Nonlinear Covariance Spectroscopy

• Nonlinear covariance spectroscopy is a technique that uses shot-to-shot fluctuations to reveal nonlinear dynamics by analyzing covariance in spectral data.
• It treats inherent noise as independent probes to extract nonlinear coupling and energy transfer pathways using covariance and Pearson coefficients.
• The method applies to ultrafast optics, vibrational, and quantum spectroscopy, enabling the resolution of weak signals without complex phase matching.

Nonlinear covariance spectroscopy is a statistical and physical methodology that leverages shot-to-shot fluctuations, rather than averaging them away, to extract sample-specific correlations induced by nonlinear dynamics. Originally developed in the context of ultrafast nonlinear optics, the approach has since been generalized to vibrational, chemical, and quantum noise spectroscopy, as well as to statistical time series and nonlinear recurrent networks. By computing covariance or higher-order correlators between fluctuating observables across an ensemble of measurements, one decodes nonlinear coupling mechanisms, energy transfer pathways, and dynamic structure that are invisible to conventional mean-value spectral or signal-averaged techniques (Tollerud et al., 2018, Suresh et al., 2020, Sparapassi et al., 2021, Jordanger et al., 2017, Jordanger et al., 2017, Shen et al., 7 Aug 2025).

1. Fundamental Principles of Nonlinear Covariance Spectroscopy

The central idea is that pulse-to-pulse noise or stochastic fluctuations in an input field or system act as independent “probes” of a nonlinear system. In the optical domain, each laser shot is treated as a distinct experiment, exposing the system to an ensemble of spectrally uncorrelated conditions. Correlations—computed via covariance or normalized Pearson coefficients—between spectral or temporal bins in the output reveal the presence and structure of nonlinear energy transfer mechanisms, as these correlations cannot arise from the originally uncorrelated input noise (Tollerud et al., 2018).

The two-point spectral covariance is defined as

$$C(\omega_1,\omega_2) = \langle \delta I(\omega_1) \delta I(\omega_2) \rangle,$$

where $$\delta I(\omega) = I(\omega) - \langle I(\omega) \rangle$$. The normalized version, the Pearson coefficient,

$$P(\omega_1,\omega_2) = \frac{\langle \delta I(\omega_1) \delta I(\omega_2) \rangle}{\sigma(\omega_1) \sigma(\omega_2)},$$

measures the strength and sign of the induced correlation, with $P=1$ for perfect correlation, $P=0$ for uncorrelated (Tollerud et al., 2018, Suresh et al., 2020).

Higher-order generalizations (three- or four-point correlations) permit access to multi-dimensional couplings and nonlinear response functions inaccessible by standard two-point analysis (Cheung et al., 2023, Sparapassi et al., 2021).

2. Experimental Realizations and Data Analysis

In ultrafast optics, two primary platforms dominate: spectrally shaped femtosecond lasers and broadband “noise burst” sources.

• Pulse-Shaping Implementation: A femtosecond Ti:Sapphire amplifier produces ultrashort pulses; a 2D spatial light modulator imprints random spectral phase shot-by-shot to create spectrally localized, uncorrelated fluctuations of controlled correlation length. The shaped beam passes through a nonlinear optical sample (e.g., α-quartz), and the spectra are resolved by diode arrays. Covariances are computed for both the reference (no sample) and the signal arm to distinguish sample-induced from instrumental correlations (Tollerud et al., 2018, Sparapassi et al., 2021).
• Modulational Instability Supercontinuum: Broadband noisy pulses are generated by modulational instability in gas-filled hollow-core photonic-crystal fiber (HC-PCF). These “noise bursts” probe Raman-active transitions in molecular gases, with output spectra acquired shot-by-shot by high-resolution spectrometers. Statistical analysis yields covariance maps with off-diagonal signatures at frequency differences matching the vibrational transitions (Suresh et al., 2020).

The typical analysis workflow is as follows (Tollerud et al., 2018, Suresh et al., 2020, Sparapassi et al., 2021):

1. Acquire large ensembles (e.g., $10^4$–$5\times10^4$) of single-shot spectra.
2. Compute mean and fluctuations for each spectral bin.
3. Form the covariance (or Pearson) map between all frequency bins.
4. Identify sample-induced features, e.g., off-diagonal streaks at separation $\pm\Omega$ for Raman-active modes.
5. Optionally, transform the map into a physically meaningful axis (e.g., Raman frequency, delay) for visualization and further quantitative extraction (peak position, linewidth, amplitude, phase).

Covariance detection is inherently phase-sensitive and enables amplitude and phase-resolved recovery of oscillatory modes, dephasing, and population transfer, which are not accessible with standard mean-value detection (Sparapassi et al., 2021).

3. Theoretical Frameworks and Nonlinear Response

The physical mechanism underlying nonlinear covariance spectroscopy is the encoding of shot-to-shot input field correlations by the system's nonlinear susceptibility. For third-order nonlinearities (e.g., stimulated Raman scattering, coherent anti-Stokes Raman scattering), the interaction generates energy or phase transfer between spectral bins separated by vibrational frequencies of the sample.

Formally, for SRS, the measured intensity covariance:

$$C_{\mathrm{SRS}}(\omega_1, \omega_2) \propto \delta(\omega_2 - \omega_1 \mp \Omega)$$

encodes vibrational modes at frequency $\Omega$ as off-diagonal ridges in the covariance map (Tollerud et al., 2018).

