Stochastic Resonance in Nonlinear Systems

• Stochastic resonance is a phenomenon where optimal noise levels enable weak, subthreshold signals to be effectively transmitted in nonlinear systems.
• It operates via noise-induced state transitions that achieve phase locking and synchronization in bistable, multistable, and non-equilibrium environments.
• Applications of stochastic resonance span neural circuits, optomechanics, biophysics, and electronic devices, offering strategies for improved signal processing.

Stochastic resonance (SR) is a noise-assisted response phenomenon in which a weak input is most effectively transmitted, detected, or synchronized at an intermediate fluctuation level rather than in the zero-noise limit. In its canonical form, SR involves a nonlinear element with thresholds or metastability, a subthreshold periodic forcing, and a stochastic drive that enables transitions to lock to the forcing timescale. Contemporary work treats SR not as a single mechanism confined to double-well models, but as a broader class of resonance phenomena spanning bistable, multistable, excitable, chaotic, and non-equilibrium systems, with realizations in neural circuits, optomechanics, trapped ions, ferroelectrics, molecular biophysics, electronic devices, and learning architectures (Castellanos, 2018, Lucarini, 2019, Wei et al., 18 May 2026, Dey et al., 16 Aug 2025).

1. Canonical formulation and physical intuition

In the standard description, SR arises when three ingredients co-occur: a bistable system or thresholded excitable element, a weak periodic input that is subthreshold, and a stochastic drive of tunable intensity. The archetypal Langevin model writes a state variable in a bistable potential as

$$m\frac{d^2 x}{dt^2} = F(t) - b\,\frac{dx}{dt} - \frac{dV(x)}{dx} + \xi(t),$$

with a quartic potential

$V(x) = -\frac{a}{2}\,x^2 + \frac{b}{2}\,x^4,$

for which the minima occur at $x=\pm \sqrt{a/b}$ and the barrier height is $\Delta V = a^2/(4b)$. The periodic drive is typically $F(t)=A\sin(\Omega t)$, and the stochastic term $\xi(t)$ provides the barrier-crossing events that the deterministic forcing alone cannot induce (Castellanos, 2018).

The intuitive picture is rate matching. If noise is too weak, transitions are rare and the forcing is not expressed in the output. If noise is too strong, crossings become frequent and desynchronized. At an intermediate level, the output becomes phase-locked to the forcing. In weak-noise theory this is summarized by a Kramers-type escape rate,

$$r_K(D) \propto \exp\!\left(-\frac{\Delta V}{D}\right),$$

with optimal resonance often associated with

$r_K(D^*) \approx \frac{\Omega}{\pi},$

so that switching occurs approximately once per half-cycle (Castellanos, 2018).

A closely related canonical model is the overdamped particle in a double-well potential,

$\dot{x}(t) = -U'(x(t)) + s(t) + \sqrt{2D}\,\xi(t),$

with

$$U(x) = \frac{\beta}{4}\,x^4 - \frac{\alpha}{2}\,x^2,$$

which makes explicit that SR is fundamentally a competition between deterministic confinement, weak forcing, and noise-driven escape (Krauss et al., 2015). In a general non-equilibrium setting, this static-potential picture is replaced by quasi-potentials and instantons. The invariant density takes the large-deviation form

$V(x) = -\frac{a}{2}\,x^2 + \frac{b}{2}\,x^4,$0

where $V(x) = -\frac{a}{2}\,x^2 + \frac{b}{2}\,x^4,$1 is the quasi-potential, and noise-activated transitions are controlled by quasi-potential barriers between metastable attractors and edge states (Lucarini, 2019).

2. Quantifiers, observables, and resonance criteria

Although SR is often introduced through signal-to-noise ratio, the literature uses several inequivalent observables depending on the system and task. In periodically driven systems, the canonical spectral measure isolates power at the forcing frequency. For single-molecule hairpins, the output signal and background noise are defined from the extension power spectral density $V(x) = -\frac{a}{2}\,x^2 + \frac{b}{2}\,x^4,$2 as

$V(x) = -\frac{a}{2}\,x^2 + \frac{b}{2}\,x^4,$3

$V(x) = -\frac{a}{2}\,x^2 + \frac{b}{2}\,x^4,$4

and

$V(x) = -\frac{a}{2}\,x^2 + \frac{b}{2}\,x^4,$5

with SR identified by a peak in $V(x) = -\frac{a}{2}\,x^2 + \frac{b}{2}\,x^4,$6 versus forcing frequency or noise level (Hayashi et al., 2012). In hardware neuron and Schmitt-trigger experiments, comparable spectral criteria are used at the drive frequency, with the SNR maximum operationally defining the optimal noise (Castellanos, 2018, Kim et al., 5 Jan 2025).

