Nonparametric Quantum Noise Spectroscopy

• Nonparametric QNS is a set of inference protocols that reconstructs noise spectra directly from measured quantum dynamics without assuming a fixed spectral model.
• It employs various techniques such as narrow-band windowing, linear inversion, Fourier-transform, and multitaper estimation to mitigate issues like leakage, bias, and aliasing.
• Experimental implementations span diverse platforms—including superconducting qubits, levitated nanospheres, and weak-measurement systems—demonstrating its broad applicability and precision in complex settings.

Nonparametric quantum noise spectroscopy (QNS) denotes a class of inference protocols that reconstruct environmental spectral quantities directly from measured quantum dynamics without imposing a fixed functional form for the noise, such as an Ornstein–Uhlenbeck, Lorentzian, or $1/f$ model. In different realizations, the unknown object may be a power spectral density $S(\omega)$, an auto-correlation function, a complex cross-spectrum, or a higher-order time-ordered polyspectrum; the probe may be a qubit, a harmonic oscillator, a qudit, or the polarization of coherent light. Across these settings, the central operational idea is to encode the environment through a known response kernel—typically a filter function determined by free evolution, pulse modulation, or repeated control blocks—and then invert the resulting overlap relation from measured heating, decoherence, or photon-count correlations (Wu et al., 2012, 1901.10445, Norris et al., 2018, Steven et al., 9 Apr 2026).

1. Definition and historical trajectory

An early nonparametric formulation identified noise directly from the non-Markovian response of a quantum system, reversing the response relationship in the frequency domain rather than fitting a predetermined spectral ansatz. In the superconducting charge-qubit setting, the method expressed the measured observable $m(s)=M^T\chi(s)v^0$ in terms of a response function $\chi(s)$ that depends linearly on the Laplace-domain noise kernels $\Gamma(s)$ and $\Omega(s)$, enabling algebraic inversion. That framework explicitly recovered Fermi’s golden rule in the long-time Markovian limit while using transient data to improve precision in the non-Markovian regime (Wu et al., 2012).

A second major line of development recast QNS in filter-function language. For a tunable harmonic oscillator coupled linearly in position to a bosonic bath, the heating increment $\chi(T)=n(T)-n(0)$ was written as a linear functional of the bath spectrum, and the oscillator itself became a spectrometer for arbitrary non-Markovian noise under the assumptions of a large hot bath and symmetric stationary correlations (1901.10445). In parallel, qubit-based spectroscopy imported concepts from classical spectral estimation. In particular, Slepian-modulated multitaper protocols established a quantum analogue of non-parametric multitaper estimation, with explicit attention to broadband leakage, local bias, aliasing, and passband design (Norris et al., 2018).

Subsequent work diversified both the mathematical machinery and the class of accessible observables. Fourier-transform noise spectroscopy showed that free-induction decay or spin-echo data alone can be inverted to recover $S(\omega)$, eliminating the need for long pulse trains (Vezvaee et al., 2022). Walsh noise spectroscopy introduced a complete digital frame of spin-flip sequences whose sequency-domain measurements can be linearly transformed into both time-domain auto-correlation and frequency-domain spectra (Wang et al., 2022). More recent extensions generalized nonparametric reconstruction to arbitrary $d$-level spectators via Weyl-basis polyspectra (Javaherian et al., 18 Feb 2025), to two-qubit complex cross-spectra and crosstalk spectra in superconducting devices (Amezcua et al., 19 Mar 2026), and to time-ordered polyspectra under arbitrary control via a control-centric formulation that dispenses with additional symmetry conditions on the control waveform (Steven et al., 9 Apr 2026).

2. Mathematical structure of the inverse problem

Despite wide variation in physical platform, most nonparametric QNS protocols reduce the inverse problem to a linear or linearized overlap between an unknown spectral object and a known kernel. In the oscillator setting, the measured heating obeys

$$\chi(T)=\int_0^\infty d\omega\,F(\omega;T)\,S(\omega),$$

with

$S(\omega)$0

This filter is sharply peaked at $S(\omega)$1, has width of order $S(\omega)$2, and integrates to $S(\omega)$3, so sweeping the tunable oscillator frequency $S(\omega)$4 samples the spectrum point by point (1901.10445).