In transient spectroscopy, a broadband pump impulsively excites a superposition of phonon modes; a noisy probe’s covariance statistics then resolve independent amplitude, phase, and dephasing of these modes as a function of pump–probe delay, yielding a single-delay 2D-Raman spectrum (covariance-based 2D Raman) (Sparapassi et al., 2021).

Quantum generalizations involve sequential weak Faraday rotation measurements on a spin system, where photon-count difference correlators are proportional to time-ordered higher-order correlation functions of sample observables. This quantum nonlinear spectroscopy protocol allows arbitrary K-point noise and nonlinear response spectra to be accessed by covariance analysis, far beyond what is possible in classical or mean-value paradigms (Cheung et al., 2023).

4. Extensions, Performance, and Applications

Nonlinear covariance spectroscopy has demonstrated the following advantages and properties:

• Noise as a Resource: Intentionally engineered or natural shot-to-shot noise enhances sensitivity to nonlinear correlations. Detection works optimally when the noise correlation length is shorter than the nonlinear coupling scale (e.g., Raman shift) (Tollerud et al., 2018, Suresh et al., 2020, Sparapassi et al., 2021).
• Single-beam and Single-detector Geometry: No requirement for multiple beams, complex phase-matching, or high-stability heterodyne setups—operation in a straightforward transmission or reflection configuration suffices (Tollerud et al., 2018).
• Rapid Acquisition and High Spectral Resolution: MHz-rate sources with dispersive Fourier transform enable sub-millisecond sampling; spectral resolution is ultimately set by the noise correlation length and spectrometer capabilities (Suresh et al., 2020).

Key applications include:

• Raman and Vibrational Spectroscopy: Infrared-inactive transitions, multi-mode phonon dynamics, and weak lines can be efficiently resolved (H₂, CH₄, quartz, mixed gases) (Suresh et al., 2020, Sparapassi et al., 2021).
• Discrimination of Resonant and Non-Resonant Processes: Two-dimensional synchronous and asynchronous covariance analysis (and 1D diagonal projections) allows separation of resonant molecular signals from non-resonant backgrounds in coherent four-wave mixing (CARS) (Nagpal et al., 2020).
• Quantum Many-Body Systems: Weak Faraday rotation covariance protocols access time-ordered correlation functions for systems with otherwise inaccessible higher-order quantum correlations (Cheung et al., 2023).
• Nonlinear Networks: In neural circuit models, the spectrum of covariance matrices (covariance spectroscopy) enables mechanistic inferences on population dynamics, connectivity, and criticality under nonlinearities and chaos (Shen et al., 7 Aug 2025).

5. Nonlinear Covariance Spectroscopy in Stochastic Time Series Analysis

Beyond physical experiments, nonlinear covariance spectroscopy finds a rigorous statistical counterpart in the local Gaussian spectrum framework. Here, local (value-conditional) covariance and its frequency-domain representation enable detection of hidden periodicities, nonlinear dependencies, or tail-specific phenomena in stationary time series.

Given a stationary sequence $$\delta I(\omega) = I(\omega) - \langle I(\omega) \rangle$$0, the local (level-dependent) spectrum

$$\delta I(\omega) = I(\omega) - \langle I(\omega) \rangle$$1

is built from the local covariance $$\delta I(\omega) = I(\omega) - \langle I(\omega) \rangle$$2, defined as the second mixed derivative of the log joint density around $$\delta I(\omega) = I(\omega) - \langle I(\omega) \rangle$$3. Peaks or troughs in $$\delta I(\omega) = I(\omega) - \langle I(\omega) \rangle$$4 at extreme $$\delta I(\omega) = I(\omega) - \langle I(\omega) \rangle$$5 reveal state-dependent oscillations invisible to the classical spectrum $$\delta I(\omega) = I(\omega) - \langle I(\omega) \rangle$$6 (Jordanger et al., 2017, Jordanger et al., 2017).

Algorithmic steps involve local Gaussian kernel estimation, selection of lag and conditioning points, and computation of the local spectrum by discrete Fourier transform, with diagnostics provided by thresholding, bootstrapping, and quantile-spectral mapping.

6. Generalizations, Limitations, and Outlook

Nonlinear covariance spectroscopy is broadly extensible to multidimensional, ultrafast, and quantum spectroscopies. It enables:

• Higher-Order Correlation Access: Three-point, four-point, and higher nonlinearities (e.g., 5th order Raman, multidimensional electronic- or X-ray spectroscopies) become experimentally tractable (Sparapassi et al., 2021, Cheung et al., 2023).
• Adaptation to Noisy Sources: Inherently noisy sources such as SASE X-ray FELs, previously unsuitable for phase-sensitive experiments, are ideally suited to covariance-based detection (Tollerud et al., 2018, Sparapassi et al., 2021).
• Materials with Ultrafast Dephasing: The time resolution is set by the coherence time of the noise or probe, enabling spectroscopy in systems with ultrafast relaxation (Suresh et al., 2020).

Limitations include the necessity for large ensembles to overcome uncorrelated detector noise, balancing spectral vs. temporal resolution via noise tuning, and the requirement for broad spectral coverage when addressing multiple transitions (Sparapassi et al., 2021). In quantum protocols, spatial resolution is diffraction-limited and higher-order correlator amplitudes fall off rapidly except near collective critical phenomena (Cheung et al., 2023).

A plausible implication is that the “noise = information” paradigm is central to the future of high-bandwidth nonlinear spectroscopy across classical, quantum, and computational systems. As detector, light source, and computational capabilities advance, the methodological infrastructure of nonlinear covariance analysis is poised for widespread cross-disciplinary adoption (Tollerud et al., 2018, Sparapassi et al., 2021, Shen et al., 7 Aug 2025).