WorksMathematics
Bayesian Time-Varying Spectral Estimation for LISA with P-Splines
Abstract
We model the evolutionary spectrum of a slowly time-varying process by a linear combination of products of time- and frequency-dependent B-splines, placing a Gaussian tensor-product penalty prior on the coefficients. The posterior distribution is obtained by updating the prior using a Whittle-type likelihood approximation of the Wilson-Daubechies-Meyer time-frequency coefficients. Separate regularisation in time and frequency allows non-stationary spectral structure to be inferred. An adaptive coarse-grained likelihood reduces the computational cost while preserving fine resolution around sharp spectral features. We first validate the method using repeated simulations of a locally stationary process with known spectrum, showing lower reconstruction errors and narrower posterior intervals as the sample size increases. We then apply it to a year-long simulated LISA data set with stationary instrumental single link spectra, an evolving time-delay interferometry (TDI) response, and a modulated Galactic foreground. We compare a flexible total power spectral density (PSD) model with models incorporating progressively stronger physical information and examine both continuous and gapped observations. Allowing time variation substantially improves reconstruction relative to a stationary model, while incorporating physical information about the TDI response further improves recovery, particularly near its moving spectral features. The results illustrate the trade-off between spectral flexibility, physical assumptions, and posterior uncertainty in non-stationary LISA noise estimation.
Introduction
The Laser Interferometer Space Antenna (LISA) [1] will observe overlapping gravitational-wave signals in instrumental noise and an unresolved Galactic foreground. Separating these contributions calls for a global fit of the sources and noise [2, 3, 4, 5]. The noise spectral model determines the weighting of the data in likelihood-based inference and must therefore be estimated alongside the signals.
LISA’s antenna motion modulates the anisotropic Galactic foreground annually [6, 7, 8], while changing arm lengths modify the time-delay interferometry (TDI) noise response and shift its spectral dips [9, 10, 11, 12]. Thermal evolution can affect optical-bench temperatures, electronics, and telescope alignment [13]. Changes in spacecraft orientation, ageing components, drag-free control and clock synchronisation, and environmental variations such as solar radiation pressure and cosmic-ray flux may introduce further non-stationarity on timescales from days to years [14, 15, 16, 17]. A single stationary power spectral density (PSD) can therefore misrepresent long observations. Noise mismodelling can affect source-parameter uncertainties even when estimators remain unbiased [18]. PSD uncertainty can propagate into source inference [19], and gaps and noise mismodelling can jointly affect LISA parameter uncertainties [20, 21]. Realistic noise and foreground modelling also matters for stochastic gravitational-wave background reconstruction [22, 23].
Flexible spectral estimation has a substantial statistical literature. Bayesian nonparametric approaches include flexible basis expansions and mixture representations, such as those based on Bernstein polynomials and B-splines [24, 25]. P-spline methods use a fixed B-spline basis with a roughness penalty to control the smoothness of the fitted function [26]. For example, Maturana-Russel and Meyer [27] developed a Bayesian P-spline approach to PSD estimation using the Whittle likelihood [28]. Penalised Whittle likelihoods are also used in frequentist spectral estimation [29]. These approaches provide flexible alternatives to fully specified parametric spectral models, while controlling the effective complexity of the estimated spectrum through priors or penalties.
In gravitational-wave data analysis, flexible spectral models are widely used for noise and foreground estimation. For example, BayesLine models the broadband noise spectrum with splines and narrow spectral features with Lorentzian components [30]. Spectral fitting is also an important component of LISA foreground estimation and global analysis [31, 32]. Recent LISA applications include flexible single-link B-spline noise models [33] and TDI-level spline models with reversible-jump inference of the number and location of knots [34]. Related approaches include other flexible noise models [35, 36], multivariate spectral estimation for gravitational-wave detectors [37, 38, 39, 40, 41], and weakly parametric stochastic-background inference [42]. These developments demonstrate the value of flexible spectral models for gravitational-wave inference, particularly when the underlying noise cannot be described adequately by a small number of physical parameters.
Modelling spectral evolution requires extending these ideas to non-stationary time series. Bayesian time-varying spectral estimation builds on the theory of evolutionary spectra [43] and locally stationary processes [44]. Existing approaches include segmentation-based models [45, 46], adaptive local Fourier bases [47, 48], direct time–frequency smoothing [49, 50], and locally stationary wavelet models [51]. Joint time–frequency inference of noise and stochastic backgrounds has also been explored using Hamiltonian Monte Carlo [52, 53]. Using this framework, Buscicchio et al. [54] investigated the reconstruction of the LISA Galactic foreground, including the effects of time segmentation and data gaps, and found strong evidence for non-stationarity. Related work by Baghi et al. [55] examines the statistical effects of averaging spectra in time and frequency, including the validity of independence approximations and optimal time segmentation for locally stationary processes.
The Wilson–Daubechies–Meyer (WDM) transform provides a time–frequency representation of the data using real-valued coefficients localised in both time and frequency [56, 57, 58, 59]. Other time-frequency representations include moving periodograms [60, 61] and multitaper methods [62]. Multiresolution time–frequency transforms have also been used in gravitational-wave burst searches [63]. Recent WDM work examines approximate diagonalisation of the noise covariance [64], relates the evolutionary spectrum to the full covariance of the WDM coefficients [65], and applies time–frequency covariance modelling to the cyclostationary Galactic foreground [8].
These developments motivate spectral models that can capture time variation over long observations without requiring a fully specified physical description of every noise contribution. For LISA, such models must accommodate the evolving Galactic foreground and the time-dependent features of the instrumental TDI response. Computational efficiency is also important given the duration and sampling rate of the observations. The combination of a localised time–frequency representation with a smooth, regularised spectral model offers a practical approach to these challenges.
In this work, we combine an approximate WDM Whittle likelihood, which neglects correlations between distinct WDM coefficients, with a tensor-product P-spline model for the logarithm of the evolutionary spectral density, with separate regularisation in time and frequency. Under suitable smoothness assumptions, this spectrum approximates the normalised WDM coefficient variances. We use the open-source WDM implementation described by Vajpeyi et al. [59], with other implementations available in [66, 67, 68, 69]. The model fits either total power or a smooth correction to a nominal instrumental spectrum, and pools neighbouring WDM powers to reduce likelihood cost. We assess reconstruction and posterior uncertainty in repeated simulations of the locally stationary, time-varying MA(1) process (LS2, described in Sec. III), and in continuous and gapped versions of a simulated LISA record, including comparison with a parametric model. We evaluate spectral recovery, not source-parameter recovery within a global fit.
Bayesian model
WDM likelihood
Let , , denote a real, zero-mean Gaussian time series with one-sided evolutionary spectral density and duration . We do not assume global stationarity. For even with , the Wilson–Daubechies–Meyer tile centres are
Here and . We exclude the DC and Nyquist frequency bins from the analysis. Following Vajpeyi et al. [59], we define the Fourier and WDM coefficients as
where indexes the Fourier frequencies .
For completeness, we give the interior Wilson basis, Meyer window, and phase factor below. We use the cosine-tapered Meyer window with , satisfying .
The interior basis functions satisfy . Since the transform is linear, the real WDM coefficients are jointly zero-mean Gaussian.
We model the one-sided evolutionary spectral density . If the process varies sufficiently slowly over the temporal extent of the WDM basis functions, and the spectrum is approximately constant across their frequency support, the midpoint approximation (see [70]) gives
The factor converts the coefficient variance to one-sided PSD units. The band-averaged approximation, which accounts for spectral variation across the WDM frequency window, is described in Appendix C [70].
Approximating the marginal WDM coefficient variances and neglecting correlations between distinct coefficients are separate assumptions. WDM orthogonality does not generally imply independent coefficients for coloured or non-stationary noise [57, 64, 70]. Under the midpoint and diagonal covariance approximations,
Here is the Kronecker delta, equal to one if and , and zero otherwise.
For a zero-mean Gaussian coefficient, its squared amplitude divided by its exact variance follows a distribution. Under the midpoint approximation, this gives
Neglecting correlations between distinct coefficients then gives the approximate WDM Whittle likelihood,
Here contains the analysed interior WDM cells, and the constant includes model-independent terms.
