---
title: Inverse Random Source Problem
url: https://www.emergentmind.com/topics/inverse-random-source-problem
type: topic
---

# Inverse Random Source Problem

An inverse random source problem concerns the recovery of statistical characteristics of a random forcing term in a partial differential equation from observations of the random field it generates. In the literature, the unknown object may be the deterministic mean profile and variance of a white-noise source, the spatial intensity \(g^2(x)\) of a source driven by fractional Brownian motion, the principal symbol of a covariance operator for a microlocally isotropic generalized Gaussian field, or the temporal strength \(|f(t)|\) in stochastic heat and wave equations [1908.03666] [1608.02880] [1812.09646] [2002.08732] [2509.15928]. The topic sits at the intersection of inverse problems, stochastic analysis, spectral theory, microlocal analysis, and numerical regularization, and it is distinguished from deterministic inverse source problems by the fact that the source is not a fixed function but a random object whose identifiable features are typically its mean, variance, covariance, or principal microlocal statistics.

## 1. Conceptual scope and model classes

Across the cited works, the defining feature is that the source term is random while the governing equation is linear, so the solution inherits randomness in a form that can be analyzed through expectations, variances, covariances, or frequency-averaged quadratic observables. This leads to a shift in the inverse target: instead of reconstructing a realization of the source, one reconstructs a statistical descriptor such as \(f(x)\), \(|g(x)|\), \(g^2(x)\), a covariance principal symbol \(\phi(x)|\xi|^{-m}\), a micro-correlation strength matrix \(A(x)\), or the temporal strength \(|f(t)|\) [1810.03144] [2106.12642] [2503.10119] [2602.19559].

| Model class | Random source model | Inverse target |
|---|---|---|
| Time-fractional diffusion | \(f(x)h(t)+g(x)B_H(t)\) or \(f(x)h(t)+g(x)\dot{W}(t)\) | \(f(x)\), \(|g(x)|\), or \(g^2(x)\) |
| Acoustic, elastic, Maxwell, biharmonic scattering | White noise or microlocally isotropic Gaussian source | Mean/variance, covariance principal symbol, or micro-correlation strength |
| Stochastic evolution equations | \(g(x)f(t)\dot{W}(t)\) | \(|f(t)|\) |

A recurrent structural distinction is between two broad settings. In one, the randomness is low-dimensional and separable, for example \(g(x)f(t)\dot W(t)\) or \(f(x)h(t)+g(x)B_H(t)\), so modal statistics or boundary-flux statistics can be used to identify scalar source factors. In the other, the source is a distribution-valued Gaussian field whose covariance operator is a classical pseudodifferential operator, and the inverse target becomes the principal symbol of that covariance or relation operator. This suggests a unifying viewpoint: inverse random source problems are inverse problems for source statistics rather than source realizations.

## 2. Forward models and random source representations

The forward models span fractional diffusion, Helmholtz, elastic, Maxwell, biharmonic wave, biharmonic Schrödinger, and stochastic heat and wave equations. A representative time-fractional diffusion model is
\[
\partial_t^\alpha u(x,t) - A u(x,t) = f(x) h(t) + g(x) B_H(t),
\]
with Caputo derivative order \(0<\alpha\le 1\), homogeneous Dirichlet boundary conditions, and a one-dimensional fractional Brownian motion \(B_H\) with Hurst index \(H\in(0,1)\) [1908.03666]. A representative elastic model is the time-harmonic Navier equation with random source
\[
\mu\Delta \boldsymbol{u} + (\lambda+\mu) \nabla\nabla\cdot\boldsymbol{u} +\omega^2 \boldsymbol{u} = \boldsymbol{g}(x)+\boldsymbol{h}(x)\dot{W}_x
\quad \text{in } \mathbb{R}^2,
\]
with the Kupradze–Sommerfeld radiation condition [1608.02880]. A representative Maxwell model is
\[
\nabla \times E = i k H, \qquad \nabla \times H = i k E + J,
\]
with \(J\) a centered complex-valued Gaussian vector field whose covariance operator is a pseudo-differential operator [2002.08732]. In the stochastic heat and wave setting, the source takes the separable form
\[
g(x)f(t)\dot{W}(t),
\]
and the inverse problem concerns the recovery of \(|f(t)|\) from boundary flux data [2509.15928].

