Dynamical systems

On the ergodicity of partially hyperbolic systems

Keith Burns and Amie Wilkinson

50 pages39 claims indexed34 referencesManuscript 2010

gpt-6.2-sol high: bw-geometry

Reviewer record, not the publication verdict. The response includes unaccepted proposals and source limits. Missing antecedent proofs are not automatically defects in the paper. The initial request for Lemmas 8.4–8.5 was our inventory error; the journal ends at Lemma 8.3. The curated report records final dispositions.

Completed visible assessment, preserved as a readable check record. It is a model assessment, not a formal certificate. This document contains no private reasoning stream.

The introduction's exact hypotheses and the one-dimensional corollaries withstand the checks performed. The main local findings are misindexed plaque memberships, an unstated metric normalization for weak-plaque chord estimates, and an orbit-indexed extension needed when the distortion bound is applied to fake Jacobians. None disproves a principal theorem. The supplied PSW correction addresses why a merely C¹ weak leaf does not preclude C¹ strong subfoliation in the C² argument; the uniform graph-transform application remains partly conditional. The low-regularity holonomy and the quantitative genuine-foliation and Pinsker inputs are not certified by this packet. Historical cutoff: 2005-10-11, supported by the received line on BW journal p. 489 and the recorded policy. The 2010 journal is the primary text under review, not a source of permission to use later results. The 2006 note citation does not establish a pre-cutoff version, and the 2008 references were not used to close any obligation. No external source or later proof was consulted. Later knowledge cannot be removed from model weights; no such knowledge was used as evidence of historical availability. The supplied images used here were readable; garbled formulas in the author extraction were not used to settle journal notation.

G1

Checked statement: The exact hypotheses and conclusions of Theorems 0.1 and 0.3 and Corollaries 0.2 and 0.4, and the separate proof obligations for ergodicity and the Kolmogorov property.

Returned assessment: no_issue_found

Theorem 0.1 uses C², partial hyperbolicity, both strict inequalities in (5), preservation of an invariant probability equivalent to Riemannian volume, and essential accessibility. Equivalence of measures, not invariance of the reference Riemannian volume or smoothness of the invariant density, is the stated hypothesis. Essential accessibility tests measurable unions of entire accessibility classes. It does not by definition apply to a set having two possibly different saturated representatives modulo null sets. Theorem 0.3 replaces C² and (5) by C¹⁺ᵟ, the auxiliary derivative bounds in (6), and a common θ satisfying 0<θ<δ and both pairs of inequalities (7)–(8). The inequalities are for unit vectors, in the convention of pp. 452–453; μ and μ̂ in (6) are rate functions, not the invariant probability. There is no dynamical coherence assumption in either branch. Under the usual Hölder convention 0<δ≤1, (7) implies (5), but (5) is not a substitute for (7)–(8) in the low-regularity holonomy argument. For the one-dimensional corollaries put a(p)=||Df|Ec(p)||, which is also the center conorm. First fix partial-hyperbolicity rates and, for Corollary 0.4, positive lower rates as in (6). With those fixed, (8) holds for sufficiently small positive θ<δ by compactness: the left-hand ratios are uniformly below 1 and the right-hand powers tend uniformly to 1. Then improve γ and γ̂ towards a and 1/a, respectively, keeping each strictly on its original side of the derivative bound. This preserves or improves (8) and gives γγ̂ uniformly as close to 1 as needed for (5) or (7). No assertion that Ec is smooth is needed for this argument.

Premises used

  • Section 5 proof interface, from the author text only: the geometric theorem is to show that the ambient Lebesgue density set of a bi-essentially saturated measurable set is saturated by both genuine strong foliations, with no invariant-probability hypothesis. The density set is Borel (one can test a countable geometric sequence of radii), is unchanged by a null modification, and is a representative modulo null sets by the Lebesgue density theorem. Essential accessibility can then be applied to that representative.
  • Ergodicity interface: continuous observables have forward limsup averages constant on genuine stable leaves and backward limsup averages constant on genuine unstable leaves, by asymptoticity and uniform continuity. For an invertible invariant probability, the pointwise ergodic theorem makes the two limits equal almost everywhere, both being conditional expectations on the invariant sigma-algebra. Thus rational sublevel sets have the two saturated representatives; the Section 5 conclusion makes each sublevel set null or conull. Rational levels give almost-sure constancy, and continuous functions are dense in L¹ of the Borel probability on the compact manifold, so the invariant sigma-algebra is trivial.
  • Kolmogorov interface: use the trivial-Pinsker criterion. This requires the separate input that each Pinsker event has a measurable genuine stable-saturated representative and a measurable genuine unstable-saturated representative modulo the invariant probability. The Hopf argument for continuous observables does not supply that input. Once supplied, measure equivalence and the same density-set theorem make each Pinsker event null or conull.