Tensor-product P-spline prior
We model the log evolutionary spectral density with a tensor-product P-spline [71, 72],
Here and are cubic B-spline basis functions and are unknown coefficients. We rescale each coordinate to and use repeated boundary knots, giving and interior knots, respectively. The basis dimensions and knot locations are fixed before sampling. Closer knots permit finer local structure.
The spline defines continuously in time and frequency. Under the midpoint approximation, its values at the WDM tile centres enter the likelihood in Eq. (8).
Let and denote the numbers of time and frequency points on the analysis grid. With , , , and , the spline model evaluated on this grid becomes
where the logarithm acts elementwise and
Substituting into Eq. (8) gives the approximate WDM likelihood in terms of the unknown spline coefficients. We complete the Bayesian model by assigning a Gaussian roughness prior to the column-stacked coefficient vector . The prior penalises curvature separately in time and frequency, as described below.
To favour smooth surfaces, we penalise curvature in time and frequency [73]. For the time basis, we define the normalised second-derivative penalty as
The frequency penalty is defined analogously. For , the combined penalty has the Kronecker-sum form [72]
Here , , is the identity matrix, and denotes the Kronecker product. The first term penalises temporal curvature for each frequency component, while the second penalises frequency curvature for each temporal component. The positive roughness scales and control the permitted curvature in time and frequency: smaller values give stronger smoothing, while larger values permit more curvature.
To make invertible, we replace its zero eigenvalues with the fixed precision and add a ridge to the remaining eigenvalues (Appendix § 1.1.5). The resulting regularised penalty is the precision matrix of the coefficient prior
We assign independent HalfNormal priors directly to the roughness scales,
Here 10 is the scale parameter. The effect of this prior depends on several factors, including the number and order of the basis functions, the knot locations, and the order of the penalty matrix. Although the basis order and penalty order are fixed across studies, each study explicitly specifies the number of basis functions and the knot layout.
Coarse-grained likelihood
We reduce the likelihood cost by grouping neighbouring WDM cells over which varies little and approximating the PSD as constant within each group. Let and denote partitions of the time and frequency indices, respectively,
Their Cartesian products define rectangular blocks on the WDM grid,
We evaluate the model at the mean time and frequency of the WDM tile centres in each group, writing
Let for cells used in the fit and 0 otherwise. For each non-empty block, we retain its mean observed power and the number of contributing WDM cells ,
Under the blockwise-constant PSD approximation, the coarse-grained WDM likelihood is
Empty blocks, such as those falling entirely within data gaps, contribute no likelihood term. To construct the blocks, we first form a smooth “pilot” estimate of the local spectrum and use it to identify regions where varies slowly. In frequency, this gives wider bins in smooth regions and finer bins around sharp spectral features. Time blocks are approximately fixed in width, with boundaries adjusted near gaps and missing data. The resulting partition is fixed before sampling, and the pilot estimate does not enter the likelihood.

Figure 1. Likelihood compression starting from the original WDM time-frequency grid. (a) The original WDM grid. (b) Missing data are excluded before pooling. Frequency bins are wider where the spectrum varies slowly and narrower around the sharp spectral feature (orange band), while time blocks remain approximately regular and stop at data gaps. (c) Each block contributes one likelihood term. The highlighted block is summarised by its mean power and contributing-cell count .
Posterior inference
We perform posterior inference using the No-U-Turn Sampler (NUTS) in NumPyro [74, 75], with JAX [76] for automatic differentiation. We sample the spline coefficients and smoothing scales jointly. These scales are and for the tensor-product model, and and for the orbit-informed model in Sec. IV B. Each chain is initialised from a penalised least-squares fit to the WDM powers. The implementation uses standard centred and non-centred parameterisations of hierarchical Gaussian priors to improve sampling efficiency [77, 78, 79]. Implementation details are provided in the tagged codebase.
The LS2 fits (Sec. III) use two chains with 500 warm-up and 500 retained draws per chain. The LISA fits (Sec. IV) use two sequential chains with 2200 warm-up and 4000 retained draws per chain.
We assess convergence using the rank-normalised potential scale reduction statistic [80], which compares agreement across chains, and bulk and tail effective sample sizes (ESS ), which measure sampling efficiency in the centre and tails of the posterior [81]. We also require zero divergent transitions, which indicate numerical integration difficulties, and an energy Bayesian fraction of missing information (E-BFMI) , which assesses exploration of the Hamiltonian energy distribution [82]. We additionally monitor maximum tree-depth saturation. We use these diagnostics to assess sampling reliability.
Evaluation criteria
For each posterior draw, we evaluate the model predictions on the time–frequency grid. At each , we summarise these values by their posterior medians and the 5th and 95th percentiles, and .
We assess reconstruction against a reference spectrum constructed from the generating model on the same grid. This reference approximates the normalised expected WDM power. Its construction is described with each study in Secs. III and IV A.
We compare the reconstruction with using the root-mean-square log error
where denotes an average over the time–frequency grid used for evaluation. We quantify posterior uncertainty using the logarithmic width of the 90% pointwise posterior interval
reported as a mean or median over the grid. We also report the fraction of 90% posterior intervals containing the reference
For the time-averaged spectrum , we average the grid predictions over the retained evaluation times. For the band-averaged spectrum , we average over the original frequency cells in a specified band . Each average is computed for each draw in linear PSD units before taking posterior medians and intervals. These descriptive summaries of an evolving spectrum do not assume global stationarity and can hide local spectral features. For a single LISA realisation, truth inclusion is descriptive. Repeated LS2 simulations allow assessment of interval calibration relative to the analytic reference.
We also assess how effectively the inferred spectrum whitens the data, without requiring an analytic reference spectrum. Using the WDM normalisation of Ref. [59], the standardised coefficients are expected to approximately follow a standard-normal marginal distribution,
We quantify agreement using the Anderson–Darling statistic , with the reference mean and variance fixed at zero and unity. Smaller values indicate closer agreement for a fixed sample size, with particular sensitivity to deviations in the tails. For independent samples, the asymptotic 5% critical value for the fixed distribution is . Values below this threshold do not reject the standard-normal hypothesis.
Because the spectrum is inferred from the same data and the coefficients may remain correlated, we interpret as a descriptive diagnostic of marginal agreement rather than a formal hypothesis test.
Controlled simulation study
Before applying the method to LISA data, we test it in a controlled setting where the true time-varying spectrum is known. This allows us to measure reconstruction accuracy and posterior uncertainty directly, and to examine how both change as the amount of data increases.
Following Tang et al. [61], we use the locally stationary time-varying MA(1) process (LS2)
where are independent standard-normal random variables. For this example, time is rescaled to , with
The function completes two cycles over the record, causing the spectral shape to vary continuously with time rather than simply rescaling a fixed spectrum. For comparison with the fitted WDM power, we use the locally stationary MA(1) spectrum evaluated at the WDM tile centres as the analytic reference (the midpoint approximation),
for , where s. The WDM grid uses frequency channels. For , the cubic time bases have functions, respectively. The cubic frequency basis has functions in every case. Interior knots are equally spaced in each coordinate, with repeated boundary knots.
We simulate 100 independent realisations at each . Figure 2 shows that reconstruction improves as the record length increases. From to , the median and median both decrease by about 65%.
Figure 2. LS2 recovery and performance. (a) Observed WDM log power, (b) the scaled analytic target , and (c) the recovered log PSD for an realisation. Panels (d)–(f) show , mean log-interval width , and two-chain NUTS wall time . Curves and shading show medians and interquartile ranges over 100 realisations at each record length. The metrics are defined in Sec. II E.
Across the 100 independent realisations at each sample size, the median fraction of evaluated time–frequency cells whose 90% posterior intervals contain the analytic reference remains close to 90%. This supports approximate interval calibration relative to the reference in this controlled LS2 study, while reconstruction errors and posterior interval widths decrease as the sample size increases.
Across all record lengths, Anderson–Darling statistics for ( to 1.11) are also comparable to those obtained using the ( to 1.07). All reported values remain below the nominal 5% critical value of 2.492, indicating no substantial departures from marginal standard normality in these summaries.
