Noise-Canceling Feedback

• Noise-canceling feedback is a multidisciplinary approach that measures disturbances via auxiliary channels and synthesizes counter-signals to suppress noise.
• It employs methods such as destructive interference, adaptive subtraction, and algorithmic filtering to stabilize systems in acoustics, quantum control, and biology.
• Its versatility is evident in applications ranging from audio feedback cancellation and active noise control in devices to decoherence suppression in quantum systems.

Searching arXiv for the cited works and closely related papers to ground the article in published sources. Noise-canceling feedback denotes a family of control, estimation, and signal-processing strategies in which a disturbance is measured directly, inferred from correlated channels, or encoded in a monitoring record, and the resulting signal is used to suppress that disturbance by destructive interference, adaptive subtraction, or trajectory shaping. In the cited literature, the term spans acoustic feedback cancellation in microphone–loudspeaker loops, active noise control in headphones and smart glasses, magnetic-field and microwave-noise suppression, quantum protocols that cancel decoherence or measurement backaction, and biological negative-feedback circuits analyzed as causal filters (Roebben et al., 1 Dec 2025, Hilgemann et al., 19 Sep 2025, Yuan et al., 7 Apr 2026, Szigeti et al., 2014, Hinczewski et al., 2016).

1. Scope and domain-specific meanings

The terminology is not uniform across fields. In audio signal processing, “noise-canceling feedback” commonly refers to acoustic feedback cancellation in a closed-loop microphone–loudspeaker system, where loudspeaker output re-enters the microphone and can produce instability and acoustic howling (Roebben et al., 1 Dec 2025). In quantum control, “no-knowledge feedback” and related “noise-canceling” protocols refer to feeding back a measurement record that contains the environmental noise itself, so that decoherence or stochastic backaction is canceled in real time (Szigeti et al., 2014). In systems biology, negative feedback is formulated as a Wiener–Kolmogorov filtering problem in which the feedback loop constructs an estimate of a noisy signal and subtracts it from the target fluctuation (Hinczewski et al., 2016). In online optimization, “canceling noise” has a narrower meaning: two function evaluations at the same round share the same perturbation, so differencing removes the oracle noise exactly (Akhavan et al., 2022).

This diversity suggests that the unifying feature is not a particular hardware topology but a structural motif: the disturbance must be represented in an auxiliary variable, channel, or state estimate closely enough that a compensating action can be synthesized.

Domain	Disturbance channel	Cancellation principle
Acoustic feedback control	Loudspeaker-to-microphone coupling	Identify $F(q)$, decorrelate regressors, or reinterpret feedback as late reverberation
Active noise control	Residual acoustic, magnetic, or electrical disturbance	Generate anti-noise or anti-field through a secondary path
Quantum feedback	Decoherence or diffusive measurement backaction	Feed back the monitored noise or choose control so the stochastic term vanishes
Biological and algorithmic filtering	Molecular fluctuations or oracle perturbations	Subtract a causal estimate or difference matched evaluations

A recurrent misconception is that noise cancellation is intrinsically acoustic. The cited work shows instead that the same abstract idea appears in audio, magnetics, microwave engineering, quantum control, biosignal denoising, and stochastic optimization (Pyragius et al., 2021, Depellette et al., 2 Jun 2025, Porr et al., 2020).

2. Acoustic feedback cancellation and identifiability

The most classical formulation appears in closed-loop audio systems with a microphone and loudspeaker in the same acoustic environment. The microphone signal is modeled as

$m[k] = F(q)\,l[k] + s[k],$

where $F(q)$ is the acoustic feedback path and $s[k]$ is the external input signal impinging on the microphone. When the forward-path amplification $G(q)$ is high enough, the loop can become unstable and produce acoustic howling, a central problem in hearing aids and public-address systems (Roebben et al., 1 Dec 2025).

