Statistical Observables in Complex Systems

• Statistical observables are mathematical entities—such as SRB measures, Hermitian operators, and observable charts—that capture asymptotic, ensemble, or geometric statistics in diverse systems.
• They replace inaccessible microscopic details with operational quantities to analyze phenomena like phase transitions, equilibrium states, and quantum randomness.
• These observables bridge empirical data and theoretical models, enabling effective inference in dynamical systems, quantum mechanics, and cosmological observations.

Searching arXiv for relevant papers on “statistical observables” and closely related usages across dynamical systems, quantum physics, statistical mechanics, and statistics. Across recent literature, “statistical observables” denotes several distinct but related constructions through which asymptotic statistics, measurement statistics, or macroscopic constraints are described. In one line of work, the term refers to observable or SRB-like probabilities that organize the asymptotic statistics of Lebesgue almost every initial state in continuous dynamics; in another, it refers to Hermitian test observables whose expectation values and moments are compared with Haar-random predictions; elsewhere it denotes observable functionals $\psi_f(P)=\mathbb E_P[f]$ that define coordinate systems directly on model space, additive observables that specify phase coexistence states, or reduced angular $N$-point spectra that connect one-sky estimators with ensemble predictions (Catsigeras et al., 2011, Bonet-Monroig et al., 2024, Plummer, 1 Apr 2026, Yoneta et al., 2021, Mitsou et al., 2019). A common thread is the replacement of inaccessible microscopic detail by quantities whose statistics are operationally, asymptotically, or geometrically meaningful.

1. Field-dependent meanings of the term

The cited literature does not use “statistical observables” in a single universal sense. Instead, the expression labels different objects depending on whether the primary problem is dynamical, quantum-statistical, inferential, geometric, or thermodynamic.

Context	Observable object	Statistical role
$C^0$ dynamics	Observable or SRB-like measure $\mu$	Describes asymptotic statistics for Lebesgue almost every initial state
Quantum-state randomness	Hermitian observable $\mathcal O$	Generates expectation values and sample moments compared with Haar moments
Singular statistical models	Observable chart $\Psi(P)$	Gives parameterization-invariant coordinates on model space
Many-body equilibrium	Outcome probabilities $p_m=\mathrm{Tr}(\rho\Pi_m)$ or additive observables $\hat X_i^{(N)}$	Encodes equilibrium statistics or macroscopic constraints

In Catsigeras–Enrich, an observable measure is a Borel probability whose $\varepsilon$-basin has strictly positive Lebesgue measure for every $\varepsilon>0$; in Bonet-Monroig, Wang, and Pérez-Salinas, the observable is a Hermitian operator with spectral decomposition $N$0; in “Observable Geometry of Singular Statistical Models,” an observable chart is a finite tuple of expectations; and in “Observable Statistical Mechanics,” the central object is the Shannon entropy of measurement outcomes (Catsigeras et al., 2011, Bonet-Monroig et al., 2024, Plummer, 1 Apr 2026, Scarpa et al., 2023).

A common misconception is that an observable must always be a directly measured scalar quantity. The literature summarized here shows a wider usage: an observable can be a probability measure, a chart on model space, a weighted cross section, a macroscopic equivalence class in a $N$1-algebraic quotient, or a family of continuous linear functionals generating a generalized Markov chain (Faigle et al., 2017, Ven, 30 Oct 2025).

2. Dynamical systems: asymptotic statistics and observable measures

For a continuous map $N$2 on a compact manifold, Catsigeras–Enrich define the empirical measures

$N$3

and the set $N$4 of all weak$N$5-limit points of $N$6. A probability $N$7 is observable, or SRB-like, if for every $N$8 the set

$N$9

has strictly positive Lebesgue measure. The resulting set $C^0$0 of observable measures is nonempty and weak$C^0$1-compact for every continuous $C^0$2, and it is the unique minimal weak$C^0$3-compact set $C^0$4 with $C^0$5 (Catsigeras et al., 2011).