The random source models are equally diverse. Some works use additive spatial white noise, such as \(J(x)=\sigma(x)W_x\) in an inhomogeneous Maxwell medium or \(\boldsymbol f(x)=\boldsymbol \sigma(x)\dot{\boldsymbol W}_x\) in elastic scattering [2503.10119] [2511.00367]. Others use microlocally isotropic generalized Gaussian fields, characterized by covariance operators whose principal symbols have the form \(\phi(x)|\xi|^{-m}\), \(\mu(x)|\xi|^{-m}\), or \(A(x)|\xi|^{-2s}\) [1812.09646] [2106.12642] [2602.19559]. In the time-fractional diffusion literature, white noise and fractional Brownian motion both appear, and the well-posedness analysis hinges on Mittag–Leffler kernels, stochastic integrals with respect to \(B_H\), and the compatibility condition \(\alpha+H>1\) [1908.03666].

Forward well-posedness is established by several distinct mechanisms. Spectral decompositions yield mild solutions for time-fractional diffusion and stochastic evolution equations [1908.03666] [2509.15928]. Lippmann–Schwinger formulations and Fredholm theory are used for elastic and fractional Helmholtz scattering [1812.09646] [2602.19559]. Resolvent estimates and meromorphic continuation are central for biharmonic Schrödinger and inhomogeneous Maxwell equations [2412.16836] [2503.10119]. A common theme is that the random source is rough enough to require distributional or mild-solution frameworks, but the linearity of the PDE still permits explicit stochastic integral representations.

## 3. Observable data and inverse targets

The observable data depend on the governing equation and on whether ensemble information or a single realization is available. In time-fractional diffusion with source \(f(x)h(t)+g(x)B_H(t)\), the data are the expectation and covariance of the final-time modal coefficients,
\[
\mathbb{E}[u_k(T)], \qquad \mathrm{Cov}(u_k(T),u_l(T)),
\]
from which \(f(x)\) and \(|g(x)|\) are recovered [1908.03666]. In the white-noise-driven stochastic fractional diffusion equation
\[
\partial_t^\alpha u(x,t)+\mathcal{A} u(x,t)=f(x)h(t)+g(x)\dot{\mathbb W}(t),
\]
the goal is similarly to reconstruct \(f(x)\) and \(|g(x)|\) from statistics of \(u(x,T)\), and only \(|g|\) is identifiable because the white-noise term is symmetric in sign [1810.03144].

In elastic scattering with additive white noise, the inverse target is the mean source \(\boldsymbol g(x)\) and the variance \(\boldsymbol h^2(x)\), reconstructed from boundary measurements of the wave field at multiple frequencies through Fredholm integral equations of the first kind [1608.02880]. In the data-assisted Helmholtz formulation, the corresponding targets are again the mean \(g(x)\) and the variance \(h^2(x)\), using empirical expectations and variances of \(\Re u\) and \(\Im u\) on a measurement circle \(\Gamma_R\) [2303.16953].

A different data regime appears when the source is a microlocally isotropic Gaussian field. In two-dimensional elastic scattering, the principal symbol of the covariance operator is recovered from the frequency-averaged amplitude
\[
\lim_{Q\rightarrow\infty}\frac{1}{Q-1}\int_1^Q\omega^{m+1}|\bm{u}(x, \omega)|^2 d\omega,
\]
measured in an observation domain away from the source, and this can be done from a single realization with probability one [1812.09646]. In Maxwell’s equations, the high-frequency limit of \(\mathbb{E}[E(x,k)E(x,k)^*]\) determines the micro-correlation strength matrix \(A(x)\), while the diagonal entries can be identified from amplitude-only data \(\mathbb{E}[|E_j(x,k)|^2]\), and even almost surely from a single realization by frequency averaging [2002.08732]. The biharmonic wave equation and the fractional Helmholtz equation follow the same pattern: the inverse target is the strength \(\mu(x)\) or the principal symbol densities \(\mu^c,\mu^r\), extracted from frequency-averaged magnitudes or far-field patterns generated by a single realization [2106.12642] [2602.19559].