A basic difficulty is that microphone and loudspeaker signals are correlated by the closed loop, so naïve adaptive identification of the feedback path is biased. Prediction error method (PEM) feedback cancellation addresses this by modeling the input signal as an autoregressive process,

$$s[k] = H(q)w[k] = \frac{1}{D(q)}w[k] = \frac{1}{1+q^{-1}\overline{D}(q)}w[k],$$

with white noise $w[k]$. In the two-channel adaptive feedback canceller (2ch-AFC), the prediction error is

$\epsilon_p[k] = B(q)l[k] + A(q)m[k],$

and the least-squares solution yields coefficients from which the feedback-path estimate is recovered by inverse filtering,

$\hat{F}(q) = \frac{-1}{A(q)}B(q).$

Correct recovery requires identifiability: the PEM solution must imply $A(q)=D(q)$, $m[k] = F(q)\,l[k] + s[k],$0, and therefore $m[k] = F(q)\,l[k] + s[k],$1 (Roebben et al., 1 Dec 2025).

The classic delay-based result attributed to Spriet et al. requires the forward-path delay $m[k] = F(q)\,l[k] + s[k],$2 to satisfy

$m[k] = F(q)\,l[k] + s[k],$3

where $m[k] = F(q)\,l[k] + s[k],$4 is the AR predictor order. The 2ch-AFC analysis generalizes this to an invertibility-based sufficient condition,

$m[k] = F(q)\,l[k] + s[k],$5

with $m[k] = F(q)\,l[k] + s[k],$6 the order of the forward-path feedforward numerator. The key object is the regressor correlation matrix $m[k] = F(q)\,l[k] + s[k],$7; identifiability is equivalent to $m[k] = F(q)\,l[k] + s[k],$8 being invertible, or equivalently to linearly independent combined microphone and loudspeaker regressors. The condition number

$m[k] = F(q)\,l[k] + s[k],$9

then serves as a practical identifiability diagnostic: small $F(q)$0 indicates better conditioning, whereas large $F(q)$1 signals ill-conditioned or unreliable identification (Roebben et al., 1 Dec 2025).

Two later developments broaden this picture. First, acoustic feedback can under mild conditions be absorbed into the late component of an effective room response. If the total loop delay is sufficiently long and the closed-loop transfer function can be reasonably approximated by an FIR filter, then the microphone signal can be written as

$F(q)$2

with $F(q)$3, where the late part $F(q)$4 includes both late reverberation and feedback. This reduces joint dereverberation and acoustic feedback cancellation to a dereverberation-only problem and enables the use of algorithms such as WPE (Liekens et al., 29 May 2026).

Second, decorrelation can be introduced deliberately to reduce adaptive-filter bias. In speech communication, the least-squares room-path estimate becomes

$F(q)$5

so correlation between loudspeaker signal $F(q)$6 and desired speech $F(q)$7 induces a bias term. A DFT-filter-bank “phase synthesizer” combines frequency shifting, phase modulation, and variable delay lines implemented as phase-only modifications to reduce that correlation. The variable-delay extension, analogous to vibrato and chorus effects, improves stability while preserving higher PESQ than earlier phase-modulation and frequency-shifting combinations (Linhard et al., 14 Oct 2025).

3. Active cancellation in classical physical systems

In classical engineering systems, noise-canceling feedback typically operates by generating an anti-noise signal that traverses a secondary path and destructively interferes with the disturbance. In headphone feedback ANC, the residual disturbance rejection is quantified by the sensitivity

$F(q)$8

with $F(q)$9 the controlled plant and $s[k]$0 the controller. A central limitation is plant uncertainty arising from differing wearing situations, human heads and ears, manufacturing tolerances, wear, temperature, and movement. Conventional norm-bounded disk models are often overly conservative because measured plant clouds in the complex plane are not circular. Data-driven multi-disk, elliptic, and convex-hull uncertainty sets provide a tighter fit and therefore allow higher attenuation while maintaining robust stability. In measurements on Bose QC45 over-ear headphones, all four controllers remained stable, and the convex-hull model improved attenuation around $s[k]$1 from about $s[k]$2 to about $s[k]$3, with reported gains of $s[k]$4–$s[k]$5 across the low-frequency band below $s[k]$6 relative to norm-bounded modeling (Hilgemann et al., 19 Sep 2025).