This construction generalizes classical SRB or physical measures. By construction every SRB measure is observable, but the converse need not hold in the $C^0$6 setting. The paper also proves that any isolated measure in $C^0$7 is SRB, and that if $C^0$8 is finite or countably infinite then there exist finitely or countably many classical SRB measures whose basins cover $C^0$9 Lebesgue a.e. The “convex-like” property of $\mu$0 is central: either $\mu$1 is a singleton or it is uncountable. This explains why isolated observable measures recover the usual physical measures, whereas continua of observable measures may still organize asymptotic behavior for typical initial data (Catsigeras et al., 2011).

A distinct but related use appears in the study of chaotic trajectories. There the dynamical observable is

$\mu$2

with $\mu$3 in an ergodic chaotic system. Typical fluctuations satisfy a central limit theorem, while rare fluctuations obey a large deviation principle with rate function obtained from the scaled cumulant generating function. The paper “Remarkable universalities in distributions of dynamical observables in chaotic systems” identifies a special class of “derived” observables,

$\mu$4

for which the sum is a boundary term of $\mu$5, the mean vanishes, the Green–Kubo sum vanishes, and the large-deviation rate function is trivial. If two observables differ by a derived function, they share exactly the same rate function (Defaveri et al., 14 May 2025).

Faigle–Gierz place observables in yet another dynamical framework: a time-discrete linear evolution $\mu$6 in a Banach space $\mu$7, together with a family of continuous linear functionals $\mu$8 satisfying $\mu$9 and $\mathcal O$0. This yields a generalized Markov chain. Joint observability, in this setting, leads to a criterion that in the quantum model two observables are jointly observable if and only if the corresponding self-adjoint operators commute (Faigle et al., 2017).

3. Quantum measurement statistics, randomness tests, and observable entropy

In quantum-state randomness verification, the observable is fixed first and the state ensemble is tested through the resulting statistics. Bonet-Monroig, Wang, and Pérez-Salinas consider a Hermitian observable

$\mathcal O$1

on an $\mathcal O$2-dimensional Hilbert space. For a pure state $\mathcal O$3, the expectation value is $\mathcal O$4. If $\mathcal O$5 is Haar-random, then in the basis that diagonalizes $\mathcal O$6 the outcome probabilities $\mathcal O$7 are distributed according to a symmetric Dirichlet law with parameters $\mathcal O$8, where $\mathcal O$9 is the multiplicity of $\Psi(P)$0. This gives a closed-form expression for the Haar moments $\Psi(P)$1 and motivates the sample moment

$\Psi(P)$2

and the average-randomness deviation

$\Psi(P)$3

If $\Psi(P)$4, the set is compatible with Haar-$\Psi(P)$5 design behavior as seen through $\Psi(P)$6; the paper terms this an $\Psi(P)$7-shadowed $\Psi(P)$8-design. Permutations of the eigenbasis and conjugation by a tomographically-complete set of unitaries are then used to strengthen the test toward full $\Psi(P)$9-design compatibility (Bonet-Monroig et al., 2024).

A complementary quantum-statistical perspective is given by Observable Statistical Mechanics. There the starting point is a PVM $p_m=\mathrm{Tr}(\rho\Pi_m)$0 for an observable $p_m=\mathrm{Tr}(\rho\Pi_m)$1, with outcome probabilities

$p_m=\mathrm{Tr}(\rho\Pi_m)$2

The observable entropy is

$p_m=\mathrm{Tr}(\rho\Pi_m)$3

The Maximum Observable Entropy Principle states that equilibrium measurement statistics maximize $p_m=\mathrm{Tr}(\rho\Pi_m)$4 under normalization and fixed average energy, leading to

$p_m=\mathrm{Tr}(\rho\Pi_m)$5

The paper compares this construction with the Diagonal Ensemble and reports that when $p_m=\mathrm{Tr}(\rho\Pi_m)$6 is highly degenerate, the exponential form accurately reproduces equilibrium outcome probabilities to high precision in extensive numerics on seven spin-$p_m=\mathrm{Tr}(\rho\Pi_m)$7 Hamiltonians (Scarpa et al., 2023).