Application to simulated LISA data
Data generation
We generate instrumental noise with lisainstrument and PyTDI [83, 84], using ESA trailing orbits with unequal, time-dependent arms. The simulation includes test-mass acceleration (TM) and optical metrology system (OMS) noise in the second-generation Michelson variables . The noise PSDs of the individual instrument links are stationary, while the corresponding TDI channel PSDs vary with time due to the evolving arm lengths and transfer functions (see Appendix B).
We add a simulated unresolved Galactic foreground to the instrumental streams. Rather than simulating individual Galactic binaries, we generate a correlated Gaussian realisation from a prescribed strain PSD and a fixed anisotropic sky distribution, represented by a HEALPix map [85].
We use backgrounds [33] to propagate this foreground through the orbit-dependent second-generation TDI response. The resulting XYZ cross-spectral matrix is used to generate correlated Gaussian Fourier coefficients, which are transformed into time-domain data. Although the intrinsic strain PSD and sky distribution are fixed, the observed TDI foreground evolves with the detector response. Appendix B provides further details.
The injected Galactic strain PSD follows the analytic model of Karnesis et al. [31],
where . We adopt the one-year running-mean prescription of Karnesis et al. [31], with a source-resolution threshold of SNR , to specify the injected PSD parameters. This prescription does not imply that individual binaries are simulated or subtracted in our study. The sky map is normalised to unit total power [33]. The same spectral family is used in the parametric model . Table I summarises the numerical settings.
| Data generation | |
| Analysed duration | 364.09 days |
| Sampling interval / rate | 2 s / 0.5 Hz |
| Orbits | ESA trailing OEM, lisaorbits v2.3 |
| TM / OMS amplitudes | 2.4 fm s Hz / 7.9 pm Hz |
| TDI / units | / Hz, Hz |
| Galactic foreground | |
| Generation | Correlated Gaussian XYZ, fixed anisotropic sky |
| Strain spectrum | One-year Karnesis prescription [31] |
| 1.99526 mHz | |
| 0.31 mHz | |
| 3.38844 mHz | |
| 1.8 | |
| Analysis | |
| Frequency band | 0.1–100 mHz |
| Original WDM grid | 2046 × 3069 cells |
| WDM time spacing | 4.267 hours |
| Gap taper | Cosine taper, 1 hour per side |
Table I. LISA simulation and analysis settings.
We transform the Michelson variables to the orthonormal A/E/T basis [86],
where is an orthogonal transformation matrix. The reference A/E/T spectra are obtained by rotating the full XYZ spectral covariance and then taking its diagonal. We show A-channel results in the main text, as E gives similar reconstruction errors and the same model ordering.
LISA observations will contain scheduled and unscheduled interruptions [87]. Gaps treated via window functions can introduce spectral leakage and correlations through the observation window [20, 88]. The time localisation of the WDM transform allows us to exclude coefficients most affected by these boundaries.
We compare the continuous simulated LISA data set with a gapped version of the same realisation. The il-
lustrative gap schedule combines regular scheduled outages with random unscheduled interruptions,1 removing 4.00% of the time-domain samples.
Gaps are applied after2 TDI. We zero-fill missing samples and apply a cosine taper to the neighbouring data over one hour on each side. This brings the data smoothly to zero at the gap edges, reducing sharp jumps and the resulting spectral leakage. A WDM coefficient uses data beyond its tile centre, so leakage can also affect rows outside the gaps. We therefore exclude rows whose centres lie in an outage or within the one-hour taper plus four WDM time spacings of either edge. This four-row guard is about 17.1 hours on each side. We also exclude rows within four time spacings of the record boundaries. The boundary diagnostic in Figure 3 checks the retained coefficients. Appendix § 1.1.4 describes their coarse-graining.
Figure 3. A-channel WDM power near outages. For each frequency, observed power is summed in four-hour distance bins and divided by the summed reference power. The orange line and shading show the median and 5th–95th percentiles across frequency. The vertical dashed line marks the fixed boundary cut: the one-hour taper plus four WDM time spacings () from either outage edge. Rows beyond this cut are retained. Shading shows variation across frequencies, not posterior uncertainty.
Because the WDM guards extend beyond the gaps themselves, the gapped analysis removes 29.18% of the WDM rows, substantially more than the raw downtime.
On the retained rows, the RMSE between gapped and continuous data, divided by the continuous-coefficient RMSE, is less than 1% in A and E, with little change in integrated power. Some residual boundary effects remain (Figure 3).
To assess the effect of tapering, we also compared with zero-filled gaps without a taper. In the A channel, the RMSE from the continuous data, normalised by the analytic reference standard deviation at each WDM cell, is 0.013 with tapering and 0.090 without tapering over 0.1–100 mHz. The largest differences occur near the TDI response nulls, indicating that tapering suppresses leakage that remains even after the WDM guard is applied. A systematic investigation of alternative gap-handling strategies is beyond the scope of this study and is left for future work.
Both continuous and gapped records use the adaptive time–frequency coarse-graining described in Sec. II C. Further details of the likelihood partitions are given in Appendix § 1.1.4.
Three spectral models
We compare three models for the total channel power from instrumental noise and the unresolved Galactic foreground. They differ in how much physical information is supplied, from a flexible spline fit to a fully parametric instrument-and-foreground model. Table II summarises the three cases. All three models provide spectral predictions for the WDM likelihood. They differ in whether the spectrum is modelled directly or incorporates physically specified instrumental and Galactic contributions.
| Model | Physical information | Fitted quantity | A/E/T treatment | Runtime [h] | |
| None | Total-PSD tensor spline | 1850 | Separate A/E/T fits | ||
| Orbits + nominal TM/OMS noise | Smooth correction to instrumental PSD | 114 | Separate A/E/T fits | ||
| Orbits + nominal TM/OMS noise + Galactic model | 2 link-noise amplitudes + 5 Galactic parameters | 7 | Joint fit |
Table II. LISA spectral models and inference costs for a 364.09-day observation (15 728 640 samples per channel). The spline models fit A, E, and T independently, whereas the parametric model uses shared parameters across channels. The main reconstruction comparisons show A. Approximate runtimes include warm-up and sampling for two sequential chains. Spline runtimes refer to a single A-channel fit, whereas the parametric runtime covers the joint A/E/T fit. These timings are indicative: computational cost depends on likelihood evaluation and NUTS sampling efficiency, not only parameter count.
The spline models and are applied independently to A, E, and T, with separate parameters and likelihoods for each channel. The following comparisons focus on A.
The parametric model instead fits A/E/T jointly, using shared instrumental and Galactic parameters. We retain T only above 3 mHz, avoiding the low-frequency regime where cross-channel correlations can become significant relative to the T-channel PSD for unequal arms [33]. Above this cut, T supplies additional information on the shared instrumental noise amplitudes. All models use marginal-power likelihoods and neglect cross-channel covariance.
a. : flexible total spectrum. The agnostic model fits the total evolutionary PSD using the tensor-product P-spline of Eq. (9). It uses no orbital information, instrumental reference, or Galactic template. The midpoint approximation uses directly in the WDM likelihood. The spline must therefore capture both the foreground modulation and the moving TDI response features.
b. : orbit-informed spline. The orbit-informed model uses the known spacecraft trajectories and fixed nominal test-mass (TM) and optical metrology system (OMS) spectra to construct a time-dependent instrumental reference,
Here and are the nominal channel spectra obtained from the TDI response (Appendix B). We hold the band-averaged instrumental reference fixed during inference (Appendix C) and infer a smooth log correction,
Thus the dimensionless log correction is
Following [90, 91, 92], is a frequency P-spline and a tensor-product P-spline of the form in Eq. (9). The constraint at each frequency makes the time-average log correction. Their smoothing amplitudes have independent priors and , using the penalty construction of Appendix § 1.1.5. As approaches zero, the correction becomes frequency-only, while the total prediction retains the time dependence of . The correction can absorb both instrumental mismatch and Galactic foreground power, so this model estimates the total spectrum rather than separating its physical components.
c. : parametric instrument and foreground. The parametric model represents the total spectrum as the sum of TM noise, OMS noise, and a Galactic foreground. The instrumental spectral shapes and orbital responses are fixed. We infer seven parameters: the two link-noise amplitudes and , and five Galactic spectral parameters from Eq. (26). The PSD prediction on the WDM grid is
Here are the nominal link-noise amplitudes used to calculate . Since PSD scales quadratically with amplitude, the squared ratios rescale these nominal components to the fitted amplitudes. The instrumental contributions and Galactic term use the band-averaged approximation and are then averaged within likelihood blocks (Appendices C and § 1.1.4). The physical response models are described in Appendix B.
