate $x$, a batch of second-stage linear programs that either certify $x$'s second-stage feasibility or supply a linear underestimate — a cut — of $Q$ around $x$. Birge & Louveaux's Introduction to Stochastic Programming (2nd ed., Springer 2011), Chapter 5 §5.1, gives the algorithm and proves its two central guarantees: a shortcut feasibility test for a special case (Theorem 1) and the algorithm's finite convergence in general (Theorem 2), which is this mission's goal.

Setting

A two-stage recourse instance consists of a first-stage feasible region $K_1 = \{x \mid Ax = b,\ x \ge 0\}$ for $x \in \mathbb{R}^{n_1}$, and, for each of $K$ finite scenarios $k = 1, \dots, K$ (occurring with probability $p_k$), second-stage data $(q_k, h_k, T_k)$ defining the recourse subproblem

$$Q(x, \xi_k) = \min_{y \ge 0} \{ q_k^\top y \mid W y = h_k - T_k x \},$$

where the recourse matrix $W$ is fixed — the same across every scenario, the case this chapter treats. $K_2 = \{x \mid Q(x,\xi_k) < \infty \text{ for all } k\}$ is the set of $x$ for which every scenario's subproblem is feasible, and the two-stage problem is

$$\min_{x} \ c^\top x + Q(x) \quad \text{s.t.} \quad x \in K_1 \cap K_2, \qquad Q(x) = \sum_{k=1}^K p_k\, Q(x, \xi_k).$$

A basis of the recourse subproblem is an injective choice of $m_2$ of $W$'s columns (where $m_2$ is $W$'s row count); each basis $b$ determines a simplex multiplier $\pi = (W_b^\top)^{-1} q_b$, and when $b$ attains the true optimum of $Q(x,\xi_k)$, LP duality gives $Q(x,\xi_k) = \pi^\top(h_k - T_k x)$ — the mechanism that turns a batch of second-stage LP solves into linear cuts on $x$.

Formalization targets

The L-shaped algorithm proceeds in three steps, repeated until neither applies:

• Step 1 solves the current master program (the $K_1$-feasible $x$, plus $\theta$ once at least one optimality cut exists, minimizing $c^\top x + \theta$ subject to every cut recorded so far — or just $c^\top x$ over $K_1$ before the first optimality cut, matching the book's convention that $\theta$ "is set equal to $-\infty$ and is not considered" until then).
• Step 2 tests each scenario's second-stage feasibility at the Step-1 optimum via an auxiliary LP; if some scenario fails (the LP's optimal value is positive), its optimal basis yields a feasibility cut and the algorithm returns to Step 1.
• Step 3, once every scenario is feasible, checks whether $\theta$ already dominates the true recourse cost at $x$ (using each scenario's optimal basis via LP duality); if not, an optimality cut is added and the algorithm returns to Step 1; if so, $x$ is optimal and the algorithm stops.

Goal — Chapter 5, Theorem 2 (p. 198)

$$\text{When } \xi \text{ is a finite random variable, the L-shaped algorithm finitely converges to}$$ $$\text{an optimal solution when it exists, or proves } K_1 \cap K_2 = \varnothing.$$

Formalized as: starting from the empty cut set, there is a finite-length run of the algorithm's Step-1/2/3 transition relation, of length bounded by the total number of distinct feasibility- and optimality-cut witnesses available, ending at a state admitting no further step — at which point either the master program has become infeasible (certifying $K_1 \cap K_2 = \varnothing$) or its optimum is second-stage feasible, passes every fresh Step-3 test, and is optimal for the two-stage problem.

Milestone — Chapter 5, Theorem 1 (p. 194)

$$\text{If } T \text{ is deterministic, } W \text{ is such that every } t \ge 0 \text{ lies in } \mathrm{pos}\,W,$$ $$\text{and } a = \min_k h_k \text{ (componentwise) is attained by some scenario } h_\ell,$$ $$\text{then } x \in K_2 \iff \exists\, y \ge 0,\ Wy = a - Tx.$$

A shortcut avoiding $K$ separate feasibility LPs at Step 2: under these structural assumptions on $W$, checking feasibility at the single componentwise-worst right-hand side certifies feasibility at every scenario simultaneously.