These two lines of work emphasize different aspects of the same general shift. Average-randomness verification asks whether a set of states is statistically compatible with Haar moments when probed by an observable, whereas Observable Statistical Mechanics asks whether equilibrium measurement statistics can be inferred directly from accessible outcomes without diagonalizing the Hamiltonian. This suggests a broader operational view in which the observable is not merely a readout but the object that defines which statistics are being tested or predicted (Bonet-Monroig et al., 2024, Scarpa et al., 2023).

4. Correlators and optimal observables as inference devices

Several papers treat observables as statistically efficient summaries rather than raw measurements. In $p_m=\mathrm{Tr}(\rho\Pi_m)$8 searches for a heavy $p_m=\mathrm{Tr}(\rho\Pi_m)$9, the optimal observable is the weighted cross section

$\hat X_i^{(N)}$0

Maximizing the ratio of signal to statistical uncertainty yields the weight

$\hat X_i^{(N)}$1

The resulting observable is completely equivalent to the $\hat X_i^{(N)}$2 fit of the differential cross section and can be used as an alternative to aggregating events into bins with further minimization of the $\hat X_i^{(N)}$3 function. In the two-parameter case, a vector of weights is constructed so that the corresponding estimators are orthogonal to the effect of the other parameter (Gulov et al., 2017).

In bipartite quantum systems, the relevant statistical observables are correlators computed from a pair of complementary observables associated with mutually unbiased bases. For two qudits, the paper defines mutual predictability,

$\hat X_i^{(N)}$4

mutual information,

$\hat X_i^{(N)}$5

and the Pearson correlation coefficient. Using only the pair $\hat X_i^{(N)}$6 and $\hat X_i^{(N)}$7, it derives analytic relations between Negativity and these correlators for noisy Bell states, Werner states, and the one-parameter Horodecki two-qutrit state (Sadana et al., 2022).

The role of these correlators is twofold. First, the relations provide state-dependent separability bounds for a fixed choice of the complementary observables. Second, for the Horodecki family they distinguish separable, bound-entangled, and NPT-entangled regions through thresholds in the measured correlators. In this usage, a “statistical observable” is a function of measurement statistics chosen because it both detects and quantifies structure that would otherwise require much more extensive state reconstruction (Sadana et al., 2022).

5. Observable charts and the geometry of singular statistical models

“Observable Geometry of Singular Statistical Models” moves the discussion from parameter space to model space. Let $\hat X_i^{(N)}$8 and let each measurable $\hat X_i^{(N)}$9 define an observable functional

$\varepsilon$0

A finite collection $\varepsilon$1 defines an observable chart

$\varepsilon$2

whose image is a finite-dimensional presentation of the model. Observable completeness then formalizes whether the chart detects identifiable directions, and observable order measures the first nonzero order at which an analytic curve becomes visible in observable coordinates (Plummer, 1 Apr 2026).

At a reference point $\varepsilon$3, first-order completeness is expressed by

$\varepsilon$4

where $\varepsilon$5 is the score. More generally, the observable order of an analytic curve $\varepsilon$6 is

$\varepsilon$7

The main theorem states that, under first-order completeness and the standard $\varepsilon$8 expansion of KL divergence,

$\varepsilon$9

In many models one even has equality (Plummer, 1 Apr 2026).

The examples make the higher-order content explicit. In a two-component Gaussian mixture near a singular point, the mean is first-order visible, the variance makes $\varepsilon>0$0 visible at order $\varepsilon>0$1, and the third cumulant makes $\varepsilon>0$2 visible at order $\varepsilon>0$3. In reduced-rank regression, the rank constraint is invisible at first order and appears at second order. The paper’s central claim is parameterization-invariance: observable charts depend only on the model image $\varepsilon>0$4, not on a chosen parameterization. This shifts “observability” from an empirical notion to an intrinsic geometric one (Plummer, 1 Apr 2026).

