Time-Domain Stochastic Reconstruction

• Time-domain stochastic reconstruction is a framework that recovers latent dynamics and state-dependent noise by directly estimating drift and diffusion from time-resolved observations.
• Methods include nonparametric kernel-based likelihood, Kramers–Moyal moment closure, and online updating techniques to robustly retrieve stochastic differential equation coefficients.
• Applications span dynamic imaging (CT and USCT), turbulence modeling, and speech enhancement, highlighting its versatility in addressing inverse problems in various fields.

Time-domain stochastic reconstruction denotes a family of methods that reconstruct latent dynamics, fields, or missing signal content directly from time-resolved observations while preserving stochastic structure. In the cited literature, this includes inferring drift and diffusion functions in Itô-Langevin models from discretely sampled paths, estimating Kramers–Moyal coefficients from conditional moments, reconstructing dynamic objects in continuous time for tomography, recovering missing turbulent trajectories with conditional diffusion models, and solving stochastic wave inverse problems in the time domain (Ohkubo, 2011, Mohan et al., 2024, Li et al., 2024). The common feature is that the unknown is represented as a stochastic process or stochastic field in time, and the reconstruction is carried out without reducing the problem to a purely deterministic static inverse problem.

1. Foundational formulations

At its classical core, time-domain stochastic reconstruction is the problem of recovering the state-dependent coefficients of a stochastic differential equation from a single discretely observed trajectory. One formulation assumes

$dX_t \;=\; f(X_t)\,dt \;+\; g(X_t)\,dW_t,$

where $f(x)$ is the drift function, $g(x)>0$ is the diffusion coefficient, and $W_t$ is standard Brownian motion. For small sampling intervals, the Euler–Maruyama approximation gives an approximately Gaussian transition density, and this leads to a path log-likelihood in which both the local mean increment and the local variance depend on $f$ and $g$ (Ohkubo, 2011).

A closely related formulation uses Kramers–Moyal coefficients. For a continuous Markov process $X(t)\in\mathbb R$, the coefficients

$$D^{(n)}(x) \;=\; \lim_{\tau\to 0}\frac{1}{n!\tau}\,\langle [X(t+\tau)-X(t)]^n \mid X(t)=x\rangle$$

are obtained from small-lag conditional moments. In practice, one estimates $K^{(n)}(x,\tau)$ for $\tau=k\Delta t$ and extrapolates linearly to $f(x)$0. If the process is Markov, then its probability distribution is governed by the Kramers-Moyal equation, so reconstruction of these coefficients amounts to reconstruction of the governing stochastic dynamics (Nikakhtar et al., 2023).

The same conceptual structure reappears in large-scale streamed settings. A scalar stochastic process $f(x)$1 may be modeled by the Itô-Langevin equation

$f(x)$2

with Gaussian white noise $f(x)$3, or equivalently by the Fokker–Planck equation for $f(x)$4. The first two Kramers–Moyal coefficients define the drift and diffusion,

$f(x)$5

and finite-$f(x)$6 conditional moments are then extrapolated to $f(x)$7 (Davis, 2023).

This suggests that the phrase does not identify a single algorithm. Rather, it refers to a class of reconstructions in which time-resolved stochastic laws are estimated directly from time-domain data, with the model class ranging from low-dimensional SDEs to distributed wave equations and neural implicit fields.

2. Nonparametric reconstruction of stochastic differential equations

One major branch is nonparametric inference of $f(x)$8 and $f(x)$9 from discrete path data. In Ohkubo’s formulation, the maximum-likelihood problem is localized by a smooth kernel $g(x)>0$0. At a query point $g(x)>0$1, weighted averages provide estimators

$g(x)>0$2

with $g(x)>0$3 and $g(x)>0$4. Because the constant-drift-and-diffusion approximation is accurate only if $g(x)>0$5 is very small, the method further expands both $g(x)>0$6 and $g(x)>0$7 up to second order about each evaluation point and uses a local linearization method, yielding a Gaussian transition density with analytic mean $g(x)>0$8 and variance $g(x)>0$9 (Ohkubo, 2011).

The localized log-likelihood

$W_t$0

is then maximized independently at each grid point. The reconstructed coefficients are $W_t$1 and $W_t$2. Ohkubo tests the method on

$W_t$3

with $W_t$4 and $W_t$5, and reports that the local-linearization scheme accurately recovers both the double-well drift and the state-dependent noise amplitude, whereas the simple constant-drift estimator shows large fluctuations (Ohkubo, 2011).

