Simple Lebesgue spectrum on a smooth three-torus

Source and verification

This article explains family 144 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.

🏷️ Koopman operators and spectral multiplicity

An invertible probability-preserving transformation $T$ defines a unitary operator $U_{T}g=g\circ T$ on complex $L^2$. Constants are invariant, so the spectral question is posed on

$$L_0^2=\left\{g\in L^2:\int gd\mu =0\right\}.$$

Let $m$ be normalized arc measure on the unit circle. Simple Lebesgue spectrum means that $U_T{|}_{L_0^2}$ is unitarily equivalent to

$$\begin{array}{c}(M_{w}h)(w)=wh(w)\\ \text{on }{L}^2(S^1,m).\end{array}$$

“Lebesgue” specifies the spectral measure type; “simple” specifies multiplicity one. A Lebesgue component inside a larger space would be a weaker assertion.

The functions $w^n$, $n\in \mathbb{Z}$, form an orthonormal basis of the circle space. Consequently the model is equivalent to one vector $f$ whose bilateral orbit $\{U_T^{n}f:n\in \mathbb{Z}\}$ is an orthonormal basis of the entire mean-zero space.

🏷️ The claimed smooth realization

Statement in the manuscript

There is a $C^{\infty }$ diffeomorphism $T$ of ${\mathbb{T}}^3$ preserving standard normalized volume, and a real $f\in L_0^2$, such that

$$\{f\circ T^n:n\in \mathbb{Z}\}$$

is an orthonormal basis of $L_0^2({\mathbb{T}}^3)$. (OpenAI, 2026)

The invariant measure is smooth standard volume. The generator $f$ is only asserted to be square integrable, not smooth. Both its orbit’s orthogonality and its completeness are necessary.

This is a discrete-time probability-preserving statement. A simple Lebesgue spectrum for a flow does not immediately give multiplicity one for a time map: identifying real frequencies modulo the reciprocal time step folds infinitely many frequency intervals onto the same circle.

🏷️ Mixing follows from the spectral model

Let $V$ be the unitary identification with the circle model. For centered $a,b\in L^2$,

$$\begin{array}{c}⟨U_T^{n}a,b⟩\\ ={\int }_{S^1}{w}^n(Va)(w)\overline{(Vb)(w)}dm(w).\end{array}$$

Cauchy–Schwarz puts the product in $L^1(m)$. The Riemann–Lebesgue lemma makes this integral tend to zero as $|n|\to \infty$. Taking centered indicators gives

$$\mu (A\cap T^{-n}B)⟶\mu (A)\mu (B).$$

Thus ordinary mixing is an elementary consequence of the claimed spectral model. Mixing of every order uses the separate manuscript on Rokhlin’s problem, rather than following directly from this two-function calculation.

The source also invokes a finite-spectral-multiplicity entropy theorem to deduce zero metric entropy. Pesin’s formula then gives

$$0=\int \sum _i\max \{{\lambda }_i(x),0\}d\mu (x).$$

Every Lyapunov exponent is nonpositive almost everywhere. Volume preservation makes their sum zero, so all three vanish. These deductions use the source’s stated external entropy results; zero exponents alone would not establish the spectral construction. (OpenAI, 2026)

🏷️ Integrable twists as approximating models

The approximating diffeomorphisms are twists in volume-preserving coordinates:

$$S(\theta ,u,a)=(\theta +\varphi (u),u,a).$$

On a large arc where $\varphi$ is monotone, angular Fourier modes have explicitly computable absolutely continuous spectral measures. Finite collections of these modes are assigned labels recording the coefficients of an observable that the construction wants to approximate.

A slow coordinate passage makes the old $u$ variable into the new angular variable. Stationary phase describes the transformed frequency indices and densities. The spare coordinate $a$ supplies a small adjustable speed variation. Adding a large integer to a lift of $\varphi$ leaves the old torus map unchanged while separating angular indices in the next passage.