In stochastic heat and wave equations with source \(g(x)f(t)\dot W(t)\), the data are local boundary fluxes on a nonempty open subset, and the target is the temporal strength \(|f(t)|\) [2509.15928]. This setting makes explicit a common misconception: inverse random source problems do not always aim at recovering a full random source realization. In several rigorous formulations, the identifiable object is necessarily a magnitude, an intensity, or a covariance descriptor.

## 4. Analytical mechanisms for uniqueness

Uniqueness proofs are typically obtained by reducing the inverse problem to explicit modal equations, integral transforms, or convolution identities. In time-fractional diffusion, spectral decomposition gives
\[
\mathbb{E}[u_k(T)] = f_k\,C_k, \qquad
\mathrm{Cov}(u_k(T),u_l(T)) = g_k g_l\,C_{k,l},
\]
where \(C_k\) and \(C_{k,l}\) are deterministic coefficients built from Mittag–Leffler kernels and fractional Brownian motion covariance. Positivity and monotonicity properties of \(E_{\alpha,\alpha}(-x)\), together with lower bounds for these coefficients, imply that \(f_k\) and \(g_k g_l\) are uniquely determined by the data [1908.03666].

When the source is modeled as a microlocally isotropic Gaussian field, the uniqueness mechanism is potential-theoretic or microlocal. In elastic scattering, the frequency-averaged amplitude converges almost surely to
\[
a\int_{\mathbb R^2}\frac{1}{|x-y|}\phi(y)\,dy,
\]
and injectivity of this integral transform yields uniqueness of the principal symbol \(\phi\) [1812.09646]. In Maxwell’s equations, the high-frequency limit yields
\[
\frac{1}{(4\pi)^2}\int_{\mathbb{R}^3} \frac{A(y)}{|x-y|^2}\,dy,
\]
so recovery of \(A(x)\) reduces to inversion of a potential-type transform [2002.08732]. For the biharmonic wave equation, the limiting observables are
\[
\frac{1}{64\pi^2}\int_D \frac{\mu(\zeta)}{|x-\zeta|^2}\,d\zeta
\quad\text{in 3D,}
\qquad
\frac{1}{32\pi}\int_D \frac{\mu(\zeta)}{|x-\zeta|}\,d\zeta
\quad\text{in 2D,}
\]
and \(\mu\) is uniquely recoverable from the resulting data [2106.12642].

For stochastic heat and wave equations, the essential step is a representation formula for the integrated normal flux:
\[
-\int_0^t \frac{\partial u}{\partial\overrightarrow{\mathbf{n}}}(z,\tau)\,d\tau
=
\int_0^t f(\tau)\,G_z(t-\tau)\,dW(\tau),
\]
which implies
\[
\mathbb{V}\left[\int_0^t \frac{\partial u}{\partial \overrightarrow{\mathbf{n}}}(z,\tau)\,d\tau \right]
=
\int_0^t f^2(\tau)\, G_z^2(t-\tau)\,d\tau.
\]
The Titchmarsh convolution theorem then yields uniqueness of \(|f|\) from local boundary flux data [2509.15928].

The fractional Helmholtz setting combines Born linearization, Green-kernel asymptotics, and microlocal analysis. There, high-frequency averages of products such as
\[
\overline{u^\infty(\hat{x},k)}\,u^\infty(\hat{x},k+\tau)
\quad\text{and}\quad
u^\infty(-\hat{x},k)\,u^\infty(\hat{x},k+\tau)
\]
recover the Fourier transforms of the principal symbol densities \(\mu^c\) and \(\mu^r\) with probability one [2602.19559]. This shows that randomness does not preclude uniqueness; rather, it changes the identifiable object from a realization to a microlocal statistic.