6. Additive, global, and post-quench observables in statistical mechanics

In the study of phase coexistence, the relevant statistical observables are additive operators

$\varepsilon>0$5

whose eigenvalues scale as $\varepsilon>0$6. The problem addressed in “Statistical ensembles for phase coexistence states specified by noncommutative additive observables” is that the standard microcanonical or restricted constructions are ill-defined or ill-behaved when some additive observables do not commute. The proposed solution is the extended generalized, or squeezed, ensemble

$\varepsilon>0$7

with $\varepsilon>0$8. Under stated concavity, regularity, self-adjointness, and surjectivity conditions, the ensemble is sharply peaked on a single macroscopic value even in first-order regions and gives direct formulas for intensive parameters from the expectation values of the additive observables (Yoneta et al., 2021).

At the level of the thermodynamic limit, “Global observables in statistical mechanics” constructs macroscopic observables inside a quotient of the full product algebra by the ideal of vanishing sequences. The key subalgebra consists of sequences that asymptotically commute with every local observable; their equivalence classes form the global algebra $\varepsilon>0$9. A particularly important commutative subalgebra is generated by macroscopic averages, or $N$00-sequences. In the commutative case, the corresponding algebra is identified with bounded functions measurable with respect to the tail $N$01-algebra. In the quantum case, global observables include translated local operators carried off to infinity and macroscopic averages such as $N$02 (Ven, 30 Oct 2025).

A more dynamical many-body usage appears in quench studies of hard-core bosons in the Aubry–André potential. The one-body observables are the site density $N$03, the momentum distribution function $N$04, and the occupations of the natural orbitals $N$05. After a sudden change in the quasi-periodic potential, the delocalized regime exhibits approximate power-law relaxation of all three observables; at the critical point the relaxation is slower; in the localized regime the density saturates quickly to a nonzero offset, while the momentum distribution and natural-orbital occupations still display an approximate power-law tail. The generalized Gibbs ensemble reproduces the post-quench stationary values of these observables in the extended regime and, when relaxation occurs, at the critical point, but fails for $N$06 and $N$07 in the localized phase (Gramsch et al., 2012).

7. Cosmological and field-theoretic observables

In cosmology, the observable is a field on the sky depending on observer position, redshift, and direction. The paper “General and consistent statistics for cosmological observations” formulates reduced angular $N$08-point spectra for arbitrary $N$09 using multilateral Wigner symbols. The observational reduced spectra are explicit estimators built from the harmonic coefficients $N$10, while the theoretical spectra are their ensemble averages. Under statistical homogeneity, isotropy, and ergodicity, the estimator covariance is cosmic variance and is written in terms of the squeezed theoretical $N$11-point spectrum (Mitsou et al., 2019).

A major technical point concerns observer-position contributions. At linear order, a field evaluated at the observer contributes only to the monopole of $N$12 for $N$13, but beyond linear order cross-terms involving the observer contribution generate terms in all multipoles with magnitude of the same order as the rest of the non-linear corrections. The paper therefore concludes that discarding observer-position terms is not justified beyond linear order and that there is no inconsistency in ensemble-averaging fields at and near the observer position (Mitsou et al., 2019).

In McDearmon’s discretized Minkowski-lattice field theory, observables are complex-valued functionals $N$14 of a fluctuating complex scalar field, with inner product

$N$15

After quotienting by null-norm functionals, these form a Hilbert space of observables. A bosonic Fock space is then constructed, together with creation and annihilation operators and the field operator

$N$16

In the free-field limit, the commutator is governed by the Pauli–Jordan function, yielding microcausality for spacelike-separated supports. Here the phrase “statistical observables” is literal: every operator is induced from a real-probability ensemble of fluctuating fields, and vacuum expectations reproduce the means and variances of the underlying functionals (McDearmon, 2024).

Taken together, these usages show that “statistical observables” is best understood as a family resemblance term rather than a single definition. Across dynamics, quantum information, singular learning theory, statistical mechanics, cosmology, and field theory, the observable is the object through which statistics become canonical: empirical limits, moments, entropies, weighted estimators, observable coordinates, additive constraints, macroscopic tails, or reduced spectra. The unifying idea is not the form of the observable, but its role in organizing statistically meaningful information.