Enhanced Noise Sensitivity: Mechanisms & Applications

• Enhanced noise sensitivity is a phenomenon where microscopic fluctuations trigger disproportionately large responses in systems with thresholds, bistability, or nonlinear dynamics.
• In sensing applications, noise can enable barrier crossing and state switching in bistable and exceptional-point sensors, thereby improving detection speed and measurement resolution.
• Mathematical analyses reveal that enhanced noise sensitivity leads to asymptotic independence and sharp finite-size scaling in critical phenomena, underpinning novel statistical behaviors.

Searching arXiv for the listed papers and closely related work to ground the article in current literature. arXiv search query: "Enhanced noise sensitivity arXiv (Choudhary et al., 2013, Caravenna et al., 14 Jul 2025) exceptional point sensing noise sensitivity stochastic resonance" Enhanced noise sensitivity denotes a class of phenomena in which fluctuations have an anomalously strong operational effect on a system’s dynamics, observables, or limiting dependence structure. In thresholded nonlinear networks, local noise can revive activity that deterministic dynamics would extinguish; in stochastic-resonance and exceptional-point sensors, noise can either activate a useful response or constrain any metrological gain; in quantum interferometric and spin-based measurements, the relevant enhancement may arise from lowering the noise floor or protecting the signal-bearing degree of freedom rather than increasing the signal amplitude itself; and in probability theory, “enhanced noise sensitivity” has a precise meaning stronger than decorrelation, namely asymptotic independence under small input resampling (Choudhary et al., 2013, Zhang et al., 2018, Caravenna et al., 14 Jul 2025).

1. Conceptual scope and recurring mechanisms

Taken together, the literature uses the term in more than one technical sense. In nonlinear dynamics, it refers to unusually strong responsiveness to local fluctuations once thresholds, bistability, or absorbing regions are present. In sensing, it typically concerns how perturbation responsivity, measurement noise, and operating-point selection interact. In probability, it denotes a strengthened limit theorem for functions of many independent random variables.

Setting	Mechanism	Operational consequence
Thresholded nonlinear dynamics	Noise makes an absorbing threshold partially permeable	Revived activity, sharp transitions, intermediate-noise coherence
Bistable and stochastic-resonant sensors	Noise-assisted barrier crossing or switching statistics	Faster detection and often a finite-noise optimum
Exceptional-point sensing	Root-law spectral response near non-Hermitian degeneracies	Large response with nontrivial SNR and robustness trade-offs
Quantum and interferometric sensing	Noise-floor suppression, squeezing, or error correction	Improved detectability without necessarily increasing peak amplitude
General probability theory	Small resampling plus vanishing influence sum	Asymptotic independence rather than mere decorrelation

A recurring structural feature is the presence of a map from microscopic fluctuations to macroscopic observables that is singular, thresholded, or otherwise highly nonlinear. This suggests that enhanced noise sensitivity is not a single mechanism but a family of fluctuation-amplification regimes whose interpretation depends on whether the quantity of interest is activity, switching rate, eigenvalue splitting, spectral contrast, or asymptotic dependence.

2. Thresholded nonlinear networks and noise-enhanced activity

A prototypical dynamical realization appears in the generalized population network studied in “Noise Enhanced Activity in a Complex Network” (Choudhary et al., 2013). The state of node $i$ evolves as

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$

with asymmetric positive and negative couplings $J_{ij}$, link probability $C$, and a thresholded Ricker-type local map

$f(x)= \begin{cases} x e^{r(1-x)}, & x>x_{\text{threshold}},\[4pt] 0, & \text{otherwise}. \end{cases}$

The threshold $x_{\text{threshold}}$ acts as a cessation or extinction threshold: once a population falls below it, the node becomes inactive.

In the deterministic system, many nodes fluctuate below threshold early on and remain inactive, so the asymptotic active sub-network is much smaller than the full network. The reported deterministic active-set size is essentially non-extensive: it does not grow proportionally with network size. The threshold therefore functions as an absorbing boundary in the effective state-space picture (Choudhary et al., 2013).