We assign independent Gaussian priors to the logarithms of the seven parameters. The Galactic priors are centred on the injected values from Sec. IV A, with standard deviations for , respectively. For the instrumental amplitudes, . The corresponding parameter posteriors are shown in Figure 8.
Figure 8. Seven-parameter posterior. Blue solid curves are continuous, orange dashed curves are gapped, and black lines mark injected values. The marginals use 8000 posterior draws per condition. Contours enclose 50% and 90% of the estimated two-dimensional posterior probability. Each axis shows the percentage departure from the injected value. The TM and OMS axes show physical amplitude departures relative to their nominal values . These constraints are conditional on the assumed instrumental and foreground model.
The displayed A-channel fits use for and for . The agnostic time knots are nonuniform, whereas the smaller orbit-informed time basis uses equally spaced knots. Frequency knots follow quantiles of a pilot-based measure of spectral variation, adapting the quantile approach of Maturana-Russel and Meyer [27]. The layouts are fixed before sampling. Preliminary fits informed these choices. Systematic optimisation of knot number and placement remains future work.
Results across models and observing conditions
We compare the three models on the same retained time–frequency points, including the TDI response dips. Table III summarises the reconstruction error and 90% truth-inclusion fraction, while Figure 4–Figure 7 show where the differences arise in frequency and time.
Figure 4. Continuous A-channel recovery. (a) Time-averaged PSD and (b) log ratio . Curves show posterior means and shading shows central 90% intervals after averaging each draw in linear PSD units. Grey regions mark TDI response dips.
| Condition | Band [mHz] | RMSE | Inc [%] | RMSE | Inc [%] | RMSE | Inc [%] | ||||
| Continuous | 0.1–3 | 0.0503 | 76.6 | 1.546 | 0.0515 | 50.5 | 1.333 | 0.0063 | 78.9 | 1.232 | 1.238 |
| 3–100 | 0.0916 | 85.7 | 0.896 | 0.0069 | 83.1 | 0.276 | 0.0004 | 100.0 | 0.326 | 0.296 | |
| 0.1–100 | 0.0907 | 85.4 | 1.025 | 0.0111 | 82.1 | 0.240 | 0.0011 | 99.4 | 0.316 | 0.289 | |
| Gapped | 0.1–3 | 0.0463 | 85.8 | 0.531 | 0.0534 | 56.4 | 0.633 | 0.0055 | 99.1 | 0.564 | 0.546 |
| 3–100 | 0.0933 | 85.7 | 1.134 | 0.0091 | 88.1 | 0.564 | 0.0007 | 100.0 | 0.617 | 0.565 | |
| 0.1–100 | 0.0923 | 85.7 | 1.254 | 0.0127 | 87.2 | 0.665 | 0.0012 | 100.0 | 0.710 | 0.668 |
Table III. A-channel reconstruction and whitening diagnostics for the three models under continuous and gapped observing conditions. The full band includes the TDI response dips. The Anderson–Darling statistic assesses marginal agreement with . The final column uses the analytic reference spectrum . For this single LISA realisation, Inc is descriptive rather than an ensemble coverage estimate.
Figure 7. Signed fractional error for continuous A-channel data. All three models use the same retained time–frequency points and colour scale, saturated outside . For display, errors are summarised within 26 equal time bins while retaining the original frequency resolution and TDI response dips.
a. : The agnostic spline captures the broad time variation in the spectrum but has difficulty following the sharp, moving TDI response features. These localised errors dominate the high-frequency reconstruction and are not always reflected by the posterior interval width. A sensitivity study using , 132, and 176 frequency basis functions shows decreasing reconstruction error with increasing basis size, although the improvement beyond our baseline is modest. A systematic optimisation of basis size is left for future work.
b. : Providing the time-dependent instrumental reference substantially improves recovery around the response features. For continuous data in the 3–100 mHz band, the RMSE decreases from 0.0916 for to 0.0069 for . At 0.1–3 mHz, however, the two models have similar RMSE, while has narrower intervals and a lower truth-inclusion fraction. The additional physical structure therefore increases posterior precision without a corresponding improvement in accuracy on this realisation.
c. : The parametric model gives the smallest reconstruction errors. This is expected under the matched simulation, since the foreground family, Galactic morphology, instrumental spectral shapes, and orbital response used by the model agree with those used to generate the data. The fit also uses the T channel. It therefore provides a matched physical benchmark rather than a test of robustness to model misspecification. Figure 8 shows the posterior distributions of the seven parameters under continuous and gapped observations. The instrumental amplitude departures are shown in physical amplitude units, rather than squared-amplitude (PSD) units.
Effect of gaps: Recovery on the retained data changes only modestly after introducing gaps. The full-band
Figure 5. Gapped A-channel recovery. (a) Posterior-median PSD for on its model-specific coarse grid, showing the moving TDI response dip. The inset enlarges the region near 60 mHz, with the expected continuous-response null shown in black. (b) Low-frequency (0.1–3 mHz) band-average ratio for the three models. Curves show posterior medians and shading gives central 90% intervals. Grey regions mark excluded times.
RMSE changes from 0.0907 to 0.0923 for , from 0.0111 to 0.0127 for , and from 0.0011 to 0.0012 for . The gapped fits generally have slightly wider posterior intervals, as shown in Figure 6b, reflecting the reduced amount of data available for inference. These comparisons do not assess recovery inside the excluded gap regions. Figure 7 shows that the remaining errors are strongly localised in time and frequency, particularly around the moving TDI response features. Accurate time- or band-averaged spectra therefore do not necessarily imply accurate recovery at every response dip.
Figure 6. A-channel reconstruction error and posterior width. (a) RMSE and (b) median() in fixed frequency bands. Solid curves show continuous data and dashed curves show gapped data. Grey regions mark TDI response dips. Bands in panel (a) show 5th–95th percentiles from paired resampling, not posterior intervals.
The Anderson–Darling statistics in Table III provide a complementary assessment of whitening performance. At low frequencies, all three models yield values comparable to those obtained using the . At higher frequencies, the agnostic model exhibits larger deviations, whereas the orbit-informed and parametric models remain close to the reference under both continuous and gapped conditions. All values remain below the nominal 5% critical value of .
Appendix A compares the two spline models with frequency-only fits. For continuous A-channel data, allowing time variation reduces the low-frequency RMSE of from 0.3978 to 0.0503, showing that time dependence is needed to recover the evolving spectrum.
Discussion and conclusions
We developed a Bayesian P-spline model for a smoothly evolving PSD, with separate regularisation in time and frequency and adaptive likelihood compression. In the controlled LS2 study, reconstruction errors and posterior interval widths decrease as more data are provided for the same evolving spectral pattern. This is evidence for recovery in that controlled setting, rather than a general asymptotic result.
The LISA study illustrates the trade-off between spectral flexibility and physical information. The agnostic spline recovers the broad time variation but has difficulty following sharp, moving TDI response features. Supplying a time-dependent instrumental reference substantially improves recovery around these features while retaining a flexible correction to the total power. At low frequencies, however, the orbit-informed model gives narrower intervals without a corresponding reduction in reconstruction error, and with lower truth inclusion. Additional physical structure can therefore reduce posterior uncertainty without improving accuracy.
The parametric model gives the smallest errors in this study, but the comparison is deliberately favourable to it. Its instrumental spectral shapes, orbital response, Galactic morphology, and foreground spectral family match those used to generate the data, and the fit also uses the T channel. It provides a matched physical benchmark rather than a test of robustness to model misspecification. The spline models also differ in basis size, regularisation, and likelihood partition, so the comparison does not isolate the effect of any single modelling choice. Preliminary fits informed their basis choices. Systematic optimisation of knot number and placement remains future work.
Gaps introduce a separate limitation. The illustrative outages remove 4.00% of the time-domain samples, but tapering and WDM guards exclude 29.18% of the WDM rows. The reported errors assess recovery only on the retained cells. Gap filling [93, 94], more localised windows [95], and explicit treatment of window-induced covariance [20, 88] could allow more data to be retained. Combining these approaches with the present spectral model remains to be investigated.
