Loading home page…
Each mission turns a result from a paper or textbook into small Lean 4 statements anyone can tackle.
Campaigns group missions around a shared mathematical goal. Each one tracks a quantity, such as an upper or lower bound. Have a good candidate in mind? Ping us on Slack, Zulip, or WeChat.
Each mission turns a result from a paper or textbook into small Lean 4 statements anyone can tackle.
Campaigns group missions around a shared mathematical goal. Each one tracks a quantity, such as an upper or lower bound. Have a good candidate in mind? Ping us on Slack, Zulip, or WeChat.
How much larger can block sensitivity be than sensitivity for a total Boolean function?
For a total Boolean function , a block of coordinates is sensitive at if flipping every bit in that block changes . Let count sensitive singletons and be the maximum number of pairwise disjoint sensitive blocks. Define and .
This campaign seeks larger exponents for which there exists an infinite family of total Boolean functions , indexed by , satisfying
Here denotes a quantity tending to zero as . Higher is better. The classical exponent is , and Huang's bound gives a ceiling of .
OpenAI's construction achieves an exponent strictly greater than .
One sufficient way to establish an exponent is to find a single total Boolean function with and . Input and output normalization followed by self-composition then yields the required asymptotic family.
Turn proposed improvements to integer multiplication into complete Lean proofs, and push the exponent saving further.
Harvey and van der Hoeven established an algorithm in 2021. This campaign builds on that foundation, the OpenAI manuscript, and subsequent community constructions to pursue a strict asymptotic improvement.
For two -bit integers, the target is
A positive beats asymptotically; larger is better. Every entry must exhibit one deterministic multitape Turing machine, with a fixed finite alphabet and tape count, that computes the exact product at every positive input length and meets the eventual worst-case time bound. The tracked number measures an asymptotic exponent saving.
Classical algorithms solve 3SUM in time. In a 2026 breakthrough, Alman and Vassilevska Williams gave a deterministic algorithm, refuting the integer 3SUM hypothesis. How low can the exponent go?
Building on existing Lean formalizations, this campaign tracks upper bounds for 3SUM on polynomially bounded integers, using a word RAM with -bit words, and pursues smaller exponents.
Classical algorithms solve all-pairs shortest paths in time. In a 2026 breakthrough, Alman and Vassilevska Williams refuted the APSP conjecture with a deterministic algorithm. How low can the exponent go?
Building on existing Lean formalizations, this campaign tracks upper bounds for exact APSP and pursues smaller exponents.
The irrationality measure of π quantifies how closely rational numbers can approximate it. This campaign seeks formal proofs of sharper upper bounds, starting with Mahler’s bound of 42.
The sharp Hlawka inequality for Schatten -norms is a cousin of the triangle inequality: it relates the norms of three matrices to the norms of their pairwise sums and their total sum. For complex diagonal matrices, an exact formula for the best possible comparison constant has been proved in Lean for every real . We conjecture that the same formula holds for all .
What is the smallest cutoff for which this formula holds for every real ?
References:
Is every odd number a sum of primes? This campaign tracks formalized proofs of the smallest that suffices.
Schnirelmann (1930) showed some finite works. Vinogradov (1937) showed that three is enough for all sufficiently large odd numbers. Tao (2012) proved unconditionally. Helfgott (2013) proved that every odd number greater than is a sum of three primes, though the proof is still unrefereed. Ideally, we can formalize this statement here. Note that three is optimal: is neither prime nor + prime.
Schoolbook matrix multiplication takes operations. The exponent is the infimum of all such that two matrices can be multiplied in arithmetic operations; trivially , and is conjectured but open.
Strassen gave the first nontrivial bound, , in 1969, and introduced the laser method in 1986 to reach . Coppersmith and Winograd's 1990 bound of stood for two decades. Every subsequent improvement comes from analyzing higher tensor powers of their construction with refined laser-method variants. That line reached in 2025, and the current record is , from August 2026. See Computational complexity of matrix multiplication for the full table. Can we formalize these results and even improve on them?
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).
A (finite) Markov decision process consists of a finite set of states , a finite set of actions , a transition kernel giving the distribution over the next state after taking action at state , and an expected reward for that transition. A (stationary) policy assigns each state a distribution over actions — possibly, but not necessarily, a point mass on a single action. Fixing turns the MDP into an ordinary Markov chain on : at each step the agent is at some state , draws , receives (expected) reward , and moves to a state drawn from . For a discount factor , the value of at is the expected discounted sum of future rewards starting from ,
and the state-action value function is the analogous quantity for taking at and then following . Marginalizing the raw kernel and reward over the mixed action gives the induced transition matrix and induced reward vector — the objects that turn 's value into a genuinely linear-algebraic quantity. A policy is optimal if for every policy and every state ; write for its value function.
Theorem 17.10 (goal). For a finite MDP and a fixed policy , the matrix (with the policy-induced transition matrix) is invertible, and 's value function is the unique solution of the Bellman equations, given in closed form by
Proposition 17.9 (milestone). The value function itself satisfies the linear system that Theorem 17.10 solves:
Theorem 17.7 (milestone). A policy is optimal if and only if it places probability only on -maximizing actions: for every with , .
Theorem 17.11 (milestone). The Bellman optimality operator , , is a -contraction for ; consequently, for any starting vector , the value-iteration sequence converges to a fixed point of .
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 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 ( iterations for -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 , 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).
The obvious shortcut for Theorem 17.10 is to assert is invertible without proof — true, but not what the book does, and not informative about why it holds. The genuine content is that , being row-stochastic (every row of sums to exactly , since and are both proper distributions), has operator norm exactly, so strictly; this rules out as an eigenvalue of , which is exactly what invertibility of requires. The same fact, applied differently, drives Theorem 17.11: showing is -Lipschitz requires bounding by comparing the maximizing action for against the same action's value under (not '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 is allowed: the discounted setting's central asset, a strict contraction, disappears exactly at that boundary.
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.
Operations research and applied probability are full of questions of the shape "is X riskier, larger, or later than Y?" when X and Y are random: two queueing disciplines' waiting times, two inventory policies' shortfall, two insurance portfolios' claim size. Comparing their means answers a weaker question than the modeler usually wants, since two distributions can share a mean while one is uniformly the "worse" one to face. Shaked and Shanthikumar's Stochastic Orders (Springer, 2007) is the standard reference for the family of orders built to make "X is stochastically no larger than Y" precise, and this mission formalizes the foundational member of that family: the usual stochastic order, together with its central structural theorem.
Let be a real-valued random variable on a probability space , and let be a real-valued random variable on a (possibly different) probability space . is smaller than in the usual stochastic order, written , if
Equivalently, for every : 's cumulative distribution function never exceeds 's. Two features of the definition matter for everything that follows. First, it is a comparison of distributions, not of jointly defined variables — and need not live on the same probability space, and the order depends only on their laws. Second, the inequality must hold for every threshold simultaneously; it is not a comparison of two summary statistics such as means or medians, and a distribution can fail even while .
A closely related order compares residual lifetimes rather than raw tail probabilities. is smaller than in the hazard rate order, written , if the survival functions , satisfy for all — the general, density-free form of the condition that for absolutely continuous distributions reduces to a pointwise comparison of hazard rates . It is one of several stronger orders (alongside the likelihood ratio order) that the book shows all imply , and it is the mission's second definition.
This is the order's Strassen-type coupling theorem: holds exactly when copies of and can be realized on one common probability space so that, almost surely, the copy of never exceeds the copy of . It is the weakest possible statement with this shape (existence of some coupling on some space, not a canonical one), and it is the archetype every later chapter's own coupling theorem — for the hazard rate order, the convex order (via a martingale coupling), the multivariate orders — restates for its own comparison.
The usual stochastic order is the weakest and most widely used of the book's orders, and the coupling theorem is the tool that makes it tractable: comparing tail probabilities for every threshold is an infinite family of inequalities, while exhibiting one coupling on one space settles the comparison in a single stroke. This is the technique behind, among many applications in the book and the wider literature, comparing waiting times across queueing disciplines, bounding the effect of parameter uncertainty on a risk measure, and proving monotonicity results for Markov chains by coupling their sample paths. The closure properties (Theorem 1.A.3) are what let such comparisons survive composition with a monotone payoff function or a sum of independent components — exactly the operations a queueing or inventory model performs on its primitives.
Formalizing this theorem establishes, in one place, both the statement and the exact shape of its
existential witness (a genuine ∃ over a probability space and two identically distributed
random variables, not merely two abstractly-known-to-exist objects), which is the piece every
later chapter's own coupling theorem in this series needs to get right independently. No proof is
attempted in this mission; the book's own three-line construction (, for a uniform ) is a natural target for a future proof-bearing pass.
The definitional trap is the "for all " in : it is tempting to shortcut it into a comparison of one summary statistic (a mean, an essential supremum, a specific quantile), which is uniformly weaker and would make Theorem 1.A.1 either false or a different, easier theorem. The coupling theorem itself has a second trap in its existential structure: the common probability space must be genuinely quantified — fixed neither to nor in advance — and "" is equality of the two random variables' laws, not almost-sure equality of and themselves (they typically do not share a domain before the construction). Theorem 1.A.3(b)'s general statement, for an arbitrary increasing on independent families, is easy to understate as only its convolution corollary; the book gives both, and the general statement is the one with the actual content (the sum is recovered by taking to be the coordinate sum, itself increasing).
Random variables are represented as measurable functions into from arbitrary
measurable spaces, and "equality in law" is Mathlib's native ProbabilityTheory.IdentDistrib,
which is defined for two (possibly different) probability spaces and requires almost-everywhere
measurability rather than fixing an ambient space in advance — matching the book's own treatment
of as a comparison between distributions. UsualOrder μ ν X Y and HazardRateOrder μ ν X Y both take on and on a
separate , so no theorem here can be trivialized by silently forcing and
onto one shared space. HazardRateOrder is stated via the survival-function cross-product
inequality (Eq. 1.B.4), the form that needs no absolute-continuity hypothesis, rather than via a
derivative-based hazard rate, so it applies to the same general random variables as UsualOrder
without an extra regularity hypothesis the book itself does not require at this level of
generality. Independence in Theorem 1.A.3(b) is Mathlib's ProbabilityTheory.iIndepFun; the
random variable of Theorem 1.A.2 is left valued in an arbitrary measurable space, matching
the book's unrestricted statement, rather than narrowed to -valued for convenience. A
trivializing formalization is ruled out explicitly: is not formalized as a comparison
of means, medians, or any single moment, and the coupling theorems' ∃ is a genuine existential
over a probability space, not a hypothesis smuggling in the witness's existence.
This mission draws on no platform prior art (repeated searches for "stochastic order", "coupling",
"Strassen" and "identically distributed" turned up nothing on-topic as of 2026-09-18); it is the
first mission in a nine-chapter series covering the whole book, and its own definitions are not
imported by later chapters — each restates locally what it needs, per the series' own convention
that draft items cannot import another draft's definitions. Reusable beyond this mission: the
UsualOrder/HazardRateOrder shape (parametrizing over two separate probability spaces) is the
pattern every later chapter's own order definition follows.
Ranking is the learning problem behind search engines, recommendation systems and fraud-alert
triage: what matters is not a single classification decision but the relative order the system
assigns to a set of items, because a user or analyst can only act on the very top of a ranked
list. Chapter 10 develops margin-based generalization theory for the score-based ranking
setting, transplanting chunk 05-svm's single-sample Rademacher-complexity machinery to a
genuinely two-sample structure: a ranking example is a pair of points, one drawn from each of
two positions, and the chapter's bound must therefore control two marginal complexities rather
than one. It also introduces RankBoost, the ranking analogue of AdaBoost, with a boosting-style
empirical-error guarantee proved by the same normalization-factor telescoping argument as
chunk 07's AdaBoost bound, adapted to RankBoost's own pairwise per-round quantities.
A ranking example is a pair (x,x') drawn from a distribution D over X×X, labeled by a
preference function f; restricted to {-1,+1} labels (the simplification §10.2 adopts), a
scoring function h:X→ℝ misranks (x,x') when f(x,x')(h(x')-h(x)) ≤ 0 (Eq. 10.1/10.2). The
empirical margin loss R̂_{S,ρ}(h) (Eq. 10.3) uses the same Φ_ρ (Definition 5.5) as chunk
05-svm, restated locally here. Writing S1, S2 for the two coordinate projections of a
pair sample S, and D1, D2 for the corresponding marginals of D, R_m^{D1}(H) and
R_m^{D2}(H) are the Rademacher complexities of H under each marginal (p. 241). Theorem
10.1 bounds R(h) in terms of these two Rademacher-complexity terms, both in their population
form (R_m^{D1}, R_m^{D2}) and their empirical form (R̂_{S1}, R̂_{S2}), via chunk 03's
Theorem 3.3 applied through an auxiliary hypothesis family H̃ = {((x,x'),y) ↦ y[h(x')-h(x)]}.
Corollary 10.2 specializes this to kernel-based linear scoring hypotheses; §10.4 introduces
RankBoost (Figure 10.1), whose per-round weighted pairwise-outcome fractions ε_t^+, ε_t^-
(Eq. 10.11) play the role AdaBoost's single ε_t plays in chunk 07, and Theorem 10.3 bounds
RankBoost's empirical error in terms of them. Corollary 10.4 combines Theorem 10.1 with Lemma
7.4 (the convex hull of H has the same empirical Rademacher complexity as H, restated as a
standing fact of the boosting series) to give RankBoost's own margin-based guarantee.
Theorem 10.1 — the mission's goal. For H a set of real-valued functions, ρ>0, δ>0,
with probability at least 1-δ, for all h∈H:
Corollary 10.2 (milestone). For a PDS kernel K with r an upper bound on K(x,x),
feature map Φ, and H = {x↦w·Φ(x) : ‖w‖≤Λ}, fixed ρ>0: R(h) ≤ R̂_{S,ρ}(h) + 4√(r²Λ²/ρ²/m) + √(log(1/δ)/(2m)).
Theorem 10.3 (milestone). RankBoost's empirical error verifies R̂_S(f) ≤ exp(-2∑_t((ε_t^+-ε_t^-)/2)²), and ≤ exp(-2γ²T) if the edge is uniformly at least γ>0.
Corollary 10.4 (milestone). Theorem 10.1's first bound, applied to h∈conv(H).
Theorem 10.1's proof is the chapter's genuine new technique, not a restatement of chunk 05's
Theorem 5.8: the two-sample decomposition (splitting the supremum over H̃ into a term on x'
alone and a term on x alone, each bounded by the Rademacher complexity under its own
marginal) is what the 2/ρ · (R_m^{D1}+R_m^{D2}) structure expresses, and collapsing it to a
single-sample bound would either be false or silently assume D1=D2 (which only holds for a
symmetric D, an assumption the theorem does not make). Corollary 10.2 is the direct
theoretical basis for the ranking SVM algorithm §10.3 derives. Theorem 10.3 mirrors chunk
07-boosting's Theorem 7.2 almost line for line in its proof technique (the same telescoping
product of normalization factors Z_t), but with genuinely different per-round quantities
(ε_t^+, ε_t^- rather than a single ε_t) that must not be conflated with AdaBoost's own, per
BRIEF.md's pitfall note. Corollary 10.4 is what makes RankBoost's output (a linear, not
convex, combination — normalized by ‖α‖_1) provably generalize independently of the number of
boosting rounds T, the ranking analogue of chunk 07's Corollary 7.5. No prior art exists on
the platform: GET /theorems?q=ranking%20loss returns zero hits.
Theorem 10.1's proof needs the two-sample structure carried through explicitly: H̃'s
Rademacher complexity splits, via the sub-additivity of sup and the fact that y_iσ_i and
σ_i have the same distribution, into a term on S2 alone and a term on S1 alone — treating
a ranking sample as an ordinary single sample (chunk 03's single-hypothesis-set machinery
applied naively) would drop this structure entirely and is exactly the pitfall BRIEF.md
names. Theorem 10.3's proof requires Z_t = ε_t^0 + 2√(ε_t^+ε_t^-) be bounded via the identity
4ε_t^+ε_t^- = (1-ε_t^0)^2 - (ε_t^+-ε_t^-)^2 and the inequality 1-x ≤ e^{-x} — the same
telescoping-normalizer technique as AdaBoost's Theorem 7.2, but RankBoost's own D_t,
ε_t^+, ε_t^- genuinely differ (they are defined via pairwise outcomes y_i(h(x'_i)-h(x_i)) ∈ {-1,0,+1}, not a single-point disagreement h(x_i)≠y_i) and must be modeled as their own
recursively-defined algorithm state, not obtained by substitution into chunk 07's AdaBoost
Lean.