With additive Gaussian white noise,

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right]+\xi_i(t),\qquad \langle \xi_i(t)\xi_j(t')\rangle=\eta\,\delta(t-t')\delta_{ij},$$

the average number of active nodes, $\langle N_{\text{active}}\rangle$, undergoes a sharp increase once the noise strength exceeds a very small critical value,

$\eta_c \approx 3\times 10^{-5}\pm 10^{-6}.$

Below $\eta_c$, noise has almost no visible effect; above $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$0, the active-node count jumps sharply upward; for sufficiently large noise, $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$1 saturates. Near the transition, the paper reports the finite-size scaling form

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$2

with fitted exponents

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$3

The mechanistic interpretation is that noise makes the extinction interval $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$4 partially permeable. The fraction of time spent in that interval, $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$5, drops sharply with increasing noise, and even a subset of noisy nodes can influence the whole network: when only a fraction $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$6 of nodes is noisy, the asymptotic active-node count scales approximately as

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$7

The same work also identifies “noise induced temporal coherence” via the synchronization metric

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$8

finding that synchrony improves at an intermediate noise strength. Qualitatively similar behavior persists for uniform noise, multiplicative noise, and an alternative thresholded modified logistic map; for uniform additive noise in $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$9, the critical value is reported as $J_{ij}$0, whereas for multiplicative noise $J_{ij}$1 tends toward zero (Choudhary et al., 2013).

3. Stochastic resonance, bistability, and noise-assisted sensing

In bistable sensors, enhanced noise sensitivity often means that noise is the enabling mechanism for converting a weak perturbation into a measurable asymmetry. In “Enhancing the speed and sensitivity of a nonlinear optical sensor with noise” (Rodriguez, 2019), the sensing element is a driven single-mode Fabry–Perot cavity with Kerr-type nonlinearity $J_{ij}$2, governed by

$J_{ij}$3

In the bistable regime, fluctuations induce random switching between two stable transmission states. The observable is the residence-time difference

$J_{ij}$4

which responds to a detuning perturbation $J_{ij}$5. The reported sensitivity is

$J_{ij}$6

and a detectability criterion is imposed through

$J_{ij}$7

Detection speed increases monotonically with the noise standard deviation $J_{ij}$8, because switching becomes more frequent, but the sensitivity peaks at a finite value, near

$J_{ij}$9

with the minimum dissipation-induced noise floor itself given by $C$0. For representative semiconductor-cavity parameters, $C$1 can be detected using roughly $C$2 residence events, corresponding to about $C$3 ns (Rodriguez, 2019).

A closely related mechanism appears in the thermal Rydberg-ensemble microwave sensor of “Nonlinearity-Enhanced Continuous Microwave Detection Based on Stochastic Resonance” (Wu et al., 2024). There the order parameter $C$4 obeys an overdamped stochastic dynamics

$C$5

with a weak periodic signal $C$6 and a double-well-like effective potential generated by strong many-body Rydberg interactions. The sensor is explicitly subthreshold without noise; at an optimal $C$7, switching between “On” and “Off” states becomes synchronized with the signal. The paper reports signal enhancement of over $C$8 dB and sensitivity improvement of over $C$9 dB compared with a heterodyne atomic sensor, with a threshold near $f(x)= \begin{cases} x e^{r(1-x)}, & x>x_{\text{threshold}},\[4pt] 0, & \text{otherwise}. \end{cases}$0 V for the $f(x)= \begin{cases} x e^{r(1-x)}, & x>x_{\text{threshold}},\[4pt] 0, & \text{otherwise}. \end{cases}$1 Hz modulation case (Wu et al., 2024).

Across these examples, the response is nonmonotonic in noise strength. This suggests that in bistable sensing, “more noise” and “better sensitivity” are not equivalent: the useful regime is one in which barrier crossing is activated without erasing the perturbation-induced asymmetry that carries the signal.

4. Exceptional points, degeneracies, and the signal–noise trade-off

Exceptional-point sensing provides a different meaning of enhanced noise sensitivity: very small perturbations can produce anomalously large spectral responses, but the same non-Hermitian structure may also amplify noise. “Quantum Noise Theory of Exceptional Point Sensors” establishes the basic issue for bosonic EP sensors (Zhang et al., 2018). For a second-order EP, a perturbation $f(x)= \begin{cases} x e^{r(1-x)}, & x>x_{\text{threshold}},\[4pt] 0, & \text{otherwise}. \end{cases}$2 yields a response scaling as

$f(x)= \begin{cases} x e^{r(1-x)}, & x>x_{\text{threshold}},\[4pt] 0, & \text{otherwise}. \end{cases}$3

while the response matrix near the EP produces signal-amplitude enhancement scaling as $f(x)= \begin{cases} x e^{r(1-x)}, & x>x_{\text{threshold}},\[4pt] 0, & \text{otherwise}. \end{cases}$4 and covariance growth with $f(x)= \begin{cases} x e^{r(1-x)}, & x>x_{\text{threshold}},\[4pt] 0, & \text{otherwise}. \end{cases}$5. The paper shows, however, that the quantum Fisher information still scales as

$f(x)= \begin{cases} x e^{r(1-x)}, & x>x_{\text{threshold}},\[4pt] 0, & \text{otherwise}. \end{cases}$6

so the sensitivity bound becomes

$f(x)= \begin{cases} x e^{r(1-x)}, & x>x_{\text{threshold}},\[4pt] 0, & \text{otherwise}. \end{cases}$7

An explicit heterodyne protocol attains the same scaling, demonstrating that signal amplification and noise amplification do not cancel identically (Zhang et al., 2018).

