Non-Uniform Specification Property

• Non-uniform specification is a relaxed version of Bowen’s property, allowing orbit segments to be concatenated with gap functions that vary with segment length or location.
• It underpins rigorous results in thermodynamic formalism, including uniqueness of equilibrium states, weak Gibbs estimates, and exponential decay of correlations.
• The property appears in diverse formulations across symbolic dynamics, hyperbolic measures, and random systems, guiding equilibrium and statistical mechanics analyses.

Searching arXiv for the cited work and closely related papers on non-uniform specification. Non-uniform specification property is a weakening of Bowen’s classical specification property in which the gluing cost for orbit segments is no longer required to be bounded by a single constant. Instead, either one works with a gap function that may increase with the segment length, or one restricts gluing to a distinguished family of “good” orbit segments while controlling the complexity or pressure of the remaining obstructions. Across several formulations, the common principle is that orbit concatenation remains possible with asymptotically negligible overhead, and this weaker combinatorial shadowing mechanism is still sufficient for strong conclusions in thermodynamic formalism, including uniqueness of equilibrium states, full support, weak Gibbs estimates, and in some settings the $K$-property, Bernoulli behavior, exponential decay of correlations, and periodic-orbit approximation (Pavlov, 2017, Climenhaga, 2015, Climenhaga et al., 2011).

1. Classical specification and its non-uniform replacements

Bowen’s classical specification property requires that for every $\varepsilon>0$ there is a fixed constant $C=C(\varepsilon)$ such that any finite collection of orbit segments can be $\varepsilon$-shadowed in order by a single orbit, with gaps of length at least $C$; in the classical form one may additionally require the shadowing orbit to be periodic (Pavlov, 2017). In symbolic dynamics this is equivalent to the existence of a uniform bridge length $\tau$ so that any finite concatenation of admissible words can be connected by words of length at most $\tau$ (Climenhaga et al., 2020, Climenhaga et al., 2011).

Non-uniform specification replaces this uniform gluing constant by data that may depend on orbit length, location in phase space, or membership in a chosen “good” subset. One prominent formulation, due to Marcus and developed by Pavlov, introduces a nondecreasing function $f\colon \mathbb N\to\mathbb N$ and requires that orbit segments of lengths $n_i$ can be concatenated provided the intermediate gaps satisfy

$m_i \ge \max\{f(n_i),f(n_{i+1})\}.$

The shadowing point need not be periodic (Pavlov, 2017). If $\varepsilon>0$0 is constant, one recovers ordinary specification.

A second family of formulations uses a decomposition of language or orbit-segment space into prefixes, good cores, and suffixes. In symbolic form one writes $\varepsilon>0$1, where every word can be factored into a prefix obstruction, a good core, and a suffix obstruction; the good family $\varepsilon>0$2 or its truncations $\varepsilon>0$3 satisfy specification, while the obstructions have small entropy or pressure (Climenhaga et al., 2011, Climenhaga et al., 2020). In continuous-time and non-symbolic settings the analogous structure is an orbit decomposition $\varepsilon>0$4 or $\varepsilon>0$5, together with specification only on $\varepsilon>0$6 or $\varepsilon>0$7 (Climenhaga et al., 2015, Climenhaga et al., 2020).

A third direction, developed for hyperbolic measures, is a single-segment periodic approximation property: for $\varepsilon>0$8-almost every point, every sufficiently long dynamical ball contains a periodic point whose period exceeds the segment length by only a sublinear term. This is also called nonuniform specification in the work of Oliveira and collaborators (Oliveira, 2010, Oliveira et al., 2011). The same paradigm appears in random dynamical systems along fibers (Bilbao, 2021).

2. Formal variants of the property

In expansive systems, one formulation fixes a scale $\varepsilon>0$9 and asks that for any $C=C(\varepsilon)$0, any points $C=C(\varepsilon)$1, and any lengths $C=C(\varepsilon)$2, if the gaps $C=C(\varepsilon)$3 satisfy

$C=C(\varepsilon)$4

then there exists $C=C(\varepsilon)$5 such that

$C=C(\varepsilon)$6

where $C=C(\varepsilon)$7 and $C=C(\varepsilon)$8 (Pavlov, 2017). This is the Marcus–Pavlov type non-uniform specification property.