MarginLossFunction restates chunk 05-svm's Definition 5.5; EmpiricalRademacherComplexity/
RademacherComplexity restate chunk 03-rademacher-vc's Definitions 3.1/3.2; IsPDS restates
chunk 06-kernels's PDS-kernel definition; ConvHull restates chunk 07-boosting's convex-hull
definition — all duplicated rather than imported since a draft item cannot import another
chunk's draft module, and none is listed as reusable in missions/README.md's "Published
definitions" table at the time of this session. D1, D2 are computed directly as
Measure.map Prod.fst D/Measure.map Prod.snd D rather than posited via a separate marginal
hypothesis, so the theorem statement itself pins down that they are genuinely the marginals of
the sampling distribution D, not independent parameters. Corollary 10.2's r is an explicit
upper bound on K(x,x) (hrK : ∀ x, K x x ≤ r) rather than a literal sSup, per BRIEF.md's
pitfall note about the possibly-infinite supremum — the theorem's conclusion is monotonic in
r, so this is not a weakening. RankBoost's D_t, ε_t^+, ε_t^-, α_t, Z_t and returned
function f are modeled as their own recursively-defined algorithm state (mirroring chunk
07's AdaBoostDist/AdaBoostEpsilon/AdaBoostEnsemble pattern exactly, but built from
RankBoost's own pairwise-outcome quantities, never by substituting into the AdaBoost Lean, per
BRIEF.md's pitfall note) — RankBoostEpsilonPlus/RankBoostEpsilonMinus are already tied to
RankBoost's own D_t and selected base ranker, so Theorem 10.3 needs no separate hypothesis
connecting them (the same trivialization guard chunk 07's own Theorem 7.2 documents). No
numerical constant is altered from the book in any of the four theorems.
Not formalized: §10.3's ranking-SVM primal/dual optimization problems (an algorithm derived
from Corollary 10.2, not a generalization-theoretic result); §10.4.2 (RankBoost as coordinate
descent, an algorithmic-equivalence argument, not a generalization bound); §10.5 (bipartite
ranking, its own distinct problem formulation with a different generalization error, Eq. 10.20,
explicitly out of scope per BRIEF.md); §10.6-10.7 (preference-based ranking, other criteria),
out of scope per BRIEF.md. The uniform-over-ρ extension mentioned after both Theorem 10.1's
and Corollary 10.2's proofs (referencing Theorem 5.9's technique from a different chapter) is
not drafted, matching chunk 09-multiclass's identical scope decision for the analogous remark.
In the Standard Model the scalar sector consists of a single complex doublet. The general Two-Higgs-Doublet Model (THDM) replaces it by two complex doublets of the same weak hypercharge . This is the minimal extension of the scalar sector that is still renormalisable and gauge invariant, and it is forced on any supersymmetric completion: the Minimal Supersymmetric Standard Model has exactly two Higgs doublets. Before any phenomenology can be done with such a model, two questions have to be answered for the given parameter point: is the scalar potential stable, i.e. bounded from below, and does its global minimum break down to the electromagnetic ?
The general THDM potential has real parameters, and answering these questions directly in field space — eight real scalar degrees of freedom, of which three are gauge — is unwieldy. Maniatis, von Manteuffel, Nachtmann and Nagel (2006) proposed a reformulation in terms of gauge-invariant bilinears, in which the gauge orbits of the Higgs fields are parametrised by a Minkowski-type four-vector confined to the closed forward light cone, and the potential becomes a quadratic polynomial on that cone. In these variables the stability question reduces to a one-variable problem: the sign of an explicit rational function on a finite set of at most ten real numbers. This mission formalizes that analysis: the gauge-orbit parametrisation (their Theorem 4), the classification of the stationary points of the potential (their Theorem 2), and the stability criterion itself (their Theorem 1).
Write the two doublets as rows of a complex matrix,
and form the hermitian matrix of gauge-invariant scalar products , i.e. . Decomposing in the Pauli basis gives four real functions
Positive semi-definiteness of is equivalent to and : the four-vector lies on or inside the forward light cone. A gauge transformation acts as with , and leaves invariant.
The most general gauge-invariant renormalisable potential is
with real parameters , and a real symmetric matrix . For one sets , so that , and
Stationary points of on the ball are governed by the functions
where is a Lagrange multiplier for the constraint . The finite set collects: every regular with ; the point when ; and every eigenvalue of at which stays finite and . It has at most ten elements, and is exactly the set of stationary values of .
For the stationary points of the full potential one uses four-vector notation , , and the metric , so that on the domain , , with .
For the potential is stable for , marginal for and unstable for . For ,
and if on with equality somewhere, the sign of — replaced by when is an eigenvalue of — decides between stability in the weak sense and instability.
Theorem 4 (gauge orbits); the four stability cases (a), (b.2), (b.3), (b.4) of Section 4 including the criterion for the marginal case; equation (4.39) identifying the stationary values of with ; equations (4.42)–(4.44) giving at those stationary points; and Theorem 2, the classification of the stationary points of .
The criterion is a decision procedure: given the parameters of a THDM, stability is settled by evaluating one rational function at the roots of another, without any search in field space. The authors use it to re-derive the known stability conditions for the MSSM potential and to settle the stability and symmetry-breaking properties of the THDM potential of Gunion et al., for which , and come out as the strong-stability conditions. Theorem 4 is what makes the whole approach legitimate: it says that nothing is lost in passing from fields to the invariants , because the fibres of that map are exactly the gauge orbits.
The results are established in the published literature; what this mission adds is machine-checked proofs. The statements are not present in Mathlib in any form, and the note added in version 3 of the paper — a condition for the marginal case that was missing in the original version — is a concrete reminder that the case analysis here is easy to get subtly wrong.
The obvious route to stability is to minimise directly; it fails because the domain is a cone with a boundary, and the minimisation over the boundary introduces a Lagrange multiplier whose admissible values are the roots of , including the degenerate "exceptional" solutions where is singular. Those exceptional solutions are not a technicality: they are where the eigenvalue clauses of , the correction in (4.43), and the junk-value behaviour of matrix inverses all live. The marginal case ( and vanishing simultaneously somewhere) is not decided by the signs alone and needs the quantitative bound .
Vectors in are plain functions Fin 3 → ℝ with an explicitly defined dot product; is a Matrix (Fin 3) (Fin 3) ℝ and its symmetry is carried as a hypothesis. Four-vectors are indexed by Unit ⊕ Fin 3 so that and are block matrices, with the first component being . Higgs configurations are complex matrices and ; a gauge transformation is with .
Stability is formalized as the honest statement that is bounded from below on the physical domain , — not as any of the sign conditions that the theorem derives — so none of the implications is true by definition. Matrix inversion in Lean returns the zero matrix at a singular argument; every occurrence of is therefore guarded by a regularity hypothesis, and the values of at an eigenvalue of are defined as limits, with the existence of those limits part of the membership condition for . The projection is characterised by its defining property (it lies in the eigenspace and is orthogonal to it) rather than by a choice of eigenbasis.
A complete development needs linear algebra over (resolvents, symmetric matrices, eigenspaces), and the spectral decomposition of positive semi-definite complex matrices for Theorem 4, and elementary real analysis (compactness of the ball, limits of rational functions) for Section 4. The gauge-orbit statement and the light-cone parametrisation are reusable for any multi-doublet scalar sector; contributions of the -doublet generalisation (Appendix B, Theorem 5) are welcome as follow-ups.
Every convergence-rate result in the preceding chapters is an upper bound: some specific estimator (the Lasso, PCA, kernel ridge regression) achieves a given error rate. A natural and much harder question is the complementary one: is that rate actually the best any procedure could achieve, no matter its computational cost? Answering this requires a theory of lower bounds that holds simultaneously for every conceivable estimator — a fundamentally different kind of argument from constructing and analyzing one particular algorithm. Wainwright's High-Dimensional Statistics: A Non-Asymptotic Viewpoint (Cambridge University Press, 2019), Chapter 15, develops this theory, unifying classical techniques (Le Cam, Assouad, Fano) under one reduction: converting continuous estimation into discrete hypothesis testing.
Let be a sample space and a class of probability distributions on . A functional assigns each distribution a parameter of interest. An estimator is a measurable map . Fix a semi-metric — symmetric, triangle-inequality-satisfying, , but possibly for — and an increasing . The minimax risk is
the smallest worst-case expected loss achievable by any measurable estimator (Eq. (15.2)).
Given a 2δ-separated set
(every pair satisfies ) with representative distributions
, Wainwright constructs a testing problem: sample
uniformly from , then ; write for the resulting joint law of
. A test function attempts to recover from ; its
error probability is .
For any increasing and any 2δ-separated set with its induced joint testing measure
,
This mission formalizes Proposition 15.1 alone (see Formalization scope below for why, and what a follow-up mission would add to reach Fano's method proper, Proposition 15.12).
Proposition 15.1 is the single reduction every subsequent technique in the chapter specializes: Le Cam's two-point method (Lemma 15.9, , bounding the testing error via total variation distance), Fano's method (Proposition 15.12, bounding it via mutual information and Fano's inequality), and Assouad's method (a different, hypercube-based packing). Formalizing it in full generality — general , general semi-metric, general -ary packing set — gives a single reusable lemma that a future mission proving any of these specific bounds can build on directly, rather than re-deriving the reduction each time.
The theorem is already proved in the source; this mission's contribution is a machine-checked formal statement (and, eventually, proof) of the reduction, in a form composing with any future formalization of the chapter's testing-error bounds (total variation, Fano, or otherwise).
The proof combines two ingredients that must each be kept in their sharpest form: Markov's inequality applied to (which only needs increasing, not convex or any specific shape — a premature specialization to would silently prove a weaker, less reusable statement), and the reduction of any estimator to a test via nearest- packing-point assignment (Eq. (15.4)), which uses the triangle inequality on in a specific direction (bounding from below via and ) to show that a small estimation error forces the induced test to be correct. Getting the direction and strictness of every inequality right — non-strict separation, but strict distance in the "test is correct" event — is where a naive restatement goes wrong.
, are arbitrary measurable spaces; the distribution class is
realized as an indexed family measure : Idx → Measure 𝒳 rather than a bare set of measures,
composing directly with θ : Idx → Ω. The semi-metric is a bare function with explicit
nonnegativity/reflexivity/symmetry/triangle-inequality hypotheses, matching the book's own
footnote definition, rather than Mathlib's PseudoMetricSpace typeclass (kept self-contained,
no extra instance machinery). The minimax risk is valued in ENNReal via the lower Lebesgue
integral ∫⁻, not the Bochner integral, specifically to avoid the non-integrable-loss junk
value 0 that a Bochner-integral formalization would silently introduce — a trivializing
formalization would use ∫ (Bochner) here, letting a non-integrable loss vanish and making the
inequality easier to satisfy than the book's actual claim; this mission does not do that. The
joint testing measure is characterized by its slice-measure equations directly on the
product space , avoiding Mathlib's general conditional/disintegration
machinery while remaining exactly equivalent to " uniform, ".
Disclosed major scope decision. BRIEF.md recommended Proposition 15.12 (the Fano bound
itself, Φ(δ)(1 - (I(Z;J)+\log 2)/\log M)) as this mission's goal. That statement requires, in
addition to everything above, a formalized notion of mutual information between a
finite-valued and a general (possibly continuous) random variable, and its use of a Fano-type
inequality (Eq. (15.31), itself deferred by the book to "Section 15.4" and not fully quoted in
the brief). Building a faithful mutual-information formalization general enough for this
setting (finite , arbitrary measurable ) — matching Mathlib's or this repo's existing,
narrower information-theoretic developments (SourceCoding.*, built for a different,
channel-coding purpose per BRIEF.md's own prior-art note) or building one from scratch — is
substantially more than this session's remaining budget after building and self-reviewing
08-pca and 07-sparse-linear. This mission instead formalizes Proposition 15.1, the
foundational reduction Proposition 15.12 itself specializes (via a particular bound on
), so that a follow-up mission can add the mutual-information/Fano
step on top of HighDimStat.Minimax.estimation_to_testing without redoing this reduction.
This mission's name, fixed from missions/README.md, still names "Fano's Method" as the
series slot this chunk occupies; its actual content is the reduction step every method in that
family (including Fano's) shares — recorded explicitly here and in STATUS.md, not left
implicit.
Every convex optimization algorithm in machine learning — from projected gradient descent to support vector machines to the mirror-descent methods of later chapters of this book — is justified by a small set of optimality certificates: a checkable condition on a candidate solution that guarantees it actually solves the problem, without searching the whole feasible set. The most widely used such certificate is the Karush–Kuhn–Tucker (KKT) system: a set of gradient and complementary-slackness equations that a solution of a convex program with differentiable data must satisfy, and that (under a mild constraint qualification) is also sufficient. It is the tool a practitioner reaches for to check optimality of a numerical solver's output, and the tool a theorist reaches for to derive an algorithm (dual ascent, augmented Lagrangian, interior point methods) in the first place. Every convex machine learning model introduced in Chapter 1 of this book — regularized least squares, support vector machines, logistic regression with constraints — is an instance of the general convex program these milestones analyze.
The result traces back to Kuhn and Tucker's 1951 paper Nonlinear Programming, with Karush's 1939 unpublished thesis establishing the same conditions independently and earlier; Slater's 1950 unpublished note is the source of the constraint qualification that bears his name and that makes the necessity direction possible. Lan's Chapter 2 gives a compact, modern, purely finite-dimensional derivation of the whole chain — separation, duality, saddle points, KKT — from first principles, self-contained in about twenty pages, aimed squarely at the convex programs that appear in machine learning.
Fix and work in with its standard inner product . A convex program (2.3.16) is
where is a nonempty closed convex set, are convex, and are affine. A point is feasible if it satisfies every and ; is optimal if it is feasible and for every feasible .
The normal cone of at is — the set of directions that make an obtuse angle with every direction into from ; it is when , recovering unconstrained first-order optimality. The Lagrangian is for multipliers , ; the Lagrange dual value is , and the Lagrange dual problem is . Weak duality, , holds unconditionally by construction. Slater's condition asks for with , ; the restricted Slater condition weakens the interior requirement to the relative interior while keeping strict inequality for every (here: every) nonlinear constraint.
Each of these five is stated the way the book states it — no constant is hard-coded, no O(·) is involved, and every hypothesis (closedness, convexity, Slater/restricted-Slater) is exactly the one the corresponding proof uses.
The KKT system is the interface between convex optimization theory and every algorithm that
exploits it: primal-dual methods track approximate KKT residuals as a stopping criterion, and the
derivation of the Lagrange dual (used throughout the book's later treatment of composite and
constrained problems) rests on strong duality, milestone strong_duality here. The saddle-point
characterization (saddle_point_sufficient/saddle_point_necessary) is the standard route to
designing an algorithm: a method that provably drives a pair to a saddle point
of is provably convergent to an optimal , without ever needing to verify optimality
directly against the primal problem.
None of the six substantive results here has a machine-checked proof on Prove2Me. The platform
holds a genuinely weaker unconstrained-in- condition (OnlineConvexOpt.ConvexBasics. kkt_optimality, Hazan's Theorem 2.2: for ,
with no inequality/equality constraints or multipliers at all) and a structurally different,
strictly more general cone-based Lagrange-duality development (VectorSpaceOpt.lagrange_duality
and VectorSpaceOpt.lagrangian_saddle_sufficient_pointed, from Luenberger, which bundle all
constraints into a single map into a convex cone in a general normed space, rather than Lan's
explicit inequality / equality split). Formalizing this mission
produces the finite-dimensional convex-program version of KKT in exactly the shape it is used and
taught: separate multiplier vectors for inequality and equality constraints, an explicit normal
cone rather than a cone-map abstraction, and both directions of the necessity/sufficiency split.
The separation theorem (Theorem 2.1) itself is routine once the projection onto a closed convex set is available. The real difficulty is entirely in the direction of the Convex Theorem on Alternative that Proposition 2.9 states: the naive idea — "insolvability of (I) should give a separating hyperplane between and directly" — fails, because these are sets in and separating them there does not produce a sign-definite multiplier vector. Lan's proof instead lifts to and separates the epigraph-like set from the open orthant-like set ; only in this lifted space does the separating normal's sign constraint (forced by 's unboundedness in the positive directions) translate into . Getting the sign of the -th coordinate strictly positive — needed to normalize and divide — is itself a small separate argument using the Slater subsystem's solution. The KKT necessity direction (the goal) then chains three of these already-nontrivial results (separation → CTA → strong duality → saddle necessity) before translating the saddle-point condition on into the gradient/normal-cone form via differentiability of at .