The evolutionary PSD enters the WDM likelihood through approximations to the marginal coefficient variances. For the physically specified LISA reference spectra, we account for variation across the WDM frequency window (Appendix C). The likelihood neglects correlations between distinct WDM coefficients [70].
The likelihood also neglects cross-channel covariance. Unequal, time-dependent arms and an anisotropic foreground can produce nonzero A/E/T cross-spectra. This information can be important for LISA stochastic-background inference [22], and a product of marginal likelihoods need not reproduce the posterior uncertainty of a full correlated likelihood [96]. The reported intervals are conditional on both covariance approximations.
A next step is to account for the cross-channel correlations neglected in the present analysis. We plan to extend our method to time-varying spectral-density matrices, allowing the A/E/T PSD and cross-spectra to be inferred jointly. The multivariate P-spline framework of Vajpeyi et al. [97] provides a starting point, although efficient likelihood construction and regularisation remain to be developed. A complementary direction is to model individual link noises and propagate them through the TDI response, as in Baghi et al. [33], to enforce consistency across channels and help separate instrumental contributions.
Data and Code availability
The spectral-estimation library is available as Log-PSplinePSD, version v0.2.0. The LISA-specific models and analysis scripts are available in the LISA study repository, version v0.1.1, archived at doi:10.5281/zenodo.23093373. The GitHub icons (☯) beside equations link to the corresponding source lines in these tagged releases. The study source release does not include the large data sets or posterior artifacts.
The simulated LISA time series, prepared continuous and gapped data sets, and posterior results for the three models are available on Zenodo at doi:10.5281/zenodo.23092878. The deposit includes a Jupyter notebook for loading the files and plotting WDM power, median posterior surfaces and PSD slices.
Acknowledgments
We thank the members of the LISA Noise Non-Stationarities Group (part of the “Deep analysis group” of the Distributed Data Processing Centre, DDPC) for helpful discussions. AV, PMR, and RM gratefully acknowledge support from the Marsden Fund Council grant MFP-UOA2531, funded by the New Zealand Government and managed by the Royal Society Te Apārangi. The work of PMR and RM was further supported by the CNRS International Research Project (IRP) OG-Science FR-NZ. GM acknowledges support from the ANR grant AAPG2024 PRC - GalaxyFIT. O. Burke acknowledges financial support from the Grant UKRI972 awarded via the UK Space Agency. This work was performed on the OzSTAR national facility at Swinburne University of Technology. The OzSTAR program receives funding in part from the Astronomy National Collaborative Research Infrastructure Strategy (NCRIS) allocation provided by the Australian Government, and from the Victorian Higher Education State Investment Fund (VHE-SIF) provided by the Victorian Government.
Stationary and time-varying contributions
We compare each full spline model with a separately fitted restriction: a frequency-only total PSD for and a frequency-only correction to the time-dependent reference for . We use the same evaluation cells and frequency bands as Figure 6. Each restricted fit uses its corresponding full fit’s data masks, frequency pooling, frequency knots, and sampler settings. The recorded convergence checks pass for all restricted fits. These comparisons describe reconstruction on this realisation, rather than evidence for selecting a model.
For continuous A-channel data, adding the time-varying component reduces low-band RMSE from 0.3730 to 0.0515 in , and from 0.3978 to 0.0503 in . The wider full-model intervals reflect the additional freedom. The agnostic fit also reduces high-band RMSE, from 0.5233 to 0.0916, but still struggles near response dips.
LISA response models and foreground simulation
We describe the instrumental reference, Galactic response, and foreground synthesis used in the LISA study. The response models are evaluated using backgrounds [33], with the ESA spacecraft trajectories specified in Table I.
Instrumental reference
We construct the instrumental reference by propagating the stationary test-mass (TM) and optical metrology system (OMS) link-noise spectra through the time-dependent TDI response.
For each noise component , denotes its stationary link-noise PSD. The TDI response maps this to the time-dependent A/E/T spectral matrix , whose diagonal gives the PSD of each channel . Summing the TM and OMS contributions gives the instrumental reference . Schematically,
Following Bayle and Hartwig [83], the nominal link-noise spectra are
where is the speed of light. The nominal amplitudes are given in Table I, with , , and .
The spacecraft trajectories determine the directed light-travel times between spacecraft. In the locally frozen response calculation, each delay is represented by the frequency-domain phase factor
We follow the second-generation TDI conventions of Tinto and Dhurandhar [11] and Bayle and Hartwig [83]. The transfer matrices map the link noises into , including the fixed simulator-specific unit conversion and discrete-time filtering.
Figure 9. A-channel component summaries and matched fit comparisons for (A) and (B) . Within each group, (i) shows posterior medians and central 90% intervals from the joint continuous-data fit: the dimensionless stationary multiplier in (A), and the geometric time-mean PSD in in (B). The geometric mean is computed for each draw over the full time grid and differs from the arithmetic average in the main results. (ii) Posterior means of the centred log correction in (A) and the log PSD after subtracting its time mean in (B). The agnostic summaries are a descriptive decomposition of one fitted tensor spline, with no separate component regularisation. (iii) RMSE and (iv) median log interval width , for time-varying and separately fitted frequency-only models on common evaluation cells. The orbit-informed frequency-only fit has a stationary correction to a time-dependent reference. The agnostic frequency-only fit has a stationary total PSD. The legends in (iii) identify the fits compared in (iii)–(iv). Solid curves are continuous and dashed curves are gapped.
Let denote the rotation defined in Eq. (27). For independent link noises with common spectrum , the corresponding A/E/T spectral matrix is
where denotes Hermitian conjugation (complex-conjugate transpose) and . The diagonal elements give the channel PSDs,
The instrumental reference is therefore
Although the link spectra are stationary, the changing delays shift the cancellations between TDI terms. Consequently, the channel spectra and their response nulls evolve over the orbit.
Galactic sky and response
We construct a fixed anisotropic sky distribution using the Galaxy model in backgrounds [33]. The map has HEALPix resolution (768 pixels) [85] and uses 400 radial integration points. The pixel powers are normalised so that , and their square roots are supplied to the response calculation. The map specifies the angular distribution, while in Eq. (26) specifies the intrinsic strain spectrum.
The backgrounds package propagates this sky distribution through the orbit-dependent second-generation TDI response. Writing the resulting XYZ response matrix as , the injected Galactic spectral matrix is
The response includes the factor converting fractional-frequency spectra to the TDI units of Table I. Rotating the full response matrix into A/E/T and taking its diagonal gives the channel response,
The sky distribution is fixed throughout the observation. Its observed modulation arises from the evolving detector response.
Gaussian foreground synthesis
We interpolate the Galactic response matrices linearly in time and log frequency, then multiply by . For each synthesis segment, the resulting spectral matrix is evaluated at the segment’s reference epoch. Its Cholesky factor is used to generate correlated complex Gaussian XYZ Fourier coefficients. Inverse Fourier transforms give real time-domain segments.
Each segment contains samples at the 2 s cadence, corresponding to approximately 12.14 days. The segments are multiplied by a sine window and combined using 50% overlap. The full XYZ spectral matrices are retained when generating the foreground, including cross-channel correlations. The foreground is added to the independently generated instrumental XYZ time series before rotation to A/E/T.
The total pointwise channel spectrum used to construct the analytic reference is
Appendix C describes how the normalised expected squared WDM coefficients approximate band-averaged versions of these physical PSDs under slow temporal evolution.
Band-averaged and midpoint PSD approximations
The LISA response models in Appendix B give pointwise evolutionary PSDs, not exact WDM coefficient variances. For an interior frequency channel, define
where and the squared cosine-Meyer window corresponding to Eq. (3b) is
Here , so has unit integral. DC and Nyquist channels, which require separate endpoint conventions, are excluded throughout our analysis.
Following Johnson et al. [70], if the process admits an oscillatory spectral representation whose complex amplitude varies slowly over the temporal extent of the WDM window, the normalised expected coefficient power is approximately the band-averaged evolutionary PSD,
This band-averaging relation assumes slow evolution over the temporal support of the WDM window; it is not an identity for arbitrary non-stationary processes [43, 70]. If also varies little across the frequency support of , the unit-integral normalisation allows the integral to be approximated by , giving the midpoint approximation in Eq. (5) [70 Eq. (39)]. For the same sampled interior window, Johnson et al.’s coefficients obey ; their Eq. (36) therefore gives half the band-averaged one-sided PSD in our convention.