In the pure-dephasing qubit setting, the experimentally accessible object is typically the coherence

$S(\omega)$5

where $S(\omega)$6 is determined by the switching function of the applied dynamical-decoupling sequence. A DD-noise-spectroscopy experiment collects multiple coherence curves $S(\omega)$7 under different known filters $S(\omega)$8, and the inverse problem becomes the recovery of a nonnegative $S(\omega)$9 consistent with all measured decays (Shitara et al., 2024).

For higher-order noise characterization, the relevant objects are no longer only second-order spectra. In the control-centric formulation, one defines the $m(s)=M^T\chi(s)v^0$0th-order time-ordered polyspectrum $m(s)=M^T\chi(s)v^0$1 through the Fourier transform of the time-ordered cumulant, and the $m(s)=M^T\chi(s)v^0$2th-order contribution to the decay exponent becomes

$m(s)=M^T\chi(s)v^0$3

with $m(s)=M^T\chi(s)v^0$4. The key point is that the fully time-ordered filter collapses to a product of single-variable filter functions, which permits frequency-comb sampling under arbitrary control scenarios (Steven et al., 9 Apr 2026).

A still broader formulation appears in weak-measurement QNS with coherent light. There, sequential weak Faraday-rotation measurements yield

$m(s)=M^T\chi(s)v^0$5

so correlated photon-count differences directly encode arbitrary time-ordered $m(s)=M^T\chi(s)v^0$6th-order correlators of the target magnetization. In that framework, second-order Fourier transforms recover $m(s)=M^T\chi(s)v^0$7, while higher-order cumulants access non-Gaussian statistics (Cheung et al., 2023).

3. Reconstruction protocols and estimator design

Different nonparametric QNS protocols mainly differ in how they engineer the kernel and how they regularize or invert the measured overlaps.

Protocol family	Primary data	Reconstruction principle
Narrow-band windowing	$m(s)=M^T\chi(s)v^0$8 or $m(s)=M^T\chi(s)v^0$9	Approximate filter as localized around a target frequency
Linear inversion	Multiple $\chi(s)$0 under known filters	Solve $\chi(s)$1
Multitaper estimation	Eigenestimates from several tapers	Uniform or adaptive weighted averaging
Fourier-transform inversion	FID or SE coherence trace	Differentiate $\chi(s)$2, then Fourier transform
Walsh/frame-based estimation	Sequency-domain decoherence $\chi(s)$3	Inverse Walsh transform and logical-to-arithmetic remapping
Variational reconstruction	DD-filtered coherence set	Optimize an overcomplete trial spectrum

The simplest estimator is a narrow-band or $\chi(s)$4-approximation. For the harmonic-oscillator spectrometer, if the filter is sufficiently narrow,

$\chi(s)$5

The same logic underlies spectral decomposition with CPMG, DYSCO, and gDYSCO filters, where the measured decay is converted into an estimate at the main-lobe frequency under the assumption that out-of-band contributions are negligible or perturbatively correctable (1901.10445, Romach et al., 2018).

Full linear inversion replaces pointwise sampling by deconvolution. In the oscillator formulation one discretizes the spectrum and solves

$\chi(s)$6

using matrix inversion with Tikhonov regularization if necessary (1901.10445). Closely related binned inversions are used for two-qubit fixed-total-time pulse sequences and for qudit spectator protocols, where repeated control blocks produce a frequency comb and measurements at $\chi(s)$7 generate linear systems for $\chi(s)$8 or for multiple Weyl-basis polyspectra (Amezcua et al., 19 Mar 2026, Javaherian et al., 18 Feb 2025).

Multitaper QNS uses discrete prolate spheroidal sequences (DPSS) to construct optimally band-limited filters. For each taper $\chi(s)$9, the eigenestimate

$\Gamma(s)$0

is formed from a distinct control setting, and one may then use either a uniform average or adaptive weights chosen to minimize mean-square bias. The analysis distinguishes broadband bias from local bias and shows that DPSS filters reduce leakage relative to conventional dynamical-decoupling filters; numerical reconstructions reported $\Gamma(s)$1–$\Gamma(s)$2 bias reduction versus DD in representative cases (Norris et al., 2018).