Missing inputs

  • Published Section 5 images are not supplied. The division of proof obligations above is checked against the October 2006 author context, not claimed as a transcription or complete audit of the journal proof.
  • BP74, Proposition 5.1: available_from=1974; date_evidence=BW journal p. 488 bibliographic entry and the author text's proposition citation; date_basis=publication year; version=1974 original and cited English translation, neither supplied. The required statement is bi-essential saturation of every Pinsker event for the invertible partially hyperbolic map and invariant probability equivalent to volume. Check the exact original regularity, rate and measure hypotheses and any extension needed to the pointwise-rate and C¹⁺ᵟ settings. This application remains conditional.
  • Classical pointwise ergodic theorem: available_from=1931, bibliographic identification only; date_evidence=Birkhoff's 1931 publication, not supplied or independently inspected; date_basis=publication year; version=the conditional-expectation formulation for an invertible probability-preserving transformation and L¹ observables. Lebesgue density theorem: available_from=not documented in the packet; date_evidence=the author text invokes it by name; date_basis=undocumented in the packet; version=the classical local Euclidean density theorem for measurable sets, applied in smooth Riemannian charts. These standard measure-theoretic inputs are stated, not source-audited. No additional equivalence with a partition-based definition of the Kolmogorov property has been audited.

Scope changes

  • The verdict concerns the introduction's statements, the dimension-one implications, and the logic of the proof interfaces. It is not an unconditional validation of either principal theorem or of the cited Pinsker result.

G2

Checked statement: Proposition 3.1(i)–(vii): uniform construction in tangent fibers, localization and growth estimates, coherence and the genuine strong plaques through the basepoint, and the two regularity branches.

Returned assessment: repairable_gap

For the extension on pp. 465–466, write the local map in exponential coordinates as L_p v+e_p(v), with e_p(0)=0 and De_p(0)=0. A cutoff χ(v/r) multiplying e_p gives the asserted C¹ convergence uniformly: the cutoff derivative is O(1/r) and sup_{||v||≤2r}||e_p(v)||=o(r) uniformly. The derivative is a uniformly small perturbation of L_p; solving the inverse equation by contraction after applying L_p⁻¹ gives global fiber diffeomorphisms. For each fixed small r the relevant higher fiber norms are uniform; they need not stay bounded as r tends to zero. The base in the construction is discrete. The estimates to be proved concern the fibers and their orbit sequences, not differentiability across fibers in the original topology of M. Existence of the global graph foliations with uniform cones remains conditional on the uniform graph-transform input. Given that input, (i) follows from exponential charts and continuity of the distributions, (ii) from agreement with f on the inner ball and room for the forward and backward images after shrinking r₁, and (iii) from one-step estimates at the orbit index in question. For strong plaques, apply the contraction estimate to every point of the connected plaque component; the image stays in the target component, which closes the induction. For weak plaques the hypotheses must be rechecked at each step. Given the stable and unstable subfoliation of their weak foliations, intersecting the transverse weak leaves gives (iv) for the center. In (v) the origin plaque is singled out by the strong orbit-contraction characterization and agreement with f near the zero orbit; there is no corresponding identification of off-basepoint fake plaques with genuine plaques. For (vi), the forward splittings s|cu and cs|u and the inverse splittings give the candidate Hölder estimates: a uniform domination ratio below 1 remains below 1 after multiplying by the inverse lower-rate factor raised to a sufficiently small positive power. Intersecting the uniformly transverse weak distributions preserves Hölder regularity. The Hölder foliation conclusion, as opposed to just the estimate on distributions, uses the nonlinear graph-transform result cited on pp. 466–467.