Open-ear smart glasses present a different architectural constraint: conventional ANC relies on an error microphone inside or at the entrance of the ear canal, but open-ear devices cannot place such a sensor in the usual position. A recent solution uses eight frame microphones and two open-ear speakers together with neural “virtual in-ear sensing.” The system estimates ANC filters from the frame signals, executes them on a low-latency DSP, and includes a separate acoustic feedback cancellation stage to remove speaker leakage from the microphone inputs. The reported mean noise reduction is $s[k]$7 without calibration and $s[k]$8 with brief user-specific calibration, in the $s[k]$9–$G(q)$0 band (Yuan et al., 7 Apr 2026).

The same anti-noise architecture extends beyond acoustics. In magnetic-field ANC, a reference fluxgate magnetometer measures ambient field noise, an error magnetometer measures the residual field, and an FPGA implements Filtered-x LMS so that Helmholtz coils generate an opposing field. Secondary-path identification is essential because DACs, current sources, coils, magnetometers, anti-aliasing filters, and ADCs all contribute delay and filtering. The system reports up to $G(q)$1 RMS suppression in the DC–$G(q)$2 band, with $G(q)$3 amplitude suppression at $G(q)$4 and $G(q)$5 at $G(q)$6 for all three field axes (Pyragius et al., 2021).

In mechanically resonant feedback systems, cancellation may be applied to the plant rather than the disturbance field. An FPGA FIR inverse filter can flatten the open-loop response by canceling both poles and zeros of the plant before the controller acts. In a high-finesse optical resonator, this approach canceled the ten largest mechanical resonances and anti-resonances, each with linewidths down to $G(q)$7, and increased the unity-gain bandwidth from $G(q)$8 to $G(q)$9 (Ryou et al., 2016).

Some implementations rely on common-mode subtraction instead of physical anti-noise generation. In an optical ring cavity, two independently controlled beams resonant with adjacent cavity modes acquire strongly correlated cavity-length noise in their PDH error signals. Forming

$$s[k] = H(q)w[k] = \frac{1}{D(q)}w[k] = \frac{1}{1+q^{-1}\overline{D}(q)}w[k],$$0

cancels the common cavity-motion term while preserving a differential phase shift on one beam. The reported differential signal shows more than $$s[k] = H(q)w[k] = \frac{1}{D(q)}w[k] = \frac{1}{1+q^{-1}\overline{D}(q)}w[k],$$1 reduction in cavity noise down to the noise floor up to $$s[k] = H(q)w[k] = \frac{1}{D(q)}w[k] = \frac{1}{1+q^{-1}\overline{D}(q)}w[k],$$2, and a simulated spin-squeezing measurement achieves phase resolution $$s[k] = H(q)w[k] = \frac{1}{D(q)}w[k] = \frac{1}{1+q^{-1}\overline{D}(q)}w[k],$$3 (Wang et al., 2020).