Fourier-transform noise spectroscopy takes a different route by inverting the time-domain coherence directly. For free-induction decay,

$\Gamma(s)$3

so the spectrum is obtained from the Fourier transform of the second derivative of $\Gamma(s)$4. Spin-echo data admit a related recursive inversion through an auxiliary function $\Gamma(s)$5. The attraction of this method is minimal control overhead, but the numerical differentiation makes denoising central to practical performance (Vezvaee et al., 2022).

Walsh noise spectroscopy converts the inverse problem into an exactly linear digital-frame reconstruction. The measured decoherence exponents $\Gamma(s)$6 are identified with a finite-time Walsh transform of a logical auto-correlation $\Gamma(s)$7; inverse Walsh transformation, followed by inversion of the dyadic shuffling map, yields the arithmetic auto-correlation $\Gamma(s)$8, after which a discrete Fourier transform gives $\Gamma(s)$9. The stated limitation is no longer leakage from a comb approximation, but sampling and discretization in time space (Wang et al., 2022).

For amplitude-control noise in the presence of low-frequency dephasing and static detuning, nonparametric estimation has also been coupled to optimal control. In that setting, the target spectrum $\Omega(s)$0 is reconstructed from tomographic measurements using dephasing-robust waveforms of the analytic form

$\Omega(s)$1

which enforce identity net rotation and $\Omega(s)$2. After discretization, the spectrum is recovered by solving a nonnegative least-squares problem for $\Omega(s)$3 (Maloney et al., 2022).

A more recent self-consistent route reconstructs $\Omega(s)$4 from DD-filtered coherence measurements by expanding the trial spectrum over an overcomplete basis of symmetrized Lorentzians and minimizing a normalized sum of mean-squared differences between experimental and trial coherences. Confidence intervals are obtained from repeated random initializations, and new measurements are selected through an integrated-filter-function heuristic that identifies spectral blind spots (Shitara et al., 2024).

4. Experimental realizations and operating regimes

A particularly explicit physical realization of nonparametric QNS is the levitated nanosphere in a Paul trap. In the proposed implementation, a silica sphere with radius $\Omega(s)$5 and charge $\Omega(s)$6 is cooled to $\Omega(s)$7, allowed to evolve freely for time $\Omega(s)$8, and then read out through sideband spectroscopy or cavity detection. The trap frequency

$\Omega(s)$9

is tunable from approximately $\chi(T)=n(T)-n(0)$0 to $\chi(T)=n(T)-n(0)$1, giving a spectroscopic range of $\chi(T)=n(T)-n(0)$2–$\chi(T)=n(T)-n(0)$3, while the resolution scales as $\chi(T)=n(T)-n(0)$4 and can reach $\chi(T)=n(T)-n(0)$5 for $\chi(T)=n(T)-n(0)$6. The same framework incorporates a Markovian background term for electric-field noise in ion traps, which is subtracted before inversion (1901.10445).

Single-qubit implementations dominate the literature because decoherence or control-noise observables are readily measurable. Nitrogen-vacancy centers and superconducting qubits appear repeatedly as benchmark platforms for DD-based, FTNS, Walsh, and variational methods. In one variational DD study, ten CPMG-filtered coherences with $\chi(T)=n(T)-n(0)$7 pulses on a shallow NV center under $\chi(T)=n(T)-n(0)$8 were processed using $\chi(T)=n(T)-n(0)$9 and $S(\omega)$0, resolving a sharp peak near $S(\omega)$1 associated with the $S(\omega)$2 Larmor frequency and indicating that previous low-frequency noise estimates had been too large by an order of magnitude (Shitara et al., 2024).