Pavlov also introduced a weaker property, non-uniform transitivity, in which one only glues two orbit segments at a time: for every pair $C=C(\varepsilon)$9 and every length $\varepsilon$0, there is a single gap $\varepsilon$1 and a point $\varepsilon$2 that shadows $\varepsilon$3 for $\varepsilon$4 iterates, waits $\varepsilon$5 iterates, then shadows $\varepsilon$6 for $\varepsilon$7 iterates (Pavlov, 2017). This is strictly weaker than the multi-segment gluing property.

In symbolic dynamics, Climenhaga–Thompson distinguish several strengths of non-uniform specification for a good collection $\varepsilon$8. In the weak form (W), any finite list of words in $\varepsilon$9 can be concatenated using connecting words of length at most $C$0; the strong form (S) requires each connecting word to have length exactly $C$1; the periodic form (Per) further requires that the resulting cylinder contain a periodic point of period equal to total length plus $C$2 (Climenhaga et al., 2011). These are still uniform within $C$3, but non-uniform at the system level because arbitrary words need not lie in $C$4.

For maps and flows with orbit decompositions $C$5, the corresponding condition is that each truncated good set $C$6 has weak specification at some scale $C$7 with a gap bound $C$8, meaning that every finite family of segments in $C$9 can be shadowed in order with transition times in $\tau$0 (Climenhaga et al., 2015). The 2025 extension to controlled specification allows the gaps themselves to depend on orbit lengths through a nondecreasing function $\tau$1, requiring

$\tau$2

with shadowing at a fixed scale $\tau$3 (Wang et al., 17 Apr 2025). This suggests a direct bridge between Marcus–Pavlov style gap functions and orbit-decomposition approaches.

For ergodic hyperbolic or expanding measures, the definition is expressed using non-uniform dynamical balls. In the endomorphism setting, Oliveira’s definition uses $\tau$4-balls

$\tau$5

where $\tau$6 is $\tau$7-slowly varying, and asks that $\tau$8 contain a periodic point of period $\tau$9 with

$\tau$0

(Oliveira, 2010). Oliveira–Tian extended this to two-sided balls $\tau$1 for diffeomorphisms (Oliveira et al., 2011). Bilbao proved the random analogue along fibers for random dynamical systems, requiring a return point $\tau$2 with period $\tau$3 and $\tau$4 (Bilbao, 2021).

3. Interaction with regularity of the potential

The role played by Bowen’s uniform distortion bound in classical theory is likewise relaxed in non-uniform settings. In Pavlov’s expansive-system theorem, one introduces the partial-sum variation

$\tau$5

and requires a nondecreasing function $\tau$6 with $\tau$7 for all $\tau$8 (Pavlov, 2017). Thus the potential is allowed to have variation growing with $\tau$9, provided it grows sufficiently slowly.

In the symbolic decomposition approach, the analogous requirement is the Bowen property only on the good collection $f\colon \mathbb N\to\mathbb N$0: $f\colon \mathbb N\to\mathbb N$1 uniformly bounded in $f\colon \mathbb N\to\mathbb N$2 (Climenhaga et al., 2011). Potentials may fail the Bowen property globally but remain regular on the good core. Climenhaga–Thompson use this to treat “grid” potentials on systems with full classical specification, as well as $f\colon \mathbb N\to\mathbb N$3-shifts, where the obstruction sets are controlled by a summability condition

$f\colon \mathbb N\to\mathbb N$4

(Climenhaga et al., 2011).

The flow and homeomorphism framework of Climenhaga–Thompson similarly replaces global regularity by the Bowen property on $f\colon \mathbb N\to\mathbb N$5, combined with a pressure gap for the bad pieces $f\colon \mathbb N\to\mathbb N$6 (Climenhaga et al., 2015). The 2025 controlled-specification variant further weakens the regularity to a $f\colon \mathbb N\to\mathbb N$7-distortion condition, with uniqueness under the asymptotic requirement

$f\colon \mathbb N\to\mathbb N$8

for each fixed $f\colon \mathbb N\to\mathbb N$9 (Wang et al., 17 Apr 2025). This suggests a systematic principle: uniqueness can survive if both the orbit-gluing cost and the distortion cost are asymptotically negligible.