A distinct nonparametric route is the moments method. Here the conditional moment $W_t$6 is approximated by a polynomial in $W_t$7,

$W_t$8

and the coefficients $W_t$9 are determined from a linear system involving global moments $f$0 and mixed moments $f$1. Linear fits $f$2 then yield

$f$3

The method recovers the Ornstein–Uhlenbeck case exactly with $f$4, reconstructs polynomial drift and diffusion in a noisy genetic model very accurately with $f$5, and can use $f$6 conditional moments to extract jump parameters in a jump-diffusion example (Nikakhtar et al., 2023).

The two approaches differ in how locality is imposed. Kernel-based localized likelihood emphasizes spatial localization in state space and handles irregular sampling through the dependence of $f$7 and $f$8 on each $f$9. The moments method trades local conditional-density estimation for global moment identities, and its performance depends strongly on the polynomial order $g$0, the conditioning of the moment matrix $g$1, and the availability of well-sampled tails. A plausible implication is that the choice between them is governed by whether the dominant difficulty is sparse temporal sampling or accurate recovery of polynomial conditional moments.

3. Streaming and large-scale conditional-moment reconstruction

Davis reformulates kernel-based conditional-moment estimation for streamed datasets through incremental, online, updating statistics. With sampling times $g$2, constant $g$3, a grid of state points $g$4, and a kernel $g$5, the method maintains, for each lag index $g$6 and grid point $g$7, a cumulative weight $g$8, raw conditional moments $g$9, and a conditional variance accumulator $X(t)\in\mathbb R$0. When a new sample $X(t)\in\mathbb R$1 arrives, these quantities are updated by exact one-step recursions rather than recomputing sums over the full history (Davis, 2023).

The core updates are

$X(t)\in\mathbb R$2

$X(t)\in\mathbb R$3

and

$X(t)\in\mathbb R$4

followed by $X(t)\in\mathbb R$5. Drift and diffusion can then be extracted by linear regression in $X(t)\in\mathbb R$6 when $X(t)\in\mathbb R$7 is small (Davis, 2023).

The computational distinction between offline kernel-based regression and online kernel-based regression is explicit. Offline KBR stores the entire time series and re-loops over all $X(t)\in\mathbb R$8 samples for each $X(t)\in\mathbb R$9, giving $$D^{(n)}(x) \;=\; \lim_{\tau\to 0}\frac{1}{n!\tau}\,\langle [X(t+\tau)-X(t)]^n \mid X(t)=x\rangle$$0 time and $$D^{(n)}(x) \;=\; \lim_{\tau\to 0}\frac{1}{n!\tau}\,\langle [X(t+\tau)-X(t)]^n \mid X(t)=x\rangle$$1 memory. Online OKBR stores only $$D^{(n)}(x) \;=\; \lim_{\tau\to 0}\frac{1}{n!\tau}\,\langle [X(t+\tau)-X(t)]^n \mid X(t)=x\rangle$$2 accumulators and updates once per new sample, so memory remains constant at $$D^{(n)}(x) \;=\; \lim_{\tau\to 0}\frac{1}{n!\tau}\,\langle [X(t+\tau)-X(t)]^n \mid X(t)=x\rangle$$3. In the reported benchmarks, time scales linearly with $$D^{(n)}(x) \;=\; \lim_{\tau\to 0}\frac{1}{n!\tau}\,\langle [X(t+\tau)-X(t)]^n \mid X(t)=x\rangle$$4 for both methods at fixed $$D^{(n)}(x) \;=\; \lim_{\tau\to 0}\frac{1}{n!\tau}\,\langle [X(t+\tau)-X(t)]^n \mid X(t)=x\rangle$$5, while memory grows proportionally to $$D^{(n)}(x) \;=\; \lim_{\tau\to 0}\frac{1}{n!\tau}\,\langle [X(t+\tau)-X(t)]^n \mid X(t)=x\rangle$$6 for KBR but remains constant at approximately $$D^{(n)}(x) \;=\; \lim_{\tau\to 0}\frac{1}{n!\tau}\,\langle [X(t+\tau)-X(t)]^n \mid X(t)=x\rangle$$7 in the examples for OKBR, enabling $$D^{(n)}(x) \;=\; \lim_{\tau\to 0}\frac{1}{n!\tau}\,\langle [X(t+\tau)-X(t)]^n \mid X(t)=x\rangle$$8 up to $$D^{(n)}(x) \;=\; \lim_{\tau\to 0}\frac{1}{n!\tau}\,\langle [X(t+\tau)-X(t)]^n \mid X(t)=x\rangle$$9 (Davis, 2023).