Other studies replace spectral criteria by task-specific performance measures. In Neurochaos Learning, SR is identified by a peak in the cross-correlation coefficient

$V(x) = -\frac{a}{2}\,x^2 + \frac{b}{2}\,x^4,$7

for subthreshold signal detection, and by a peak in macro F1-score for classification as the input noise amplitude $V(x) = -\frac{a}{2}\,x^2 + \frac{b}{2}\,x^4,$8 is swept (NB et al., 2021). In recurrent three-neuron motifs, the relevant quantity is not SNR but mutual information between successive network states,

$V(x) = -\frac{a}{2}\,x^2 + \frac{b}{2}\,x^4,$9

which can peak at finite input noise when the gain in state-space exploration exceeds the increase in stochastic information loss (Krauss et al., 2018).

A distinct line of work argues that adaptive SR should not require access to the unknown input. In that framework, the detector output autocorrelation

$x=\pm \sqrt{a/b}$0

or its normalized form serves as an input-free optimization objective. For a symmetric binary detector, the one-lag output autocorrelation becomes

$x=\pm \sqrt{a/b}$1

with $x=\pm \sqrt{a/b}$2 the success probability of correct threshold decisions; since mutual information and input–output correlation are monotone in $x=\pm \sqrt{a/b}$3 in that model, the optimal noise level inferred from output autocorrelation coincides with the optimum from mutual information or cross-correlation (Krauss et al., 2015).

Threshold-based statistical formulations provide yet another viewpoint. In a double-threshold detector with noisy observations $x=\pm \sqrt{a/b}$4, the retained Fisher Information under thresholding can itself show an SR-like peak as the noise standard deviation varies, and average mean squared error becomes the practical tuning objective. In the multiscale wavelet formulation, the signal estimate is reconstructed from threshold exceedance probabilities after discrete wavelet decomposition, with the paper reporting that optimal noise levels decrease and AMSE can be strongly reduced in the multiscale domain (Vimalajeewa et al., 9 Oct 2025).

3. Beyond the static double-well: periodic, multistable, and autonomous SR

A persistent misconception is that SR is restricted to static bistable potentials. Several of the cited studies directly contest that restriction. In underdamped periodic potentials, SR was numerically demonstrated in the high-frequency regime through a peak in hysteresis loop area or input energy loss per cycle. The key mechanism is the existence of two dynamical states of the driven trajectory—an in-phase state and an out-of-phase state—between which noise induces switching (Saikia et al., 2010). The later revisit extended this picture to the periodic bistable potential

$x=\pm \sqrt{a/b}$5

and further to tilted washboard potentials after correcting the hysteresis loop area for the average drift term (Reenbohn et al., 2012).

Optomechanical systems generalize SR from bistability to multistability and mixed forcing channels. In a membrane-in-the-middle system with quadratic optomechanical coupling, the effective mechanical potential becomes triple-welled, and stochastic switching among three stable states can be synchronized either by an additive mechanical force or by a multiplicative optical modulation. The two channels produce qualitatively different SR patterns, and when both are present the response can be enhanced; if their frequencies are slightly detuned, a beating envelope appears in the synchronized switching (Fan et al., 2016). In a bistable optomechanical system driven by both optical and mechanical weak signals, constructive interference between the two SR channels and a beating-induced robustness to initial phase mismatch were likewise identified, and a quantum-stochastic-resonance regime induced solely by quantum fluctuations was numerically observed with quantum trajectories (Xie et al., 2018).

Autonomous SR in non-equilibrium systems pushes the concept further. A general large-deviation framework based on quasi-potentials shows that periodically modulated escape rates in noisy non-equilibrium systems can still be reduced to a two-state approximation, with the forcing effect encoded by the first-order quasi-potential correction $x=\pm \sqrt{a/b}$6 obtained from

$x=\pm \sqrt{a/b}$7

In that framework, SR is amplified or suppressed according to the barrier-modulation asymmetry between the two metastable states, and the forcing is relevant only through its projection on $x=\pm \sqrt{a/b}$8 (Lucarini, 2019).