All milestones are stated over EuclideanSpace ℝ (Fin n) with Convex/ConvexOn from Mathlib.
Affine equality constraints are represented by explicit witnesses with , so that is
available without a separate affine-differentiability lemma. The normal cone is Lan's
own primal-space object (normalCone); the restricted Slater condition uses Mathlib's
intrinsicInterior ℝ X for , distinct from the plain interior X used by
the un-restricted Slater condition of Theorems 2.6/2.7(b). Primal optimal values and Lagrange dual
values are stated via IsGLB/pointwise-inequality forms rather than raw sInf, so that no
hypothesis is silently made true by an empty or unbounded set defaulting sInf to a junk value —
in every milestone here the relevant set is guaranteed nonempty by the Slater-point hypothesis
already present.
A trivializing formalization this mission rules out: stating kkt_necessary/kkt_sufficient
with (no set constraint) and empty (no functional constraints) would
collapse the normal-cone condition to and make the whole KKT apparatus vacuous
of any duality content; the milestones here keep , and as genuine free parameters (the
theorems are stated for arbitrary m, p : ℕ, including but not restricted to the degenerate case)
so that the constrained content of Lan's theorem is what gets proved.
Reusable beyond this mission: normalCone and lagrangian are generic enough that any later
chapter of this book needing Lagrangian duality or normal-cone stationarity (none of the current
first-wave chapters 03/06 needs them directly) could import them once published rather than
redeclaring. Contributions most welcome on the two hardest milestones,
cta_solvable_of_insolvable and kkt_necessary, since they carry the mission's real difficulty;
separation_point_closed can likely be discharged quickly via Mathlib's
geometric_hahn_banach_point_closed.
Every statistical guarantee for a learning algorithm ultimately rests on a concentration inequality: a bound on how far an empirical average can stray from its expectation. For real-valued averages, Bernstein's inequality (1924, refined through the 20th century) is the classical tool — sharper than Hoeffding's inequality whenever the summands' variance is small compared to their range. Modern learning theory, however, frequently needs to control averages of objects that are not real numbers but elements of a Hilbert space: feature vectors , gradients of a loss, or the values a kernel machine's empirical risk functional takes. Steinwart & Christmann, Support Vector Machines (Springer 2008, Information Science and Statistics), Chapter 6, develop exactly the vector-valued extension needed for the book's own SVM consistency proofs: Theorem 6.14, Bernstein's inequality for independent random variables taking values in a separable Hilbert space, and its corollaries.
Fix a probability space and a separable real Hilbert space . Random variables are independent if the family is mutually independent (not merely pairwise), and has essential supremum bound , , if for -almost every . Writing for , the quantities of interest are the mean (a Bochner integral) and the variance bound , an upper bound on .
The classical scalar case (Theorem 6.12, itself a refinement of Hoeffding's inequality, Theorem 6.10) bounds for real-valued, mean-zero, range- and variance-bounded . The tool behind both the scalar and the vector-valued case is a general exponential-moment inequality (Theorem 6.13) valid for independent, integrable random variables taking values in any separable Banach space :
for independent, mean-zero with and . This is the weakest stable form of the claim: it is stated for a general separable Hilbert space (not a fixed finite dimension), with the tail written as a sum of three explicit terms rather than folded into an unspecified constant, so it survives specialization to any concrete without modification.
Theorem 6.13 (above) is the direct tool Theorem 6.14's own proof invokes ("we will prove the assertion by applying Theorem 6.13"); Theorem 6.12 is the scalar analogue Theorem 6.14 generalizes, included to make the generalization's exact form (three terms, not two) checkable against its source; Corollary 6.15, Hoeffding's inequality in Hilbert spaces, is the immediate mean-free-of-a-variance-bound consequence of Theorem 6.14 obtained by centering, reused directly in the book's own oracle-inequality proof (§6.4).
Theorem 6.14 is the concentration inequality behind the book's later empirical-process arguments for SVMs: whenever an SVM's analysis needs to bound the deviation of an empirical average of Hilbert-space-valued quantities (feature-map evaluations, loss gradients) from its mean, this is the tool invoked, via Corollary 6.15 in the book's own oracle-inequality derivation. Its distinct contribution over simply applying the scalar Theorem 6.12 coordinate-by-coordinate (which is not available without a fixed, finite orthonormal basis, and even then would produce dimension- dependent bounds) is that the bound here is entirely dimension-free: it depends on only through the variance bound and the range bound , not through .
The scalar Bernstein inequality is classical (Bernstein 1924, refined by Bennett 1962 and others
to the sharper multiplicative form used here); its Banach- and Hilbert-space generalizations
(via the martingale-difference / Yurinskii-type argument Theorem 6.13's proof uses) are standard
in the empirical-process-theory literature by the time of this book (see, e.g., Pinelis 1994 for
closely related Banach-space martingale inequalities). No machine-checked Lean proof of the
Hilbert-space form is known to exist on the platform at the time of writing (the platform's own
HighDimProb.Concentration.bernstein_unweighted, from the Vershynin series, is a different,
sub-exponential-norm scalar statement — see Formalization scope); this mission asks for the
book's own bounded-summand, explicit-constant Hilbert-space form.
The natural first attempt generalizes the scalar proof's Markov-inequality argument directly: bound using independence. This breaks immediately because is not linear, so does not factor over the way does in the scalar case — there is no vector-valued analogue of the moment generating function that tensorizes under independence directly. Theorem 6.13's proof resolves this with a martingale-difference decomposition (writing the deviation as a telescoping sum of conditional-expectation differences across the filtration generated by ) rather than a direct product-of-moment-generating-functions argument, at the cost of needing separable (for the conditional expectations and the resulting sums to be well-defined and measurable). Deriving Theorem 6.14 from Theorem 6.13 then requires controlling the two extra terms Theorem 6.13 introduces (the mean-norm term and the per-summand correction) using only the Hilbert-space-specific facts for (from independence and mean-zero) and the scalar bound on — which is exactly where the tail's extra term originates, and is not obtainable by naively reusing the scalar Theorem 6.12's two-term optimization over unchanged.
(and, for Theorem 6.13, ) is an arbitrary separable real (Hilbert, resp. Banach) space —
not fixed to a Euclidean space of any dimension — matching the book's own generality, which is
essential since the theorem's dimension-independence is part of its content. is formalized as the almost-sure bound ∀ᵐ ω ∂P, ‖ξ i ω‖ ≤ B, matching the book's
convention rather than requiring the bound to hold for literally every .
Independence is the mutual independence of the whole family (Mathlib's iIndepFun), matching
Theorem 6.13's proof, which uses independence of from for every
simultaneously, not merely pairwise independence.
A trivializing formalization would state the conclusion in terms of the scalar random variable rather than the vector-valued average's norm — this would collapse the theorem to an easier scalar statement about a nonnegative random variable and lose the whole point of the vector-valued generalization; it is ruled out here by writing the norm of the sum (not a sum of norms) inside the probability. Likewise, dropping the middle term of the tail bound (present here, absent from the scalar Theorem 6.12) would silently understate the genuine dimension-independent cost of vector-valued concentration; all three terms are kept.
Theorem 6.13's general Banach-space statement, and the scalar Theorem 6.12, are reusable beyond
this mission (Theorem 6.13 is the tool any future Banach-space-valued concentration mission in
this series would reach for first). The platform's existing HighDimProb.Concentration. bernstein_unweighted/hoeffding_rademacher (Vershynin series) are not reused as kind: reference items here: they use a sub-Gaussian/sub-exponential-Orlicz-norm parameterization and a
universal (unspecified) constant, a genuinely different hypothesis structure from this chapter's
explicit -bounded, exact-constant form — reusing them would misrepresent this chapter's
own, sharper statement. Completing the four sorrys (Theorems 6.12-6.14, Corollary 6.15) is
welcome; the martingale-difference argument behind Theorem 6.13 is the natural starting point,
since the other three all reduce to it directly.
Quantum field theory begins, mathematically, with a question that has a completely precise answer: what is the state space of a single relativistic particle? Non-relativistic quantum mechanics answers and moves on. Relativity does not allow that answer, because the state space must carry an action of the symmetry group of Minkowski spacetime — the Poincaré group — and the choice of Hilbert space is dictated by which such action one wants. The construction that results is the foundation on which Fock space, creation and annihilation operators, free fields and eventually interacting theories are built, and it is where the objects that reappear everywhere in the subject are introduced: the mass shell, the Lorentz-invariant measure on it, and the double cover that is responsible for spin.
This mission formalizes that construction as it is presented in S. Chatterjee's Lectures on Quantum Field Theory (Stanford, 2018–19), Lectures 9–11 and Lecture 25: the one-particle space of a massive scalar boson, the one-particle space of an electron, and the statement that both carry inner products invariant under the Poincaré action.
Minkowski spacetime is with the bilinear form
A Lorentz transformation is a linear map with ; the restricted Lorentz group consists of those with and . The Poincaré group is with the group law .
For a mass , the four-momentum of a particle satisfies and , so it lies on the mass shell
a three-dimensional manifold parametrised by the spatial momentum through with . On there is, up to a multiplicative constant, exactly one measure invariant under ; with the normalisation used in the lectures it is the measure determined by
The state space of a massive scalar boson is , acted on by .
For an electron the wave function takes values in and the group acts through the double cover. To each four-vector one attaches the Hermitian matrix
and for the transformation of is defined by ; the map is a surjective two-to-one homomorphism onto . Writing , each is reached from by a unique positive-definite , the pure boost, and the electron inner product is the -weighted one,
with the group acting by .
For , , and :
for in the weighted space of -valued functions. The goal fixes no constants beyond the normalisation of , and it is the statement that the spaces defined in the mission really are the one-particle spaces of the theory: a Hilbert space together with a Poincaré action by isometries.
The milestone list follows the lectures: the parametrisation of ; invariance of under ; the integration formula for (eq. (10.1)); invariance of ; uniqueness of the invariant measure up to a constant; the composition law and unitarity of the scalar representation; and bijectivity of onto Hermitian matrices; as a multiplicative map into ; surjectivity of with fibres ; existence and uniqueness of the pure boost ; Lemma 25.1; and the identification of the weighted electron space with the plain -valued space via .
The objects here are used unchanged for the rest of a QFT course: the bosonic and fermionic Fock spaces are built on these one-particle spaces, and the free scalar and Dirac fields are operator-valued distributions written as integrals against on . Formalizing them fixes, once and for all, the conventions later work must match — the normalisation of , the sign convention of the metric, which of the two elements of acts, and the weight in the electron inner product.
What this mission adds beyond the lectures is a machine-checked development of material usually treated as routine but rarely written out: the uniqueness of the invariant measure, the covering map and its fibres, and the existence-uniqueness of the pure boost are all stated in the source either without proof or as exercises. Mathlib has the general theory of spaces, push-forward measures, Hermitian and positive-definite matrices, and , but it has no mass shell, no invariant measure on it, and no covering map onto the restricted Lorentz group; all of that is constructed here and is reusable by any later mission on free fields or Fock spaces.
The obvious route to the invariant measure — "restrict Lebesgue measure to the submanifold " — does not work: the induced Riemannian volume of the hyperboloid in the Euclidean metric is not Lorentz invariant. The lectures instead take a scaling limit of Lebesgue measure on the invariant annuli ; the formalization takes the resulting formula (10.1) as the definition and must then prove invariance, which amounts to a change-of-variables computation whose Jacobian is exactly -dependent. Uniqueness is harder: it is a statement about invariant measures on a homogeneous space of a non-compact group, with no finiteness available.
On the spinor side, the central difficulty is that the weight is unbounded on , so the electron space is not the naive -valued — the two spaces consist of different functions, and are related only through the measurable field of isomorphisms . A formalization that silently uses the unweighted space would prove a different, and false, unitarity statement.
Four-vectors are functions , with the metric signature ; Lorentz transformations are real matrices, and membership in is the predicate (, , form preserved). is a Borel measure on all of carried by , defined as the push-forward of ; the boson space is the library space of that measure. The pure boost is given by the closed formula — the positive-definite square root of — rather than by a choice function, and a milestone certifies that it is the unique positive-definite element of carrying to . The electron space is the set of -valued measurable functions of finite weighted norm, with the weighted pairing given explicitly; the milestone identifying it with the plain space via the inverse boost is what supplies its Hilbert-space structure.
The unitarity statements are formalized as equalities of integrals over pairs of wave functions rather than as statements about abstract operators, so that no trivializing reading is available: in particular the electron clause is stated for the weighted pairing, which is not the standard inner product, and the hypotheses (, , in the respective spaces) are satisfiable, so no clause holds vacuously. Total-function conventions of the library (inverse of a singular matrix is ; integral of a non-integrable function is ) are visible in the statements and are recorded in each item's read-back.
Contributions of any of the milestones are welcome; the measure-theoretic milestones (invariance and uniqueness) and the covering milestones are independent of each other and can be attacked in parallel.
Chapter I's usual stochastic order compares "how large" two random variables tend to be. A different, equally common question in operations research is "how spread out" a random variable is: two portfolios with the same expected loss can differ sharply in how much that loss varies, and a risk-averse decision maker or a convex cost function cares about exactly that difference. Shaked and Shanthikumar's Stochastic Orders (Springer, 2007) formalizes this comparison as the convex order, the book's mean-preserving-spread order (known outside OR and probability circles as the Rothschild–Stiglitz order from economics). This mission formalizes the order and its central structural result: Strassen's martingale-coupling characterization.
Let be a real-valued random variable on a probability space , and let be a real-valued random variable on a (possibly different) probability space . is smaller than in the convex order, written , if
Convex functions take their relatively largest values on the "extreme" regions outside some interval, so says is more likely than to take extreme values: is "more variable" than . Unlike the usual stochastic order, forces the two means to agree (, taking , both convex) — the convex order compares spread holding location fixed, exactly the mean-preserving-spread reading. Two equivalent forms make it tractable: comparing tail integrals of the survival/distribution functions (Theorem 3.A.1), and comparing mean absolute deviations from every point (Theorem 3.A.2).
Furthermore, can be chosen so that the conditional law stochastically increases with (in ). This is the convex-order analogue of Chapter I's Theorem 1.A.1: instead of one variable dominating the other pathwise, 's copy is a fair (martingale) randomization of 's copy whose spread only grows in the conditioning value. The book itself calls the constructive direction "not easy to prove"; this mission states the theorem faithfully, including the "Furthermore" strengthening, without attempting a proof.
The convex order is the standard way operations research and actuarial science formalize "more variable, same average": comparing the riskiness of two portfolios with matched expected return, the effect of aggregation or diversification on total claim size, or the value of information in a stochastic program (where a random variable degenerates to its mean exactly when the decision maker learns everything, the two extremes of a convex-order chain). Strassen's characterization is what turns "compare against every convex function" — an intractable universal quantifier — into a single explicit construction: exhibit one martingale coupling and the comparison is settled for every convex function at once, by Jensen's inequality. This is the technique behind bounding the effect of information or risk aggregation without checking convexity function by function, and the "Furthermore" monotonicity clause is what makes the coupling itself informative about how the spread grows with the conditioning variable, not merely that some fair coupling exists.
Formalizing this theorem fixes, for the whole book series, the exact shape every later coupling theorem for a variability-type order (the increasing convex/concave orders of Chapter IV via a submartingale, the multivariate convex order of Chapter VII) is expected to restate. No proof is attempted; the book's own remark that the constructive direction is "not easy" marks it as a genuine target for a future proof-bearing pass, not a formality.
The definitional trap mirrors Chapter I's: "for every convex " must be a genuine
universal quantifier over Mathlib's own convexity predicate, with the existence of both
expectations stated as an explicit integrability hypothesis inside the quantifier, not assumed
globally or dropped. The coupling theorem doubles the difficulty of Theorem 1.A.1's: the
conditional-expectation condition a.s. is with respect to the
-algebra generated by , not an informal "expected value given ", and must
use the martingale API's own convention (Mathlib's condExp) precisely. The "Furthermore" clause
compounds this: it is a claim about the regular conditional distribution of given
for every point , not merely the conditional expectation, and dropping it silently
(as a "remark" rather than part of the theorem) would understate what Theorem 3.A.4 actually
claims — the chapter brief flags this explicitly as a trap, and it is kept as a conjunct of the
same existential witness here.