This separation allows later perturbations to distinguish labels even when their temporal frequency distributions overlap. The construction does not assume a uniformly bounded number of modes or uniformly bounded derivatives of all coordinate changes. Instead, each large parameter is selected after the current finite data and desired errors are fixed. (OpenAI, 2026)

🏷️ Predicting labels while preserving spectral mass

Suppose label $r$ carries density $p_r(z)$ and desired coefficient $b_r$. Put

$$\begin{array}{c}p(z)=\sum _r{p}_r(z),\\ \beta (z)=\frac{{\sum }_r{b}_r{p}_r(z)}{p(z)}\end{array}$$

where $p(z)>0$. The quantity $\beta$ is the conditional mean of the coefficient given temporal frequency.

Small sinusoidal biases change the label densities. Centering a block at a smooth approximation to $\beta$ makes the loss in $\int \sqrt{p}$ quadratic in the block’s size, while the same block improves prediction of the label. Repeating many small blocks therefore aims to make prediction accurate at a summably small spectral cost.

Why monitor $\int \sqrt{p}$? If a vector has spectral density $p$ and norm near one, replacing its density by one through spectral normalization changes its physical $L^2$ norm by a quantity governed by

$$\begin{array}{rl}& \int (\sqrt{p}-1{)}^{2}dm\\ & =\int pdm+1-2\int \sqrt{p}dm.\end{array}$$

Zeros of the density require a separate term. For the current vector $G$, let $E=\{p=0\}$, and let $v$ be an auxiliary vector with spectral density one supplied by the source. Spectral functional calculus defines

$$\begin{array}{rl}{f}^{\prime }& =\left(\frac{1_{E^c}}{\sqrt{p}}\right)(U)G\\ & +1_E(U)v.\end{array}$$

The two terms have disjoint spectral supports and densities $1_{E^c}$ and $1_E$. The second fills the missing frequencies; division by $\sqrt{p}$ alone would not do so. The displayed distance formula includes this zero-set contribution. Keeping $\int \sqrt{p}$ near one makes the correction small, and the resulting density exactly one gives an orthonormal orbit.

🏷️ Completeness in the limit

Orthogonality alone can leave a large unused invariant subspace. The manuscript enumerates a dense family of observables and uses the label-prediction construction to approximate each one by finite linear combinations of orbit vectors.

A diagonal construction balances three requirements: convergence of the diffeomorphisms in every smooth norm, convergence of the generating vectors in $L^2$, and preservation of the previous approximation tasks. The limiting vector has an orthonormal orbit whose closed span contains the dense family.

These finite prediction estimates and the simultaneous smooth/spectral limit are the substantive research inputs. The spectral consequences above explain their significance without substituting for their verification.

🏷️ Key steps and proof outlines

Prediction and spectral preservation have different scales

The key estimate in “Finite blocks and prediction” makes the simultaneous requirements quantitative. Let packet label $r$ carry density $p_r$ and target coefficient $b_r$, scaled so $|b_r|\le 1$. Put $p={\sum }_r{p}_r$ and $\beta ={\sum }_r{b}_r{p}_r/p$ on $\{p>0\}$. Retain the normalization $\int p\le 2$ at every block. At the start of each block, choose a smooth $q$ approximating this conditional coefficient mean and hold it fixed throughout that block.

For block strength parameter $0<\delta \le 1/100$, the source constructs a multiplier $D$ whose model conditional mean is $b_r$ at every frequency and whose conditional mean squared error is at most $2/\delta$. If $A$ is the previous estimator sum after $k$ blocks, the cross term against $A-k{b}_r$ vanishes pointwise in frequency. After integration, the next block increases the unnormalized prediction error by at most $4/\delta$ plus a chosen approximation error. Meanwhile the centered density perturbation loses only $O({\delta }^2)$ in $\int \sqrt{p}$.

