Stoch-IDENT: SPDE Drift-Diffusion ID

• Stoch-IDENT is a framework for identifying stochastic partial differential equations by decomposing the problem into drift estimation on sample means and quadratic recovery for diffusion from residual statistics.
• It leverages sparse deterministic PDE discovery for the drift and quadratic sparse recovery for second-order moments, enabling reliable inference even under additive or multiplicative noise.
• The method provides theoretical guarantees through spectral richness conditions and employs tailored algorithms like Subspace Pursuit and Quadratic Subspace Pursuit for robust coefficient recovery.

Searching arXiv for Stoch-IDENT and closely related system-identification papers to ground the article. Stoch-IDENT is most specifically a framework for identifying stochastic partial differential equations (SPDEs) from sampled solution trajectories by decomposing the inference problem into a drift component and a diffusion component. In its most explicit usage, the method targets SPDEs driven by a time-dependent Wiener process, accommodates both additive and multiplicative noise, and combines sparse deterministic PDE discovery on sample means with quadratic sparse recovery on second-order residual statistics (Cui et al., 26 Aug 2025). The associated theory connects identifiability to the spectral richness of the solution’s mean and covariance, and treats linear constant-coefficient, parabolic, hyperbolic, and stochastic Schrödinger settings within a unified identification perspective (Cui et al., 26 Aug 2025).

1. Terminology and scope

In the most precise contemporary sense, Stoch-IDENT refers to the method introduced in “Stoch-IDENT: New Method and Mathematical Analysis for Identifying SPDEs from Data” (Cui et al., 26 Aug 2025). There, the unknown equation is written in the template form

$$du = \sum_{k=1}^K a_k F_k\,dt + \sum_{j=1}^J b_j G_j\,dW(t),$$

where $F_k$ are candidate drift features, $G_j$ are candidate diffusion features, and the coefficient vectors are assumed sparse. The identified model is

$$du = \sum_{k=1}^K \widehat a_k F_k\,dt + \sum_{j=1}^J \widehat b_j G_j\,dW(t).$$

The label is, however, not unique across the stochastic-identification literature. It has also been used for meta-state-space learning, where a deterministic meta-state parameterizes the full state probability distribution of a stochastic dynamical system (Beintema et al., 2023); for online identification of stochastic continuous-time Wiener models from sampled data via stochastic approximation (Abdalmoaty et al., 2024); and for exact moment-based structural identifiability analysis of partially observed near-linear SDEs (Browning et al., 25 Mar 2025). A broader historical backdrop is provided by stochastic subspace identification, which studies valid finite-data realization, asymptotics, and $\mathcal H_2$/$\mathcal H_\infty$ model error bounds (Li et al., 2012). This suggests that “Stoch-IDENT” is best read contextually: in PDE discovery it denotes a specific SPDE-identification pipeline, while elsewhere it may denote a wider class of stochastic system-identification methodologies.

2. SPDE formulation and the drift–diffusion split

The defining construction of Stoch-IDENT is the separation of the unknown SPDE into a mean-identifiable drift and a covariance-identifiable diffusion (Cui et al., 26 Aug 2025). Observations consist of sampled solution trajectories

$U_n(t_i,x_m)=u_n(t_i,x_m)$

on a space-time grid

$$\Gamma = \{(t_i,x_m): i=0,\dots,I,\; m=0,\dots,M\}.$$

The target equation may be written in either Itô or Stratonovich form; the theoretical analysis also uses

$$du = \sum_{k=1}^K a_kF_k\,dt + \sum_{j=1}^J b_jG_j\circ dW(t).$$

The first step uses the fact that the stochastic integral has mean zero. Taking expectation yields

$$d\mathbb E[u] = \sum_{k=1}^K a_k \mathbb E[F_k]\,dt.$$

Hence the drift can be estimated from averaged trajectories by reducing the SPDE problem to deterministic PDE identification on the sample mean. After a drift estimate $F_k$0 is obtained, a residual is formed over each time increment,

$F_k$1

and second-order information in $F_k$2 is used to recover diffusion coefficients.

This decomposition is the central conceptual distinction of Stoch-IDENT. Drift recovery is linear in the unknown coefficients after expectation, whereas diffusion recovery is intrinsically quadratic because the second moments of the stochastic integral depend on pairwise products $F_k$3 rather than linearly on $F_k$4 itself (Cui et al., 26 Aug 2025). For that reason, the method is designed not merely to estimate a squared diffusion amplitude, but to recover the diffusion operator or term up to the equivalence allowed by second-order statistics.

The framework is stated to handle linear and nonlinear high-order SPDEs, additive and multiplicative noise, and time-dependent Wiener forcing (Cui et al., 26 Aug 2025). The experimental set includes stochastic transport, KdV, Burgers, nonlinear Schrödinger, Allen–Cahn, KPZ, and stochastic heat equations with multiplicative or mixed noise.