Qudit spectators enlarge the accessible operator space. Using Weyl-basis decompositions $S(\omega)$3, the protocol reconstructs Gaussian stationary noise polyspectra from repeated pulse blocks and measurements of qudit observables. The reduced-Weyl framework targets $S(\omega)$4-type dephasing for qutrit, ququad, and quoct probes, whereas the full-Weyl framework for a $S(\omega)$5 Antimony quoct reconstructs multiple real polyspectral functions associated with $S(\omega)$6-dephasing. Reported simulations include qutrit estimation within $S(\omega)$7 error for moderate $S(\omega)$8 and full-Weyl Poissonian tests with $S(\omega)$9 and $d$0 (Javaherian et al., 18 Feb 2025).

Two-qubit superconducting implementations push nonparametric QNS into spatiotemporally correlated noise. The pulse-based protocol reconstructs $d$1, $d$2, $d$3, and the real and imaginary parts of the cross-spectrum $d$4, while also extracting static detunings $d$5 and static $d$6-crosstalk $d$7. Fixed-total-time pulse sequences with cosine and sine switching functions create narrow, overlapping frequency bins without repetition-based comb sharpening, and SchWARMA-engineered benchmarks reported mean absolute errors on the order of $d$8–$d$9 over $$\chi(T)=\int_0^\infty d\omega\,F(\omega;T)\,S(\omega),$$0 bins (Amezcua et al., 19 Mar 2026).

Optical weak-measurement implementations target a different regime entirely. In quantum nonlinear spectroscopy via weak Faraday rotation, a coherent light beam acts as a pseudo-spin sensor for a transparent magnetic sample. Because one beam can probe $$\chi(T)=\int_0^\infty d\omega\,F(\omega;T)\,S(\omega),$$1–$$\chi(T)=\int_0^\infty d\omega\,F(\omega;T)\,S(\omega),$$2 spins uniformly, the protocol is naturally suited to many-body systems, although its higher-order signal-to-noise ratio decays rapidly unless the fluctuations become collective, as near criticality (Cheung et al., 2023).

5. Cross-spectra, polyspectra, and non-Gaussian structure

A common restriction in early QNS was the exclusive focus on a single second-order spectrum $$\chi(T)=\int_0^\infty d\omega\,F(\omega;T)\,S(\omega),$$3 under Gaussian, stationary assumptions. More recent work makes clear that the nonparametric program extends far beyond that case.

In qudit-spectator spectroscopy, the unknown quantities are explicitly described as noise polyspectra in the Weyl basis. Repeated reference sequences of duration $$\chi(T)=\int_0^\infty d\omega\,F(\omega;T)\,S(\omega),$$4 produce a frequency comb at $$\chi(T)=\int_0^\infty d\omega\,F(\omega;T)\,S(\omega),$$5, and measurements for $$\chi(T)=\int_0^\infty d\omega\,F(\omega;T)\,S(\omega),$$6 different $$\chi(T)=\int_0^\infty d\omega\,F(\omega;T)\,S(\omega),$$7 blocks generate a linear system for discrete values $$\chi(T)=\int_0^\infty d\omega\,F(\omega;T)\,S(\omega),$$8. In the qutrit reduced-Weyl case, the signal depends on three even real functions $$\chi(T)=\int_0^\infty d\omega\,F(\omega;T)\,S(\omega),$$9, $S(\omega)$00, and $S(\omega)$01; in the $S(\omega)$02 Antimony full-Weyl case, nine complex spectra reduce to four real functions, of which three are reconstructed from the measurement equations (Javaherian et al., 18 Feb 2025).

Weak Faraday-rotation QNS treats arbitrary time-ordered correlators as the primary observables. By choosing whether each pulse measures $S(\omega)$03 or $S(\omega)$04, one selects anti-commutator or commutator superoperators and thereby accesses any of the $S(\omega)$05 distinct time-ordered $S(\omega)$06th-order correlators. Standard moment–cumulant relations then yield irreducible cumulants, so bispectra and trispectra emerge directly from measured multi-shot correlations rather than from a parametric fit to a presumed nonlinear model (Cheung et al., 2023).