Near sharp LISA TDI response nulls, the centre-frequency PSD can be unrepresentative of the power sampled by the finite WDM window. We therefore define the reference predictions by band averaging,
These frequency-window averages generally differ from the pointwise physical PSDs. We evaluate the integrals using 16-node Gauss–Legendre quadrature, with non-negative squared-window weights normalised to sum to one. The fixed enters , whereas fits a smooth midpoint surface directly.
For , the nominal instrumental components are band averaged in the same way, and the Galactic contribution is
The parametric model combines this frequency projection with likelihood-block averaging (Appendix § 1.1.4). Frequency projection improves the approximation to marginal WDM powers. It does not recover off-diagonal WDM covariance and is distinct from coarse-graining the observed WDM powers.
Likelihood coarse-graining
Section II C defines the coarse-grained likelihood. Here we summarise the partition used in the reported analyses and assess its impact in the controlled LS2 study. The tagged analysis code provides the full implementation of the partitioning procedure.
Partition construction. The partition is constructed before sampling from a smooth pilot estimate of the WDM power. The pilot determines only the block boundaries. All likelihood calculations use the original, unsmoothed WDM powers.
For the LISA analysis, we divide the data into 32 broad time intervals and compute the median log power at each frequency. A 31-frequency-cell running median is then used to obtain a smoother pilot spectrum, following the idea of a robust initial spectral estimate in Gupta and Cornish [98]. Starting from low frequency, neighbouring frequency cells are grouped while the pilot log power varies by no more than 0.25 within the bin, with a maximum width of 24 original frequency cells. This gives wider bins in smooth regions and finer bins around sharp spectral features.
The frequency bins are combined with groups of four neighbouring WDM time rows. For each rectangular block , the mean power and contributing-cell count are computed from the original WDM powers as defined in Eq. (17). The spline is evaluated at the mean time and frequency of the original cells defining the block. The implementation stores summed power and counts, giving the same likelihood as Eq. (18).
For the continuous A-channel data, the 3069 original frequency cells are reduced to 676 frequency bins. Together with the four-row time grouping, this corresponds to approximately an 18-fold reduction in the number of likelihood terms away from boundaries and missing-data regions. Figure 10 shows the frequency dependence of this partition alongside the time-averaged PSD. The partition responds to frequency variation in the pilot log power across all 32 time profiles, rather than to the level of the time-averaged PSD. The plotted quantity is the number of original frequency cells in each coarse frequency bin. For a fully retained four-row time block, the corresponding likelihood count is . Boundaries and exclusions can reduce the contributing-cell count.
Figure 10. Frequency coarse-graining for the continuous A-channel fit. (a) Time-averaged true PSD from Figure 4, with the same axis limits. (b) Number of original frequency cells per coarse frequency bin, , ranging from 1 to the imposed maximum of 24 cells. The likelihood also groups four time rows, so fully retained blocks have contributing WDM cells. The exact frequency-cell counts are not smoothed. One original frequency cell spans 32.55 μHz. The partition is fixed from the pilot estimate before sampling.
Model-specific coarse-graining. For the LS2 and fits, the PSD itself is approximated as constant within each block. For , each original power is first divided by the fixed instrumental reference before pooling. The blockwise-constant approximation therefore applies to the smooth correction rather than to the sharp instrumental response (Sec. IV B). Thus Eq. (18) uses mean reference-normalised powers and the predicted correction for this model. For , the deterministic model prediction is averaged over the original cells belonging to each block. Separate partitions are constructed for the continuous and gapped data. The joint A/E/T parametric fit uses a shared frequency partition.
Impact of coarse-graining. We assess the particular coarse-graining rule used in this work with paired LS2 simulations. Time is pooled in groups of four WDM rows, while the adaptive frequency bins span 1–5 original cells, with median width one.
Relative to the original likelihood, this partition gives a median 3.40× reduction in sampling time and changes the RMSE by +0.0067 nats, with a 95% bootstrap interval of [+0.0040, +0.0128] nats. The corresponding truth-inclusion fractions are 0.942 and 0.937.
These checks characterise the coarse-graining used in this analysis rather than coarse-graining in general. Here coarse-graining is used primarily to reduce the computational cost of the time-varying fits, and we do not attempt to optimise the partitioning rule. The trade-off between block size, computational cost, and reconstruction accuracy, together with alternative pilot estimates and adaptive partitioning strategies, is left for future work.
Penalty construction and regularisation
The regularised penalty in Eq. (13) follows Maturana-Russel and Meyer [27], Lang and Brezger [99], Boer [100]. The joint time–frequency penalty is
as defined in Eq. (12). Since and are symmetric, they admit the following eigendecompositions
where and are orthogonal matrices whose columns are the corresponding eigenvectors, and and are diagonal matrices containing the corresponding eigenvalues.
The Kronecker-sum penalty is diagonal in the basis
where and are the eigenvalues of the time and frequency penalties. Some are zero because the second-derivative penalties leave low-order modes unpenalised, making singular. We assign these modes a fixed weak precision and add a small ridge to the remaining modes,
where selects the zero-eigenvalue modes. We use and . The resulting is invertible and defines the proper Gaussian prior in Eq. (13).
References
References
- [1]P. Amaro-Seoane et al., Laser interferometer space antenna (2017), LISA mission proposal, arXiv:1702.00786 [astro-ph.IM].
- [2]T. B. Littenberg and N. J. Cornish, Physical Review D 107, 063004 (2023), arXiv:2301.03673 [gr-qc].
- [3]M. L. Katz, N. Karnesis, N. Korsakova, J. R. Gair, and N. Stergioulas, Physical Review D 111, 024060 (2025).
- [4]S. H. Strub, L. Ferraioli, C. Schmelzbach, S. C. Stähler, and D. Giardini, Global analysis of LISA data with galactic binaries and massive black hole binaries (2024), arXiv:2403.15318 [gr-qc].arxiv.org/abs/2403.15318
- [5]S. Deng, S. Babak, M. Le Jeune, S. Marsat, É. Plagnol, and A. Sartirana, Modular global-fit pipeline for LISA data analysis (2025), arXiv:2501.10277 [gr-qc].arxiv.org/abs/2501.10277
- [6]M. C. Digman and N. J. Cornish, The Astrophysical Journal 940, 10 (2022).
- [7]R. Buscicchio, A. Klein, V. Korol, F. Di Renzo, C. J. Moore, D. Gerosa, and A. Carzaniga, The European Physical Journal C 85, 887 (2025).
- [8]S. Mathur and N. J. Cornish, Time-frequency analysis for lisa: Galactic foreground modulation (2026), arXiv:2610.07977 [gr-qc].arxiv.org/abs/2610.07977
- [9]J. W. Armstrong, F. B. Estabrook, and M. Tinto, ApJ 527, 814 (1999).
- [10]M. Vallisneri, Phys. Rev. D 72, 042003 (2005), arXiv:gr-qc/0504145 [gr-qc].arxiv.org/abs/gr-qc/0504145
- [11]M. Tinto and S. V. Dhurandhar, Living Reviews in Relativity 24, 1 (2021), arXiv:gr-qc/0409034 [gr-qc].
- [12]J.-B. Bayle, Simulation and Data Analysis for LISA: Instrumental Modeling, Time-Delay Interferometry, Noise-Reduction Performance Study, and Discrimination of Transient Gravitational Signals, PhD thesis, Université de Paris (2019).