Premises used

  • C² branch of (vii): with the summands taken orthogonal up to an arbitrarily small metric error, the projective contraction for u against cs is bounded with slack by ν̂/γ̂, and the inverse base derivative along a cu leaf is bounded with slack by 1/γ. The required product at exponent 1 is thus governed by ν̂/(γγ̂)<1 uniformly. Use f⁻¹ for s inside cs, obtaining ν/(γγ̂). Although p. 467 mentions a>1, p. 466 also requires a≤k−1; at k=2 it is a=1 that is used. The strict inequality already holds there, so no C²⁺ assumption is needed. The supplied PSW correction addresses the C¹ weak-leaf issue: restrict the ambient C¹ bundle map to the C¹ weak leaves, obtain the C¹ invariant strong bundle as an ambient subbundle over those leaves, and project each weak graph by a linear projection onto a smooth plane before applying Frobenius. The projected strong distribution is C¹ via the restriction of the ambient linear projection, which does not require differentiating the weak graph twice. Uniformity requires common graph cones, bounds for the graph-transform derivatives, and uniformly bounded chart sizes and inverses.
  • C¹⁺ᵟ branch of (vii): openness transfers (6)–(8) to the fiber maps for fixed θ, with a common strict margin, using derivative closeness and cone closeness also for minimum and maximum norms on the strong bundles. This transfers the hypotheses, not the holonomy conclusion. In particular the C² invariant-section calculation is not a proof of C¹ holonomy for this branch. The needed result concerns fake strong holonomy within invariant fake weak leaves of the fiber model, in both directions and with uniform estimates; a result only asserting regularity for globally dynamically coherent maps of compact manifolds would require a further application argument.

Missing inputs

  • HPS77 and KH95: available_from=1977 and 1995, respectively; date_evidence=BW journal p. 488; date_basis=publication years; versions=HPS LNM 583, in particular Theorem 6.1(e), and KH's 1995 Hadamard–Perron proof. Neither text is supplied. Needed: global invariant graph foliations for a bi-infinite sequence of fiber diffeomorphisms uniformly close in C¹ to the split linear sequence, under uniform domination and strong contraction/expansion; uniqueness in a common bounded-slope class, strong-leaf characterization and subfoliation of the weak graphs. Also needed from the section theorem is uniform C¹ regularity of an invariant section for a C¹ graph transform over a C¹ invertible base when fiber contraction times the inverse base derivative has uniform supremum less than 1. Exact source hypotheses and the full noncompact uniform application were not verified.
  • PSW97 and PSW00: available_from=1997 and 2000; date_evidence=BW journal p. 489 and the supplied correction's 2000 imprint; date_basis=publication years; versions=Duke Math. J. 86 original, and Duke Math. J. 105 correction of Theorem B. Original text unavailable. Theorem A input needed in (vi) is Hölder regularity for invariant distributions and graph-transform foliations with C¹⁺ᵟ dynamics and the strict domination-versus-base-rate inequality quoted on p. 466. The corrected Theorem B input in (vii) is C¹ strong subfoliation of invariant weak leaves under C² dynamics and the strict center-bunching ratio. The supplied correction supports the application mechanism but not verification of every hypothesis in the original theorem or of the uniform extension. The alternative references listed on p. 466 were not used to close these steps.
  • BW06b: available_from=2006 at the resolution of the journal bibliography, with availability of the needed version by 2005-10-11 unresolved; date_evidence=BW journal pp. 467 and 488; date_basis=preprint year recorded in the journal; version=A note on unstable holonomy between centers, not supplied. Needed: uniform C¹ subfoliation, or the precise holonomy regularity and chart implications yielding it, under C¹⁺ᵟ and (6)–(8) for the disjoint-fiber model and its inverse. Request the proof and a version record showing that the needed result was available by the cutoff. BW05 is a different preprint; neither the note's Lemma 1.1 nor the journal's Lemma 1.1 is supplied in primary-page form here.

