Nonexpansive fixed points in reflexive spaces

Source and verification

This article explains family 328 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

A strict contraction has a fixed point by geometric convergence. A nonexpansive map has no contraction factor to spend. Family 328 claims that reflexivity alone supplies enough compactness for a fixed point on every nonempty closed bounded convex domain, in the original norm. The proposed proof adds an adaptive tree argument to the classical minimal-invariant-set reduction. (OpenAI, 2026)

🏷️ Preliminaries: norm and weak compactness

A map $F:C\to C$ is nonexpansive if

$$\Vert F(x)-F(y)\Vert \le \Vert x-y\Vert .$$

A real Banach space $X$ is reflexive when the canonical map $X\to X^{\ast \ast }$ is onto. In a reflexive space, every closed bounded convex set is compact for the weak topology, the topology of convergence under every continuous linear functional.

Weak compactness is weaker than norm compactness. Also, a norm-continuous nonlinear map need not be weakly continuous. Thus one cannot simply apply a weak-topology version of a continuous-map fixed-point theorem.

Manuscript claim

Every nonexpansive selfmap of a nonempty norm-closed bounded convex subset of a real reflexive Banach space has a fixed point. The given norm is retained; uniform convexity, normal structure, separability, and weak continuity of the map are not assumed. (OpenAI, 2026)

Changing to a favorable equivalent norm would not prove this statement: it may change whether $F$ is nonexpansive.

🏷️ Approximate fixed points are easy to produce

Fix $a\in C$ and $0<r<1$. The map

$$F_r(x)=(1-r)a+rF(x)$$

takes $C$ into itself by convexity and has Lipschitz constant at most $r$. Banach’s contraction theorem supplies $x_r=F_r(x_r)$. Hence

$$\begin{array}{c}\Vert x_r-F(x_r)\Vert \\ =(1-r)\Vert a-F(x_r)\Vert \\ \le (1-r)\mathrm{diam}C.\end{array}$$

Letting $r\uparrow 1$ gives approximate fixed points.

The issue is passing to an actual one. A weakly convergent subsequence and norm-small defects do not automatically imply $F(x)=x$ without an additional closure argument.

🏷️ Minimal invariant sets and diametral behavior

Suppose a counterexample exists. Starting from one point and repeatedly taking the closed convex hull of the set and its image reduces to a separable invariant domain. Weak compactness then gives a minimal nonempty closed convex invariant subset $K$. A singleton would be a fixed point, so normalize its diameter to one.

The classical Goebel–Karlovitz lemma says that any approximate fixed point sequence $(u_j)$ in this minimal set satisfies

$$\Vert u_j-z\Vert ⟶1(z\in K).$$

Here is the mechanism behind that conclusion. Define

$$\begin{array}{rl}s(z)& =\underset{w\in K}{\sup }\Vert z-w\Vert ,\\ r(z)& =\underset{j}{\mathrm{limsup}}\Vert z-u_j\Vert .\end{array}$$

Both functions are convex and continuous. Minimality gives $K=\overline{\mathrm{conv}}F(K)$, so nonexpansiveness makes their nonempty sublevel sets invariant. Consequently both functions are constant; $s=1$.

If $r=R$, a weak cluster point $u$ of the approximate fixed points satisfies $\Vert z-u\Vert \le R$ for every $z$, by weak lower semicontinuity. Thus $1=s(u)\le R$, while bounded diameter gives $R\le 1$. Applying the same reasoning to every subsequence upgrades the limsup to convergence.

This means that an approximate fixed point eventually lies almost a full diameter from every prescribed point. It does not contradict weak compactness on its own.

🏷️ One functional detecting a compact family

The source strengthens the usable form of diametral behavior. Given a norm-compact set $E\subset K$ and $\epsilon >0$, every sufficiently accurate approximate fixed point $y$ admits one norm-one functional $f$ with

$$f(y-z)\ge 1-\epsilon (z\in E).$$

To see why, choose an $\epsilon /2$-net $a_1,\dots ,a_m$ and let $a$ be their average. Make $\Vert y-a\Vert \ge 1-\epsilon /(2m)$ and choose $f$ norming $y-a$.

Each $f(y-a_i)\le 1$, while their sum is at least $m-\epsilon /2$. Therefore every term is at least $1-\epsilon /2$. The finite net extends the estimate to all of $E$. This supplies a common detector, not a different functional for each point.

🏷️ The manuscript’s adaptive tree argument

The new part of the proposed proof must select compact families and approximate fixed points in a compatible order.

An anchor construction provides a point recoverable by finite convex combinations of increasingly accurate approximate-fixed-point images. The required accuracy may depend on all earlier choices. These images label a tree whose nodes also carry compact families of predicted root outputs, retaining parameters for convex coefficients that have not yet been chosen.

The proof replaces leaves from right to left. Although the maps at the nodes are nonlinear, consecutive actual root outputs have differences that telescope. Weighted bounds on these differences, together with the common detectors, produce vectors detected along finite branches.

The source uses a summable error budget without assuming a positive lower bound for branch weights. Reflexivity then supplies a vector in the weakly compact ball of $\overline{\mathrm{span}}(K-K)$ detected along an infinite branch. At the same time, the selected functionals tend pointwise to zero there, producing the contradiction.