Device-level narrowband compensation follows the same logic. In Si/SiGe gated quantum dots operated at $$s[k] = H(q)w[k] = \frac{1}{D(q)}w[k] = \frac{1}{1+q^{-1}\overline{D}(q)}w[k],$$4, an automated feedback loop identifies a dominant $$s[k] = H(q)w[k] = \frac{1}{D(q)}w[k] = \frac{1}{1+q^{-1}\overline{D}(q)}w[k],$$5 interference, synthesizes a phase-coherent anti-noise tone, minimizes beat frequency, and adaptively tunes phase and amplitude. The dominant $$s[k] = H(q)w[k] = \frac{1}{D(q)}w[k] = \frac{1}{1+q^{-1}\overline{D}(q)}w[k],$$6 component is reduced by about $$s[k] = H(q)w[k] = \frac{1}{D(q)}w[k] = \frac{1}{1+q^{-1}\overline{D}(q)}w[k],$$7, and the integrated noise power from $$s[k] = H(q)w[k] = \frac{1}{D(q)}w[k] = \frac{1}{1+q^{-1}\overline{D}(q)}w[k],$$8 to $$s[k] = H(q)w[k] = \frac{1}{D(q)}w[k] = \frac{1}{1+q^{-1}\overline{D}(q)}w[k],$$9 drops from $w[k]$0 to $w[k]$1 with $w[k]$2 anti-noise phase (Bharadwaj et al., 9 Nov 2025). Microwave amplitude-noise cancellation uses an FPGA to reproduce amplitude sidebands with tunable gain and delay and then recombines them with the original $w[k]$3 tone; at a $w[k]$4 offset the reported suppression is about $w[k]$5, reducing the externally induced cavity-heating rate in an optomechanics experiment by a factor of $w[k]$6 and the minimum oscillator occupation by a factor of $w[k]$7 (Depellette et al., 2 Jun 2025).

An algorithmic rather than physical anti-noise generator appears in real-time EEG denoising. A compound electrode provides a main EEG channel $w[k]$8 and a noise-reference channel $w[k]$9. A Deep Neural Filter maps the tapped-delay-line reference to a learned remover signal $\epsilon_p[k] = B(q)l[k] + A(q)m[k],$0, and the residual

$\epsilon_p[k] = B(q)l[k] + A(q)m[k],$1

is both the filtered output and the online learning error. In recordings with jaw-EMG contamination, the reported average SNR improvement is $\epsilon_p[k] = B(q)l[k] + A(q)m[k],$2, versus $\epsilon_p[k] = B(q)l[k] + A(q)m[k],$3 for LMS adaptive FIR filtering, with a maximum improvement of $\epsilon_p[k] = B(q)l[k] + A(q)m[k],$4 (Porr et al., 2020).

4. Quantum formulations of noise-canceling feedback

Quantum versions of noise-canceling feedback replace destructive interference in a physical field by cancellation of stochastic terms in conditioned dynamics. In no-knowledge quantum feedback, a continuously monitored open system obeys

$\epsilon_p[k] = B(q)l[k] + A(q)m[k],$5

For Hermitian coupling $\epsilon_p[k] = B(q)l[k] + A(q)m[k],$6, homodyne detection at phase $\epsilon_p[k] = B(q)l[k] + A(q)m[k],$7 gives

$\epsilon_p[k] = B(q)l[k] + A(q)m[k],$8

so the measurement record contains only the white-noise input from the environment. With perfect efficiency $\epsilon_p[k] = B(q)l[k] + A(q)m[k],$9, the conditional evolution becomes

$\hat{F}(q) = \frac{-1}{A(q)}B(q).$0

Applying the feedback Hamiltonian $\hat{F}(q) = \frac{-1}{A(q)}B(q).$1 cancels the stochastic term exactly and leaves

$\hat{F}(q) = \frac{-1}{A(q)}B(q).$2

For $\hat{F}(q) = \frac{-1}{A(q)}B(q).$3, the residual decoherence is suppressed by the factor $\hat{F}(q) = \frac{-1}{A(q)}B(q).$4 (Szigeti et al., 2014).

A more general treatment of diffusive trajectories formulates noise-canceling quantum feedback as the choice of a coherent feedback increment $\hat{F}(q) = \frac{-1}{A(q)}B(q).$5 that removes the $\hat{F}(q) = \frac{-1}{A(q)}B(q).$6 term itself. Under the idealized assumptions of pure states, unit measurement efficiency, and zero time delay, perfectly noise-canceling feedback always exists. In the pure-state case, the cancellation condition is

