Fluctuation-Response Inequalities (FRIs)

• Fluctuation-Response Inequalities (FRIs) are bounds that limit how observable responses change under perturbations using fluctuation measures and statistical divergences.
• They generalize fluctuation-dissipation relations by incorporating information inequalities like Cramér-Rao and Chapman-Robbins across various dynamical frameworks.
• FRIs find practical applications in nonequilibrium statistical mechanics, experimental data analysis, and quantum systems, linking response, fluctuations, and entropy production.

Fluctuation-response inequalities (FRIs) are inequalities that constrain how strongly an observable can change under a perturbation in terms of fluctuation measures and statistical distinguishability costs of the underlying states, paths, or output records. In the classical out-of-equilibrium formulation, the magnitude of the response is bounded by the cumulant generating function of the observable and the Kullback-Leibler divergence between perturbed and unperturbed distributions (Dechant et al., 2018). Subsequent work recast FRIs through information inequalities such as Cramér-Rao and Chapman-Robbins (Hasegawa et al., 2018), extended them to sub-Gaussian and subexponential observables (Wang, 2020), to finite-time Markov jump and Langevin dynamics (Kwon et al., 2024, Chun et al., 23 Jan 2026), to nonlinear and finite-frequency response (Zheng et al., 23 Sep 2025, Zheng et al., 20 Feb 2026, Dechant, 17 Oct 2025), and to quantum and open-quantum systems (Wang, 2022, Gu et al., 5 May 2026). In nonequilibrium statistical mechanics, FRIs occupy the inequality layer between exact fluctuation-response relations (FRRs), fluctuation-dissipation theorems (FDTs), and uncertainty relations, thereby linking response, fluctuations, entropy production, traffic, dynamical activity, and Fisher information.

1. Foundational formulations and information-theoretic structure

The classical out-of-equilibrium FRI introduced a general inequality for two probability distributions $P^a(\omega)$ and $P^b(\omega)$ and an observable $r(\omega)$: $$\left| r^b-r^a \right| \leq \inf_{h>0}\frac{1}{h}\left( K_{\Delta r}^a(h\sigma)+D_{\mathrm{KL}}^{b\Vert a} \right),$$ where $\Delta r=r-r^a$, $K_{\Delta r}^a(h)=\ln \mathbb E_a[e^{h\Delta r}]$, $\sigma=\operatorname{sign}(r^b-r^a)$, and $D_{\mathrm{KL}}^{b\Vert a}$ is the KL divergence (Dechant et al., 2018). When $r$ is approximately Gaussian in the reference distribution, this reduces to

$$\left| r^b-r^a \right| \leq \sqrt{2D_{\mathrm{KL}}^{b\Vert a}\,\mathrm{Var}_a(r)}.$$

The same framework shows that, for small perturbations, the observable change is first order in the perturbation while the KL divergence is typically second order, which yields a linear-response FRI in terms of variance and relative entropy (Dechant et al., 2018).

A parallel line of work placed FRIs within statistical estimation theory. For stochastic processes described by Langevin equations, the Cramér-Rao inequality gives

$P^b(\omega)$0

with $P^b(\omega)$1 the Fisher information of the path measure. In that formulation, applying the Cramér-Rao inequality yields the FRI, while the thermodynamic uncertainty relation (TUR) appears as a particular case in which the Fisher information is the total entropy production (Hasegawa et al., 2018). The same analysis also derives a Chapman-Robbins version for finite perturbations in terms of the Pearson divergence between process measures, thereby extending the fluctuation-response trade-off beyond infinitesimal response (Hasegawa et al., 2018).

These formulations already clarify two structural properties that remain central throughout the literature. First, FRIs are inequalities, not identities: they bound response by fluctuation or information quantities but do not generally determine it exactly. Second, the same mathematical backbone can yield either response bounds or uncertainty relations, depending on how the perturbation and estimator are chosen. This suggests that FRIs are best understood as an information-geometric envelope around response theory rather than as a single model-specific formula.

2. Norm-based FRIs beyond Gaussian response

A major generalization replaces variance-based control by concentration norms adapted to non-Gaussian statistics. For a centered random variable $P^b(\omega)$2, the sub-Gaussian condition is

$P^b(\omega)$3

and if $P^b(\omega)$4 under $P^b(\omega)$5 is sub-Gaussian, then

$P^b(\omega)$6

For subexponential variables, defined by

$P^b(\omega)$7

one obtains the same square-root form when $P^b(\omega)$8, and a linear-plus-inverse-linear bound for larger KL divergence (Wang, 2020). The paper characterizes these inequalities as applying to arbitrary sub-Gaussian or subexponential observables, as nonperturbative, and as relying on the norm and divergence rather than on explicit Gaussianity (Wang, 2020).