Random variables are again measurable functions into from arbitrary measurable
spaces, ConvexOrder μ ν X Y taking on and on a separate ,
matching Chapter I's convention and this series' own pattern. Convexity is Mathlib's ConvexOn ℝ Set.univ φ; "equality in law" is again ProbabilityTheory.IdentDistrib. The martingale condition
uses Mathlib's conditional-expectation notation ρ[Ŷ | m] =ᵐ[ρ] X̂ with m the -algebra
MeasurableSpace.comap X̂ inferInstance generated by X̂ — the martingale API's own convention,
not a hand-rolled gloss. The "Furthermore" clause uses Mathlib's ProbabilityTheory.condDistrib,
the regular conditional distribution kernel of Ŷ given X̂ (available since ℝ is a standard
Borel space), and states "increasing in x in ≤st" via the kernel's survival function being
monotone in x at every threshold — restating, locally to this chapter's namespace, the same
tail-probability shape Chapter I's UsualOrder uses (drafts cannot import another mission's
definitions, per this series' convention). A trivializing formalization is ruled out explicitly:
the martingale and monotonicity conjuncts are both kept as genuine content of the existential
witness in the goal theorem, not weakened to a bare coupling or dropped as an optional remark.
This mission draws on no platform prior art (searches for "convex order", "Strassen",
"martingale", "Jensen" returned only unrelated geometry/algorithm/physics results as of
2026-09-18); Mathlib's Probability/Martingale/* supplies the conditional-expectation and kernel
machinery the coupling condition and its "Furthermore" clause are built from, but no platform
theorem states the convex order or its coupling characterization itself. Reusable beyond this
mission: the condExp/condDistrib-based martingale-coupling pattern is the shape Chapter IV's
analogous submartingale coupling (Theorem 4.A.5) and Chapter VII's multivariate convex order
(Theorem 7.A.1) are each expected to restate independently.
A device's mean residual life at age — its conditional expected remaining lifetime given that it has survived to — 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 , a natural "the survivor of is worn less, on average, than the survivor of " 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 : 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.
Fix a probability space and a real-valued random variable with survival function and finite mean. The mean residual life function of at is
For a second random variable on with mrl function , is smaller than in the mean residual life order, , if for every . 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 for all , where is 's survival function. A random variable is DMRL (decreasing mean residual life) if its mrl function is decreasing in .
Combined with the companion milestone below, this is a genuine conditional equivalence: under the monotone-ratio hypothesis, and coincide, and in particular under that condition. Without the hypothesis, the book states explicitly (the paragraph immediately preceding Theorem 2.A.1) that neither nor implies the other.
Each milestone is stated exactly as the book states it: no constant is hard-coded, no or asymptotic approximation is involved, and the goal's monotone-ratio hypothesis is the genuine ratio , not two separately-monotone functions (a different, unrelated condition the book itself does not state).
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.
The mrl function is a genuinely two-case object: a real conditional expectation on , and a hard outside that region. The goal theorem's proof (not formalized here; only the statement is a milestone) differentiates and , uses the identity 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 (a pointwise comparison of a derived function) visibly distinct from the function-class shape of 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.
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 directly via positivity of the survival probability
(its defining equivalent under the survival function's monotonicity) rather than through the
derived quantity 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 and 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 (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 " is decreasing in " 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 and individually (rather than
genuine monotonicity of the ratio on the region where ) 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 " increases in
" verbatim.
StochasticOrders.Usual), for the usual and hazard rate orders this
chapter's own restated definitions parallel.Von Neumann's minimax theorem (Chapter VIII) settles two-player zero-sum games with scalar payoffs. In 1956, Blackwell asked the natural generalization: what can a player guarantee in a repeated game with vector-valued payoffs, where "winning" means driving the average payoff into a target set rather than above a target value? For decades the resulting theory — approachability — and the regret-minimization theory this book develops were believed to be different, with approachability seen as the stronger notion. Chapter 13 closes that gap: approachability and online convex optimization are shown to be algorithmically equivalent, each reducible to the other with no loss of efficiency, and along the way this equivalence yields a constructive, rate-quantified proof of Blackwell's own theorem.
A generalized vector game (Definition 13.2) is given by bounded convex closed decision sets and a vector payoff . A set is approachable (Definition 13.3) if some non-anticipating algorithm, playing in against any sequence , drives the average payoff's distance to to zero. Blackwell's theorem (13.4) characterizes exactly which are approachable via a purely geometric condition: every column-player strategy admits a row-player best response landing the payoff in .
Section 13.2 constructs an explicit approachability algorithm from any OCO algorithm: given a best-response oracle realizing Blackwell's condition, Algorithm 37 runs the OCO algorithm on the proxy losses (the support function letting distance-to- be written, via Lemma 13.5's minimax duality, as a convex optimization problem over the unit ball), queries the oracle at the OCO algorithm's play , and averages the resulting rewards.
This chapter's headline claim — approachability and OCO are equivalent — is proved in two
directions in the book (§13.2 and §13.3); this mission drafts the direction the book itself
foregrounds as "the more interesting implication" and constructively proves: any sublinear-regret
OCO algorithm converts directly into an explicit approachability algorithm with an explicit
convergence rate, giving a self-contained, algorithmic proof of a 1956 game-theory theorem using
1990s–2000s online-learning machinery. Historically, this equivalence resolved a standing
misconception (approachability believed strictly stronger) and reframes Blackwell's theorem as a
special case of regret minimization rather than a separate theory requiring its own toolkit. No
prior art was found on the platform for Blackwell approachability (planning search:
q=Blackwell — the one hit, PRNGCompression.prng_no_free_lunch's cousin, an unrelated
Rao-Blackwellization result, is not a substitute); this mission drafts both items fresh.
Theorem 13.4's statement is a clean geometric implication, but the book is explicit that its
proof is entirely carried by Theorem 13.7 plus an unstated "explicit conclusion" left as an
exercise (the passage from a finite-horizon rate bound to the asymptotic Dist → 0 claim, using
any of the book's own sublinear-regret OCO algorithms as a witness). Theorem 13.7's own proof
combines three nontrivial facts: Lemma 13.5's minimax-duality rewriting of as a linear optimization over the unit ball (itself proved via Sion's minimax theorem, not
excerpted here), the best-response oracle's defining inequality (13.2) applied pointwise at each
round's , and the OCO algorithm's own regret guarantee applied to the specific proxy-loss
sequence built from the realized game trajectory — a genuine composition of three separate
pieces of machinery from earlier in the book (Chapters III–VIII), not a routine substitution.
IsApproachable is declared as its own definition (per BRIEF.md's explicit instruction, since
Theorem 13.4 depends on it), with the non-anticipation clause made explicit (matching the
series' IsOnlineAlgorithm convention from Chunk 03) even though the book's own Definition 13.3
states it only informally ("x_t ← A(y_1,\dots,y_{t-1})"). SupportFunction is h_S exactly as
displayed, as a real supremum (a genuine maximum given the chapter's standing "closed, bounded"
hypothesis on S). throughout is Euclidean distance, rendered as Mathlib's
Metric.infDist — confirmed the chapter uses no other distance notion (checked §13.1-13.3
directly, per the pitfall BRIEF.md flags). Theorem 13.7 transcribes Algorithm 37's
construction faithfully, including its one-round offset (using the
previous round's realized reward to build the current round's proxy loss, while the
conclusion averages the current round's rewards) — exactly as the book's own pseudocode has it,
not smoothed over.
Scope decision on Theorem 13.4's biconditional. The book states Theorem 13.4 as an ↔ but
proves, and explicitly flags as proved, only the sufficiency direction (←): "The necessity of
this condition is left as an exercise... Our reductions henceforth give an explicit proof of
Blackwell's theorem [meaning: of the sufficiency direction]." Per CAPTAIN_BRIEF.md rule 6 and
BRIEF.md's explicit instruction, this mission drafts only that direction, named as such in the
goal item's own docstring; see STATUS.md.
Not formalized (out of scope for this mission, given the remaining budget and the explicit
"exercise" status of several results on these pages): the necessity direction of Theorem 13.4;
Lemma 13.5 (minimax duality for Dist, itself relying on Sion's theorem, not separately
formalized here); Lemma 13.6 (the equivalent best-response-oracle condition); §13.3's entire
approachability-to-OCO direction (Theorem 13.9, Lemma 13.8, the cone/polar-cone machinery of
§13.3.1) and §13.3.3 (existence of a best-response oracle for the constructed set); the "explicit
conclusion" of Blackwell's theorem from Theorem 13.7, left as an exercise by the book itself.
Online convex optimization (OCO) asks a learner to repeatedly pick a point in a convex set , pay a cost that an adversary reveals only after the choice is made, and be judged against the best fixed point in hindsight. Chapter III of this series formalized the simplest general-purpose answer, online gradient descent (OGD): take a gradient step, project back onto . OGD's analysis, however, is tied to the Euclidean geometry of the projection step — it treats every coordinate of alike, and its regret bound degrades badly when 's natural geometry is not Euclidean (the probability simplex under the norm is the standard example, where a Euclidean-projection algorithm's regret scales with in the dimension , while an algorithm that exploits the simplex's own geometry attains regret scaling only with ).
Regularized Follow the Leader (RFTL) is the meta-algorithm this chapter introduces to fix this: rather than fixing a specific geometry, RFTL is parameterized by an arbitrary regularization function , and its regret bound depends on only through two scalar quantities the mission makes explicit — the range of over , and a -dependent "local norm" of the gradients. Choosing to match 's geometry (entropy regularization on the simplex, for instance) recovers the sharp bounds that plain OGD cannot. RFTL and its close relative Online Mirror Descent (OMD), also introduced here, are the ancestors of essentially every regularization-based online learning algorithm in use today, including the multiplicative-weights/Hedge algorithm of Chapter I as a special case (entropy regularization on the simplex) and the exponentiated-gradient algorithm this book's own Chapter VIII reuses (Corollary 5.7, a further specialization of Theorem 5.2 this mission's Theorem 5.2 underlies). The naive "Follow the Leader" strategy this chapter opens by refuting — always play the empirically best point so far — is a natural first idea and provably fails: the book gives an explicit two-point cost sequence on which it incurs regret linear in the horizon. Regularization is the fix, and quantifying exactly how much it costs and buys is this chapter's content.
Fix a convex, nonempty decision set in a real inner product space and a sequence of convex cost functions . As in Chapter III, regret after rounds is
A regularization function is a strongly convex, smooth, twice differentiable function with a positive-definite Hessian on the interior of . Its Bregman divergence measures the gap between and its own first-order Taylor approximation:
By the mean value theorem, for some point on the segment , where is the norm induced by the Hessian ; its dual norm, denoted , is the local norm at . Writing for the local norm between consecutive iterates , the -diameter of is .
The RFTL algorithm (Algorithm 13), with step size , plays , then at every round updates
The agile Online Mirror Descent algorithm (Algorithm 14, agile version) instead maintains a dual point with , updates it by , and projects via the Bregman divergence, (with defined the same way from ). RFTL and the lazy variant of OMD coincide for linear costs (Lemma 5.5, not formalized here — it is not used by either target); the agile variant's analysis is genuinely different and is the mission's second target.
This is the mission's goal: RFTL, run with any admissible regularizer, attains a regret bound governed only by the cumulative squared local norm of the gradients and 's range over . The bound is proved via two milestones: Lemma 5.3 (regret controlled by the total "prediction drift" plus ), which in turn rests on Lemma 5.4 (a "follow-the-leader beats be-the-leader" comparison inequality, proved by induction on the horizon).
A structurally similar bound for the agile variant, included as its own goal-level item since — as the book states explicitly — its proof technique is unrelated to RFTL's, not a corollary of it.
Both targets are the book's own tightest, non-asymptotic statements: neither is weakened to an form, and the book's own further (unnumbered) corollary specializing Theorem 5.2 to a uniform local-norm bound is left out, matching this series' convention of formalizing only the numbered results.
The results themselves. Theorem 5.2 is the general regret theorem behind every regularization scheme in online learning: instantiating recovers the projected-gradient bound of Chapter III (Euclidean ), the multiplicative-weights bound of Chapter I (entropy on the simplex), and — through the exponentiated-gradient specialization (Corollary 5.7, not itself a target here) — the row-player regret bound this book's own Chapter VIII cites as "Eq. (8.1)" in its reduction of zero-sum games to regret minimization. Theorem 5.6 gives the same guarantee for an algorithm (agile OMD) that, unlike RFTL, maintains a feasible point at every round, which the book notes is preferable in the adaptive-regret setting of Chapter X.
Formalizing it. Both theorems have complete, elementary proofs in the source (no gaps, no "with high probability", no hidden regularity conditions); the mission's work is converting the analytic argument — the Bregman-divergence identity, the generalized Cauchy-Schwarz inequality bounding the drift term by the local norm, and the two induction arguments underlying Lemma 5.4 — into machine-checked statements. No formalization of RFTL, OMD, or the local-norm machinery exists on the platform (checked below); the closest Formalpedia entries state a related but distinctly narrower result.
The central obstacle is that the regularizer is a hypothesis, not a fixed function: the theorem must hold for every admissible simultaneously, so nothing about beyond its stated properties (strong convexity, smoothness, twice differentiability) may be used. A newcomer's first instinct — bound the local norm by a fixed multiple of the Euclidean dual norm — fails in general and is exactly the bound RFTL is designed to avoid needing; the whole point of the local-norm formulation is that it can be tight for regularizers (like entropy) whose Hessian is very far from a multiple of the identity. A second obstacle is Lemma 5.4's induction, which compares (a minimizer over terms) against using the minimality of at exactly the right instantiation — an argument that looks almost circular until the induction hypothesis is applied at , not at the theorem's free variable.
ranges over an arbitrary real, complete inner product space (a real Hilbert space),
matching Chapters III and IV, not a fixed . The RFTL and agile-OMD update
rules are represented relationally (IsArgMinOn), since Mathlib has no canonical argmin
operator for a general convex set — mirroring IsMetricProjection's precedent from Chapter
III. The Hessian at the mean-value-theorem's intermediate point is represented via the
second Fréchet derivative of 's gradient map (HasFDerivAt), since Mathlib has no
dedicated Hessian type; the local dual norm is then any value satisfying the resulting
existential characterization (IsLocalDualNormSq), stated once and shared by both targets.
A boundedness hypothesis on over is added to Lemma 5.3's statement to keep the
-diameter from collapsing to Mathlib's junk value for an unbounded supremum — a
condition every regularizer the book actually uses (strongly convex and smooth over a
bounded ) already satisfies, so it narrows nothing.
Trivializing formalization ruled out. A regret bound stated for an "algorithm" defined
loosely enough to include the after-the-fact optimal choice would be vacuous; IsRFTLRun
and IsOMDAgileRun instead pin down the exact history-dependent update rule of Algorithms
13 and 14 (the current gradient sequence, the current regularizer, and nothing else) as a
hypothesis, so a proof must genuinely use the specific update. Theorem 5.2 and Theorem 5.6
are kept as two separate items rather than one theorem parameterized by an algorithm choice,
since — per the chapter's own remark that their analyses are unrelated — a merged statement
would either need to branch internally on the algorithm or silently identify two genuinely
different update rules.
Reuse and prior art. OnlineConvexOpt.FirstOrder.RegretT (Chapter III, published) is
imported and reused verbatim, keeping the regret functional identical across the whole
book. Definitions specific to Chapter IV (OnlineConvexOpt.SecondOrder, not yet published)
are not imported per this series' convention that a draft cannot import another draft;
quadForm is redeclared locally instead. On the platform, BanditAlgorithm.ftrl_regret_bound,
BanditAlgorithm.mirror_descent_regret_bound, and BanditAlgorithm.ftrl_simplex_exp_weights_regret
(the Bandit Algorithms series, Chapter XII) state regret bounds for FTRL and Mirror Descent
in the linear-cost, bandit-idiom setting (a fixed linear loss at
each round, regret compared via a Bregman-divergence potential at fixed points). Hazan's
Theorem 5.2 and 5.6 are for general convex and use the book's own local-norm
object, which has no counterpart in those statements; they are read in full and are not
faithful substitutes (different hypothesis class), so this mission drafts its own,
independent items rather than reusing them.
Online decision-making problems — multi-armed bandits, contextual bandits, structured bandits, and episodic reinforcement learning — look superficially different but share a common shape: a learner repeatedly acts, observes feedback, and is scored by regret against the best action in hindsight. Foster, Kakade, Qian and Rakhlin's Foundations of Reinforcement Learning and Interactive Decision Making (Foster & Rakhlin, arXiv:2312.16730v1) develops a unifying account of this shape and asks a sharper question than "does this specific algorithm work?": for a given class of possible environments, what is the best regret any algorithm can achieve? The Decision-Estimation Coefficient (DEC), introduced by Foster, Kakade, Qian and Rakhlin (2021, "The Statistical Complexity of Interactive Decision Making") and refined by Foster, Golowich, Qian, Rakhlin and Sekhari (2023), was proposed as the answer: a single real-valued complexity measure of a model class that simultaneously (i) drives a generic optimal-up-to-constants algorithm (Estimation-to-Decisions, E2D), and (ii) lower-bounds the regret of every algorithm. Item (ii) is what turns the DEC from "a complexity measure that happens to work for the algorithms we know" into a genuine characterization of statistical difficulty, in the same sense that minimax rates characterize the difficulty of estimation problems in classical statistics. This mission formalizes that lower bound.