Subsequent work has focused on how to retain root-law responsivity while avoiding the most pathological noise growth. In the $f(x)= \begin{cases} x e^{r(1-x)}, & x>x_{\text{threshold}},\[4pt] 0, & \text{otherwise}. \end{cases}$8-symmetric electromechanical accelerometer of “Enhanced Signal-to-Noise Performance of EP-based Electromechanical Accelerometers”, the observable is not the isolated-system eigenfrequency but the transmission-peak spectrum of an open system (Kononchuk et al., 2022). The transmission peak degeneracy occurs at

$f(x)= \begin{cases} x e^{r(1-x)}, & x>x_{\text{threshold}},\[4pt] 0, & \text{otherwise}. \end{cases}$9

and near the TPD,

$x_{\text{threshold}}$0

Because the bi-orthogonal basis remains complete at the TPD, the Petermann-factor divergence associated with eigenbasis collapse is avoided; the linewidth remains finite,

$x_{\text{threshold}}$1

Experimentally, the noise-equivalent acceleration improves from

$x_{\text{threshold}}$2

at $x_{\text{threshold}}$3 g to

$x_{\text{threshold}}$4

at $x_{\text{threshold}}$5 g, corresponding to a three-fold SNR improvement (Kononchuk et al., 2022).

A nonlinear route is developed in “Noise Resilient Exceptional-Point Voltmeters based on Neuromorphic functionalities” (Suntharalingam et al., 2023). There a self-oscillating nonlinear electronic dimer supports a nonlinear exceptional-point degeneracy between amplitude-death and oscillation-death steady states. The measured splitting obeys

$x_{\text{threshold}}$6

and the experiment reports two-orders signal-to-noise enhancement near the NLEPD. Allan deviation in raw frequency grows approaching the degeneracy, but the normalized voltage-equivalent noise

$x_{\text{threshold}}$7

decreases because sensitivity grows faster than the measured fluctuations (Suntharalingam et al., 2023).

The favorable interpretation is not universal. “Noise Constraints for Nonlinear Exceptional Point Sensing” shows that once noise is treated self-consistently in a nonlinear two-cavity model, it shifts the EP, lowers its effective order, and destroys the expected SNR divergence (Zheng et al., 2024). The noise-free third-order scaling is replaced by effective square-root behavior, the dominant peak frequency scales approximately as $x_{\text{threshold}}$8, and the standard Petermann-factor estimate from the instantaneous $x_{\text{threshold}}$9 Hamiltonian underestimates the actual fluctuations. The correct fluctuation theory is governed by a Bogoliubov–de Gennes Hamiltonian, yielding

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right]+\xi_i(t),\qquad \langle \xi_i(t)\xi_j(t')\rangle=\eta\,\delta(t-t')\delta_{ij},$$0

in quantitative agreement with stochastic simulations (Zheng et al., 2024).

Recent work extends both sides of this debate. A closed four-mode superconducting simulator realizes a $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right]+\xi_i(t),\qquad \langle \xi_i(t)\xi_j(t')\rangle=\eta\,\delta(t-t')\delta_{ij},$$1 dimer without engineered dissipation and observes enhanced sensitivity near the EP, but explicitly in a classical-noise-dominated rather than quantum-limited regime (Assouly et al., 14 Jun 2026). A four-channel dissipative-coupling model supports fourth-order exceptional surfaces with

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right]+\xi_i(t),\qquad \langle \xi_i(t)\xi_j(t')\rangle=\eta\,\delta(t-t')\delta_{ij},$$2

giving eigenvalue splittings about one order of magnitude larger than second-order EPs and sensitivity-factor enhancement reaching hundreds of times, but at the price of significantly larger relative error under unwanted longer-range couplings (Zhang et al., 9 Jun 2026). The consistent lesson is that exceptional points enhance response, not automatically precision.