4. Uniqueness of equilibrium states and the logarithmic threshold

The central theorem of Pavlov’s paper states that if $n_i$0 is expansive with expansivity constant $n_i$1, $n_i$2, $n_i$3 has non-uniform specification at scale $n_i$4 with gap bounds $n_i$5, $n_i$6 has partial-sum-variation bounds $n_i$7 at the same scale, and

$n_i$8

then $n_i$9 has a unique equilibrium state $m_i \ge \max\{f(n_i),f(n_{i+1})\}.$0, and $m_i \ge \max\{f(n_i),f(n_{i+1})\}.$1 is fully supported (Pavlov, 2017). Under non-uniform transitivity one obtains the same uniqueness and full-support conclusion under the stronger condition

$m_i \ge \max\{f(n_i),f(n_{i+1})\}.$2

(Pavlov, 2017).

A distinctive feature of this theorem is its sharp threshold. Pavlov shows that logarithmic growth is optimal: if either $m_i \ge \max\{f(n_i),f(n_{i+1})\}.$3 or $m_i \ge \max\{f(n_i),f(n_{i+1})\}.$4 is bounded away from zero, uniqueness may fail (Pavlov, 2017). The examples recorded there include the “Double Hofbauer model,” where $m_i \ge \max\{f(n_i),f(n_{i+1})\}.$5 and $m_i \ge \max\{f(n_i),f(n_{i+1})\}.$6, and subshifts due to Kwietniak–Pavlov with multiple measures of maximal entropy once $m_i \ge \max\{f(n_i),f(n_{i+1})\}.$7 and $m_i \ge \max\{f(n_i),f(n_{i+1})\}.$8 (Pavlov, 2017). Thus $m_i \ge \max\{f(n_i),f(n_{i+1})\}.$9 is the optimal transition point.

The proof strategy is a contradiction argument. Assuming two distinct ergodic equilibrium states $\varepsilon>0$00, one uses mutual singularity to find disjoint large-mass compact sets, constructs large families of orbit pieces with prescribed visitation properties via a maximal ergodic theorem argument, and then glues orbit segments from the two families using non-uniform specification (Pavlov, 2017). The control of Birkhoff-sum distortion is provided by $\varepsilon>0$01, while the cost of inserted gaps is measured by $\varepsilon>0$02. The resulting separated set has partition sum exceeding $\varepsilon>0$03 by a power-law factor, contradicting the upper bound on partition sums implied by the sublogarithmic hypothesis (Pavlov, 2017). A related argument proves full support.

The decomposition-based uniqueness theorems have a different formal shape but the same architecture. For shifts, Climenhaga–Thompson prove that if $\varepsilon>0$04, every $\varepsilon>0$05 has (S)-specification, $\varepsilon>0$06, and the obstruction series is summable, then $\varepsilon>0$07 has a unique equilibrium state $\varepsilon>0$08 with weak Gibbs bounds on $\varepsilon>0$09 (Climenhaga et al., 2011). For homeomorphisms and flows, if $\varepsilon>0$10 or $\varepsilon>0$11 satisfies weak specification, $\varepsilon>0$12 has the Bowen property on $\varepsilon>0$13, and the pressure of $\varepsilon>0$14 is strictly below $\varepsilon>0$15, then uniqueness follows (Climenhaga et al., 2015). In shift spaces satisfying non-uniform specification in the tower sense, one can model the system by a strongly positive recurrent countable-state Markov shift, which yields uniqueness and stronger statistical conclusions (Climenhaga, 2015).