- [13]M. Armano, H. Audley, J. Baird, P. Binetruy, M. Born, D. Bortoluzzi, N. Brandt, E. Castelli, A. Cavalleri, A. Cesarini, A. M. Cruise, K. Danzmann, M. de Deus Silva, I. Diepholz, G. Dixon, R. Dolesi, L. Ferraioli, V. Ferroni, E. D. Fitzsimons, R. Flatscher, M. Freschi, A. García, R. Gerndt, L. Gesa, D. Giardini, F. Gibert, R. Giusteri, C. Grimani, J. Grzymisch, F. Guzman, I. Harrison, M.-S. Hartig, G. Heinzel, M. Hewitson, D. Hollington, D. Hoyland, M. Hueller, H. Inchauspé, Q. Jennrich, P. Jetzer, U. Johann, B. Johlander, N. Karnesis, B. Kaune, C. J. Killow, N. Korsakova, J. A. Lobo, L. Liu, J. P. López-Zaragoza, R. Maarschalkerweerd, D. Mance, V. Martín, L. Martin-Polo, F. Martin-Porqueras, J. Martino, P. W. McNamara, J. Mendes, L. Mendes, N. Meshksar, A. Monsky, M. Nofrarias, S. Paczkowski, M. Perreur-Lloyd, A. Petiteau, P. Pivato, E. Plagnol, P. Prat, U. Ragnit, J. Ramos-Castro, J. Reiche, J. A. Romera Perez, D. I. Robertson, H. Rozemeijer, F. Rivas, G. Russano, P. Sarra, A. Schleicher, J. Slutsky, C. F. Soupertа, T. J. Sumner, D. Texier, J. I. Thorpe, D. Vetrugno, S. Vitale, V. Wang, G. Wanner, H. Ward, P. J. Wass, W. J. Weber, L. Wissel, A. Wittchen, and P. Zweifel, Phys. Rev. Lett. 126, 131103 (2021).
- [14]M. Armano, H. Audley, G. Auger, J. T. Baird, P. Binetruy, M. Born, D. Bortoluzzi, N. Brandt, A. Bursi, M. Caleno, A. Cavalleri, A. Cesarini, M. Cruise, K. Danzmann, M. de Deus Silva, I. Diepholz, R. Dolesi, N. Dunbar, L. Ferraioli, V. Ferroni, E. D. Fitzsimons, R. Flatscher, M. Freschi, J. Gallegos, C. García Marrirrodriga, R. Gerndt, L. Gesa, F. Gibert, D. Giardini, R. Giusteri, C. Grimani, J. Grzymisch, I. Harrison, G. Heinzel, M. Hewitson, D. Hollington, M. Hueller, J. Huesler, H. Inchauspe, O. Jennrich, P. Jetzer, B. Johlander, N. Karnesis, B. Kaune, C. J. Killow, N. Korsakova, L. Lloro, L. Liu, J. P. López-Zaragoza, R. Maarschalkerweerd, S. Madsen, D. Mance, V. Martín, L. Martin-Polo, J. Martino, F. Martin-Porqueras, I. Mateos, P. W. McNamara, J. Mendes, L. Mendes, A. Moroni, M. Nofrarias, S. Paczkowski, M. Perreur-Lloyd, A. Petiteau, P. Pivato, E. Plagnol, P. Prat, U. Ragnit, J. Ramos-Castro, J. Reiche, J. A. Romera Perez, D. I. Robertson, H. Rozemeijer, F. Rivas, G. Russano, P. Sarra, A. Schleicher, J. Slutsky, C. Souperta, T. J. Sumner, D. Texier, J. I. Thorpe, C. Trenkel, D. Vetrugno, S. Vitale, G. Wanner, H. Ward, P. J. Wass, D. Wealthy, W. J. Weber, A. Wittchen, C. Zanoni, T. Ziegler, P. Zweifel, and LISA Pathfinder Collaboration, Phys. Rev. Lett. 118, 171101 (2017), arXiv:1702.04633 [astro-ph.IM].DOI
- [15]K. Danzmann and A. Rüdiger, Classical and Quantum Gravity 20, S1 (2003).
- [16]M. Pagone and C. Novara, arXiv e-prints, arXiv:2405.02339 (2024), arXiv:2405.02339 [astro-ph.IM].
- [17]F. Dimiccoli, R. Dolesi, M. Fabi, V. Ferroni, C. Grimani, M. Muratore, P. Sarr, M. Villani, and W. J. Weber, A&A 700, A102 (2025), arXiv:2411.18030 [astro-ph.IM].
- [18]O. Edy, A. Lundgren, and L. K. Nuttall, Physical Review D 103, 124061 (2021), arXiv:2101.07743 [astro-ph.IM].
- [19]C. Talbot and E. Thrane, Physical Review Research 2, 043298 (2020), arXiv:2006.05292 [astro-ph.IM].
- [20]O. Burke, S. Marsat, J. R. Gair, and M. L. Katz, Physical Review D 111, 124053 (2025), arXiv:2502.17426 [gr-qc].
- [21]J.-M. Yan and Z.-K. Guo, arXiv e-prints, arXiv:2609.37248 (2026), arXiv:2609.37248 [gr-qc].
- [22]J. Kume, M. Peloso, M. Pieroni, and A. Ricciardone, J. Cosmology Astropart. Phys. 2025, 030 (2025), arXiv:2410.10342 [gr-qc].
- [23]N. Karnesis, Q. Baghi, J.-B. Bayle, and N. Galanis, arXiv (2026), arXiv:2601.19741.
- [24]C. Kirch, M. C. Edwards, A. Meier, and R. Meyer, Bayesian Analysis 14, 1037 (2019).
- [25]M. C. Edwards, R. Meyer, and N. Christensen, Statistics and Computing 29, 67 (2019).
- [26]P. H. C. Eilers and B. D. Marx, Statistical Science 11, 89 (1996).
- [27]P. Maturana-Russel and R. Meyer, Computational Statistics 36, 2055 (2021).
- [28]P. Whittle, Journal of the Royal Statistical Society: Series B 15, 125 (1953).
- [29]Y. Pawitan and F. O’Sullivan, Journal of the American Statistical Association 89, 600 (1994).
- [30]T. B. Littenberg and N. J. Cornish, Physical Review D 91, 084031 (2015).
- [31]N. Karnesis, M. C. Babak, M. Pieroni, N. Cornish, and T. Littenberg, Physical Review D 104, 043019 (2021).
- [32]R. Rosati and T. B. Littenberg, Physical Review D 112, 084060 (2025).
- [33]Q. Baghi, N. Karnesis, J.-B. Bayle, M. Besançon, and H. Inchauspé, J. Cosmology Astropart. Phys. 2023, 066 (2023), arXiv:2302.12573 [gr-qc].
- [34]A. Santini, M. Muratore, J. Gair, and O. Hartwig, Phys. Rev. D 112, 084050 (2025), arXiv:2507.06300 [gr-qc].
- [35]N. Aimen, P. Maturana-Russel, A. Vajpeyi, N. Christensen, and R. Meyer, Phys. Rev. D 113, 024022 (2026).
- [36]C. Hoy, J. Bayley, O. Burke, and L. K. Nuttall, Noisy pipelines ii: Non-parametric noise inference for lisa (2026), arXiv:2611.03911 [gr-qc].arxiv.org/abs/2611.03911
- [37]A. Vajpeyi, R. Meyer, P. Maturana-Russel, and J. Liu, Multivariate bayesian p-spline estimation of spectral density matrices, with application to lisa tdi noise (2026), arXiv:2607.04833 [stat.ME].arxiv.org/abs/2607.04833
- [38]J. Liu, A. Vajpeyi, R. Meyer, K. Janssens, J. E. Lee, P. Maturana-Russel, N. Christensen, and Y. Liu, arXiv preprint 2409.13224 (2024).
- [39]J. Liu, A. Vajpeyi, R. Meyer, J. E. Lee, and P. Maturana-Russel, arXiv e-prints, arXiv:2608.22245 (2026), arXiv:2608.22245 [astro-ph.IM].
- [40]Y. Liu, R. Meyer, N. Christensen, J. E. Lee, J. Liu, P. Maturana-Russel, and A. Vajpeyi, Phys. Rev. D (2026), submitted, arXiv:2607.26619 [gr-qc].
- [41]O. Burke, J. Bayley, C. Hoy, and L. Nuttall, Noisy pipelines i: An end-to-end pipeline to perform multivariate noise inference for lisa (2026), arXiv:2610.28726 [gr-qc].arxiv.org/abs/2610.28726
- [42]F. Pozzoli, R. Buscicchio, C. J. Moore, F. Haardt, and A. Sesana, Physical Review D 109, 083029 (2024).
- [43]M. B. Priestley, Journal of the Royal Statistical Society: Series B 27, 204 (1965).
- [44]R. Dahlhaus, The Annals of Statistics 25, 1 (1997).