The centering at the current $\beta$ is crucial: it removes the first-order spectral loss. The proof uses the elementary inequality

$$\sqrt{x}\ge 1+\frac{x-1}{2}-(x-1{)}^2(x\ge 0)$$

to reduce the square-root estimate to finite moment tests. It does not approximate the nonlinear square-root functional by an unjustified first-order expansion near possible zeros of the density. (OpenAI, 2026)

Finite phase tests realize the model deterministically

The probabilistic-looking model uses several circle phases. The actual construction substitutes phases $e^{2\pi i{L}_{j}z}$ with successively large integer frequencies $L_j$. Only finitely many Laurent-polynomial expressions in those phases must be tested in one block. Expanding such an expression gives finitely many Fourier integrals with integrable coefficient functions. The Riemann–Lebesgue lemma makes each nonconstant-frequency integral small when the next $L_j$ is chosen large enough; constant terms reproduce the independent-phase average.

This realizes both prediction and spectral tests simultaneously. The list of coefficient functions, the previous estimator’s norm, and the needed error tolerances are fixed before those frequencies are chosen. Large frequencies may worsen later derivative bounds, but the next stage is allowed to choose its parameters after seeing the current finite data. No uniform frequency count or uniform derivative bound across all stages is asserted.

Averaging many weak blocks gives both desired limits

Let $A_K$ be the sum of the $K$ successive multipliers. With a per-block auxiliary error at most ${\delta }^2$, the source obtains

$$\begin{array}{rl}\int \sqrt{p^{(K)}}& \ge \int \sqrt{p^{(0)}}-400K{\delta }^2,\\ \sum _r\int p_r^{(K)}|A_K/K-b_r{|}^2& \le \frac{5}{K\delta }.\end{array}$$

Given prediction tolerance $\epsilon$ and spectral-loss budget $\tau$, first choose $\delta$ small and then $K=\lceil 10/({\epsilon }^2\delta )\rceil$. The prediction denominator $K\delta$ is large, while $K{\delta }^2$ is of order $\delta /{\epsilon }^2$ and can be made smaller than $\tau$. This is the precise balance that permits useful information acquisition at arbitrarily small spectral cost.

Disjoint angular-index supports make the different label cyclic subspaces orthogonal. Consequently the prediction integral equals the physical squared norm of the error between ${\sum }_r{b}_r{F}_r$ and the spectral multiplier $(A_K/K)(U){\sum }_r{F}_r$. Thus the estimate produces an approximation of an actual observable by the cyclic space of the packet sum, not merely an abstract classifier for labels.

Spectral normalization and dense approximation close the construction

At each stage, correcting density $p$ to one uses division by $\sqrt{p}$ on $\{p>0\}$ and a separate auxiliary vector on $\{p=0\}$. Orthogonality of those spectral supports makes the corrected density exactly one. Its orbit is therefore orthonormal because its correlation at time $n$ is $\int w^{n}dm$, zero for nonzero integers.

Completeness requires a second diagonal requirement. Enumerate a dense set $g_1,g_2,\dots$ in the mean-zero space, and at increasingly accurate stages approximate each by a finite orbit polynomial of the current generator. For any fixed finite polynomial, convergence of the maps and $L^2$ convergence of the vectors preserve the approximation after sufficiently small later changes. Summable error budgets ensure that every $g_j$ belongs to the limiting cyclic space. Orthonormality plus this dense containment gives the claimed basis. This last passage uses the source’s simultaneous smooth-map and physical-vector convergence estimates; orthogonality alone would leave multiplicity uncontrolled.

See Also

Rokhlin’s multiple-mixing problem supplies the additional all-orders consequence. Smooth positive-volume realization asks which abstract systems admit any smooth volume model, rather than prescribing one spectral type.

🏷️ Manuscripts and verification scope

• A smooth three-torus diffeomorphism with simple Lebesgue spectrum — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The scope describes the full mean-zero spectral model, its real cyclic generator, and ergodicity. The entropy, Lyapunov, and higher-order mixing consequences use the additional results identified in the manuscript. 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. A smooth three-torus diffeomorphism with simple Lebesgue spectrum.