Scope changes

  • The construction around the zero orbit, the rate arithmetic, and the regularity interfaces were checked. Global graph-transform existence and uniqueness, the full Hölder foliation theorem, and the low-regularity holonomy theorem were not re-proved or treated as verified imports.

Located concern

BW journal printed pp. 464–465, Proposition 3.1(iii)(a)–(b), compared with p. 468

The printed iterated plaque memberships keep the subscript p, although the plaques were defined only in B(p,r). The intermediate weak-plaque hypotheses have the same indexing problem.

Take a linear toral Anosov map times a circle identity, with its orthogonal product metric and constant strict partially hyperbolic rates. Choose a nonfixed p, and use the allowed rescaling so that d(p,f(p))>R while balls of radius R are in strong foliation boxes. For n=1 and q=q′=p, the stable hypothesis holds for every admissible r₁. But q₁=f(p) is outside the domain of the plaque with subscript p, for every r<R. The same defect does not occur in the one-step invariance statement (ii).

Use p_j and p_{−j} in the intermediate hypotheses, and p_n and p_{−n} in the conclusions. For example, the stable conclusion uses Ŵ^s_{p_n}(q_n,r₁). Iterate the one-step estimates on those copies, as the proof on p. 468 indicates.

The literal membership statement is not well-defined in general. The orbit-index correction restores the intended local statement and is already consistent with the use on p. 470. This is not a counterexample to the intended proposition or the principal theorems.

Located concern

BW journal printed pp. 457 and 468, the adapted metric convention and the one-step cs and cu distance estimates

Bounds on the three individual summands do not by themselves give the claimed weak-plaque bounds for ambient distance in an arbitrary adapted metric as defined on p. 457.

For the toral map given by the matrix with rows (2,1) and (1,1), times a circle identity, give unit stable and center directions inner product ρ close to 1, with the unstable direction orthogonal to both. Rates ν=ν̂=1/2, γ=0.9 and γ̂=1/1.1 obey (1)–(5). With λ=(3−√5)/2, the squared norm ratio on e_c−e_s is (1+λ²−2λρ)/(2−2ρ), which exceeds 1.1² for ρ sufficiently close to 1. At sufficiently small cone width the tangent of any proposed cs plaque has a direction near e_c−e_s, so the asserted one-step ambient distance estimate with bound 1.1 cannot follow in this fixed metric.

New local repair: before choosing scales, make the three summands orthogonal in the continuous product metric that keeps the original metric on each summand. Approximate this metric uniformly by a smooth one; strict uniform margins preserve the original separate bounds and the required weak-sum bounds. In a uniformly small exponential chart a secant of a small-slope weak plaque is close in direction to the corresponding weak plane. The coordinate map differs from its linearization by a map of uniformly small Lipschitz constant on the chart ball, and chart and metric comparison factors tend uniformly to 1 with the chart size. The strict weak-sum margins then give the stated chord estimates for f and f⁻¹.

This is a missing metric normalization for the local proof. It can be made in the freely chosen adapted metric without altering bunching, the measure class, or density points.

Located concern

BW journal printed pp. 466–467; PSW 2000 correction printed p. 105

The PSW correction's block notation does not literally support two descriptions in its graph-transform calculation.

A in (C+KP)(A+BP)⁻¹ is the block on the approximating unstable summand, so m(A) is not the base conorm along the weak leaf. Also the graph transform over the disjoint union of cu leaves sends the fiber at x to the fiber at f(x), rather than covering the identity on points of that union.

Use the map covering f on the disjoint union and the conorm of Df on Ecu for the base estimate. In the adapted orthogonal splitting that conorm is the center conorm; suitable close graph coordinates and strict bunching preserve the resulting inequality. Retain the projection argument on p. 106.

These notational slips do not defeat the bunching estimate or the remedy for the C¹ weak-leaf issue. The uniform invariant-section application still needs the inputs identified above.

G3

Checked statement: Section 4 orbitwise bounds, uniformity of distortion and choice of scales, through their uses in the julienne construction and the Section 5 density argument.