The control-centric theory of time-ordered polyspectra generalizes this higher-order perspective within the filter-function formalism. Its central claim is that, once control is treated as the primary object, time-ordering no longer burdens the filters themselves: repeated control blocks generate a multidimensional sampling comb for the time-ordered polyspectra without imposing auxiliary symmetry constraints on the control waveform. Simulations reported reconstruction fidelity $S(\omega)$07 for Gaussian $S(\omega)$08 and relative error $S(\omega)$09 for bispectral reconstruction in a quantum non-Gaussian bath (Steven et al., 9 Apr 2026).

Second-order multichannel generalizations sit between ordinary PSD estimation and full higher-order spectroscopy. The two-qubit superconducting protocol reconstructs a complex cross-spectrum, with $S(\omega)$10 and $S(\omega)$11 separated by suitable pulse configurations. This makes explicit that “nonparametric QNS” is not confined to one real-valued spectrum of one sensor, but includes spatial correlations and fluctuating crosstalk in multi-qubit devices (Amezcua et al., 19 Mar 2026).

6. Assumptions, error sources, and interpretive boundaries

Nonparametric QNS avoids a fixed functional form for the target spectrum, but it does not eliminate dynamical, statistical, or control-theoretic assumptions. In the harmonic-oscillator spectrometer, the derivation assumes a large hot bath, negligible dissipation, and a real stationary two-point correlation $S(\omega)$12 whose Fourier transform is nonnegative and symmetric. The same work notes that colder environments require dissipative terms and quadrature-decoherence measurements, and that finite back-action, sphere-charging drifts, and trap nonlinearities bound the usable free-evolution time $S(\omega)$13 (1901.10445).

Many qubit protocols assume stationary Gaussian noise, weak coupling, and a second-order cumulant truncation. Fourier-transform noise spectroscopy additionally relies on taking second derivatives of noisy coherence data, which makes low-pass filtering, early-time interpolation, and long-time fitting essential; finite $S(\omega)$14 sets the low-frequency limit, while finite sampling interval $S(\omega)$15 sets the high-frequency cutoff $S(\omega)$16 (Vezvaee et al., 2022). In multitaper spectroscopy, the primary technical issues are broadband leakage, local bias, and aliasing; although DPSS filters suppress side lobes, Nyquist constraints and out-of-band power remain central (Norris et al., 2018).

Repeated-sequence comb methods have their own asymptotic requirements. In qudit spectroscopy, $S(\omega)$17 must be large enough that the comb approximation holds, specifically $S(\omega)$18 bath correlation time, and control complexity increases with $S(\omega)$19 because selective pulses $S(\omega)$20 are required across the level structure (Javaherian et al., 18 Feb 2025). In weak Faraday-rotation spectroscopy, the signal-to-noise ratio for $S(\omega)$21th-order signals scales as

$S(\omega)$22

so higher-order cumulants are essentially inaccessible for uncorrelated spins unless critical collective fluctuations amplify the correlators (Cheung et al., 2023).

Multi-qubit implementations introduce additional calibration issues. In the two-qubit superconducting protocol, static angles must satisfy $S(\omega)$23 and $S(\omega)$24 to avoid phase wrapping; finite $S(\omega)$25-pulse widths generate off-axis errors; and $S(\omega)$26 must remain much shorter than $S(\omega)$27 so that relaxation does not contaminate the dephasing signal (Amezcua et al., 19 Mar 2026). Variational DD-based reconstructions address some of these practical limitations by incorporating measurement-error robustness and repeated randomized runs for uncertainty quantification, but they still require basis selection, optimizer convergence, and sufficient filter sensitivity across the frequency range of interest (Shitara et al., 2024).

This suggests that, in current usage, “nonparametric” refers primarily to the absence of a fixed spectral model, not to the absence of structure altogether. The structure persists in the assumed noise statistics, in the control-generated filter map, in discretization and regularization choices, and in the finite-resolution limits set by experiment. Within those boundaries, nonparametric QNS has evolved from single-spectrum inversion of transient qubit response into a broad framework for reconstructing PSDs, cross-spectra, and time-ordered polyspectra across oscillators, qubits, qudits, and optical probes.