$\hat{F}(q) = \frac{-1}{A(q)}B(q).$7

and the resulting evolution is generated deterministically by an effective non-Hermitian Hamiltonian. The formalism is applied to half-parity and fluorescence measurements for entangled-state preparation and stabilization, and to time-continuous $\hat{F}(q) = \frac{-1}{A(q)}B(q).$8-to-$\hat{F}(q) = \frac{-1}{A(q)}B(q).$9 magic-state distillation. In that setting, adding noise-canceling feedback increases successful postselection probabilities by about $A(q)=D(q)$0–$A(q)=D(q)$1 times and extends the effective input-error threshold from the standard

$A(q)=D(q)$2

to about

$A(q)=D(q)$3

The same work is explicit that perfect cancellation is not generic for mixed states (Karmakar et al., 8 Jul 2025).

A third quantum line of work treats noise suppression through coherent feedback without measurement. A plant on $A(q)=D(q)$4 is coupled coherently to a controller on $A(q)=D(q)$5, and the interaction Hamiltonian is modulated as

$A(q)=D(q)$6

If the composite feedback generator has a unique steady state $A(q)=D(q)$7, then transient errors are asymptotically erased, and under persistent Markovian noise the reduced plant error satisfies a long-time bound of order

$A(q)=D(q)$8

with $A(q)=D(q)$9 the feedback strength. The paper illustrates this with a two-qubit plant and a two-level controller (Zhang et al., 2024).

One implication of these results is that quantum noise cancellation does not always require state estimation in the classical feedback sense. In no-knowledge feedback, the record is useful precisely because it contains no system information; in noise-canceling diffusive feedback, the target is cancellation of stochastic backaction rather than reconstruction of the state (Szigeti et al., 2014, Karmakar et al., 8 Jul 2025).

5. Biological, stochastic, and network-theoretic formulations

In gene regulatory networks, negative feedback can be formalized as a noise-canceling filter. For a linearized two-species network with target $m[k] = F(q)\,l[k] + s[k],$00 and mediator $m[k] = F(q)\,l[k] + s[k],$01, the feedback estimate is written as

$m[k] = F(q)\,l[k] + s[k],$02

and the output fluctuation is

$m[k] = F(q)\,l[k] + s[k],$03

The Wiener–Kolmogorov problem is to choose a causal kernel $m[k] = F(q)\,l[k] + s[k],$04 minimizing the output variance. In a TetR autorepression circuit, the WK-optimal Fano factor is

$m[k] = F(q)\,l[k] + s[k],$05

while the Lestas–Vinnicombe–Paulsson lower bound is

$m[k] = F(q)\,l[k] + s[k],$06

For large $m[k] = F(q)\,l[k] + s[k],$07, the standard deviation scales as $m[k] = F(q)\,l[k] + s[k],$08. The paper further shows that steep Hill-like nonlinear regulation asymptotically reaches the same WK optimum rather than surpassing it (Hinczewski et al., 2016).

In online zero-order optimization, the phrase “canceling noise” refers to a two-point oracle model in which the two noisy evaluations at a given round share the same perturbation,

$m[k] = F(q)\,l[k] + s[k],$09

Then the difference $m[k] = F(q)\,l[k] + s[k],$10 removes noise exactly, and the resulting $m[k] = F(q)\,l[k] + s[k],$11-sphere estimator is unbiased for the gradient of an $m[k] = F(q)\,l[k] + s[k],$12-smoothed objective (Akhavan et al., 2022). This is a very different setting from physical feedback control, but it preserves the same algebraic structure of matched perturbation plus differencing.

Noise-canceling dynamics also arise in stochastic oscillator theory. In event-triggered feedback for noise-driven phase oscillators, each event induces an instantaneous increase or decrease in frequency followed by exponential relaxation,

$m[k] = F(q)\,l[k] + s[k],$13

Positive feedback can produce bistable dynamics and alter excitability, while both positive and negative feedback can lead to more regular output for particular noise strengths. The sign of the induced serial correlation is also feedback-dependent: positive feedback yields extended positive interval correlations, whereas negative feedback yields short-ranging negative correlations (Kromer et al., 2014).