Returned assessment: repairable_gap

The forward product uses indices 0 through n−1; the backward product has reciprocal factors at −1 through −n. For Lemma 4.1, positivity on the compact base gives a positive minimum of α, so log α has a uniform Hölder constant. Summing the log differences gives exactly the claimed two-sided bound, including for the inverse product. For Lemma 4.2, with the corrected moving plaques and one-step estimates, a center offset remains at distance at most γ̂_j(p)⁻¹σ_n(p) from the forward image of its stable basepoint. This is at most (σ/γ̂)_j(p), because the remaining σ factors are at most 1. The stable basepoint is at distance at most ν_j(p) from p_j. These two offsets are each less than 1 in the open plaques and bounded by uniform geometric rates, since sup ν and sup(σ/γ̂) are less than 1. These bounds also verify the local hypotheses at the next step. For Lemma 4.3, the backward pullback of a terminal radius τ_n(p) has offset at time j at most τ_n(p) times the product of ν̂(p_i) over j≤i<n, both for the genuine unstable plaque and for the fake one, conditional on their respective contraction estimates. This establishes the induction in the radius-3 region. When τ<1, the offset is at most (sup τ)^j because τ_n≤(sup τ)^n and j≤n. When τ=1, only the uniform radius-3 assertion is used. Thus for τ<1 one may use κ=max{sup ν,sup(σ/γ̂),sup τ}<1 and a fixed additive bound on the three offsets. Lemma 4.4 follows since the sum of Hölder powers of the pairwise distances is at most (2C)^θ/(1−κ^θ). The constants may depend on f, the fixed metric, cone and rate choices, and the exponent and uniform Hölder bound, but not on p, n or the test points. No Hölder assumption on σ or τ is needed.

Premises used

  • Scale selection in the Section 5 author context: ν<min(γ,γγ̂) permits ν/γ<σ<min(γ̂,1), then ν<τ<σγ. Continuous intermediate functions may be chosen by averaging endpoints. Here σγ<γγ̂<1, and ν̂τ<σ because ν̂<γ̂. Strictness and compactness give uniform margins for products and uniform lower bounds for positive successive-radius ratios. In the C¹⁺ᵟ branch this choice uses the implication from (7) to (5), not a weakening of the assumption needed for (vii). For the reversed argument the hatted and unhatted rates exchange and are evaluated at f⁻¹(p), including the lower strong rates in (6).
  • The author context uses fake center and fake unstable pieces for a julienne contained in a single C¹ fake cu plaque. Fake stable holonomy between fake cu plaques preserves the julienne sequences up to bounded index shifts; (vii) supplies the uniform bounded-Jacobian interface for transferring densities. Fake unstable regularity inside fake cu is also used for volume regularity. The starting centers lie on the genuine stable plaque through p, where agreement with the fake plaque is precisely the basepoint uniqueness in (v), not agreement on every nearby leaf.
  • Section 8 author-context interface: the comparison from C_n to D_n uses genuine unstable fibers, essential genuine unstable saturation and bounded-Jacobian absolute continuity. D_n to E_n changes genuine to fake fibers by geometric internesting, not by treating fake fibers as saturated. The final comparison from G_n to a fake cu julienne uses the genuine stable foliation and a fully genuine stable-saturated representative. For that representative the essential genuine unstable saturation is unchanged. Both density-transfer steps require the genuine foliation's measure comparison on the indicated C¹ transverse disks.

Missing inputs

  • The journal's Lemma 1.1 and its adapted-metric proof, invoked on BW journal p. 470, are not supplied as images. The needed statement contracts distances and retains local genuine strong plaque membership while the comparison orbit remains in the designated small balls, with forward factors ν(p_i) and backward factors ν̂(p_i) at the correct indices. The author-context lemma has this role; it is not the holonomy note's lemma.
  • The journal Section 5–8 images are not supplied. The uses of bounded-index internesting, uniform fake holonomy, regularity of actual julienne volumes, and genuine-foliation disintegration were traced in the author context, not certified line by line for the journal proof.
  • Genuine strong foliation input: needed is transverse absolute continuity and the associated bounded-Jacobian Fubini comparison for genuine stable and unstable foliations of a C¹⁺ᵟ partially hyperbolic map with continuous pointwise rates, uniformly over sufficiently small boxes and transversals of uniformly positive angle, including the C¹ fake weak plaques used as transversals. For the genuine unstable distortion one also needs a uniform Hölder estimate on the genuine unstable tangent bundle. BP74 and PS72: available_from=1974 and 1972; date_evidence=BW journal p. 488; date_basis=publication years; versions=the original published partial-hyperbolicity results, described in the author context as treating constant rates; texts not supplied. The author context also invokes Pesin theory without giving a precise theorem for this quantitative interface: available_from=unresolved for the needed formulation; date_evidence=the general attribution in the author text; date_basis=undocumented there; version=unspecified. The cited direct pointwise-rate proof AV08 has available_from=2008 in the supplied record, date_evidence=BW journal pp. 487–488, date_basis=preprint year, version=Flavors of partial hyperbolicity; no pre-cutoff version containing the needed result is evidenced. These inputs are not closed by the present audit.