5. Noise-floor engineering in interferometric and polarimetric measurement

A distinct strand of work enhances sensitivity by redesigning the transduction of fluctuations into detector outputs. In “Enhanced Frequency Noise Discrimination Using Cavity-coupled Mach-Zehnder Interferometer”, the discriminator output is

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right]+\xi_i(t),\qquad \langle \xi_i(t)\xi_j(t')\rangle=\eta\,\delta(t-t')\delta_{ij},$$3

so the steep phase response of a resonant ring cavity is converted into a balanced-detector error signal (Idjadi et al., 2023). In the reported numerical comparison, the normalized frequency-discrimination gains are

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right]+\xi_i(t),\qquad \langle \xi_i(t)\xi_j(t')\rangle=\eta\,\delta(t-t')\delta_{ij},$$4

Thus the cavity-coupled MZI exceeds the conventional unbalanced MZI by more than two orders of magnitude while remaining passive and eliminating fast phase modulation. On the programmable photonic platform used for the proof-of-concept, loss limited the ring to $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right]+\xi_i(t),\qquad \langle \xi_i(t)\xi_j(t')\rangle=\eta\,\delta(t-t')\delta_{ij},$$5k and the measured discriminator gain to about $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right]+\xi_i(t),\qquad \langle \xi_i(t)\xi_j(t')\rangle=\eta\,\delta(t-t')\delta_{ij},$$6 (Idjadi et al., 2023).

In spin-noise spectroscopy, “Resources of polarimetric sensitivity in spin noise spectroscopy” exploits high-extinction polarization geometries rather than higher detector flux (Glasenapp et al., 2013). For transmitted photon flux $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right]+\xi_i(t),\qquad \langle \xi_i(t)\xi_j(t')\rangle=\eta\,\delta(t-t')\delta_{ij},$$7, a small Faraday rotation $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right]+\xi_i(t),\qquad \langle \xi_i(t)\xi_j(t')\rangle=\eta\,\delta(t-t')\delta_{ij},$$8 produces

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right]+\xi_i(t),\qquad \langle \xi_i(t)\xi_j(t')\rangle=\eta\,\delta(t-t')\delta_{ij},$$9

in the ideal shot-noise-limited case. Moving the analyzer toward extinction allows the incident sample power $\langle N_{\text{active}}\rangle$0 to increase while detector power stays fixed, thereby raising the spin-noise signal without increasing detector shot noise. Experimentally, at $\langle N_{\text{active}}\rangle$1 nm, the high-extinction geometry increased the signal from about $\langle N_{\text{active}}\rangle$2 to $\langle N_{\text{active}}\rangle$3 of shot noise, roughly a $\langle N_{\text{active}}\rangle$4 amplitude increase, and extended measurements to $\langle N_{\text{active}}\rangle$5 nm, albeit with increased perturbation and linewidth broadening (Glasenapp et al., 2013).

The same logic appears in Fourier-domain quantum interferometry. “Quantum-enhanced sensing via spectral noise reduction” derives, for $\langle N_{\text{active}}\rangle$6-photon correlated probes,

$\langle N_{\text{active}}\rangle$7

so the spectral peak height is independent of $\langle N_{\text{active}}\rangle$8 while the noise floor scales as $\langle N_{\text{active}}\rangle$9 (Dalidet et al., 18 Feb 2026). The resulting signal-to-noise ratio is

$\eta_c \approx 3\times 10^{-5}\pm 10^{-6}.$0

which gives a $\eta_c \approx 3\times 10^{-5}\pm 10^{-6}.$1 dB improvement for $\eta_c \approx 3\times 10^{-5}\pm 10^{-6}.$2. Experimentally, the single-photon and two-photon channels were acquired simultaneously under identical noise conditions; at about $\eta_c \approx 3\times 10^{-5}\pm 10^{-6}.$3 sound volume, the classical peak is buried in the PSD background while the two-photon peak remains resolvable (Dalidet et al., 18 Feb 2026).

Taken together, these architectures show that enhanced noise sensitivity can be achieved by steepening a transfer function, reallocating optical power with fixed detector flux, or suppressing the spectral background. The common feature is not necessarily a larger signal, but a more favorable conversion between perturbation and detectable contrast.

