Markov type and superreflexivity

Source and verification

This article explains family 327 of the October 6, 2026 OpenAI collection. Research conclusions are attributed to its manuscripts. The review covers statements, introductions, and selected proof mechanisms; it is not an independent verification of the complete proofs.

Overview

Markov type bounds how fast a stationary reversible random walk can spread after mapping its states into a metric space. Family 327 claims that a real Banach space has a nontrivial such bound exactly when it is superreflexive. The proof constructs walks with linear displacement in every nonsuperreflexive space, contradicting the sublinear moment growth required by Markov type. (OpenAI, 2026)

🏷️ Preliminaries: reversible walks and equivalent norms

Let $P$ be a stochastic matrix on a finite state set $S$, and let $\pi$ be a probability distribution. Reversibility is the detailed-balance identity

$${\pi }_i{P}_{ij}={\pi }_j{P}_{ji}.$$

It implies stationarity when the initial state has law $\pi$.

A metric space has Markov type $p$ with constant $K$ if every such stationary walk, every map $f:S\to X$, and every integer $t\ge 1$ satisfy

$$\begin{array}{c}\mathbb{E}d(f(Z_t),f(Z_0){)}^p\\ \le K^{p}t\mathbb{E}d(f(Z_1),f(Z_0){)}^p.\end{array}$$

The same $K$ must work for every chain, map, and time. For $p=1$, the triangle inequality and stationarity already give the assertion with $K=1$. Thus the phrase nontrivial means $p>1$.

A real Banach space is superreflexive if it admits an equivalent uniformly convex norm. Uniform convexity says that for every $\epsilon >0$, two unit-ball vectors at distance at least $\epsilon$ have a midpoint whose norm is at most $1-\delta (\epsilon )$.

Equivalent norms need not have the same Markov-type constant, but preserve the existence of one. If

$$a\Vert x\Vert \le |x|\le b\Vert x\Vert ,$$

a Markov-type inequality in $|\cdot |$ transfers to $\Vert \cdot \Vert$ with constant at most $(b/a)K$. This follows by bounding the left side using $a^{-p}$ and the one-step cost using $b^p$.

Manuscript claim

A real Banach space has Markov type $p$ for some $p>1$ if and only if it is superreflexive. The new direction in the paper is from Markov type to superreflexivity. The converse uses classical power-type smooth renorming and its Markov-type estimate. (OpenAI, 2026)

The theorem is existential in both the exponent and the equivalent norm. It does not identify an optimal convexity modulus from a given Markov-type constant.

🏷️ Finite representability and an ordered norm

A space $Y$ is finitely representable in $X$ if every finite-dimensional part of $Y$ embeds linearly into $X$ with distortion arbitrarily close to one. The classical James–Enflo characterization says that $X$ is superreflexive exactly when every space finitely representable in it is reflexive.

Starting from failure of superreflexivity, the manuscript constructs a norm $N$ on finitely supported sequences with two useful features. Inserting zeros or moving coordinates to an increasing list of indices preserves the norm. Merging adjacent entries is contractive:

$$\begin{array}{c}N(\dots ,a_i+a_{i+1},\dots )\\ \le N(\dots ,a_i,a_{i+1},\dots ).\end{array}$$

Each finite coordinate span is finitely representable in $X$.

Repeated merging controls sums over consecutive blocks. If the prefix sums of a coefficient array repeatedly move between two well-separated levels, these block sums detect the oscillation. This supplies the geometric lower bound used by the walk.

🏷️ From edge labels to actual increments

The finite combinatorial construction assigns vector labels to oriented transitions. Reversing a transition negates its label. The manuscript’s target is a walk of length $n$ with unit-norm labels for which

$$\mathbb{P}\left(N\left(\sum _{j=1}^n{\xi }_j\right)\ge n/24\right)\ge 3/20.$$

It obtains the oscillating prefix sums by ordered motions on finite coordinate sets, with a Ramsey selection matching the marginal laws needed for reversibility.

There is a crucial obstruction to immediately invoking Markov type. Arbitrary edge labels need not be differences $f(v)-f(u)$; their sum around a cycle may be nonzero. The paper therefore enlarges the state space by finite integer coordinates recording traversed oriented edges. Away from its boundary, a vector-valued potential on this enlargement has precisely the prescribed increments.

Moves that would cross the boundary are rejected. This preserves reversibility. Taking the coordinate box sufficiently large makes the probability of hitting its boundary during the fixed time horizon small. The transfer lemma then carries a Markov-type bound to the edge-label construction. This step is needed even when the original chart states coincide across a marked transition.

🏷️ The contradiction in one estimate

Once the edge labels have been transferred to genuine increments and embedded into $X$ with arbitrarily small distortion, the claimed displacement estimate gives

$$\mathbb{E}N{\left(\sum _{j=1}^n{\xi }_j\right)}^p\ge \frac{3}{20}{\left(\frac{n}{24}\right)}^p.$$

Unit one-step labels would give the Markov-type upper bound $K^{p}n$. Therefore

$$K^p\ge \frac{3}{20\cdot 24^p}{n}^{p-1}.$$

For $p>1$, the right side diverges. Approximation errors from finite representability and the lattice boundary are sent to zero for each fixed $n$ before increasing $n$.

The exponent comparison is elementary. The difficult source inputs are the ordered-norm construction, the reversible walk with the stated displacement, and the transfer from labels to a genuine state map.