A fluid-dynamical realization of this broader autonomous picture was reported in inertia-less viscoelastic channel flow just above a pure elastic instability. There, SR was associated with three ingredients: a chaotic streamwise velocity spectrum, a white-noise spanwise spectrum, and weak elastic waves. SR existed only in a lower sub-region $x=\pm \sqrt{a/b}$9, where a low-frequency high-energy peak in the streamwise velocity spectrum was observed; it disappeared when the spanwise spectrum itself became chaotic at larger $\Delta V = a^2/(4b)$0 (Li et al., 2023).

4. Noise color, correlation structure, and engineered fluctuations

The stochastic ingredient in SR is not exhausted by a scalar noise intensity. Its spectral content, correlation time, and quadrature structure can qualitatively reshape resonance. In a hardware artificial neuron implemented with a monostable Schmitt trigger, white and pink noise were compared under identical subthreshold sinusoidal drive. For an input amplitude of about $\Delta V = a^2/(4b)$1, pink noise yielded a maximum output SNR about 20 times larger than white noise, and the SNR-versus-noise curve was broader, indicating greater robustness to mistuning (Castellanos, 2018). This result was interpreted as consistent with the ubiquity of $\Delta V = a^2/(4b)$2-like backgrounds in neural systems.

Colored noise can also suppress SR. In globally coupled overdamped bistable oscillators with both pairwise and 2-simplex coupling, Ornstein–Uhlenbeck colored noise reduced the resonance peak and shifted the optimal noise intensity to higher values relative to white noise. Higher-order interactions did not reverse that trend; rather, they further exacerbated the suppression and were linked to a four-stage evolution of network synchronization as the noise intensity increased (Bi et al., 6 Mar 2026).

A different strategy is to reshape intrinsic fluctuations instead of injecting auxiliary noise. In a trapped-ion Duffing oscillator, classical quadrature squeezing was used to squeeze phase noise and anti-squeeze amplitude noise. Because switching occurs along the amplitude coordinate near the bistability point, this redistribution enhanced SR without an added external noise source. Under identical electric-field detection conditions, the squeezing-induced protocol achieved a maximum SNR of 19.23 dB at $\Delta V = a^2/(4b)$3, compared with 14.77 dB for conventional noise-induced SR, corresponding to an improvement of $\Delta V = a^2/(4b)$4 dB; the minimum detectable field at SNR = 0 dB improved from $\Delta V = a^2/(4b)$5 to $\Delta V = a^2/(4b)$6 (Wei et al., 18 May 2026).

These comparisons also delimit a conceptual boundary. Some work on high-frequency transcranial stimulation argues that resonance-like enhancement can occur with deterministic carriers rather than stochastic ones, under the label deterministic amplitude resonance. In that framework, threshold nonlinearities plus low-pass filtering can restore subthreshold signals using high-frequency triangle or sine carriers, formally yielding infinite output SNR in the absence of stochastic noise. This does not refute classical SR, but it does show that noise is not the only route to resonance-like gain in nonlinear systems (Potok et al., 2023).

5. Experimental realizations across physical and engineered systems

SR has been demonstrated in a wide range of experimental platforms, often with system-specific observables and constraints. In single-molecule biophysics, optical-tweezers experiments on DNA hairpins established SR through the folding–unfolding dynamics of a periodically driven bistable free-energy landscape. For a 20 bp hairpin at coexistence, the SNR of extension fluctuations peaked at $\Delta V = a^2/(4b)$7 Hz, close to the matching estimate $\Delta V = a^2/(4b)$8 Hz, and analogous peaks were used to infer kinetics in shorter hairpins with poorer direct signal quality (Hayashi et al., 2012).

In nanoelectronics, a single-electron turnstile in the Coulomb-blockade regime was mapped onto the McNamara–Wiesenfeld adiabatic two-state framework. Monte Carlo simulations showed the characteristic nonmonotonic SNR-versus-noise curve under Gaussian colored gate noise, indicating that SR can enhance the detectability of weak periodic charge motion (0710.2718). In nonlinear analog electronics, both a unijunction transistor relaxation oscillator biased near homoclinic bifurcation and an op-amp Schmitt trigger produced SR under weak periodic drive plus injected Gaussian noise; in both cases, normalized variance or SNR exhibited clear optima as the noise amplitude was tuned (Bose et al., 2014, Kim et al., 5 Jan 2025).