6. Quantum correlations, error correction, and sensing under dephasing

Quantum metrology often formulates enhanced sensitivity in explicitly noise-referenced terms. “Magnetic sensitivity beyond the projection noise limit by spin squeezing” demonstrates this in a magnetically sensitive ensemble of $\eta_c \approx 3\times 10^{-5}\pm 10^{-6}.$4 laser-cooled $\eta_c \approx 3\times 10^{-5}\pm 10^{-6}.$5 atoms (Sewell et al., 2011). A synthesized optical QND protocol based on two pulses with opposite polarization squeezes a mixed alignment-orientation variable while canceling probe-induced back-action. The conditional variance is written as

$\eta_c \approx 3\times 10^{-5}\pm 10^{-6}.$6

and the experiment reports $\eta_c \approx 3\times 10^{-5}\pm 10^{-6}.$7 dB noise reduction below projection noise, $\eta_c \approx 3\times 10^{-5}\pm 10^{-6}.$8 dB Wineland squeezing, and

$\eta_c \approx 3\times 10^{-5}\pm 10^{-6}.$9

Using alignment-to-orientation conversion for readout, the squeezed state yields up to a $\eta_c$0 gain and a maximum improvement of about $\eta_c$1 without noise subtraction (Sewell et al., 2011).

A complementary strategy is active protection rather than passive suppression. “Demonstration of quantum error correction for enhanced sensitivity of photonic measurements” encodes a birefringent phase $\eta_c$2 in the Bell state

$\eta_c$3

and corrects a single bit-flip on the sensing photon by phase shifts, a Hadamard, interference on a PBS, and post-selection (Cohen et al., 2016). The corrected perturbed and unperturbed states are both mapped back onto the same protected entangled state, and the proof-of-principle experiment reports recovery of about $\eta_c$4 of the sensitivity, essentially independent of the noise rate (Cohen et al., 2016).

Theoretical treatments of correlated dephasing refine when entanglement remains useful. In Ramsey spectroscopy with pure classical dephasing,

$\eta_c$5

the fundamental bound derived in “Entanglement Enhanced Sensing with Qubits affected by non-Markovian Dephasing” is

$\eta_c$6

so nonzero zero-frequency noise precludes asymptotic improvement beyond $\eta_c$7 (Kaufmann et al., 24 Apr 2026). When the noise has vanishing zero-frequency component and short-time dephasing obeys $\eta_c$8 or $\eta_c$9, optimized spin-squeezed probes achieve

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$00

respectively. The paper’s central point is that temporal correlations across shots fundamentally alter the usual reset-and-average intuition (Kaufmann et al., 24 Apr 2026).

A more device-oriented formulation appears in the two-qubit magnetometer of “Enhanced Sensitivity and Noise Resilience in Two-Qubit Quantum Magnetometers”, where the signal is inferred from

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$01

and benchmarked against

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$02

The reported message is not unlimited enhancement with time, but a finite, tunable, noise-resistant precision optimum set by the competition between coherent phase accumulation, coupling-assisted entanglement, and dephasing (Shishavan et al., 18 Aug 2025).

7. Mathematical noise sensitivity and asymptotic independence

In probability theory, enhanced noise sensitivity has a precise asymptotic meaning. “Enhanced noise sensitivity, 2D directed polymers and Stochastic Heat Flow” generalizes the Benjamini–Kalai–Schramm framework from Boolean functions to broad classes of functions of independent random variables (Caravenna et al., 14 Jul 2025). For a noisy resampling $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$03, classical noise sensitivity is

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$04

For general-valued $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$05, this is weaker than independence. The strengthened notion requires that for all $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$06,

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$07

which implies asymptotic independence of the noisy and original observables.

The basic analytic quantities are the probabilistic gradient

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$08

the $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$09 influence

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$10

and the aggregate influence

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$11

Under the paper’s hypercontractive assumption, the normalized covariance satisfies

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$12

and in the optimal hypercontractive regime,

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$13

The exponent $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$14 is proved optimal (Caravenna et al., 14 Jul 2025).

The principal application is to critical two-dimensional directed polymers. For the partition function $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$15 in the critical scaling window,

$$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$16

the space-averaged partition functions satisfy classical and enhanced noise sensitivity. Consequently, the limiting critical 2D Stochastic Heat Flow $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$17 is independent of the white noise $$x_i(n+1)=f\!\left[x_i(n)\left(1+\sum_j J_{ij}x_j(n)\right)\right],$$18 arising from the disorder under diffusive scaling. This is substantially stronger than decorrelation: the macroscopic limit is asymptotically independent of the rescaled microscopic noise field from which it originates (Caravenna et al., 14 Jul 2025).

In this mathematical setting, enhanced noise sensitivity is therefore not a matter of better sensing performance, but a limit-theoretic statement about how weak resampling destroys dependence. The contrast with the sensing literature is instructive: the same phrase can denote either operational fluctuation amplification or rigorous asymptotic independence, and the distinction is essential for interpreting claims across fields.