5. Statistical properties and structural consequences

Non-uniform specification is not merely a uniqueness criterion. In Pavlov’s expansive-system setting, the unique equilibrium state obtained under non-uniform specification has the Kolmogorov $\varepsilon>0$16-property. The argument applies the uniqueness theorem to the product system $\varepsilon>0$17 with the product potential $\varepsilon>0$18, following Ledrappier’s method (Pavlov, 2017).

For shift spaces admitting the tower/Markov coding of Climenhaga, non-uniform specification leads to a strongly positive recurrent countable-state Markov shift $\varepsilon>0$19 and a 1-block code $\varepsilon>0$20 carrying all high-pressure measures (Climenhaga, 2015). Since the lifted potential is strongly positive recurrent, the equilibrium measure on $\varepsilon>0$21 is unique, and therefore so is the equilibrium state on $\varepsilon>0$22 (Climenhaga, 2015). This coding yields strong statistical consequences: Bernoulli up to a period in the two-sided case, exponential decay of correlations for Hölder observables, a central limit theorem, and analyticity of pressure near a Hölder perturbation (Climenhaga, 2015). The abstract states that these properties are new even for uniform specification (Climenhaga, 2015).

In the decomposition approach of Climenhaga–Thompson, the unique equilibrium state satisfies a weak Gibbs bound on cylinders from the good set $\varepsilon>0$23: $\varepsilon>0$24 (Climenhaga et al., 2011). In the periodic-specification case, the equilibrium state is the weak$\varepsilon>0$25 limit of weighted periodic orbit measures (Climenhaga et al., 2011). For homeomorphisms with non-uniform structure, the unique equilibrium state also satisfies a level-2 large deviations upper bound (Climenhaga et al., 2015).

The survey “Beyond Bowen’s Specification Property” organizes these consequences as part of a general program extending Bowen’s methods beyond uniform hyperbolicity, including applications to partially hyperbolic systems and geodesic flows beyond negative curvature (Climenhaga et al., 2020). A plausible implication is that non-uniform specification functions less as a single property than as a family of combinatorial templates for isolating thermodynamically dominant orbit segments.

6. Examples, applications, and sharp counterexamples

Several classes of systems illustrate different forms of non-uniform specification.

Setting	Form of non-uniformity	Consequence recorded in the sources
Expansive systems with gap function $\varepsilon>0$26 and variation $\varepsilon>0$27	Sublogarithmic growth of $\varepsilon>0$28	Unique fully supported equilibrium state; $\varepsilon>0$29-property in the specification case (Pavlov, 2017)
Shift spaces with good set $\varepsilon>0$30 and low-pressure obstructions	Specification only on $\varepsilon>0$31	Unique equilibrium state with weak Gibbs property (Climenhaga et al., 2011)
Shift spaces modeled by SPR Markov shifts	Non-uniform specification in the language	Bernoulli up to a period, exponential decay of correlations, CLT, analyticity of pressure (Climenhaga, 2015)
Hyperbolic or expanding measures	Periodic shadowing with sublinear overhead	Approximation by periodic orbit measures (Oliveira, 2010, Oliveira et al., 2011)
Random dynamical systems	Fiberwise return-point shadowing with sublinear lag	Approximation by finitely supported fiber measures (Bilbao, 2021)

In Pavlov’s paper, bounded-density shifts are a principal example: if words are forbidden when the total symbol sum over length $\varepsilon>0$32 exceeds $\varepsilon>0$33, then the shift has non-uniform specification with

$\varepsilon>0$34

Whenever $\varepsilon>0$35, this yields a unique measure of maximal entropy (Pavlov, 2017). The same paper also gives examples of potentials that fail the Bowen property globally but satisfy $\varepsilon>0$36, and therefore still have unique equilibrium states (Pavlov, 2017).

Climenhaga–Thompson’s symbolic theory applies to $\varepsilon>0$37-shifts, non-Bowen “grid” potentials, piecewise monotonic interval maps, and more general settings where the obstructions can be isolated in $\varepsilon>0$38 and $\varepsilon>0$39 and made summable in pressure (Climenhaga et al., 2011). The tower construction of Climenhaga applies to shifts of quasi-finite type, synchronised shifts, coded shifts, and factors of $\varepsilon>0$40-shifts and $\varepsilon>0$41-gap shifts (Climenhaga, 2015).