Significance

Van Slyke and Wets' method (and Benders decomposition more generally, of which it is the recourse-problem specialization) underlies essentially every large-scale two-stage stochastic program solved in practice, and its finite-convergence guarantee — not merely that an optimum exists, but that this specific cutting-plane procedure reaches it in finitely many outer iterations — is what makes the method a decision procedure rather than a heuristic. The proof's content is an explicit finiteness argument (the number of distinct simplex bases of the recourse subproblem and the feasibility-test LP is finite, so the algorithm cannot generate infinitely many distinct cuts before either exhausting the feasible region or converging), not a general compactness or fixed-point argument; formalizing it means formalizing the cutting-plane mechanism itself as a transition system and proving termination combinatorially, over the finite type of available bases, rather than proving only that some optimal $x$ exists.

Difficulty

The natural shortcut — state only "an optimal $x$ exists, or $K_1 \cap K_2 = \varnothing$" — is not Theorem 2's actual content and is not what this mission targets: that weaker claim would already follow from $K_1 \cap K_2$ being a nonempty polyhedron (or empty), with no reference to the algorithm at all, and would not require the finiteness-of-bases argument the book's proof turns on. The genuine difficulty is representing Steps 1-3 faithfully as a relation on accumulating cut sets, and pinning the termination bound to the actual combinatorial object the book cites (the finite set of bases of the two LPs the algorithm solves at each iteration) rather than to a numeral or an abstract compactness bound. A second, quieter difficulty is Step 1's own optimum: once optimality cuts exist, the master program optimizes $c^\top x + \theta$ jointly, but before the first one it optimizes $c^\top x$ alone; conflating the two (e.g. always requiring $\theta$ to be part of the optimum) does not match Step 1 as the book states it.

Formalization scope

First-stage and second-stage vectors are Fin n1 → ℝ / Fin n2 → ℝ; the finite scenario set is Fin K with probability vector p. A basis is {b : Fin m2 → Fin n2 // Function.Injective b} (m2 = the recourse matrix's row count), matching "an injective choice of $m_2$ columns of $W$"; its finiteness is definitional, from Fin m2 → Fin n2 being finite. Simplex multipliers use Matrix.inv, whose junk value 0 on a singular matrix is never reachable in a proof because multipliers are only ever used through an IsOptimalAt/IsFeasBasisOptimalAt hypothesis that pins the basis to one genuinely attaining the LP's true optimum. The recourse value $Q(x,\xi_k)$ is EReal-valued (reusing this series' Instance/QVal convention from Chunk 03), so an optimality-cut witness's claimed value is compared to it by an explicit EReal cast, never by EReal arithmetic. The algorithm's state is a pair of finite sets of witnesses recorded so far (Finset (Fin K × FeasBasis n2 m2) × Finset (Fin K → Basis n2 m2)); Step is an inductive relation with one constructor per Step-2 and Step-3 branch, each requiring its witness not already recorded, and the goal states a bounded-length Step-path from the empty state to a state admitting no further Step. This mission does not restate Chapter 3's polyhedrality fact about $K_2$ as a separate lemma: the finiteness fact it is invoked for is already exposed directly and structurally by the finite Fintype bound on the number of bases, so no additional axiom stands in for it (see MODERATION_NOTES.md). Lemmas 3-9 and Theorem 10 of §5.2 (Regularized Decomposition, a different algorithm) are out of scope. The trivializing formalization this mission rules out is exactly the one named under Difficulty above: a bare existence-of-optimal-or- infeasible-$x$ statement with no reference to Steps 1-3 or to a finite bound on the number of iterations — such a statement would be true of any nonempty polyhedron and would not be Theorem 2.

Selected references

• R. Van Slyke and R. Wets, L-Shaped Linear Programs with Applications to Optimal Control and Stochastic Programming, SIAM Journal on Applied Mathematics, 17(4), 1969, pp. 638-663. https://doi.org/10.1137/0117061
• J. Birge and F. Louveaux, Introduction to Stochastic Programming, 2nd ed., Springer, 2011, Chapter 5. https://doi.org/10.1007/978-1-4614-0237-4
• G. Dantzig, Linear Programming under Uncertainty, Management Science, 1(3-4), 1955, pp. 197-206. https://doi.org/10.1287/mnsc.1.3-4.197