## 5. Ill-posedness, instability, and stability theory

Uniqueness is not equivalent to stability, and this distinction is explicit throughout the literature. In the time-fractional diffusion problem driven by fractional Brownian motion, the inverse mappings \(f\mapsto \mathbb{E}[u(T)]\) and \(g^2\mapsto \mathrm{Var}(u(T))\) are described as severely ill-posed: the relevant coefficients decay like negative powers of \(\lambda_k\), so high-frequency modes amplify noise strongly [1908.03666]. The stochastic fractional diffusion problem with white noise exhibits the same pattern, with reconstruction formulas for \(f_n\) and \(g_mg_n\) becoming unstable as \(n\to\infty\) because the forward kernels decay in \(\lambda_n\) [1810.03144].

In elastic scattering with additive white noise, the inverse equations are Fredholm integral equations of the first kind. Their discretizations have rapidly decaying singular values, and the raw variance equations are observed to have singular values that decay exponentially, while certain variance-difference combinations decay more slowly and are more numerically stable [1608.02880]. This is a concrete manifestation of compact forward operators and severe statistical ill-posedness.

A different stability regime arises in high-frequency scattering. For the biharmonic Schrödinger equation, the recovery of the micro-correlation strength from correlation far-field data on a finite frequency interval satisfies increasing stability estimates of mixed Lipschitz-logarithmic type,
\[
\|h\|_{L^2(\mathbb{R}^d)}^2 \le C K^{d t + 2m}\,\varepsilon^2 + C K^{-2\beta},
\]
with the second term improving as the maximal frequency \(K\) grows [2412.16836]. In inhomogeneous Maxwell media with white-noise source, the strength \(\sigma(x)\) is recovered with a logarithmic stability estimate,
\[
\|\sigma\|_{L^2(\mathbb{R}^3)} \le C\big(-\log\varepsilon\big)^{-\frac{7+2s}{4s}},
\]
from random Dirichlet data at a fixed frequency [2503.10119]. These results show that the severity of instability depends on the PDE, the data type, and the asymptotic regime.

A persistent misunderstanding is that single-realization formulations must be intrinsically less stable than ensemble formulations. The literature suggests a more nuanced picture. Single-realization recovery is possible when frequency averaging or ergodic-type effects convert random quadratic observables into deterministic limits, but the resulting inverse maps may still be logarithmically unstable or require large frequency bands [1812.09646] [2002.08732] [2106.12642]. A plausible implication is that identifiability and practical robustness are governed by different mechanisms: ergodicity may provide uniqueness almost surely, while regularization remains necessary to control noise amplification.

## 6. Computational reconstruction and empirical performance

The numerical literature reflects the analytical structure of the inverse problem. In elastic waves with additive white noise, the explicit mild solution leads to Fredholm integral equations for the mean and variance, and the regularized Kaczmarz method is used to reconstruct \(\boldsymbol g\) and \(\boldsymbol h^2\) from multi-frequency boundary data [1608.02880]. In the stochastic Helmholtz equation, this idea is extended into a two-stage framework: the first stage uses a regularized block Kaczmarz scheme to generate initial approximations of \(g\) or \(h^2\), and the second stage treats the result as an image-to-image translation problem using PCA, DMD, U-Net, and pix2pix [2303.16953].

The data-assisted results are explicitly quantified. In the homogeneous medium case for mean reconstruction, the reported average relative \(L^1\) errors are approximately \(0.62\) for PCA, \(0.63\) for DMD, \(0.20\) for U-Net, and \(0.22\) for pix2pix; for variance reconstruction they are approximately \(0.90\), \(0.77\), \(0.28\), and \(0.30\), respectively [2303.16953]. In the inhomogeneous medium case, U-Net and pix2pix remain the best-performing methods, and the paper emphasizes that the second stage learns to compensate for sampling noise and modeling errors without explicit knowledge of the medium. This suggests that physics-based coarse inversion and data-driven refinement are complementary rather than competing paradigms.