In the broader survey literature, non-uniform specification via decompositions is applied to Bonatti–Viana partially hyperbolic diffeomorphisms, geodesic flows in nonpositive curvature, Manneville–Pomeau-type maps, derived-from-Anosov systems, and rank-1 geodesic flows (Climenhaga et al., 2015, Climenhaga et al., 2020). In these examples the good set typically consists of orbit segments spending enough time in uniformly hyperbolic or regular regions, while the bad set has smaller pressure (Climenhaga et al., 2015, Climenhaga et al., 2020).

The sharpness of the theory is equally important. Pavlov’s logarithmic threshold shows that once gluing or distortion grows at logarithmic scale, uniqueness can fail (Pavlov, 2017). The more recent quantitative work of Lin–Tian–Yu considers a gap function $\varepsilon>0$42 with asymptotic linear growth rate

$\varepsilon>0$43

and demonstrates that when $\varepsilon>0$44, several classical specification consequences may fail: transitive points may not have full Bowen entropy, the conditional variational principle may fail, and the intermediate entropy property may be destroyed (Lin et al., 25 Aug 2025). This does not alter Pavlov’s logarithmic uniqueness theorem, but it shows that for other topological and multifractal consequences, a positive linear gap rate can fundamentally change the picture.

7. Measure-theoretic and random forms

The measure-theoretic version of nonuniform specification arose in the study of non-uniform hyperbolicity. Oliveira proved that every ergodic expanding measure for a $\varepsilon>0$45 endomorphism with non-flat critical set and strong transitivity on its support has the nonuniform specification property (Oliveira, 2010). The proof uses abundance and non-lacunarity of hyperbolic times: positive Lyapunov exponents imply infinitely many hyperbolic times with positive lower density, integrability of the first hyperbolic time, and then sublinear gaps between consecutive hyperbolic times (Oliveira, 2010). This makes it possible to find periodic points in almost every dynamical ball with period $\varepsilon>0$46.

Oliveira–Tian extended this result to $\varepsilon>0$47-diffeomorphisms for any ergodic hyperbolic measure, and to $\varepsilon>0$48 diffeomorphisms when the Oseledec splitting is dominated on the support of the measure (Oliveira et al., 2011). They also established a generalized non-uniform specification allowing simultaneous shadowing of finitely many orbit segments in prescribed order (Oliveira et al., 2011). Consequences include quantitative Poincaré recurrence, weak-$\varepsilon>0$49 approximation by periodic orbit measures, and periodic-orbit formulas for topological pressure in positively expansive settings where high-pressure measures are hyperbolic (Oliveira et al., 2011).

Bilbao carried the same philosophy to random dynamical systems. For an ergodic nonuniformly expanding invariant measure of a skew product $\varepsilon>0$50, the system has nonuniform specification along fibers: for $\varepsilon>0$51-almost every $\varepsilon>0$52, every $\varepsilon>0$53 and every sufficiently small $\varepsilon>0$54, there exists a return point $\varepsilon>0$55 of period at most $\varepsilon>0$56, with $\varepsilon>0$57 (Bilbao, 2021). The proof again relies on hyperbolic times, now fiberwise, together with non-lacunarity and a ball-mixing argument using topological exactness in the fibers (Bilbao, 2021). The corollaries state that any expanding measure is a weak$\varepsilon>0$58 limit of invariant measures whose fiber disintegrations are supported on finitely many return points, and that empirical measures of these return orbits converge to the disintegrated fiber measures (Bilbao, 2021).

These measure-centered formulations are weaker than Bowen’s full multi-segment specification, since they only guarantee periodic approximation of a single typical orbit segment. Nonetheless, they preserve a core dynamical feature: typical finite-time statistics can be realized by true periodic data with asymptotically negligible overhead (Oliveira, 2010, Oliveira et al., 2011, Bilbao, 2021). This suggests why they remain effective for recurrence estimates and periodic approximation even when they are not by themselves the exact hypotheses used in abstract uniqueness theorems.