These norm-based FRIs yield thermodynamic consequences. For two equilibrium states with Hamiltonians $P^b(\omega)$9 and $r(\omega)$0,

$r(\omega)$1

and the resulting entropy-energy fluctuation relation is

$r(\omega)$2

The same framework produces generalized TURs. In the sub-Gaussian regime,

$r(\omega)$3

and in the subexponential regime,

$r(\omega)$4

for moderate KL divergence, with a looser bound otherwise (Wang, 2020).

Operationally, this line of work also addresses experimental implementation. For sub-Gaussian observables, the error incurred by replacing expected values with sample means admits a nonasymptotic concentration bound: $r(\omega)$5 with probability at least $r(\omega)$6. A plug-in estimator for the sub-Gaussian norm is proposed as

$r(\omega)$7

where $r(\omega)$8 are centered data points (Wang, 2020).

3. Markov jump processes, exact FRRs, and finite-time FRIs

For continuous-time Markov jump processes, a general finite-time theory derives FRIs directly from the path-wise Cramér-Rao bound. With transition rates parameterized as

$r(\omega)$9

where $$\left| r^b-r^a \right| \leq \inf_{h>0}\frac{1}{h}\left( K_{\Delta r}^a(h\sigma)+D_{\mathrm{KL}}^{b\Vert a} \right),$$0 and $$\left| r^b-r^a \right| \leq \inf_{h>0}\frac{1}{h}\left( K_{\Delta r}^a(h\sigma)+D_{\mathrm{KL}}^{b\Vert a} \right),$$1, the response of a time-integrated observable $$\left| r^b-r^a \right| \leq \inf_{h>0}\frac{1}{h}\left( K_{\Delta r}^a(h\sigma)+D_{\mathrm{KL}}^{b\Vert a} \right),$$2 to a perturbation parameter $$\left| r^b-r^a \right| \leq \inf_{h>0}\frac{1}{h}\left( K_{\Delta r}^a(h\sigma)+D_{\mathrm{KL}}^{b\Vert a} \right),$$3 satisfies

$$\left| r^b-r^a \right| \leq \inf_{h>0}\frac{1}{h}\left( K_{\Delta r}^a(h\sigma)+D_{\mathrm{KL}}^{b\Vert a} \right),$$4

with $$\left| r^b-r^a \right| \leq \inf_{h>0}\frac{1}{h}\left( K_{\Delta r}^a(h\sigma)+D_{\mathrm{KL}}^{b\Vert a} \right),$$5 the Fisher information of the path probability. This yields the explicit FRIs

$$\left| r^b-r^a \right| \leq \inf_{h>0}\frac{1}{h}\left( K_{\Delta r}^a(h\sigma)+D_{\mathrm{KL}}^{b\Vert a} \right),$$6

where $$\left| r^b-r^a \right| \leq \inf_{h>0}\frac{1}{h}\left( K_{\Delta r}^a(h\sigma)+D_{\mathrm{KL}}^{b\Vert a} \right),$$7 is the dynamical activity, and for observables in the set $$\left| r^b-r^a \right| \leq \inf_{h>0}\frac{1}{h}\left( K_{\Delta r}^a(h\sigma)+D_{\mathrm{KL}}^{b\Vert a} \right),$$8,

$$\left| r^b-r^a \right| \leq \inf_{h>0}\frac{1}{h}\left( K_{\Delta r}^a(h\sigma)+D_{\mathrm{KL}}^{b\Vert a} \right),$$9

These inequalities are valid for finite times and for current-like, state-dependent, and mixed observables (Kwon et al., 2024).

A complementary development establishes exact FRRs in nonequilibrium steady states of Markov jump processes. For arbitrary steady-state currents $\Delta r=r-r^a$0, the covariance admits the exact representations

$\Delta r=r-r^a$1

with $\Delta r=r-r^a$2 the traffic, $\Delta r=r-r^a$3 the edge current, and $\Delta r=r-r^a$4 static responses to symmetric and antisymmetric rate perturbations, respectively (Aslyamov et al., 2024). For time-integrated state observables, structurally identical FRRs are obtained: $\Delta r=r-r^a$5 together with finite-time lower bounds that become equalities in the long-time limit (Ptaszynski et al., 2024).

From these exact FRRs one obtains a hierarchy of FRIs and uncertainty relations. For symmetric perturbations, the ratio between any current response and its variance is bounded by entropy-production-type quantities, including partial EPR and pseudo-EPR; for antisymmetric perturbations, the corresponding bound is controlled by traffic rather than EPR (Aslyamov et al., 2024). For mixed state-current covariances, exact FRRs and inverse FRRs express covariances in terms of local responses and, conversely, responses in terms of covariances. In that setting, the breaking of Onsager symmetry can occur only in the presence of state-current correlations (Ptaszynski et al., 10 Jun 2025).