Chapter 6 of the book (pp. 93–128) introduces Decision Making with Structured Observations (DMSO), a protocol general enough to subsume the contextual-bandit, structured-bandit and episodic tabular-RL protocols of earlier chapters. Over rounds, the learner selects a decision from a decision space ; nature draws a reward-observation pair from a fixed, unknown model , where a model maps each decision to a distribution over a reward space and an observation space . The learner has access to a model class containing (realizability). For , write for the mean reward function and for the optimal decision; regret is , exactly as in the bandit chapters, now for the general model class.
Because observations, not just mean rewards, now carry information, the DEC needs a way to measure distance between the full conditional distributions and , not just between scalars and . The chapter uses the squared Hellinger distance , one of a family of Csiszár -divergences that also includes total variation () and Kullback-Leibler () divergence. For a reference model and scale , the general Decision-Estimation Coefficient is the min-max game value
and . This mission's Lean development (FoundationsRL.GeneralDM)
formalizes discrete versions of , , for a finite outcome type, the DMSO
regret, and this DEC.
The goal is Proposition 28 (DEC Lower Bound), p. 105:
Here is the constrained DEC (§6.5.1), a variant of the offset DEC above that hard-constrains the information gain rather than subtracting it — a technical refinement needed to make the lower-bound direction go through — and the "localization condition" is a genuine hypothesis of the proposition, not a footnote. Unlike almost every other target in this series of missions, the statement quantifies over every algorithm rather than naming one: it is a genuine impossibility result. Two supporting divergence facts are included as milestones because the DEC's information-theoretic argument rests on them: Lemma 19 () and Lemma 20 (a bounded-likelihood-ratio refinement bounding in terms of ). The chapter's own matching upper bound, Proposition 26 (the E2D regret bound for the general DMSO protocol, the direct analogue of Chapter 4's Proposition 13), is included as a milestone to give the reader the matching pair the chapter presents together. Finally, Corollary 1 restates the lower bound in terms of the localized offset DEC (combining Proposition 28 with Proposition 27), included as a milestone showing the lower bound's reach beyond the constrained DEC alone.
Proposition 28 is what makes the DEC a genuine characterization of the statistical complexity of interactive decision making, rather than merely a sufficient condition for a particular algorithm family to succeed. Combined with the (uncited, technically deeper) matching upper bound for the constrained DEC — Proposition 29, stated but not proved in the book — it shows that for any finite model class, the constrained DEC is necessary and sufficient for low regret up to a factor in the localization radius: no complexity measure that is substantially different from the DEC can characterize the same problems. This is the general decision-making analogue of how minimax rates pin down statistical estimation, now for interactive protocols with adaptive feedback.
Formalizing the lower bound is new work: no result of this shape exists on the Prove2Me platform
(searches for "decision-estimation", "general divergence", "constrained DEC" and "Hellinger" — the
last of which surfaces two related-but-distinct affinity/Le Cam bounds from a different mission on
bandit lower bounds — return no faithful prior art; see MODERATION_NOTES.md). The formal
statement is the boxed proposition; the book gives a self-contained but simplified proof (two
named simplifying assumptions, §6.5.3) and cites Foster, Golowich, Qian, Rakhlin & Sekhari (2023)
for the unrestricted argument. This mission's Lean items are draft statements (:= by sorry), not
proofs; formalizing the proof itself — a two-point adaptive testing argument using the chain rule
for KL divergence and a change-of-measure step — is the open contribution this mission proposes.
The obvious first attempt is to try to prove the lower bound by exhibiting one fixed pair of hard models , as in classical two-point minimax lower bounds (Le Cam's method, Fano's inequality). This fails here because the decision-making protocol is interactive and adaptive: the algorithm's queries depend on what it has observed, so a model pair chosen obliviously (before seeing the algorithm) cannot in general be made indistinguishable to every algorithm — an adaptive algorithm can be constructed that distinguishes any two fixed models quickly by querying where they differ. The book's proof instead selects the "hard" alternative model as a function of the algorithm's own strategy (via the constrained DEC's arg max, Eq. (6.36)), so that the pair is hard specifically for the algorithm under consideration, then uses the chain rule for KL divergence plus the change-of-measure identity between the algorithm's induced distributions under and to conclude that the algorithm's realized decisions must look similar under both models — hence it cannot get low regret on both simultaneously. Every step of this argument depends on the exact game structure of the constrained DEC, not just its numerical value; a formalization that leaves as an unconstrained real parameter (rather than the actual - game with its information-gain constraint) would make the lower bound's conclusion vacuous, since the hypothesis would no longer track any actual property of .
The decision space and the outcome (reward, observation) alphabet are both taken as
finite types (Fintype); a model is a conditional probability
vector, and a reward-extraction map recovers the mean reward
. hellingerSq, totalVariationDiscrete,
klDivDiscrete specialize the book's general dominating-measure divergence formula (Eq. (6.5)) to
the counting measure on this finite type; klDivDiscrete returns an ENNReal so its
case (when is not absolutely continuous w.r.t. ) is represented honestly. The DEC, the
constrained DEC and the localized subclass are literal sInf-of-sSup/sSup-of-sSup
transcriptions of the book's min-max games — the same convention this series uses for the
Chapter-4 DEC — not opaque free real numbers, which rules out the trivializing formalization named
above.
Three deviations from this series' usual convention of pinning every constant to the value the
book's own proof derives are deliberate and disclosed. First, the numerical constant in
is explicitly called "not important" by the authors themselves
(footnote a, p. 105); it is existentially quantified (∃ c > 0) rather than pinned to a numeral.
Second — added at moderation, round 2, 2026-09-19, after the constant was found to be pinned
incorrectly — the lower bound's own multiplicative constant is also existentially quantified
(∃ c' > 0) rather than pinned to 1/20. The book's printed proof (§6.5.3, pp. 107–110) derives
1/20 (p. 110, not p. 109 as an earlier draft of this mission stated) only under two named
simplifying assumptions the theorem's hypotheses do not carry (p. 107, "Simplifications": a
class-wide bounded-curvature hypothesis, Eq. (6.34); and a bound on the unaugmented
rather than the officially-defined,
augmented this mission's decC implements). Since
augmenting either supremum's domain can only raise its value, the printed proof's bound on the
narrower, unaugmented quantity does not license a pinned 1/20 against the fully general decC
this theorem states; the book itself attributes the proof of the general statement to an external
reference (Foster, Golowich, Qian, Rakhlin & Sekhari 2023) not in this document. The existential
c' matches the book's own unpinned for Proposition 28 as printed on pp. 105–106.
Third, "any algorithm" and are formalized, as throughout
this series, without a full stochastic-process/history model: regret is a deterministic quantity
evaluated at a fixed realized decision-distribution sequence , rather than an expectation over an adaptive, history-dependent algorithm's own
randomness. Formalizing the fully adaptive, measure-theoretic version of "any algorithm" — with an
explicit filtration and expectation over the induced process law — is future work a solver
could add; the current statement is faithful to the book's deterministic-per-realization content
but not to its full generality over randomized, history-dependent strategies. The DMSO protocol
(Def_FoundationsRL_GeneralDM_Protocol) and the DEC (Def_FoundationsRL_GeneralDM_DEC) are
restated locally rather than imported from Chapter 4's mission (FoundationsRL.Structured), since
draft items cannot import another chunk's drafts; contributions extending either mission to reuse
the other's substrate once both are published are welcome.
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.
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.
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).
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.
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.
(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.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).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.Two-stage stochastic programs with recourse — choose a first-stage decision now, observe a random outcome , then choose a second-stage recourse decision to repair whatever 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 ranges over a finite set of scenarios, the recourse function that averages the second-stage cost over scenarios is piecewise linear and convex in , 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 (and a scalar approximating the recourse cost) and, at each candidate , a batch of second-stage linear programs that either certify 's second-stage feasibility or supply a linear underestimate — a cut — of around . 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.
A two-stage recourse instance consists of a first-stage feasible region for , and, for each of finite scenarios (occurring with probability ), second-stage data defining the recourse subproblem
where the recourse matrix is fixed — the same across every scenario, the case this chapter treats. is the set of for which every scenario's subproblem is feasible, and the two-stage problem is
A basis of the recourse subproblem is an injective choice of of 's columns (where is 's row count); each basis determines a simplex multiplier , and when attains the true optimum of , LP duality gives — the mechanism that turns a batch of second-stage LP solves into linear cuts on .
The L-shaped algorithm proceeds in three steps, repeated until neither applies:
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 ) or its optimum is second-stage feasible, passes every fresh Step-3 test, and is optimal for the two-stage problem.
A shortcut avoiding separate feasibility LPs at Step 2: under these structural assumptions on , checking feasibility at the single componentwise-worst right-hand side certifies feasibility at every scenario simultaneously.
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 exists.
The natural shortcut — state only "an optimal exists, or " — is not Theorem 2's actual content and is not what this mission targets: that weaker claim would already follow from 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 jointly, but before the first one it optimizes alone; conflating the two (e.g. always requiring to be part of the optimum) does not match Step 1 as the book states it.
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 columns of
"; 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
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 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- 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.
Von Neumann introduced amenable groups in 1929 to explain the Hausdorff–Banach–Tarski paradox, and showed that the class of amenable groups contains all finite and all abelian groups and is closed under four processes: (I) subgroups, (II) quotients, (III) extensions and (IV) directed unions. Day named the smallest class with these properties , the elementary amenable groups. For fifty years these were the only amenable groups anyone could exhibit, and von Neumann's question whether every non-amenable group contains a free subgroup on two generators — whether equals the class of groups without such a subgroup — was open. (It was answered in the negative by Ol'shanskii in 1980, the year of this paper, by different methods.)
Ching Chou's Elementary amenable groups (Illinois J. Math. 24 (1980) 396–407, doi:10.1215/ijm/1256047608) gives the structure theory of that everything later relies on. Its central result is that the class can be built from finite and abelian groups by extensions and directed unions alone — subgroups and quotients add nothing (Proposition 2.2). From that description three things follow: periodic elementary amenable groups are locally finite, so the periodic non-locally-finite groups of Golod and Novikov–Adjan show (Theorem 2.3); a finitely generated simple elementary amenable group is finite (Corollary 2.4); and Wolf's conjecture holds in : a finitely generated elementary amenable group is almost nilpotent or has exponential growth (Theorem 3.2, extending Milnor and Wolf's theorem for solvable groups). A final section introduces a packing property (P) of groups and proves it for every elementary amenable group (Proposition 4.2) and every residually elementary amenable group (Corollary 4.7).
On this platform the definition of is already published
(the bundle Chou_ElementaryAmenable, from the mission
Cannon–Floyd–Parry: Thompson's group F and the simplicity of its commutator subgroup),
together with the theorem that Thompson's group is not elementary amenable
(Theorem 4.10) and, from Brin and Squier, that has no free subgroup on two
generators (CannonFloydParry.no_free_subgroup_of_rank_two). This mission formalizes
Chou's paper on top of that definition.
The class and its constructible core. Chou.ElementaryAmenable G is an inductive
predicate on groups: finite groups and abelian groups are in the class, and the class is closed
under isomorphism, subgroups, quotients, extensions and directed unions of subgroups; each rule
is a constructor of the published bundle Chou_ElementaryAmenable, where it is stated precisely. Chou builds the hierarchy
by transfinite recursion, applying only extensions and directed unions to the finite and
abelian groups, and proves that is closed under subgroups and
quotients, hence equals . The union is realised here without
ordinals, as the inductive predicate Chou.Constructible, whose constructors are of_finite,
of_commGroup, of_mulEquiv, extension and directedUnion; Chou's transfinite induction
over becomes structural induction over a derivation, with the same case analysis.
Periodic and locally finite groups. A group is periodic if every element has finite order
(Mathlib's IsMulTorsion) and locally finite if every finitely generated subgroup is finite
(Chou.IsLocallyFinite). Day's class is Chou.NoFreeSubgroupOfRankTwo: no homomorphism
from the free group on two generators into is injective.
Growth. For a finite generating set of , Chou.wordBall S n is the set of products
of at most factors, each in or with inverse in . has exponential growth if for
some finite generating set the ball of radius has at least elements for some
and all ; it is exponentially bounded if for some finite generating set and every
the balls are eventually smaller than . Chou works with for products of exactly
elements of a finite generating set ; for symmetric and containing the identity the
two agree, and Wolf's observation that the growth type is independent of the generating set is
one of the milestones. "Almost nilpotent" is Mathlib's Group.IsVirtuallyNilpotent: a
nilpotent subgroup of finite index. A free subsemigroup on two generators means two elements
such that distinct positive words in are distinct in
(Chou.HasFreeSubsemigroupOfRankTwo).
Packings. A pair of subsets is a packing of if is a
bijection (Chou.IsPacking), and has property (P) if every finite
subset lies in a finite for which some is a packing (Chou.HasPackingProperty).
is residually elementary amenable if every survives in some elementary amenable
quotient (Chou.ResiduallyElementaryAmenable).
The goal is Chou's description of the class, Proposition 2.2 (b) (p. 397): “ is the smallest
class of groups which contains all finite groups and all abelian groups and is closed under
processes (III) and (IV).” It is stated as the equivalence ElementaryAmenable G ↔ Constructible G.
The milestones follow the paper's order.
Section 2. Proposition 2.1 in two halves — the constructible groups are closed under subgroups and under quotients — which is the whole proof of the goal. Theorem 2.3: periodic elementary amenable groups are locally finite; and its consequence that is nonempty. Corollary 2.4: finitely generated simple elementary amenable groups are finite.
Section 3. Lemma 3.1 (an extension of almost nilpotent by almost nilpotent is almost nilpotent or of exponential growth), Theorem 3.2 and Rosenblatt's sharpening Theorem 3.2′, together with the facts Chou uses on the way: a finite-by-nilpotent group is almost nilpotent; a free subsemigroup forces exponential growth; Wolf's independence of the generating set; and Milnor's existence of the growth rate, in the form "exponentially bounded means not of exponential growth".
Section 4. Property (P) for finite groups, for , for finitely generated abelian groups; Lemma 4.1 (directed unions and extensions preserve (P)); Proposition 4.2 (every elementary amenable group has (P)); Lemma 4.6 (a) and Corollary 4.7 (residually elementary amenable groups have (P)); and the free groups.
Chou's Section 3 rests on results the paper cites rather than proves, none of which is in Mathlib. They are stated here as milestones in their own right, so that the dependence is visible and each is a well-defined target: the Milnor–Wolf theorem (a finitely generated solvable group that is exponentially bounded is almost nilpotent); M. Hall's theorem that a finitely generated group has finitely many subgroups of each finite index; that finitely generated nilpotent groups are finitely presented and that a group with a finitely presented subgroup of finite index is finitely presented; Milnor's Lemmas 1–2 in the form Chou states on p. 400 (in a finitely generated exponentially bounded group, a normal subgroup with finitely presented quotient is finitely generated); and Rosenblatt's variant of Lemma 3.1. Section 4 needs one more: free groups are residually finite. Lemma 3.1 and Theorems 3.2, 3.2′ can be closed only once these are; every other milestone is provable from Mathlib and the published library.
Two remarks on Theorem 2.3. Chou's witness for is a periodic group that is not locally finite (Golod; Novikov–Adjan), whose existence is not formalized. The platform already holds a different witness: Thompson's group is not elementary amenable (Cannon–Floyd–Parry, Theorem 4.10) and has no free subgroup on two generators (Brin–Squier); both are published and proved, and the milestone is proved from them. The inclusion itself is von Neumann's theorem that amenable groups contain no free subgroup of rank two, which passes through the definition of amenability and is not part of this mission.
The ordinal-indexed hierarchy and the remark that it stabilises at some (Proposition 2.2 (a)) are replaced by the inductive predicate. Chou's two examples of finitely generated groups in that are not almost solvable (p. 402), the Golod–Shafarevich discussion, and Lemma 4.6 (b) (ordinal-indexed normal series) are omitted. Propositions 4.3–4.5 on almost convergent sets, and Milnor's remark that exponentially bounded groups are amenable, need invariant means on ; amenability itself is the subject of Garrido I.