For the time-fractional diffusion equation with Gaussian random source, the inverse problem is reduced to phase retrieval after proving that the variance of the Fourier transform of the boundary data determines the Fourier modulus of the diffusion coefficient. The PhaseLift method with random masks is then used to recover the coefficient from its Fourier modulus [2012.11042]. This is notable because it isolates a specific nonlinearity—the missing Fourier phase—and treats it with a convex lifting method rather than a bespoke PDE iteration.

The helium production-diffusion equation driven by fractional Brownian motion is treated numerically through spectral truncation. Expectations and covariances of the final-time data are estimated from \(P=1000\) sample paths, the relevant deterministic integrals are computed by composite Simpson’s rule, and the first \(N_1=30\) modes are retained in the reconstruction. The reported reconstructions are best for smooth \(g\), degrade for piecewise smooth \(g\), and are worst for discontinuous \(g\), while larger Hurst index \(H\) gives better results [2206.02359].

A recent three-dimensional elastic-wave method departs from multi-frequency iteration entirely. It reconstructs the variance matrix of a white-noise source from correlation boundary measurements at a single frequency by solving a family of \(3\times 3\) linear systems in Fourier space and then applying inverse Fourier transform. The method is non-iterative, requires only a single frequency, and comes with an explicit error estimate balancing Monte Carlo error, boundary data noise, conditioning, and high-frequency truncation [2511.00367]. In the numerical examples, the reported relative \(L^2(D)\) errors for the three variance components are \(3.5\%\), \(4.4\%\), and \(3.9\%\) [2511.00367].

## 7. Applications, limitations, and evolving directions

The application domains are broad. Fractional diffusion models are linked to anomalous diffusion in heterogeneous media, viscoelastic materials, and non-Markovian diffusion processes [1908.03666]. Elastic and Maxwell formulations address random body forces, random electric current densities, and inhomogeneous background media [1812.09646] [2002.08732] [2503.10119]. Biharmonic wave and biharmonic Schrödinger models connect to flexural-wave scattering and high-order dispersive systems [2106.12642] [2412.16836]. A plausible implication is that inverse random source methodology is less tied to a specific PDE than to a shared statistical structure: linear propagation, explicit or asymptotic Green representations, and recoverable second-order source information.

The main limitations are also consistent across models. Many results assume Gaussianity, microlocal isotropy, compact support, and either full or sufficiently rich observation geometry. Amplitude-only results often recover only diagonal covariance information, leaving off-diagonal entries open [2002.08732]. Some formulations recover only \(|g|\), \(g^2\), or \(|f|\) rather than a signed quantity, and this is not a defect of the analysis but a consequence of the symmetry of the driving noise [1908.03666] [2509.15928]. Other restrictions are technical: high-frequency limits, finite-interval analytic continuation, non-trapping conditions, or positivity and monotonicity of Mittag–Leffler kernels.

Several open directions recur explicitly in the cited works. These include inverse problems for more general random sources such as space-time random fields, relaxing data assumptions to partial observations or noisy measurements, extending uniqueness beyond \(\alpha\le 1\) in time-fractional diffusion, allowing correlated real and imaginary parts in complex Gaussian sources, treating random potentials or random media jointly with random sources, and developing robust regularization strategies tailored to fractional or high-order PDEs [1908.03666] [2002.08732] [2412.16836] [2602.19559]. The emerging picture is that inverse random source problems are evolving from uniqueness theory toward a more integrated theory combining microlocal identifiability, probabilistic ergodicity, finite-frequency stability, and computational inversion.

Source: https://www.emergentmind.com/topics/inverse-random-source-problem