The distinction between FRRs and FRIs is especially transparent in this literature. FRRs are exact equalities that express covariances through response coefficients, whereas FRIs arise by applying inequalities such as Cauchy-Schwarz or by discarding part of a mode decomposition. This suggests a structural hierarchy in which exact response-covariance identities generate lower or upper bounds once only partial response information is retained.

4. Dynamical, nonlinear, and finite-frequency extensions

For nonequilibrium Langevin dynamics, a unified fluctuation-response relation and a finite-time FRI are derived for the one-dimensional overdamped process

$\Delta r=r-r^a$6

and the general time-averaged observable

$\Delta r=r-r^a$7

The central finite-time FRI is

$\Delta r=r-r^a$8

valid at any time and for arbitrary initial conditions. In steady state this becomes a spatial integral with $\Delta r=r-r^a$9 on the left-hand side and becomes tight as $K_{\Delta r}^a(h)=\ln \mathbb E_a[e^{h\Delta r}]$0. Applying Cauchy-Schwarz yields response uncertainty relations, including the response thermodynamic uncertainty relation

$K_{\Delta r}^a(h)=\ln \mathbb E_a[e^{h\Delta r}]$1

and, for uniform perturbations and current-type observables,

$K_{\Delta r}^a(h)=\ln \mathbb E_a[e^{h\Delta r}]$2

The paper explicitly states the hierarchy $K_{\Delta r}^a(h)=\ln \mathbb E_a[e^{h\Delta r}]$3 and illustrates the resulting long-time diffusion bounds for the $K_{\Delta r}^a(h)=\ln \mathbb E_a[e^{h\Delta r}]$4-ATPase molecular motor (Chun et al., 23 Jan 2026).

FRIs have also been extended to nonlinear response. For stochastic Markov dynamics with trajectory probability $K_{\Delta r}^a(h)=\ln \mathbb E_a[e^{h\Delta r}]$5, the $K_{\Delta r}^a(h)=\ln \mathbb E_a[e^{h\Delta r}]$6-th order response is written as

$K_{\Delta r}^a(h)=\ln \mathbb E_a[e^{h\Delta r}]$7

where

$K_{\Delta r}^a(h)=\ln \mathbb E_a[e^{h\Delta r}]$8

and $K_{\Delta r}^a(h)=\ln \mathbb E_a[e^{h\Delta r}]$9 has the complete Bell polynomial form in the score function and its derivatives. The corresponding nonlinear FRI is

$\sigma=\operatorname{sign}(r^b-r^a)$0

This yields higher-order response uncertainty relations and generalizes the linear FRI recovered at $\sigma=\operatorname{sign}(r^b-r^a)$1 (Zheng et al., 23 Sep 2025).

A distinct extension moves to the frequency domain. For steady-state Markov processes with time-dependent perturbations, a general matrix inequality takes the form

$\sigma=\operatorname{sign}(r^b-r^a)$2

and for barrier and entropic perturbations one obtains

$\sigma=\operatorname{sign}(r^b-r^a)$3

For state-current observables, the spectral signal-to-noise ratio is additionally bounded by the entropy production rate (Zheng et al., 20 Feb 2026). In an even broader Markovian setting covering over- and underdamped Langevin systems and jump processes, the finite-frequency FRI is

$\sigma=\operatorname{sign}(r^b-r^a)$4

with scalar form

$\sigma=\operatorname{sign}(r^b-r^a)$5

and the integrated broad-band SNR satisfies

$\sigma=\operatorname{sign}(r^b-r^a)$6

That bound becomes an equality for appropriately chosen observables or perturbations in linear systems, both overdamped and underdamped and both in and out of equilibrium (Dechant, 17 Oct 2025). For nonautonomous Markov jump processes, exact dynamical FRRs decompose finite-time covariance into an initial-variability term plus response-kernel integrals, and known autonomous FRIs are identified as the zero-frequency mode (Aslyamov et al., 27 Apr 2026).

5. Quantum FRIs and open-system generalizations

The quantum fluctuation-response inequality (QFRI) bounds the mean difference of an observable between two quantum states in terms of quantum relative entropy. For density operators $\sigma=\operatorname{sign}(r^b-r^a)$7 and observable $\sigma=\operatorname{sign}(r^b-r^a)$8,

$\sigma=\operatorname{sign}(r^b-r^a)$9

When the spectrum of $D_{\mathrm{KL}}^{b\Vert a}$0 is bounded, the sub-Gaussian property yields the explicit bound

$D_{\mathrm{KL}}^{b\Vert a}$1

with $D_{\mathrm{KL}}^{b\Vert a}$2 the sub-Gaussian norm of the centered observable under $D_{\mathrm{KL}}^{b\Vert a}$3. For observables with spectrum in $D_{\mathrm{KL}}^{b\Vert a}$4, the norm satisfies $D_{\mathrm{KL}}^{b\Vert a}$5 (Wang, 2022).