Given independent random variables , when does converge? Not absolutely — that question is settled by and is usually too strong. The interesting question is when the partial sums converge for almost every outcome, and here independence buys something that holds for no general sequence: convergence is not a delicate matter of cancellation but is decided, once and for all, by three numerical series.
Chapter 2 of Rick Durrett's Probability: Theory and Examples (Version 5, 2019) reaches this in section 2.5. Kolmogorov's three-series theorem fixes a truncation level , replaces each by , and asserts that converges almost surely if and only if
Three deterministic conditions on the distributions decide an almost-sure question about paths, and the answer does not depend on which is chosen. Through Kronecker's lemma this is also the route to the strong law of large numbers, which is how the chapter uses it.
Let be independent real random variables on a probability space, with partial sums . Say that converges almost surely when for almost every the sequence has a real limit; following Durrett, " converges" means exists, not that it converges absolutely.
Three tools from the same section support the theorem. Kolmogorov's maximal inequality strengthens Chebyshev from to the maximum of the whole path,
for independent, centred, square-integrable summands. From it comes the convergence criterion: if and then converges almost surely. Kronecker's lemma is the deterministic bridge to averages: if and converges then . And the Hewitt–Savage 0-1 law says that for an i.i.d. sequence every permutable event — one unchanged by rearranging finitely many coordinates — has probability or .
Both directions are asserted, as Durrett states the theorem. The truncation level is arbitrary and fixed in the statement; that the three conditions hold for one exactly when they hold for every is a consequence, not an assumption.
Kolmogorov's maximal inequality (2.5.5); the convergence criterion under summable variances (2.5.6); Kronecker's lemma (2.5.9); and the Hewitt–Savage 0-1 law (2.5.4).
The result itself. The three-series theorem is the complete answer to a question that has no complete answer without independence, and the shape of the answer is the interesting part: a pathwise, almost-sure property is equivalent to three conditions each computable from the marginal distributions alone. Each of the three does a separate job — the first says and its truncation differ only finitely often, so Borel–Cantelli lets them be exchanged; the second controls the drift of the truncated sums; the third controls their fluctuation. The theorem is also the standard route to the strong law: applying it to and then Kronecker's lemma gives , which is why section 2.5 sits where it does.
Formalizing it. Mathlib has the strong law of large numbers (strong_law_ae), both
Borel–Cantelli lemmas, and Kolmogorov's 0-1 law for the tail σ-field. It has none of the
following: Kolmogorov's maximal inequality, the almost-sure convergence criterion for random
series with summable variances, Kronecker's lemma, the Hewitt–Savage 0-1 law, or the three-series
theorem. The mission therefore contributes the whole of section 2.5, and the pieces are reusable
well beyond it — the maximal inequality and Kronecker's lemma in particular are standard tools with
no probabilistic content in the second case at all.
The maximal inequality is the step where the argument stops being routine. Chebyshev bounds and no more; controlling the maximum over the whole path needs the first passage decomposition and the observation that is measurable with respect to the first variables while is independent of them, so the cross terms vanish. That is a stopping-time argument in disguise, and it is what makes the whole section work.
The sufficiency half of the goal is then assembly: the third series and the convergence criterion give convergent, the second adds the means back, and the first plus Borel–Cantelli replaces by . Necessity is the harder direction, and Durrett does not prove it in Chapter 2 at all — he defers it to Example 3.4.12, where it follows from the Lindeberg–Feller central limit theorem. A solver attacking the goal should expect the reverse implication to need machinery from outside this section.
The Hewitt–Savage law has a difficulty of its own kind: the natural statement is about a σ-field of events on a sequence space, and the proof approximates a permutable event by cylinder events and then applies the permutation that swaps the first coordinates with the next .
Random variables are measurable real-valued functions on a probability space and independence is
Mathlib's iIndepFun. Variance is Mathlib's variance, and square-integrability is stated as
membership in where the maximal inequality and the convergence criterion need it. The
three-series theorem itself assumes no integrability: the truncated variables are bounded, so
their means and variances exist automatically, which is exactly why the truncation is there.
" converges" is formalized as convergence of the sequence of partial sums to a real
limit, not as Summable, which in Mathlib means unconditional and hence absolute convergence for
real series. This distinction is not pedantic here: condition (ii) of the theorem is convergence of
in Durrett's sense and would be a strictly stronger condition if read as
summability. Conditions (i) and (iii) are series of non-negative terms, where the two notions
agree, and are stated as Summable.
Almost-sure convergence of is "for almost every there exists a real with " — the limit is not asserted to be measurable in , and does not need to be for the statement to say what it should.
For the Hewitt–Savage law the sequence space is the countable product carrying
the infinite product of copies of one law, which is Mathlib's Measure.infinitePi, and a
permutable event is a measurable set invariant under every finitely supported permutation of the
coordinates. That is Durrett's exchangeable σ-field stated directly rather than constructed as a
σ-field object.
Contributions welcome beyond the listed items: the converse direction via Lindeberg–Feller (Example 3.4.12); the derivation of the strong law from the three-series theorem and Kronecker's lemma; the Marcinkiewicz–Zygmund law (2.5.12); and the rates of convergence of section 2.5.1.
Many hard combinatorial optimization problems — finding the maximum cut of a graph, deciding the ground state of an Ising spin system, bounding the correlation of a physical system — can be written as maximizing a bilinear form over sign vectors . Exhaustive search over sign patterns is intractable, so practitioners relax the problem: replace each sign by a unit vector in a higher-dimensional space and optimize the resulting inner products instead. This relaxation, a semidefinite program, is convex and solvable in polynomial time. The question is how much is lost in the relaxation — whether its optimal value can be far from the true, combinatorial optimum.
Grothendieck's inequality, proved by Alexander Grothendieck in 1953 in the context of Banach space theory (Résumé de la théorie métrique des produits tensoriels topologiques, Bol. Soc. Mat. São Paulo 8 (1953), 1–79), answers this for a broad class of such relaxations: replacing signs by unit vectors in an arbitrary Hilbert space changes the optimal value by at most an absolute, dimension-free constant factor. The inequality has since become a standard tool across combinatorial optimization, Banach space geometry, and (via the Goemans-Williamson algorithm for maximum cut, Section 3.6 of the source) approximation algorithms; see U. Haagerup, The Grothendieck inequality for bilinear forms on -algebras, Adv. Math. 56 (1985) for the tightest known constant, and Alon–Naor, Approximating the cut-norm via Grothendieck's inequality, SIAM J. Comput. 35 (2006), for the algorithmic connection this mission's Theorem 3.5.6 sets up.
Fix positive integers . Consider a real matrix . Say is normalized if for every choice of numbers ,
This says , viewed as a bilinear form on , has sup-norm at most . Now let be any real Hilbert space — a real vector space equipped with an inner product complete in the induced norm — and consider vectors and , each of unit norm . Replacing the scalar product by the inner product in the same bilinear form gives , a real number depending on the choice of and of the unit vectors. The question is how large this can be, uniformly over every such choice.
for every real Hilbert space and unit vectors , where is a constant depending on neither , its dimensions, nor . This mission's goal formalizes the book's own first-pass bound (Section 3.5), proved by a Gaussian truncation argument; it does not fix a numeral for , only that some absolute constant works, matching the shape of the true statement rather than a specific numeral that a sharper argument (the book's own Section 3.7 gives ) would immediately obsolete. See Formalization scope below for why this is the goal, not the sharper bound.
The result itself. Grothendieck's inequality is the single fact that makes semidefinite relaxation a provably good algorithmic strategy rather than a heuristic: whatever the true, hard-to-compute combinatorial optimum of a -valued bilinear optimization is, the tractable Hilbert-space relaxation cannot overshoot it by more than the constant . Milestone Theorem 3.5.6 makes this concrete for positive-semidefinite matrices, showing the semidefinite relaxation SDP of the integer program INT satisfies INT SDP INT — the guarantee underlying the Goemans-Williamson 0.878-approximation algorithm for maximum cut (Theorem 3.6.5 of the source, out of scope for this mission; see Formalization scope).
Formalizing it. The inequality and its two chapter milestones are proved but not previously
formalized on this platform (checked by concept search for "Grothendieck", "semidefinite",
"positive-semidefinite", and "max-cut" — no hits beyond the unrelated Grothendieck-Teichmüller
group). What remains after this mission is the sharper argument of Section 3.7
(the "kernel trick"), a separate, heavier development building on positive-definite kernels,
and full proofs of every milestone below (currently open sorry goals).
The statement of Grothendieck's inequality contains no randomness, yet every known elementary proof is probabilistic; this is itself a striking feature of the result. The obvious approach — bound directly by exploiting the normalization hypothesis on — fails because the normalization hypothesis only controls against sign vectors, and there is no way to project an arbitrary unit vector in a Hilbert space onto without losing information. The book's proof instead represents each unit vector via a scalar Gaussian random variable for a single Gaussian vector , recovering the inner products in expectation (Exercise 3.3.5); but these Gaussian variables are unbounded, so the normalization hypothesis (which bounds against bounded inputs) cannot be applied to them directly. The core technical step is a truncation argument: splitting each Gaussian variable into a bounded part and a small--norm unbounded remainder, applying the hypothesis to the bounded parts, and bounding the remainder terms by treating them as elements of the Hilbert space and invoking the very inequality being proved (Theorem 3.5.1 itself, applied with ) as a self-referential bootstrap — this is why the proof fixes as the smallest valid constant before starting, rather than building it up from scratch.
The goal and both milestones work with the real matrix and real inner product space directly;
H is required to be a complete real inner product space (NormedAddCommGroup,
InnerProductSpace ℝ, CompleteSpace), matching the book's "any Hilbert space." No
dimension bound on is imposed — the inequality's content is exactly that does not grow
with .
This mission does not formalize the sharper bound of Section 3.7, nor Theorem 3.6.5 (the 0.878-approximation guarantee for maximum cut via randomized rounding): the latter's statement quantifies over "the result of a randomized rounding of the solution of the semidefinite program," which would drag a specific algorithm into the audited statement rather than keeping it a self-contained mathematical claim (the statement/proof-separation trap this series' triage rubric flags). Grothendieck's identity (Lemma 3.6.6), the key fact behind that rounding step, is included on its own as a milestone, stated with an explicit, named random sign variable rather than an opaque "rounding procedure."
A trivializing formalization would state the goal with allowed to depend on , , ,
or — every such bound is easy (e.g. ) and carries none of the
theorem's content; the Lean statement rules this out by quantifying before every other
object. and (Theorem 3.5.6) are defined from scratch in this
chunk's namespace, using Matrix.PosSemidef from Mathlib for the positive-semidefiniteness
hypothesis (which bundles the real-symmetric condition); Mathlib has no ready-made SDP-value
construction to reuse. The sub-gaussian (Orlicz ) norm used by Theorem 3.1.1 is reused,
unchanged, from the 01-concentration mission in this series (HighDimProb.Concentration.SubgaussianNorm)
rather than redefined.
Any dataset of points can be described exactly by embedding it in for large enough — but a large is expensive: nearest-neighbor search, clustering, and streaming algorithms all scale with the ambient dimension, not with . The question that opens this mission is whether the dimension can be cut down while leaving the data's geometry — the pairwise distances between points — essentially untouched.
Johnson and Lindenstrauss answered this in 1984, while studying extensions of Lipschitz maps into Hilbert space (W. Johnson, J. Lindenstrauss, Extensions of Lipschitz mappings into a Hilbert space, Contemp. Math. 26 (1984), 189–206): points in any Euclidean space, of any dimension , can be mapped by a single linear map into a space of dimension only , distorting every pairwise distance by at most a factor of . The map does not depend on the data beyond its cardinality — a single random object works simultaneously for the whole point set with high probability. This is now one of the standard tools of randomized dimension reduction, cited across nearest-neighbor search, streaming linear algebra, compressed sensing, and machine learning pipelines that need to shrink feature dimension before a downstream algorithm runs.
Fix a probability space . A random orthogonal projection of rank in is a map , continuous and linear for each , such that almost surely is idempotent (), self-adjoint, and has range of dimension — i.e. is the orthogonal projection onto some -dimensional subspace . It is uniformly distributed in the Grassmannian (written ) when its law is rotation invariant: for every orthogonal transformation of , the conjugated map has the same law as . Conjugating a projection by is exactly the projection onto the image of its range under , so this says the law of the random subspace is invariant under the full orthogonal group — the operational definition Vershynin himself uses for a "uniformly distributed" random subspace, since no coordinate-free formula for such a subspace's law is given directly.
A companion notion drives the proof: a random vector is uniform on the Euclidean sphere of radius , , when it lies on that sphere almost surely and its law is likewise rotation invariant. And a real random variable is sub-gaussian with sub-gaussian () norm , the standard non-asymptotic measure of how light-tailed 's distribution is (a bounded or Gaussian random variable has finite norm; the tail probability then decays at least as fast as ).
for every finite , every , and every random orthogonal projection of rank uniformly distributed in . The universal quantifier over pairs sits inside the single probability event — this is the union-bound content that makes the statement a genuine simultaneous guarantee for the whole point set, not a restatement of the single-vector lemma below for one fixed pair. Both constants are the book's own unnamed absolute constants, never depending on , , , or ; this is the weakest stable form of the claim (no numeral is hard-coded for or ), matching the book's own statement exactly.
The lemma gives a universal, data-oblivious dimension-reduction guarantee: the target dimension depends only on the number of points and the desired distortion, never on the ambient dimension or on the geometry of the specific point set. This is what makes it usable as a black-box preprocessing step ahead of an algorithm whose cost scales with — the projection is drawn once, without looking at the data, and works with high probability for every pairwise distance simultaneously. The bound is also known to be essentially optimal in : Alon (Problems and results in extremal combinatorics, Discrete Math. 273 (2003)) showed a lower bound of on the target dimension, so the dependence cannot be removed.
The theorem itself has been proved for decades and admits several proof strategies (this book's route through Lipschitz concentration on the sphere; the original volume/measure-concentration argument; later "sparse" or structured variants of the projection for faster computation). This mission formalizes the classical dense-Gaussian-projection proof route as Vershynin presents it, building the sphere-concentration engine (Theorem 5.1.4) and the single-vector projection lemma (Lemma 5.3.2) that the union-bound argument for the goal rests on. No machine-checked formal proof of this chain is known to exist on the platform prior to this mission (see Formalization scope below); what is contributed is the statement infrastructure — the goal and its two direct supporting lemmas, stated with explicit, unpinned absolute constants — for solvers to close.
The natural first idea — bound the distortion of a single fixed vector under a random projection, then take a union bound over the pairwise differences — is exactly the strategy Lemma 5.3.2 and the goal use, but it does not by itself explain why the single-vector concentration bound (Lemma 5.3.2(b)) holds with the stated sub-gaussian-type tail. That bound is not elementary: it reduces to a uniform concentration statement for an arbitrary Lipschitz function of a uniformly random point on a high-dimensional sphere (Theorem 5.1.4), since , viewed as a function of a rotated copy of , is a -Lipschitz function on the sphere. Proving that every Lipschitz function concentrates — not just linear ones, for which sub-gaussianity was already established in Chapter 3 — needs a genuinely different tool: comparing the sub-level sets of an arbitrary Lipschitz function to spherical caps via an isoperimetric inequality on the sphere. This geometric input is what makes the concentration phenomenon behind Johnson-Lindenstrauss a dimension-free fact rather than a special property of coordinate projections.
is a Finset of points in EuclideanSpace ℝ (Fin n), matching "a set of points"; is
read off as X.card. The random subspace is represented throughout by the
orthogonal projection onto it (IsUniformProjection), following the book's own statements,
which are phrased in terms of rather than ; the scaled map of the goal is
written Real.sqrt (n/m) • P ω applied to x - y, using linearity of to realize
. Both "uniform on the sphere" and "uniform in the Grassmannian" are defined
operationally by rotation invariance of the underlying law, since Mathlib has no ready-made
normalized surface measure on a general-radius Euclidean sphere or Haar-measure construction on
the Grassmannian/orthogonal group to build a canonical uniform object from; rotation invariance
uniquely determines the corresponding measure among those supported on the relevant set, so the
operational and constructive definitions coincide extensionally. Every "absolute constant" in the
book ( in Theorem 5.3.1's sample-complexity hypothesis, in every failure-probability bound,
and the sub-gaussian constant of Theorem 5.1.4) is existentially quantified ahead of the
dimension, sample size, and every other object, and pinned to no numeral — a formalization that
hard-coded a specific numeral for any of these would be invalidated by the next sharper constant
in the literature and would not match what the book actually proves.
A trivializing formalization is one that states the conclusion for a single fixed pair rather than universally over all pairs inside one event; that would collapse the union-bound content that makes this a dimension-reduction statement for a whole point set (with points), rather than a restatement of the single-vector Lemma 5.3.2(b). This mission's goal statement is built to rule that out explicitly (see Formalization targets above).
