Cotype, duality, and the approximation property

Source and verification

This article explains family 326 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

Cotype prevents vectors from being too large collectively while their random signed sum remains small. Having finite cotype for both a space and its dual still does not, by itself, force nontrivial type. Family 326 claims that the ordinary approximation property supplies the missing hypothesis, without requiring uniformly bounded approximating operators. (OpenAI, 2026)

🏷️ Preliminaries: random signs and approximation

Let ${\epsilon }_i$ be independent signs taking values $\pm 1$ with probability $1/2$. A real Banach space $Y$ has cotype $q$, $2\le q<\infty$, if

$$\begin{array}{c}{\left(\sum _i\Vert y_i{\Vert }^q\right)}^{1/q}\\ \le C_q(Y){\left(\mathbb{E}{\Vert \sum _i{\epsilon }_i{y}_i\Vert }^2\right)}^{1/2}\end{array}$$

for every finite vector family. Type $p>1$ is the inequality in the other direction, with the right-hand coefficient norm ${\left({\sum }_i\Vert y_i{\Vert }^p\right)}^{1/p}$.

For Hilbert space, expanding the square makes cross terms vanish:

$$\mathbb{E}{\Vert \sum _i{\epsilon }_i{y}_i\Vert }^2=\sum _i\Vert y_i{\Vert }^2.$$

Thus Hilbert spaces have both type two and cotype two with constant one. In ${\ell }_1$, coordinate vectors instead have signed-sum norm $n$, so type $p>1$ would require $n\le D{n}^{1/p}$ for every $n$, which is impossible.

The approximation property (AP) says that the identity can be approximated uniformly on each compact subset by bounded finite-rank linear maps. The bounded approximation property (BAP) additionally asks for one norm bound on all required maps. AP makes no such promise.

🏷️ The Rademacher projection

For an $X$-valued function on the finite sign cube, define

$$({\mathrm{Rad}}_{s}f)(\epsilon )=\sum _{i=1}^s{\epsilon }_i{\mathbb{E}}_{\eta }[{\eta }_{i}f(\eta )].$$

This keeps only first-order sign coefficients. Put

$$K(X)=\underset{s\ge 1}{\sup }\Vert {\mathrm{Rad}}_s{\Vert }_{L_2(X)\to L_2(X)}.$$

The measures are uniform probability measures. The space is $K$-convex if $K(X)<\infty$. Classical results identify this with nontrivial Rademacher type and with failure to contain the finite-dimensional ${\ell }_1^s$ spaces with uniformly bounded distortion.

Manuscript claim

If a nonzero real Banach space $X$ has AP, then it is $K$-convex exactly when $X$ and $X^{\ast }$ both have finite Rademacher cotype. The two finite cotype exponents may differ. The forward implication needs no AP assumption. (OpenAI, 2026)

The source explicitly retains AP in the reverse implication: without approximation, there are spaces whose primal and dual have cotype two but which contain ${\ell }_1$. It also gives an existential uniform bound for $K(X)$ depending only on the two exponents and cotype constants. It does not give a numerical formula for that bound.

🏷️ Uniform Walsh decay

A Walsh transform is the discrete Fourier transform on binary sign coordinates. Its entries are $\pm 2^{-n}$, with signs determined by parity. Applying it to vector families tests how the geometry of the target interacts with cancellation.

The first part of the manuscript proves uniform decay of suitable normalized Walsh operator norms for finite-rank maps $T:G\to F$, assuming cotype estimates on $G^{\ast }$ and $F$. A factorization of one fixed finite-rank map through Hilbert space supplies a starting decay estimate, but its constant depends on the map. The argument must remove this dependence.

Its mechanism is a self-improvement contradiction. Cotype supplies bounded sign-indexed moment lifts: a tuple $(x_i)$ is represented by vectors $P_{\epsilon }$ with prescribed first moments

$${\mathbb{E}}_{\epsilon }[{\epsilon }_i{P}_{\epsilon }]=a{x}_i$$

for a controlled positive $a$. If Walsh decay is too slow, the source converts these lifts into scalar functions with specified row and column moments. Their Fourier expansion must contain two independent Walsh factors. Iterating the resulting lower bound would contradict the Hilbert-factorization estimate for the original fixed operator.

This is a quantitative operator lemma. Merely knowing that each finite-rank map factors through some Hilbert space does not yield its uniform version.

🏷️ Why ordinary AP is enough

A direct approximation of the identity by maps $S_j$ would introduce $\Vert S_j\Vert$, which AP does not control. The second part instead constructs, for each finite-dimensional $E\subset X$, a compact operator $J:G\to X$ and a contraction $\iota :E\to G$ such that

$$J\iota x=x(x\in E).$$

The auxiliary domain has controlled dual moment estimates.

Compactness now changes the approximation problem. The closure of $J(B_G)$ is compact, so AP gives finite-rank $S_j$ with

$$\begin{array}{c}\Vert S_{j}J-J\Vert \\ =\underset{\Vert g\Vert \le 1}{\sup }\Vert S_{j}Jg-Jg\Vert \\ ⟶0.\end{array}$$

Although $\Vert S_j\Vert$ may diverge, $\Vert S_{j}J\Vert$ remains bounded because the composites converge in operator norm. Uniform Walsh decay passes first to $J$, then through $\iota$ to the finite vector configurations in $E$.