This QFRI has several stated applications. In quantum hypothesis testing, it yields the nonasymptotic bound

$D_{\mathrm{KL}}^{b\Vert a}$6

which is described as stronger and more informative than the bound based on quantum Pinsker’s inequality, while also being measurement-dependent through $D_{\mathrm{KL}}^{b\Vert a}$7 (Wang, 2022). The same paper applies QFRI to thermodynamic inference and to quantum speed limits of the form

$D_{\mathrm{KL}}^{b\Vert a}$8

Open-system and trajectory-level extensions bring dynamical activity to the forefront. For Lindblad dynamics

$D_{\mathrm{KL}}^{b\Vert a}$9

with $r$0, the quantum FRI takes the form

$r$1

where

$r$2

is the quantum analog of dynamical activity through channel $r$3. In this formulation, dynamical activity is the central kinetic quantity, and the bound does not involve entropy production (Kwon et al., 2024).

A further finite-frequency open-quantum formulation is developed in an input-output setting. For any downstream measurement of the emitted field, the measured response-to-noise matrix satisfies

$r$4

Here the left side is detector-facing, while the intermediate ceiling is the output-field quantum Fisher information rate and the final ceiling is a signal-activity matrix. For kinetic modulation, the activity reduces to stationary channel fluxes. The paper emphasizes that the result is detector-facing but unraveling-independent (Gu et al., 5 May 2026).

6. Applications, interpretation, and recurrent misconceptions

FRIs have immediate applied consequences in transport, inference, and spectroscopy. In steady-state particle transport, the original classical FRI yields a bound of differential mobility by diffusivity: $r$5 and, under a virtual perturbation proportional to the local mean velocity, recovers the steady-state TUR

$r$6

for time-integrated currents (Dechant et al., 2018). In nonequilibrium Langevin dynamics, response-based bounds constrain the long-time diffusion coefficient of the $r$7-ATPase molecular motor, while finite-frequency Markov-process FRIs provide a route to infer entropy production rate from power spectrum measurements (Chun et al., 23 Jan 2026, Zheng et al., 20 Feb 2026).

In network problems, exact FRRs and the associated FRIs simplify fluctuation calculations in large Markov networks and provide mechanistic interpretations of correlations. For state observables, the formalism is used to explain positive and negative correlations of occupation times in a double quantum dot device (Ptaszynski et al., 2024). For mixed state-current covariances, FRRs are used to explain fluctuations in quantum dot devices and enzymatic reaction schemes and to discuss their potential relevance for model inference (Ptaszynski et al., 10 Jun 2025). These applications rely on the fact that the response decomposition is local in edge space even when the observable is global.

Experimental implementation and data analysis are recurring themes rather than afterthoughts. The sub-Gaussian and subexponential framework supplies nonasymptotic sample-mean error bounds and a plug-in estimator for the relevant norm, explicitly addressing finite-data use of FRIs (Wang, 2020). Spectral formulations are expressed directly in terms of measurable susceptibilities, power spectra, or output-current covariances (Zheng et al., 20 Feb 2026, Gu et al., 5 May 2026). This suggests that modern FRIs are increasingly formulated in detector-level variables rather than only in idealized ensemble averages.

Several misconceptions recur in the literature. One is to identify FRIs with exact fluctuation-dissipation-type equalities. In fact, the exact equalities are FRRs, whereas FRIs are bounds that remain valid for broader observable classes, finite times, or reduced response information (Kwon et al., 2024, Aslyamov et al., 27 Apr 2026). Another is to treat FRIs as inherently Gaussian or near-equilibrium statements. The sub-Gaussian, subexponential, nonlinear Bell-polynomial, and finite-frequency theories explicitly move beyond that restriction (Wang, 2020, Zheng et al., 23 Sep 2025, Dechant, 17 Oct 2025). A third concerns saturation. In one information-inequality analysis, stochastic total entropy production is the only quantity that can attain equality in the TUR (Hasegawa et al., 2018), whereas in finite-frequency Markovian dynamics equality can occur for appropriately chosen observables or perturbations in linear systems (Dechant, 17 Oct 2025). Saturation is therefore highly structure-dependent rather than generic.

Taken together, these results indicate that FRIs are not a single inequality but a family of bounds organized by the choice of observable class, perturbation class, distinguishability measure, and dynamical level—state, path, spectrum, or quantum output field. That family now spans KL-divergence bounds, Fisher-information bounds, activity- and traffic-controlled bounds, norm-based non-Gaussian bounds, and frequency-resolved matrix inequalities, all serving the same core purpose: to quantify the maximum admissible response compatible with fluctuations and nonequilibrium structure.