The numerical examples emphasize what constant-memory estimation changes in practice. For the Ornstein–Uhlenbeck process with $K^{(n)}(x,\tau)$0, $K^{(n)}(x,\tau)$1, KBR and OKBR are identical for $K^{(n)}(x,\tau)$2 at $K^{(n)}(x,\tau)$3, but OKBR with $K^{(n)}(x,\tau)$4 cleanly recovers the coefficients over the full range. For the tri-stable toy model $K^{(n)}(x,\tau)$5, $K^{(n)}(x,\tau)$6, KBR at $K^{(n)}(x,\tau)$7 fails to resolve the weak attractor at $K^{(n)}(x,\tau)$8, whereas OKBR at $K^{(n)}(x,\tau)$9 resolves the drift zero-crossing at $\tau=k\Delta t$0, revealing the metastable state. For multiplicative and correlated noise, a parametric fit $\tau=k\Delta t$1 improves from $\tau=k\Delta t$2, $\tau=k\Delta t$3, $\tau=k\Delta t$4, $\tau=k\Delta t$5 with KBR at $\tau=k\Delta t$6 to $\tau=k\Delta t$7, $\tau=k\Delta t$8, $\tau=k\Delta t$9, $f(x)$00 with OKBR at $f(x)$01. On an empirical turbulence dataset of $f(x)$02 hot-wire velocity increments at $f(x)$03, OKBR reproduces previously published $f(x)$04 and $f(x)$05 (Davis, 2023).

The stated limitations are also structural. The method is currently one-dimensional in $f(x)$06; extension to multi-dimensional $f(x)$07 is straightforward but carries $f(x)$08 accumulators. Only the first two moments are treated, and online formulas for higher moments such as skewness and kurtosis are described as nontrivial. Jump-diffusion processes would benefit from higher-order online estimates. At the same time, the method is described as mathematically exact, single-pass, constant-memory, and amenable to real-time streaming applications and memory-constrained embedded systems (Davis, 2023).

4. Time-domain inverse problems in imaging, wave propagation, and turbulence fields

In dynamic X-ray CT, time-domain stochastic reconstruction appears as stochastic optimization over a continuous time-space object rather than as direct inference of SDE coefficients. Mohan et al. formulate a 4D inverse problem for a continuous density $f(x)$09, approximate it with a distributed implicit neural representation $f(x)$10, and define ray-sampled projections

$f(x)$11

where $f(x)$12. The DINR network uses input normalization to $f(x)$13, a random Fourier embedding $f(x)$14, and an $f(x)$15-layer MLP with Swish nonlinearities; typical choices are $f(x)$16 and $f(x)$17. Training uses distributed stochastic optimization: each of $f(x)$18 GPUs samples a small random set $f(x)$19 of projection indices, computes a local loss, synchronizes gradients, averages them, and performs the same Adam update. Because each GPU holds only the MLP and the activations for a few hundred rays, memory remains at $f(x)$20 per GPU instead of $f(x)$21 for a dense 4D voxel grid (Mohan et al., 2024).

The temporal rationale is explicit. Conventional 4DCT groups projections into static time frames; if the frame uses many views, motion blur appears, and if it uses few views, limited-angle streaks emerge. DINR instead treats each projection at its exact acquisition time $f(x)$22 and randomly interleaves views across epochs. The reported consequence is near-single-projection temporal resolution, approximately $f(x)$23 per frame in the experiments, with high fidelity. On the LLNL D4DCT MPM simulations, DINR-512 achieves the best PSNR, approximately $f(x)$24 to $f(x)$25 over PINR or TIMBIR, and strong scaling on Lassen from $f(x)$26 to $f(x)$27 V100 GPUs is near-linear (Mohan et al., 2024).

An ultrasound analogue is time-domain stochastic full-waveform inversion with randomized super-shots. In ring-array USCT, the physical model is the constant-density, lossless acoustic wave equation

$f(x)$28

with unknown speed of sound $f(x)$29. The reconstruction introduces super-shots, in which multiple transmitters fire simultaneously, each assigned an independent Rademacher weight $f(x)$30, together with multiple stochastic ensembles $f(x)$31. Gradients are computed by the adjoint-state method, and the update is stochastic gradient descent with an inexact line search: $f(x)$32 For the reported implementation, $f(x)$33, the pixel size is $f(x)$34, $f(x)$35, and $f(x)$36 (Forte, 13 Aug 2025).