The anchor, compact-output construction, and weighted switching estimate are the manuscript’s main proof obligations. The elementary approximate-fixed-point and diameter arguments above do not replace them.

🏷️ A consequence for commuting maps

Using Bruck’s theorem, the source deduces that any pairwise commuting family of nonexpansive selfmaps of $C$ has a common fixed point; its fixed set is a nonexpansive retract of $C$. The main theorem supplies the hereditary and conditional fixed-point hypotheses on closed convex invariant subsets. This is an application of an additional classical theorem, not a claim that arbitrary intersections of fixed sets are convex.

🏷️ Key steps and proof outlines

An approximate fixed point is an anchor, not a weak fixed point

For a chosen $a\in K$ and integer $n>1$, the map $u\mapsto a/n+(1-1/n)Tu$ is a strict contraction. Its fixed point $R_n(a)$ therefore exists, even though $T$ itself need not be weakly continuous. The equation gives a small displacement $\Vert T{R}_n(a)-R_n(a)\Vert \to 0$. This supplies approximate fixed points without passing $T$ through a weak limit.

Under the assumed absence of a fixed point, the source reduces to a minimal weakly compact invariant convex set of diameter one. The diametral property says that approximate fixed points eventually lie almost distance one from each fixed point of this set. To turn many separate distance statements into one linear detector, average a finite net of the relevant compact set and take a norming functional for the difference from that average. Each summand is at most one; if their average is nearly one, every summand must be nearly one. Approximation by the net gives a functional detecting the entire compact set. (OpenAI, 2026)

Predicting future outputs resolves the order of choices

The tree construction repeatedly selects points and norm-one functionals. A functional selected now must also control root outputs that depend on convex weights chosen later. The source handles this by putting all admissible earlier-candidate weights into a compact parameter space. The resulting family of predicted root outputs is compact and can therefore be included in the detector set before making the current selection.

Discarded candidates matter: their final parameter is zero, while retained candidates receive their eventual actual weights. Thus the prediction family contains the eventual switching configuration without requiring advance knowledge of which candidates survive. Error budgets are assigned to every candidate before its selection, with summable sizes such as $2^{-\ell -8}$ for candidate number $\ell$. This quantifier order prevents a later small branch weight from invalidating an earlier estimate.

Actual root changes telescope through nonlinear maps

Fix a finite depth $N$. Choose one approximate fixed point $b$ that simultaneously satisfies the finitely many detector conditions and makes all tree-map displacements small. Replace the leaf outputs by $b$ from right to left, recomputing the nonlinear tree after every replacement. Let $U_v,V_v$ be the root outputs before and after the whole descendant block of node $v$ changes, and let $W_v$ be the product of the convex weights along its branch.

Nonexpansiveness propagates errors. For untouched subtrees, the errors are bounded by a tail of the prescribed level tolerances. For fully switched subtrees, they are bounded by the sum of the displacements of the maps at $b$. Along the branch to $v$, off-branch errors have the exact total

$$\sum _i(W_{p_i}-W_{p_{i+1}})=1-W_v.$$

Combining these estimates with detection gives $f_v(U_v-V_v)\ge W_v-5e_v$. For each leaf $s$, define $d_s=(U_s-V_s)/W_s$. Sensitivity of the tree gives $\Vert d_s\Vert \le 1$. Crucially, these are actual consecutive root differences, so

$$\sum _{s\text{ below }v}{W}_s{d}_s=U_v-V_v$$

holds by telescoping, despite the nonlinear internal maps.

Weighted deficits produce one branch and one vector

Since $f_v(d_s)\le 1$, every deficit is nonnegative. Summing the previous inequality yields

$$\sum _{s\in L_N}{W}_s\sum _{\begin{array}{c}v\text{ on the path to }s\\ v\ne r\end{array}}(1-f_v(d_s))\le 5\sum _{v\ne r}{e}_v\le \frac{5}{128}.$$

A leaf with any detector at most $1/2$ contributes a path deficit at least $1/2$. Since leaf weights sum to one, some leaf has all its non-root detectors greater than $1/2$. No positive lower bound on individual leaf weights is used.

Finite branching then gives an infinite feasible branch. Weak compactness of the unit ball turns its nested finite collections of inequalities into one vector $d$ satisfying $f_{v_k}(d)\ge 1/2$ for all $k$. Meanwhile, detection of an increasing dense list of points makes $f_{v_k}$ tend to zero on every difference in a dense spanning set, hence pointwise on its closed span. This contradicts the common vector. These are the source’s closure steps; explaining their interfaces does not independently certify the manuscript’s full tree construction.

See Also

• Sphere-isometry extension uses a condensing return map and Darbo’s theorem, with different fixed-point hypotheses.
• Superreflexivity concerns a stronger renorming property; this family’s assertion retains every reflexive space’s original norm.

🏷️ Manuscripts and verification scope

• Fixed Points of Nonexpansive Maps in Reflexive Banach Spaces — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The selected theorem retains the original norm and includes zero and nonseparable spaces without weak continuity of the map. 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. Fixed Points of Nonexpansive Maps in Reflexive Banach Spaces.