Optomechanics and trapped ions provide mechanically resolved realizations. In bistable and tristable optomechanical systems, SR has been analyzed through stochastic switching of the mechanical coordinate under optical and mechanical weak signals, with interference, beating, and multistable switching patterns available depending on the architecture (Fan et al., 2016, Xie et al., 2018). In the trapped-ion Duffing oscillator, fluorescence readout of bistable motional states enabled direct field sensing through SR, with a 60 s trace sampled at 50 Hz and a defined spectral SNR metric at the 1 Hz target frequency (Wei et al., 18 May 2026).

Ferroelectric thin-film capacitors extend SR into nonvolatile switching. In PbZr$\Delta V = a^2/(4b)$9Ti$F(t)=A\sin(\Omega t)$0O$F(t)=A\sin(\Omega t)$1 capacitors driven by sub-coercive voltages, synchronous polarization switching was observed at optimal injected noise, and the Kramers time was directly measured by a first-passage protocol. For one sample at $F(t)=A\sin(\Omega t)$2 Hz, the figures of merit peaked at $F(t)=A\sin(\Omega t)$3 V, with $F(t)=A\sin(\Omega t)$4 ms, in agreement with the SR condition $F(t)=A\sin(\Omega t)$5 (Dey et al., 16 Aug 2025).

6. Neural and information-processing perspectives, limitations, and open problems

Neural and brain-inspired systems have become a major domain for reframing SR as an information-processing resource. A single artificial neuron implemented as a Schmitt-trigger circuit exhibited SR under subthreshold periodic forcing, and pink noise outperformed white noise by a large margin in the measured output SNR (Castellanos, 2018). In three-neuron recurrent motifs with Boltzmann updates, added Gaussian input noise could maximize the mutual information between successive network states, especially in strongly connected motifs whose noiseless dynamics were trapped in a small repertoire of states (Krauss et al., 2018). In Neurochaos Learning, intermediate noise improved both subthreshold signal detection and classification, with the reported maxima occurring at $F(t)=A\sin(\Omega t)$6 and $F(t)=A\sin(\Omega t)$7 for single-neuron detection, at average macro F1 = 0.907 on CCD at $F(t)=A\sin(\Omega t)$8, and at average macro F1 = 0.911 on spoken digits at $F(t)=A\sin(\Omega t)$9 (NB et al., 2021).

This suggests a shift from the classical “detector-plus-noise” picture toward a broader view in which SR can optimize feature richness, state-space exploration, and temporal coherence. In perceptual neuroscience, high-frequency deterministic tACS with triangle or sine waveforms reduced visual contrast detection thresholds with effect sizes comparable to high-frequency tRNS, which the authors interpreted as evidence that resonance-like enhancement need not rely on stochastic carriers alone (Potok et al., 2023). That claim should be read carefully: it broadens the taxonomy of resonance phenomena, but it also complicates the boundary between SR proper and deterministic amplitude resonance.

Several limitations recur across the literature. Exact reproducibility is often constrained by incomplete reporting of thresholds, transfer functions, or acquisition parameters; hardware studies may identify optimal noise empirically without directly validating Kramers prefactors or transition-state structure (Castellanos, 2018). Multistable and non-equilibrium systems can be reduced to two-state approximations only under regime-specific assumptions, and the relevant escape objects may be chaotic attractors or edge states rather than simple fixed points (Lucarini, 2019). In higher-order networks, colored-noise effects remain primarily numerical, with analytic escape-rate theory still underdeveloped (Bi et al., 6 Mar 2026). In trapped-ion sensing, squeezing-induced SR improves performance, but mapping quadrature-shaped thermal noise onto an effective Kramers description is still a theoretical program rather than a closed solution (Wei et al., 18 May 2026).

The most stable conclusion across these diverse settings is therefore narrow but durable: SR is not simply “noise helps.” It is the existence of an optimal fluctuation regime—defined relative to nonlinear state structure, forcing timescale, and measurement observable—that makes weak signals, switching events, or information transfer most coherent. What counts as the relevant state structure may be a double well, a threshold, a dynamical attractor pair, a tristable landscape, or a quasi-potential barrier in a non-equilibrium flow. What counts as the relevant observable may be SNR, hysteresis-loop area, residence-time synchronization, mutual information, autocorrelation, Fisher Information, or task-level F1-score. The common structure is intermediate-noise optimality, but the mechanism and the metric are system-dependent.