- [45]O. Rosen, S. Wood, and D. S. Stoffer, Journal of the American Statistical Association 107, 175 (2012).
- [46]O. Rosen, D. S. Stoffer, and S. Wood, Journal of the American Statistical Association 104, 249 (2009).
- [47]H. C. Ombao, J. A. Raz, von Sachs, and B. A. Malow, Journal of the American Statistical Association 96, 543 (2001).
- [48]H. Ombao, R. von Sachs, and W. Guo, Journal of the American Statistical Association 100, 519 (2005).
- [49]W. Guo, M. Dai, H. C. Ombao, and R. von Sachs, Journal of the American Statistical Association 98, 643 (2003).
- [50]P. J. Wolfe, S. J. Godsill, and W.-J. Ng, Journal of the Royal Statistical Society: Series B 66, 575 (2004).
- [51]G. P. Nason, R. von Sachs, and G. Kraissofant, Journal of the Royal Statistical Society: Series B 62, 271 (2000).
- [52]F. Pozzoli, R. Buscicchio, A. Klein, and D. Chirico, Bayesian inference with Hamiltonian Monte-Carlo for Astrophysical Stochastic background (2025), arXiv:2506.22542 [astro-ph.IM].arxiv.org/abs/2506.22542
- [53]R. M. Neal, in Handbook of Markov Chain Monte Carlo, edited by S. Brooks, A. Gelman, G. L. Jones, and X.-L. Meng (CRC Press, 2011) pp. 113–162.
- [54]R. Buscicchio, F. Pozzoli, D. Chirico, and A. Sesana, The first year of lisa galactic foreground (2025), arXiv:2511.03604 [astro-ph.IM].arxiv.org/abs/2511.03604
- [55]Q. Baghi, N. Karnesis, and J.-B. Bayle, Statistics of time and frequency-averaged spectra in gravitational-wave background searches (2026), arXiv:2602.12781 [gr-qc].arxiv.org/abs/2602.12781
- [56]V. Necula, S. Klimenko, and G. Mitselmakher, Journal of Physics: Conference Series 363, 012032 (2012).
- [57]N. J. Cornish, Physical Review D 102, 124038 (2020).
- [58]A. D. Johnson, K. Chatziioannou, and J. Summers, The WDM time-frequency transform in gravitational-wave data analysis I: Formalism (2026), arXiv:2606.21473 [gr-qc].arxiv.org/abs/2606.21473
- [59]A. Vajpeyi, G. Mentasti, Q. Baghi, O. Burke, and L. Speri, An explicit and differentiable Wilson-Daubchies-Meyer transform for gravitational-wave data analysis (2026), arXiv:2606.20269 [gr-qc].arxiv.org/abs/2606.20269
- [60]F. Häfner and C. Kirch, Journal of Time Series Analysis 38, 895 (2017).
- [61]Y. Yang, T. Kirch, J. E. Lee, and R. Meyer, Journal of the American Statistical Association , 1 (2026), arXiv:2303.11561 [stat.ME].
- [62]D. J. Thomson, Proceedings of the IEEE 70, 1055 (1982).
- [63]S. Chatterji, L. Blackburn, G. Martin, and E. Katsavounidis, Classical and Quantum Gravity 21, S1809 (2004).
- [64]N. J. Cornish, Physical Review D 114, 044035 (2026), arXiv:2511.10632 [gr-qc].
- [65]N. J. Cornish, Physical Review D 114, 064054 (2026), arXiv:2607.13168 [gr-qc].
- [66]C. J. Moore and T. Kinowski, WDM_GW_wavelets (2026).arxiv.org/abs/2611.17396
- [67]M. Digman and K. Breivik, WDMWaveletTransforms (2022).
- [68]N. Cornish, WDM_Transform.
- [69]W. M. Farr, WDMWavelets.jl (2025).
- [70]A. D. Johnson, K. Chatzioannou, and R. Rosati, arXiv e-prints , arXiv:2610.19002 (2026), arXiv:2610.19002 [gr-qc].
- [71]P. H. Eilers and B. D. Marx, Chemometrics and Intelligent Laboratory Systems 66, 159 (2003).
- [72]S. N. Wood, Generalized Additive Models: An Introduction with R, 2nd ed. (CRC Press, 2017).
- [73]S. N. Wood, Statistical Science 27, 985 (2017).
- [74]M. D. Hoffman, A. Gelman, et al., J. Mach. Learn. Res. 15, 1593 (2014).
- [75]D. Phan, N. Pradhan, and M. Jankowiak, Composable effects for flexible and acceleration-efficient probabilistic programming (2019), arXiv:1912.11554 [cs.PL].
- [76]J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Leary, et al., JAX: composable transformations of Python+NumPy programs (2018).
- [77]R. M. Neal, The Annals of Statistics 31, 705 (2003).
- [78]O. Papaspiliopoulos, G. O. Roberts, and M. Sköld, Statistical Science 22, 59 (2007).
- [79]M. Betancourt and M. Girolami, in Current Trends in Bayesian Methodology with Applications, edited by S. K. Upadhyay, U. Singh, D. K. Dey, and A. Loganathan (Chapman and Hall/CRC, 2015) pp. 79–101.
- [80]A. Gelman and D. B. Rubin, Statistical Science 7, 457 (1992).
- [81]A. Vehtari, A. Gelman, D. Simpson, B. Carpenter, and P.-C. Bürkner, Bayesian Analysis 16, 667 (2021).
- [82]M. Betancourt, arXiv e-prints , arXiv:1604.00695 (2016), arXiv:1604.00695 [stat.ME].
- [83]J.-B. Bayle and O. Hartwig, Physical Review D 107, 083019 (2023).
- [84]M. Staab, J.-B. Bayle, and O. Hartwig, PyTDI (2023).DOI
- [85]K. M. Górski, E. Hivon, A. J. Banday, B. D. Wandelt, F. K. Hansen, M. Reinecke, and M. Bartelmann, ApJ 622, 759 (2005), arXiv:astro-ph/0409513 [astro-ph].
- [86]T. A. Prince, M. Tinto, S. L. Larson, and J. W. Armstrong, Phys. Rev. D 66, 122002 (2002).
- [87]M. Colpi et al. (LISA), LISA Definition Study Report (2024), arXiv:2402.07571 [astro-ph.CO].arxiv.org/abs/2402.07571
- [88]O. Burke, F. Pozzoli, M. Muratore, and J. R. Gair, Mind the second gap: The impact of gaps on noise and signal parameter inference (2026), arXiv:2609.35150 [gr-qc].
- [89]O. Burke, M. Muratore, and G. Woan, Phys. Rev. Appl. 25, 034041 (2026).
- [90]D.-J. Lee and M. Durán, Statistical Modelling 11, 49 (2011).
- [91]D.-J. Lee, M. Durán, and P. H. C. Eilers, Computational Statistics & Data Analysis 61, 22 (2013).
- [92]C. Gu, Journal of Statistical Software 58, 1 (2014).
- [93]Q. Baghi, J. I. Thorpe, J. Slutsky, J. Baker, T. D. Canton, N. Korsakova, and N. Karnesis, Physical Review D 100, 022003 (2019).
- [94]N. Pearson and N. J. Cornish, Physical Review D 113, 064033 (2026), arXiv:2509.05479 [gr-qc].
- [95]N. Pearson, M. Çalışkan, S. Bini, and N. J. Cornish, A compact time-frequency representation for gravitational-wave data analysis (2026), arXiv:2609.05601 [gr-qc].arxiv.org/abs/2609.05601
- [96]M. Ribatet, D. Cooley, and A. C. Davison, Statistica Sinica 22, 813 (2012).
- [97]A. Vajpeyi, R. Meyer, P. Maturana-Russel, and J. Liu, Multivariate Bayesian P-spline estimation of spectral density matrices, with application to LISA TDI noise (2026), arXiv:2607.04833 [stat.ME].arxiv.org/abs/2607.04833
- [98]T. Gupta and N. J. Cornish, Phys. Rev. D 109, 064040 (2024), arXiv:2312.11808 [gr-qc].
- [99]S. Lang and A. Brezger, Journal of Computational and Graphical Statistics 13, 183 (2004).
- [100]M. P. Boer, Statistical Modelling 23, 465 (2023).