At the network-design level, noise cancellation can be “hard-wired” into the topology itself. In arrays of second-order phase oscillators, the fluctuation variance is expressed in terms of the graph Laplacian $m[k] = F(q)\,l[k] + s[k],$14 and the noise covariance. For white noise,

$m[k] = F(q)\,l[k] + s[k],$15

and under the cost constraint

$m[k] = F(q)\,l[k] + s[k],$16

the optimal architecture depends strongly on spatial and temporal correlation structure. When the input fluctuations become more correlated in space or time, optimal networks become sparser and more hierarchically organized, resembling plant or animal vasculature (Ronellenfitsch et al., 2018).

6. Limits, diagnostics, and recurring misconceptions

A central limitation across domains is that formal cancelability and practical cancelability are not the same. In acoustic feedback cancellation, satisfying the order condition $m[k] = F(q)\,l[k] + s[k],$17 does not guarantee numerically reliable identification if the regressor matrix is close to singular; the condition number $m[k] = F(q)\,l[k] + s[k],$18 is therefore a practical indicator of whether cancellation will be stable and accurate (Roebben et al., 1 Dec 2025). In robust headphone ANC, a controller may be safe yet unnecessarily weak if the uncertainty model is too conservative; the measurements on over-ear and in-ear headphones show that disk-based uncertainty can overestimate the physically relevant plant set (Hilgemann et al., 19 Sep 2025).

Latency and secondary-path modeling are equally recurrent constraints. FxLMS-based magnetic ANC depends on accurate secondary-path identification and on coherence between reference and error sensors, with achievable cancellation limited by

$m[k] = F(q)\,l[k] + s[k],$19

Reference-sensor contamination by the actuator field can also produce an unwanted feedback loop (Pyragius et al., 2021). Open-ear smart-glasses ANC is constrained by the absence of a conventional in-ear error microphone and by microsecond-scale DSP latency; the reported performance drops markedly when processing delay increases (Yuan et al., 7 Apr 2026). In microwave amplitude-noise cancellation, the cancellation bandwidth is set by delay mismatch, and the residual floor is set by the uncanceled phase noise of the source (Depellette et al., 2 Jun 2025).

Quantum work imposes even stricter idealizations. Perfect diffusive noise cancellation requires pure states, unit measurement efficiency, and zero feedback delay, and exact mixed-state cancellation is not generic (Karmakar et al., 8 Jul 2025). No-knowledge feedback similarly suppresses decoherence in proportion to the detected fraction of the bath noise, leaving a residual $m[k] = F(q)\,l[k] + s[k],$20 when $m[k] = F(q)\,l[k] + s[k],$21 (Szigeti et al., 2014). Coherent quantum feedback guarantees asymptotic recovery or bounded suppression only under explicit structural conditions on the composite generator and interaction Hamiltonian (Zhang et al., 2024).

Another common misconception is that all “noise-canceling feedback” is literally feedback. Some high-performance laboratory schemes are explicitly feed-forward: environmental disturbance is sensed by witness channels and a cancellation signal is applied before the target channel is contaminated. In a suspended Fabry–Perot interferometer, Wiener filtering and FxLMS-based online adaptive filtering reduce seismic contamination by using seismometers and accelerometers as witnesses rather than relying on ordinary error-driven feedback (Driggers et al., 2011). The broader literature therefore contains both strict feedback loops and feedback-like disturbance-rejection architectures.

Finally, noise cancellation does not always mean removal of mean signal components. In quantum imaging with undetected squeezed photons, the object can be reconstructed from quadrature noise itself; the uncertainty

$m[k] = F(q)\,l[k] + s[k],$22

is object-dependent, and phase-cycled combinations suppress background contributions while retaining image information encoded in fluctuations (Samimi et al., 2024). A plausible implication is that the concept of noise-canceling feedback is most coherent when defined operationally—by how disturbance information is represented and opposed—rather than by any single physical realization or disciplinary vocabulary.