Every finite configuration lies in such an $E$. Uniform copies of ${\ell }_1^s$ would violate the resulting decay; the classical characterization then yields $K$-convexity. The special compact-domain construction is therefore the bridge that genuinely uses AP rather than silently assuming BAP.

🏷️ An entropy consequence

For symmetric convex bodies $K,L$, let $N(K,L)$ be the least number of translates of $L$ covering $K$. Under fixed primal and dual cotype bounds for either body’s normed space, the manuscript derives dimension-free comparison with the covering numbers of the polar pair. Finite-dimensional spaces automatically have AP, while the uniform cotype bounds supply uniform $K$-convexity.

This is a restricted geometric class. It does not assert universal entropy duality for arbitrary symmetric bodies.

🏷️ Key steps and proof outlines

Cotype produces moment lifts by finite-dimensional duality

For a finite tuple $(x_i)$, a moment lift is a sign-indexed function $H(\epsilon )$ satisfying $\mathbb{E}{\epsilon }_{i}H(\epsilon )=x_i$. The cotype hypothesis on a dual space bounds the adjoint of this moment map. Finite-dimensional separation then gives a lift with controlled $L_2$ norm. This is stronger than choosing unrelated preimages for each coordinate: all prescribed moments must be realized by one sign-cube function.

The source uses such lifts on both sides of a finite-rank operator. Their scalar pairings produce functions with controlled row and column moments. Walsh expansion then turns the two moment constraints into information about the Fourier support. See “Walsh transforms and moment lifts.” (OpenAI, 2026)

A rank-two Fourier coefficient prevents one-sided collapse

The scalar coefficient lemma shows that the relevant Fourier support cannot consist only of rank-at-most-one indices. Its proof assumes that collapse and encodes the rectangular support patterns in a damped polynomial. The prescribed mixed derivative is nonzero, whereas splitting the polynomial into pieces of low degree in one variable forces a contradictory derivative bound.

Compactness upgrades nonvanishing to a uniform positive lower bound on some rank-two coefficient over the admissible scalar data. This supplies two independent Walsh directions. One direction alone would not generate the dimensional growth needed by the iteration. The lemma therefore identifies the specific new cancellation supplied by using cotype of both the domain dual and the range.

Self-improvement removes dependence on a fixed operator

For one finite-rank map $T$, factorization through a finite-dimensional Hilbert space gives decay of its normalized Walsh norms, but with a constant depending on $T$. The main theorem needs a rate depending only on cotype data and the operator norm.

Assume no such uniform decay exists. Moment lifts and the rank-two coefficient lemma turn slow decay at one dimension into a lower bound at a larger dimension. Iterate this growth while keeping the original operator fixed. Eventually the lower bound contradicts that operator’s Hilbert-factorization decay. The contradiction is not a uniform estimate on all Hilbert factorizations; it is an iteration that defeats the one fixed factorization after enough dimension growth.

The quantitative constants in the moment lifts and scalar lemma remain fixed during this iteration. Allowing them to deteriorate with the current Walsh dimension would remove the contradiction.

Compact factorization is where ordinary AP enters

Fix a finite-dimensional $E\subset X$. The source builds nested finite-dimensional enlargements $E_0\subset E_1\subset \cdots$ so tuples in $E_l$ have controlled moment lifts in $E_{l+1}$. Compactness of finite-dimensional unit balls allows a finite net of lift requirements at each stage; approximation then extends the selected requirements to all tuples with a controlled constant.

Weights tending to zero combine these levels into an auxiliary domain $G$ and a compact operator $J:G\to X$. A contraction $\iota :E\to G$ satisfies $J\iota =I_E$. The weights provide compactness, while the successive lift properties give the dual moment estimates required by the finite-rank Walsh theorem.

AP approximates the identity uniformly on the compact closure of $J(B_G)$. Hence finite-rank maps $S_l$ satisfy $\Vert S_{l}J-J\Vert \to 0$. Although $\Vert S_l\Vert$ need not be bounded, the composites obey $\Vert S_{l}J\Vert \le \Vert J\Vert +o(1)$. Apply the uniform finite-rank theorem to these composites and pass in operator norm to $J$. Restriction through $\iota$ transfers the estimate to $E$.

Uniform finite tests exclude the l1 obstruction

Every finite vector configuration in $X$ lies in some $E$. The resulting Walsh decay is therefore uniform over all such configurations. Uniformly distorted copies of ${\ell }_1^s$ would violate that decay as $s$ increases, because independent sign cancellation does not reduce their coordinate-sum norm in the required way. The classical characterization then yields $K$-convexity.

For the stated uniform bound on $K(X)$, the construction depends only on the two cotype exponents and constants, not on the particular finite-dimensional $E$. The entropy-duality consequence applies this uniform conclusion in finite dimensions, where AP is automatic. It remains restricted to the bodies with the specified primal and dual cotype control.

See Also

• AP without BAP explains why those approximation hypotheses cannot be interchanged.
• Metric-entropy duality treats the claimed failure of universal constants outside the controlled geometric regime.
• Markov type tests more than independent signed sums and detects a stronger renorming property.

🏷️ Manuscripts and verification scope

• The cotype–cotype conjecture under the approximation property — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The selected equivalence retains the ordinary approximation property. No bounded approximation hypothesis is stated. 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. The cotype–cotype conjecture under the approximation property.