Reusable infrastructure: subgaussianNorm (the Orlicz norm, restated per Vershynin
Definition 2.5.6) and the rotation-invariance idiom for "uniformly distributed" random geometric
objects are of independent interest to any later chapter needing sub-gaussian random vectors or
random subspaces/projections (e.g. Chapters 4, 6, 7, 9, 11 of this same book series). Solvers'
contributions are welcome on: the isoperimetric inequality on the sphere and its use to prove
Theorem 5.1.4 (the mission's hardest open leaf); the coordinate-projection computation underlying
Lemma 5.3.2(a); and the concentration-plus-union-bound argument closing the goal from the three
supporting lemmas.
A recurring question in economics and operations research is: when a decision problem depends on a parameter, does the optimal decision move monotonically as the parameter changes? A firm's optimal input mix as a price rises, a consumer's optimal consumption bundle as income grows, a Cournot firm's optimal output as a rival's output changes — in each case one wants "more of the parameter implies (weakly) more of the optimum" without assuming convexity, differentiability, or a unique optimizer. The classical tool for such comparative statics questions is the implicit function theorem, which needs smoothness and a nondegenerate Hessian and breaks down the moment the optimum is not unique or the objective is not differentiable. Topkis [1978] showed that a purely order-theoretic condition — supermodularity of the objective jointly in the decision variable and the parameter — is sufficient on its own, with no smoothness, uniqueness, or convexity assumed at all, and Milgrom and Roberts [1990a, 1994] later showed this lattice-theoretic approach subsumes and strengthens the classical monotone-comparative-statics results in economics. This mission formalizes the two central results this book calls "Topkis's theorem" (Theorem 2.8.1 and Theorem 2.8.2), together with the structural fact about maximizers of a supermodular function (Theorem 2.7.1) that both rest on, and the strengthening to strictly ordered optimal selections (Theorem 2.8.4).
Let be a lattice: a partially ordered set in which every pair
has a join and a meet . A real-valued function
is supermodular on if
for all ; this is
the same relativized notion (SupermodularOn) used, with , throughout chunk I of
this series.
Now let also be a partially ordered set (the parameter set), and let be a real-valued function of the pair . has increasing differences in if, for every in , the map is monotone (order-preserving) in ; equivalently, the marginal gain from raising is itself increasing in . Replacing "monotone" with "strictly monotone" gives strictly increasing differences. To compare the resulting sets of optimizers rather than single points, this mission reuses the induced set ordering from chunk I: for , holds when and for all , .
Let and be lattices, let be a sublattice of the product lattice , and let be the section of at . If is supermodular on (jointly in the pair ), then
is increasing in , with respect to , on .
With increasing in (with respect to ), supermodular in for each fixed , and having increasing differences in on , the same conclusion — increasing in — holds. Theorem 2.8.2's joint-supermodularity hypothesis on a sublattice of automatically forces both of Theorem 2.8.1's hypotheses, so 2.8.1 is the logically weaker, more elementary statement from which 2.8.2's proof proceeds.
Under the hypotheses of Theorem 2.8.1 but with strictly increasing differences, every individual optimal solution at a larger parameter value dominates every individual optimal solution at a smaller one: , , and together force — a genuinely stronger conclusion than alone gives.
A supporting result is formalized as a milestone because both goals' proofs use it directly: Theorem 2.7.1, that is a sublattice of whenever is supermodular on — the structural fact that makes it meaningful to compare optimal-solution sets with in the first place.
The result itself. Theorem 2.8.2 is the book's own headline theorem, cited throughout the rest of the monograph: it underlies the assortative-matching existence theorem (Chapter 3), monotone optimal policies in Markov decision processes (Chapter 3), and equilibrium comparative statics in supermodular games (Chapter 4) — each a later mission in this series. Its distinguishing feature relative to the implicit function theorem is that it needs no differentiability, no uniqueness of the optimizer, and no interiority: it applies equally to discrete decision problems (integer programming, combinatorial selection) and continuous ones.
Formalizing it. Nothing in Mathlib currently states a parametric monotone-comparative-
statics result of this shape: the closest neighboring material (order-preserving maps,
MonotoneOn, lattice structures) supplies only the vocabulary, not the theorem. This
mission is the first formalization of Topkis's theorem on this platform and introduces
the increasing-differences vocabulary (IncreasingDifferencesOn,
StrictlyIncreasingDifferencesOn) that later missions in this series (matching, MDPs,
supermodular games) reuse directly.
The natural first idea — differentiate in , set the gradient to zero, and use the implicit function theorem on the resulting first-order condition — fails immediately because nothing here is assumed differentiable, and need not be a single point. The correct argument instead compares two arbitrary elements , directly through the supermodularity inequality applied to the pair against (a chain of inequalities Topkis calls "Lemma 2.8.1"), using increasing differences only to move the parameter from to inside that chain — at no point is a derivative, a selection function, or an interior point used. A second subtlety is that "increasing" in the conclusion is with respect to the induced set order , not a claim that some selection is monotone: proving the stronger, pointwise-ordered conclusion (Theorem 2.8.4) genuinely needs the strict form of increasing differences, not merely increasing differences plus an extra hypothesis.
and are kept as abstract Lattice/PartialOrder types throughout, matching the
book's own generality — Theorem 2.8.1's and 2.8.2's /
corollary via second partial derivatives (discussed in the book's prose immediately after
Theorem 2.8.2, p. 77) is not itself a numbered theorem and is not formalized here.
Supermodularity, increasing differences, and strictly increasing differences are each
formalized as a single relativized definition (SupermodularOn f S,
IncreasingDifferencesOn f S, StrictlyIncreasingDifferencesOn f S) so the same
declaration expresses both "supermodular on the whole lattice " (used by Theorem 2.7.1
and Theorem 2.8.1's per- hypothesis) and "jointly supermodular on a sublattice of
" (Theorem 2.8.2) — a formalization that instead only ever supermodularized
for fixed would collapse Theorem 2.8.2's genuinely joint hypothesis into
a restatement of Theorem 2.8.1, which is exactly the trivialization this mission's chunk
brief warns against. is written out as the set
of that dominate every other element of under , and every
conclusion is stated only for pairs at which both argmax sets are assumed
nonempty — matching the book's own restriction to , since holds
vacuously whenever either side is empty. This mission depends on chunk I's
InducedSetOrder; it introduces no reusable infrastructure beyond its own three
definitions, which later missions in the series (matching, MDPs, supermodular games) are
expected to import directly rather than redefine.
This mission formalizes §4 of Cannon, Floyd and Parry's Introductory notes on Richard Thompson's groups, together with the definition of Thompson's group from their §1. The goal is their Theorem 4.5: the commutator subgroup is simple.
In the 1960s Richard Thompson defined three groups, now written , and , whose properties have kept them in use ever since as a source of examples at the edge of what groups can do. is the smallest of the three and the least understood. It is finitely presented (§3 of the source) and torsion-free, it has no free subgroup of rank two, and whether it is amenable — whether it carries a finitely additive left-invariant probability measure defined on all its subsets — is open. Cannon, Floyd and Parry record (§4, p. 227) that Geoghegan raised the question and conjectured in 1979 both that contains no non-Abelian free subgroup and that is not amenable.
That question is what makes worth pinning down precisely. Write for the class of
amenable discrete groups, for the elementary amenable ones, and for the groups with
no free subgroup of rank two. That was noted by
Day and follows from
von Neumann; whether it is strict is the von
Neumann–Day problem. It is: Olshanskii proved in a 1984 ICM address and
Gromov gave an independent proof — but by
examples that are not finitely presented. Brin and Squier proved in 1985 that , and
is not elementary amenable (Theorem 4.10 of the source, CannonFloydParry.not_elementaryAmenable_F). So
is a finitely presented group in if it is amenable and in
if it is not — a question with no other finitely presented candidate.
Call a real number dyadic if it has the form with and .
Thompson's group , as §1 of the source defines it, is the set of piecewise linear homeomorphisms of the closed unit interval onto itself that are differentiable except at finitely many dyadic rationals, and whose derivatives, where they exist, are powers of . Since those derivatives are positive, every element preserves orientation, so the elements of are increasing. Composition of two such maps is again one, and so is the inverse of one, so is a group.
The formalization calls such a map piecewise linear over the dyadics, and defines as the subgroup generated by those maps — so that closure under composition and inverses is a theorem rather than part of the construction, as the source has it. What the model fixes rather than derives is under Formalization scope below.
Two particular elements generate it. Write
An element of is trivial near if it fixes every point of some interval , and trivial near if it fixes every point of some . The support of is the set of points of that moves. The commutator convention throughout is , and denotes the commutator subgroup.
This is the capstone of §4: it says the commutator subgroup has no normal subgroup other than itself and the trivial one. It is the goal because the rest of the section feeds it — both halves of Theorem 4.1, Theorem 4.3, and both supporting lemmas below are consumed by its proof.
So has no interesting proper quotients at all. With the first part of Theorem 4.1 this forces every nontrivial normal subgroup of to contain .
That the piecewise-linear maps are already closed under composition and inverses, so that consists of exactly those maps; a transitivity lemma on dyadic partitions of ; the fact that the subgroup of elements supported in a dyadic interval of dyadic length is isomorphic to itself; triviality of the center; that contains no non-Abelian free group; and that admits a total order invariant under multiplication on both sides.
What the results give. Theorem 4.1 identifies concretely — a subgroup defined by a global algebraic condition turns out to be cut out by local behavior at the two endpoints — and computes the abelianization, making the pair of endpoint slopes a complete invariant of modulo commutators. Theorem 4.3 and the simplicity of together determine the whole normal subgroup lattice: every normal subgroup of is trivial or contains . That lattice is the input to the elementary-amenability argument.
What formalizing adds. All of these are proved in the source; none is in Mathlib, which has no piecewise-linear homeomorphism API and no Thompson group. Four of the milestones are proved as part of this proposal: that the piecewise-linear maps form a subgroup, that elements of permute the dyadic rationals, that embeds in the group Brin and Squier work with, and the absence of a free subgroup of rank two, which follows from the already-formalized Brin–Squier theorem via that embedding. The rest are open. The piecewise-linear machinery built along the way — local affineness, dyadic-breakpoint bookkeeping, extension by the identity — is reusable for , for , and for the wider family of piecewise-linear homeomorphism groups.
The obvious approach to the goal is to argue that a normal subgroup of containing a nontrivial element must be everything, by conjugating that element around. It fails on its own: an element of is pinned down only by being trivial near the two endpoints, and one still has to manufacture — inside , not merely inside — an element carrying a prescribed pair of neighborhoods into those. That construction is what the dyadic-partition transitivity lemma supplies, and it is where the combinatorics of dyadic subdivision enters.
The second difficulty was that the source proves §4 using the tree-diagram normal form of §2.
That section is now formalized in its own mission, Cannon–Floyd–Parry §2: tree diagrams and the
normal form (mission ffd1e4ea-9f9a-4cb6-8419-78e70f2545e8), all of whose milestones are proved.
Corollary 2.6 — milestone 5 here, the same theorem object — is closed from there, and Theorem
2.5 (represents_word_exponents) and the normal form (existsUnique_normalForm) are available to
a solver attacking Theorem 4.3, so the source's argument can now be followed. A solution file
imports only definitions, so whatever it uses from §2 must be reproved inline; the §2 solutions
are public and written to be reused that way. The piecewise-linear route — dyadic-partition
transitivity, Lemma 4.4 and Theorem 4.1 — remains an alternative, and is what Theorem 4.5's own
argument uses.
The unit interval is as a subtype, and an element of is an order isomorphism of it, so orientation preservation is built into the representation rather than derived — faithful to the source's set, but assuming one sentence CFP prove. Piecewise linearity is stated as: there is a finite set of dyadic reals such that the map is affine, with slope a power of two, on every closed interval whose interior misses . Intercepts are not required to be dyadic — that is derived by induction along the breakpoints, not part of the definition.
The definition is not vacuous: and of Example 1.1 are constructed explicitly, and that is not the trivial group is one of the milestones below — so no statement here is satisfied by the trivial group. In particular the goal, which asserts simplicity and therefore nontriviality, is not trivially false.
A companion definition places the same data on the real line, each element extended by the identity outside ; that line realisation is what the bridge statement connects to Brin and Squier's group.
The Riemann hypothesis (RH) asserts that every zero of the Riemann zeta function in the strip has . One of the few reformulations that turns RH into a positivity statement, rather than a statement about the location of points, goes back to A. Weil (1952): the explicit formula expresses a sum over the zeros of as a sum of local contributions over the places of , and RH is equivalent to the resulting functional being positive on elements of the form .
Connes' 1999 programme paper Noncommutative geometry and the Riemann zeta function takes this reformulation as its endpoint. It builds a geometric framework — the adele class space carrying an action of the idele class group — in which the explicit formula appears as a Lefschetz formula, the zeros of -functions appear spectrally, and the paper's concluding assertion (§3, p. 22) is that the validity of the global trace formula implies, and is in fact equivalent to, positivity of the Weil distribution, i.e. RH for all -functions with Grössencharakter.
This mission formalizes the arithmetic core of that endpoint in its simplest instance: the global field with trivial Grössencharakter, so that the -function is itself. Concretely it asks for (i) the Riemann–Weil explicit formula for in the shape of Connes' equation (11), and (ii) both directions of the equivalence between positivity of the resulting Weil distribution and RH.
A rough timeline of the statements involved: Riemann (1859) gave the first explicit formula; von Mangoldt (1895) proved it rigorously; Weil (Sur les "formules explicites" de la théorie des nombres premiers, 1952) extended it to all global fields and isolated the positivity criterion; Bombieri (Remarks on Weil's quadratic functional in the theory of prime numbers, 2000) studied the associated quadratic functional in detail; Connes (1996–1999) gave the trace-formula interpretation formalized in part here.
All objects live on the group , the module of the idele class group of , written additively through , .
A test function is a map that is and has compact support (IsTest).
Its transform is
which is Connes' in the coordinate , shifted by so that is the variable of (mellinHat). On the critical line, is the ordinary Fourier transform.
The involution is , i.e. (starInv), and convolution is (conv).
The Weil distribution of a test function collects the pole terms, the finite places and the archimedean place:
where is the von Mangoldt function and (weilDistribution, with the three pieces named mellinHat, primeSum, archTerm). The middle sum is Weil's contribution of the finite places , the last integral the contribution of the real place.
The spectral side is
the sum over the zeros of with , each counted with its multiplicity (zeroSum, IsCriticalZero, zeroMult).
This is the direction that yields RH, and it is the weakest form of the endpoint of the paper: it fixes no rate, no test-function normalization beyond , and no numerical constant.
with the sum over zeros asserted to be (unconditionally) summable.
Together with the goal this is the equivalence asserted on p. 22 of the paper, in the case , trivial Grössencharakter.
The -identity on the critical line, and the fact that is again a test function.
Weil's positivity criterion is one of the standard equivalent forms of RH, and the only one in which the arithmetic input (the primes, through ) and the archimedean input (the -factor) enter as separate, explicitly computable local terms. Formalizing it produces a machine-checked bridge between the zeros of and prime sums: the explicit formula milestone is the reusable object here, since essentially every analytic application of zeta zeros — zero-density estimates, prime-counting error terms, pair-correlation statistics — is an instance of it.
Status honesty: neither RH nor the positivity statement is known; the explicit formula and both implications relating positivity to RH are classical theorems, proved but not, as far as the catalog shows, formalized in Lean. Mathlib currently provides , its functional equation, the von Mangoldt function and , but no explicit formula of any kind. What this mission adds on top of the paper is therefore the formal proof of known results, not new mathematics.
The obvious route to the explicit formula — integrate over a vertical line, move the contour to the reflected line, collect residues — fails to be routine at exactly two points. First, moving the contour requires control of on horizontal segments between zeros, which is where the classical proof invests most of its work; Mathlib has bounds near but nothing of this shape inside the strip. Second, the sum over zeros must be shown to converge unconditionally, which needs a zero-counting bound of Riemann–von Mangoldt type () that is not in Mathlib either.
For the goal implication, the naive idea — pick a test function whose transform is supported near a hypothetical off-line zero — is unavailable: is entire whenever has compact support, so it cannot be localized. The classical argument instead exploits the symmetry of the zero set and makes the off-line quadruple contribute a negative amount in the limit along a family of test functions.