Scope changes

  • Checked the orbit estimates and scale arithmetic directly, including endpoint cases and the distinction between τ=1 and τ<1. Later julienne proofs were used to identify consumers of those estimates, not as a substitute for missing journal pages.

Located concern

BW journal printed p. 470, proof of Lemma 4.2

The displayed comparison of γ̂_j(p)⁻¹σ_n(p) with γ̂_j(p)⁻¹σ_j(p) is strict even at the endpoint j=n.

At j=n the two quantities are identical, including when n=j=0. The original plaque radius bound is open, so equality of these two comparison bounds does not cause an orbit to escape.

Use ≤ for all 0≤j≤n and then bound the common expression by (sup σ/γ̂)^j, meaning the j-th power of the supremum of the pointwise ratio. Keep the open-plaque inequalities for the radius-2 containment.

No change to Lemma 4.2 or to the downstream neighborhood estimates.

Located concern

BW journal printed pp. 468–470, Lemmas 4.1 and 4.4; fake-Jacobian use identified in the October 2006 author context, PDF p. 38

Lemma 4.4 as stated is about a single Hölder function on M. A Jacobian on a fake distribution is naturally a function on an orbit-indexed copy of a neighborhood; it need not descend to one function on M.

The fake foliations depend on the chosen center and the extension. Overlap of two physical balls does not identify their fake bundles. The chain rule along a tracked orbit uses the function on the j-th copy at time j, and compares both orbits in that same copy.

New direct extension of the distortion calculation: write the two Jacobians as products of α_j evaluated in the j-th copy, using local invariance of the fake tangent bundle along both tracked orbits. If the α_j have a common positive lower bound and a common Hölder bound, the difference of the log products is at most H times the same sum of distance powers, with H independent of j. The common lower bound follows from the conorm of Df on the compact manifold, and the Hölder bound follows from the uniform within-copy tangent Hölder estimate in (vi) and the Hölder regularity of Df. Exponentiation gives the required uniform comparison without a global fake Jacobian on M.

The printed Section 4 distortion lemma for a single function is valid. The downstream interface needs this orbit-indexed version or its explicit calculation, conditional on (vi); it does not require fake distributions to agree on overlaps.

Outside the audit

  • Full published proof on journal pages not supplied, including the journal's Lemma 1.1 and the detailed proofs of the julienne propositions.
  • The separate supporting holonomy note and any pre-cutoff version history for it.
  • Full cited graph-transform, foliation regularity, genuine absolute continuity and Pinsker proofs, except for the supplied PSW correction.
  • Examples, open questions, and other assertions outside G1–G3.
Source record

Additional supporting archive items are identified in the original report. View the manuscript hashes and repository snapshot.

Cite this review

If this review or its supporting reports inform your work, please cite it.

Igor Rivin. Review of On the ergodicity of partially hyperbolic systems. Math Referee, .

Download .bib
View BibTeX
@misc{rivin2026review-human-burns-wilkinson-ergodicity,
  author = {Rivin, Igor},
  title = {{Review of On the ergodicity of partially hyperbolic systems}},
  year = {2026},
  month = {November},
  date = {2026-11-21},
  howpublished = {Math Referee},
  url = {https://mathreferee.dimensionreducers.ai/papers/human-burns-wilkinson-ergodicity/panel/},
  note = {Review report}
}