3. Identifiability theory

Theoretical identifiability in Stoch-IDENT is formulated through Fourier-domain structure and spectral richness of the solution (Cui et al., 26 Aug 2025). For linear constant-coefficient problems on the torus,

$F_k$5

with

$F_k$6

the Fourier transform converts the SPDE into modewise stochastic evolution. In the multiplicative case the ratio form

$F_k$7

makes explicit how drift and diffusion enter with different statistical signatures.

For drift identifiability, the crucial set is

$F_k$8

If $F_k$9 is sufficiently large and $G_j$0 is not contained in an algebraic hypersurface of degree $G_j$1 consisting only of even-order or only odd-order terms, then the drift coefficients $G_j$2 are uniquely determined by the solution at two time instants for sufficiently small $G_j$3 (Cui et al., 26 Aug 2025). The condition encodes the requirement that the initial data have enough nonzero and sufficiently generic Fourier modes.

Diffusion identifiability is subtler. Once drift is known, second-order quantities such as

$G_j$4

become polynomial expressions in the pairwise products $G_j$5. The resulting inverse problem is quadratic, so the diffusion coefficients are identifiable only up to an equivalence class from solution statistics at two time instants (Cui et al., 26 Aug 2025). The paper states explicitly that if the stochastic convolution or Brownian increment were known pathwise, diffusion would be uniquely identifiable in the classical sense; Stoch-IDENT does not assume such pathwise access.

A related theoretical strand studies approximation of the trajectory data space. For admissible parabolic operators, there exists a linear space $G_j$6 with

$G_j$7

such that

$G_j$8

Hyperbolic problems admit only polynomial-dimension approximation, and stochastic Schrödinger equations exhibit a split behavior: the mean trajectory admits logarithmic-dimensional approximation, whereas the full solution trajectory admits only polynomial-dimensional approximation (Cui et al., 26 Aug 2025). These results generalize deterministic PDE-identification intuition to the stochastic setting and indicate that parabolic SPDEs are structurally more compressible than hyperbolic ones.

4. Estimation algorithms

Drift estimation is performed by a sample-mean generalization of Robust-IDENT (Cui et al., 26 Aug 2025). Starting from the integral form over each time step,

$G_j$9

a left Riemann or Euler–Maruyama approximation gives

$$du = \sum_{k=1}^K \widehat a_k F_k\,dt + \sum_{j=1}^J \widehat b_j G_j\,dW(t).$$0

Expectation removes the stochastic term:

$$du = \sum_{k=1}^K \widehat a_k F_k\,dt + \sum_{j=1}^J \widehat b_j G_j\,dW(t).$$1

This yields a sparse linear regression

$$du = \sum_{k=1}^K \widehat a_k F_k\,dt + \sum_{j=1}^J \widehat b_j G_j\,dW(t).$$2

followed by the constrained problem

$$du = \sum_{k=1}^K \widehat a_k F_k\,dt + \sum_{j=1}^J \widehat b_j G_j\,dW(t).$$3

The support is recovered by Subspace Pursuit with the standard expand–shrink cycle: column normalization, correlation-based support initialization, residual-driven support expansion, least-squares refit on the expanded support, shrinkage to the $$du = \sum_{k=1}^K \widehat a_k F_k\,dt + \sum_{j=1}^J \widehat b_j G_j\,dW(t).$$4 largest coefficients, and repetition until the residual ceases to decrease.

Diffusion estimation is the main algorithmic novelty. After drift recovery, Stoch-IDENT forms quadratic measurements from expected squared residuals and solves

$$du = \sum_{k=1}^K \widehat a_k F_k\,dt + \sum_{j=1}^J \widehat b_j G_j\,dW(t).$$5

where $$du = \sum_{k=1}^K \widehat a_k F_k\,dt + \sum_{j=1}^J \widehat b_j G_j\,dW(t).$$6 contains diffusion-feature products and $$du = \sum_{k=1}^K \widehat a_k F_k\,dt + \sum_{j=1}^J \widehat b_j G_j\,dW(t).$$7 is derived from residual second moments. This is a sparse regression problem with quadratic measurements. The proposed solver, Quadratic Subspace Pursuit (QSP), adapts the SP logic to nonlinear quadratic sensing (Cui et al., 26 Aug 2025).

At iteration $$du = \sum_{k=1}^K \widehat a_k F_k\,dt + \sum_{j=1}^J \widehat b_j G_j\,dW(t).$$8, for each candidate index $$du = \sum_{k=1}^K \widehat a_k F_k\,dt + \sum_{j=1}^J \widehat b_j G_j\,dW(t).$$9 outside the current support, QSP evaluates the one-variable fit

$\mathcal H_2$0

Indices with the smallest such errors are added to form an expanded support. Then a nonlinear refit solves

$\mathcal H_2$1

The support is shrunk back to the $\mathcal H_2$2 largest coefficients, the nonlinear regression is repeated on the reduced support, and the procedure stops when the residual no longer improves. The reported implementation uses nonlinear conjugate gradient with Hager–Zhang updates, and QSP typically converges in about $\mathcal H_2$3–$\mathcal H_2$4 iterations (Cui et al., 26 Aug 2025).