🏷️ Distinct probabilistic invariants

Rademacher type only tests independent signed sums and does not characterize superreflexivity. Markov convexity is also different: it compares walks sharing a past and then evolving independently. This family concerns ordinary stationary reversible Markov type.

Nor does the claim about nontrivial Markov type conflict with the next family’s cotype assertion for ${\ell }_1$. Metric Markov cotype is a different inequality with new points that may be chosen after the original data are given.

🏷️ Key steps and proof outlines

Ramsey selection creates an ordered norm with uniform finite control

Failure of superreflexivity supplies finite arrays detecting nonreflexive geometry. For each fixed arity, color increasing subtuples according to the norms of finitely many rational coefficient combinations, rounded to a fine grid. A finite Ramsey theorem yields a long subarray on which these colors agree. Repeat for increasing arities, always thinning the previous subarray so earlier homogeneity survives.

A diagonal subsequence over rational coefficient tuples gives a limiting norm $N_0$. A common ${\ell }_1$ Lipschitz bound extends the limit to real coefficients. The norm does not depend on the particular increasing positions of nonzero coordinates. A lower bound by the largest absolute prefix sum ensures positive definiteness, so the construction has not collapsed to a seminorm. Uniform convergence on each finite coefficient cube also proves finite representability in the original space. (OpenAI, 2026)

Subdivision makes adjacent merging contractive

Replace a coefficient $a_i$ by many equal pieces $a_i/J_i$. Averaging over increasing placements and using convexity shows that further subdivision cannot increase the norm. Pass to the subdivision limit to obtain the ordered norm $N$.

If two adjacent coefficients have the same sign and rational ratio, subdivide them into a common unit size. Their separate subdivisions and the subdivision of their sum give the same ordered array, so merging preserves the norm. Continuity extends this to all same-sign pairs.

For arbitrary signs, fix their sum $s$ and consider the convex function

$$t⟼N(\dots ,t,s-t,\dots ).$$

It is constant at the merged value on the interval between $0$ and $s$. Convexity forces it to be at least that value everywhere: a smaller exterior value would lower a point inside the constant interval by convex interpolation. Letting $s\to 0$ covers the zero-sum case. This proves the merging inequality for every adjacent pair.

Ordered motion converts repeated level crossings into large norm

The finite walk is built from charts encoding ordered motions of coordinates. Reduced words track which movements survive cancellation. The construction arranges that prefix sums of the accumulated coefficient array repeatedly pass between two separated levels. Contractive merging groups the coefficients between successive crossings; the lower prefix-sum geometry then forces the norm of the total label sum to be proportional to the number of useful crossings.

Ramsey selection and finite averaging make the chart transition marginals compatible with reversal. This is necessary for a stationary reversible walk, rather than a directed process chosen solely to create displacement. The source proves that with a fixed positive probability there are enough useful crossings to give the stated $n/24$ lower bound after $n$ steps. The constants are independent of the size of the coordinate state space.

A finite coordinate lift repairs nonzero cycle sums

Assign one integer coordinate to each pair of reversed oriented transitions. A forward traversal adds the corresponding basis vector, and the reverse traversal subtracts it. Enlarge the chain to $S\times \{-R,\dots ,R{\}}^r$, where $r$ is the number of such transition pairs. Moves leaving the box are rejected and become holding moves.

The product of the original stationary distribution and uniform box measure satisfies detailed balance: every allowed lifted transition has a reverse transition with the same weight, and holding moves are automatically reversible. Define a vector-valued potential by summing each integer coordinate times its assigned vector label. On allowed transitions, its difference is exactly the desired edge label, even when the original chart state does not change.

For fixed time horizon $n$, a starting point farther than $n$ from every coordinate boundary sees no rejection. The fraction of other starting points is at most a constant times $rn/R$. Sending $R\to \infty$ makes the boundary error vanish. Thus a Markov-type inequality for genuine state maps transfers to the abstract edge-label walk.

The exponent contradiction is taken after both approximations

Embed the finite coordinate span of $N$ into $X$ with distortion tending to one. For each fixed $n$, first remove this distortion error and the finite-box boundary error. Markov type then bounds the $p$th moment of the label sum by $K^{p}n$, while the successful crossing event gives a constant times $n^p$.

Only after these finite approximations are removed does $n$ tend to infinity. Since $p>1$, the two growth rates are incompatible. This order avoids assuming a single embedding or a single coordinate box works uniformly for every horizon. The converse direction uses the known smooth renorming theorem and transfer between equivalent norms, as distinguished in the statement.

See Also

• Type, cotype, and approximation introduces the independent-sign geometry.
• Metric Markov cotype of ℓ₁ explains the modification-of-values inequality and its extension consequence.
• Asymptotic midpoint geometry concerns finite-codimensional convexity rather than a uniformly convex equivalent norm.

🏷️ Manuscripts and verification scope

• Nontrivial Markov Type Forces Superreflexivity — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The documented equivalence is existential in the exponent and equivalent norm and includes the zero space. No Lean build, axiom audit, or manuscript-to-formal-statement equivalence check was performed for these notes.

OpenAI collection provides prerequisites and the reading sequence; coverage register records completed and pending families.

🐻  OpenAI 2026. Nontrivial Markov Type Forces Superreflexivity.