The results are deliberately qualified. For two clinical slices, image contrast and anatomy recovery are comparable between frequency-domain deterministic FWI and time-domain stochastic FWI, but the time-domain method is significantly slower: frequency-domain deterministic FWI requires $f(x)$37–$f(x)$38 on one GPU, whereas time-domain stochastic FWI with $f(x)$39 and $f(x)$40 iterations requires at least $f(x)$41 on the same hardware. A single super-shot is described as inadequate, and the paper ultimately recommends FD or FD-stochastic FWI for routine breast USCT (Forte, 13 Aug 2025).

A different continuum setting is the time-domain stochastic acoustic wave equation with a random source. Here the unknown $f(x)$42 solves

$f(x)$43

on $f(x)$44, with spatial white noise $f(x)$45. The direct problem is truncated by a time-domain perfectly matched layer, while the white noise is approximated by a piecewise constant field

$f(x)$46

The approximation error satisfies

$f(x)$47

and the full PML error estimate combines this with an exponentially small truncation term in the layer thickness. The inverse random source problem is then posed from time-domain boundary correlations, and a logarithmic stability estimate is derived for the recovery of $f(x)$48, emphasizing severe ill-posedness (Guo et al., 29 Jun 2025).

In turbulence modeling, the time-domain object may itself be a stochastic velocity field. Antoni, Kürpick, Lindner, Marheineke, and Wegener construct an inhomogeneous Gaussian random field from $f(x)$49-$f(x)$50 data by combining a moving average in time with a Fourier-type integral in space,

$f(x)$51

The model is shown to have mean zero, to satisfy the prescribed $f(x)$52 and $f(x)$53 constraints asymptotically, to be asymptotically incompressible, and to obey an inhomogeneous ergodicity result under local space-time averaging (Antoni et al., 2023).

5. Missing data, unresolved scales, and stochastic signal restoration

A Bayesian route treats time-domain and frequency-domain representations jointly as latent variables. In BRFP, the unknown time series $f(x)$54 and its discrete Fourier transform $f(x)$55 are embedded in a joint Gaussian state $f(x)$56, where $f(x)$57 and $f(x)$58. Time-domain observations and frequency-domain observations enter through a linear Gaussian observation model $f(x)$59, and Gaussian conditioning yields a closed-form posterior

$f(x)$60

Missing data are handled by binary selection matrices $f(x)$61 and $f(x)$62, while uncertainty is read directly from the posterior covariance. Reported experiments include recovery from only $f(x)$63 of the entries in time and frequency on synthetic data, ECG with $f(x)$64 irregular time sampling plus noise $f(x)$65, audio with up to $f(x)$66 missingness, and interferometric imaging from $f(x)$67 of Fourier visibilities, where the ALMA reconstruction matches CLEAN with $f(x)$68 while also providing pixel-wise standard deviations (Tobar et al., 2020).

Conditional diffusion models provide a different answer to missing-data reconstruction. For a discretized time series $f(x)$69, one splits the indices into measured $f(x)$70 and gap $f(x)$71, and reconstructs by sampling from $f(x)$72. The forward process noises only the gap values,

$f(x)$73

while the reverse model

$f(x)$74

is implemented with a U-Net conditioned on the observed context. Reconstruction quality is then assessed with structure functions $f(x)$75, flatness $f(x)$76, increment PDFs, acceleration PDFs, and pointwise MSE. The method is reported to reproduce fat tails up to $f(x)$77–$f(x)$78, to handle arbitrary gap shapes, and to achieve $f(x)$79–$f(x)$80 lower MSE for large gaps than Gaussian-process regression, while retaining highly non-Gaussian and intermittent statistics in both DNS tracers and NOAA Global Drifter Program trajectories (Li et al., 2024).

A multiscale version conditions on a wavelet coarse-grained trajectory $f(x)$81 and reconstructs the unresolved fine-scale component $f(x)$82 with a guided reverse-diffusion process. Conditioning enters through the gradient of a coarse-scale likelihood or, equivalently, through Diffusion Posterior Sampling. The reported diagnostics are high-order structure functions $f(x)$83, flatness $f(x)$84, local exponents $f(x)$85, and cross-scale correlations

$f(x)$86

The reconstructed trajectories recover these quantities in excellent agreement with DNS and preserve broad, non-Gaussian conditional PDFs of acceleration that Gaussian-process reconstructions in wavelet representation suppress (Wang et al., 4 Jun 2026).