5. Sample-average approximation, model selection, and statistical diagnostics

Because the exact expectations in the drift and diffusion objectives are unavailable, Stoch-IDENT uses sample-average approximation (SAA):

$\mathcal H_2$5

Analogous empirical averages replace the quadratic diffusion statistics. Under a local-minimizer assumption for a $\mathcal H_2$6-sparse solution $\mathcal H_2$7, the sampled drift problem converges almost surely to the population problem:

$\mathcal H_2$8

A similar result is stated for the diffusion SAA problem (Cui et al., 26 Aug 2025).

The paper also supplies score bounds linking coefficient error to objective value. For drift,

$\mathcal H_2$9

For diffusion,

$\mathcal H_\infty$0

These inequalities justify choosing the candidate model with the smallest score and make explicit that diffusion error propagates through drift error as well (Cui et al., 26 Aug 2025).

A practical diagnostic is an additive-noise detection shortcut. Space-averaged drift residuals are tested for normality with the D’Agostino–Pearson test, and the pathwise $\mathcal H_\infty$1-values are aggregated by Stouffer’s method:

$\mathcal H_\infty$2

If the residuals appear Gaussian, the method treats the noise as additive and estimates only a scalar noise amplitude $\mathcal H_\infty$3, bypassing the full quadratic diffusion-recovery stage (Cui et al., 26 Aug 2025). A common misconception is therefore that Stoch-IDENT always requires QSP; in fact, the paper explicitly provides this shortcut for cases that appear statistically consistent with additive noise.

6. Numerical behavior and relation to adjacent identification frameworks

The empirical study evaluates precision, recall, accuracy, F1 score, relative in-coefficient error, and relative out-coefficient error across a range of SPDEs (Cui et al., 26 Aug 2025). Several recurring patterns are reported. More trajectories improve both support recovery and coefficient accuracy. Pure additive noise is the easiest regime and mixed noise the hardest. Drift is generally easier to identify than diffusion. Correct SPDE structure can be recovered even from coarse grids, although diffusion coefficients are more sensitive to numerical differentiation error. In stochastic Allen–Cahn, the model recovered from $\mathcal H_\infty$4 trajectories reproduces the stripe pattern of the true dynamics. In stochastic KPZ, random initial conditions help drift identification but do not help diffusion much in the additive-noise case. For stochastic nonlinear Schrödinger dynamics, correct feature recovery improves with sample size, and the method handles complex-valued systems by separating real and imaginary parts (Cui et al., 26 Aug 2025).

Relative to adjacent literature, Stoch-IDENT occupies a distinct point in the stochastic-identification landscape. “Meta-State-Space Learning: An Identification Approach for Stochastic Dynamical Systems” replaces the random hidden state by a deterministic meta-state that parameterizes the full state PDF and learns deterministic meta-dynamics with an ANN and mixture-density output map (Beintema et al., 2023). “Online Identification of Stochastic Continuous-Time Wiener Models Using Sampled Data” instead develops a recursive output-error predictor and stochastic approximation scheme for continuous-time Wiener systems, emphasizing robustness to misspecified disturbance spectra (Abdalmoaty et al., 2024). “Exact identifiability analysis for a class of partially observed near-linear stochastic differential equation models” studies structural identifiability of SDEs through exact moment recurrences and input-output equations involving observed moments (Browning et al., 25 Mar 2025). In PDE discovery, “PDE Identification Using Noise Adaptive Differentiation in Strong Form (S-IDENT)” addresses noisy strong-form deterministic PDE identification with Savitzky–Golay differentiation, SURE-based window selection, Subspace Pursuit, trimming, and reduction-in-residual model selection (He et al., 30 Jun 2026). Stoch-IDENT differs from S-IDENT in that its central inverse problem is genuinely stochastic: mean dynamics identify drift, covariance identifies diffusion, and diffusion recovery is quadratic rather than linear (Cui et al., 26 Aug 2025).

Historically, the broader field also includes stochastic subspace identification, where finite-sample validity, asymptotic normality of covariance and block-Hankel estimates, and $\mathcal H_\infty$5/$\mathcal H_\infty$6 error bounds are central (Li et al., 2012). That contrast is instructive. Subspace-identification work focuses on innovations-form state-space realization and asymptotic model error; Stoch-IDENT, in its SPDE sense, focuses on sparse recovery of differential operators from ensembles of stochastic trajectories. The shared theme is identifiability under stochastic forcing, but the mathematical objects, statistics, and reconstruction targets differ substantially.

Overall, Stoch-IDENT denotes a technically specific program: infer the drift of an SPDE from the sample mean, infer the diffusion from quadratic residual structure, and analyze identifiability through spectral richness of the solution. Its main theoretical message is that stochasticity does not merely perturb deterministic PDE discovery; it changes the inverse problem qualitatively, especially on the diffusion side. Its main algorithmic message is that SPDE discovery can be organized as a coupled linear-and-quadratic sparse recovery problem with distinct statistical signatures for drift and diffusion (Cui et al., 26 Aug 2025).