Conventions the Lean statements commit to. Test functions are -valued on , ContDiff ℝ (⊤ : ℕ∞) (so , not analytic) with HasCompactSupport; the multiplicative group is always written additively. The transform carries the -shift shown above, so the critical line is and are the two pole terms. Zeros are indexed by the subtype and weighted by (analyticOrderAt riemannZeta s).toNat; the trivial zeros are excluded. Integrals are Bochner integrals and sums are tsum, so both take the junk value when the integrand is not integrable or the family is not summable — for that reason the explicit formula is stated as a HasSum, which carries summability, rather than as an equation between tsums. The archimedean term is written with , logDeriv Complex.Gamma, rather than as a principal value, to avoid a second regularization convention.
The goal is not trivially satisfiable: its hypothesis quantifies over a nonempty class (smooth bump functions exist), and its conclusion is RH for , so no vacuous reading is available.
Infrastructure a complete development needs, all reusable beyond this mission: growth bounds for inside the critical strip, a Riemann–von Mangoldt zero-counting bound, Fourier analysis of functions (Paley–Wiener style decay of ), and the Hadamard product / functional equation package for the completed zeta function.
Out of scope, and deliberately so: Connes' operator-theoretic trace formula (equations (41) and (45) of the paper) and the spectral realization theorem of p. 17. Both are statements about traces of operators on Hilbert space, and Mathlib presently has no trace-class operator theory to state them faithfully. The mission therefore formalizes the arithmetic side of the paper's endpoint; contributions that build the missing operator theory, or that extend the statements from to Dirichlet -functions and Hecke -functions with Grössencharakter, are welcome.
Before any geometry of surfaces can be done, one has to say what a surface is, in a way that supports calculus: a subset of that is locally the smooth, non-degenerate image of an open piece of the plane. Every statement in the later theory — the first and second fundamental forms, the Gauss map, curvature, geodesics — is written in local coordinates, and is therefore meaningful only once one knows that the answer does not depend on the coordinates chosen. That independence is the content of the change-of-parameters theorem, which is what makes "differentiable function on a surface" and "geometric quantity of a surface" well-defined notions.
The reference for this mission is Manfredo P. do Carmo, Differential Geometry of Curves and Surfaces, 2nd edition (Dover, 2016), §2-2 "Regular Surfaces; Inverse Images of Regular Values" (pp. 54–71) and §2-3 "Change of Parameters; Differentiable Functions on Surfaces" (pp. 72–85): Definition 1 (p. 54), Propositions 1–4 of §2-2 (pp. 59, 61, 63, 65) and Proposition 1 of §2-3 (p. 74).
This is the fourth mission of a series formalizing do Carmo's book, sharing the namespace
DoCarmoDG with the others.
A subset is a regular surface when every has an open neighbourhood such that is the image of a map , defined on an open set , satisfying the three conditions of do Carmo's Definition 1:
Such an is a parametrization, or system of local coordinates, and is a coordinate neighbourhood.
Given a differentiable on an open set , a value is a regular value of when is surjective — equivalently, nonzero — at every with (do Carmo Definition 2, §2-2).
If and are two parametrizations of a regular surface with , then
is a diffeomorphism: is differentiable, bijective, and is differentiable.
The graph of a differentiable function of two variables is a regular surface (Proposition 1); the inverse image of a regular value is a regular surface (Proposition 2); a regular surface is locally the graph of a differentiable function of one of the three coordinate pairs (Proposition 3); and an injective map satisfying conditions 1 and 3 whose image lies in a regular surface automatically has a continuous inverse (Proposition 4).
Proposition 2 is the practical criterion: it is what shows in one line that spheres, ellipsoids, tori and the level sets of generic polynomials are regular surfaces, and it is applied throughout the book. Proposition 3 is the structural statement that a regular surface is locally a graph, which is the form in which most local computations are carried out; Proposition 4 removes the awkward homeomorphism clause from the verification of examples. The change-of-parameters theorem is what allows every subsequent definition — differentiable function on a surface, tangent plane, first fundamental form, curvature — to be given in coordinates and then shown to be independent of them, and it is also the reason a regular surface carries a smooth structure at all.
Mathlib has smooth manifolds, the implicit and inverse function theorems, and ContDiffOn, but it
does not contain do Carmo's concrete definition of a regular surface as a subset of
or these four propositions about it. Establishing them is what allows the rest of this series to
work with patches while knowing that the objects so defined are coordinate-independent.
Everything here rests on the inverse function theorem, but each proposition needs it in a slightly different form. Proposition 2 requires completing to a local diffeomorphism and reading off the level set — with the complication that which partial derivative is nonzero varies from point to point, so the coordinate that is solved for is not fixed in advance. Proposition 3 needs the same case distinction on which Jacobian minor of is nonzero, and this is exactly why the conclusion is a disjunction over the three coordinate pairs. Proposition 4 is where the homeomorphism condition is shown to be redundant, and the argument goes through the local factorization .
The change-of-parameters theorem is not a direct application of the inverse function theorem to : the map is defined only on a subset of the plane and is, a priori, merely continuous. One first extends to a local diffeomorphism of a neighbourhood in and then composes; the continuity of (condition 2 of Definition 1) is what makes the domain of open, and it cannot be dispensed with.
A surface is a set S : Set (EuclideanSpace ℝ (Fin 3)), and a parametrization is a map
x : ℝ × ℝ → EuclideanSpace ℝ (Fin 3) together with an open U : Set (ℝ × ℝ). Smoothness is
ContDiffOn ℝ (⊤ : ℕ∞), matching do Carmo's "differentiable" for ; regularity is
injectivity of the Fréchet derivative at each point of U, which is do Carmo's condition 3; and
the homeomorphism condition is stated as injectivity on U together with the existence of a
continuous left inverse on the image, which is the content of "the inverse is continuous". The
neighbourhood clause of Definition 1 is x '' U = V ∩ S for an open V containing the point.
Graphs are formalized as three separate sets, one for each of , and
, so that Proposition 3 can state its disjunction faithfully; in that statement the
neighbourhood is an open set W of and the claim is W ∩ S = W ∩ graph.
The goal states the diffeomorphism property of explicitly — two maps, mutually inverse on the
relevant domains, both ContDiffOn, together with the openness of those domains — rather than
through a bundled structure, so that no library convention is assumed. There is no trivializing
reading: the domains are those forced by the two parametrizations, and in the degenerate case
where the images do not overlap the statement reduces to a true but empty claim about the empty
set, while the substance is in the overlapping case.
Until 1827 the curvature of a surface in space was understood as a statement about how the surface sits inside : it was computed from the way the unit normal turns, that is, from the second fundamental form. Gauss's Disquisitiones generales circa superficies curvas showed that one particular combination of those extrinsic quantities — the product of the principal curvatures — can be recomputed from measurements made entirely inside the surface, using only lengths of curves drawn on it. This is the Theorema Egregium, and it is the reason the subject splits into extrinsic and intrinsic geometry; the latter is what becomes Riemannian geometry.
The reference for this mission is Manfredo P. do Carmo, Differential Geometry of Curves and Surfaces, 2nd edition (Dover, 2016), §4-3, "The Gauss Theorem and the Equations of Compatibility" (pp. 235–240). The theorem is stated on page 237 and derived from the Gauss formula, equation (5) of that section; the Mainardi–Codazzi equations (6) and (6a) on page 238 complete the list of compatibility equations.
This is the second mission of a series formalizing do Carmo's book; it shares the namespace
DoCarmoDG with the first, on the local theory of curves.
A regular parametrized patch is a map , defined and smooth on an open set with coordinates , whose partial derivatives satisfy at every point of ; the last condition says that is injective, so spans a -dimensional tangent plane at each point and
is a unit normal field along the patch.
The first fundamental form is the restriction of the ambient inner product to the tangent plane; in the parametrization it is recorded by the three functions
and . The second fundamental form is recorded by
and the Gaussian curvature is
The three second derivatives decompose in the basis ; the tangential coefficients are the Christoffel symbols of the patch, and the normal coefficients are , , , which is do Carmo's system (1) of §4-3:
Two patches over the same parameter domain are isometric when their first fundamental forms coincide, , , at every point: lengths of curves, angles and areas computed in the parameter domain then agree, and a local isometry between the two surfaces is obtained by matching parameters.
The Gaussian curvature of a regular patch is determined by its first fundamental form alone, although its definition uses the second fundamental form, i.e. the position of the surface in space.
The existence and uniqueness of the Christoffel symbols; the linear system (2) expressing them through and their first derivatives; the Gauss formula (5),
the Mainardi–Codazzi equations (6) and (6a); the closed formula for in an orthogonal parametrization (Exercise 1, p. 240); the invariance of under a change of parameters; and, as a corollary, that no neighbourhood of a point of the unit sphere is isometric to a piece of a plane (Exercise 4, p. 240).
The theorem is what makes intrinsic geometry possible: a quantity defined through the embedding turns out to be computable from the metric, so it survives every isometric deformation. Concrete consequences include the impossibility of a distortion-free map of the sphere — the reason every cartographic projection distorts lengths — and the equality of the Gaussian curvatures of the catenoid and the helicoid at corresponding points, which do Carmo notes immediately after the theorem. In the structure of the book, the Gauss formula is also the identity that makes the global Gauss–Bonnet theorem of §4-5 a statement about intrinsic data.
Mathlib has inner product spaces, iterated derivatives and the smooth manifold library, but it does not contain the first and second fundamental forms of a parametrized surface, the Christoffel symbols of a patch, the Gaussian curvature in this sense, or the compatibility equations. This mission produces that vocabulary together with machine-checked proofs of the classical identities. The mathematics is Gauss's, from 1827; what is open is the formalization.
The proof is a computation, but not a short one: one differentiates the system (1), uses , re-expands every second derivative through (1) again, and equates coefficients in the basis . Formally, the cost sits in three places: justifying the interchange of the mixed partial derivatives; establishing that the coefficient functions obtained pointwise from linear algebra are differentiable in the parameters; and carrying out the coefficient comparison in a basis that is not orthonormal, where one must use that rather than take inner products with an orthonormal frame.
The naive route to the Theorema Egregium — "solve the system (2) for the , then quote the Gauss formula" — is the right one, but the first step must actually be carried out: the system (2) determines the symbols only because each of its three blocks has determinant , and that is where the regularity hypothesis is used.
A patch is a curried map x : ℝ → ℝ → EuclideanSpace ℝ (Fin 3), so that the partial derivatives
and are ordinary one-variable derivatives, and the domain is an open set
U : Set (ℝ × ℝ); smoothness is ContDiffOn ℝ (⊤ : ℕ∞) of the uncurried map on U, matching
do Carmo's use of "differentiable" for . Regularity is stated as
on U, with the vector product defined componentwise. All quantities
(, , , , , , , ) are total functions of the parameters, taking junk
values off U; every statement restricts to points of U.
Christoffel symbols are not defined by a formula: a statement that mentions them quantifies over
functions assumed to satisfy do Carmo's decomposition (1) on U, and a separate
milestone asserts that such functions exist and are unique on U. The symmetry
is built into the notation, as in the book.
Isometry is formalized as equality of , , over a common parameter domain rather than as a map between surfaces; together with the milestone on invariance under change of parameters, this recovers do Carmo's statement that is invariant under local isometries. The formalization deliberately keeps the surface concrete (a patch, not an abstract manifold), which is what makes the compatibility equations expressible as identities between explicit derivatives.
This mission's definition file builds on the vector-product definition introduced in mission I of this series (Fundamental Theorem of the Local Theory of Curves), so mission I must be submitted first: its definitions have to be published before the definition file of this mission can compile.
There is no vacuous reading: the hypotheses are satisfiable — every regular patch, for instance a graph or a surface of revolution, satisfies them — and the conclusion compares two curvature functions pointwise.
The differential geometry of curves in is the entry point of every course and every textbook in the subject, and it is the first place where a geometric object is shown to be completely determined by a small list of numerical invariants. A space curve traced out by a particle moving at unit speed bends (curvature) and twists (torsion); the assertion that these two scalar functions determine the curve completely, up to a motion of space, is the prototype of every later "fundamental theorem" of the subject — for surfaces (Bonnet), for Riemannian metrics, and for submanifolds in general.
The reference for this mission is Manfredo P. do Carmo, Differential Geometry of Curves and Surfaces, 2nd edition (Dover, 2016), Chapter 1, Sections 1-4 and 1-5. The statement targeted here is the one printed on page 19 under the heading Fundamental Theorem of the Local Theory of Curves; the uniqueness half is proved on pages 20–22, and the existence half is deferred by do Carmo to the appendix of Chapter 4, where it is obtained from the existence and uniqueness theorem for linear systems of ordinary differential equations.
This is the first mission of a series formalizing do Carmo's book. The series shares one Lean
namespace, DoCarmoDG, so that later missions on regular surfaces, the Gauss map, and
Gauss–Bonnet build on the vocabulary fixed here.
Let be an open interval and let be a smooth map. The curve is parametrized by arc length if for every ; the parameter is then the arc length measured along the curve.
For such a curve one sets
The vector is the unit tangent and the scalar is the curvature at . Differentiating gives , so is orthogonal to . At a point where one defines the normal vector and the binormal vector
where is the vector product of (do Carmo §1-4). The triple is a positively oriented orthonormal basis, the Frenet trihedron. Since has constant length and is orthogonal to , the derivative is a multiple of , and the torsion is defined by
This is do Carmo's sign convention; many authors write for the same quantity, and the mission is committed to do Carmo's. With these conventions the Frenet formulas read
A rigid motion of is a map where is an orthogonal linear map with positive determinant and (do Carmo §1-5, Exercise 6).
Given smooth functions with :
The two halves are also stated separately as milestones, since they are proved by entirely different means: uniqueness by a Gronwall-free energy argument on the Frenet trihedron, existence by solving a linear ODE system.
The orthonormality of the Frenet trihedron, the Frenet formulas themselves, the characterization of straight lines by and of plane curves by , the closed formula , and the invariance of arc length, curvature and torsion under rigid motions.
The theorem is the model case of a classification result: a geometric object modulo a symmetry group is faithfully encoded by a complete set of local invariants. Downstream it is what licenses the standard practice of "prescribing curvature and torsion" — constructing curves with specified geometric behaviour, computing with the Frenet apparatus rather than with the curve itself, and recognizing that any identity among , and their derivatives is a genuine statement about the curve and not about its parametrization. In do Carmo's own development the local canonical form (§1-6) and the global results of §1-7 both rest on the Frenet apparatus fixed here.
Mathlib contains the analytic ingredients — the Picard–Lindelöf theorem, existence and uniqueness for linear ODE systems, orthonormal bases and the orthogonal group of a real inner product space — but it does not contain the Frenet trihedron of a space curve, the torsion of a space curve, or this theorem. What this mission produces is therefore a reusable formal vocabulary for the local theory of space curves, plus machine-checked proofs of the classical statements about it. The mathematics is completely classical and has been known since Frenet (1847) and Serret (1851); what is open here is the formalization, not the mathematics.
The uniqueness half is a short argument on paper — the function has vanishing derivative by the Frenet formulas — but formally it requires first establishing that the Frenet frame is differentiable and satisfies those formulas, which needs and the smoothness of away from its zeros, and then a connectedness argument on the interval.
The existence half cannot be done by exhibiting a formula: the curve is produced by solving the linear system for the frame , checking that the solution stays orthogonal (this is where the skew-symmetry of enters), and then integrating the first row. Recovering that the resulting curve has exactly the prescribed curvature and torsion, as computed by the definitions rather than as postulated by the ODE, is the step where most of the formal work sits.
The obvious shortcut — defining torsion by the closed formula — is not taken here: the definition is the book's, , and the closed formula is a milestone to be proved.
Curves are total functions ℝ → EuclideanSpace ℝ (Fin 3) that are assumed smooth only on the
open interval Set.Ioo a b; smoothness is ContDiffOn ℝ (⊤ : ℕ∞), matching do Carmo's use of
"differentiable" to mean . Because the interval is open, the ordinary deriv agrees
with the derivative along the interval at every interior point, and all derivatives in the
statements are plain iterated deriv. Curvature, normal, binormal and torsion are defined
exactly as above; at points where the normal vector takes the junk value , so every
statement that mentions , or carries the hypothesis explicitly.
The vector product is defined componentwise on EuclideanSpace ℝ (Fin 3), and a rigid motion is
a LinearIsometryEquiv of EuclideanSpace ℝ (Fin 3) with positive determinant followed by a
translation.
Degenerate intervals are not excluded: if the interval is empty and the statements hold vacuously, which is why the goal is not formulated as a statement about a single point but as a statement about all of — no hypothesis is vacuous for , and the existence clause is a genuine construction.
Contributions welcome: the Frenet apparatus and the ODE construction are the reusable parts, and both are prerequisites for the later missions of this series.