Time-domain stochastic restoration also appears in speech enhancement. SRTNet decomposes the clean waveform $f(x)$87 and noisy waveform $f(x)$88 into an initial deterministic estimate $f(x)$89, a clean residual $f(x)$90, and a noisy residual $f(x)$91. A conditional diffusion process is then defined on $f(x)$92,

$f(x)$93

and the reverse process predicts the effective noise with a time-domain diffusion network conditioned on $f(x)$94. On VoiceBank-DEMAND, the reported scores are PESQ $f(x)$95, CSIG $f(x)$96, CBAK $f(x)$97, and COVL $f(x)$98, exceeding the cited CDiffuSE large baseline; on WSJ0+MUSAN, the method loses only $f(x)$99 PESQ out of domain. Training for $g(x)>0$00 steps requires $g(x)>0$01 on two RTX 3090 GPUs versus $g(x)>0$02 for the cited baseline (Qiu et al., 2022).

6. Latent stochastic dynamics, evaluation criteria, and recurring limitations

Several recent methods reconstruct dynamical systems from sparse or partial observations by explicitly learning latent stochastic dynamics. Stochastic NODE-DMD models a low-dimensional latent state $g(x)>0$03 with

$g(x)>0$04

combining a linear DMD drift, a neural nonlinear residual, and Gaussian diffusion. The field is represented continuously as

$g(x)>0$05

where the mode functions are given by an implicit neural network $g(x)>0$06. Across four benchmarks, including a synthetic setting and three physics-based flows, the method surpasses a baseline in reconstruction accuracy when trained from only $g(x)>0$07 observation density, aligns learned modes and continuous-time eigenvalues with ground truth, and learns a calibrated distribution over latent dynamics on datasets with multiple realizations (Kim et al., 25 Nov 2025).

DPDSR takes a discrete-time VAE formulation with explicit latent noise: $g(x)>0$08 A state encoder estimates $g(x)>0$09 from the observed sequence, a second encoder approximates the posterior over $g(x)>0$10, and the ELBO combines a data reconstruction term, a latent-state reconstruction term, and a KL penalty. Training uses teacher forcing every $g(x)>0$11 steps. The reported qualitative regimes are sharply separated: small $g(x)>0$12 leads to learned deterministic dynamics that become chaotic and posterior KL approaching zero, whereas large $g(x)>0$13 leads to stable fixed points or limit cycles and large posterior KL. Across six test problems, evaluated with distribution distance $g(x)>0$14, spectral distance $g(x)>0$15, $g(x)>0$16-step prediction error $g(x)>0$17, and $g(x)>0$18 for ECG, DPDSR achieves the lowest cumulative score (Sip et al., 1 Oct 2025).

Across the literature, the diagnostic toolkit is correspondingly broad. SDE-reconstruction papers emphasize recovered drift and diffusion functions, conditional moments, and finite-$g(x)>0$19 extrapolation. Large-scale streamed methods emphasize memory complexity, time complexity, and the ability to recover weak attractors or tail structure from $g(x)>0$20 up to $g(x)>0$21. Tomographic and imaging methods use PSNR, SSIM, contrast-to-noise behavior, strong-scaling measurements, and wall-clock time. Bayesian and diffusion-based signal reconstructions use posterior variances, KL-divergence, $g(x)>0$22, pointwise MSE, flatness, high-order structure functions, acceleration PDFs, and cross-scale correlations (Davis, 2023, Mohan et al., 2024, Tobar et al., 2020, Li et al., 2024, Wang et al., 4 Jun 2026).

Several limitations recur. Multidimensional drift–diffusion estimation is straightforward in principle but incurs $g(x)>0$23 accumulators in the online kernel framework. Higher-order online moments remain nontrivial. Diffusion-model methods require substantial compute, for example roughly $g(x)>0$24 denoising steps per reconstruction in the gappy-trajectory setting and many GPU-hours in both dynamic CT and ultrasound. Hyperparameters such as kernel bandwidth, Fourier-embedding scales $g(x)>0$25, network size, forcing interval $g(x)>0$26, or diffusion schedules materially affect the smoothness-detail tradeoff. Time-domain formulations also do not guarantee superior throughput: in ring-array USCT, time-domain stochastic inversion matches frequency-domain image quality only at an order-of-magnitude larger wall-clock time, and temporal resolution in continuous-time CT cannot exceed the projection-acquisition cadence (Davis, 2023, Forte, 13 Aug 2025, Mohan et al., 2024).

A common misconception is that stochastic reconstruction is synonymous with diffusion-model sampling. The cited work shows a much wider landscape: kernel-based maximum likelihood, Kramers–Moyal moment closure, online conditional-moment updates, Gaussian-linear Bayesian conditioning, stochastic full-waveform inversion with randomized source encoding, stochastic integral models for turbulence, conditional diffusion models, and latent stochastic dynamical systems all fall under the same general aim of reconstructing time-domain stochastic structure from incomplete or indirect observations.