Neural-Operator Priors

• Neural-operator priors are inductive biases that constrain mappings between function spaces using embedded transforms, discretization‐agnostic architectures, or explicit stochastic formulations.
• They enhance performance by integrating hard, architectural, or probabilistic priors that enforce consistency, physical structure, and improved error bounds in operator learning.
• Applications in precision learning, spectral super-resolution, and inverse problems demonstrate that operator priors reduce approximation errors and boost stability across resolutions.

Neural-operator priors are inductive biases, constraints, or probability measures that act on mappings between function spaces rather than only on finite-dimensional parameter vectors. In the surveyed literature, the term encompasses at least three technically distinct constructions. First, a prior may be imposed by embedding a mathematically known transform directly into the computational graph, so that only the unknown residual map is learned; this is the operator-prior view articulated under the name precision learning (Maier et al., 2017). Second, a prior may be implicit in the architecture class itself, through discretization-agnostic integral operators, quadrature-consistent attention, spectral truncation, or geometry-based locality, all of which constrain the admissible continuum map $NO_\theta:\mathcal{F}\to\mathcal{G}$ (Berner et al., 12 Jun 2025). Third, a prior may be explicit and probabilistic, as in infinite-width neural operators that induce Gaussian-process priors on function spaces, or in inverse problems where a neural operator is coupled to a separate generative or physics prior (Souza et al., 19 Oct 2025). Taken together, these lines of work define neural-operator priors as a broad research program for making operator learning more physically structured, mathematically legible, and better aligned with scientific inverse problems.

1. Conceptual scope and taxonomy

A minimal definition is that a neural-operator prior constrains a learned operator $\mathcal{G}:\mathcal{X}\to\mathcal{Y}$, with $\mathcal{X}$ and $\mathcal{Y}$ function spaces, by hard-wiring known operator structure, by choosing an architecture whose continuum limit already encodes specific assumptions, or by placing an explicit stochastic prior over operators or operator parameters (Berner et al., 12 Jun 2025). This differs from ordinary neural-network priors because the object of interest is not merely a map $x\mapsto f(x)$ on vectors, but a function-to-function map $f\mapsto g$, often intended to remain meaningful across discretizations and query locations.

The literature in this area can be organized into three main categories.

Category	Mechanism	Representative papers
Hard operator prior	Insert known operators as fixed layers	(Maier et al., 2017, Zhang et al., 22 Nov 2025)
Architectural operator prior	Choose function-space layers with discretization-agnostic structure	(Berner et al., 12 Jun 2025)
Probabilistic operator prior	Define GP or structured stochastic priors over operators or parameters	(Souza et al., 19 Oct 2025, Karaletsos et al., 2020)

The first category treats prior knowledge as exact operator structure. If the target map factors as

$f(\cdot)=g(u(\cdot)),$

and $u$ or $g$ is known, the network should not relearn that operator from data; instead, the known transform should be inserted directly, leaving only the unknown complement to be learned (Maier et al., 2017). The second category treats the architecture itself as the prior: locality is defined in the underlying domain, sums are replaced by quadrature-weighted integrals, and the same finite parameter vector $\theta\in\mathbb{R}^p$ is used across resolutions (Berner et al., 12 Jun 2025). The third category treats the prior literally as a probability law, either on function-valued outputs induced by infinite-width neural operators or on structured parameterizations that can later be transferred to operator-learning settings (Souza et al., 19 Oct 2025).

A plausible implication is that “neural-operator prior” is best understood as a family of constructions rather than a single formalism. Some papers place the prior inside the architecture, some in an analytic projection step surrounding the operator, and some in the stochastic law induced before training.

2. Embedded known operators as hard priors

The earliest explicit formulation of this perspective is the precision-learning framework, which considers a target mapping that decomposes as $\mathcal{G}:\mathcal{X}\to\mathcal{Y}$0 and compares three cases: approximate $\mathcal{G}:\mathcal{X}\to\mathcal{Y}$1 while keeping $\mathcal{G}:\mathcal{X}\to\mathcal{Y}$2 known, keep $\mathcal{G}:\mathcal{X}\to\mathcal{Y}$3 known while approximating $\mathcal{G}:\mathcal{X}\to\mathcal{Y}$4, or approximate both parts (Maier et al., 2017). In this setting, the architectural question is not merely whether a network can approximate $\mathcal{G}:\mathcal{X}\to\mathcal{Y}$5, but whether known intermediate operators should be inserted exactly into the graph. The answer given is affirmative: replacing a learned stage by the exact operator removes one approximation contribution from the worst-case bound.

The central bound is

$\mathcal{G}:\mathcal{X}\to\mathcal{Y}$6

with sigmoid Lipschitz constant $\mathcal{G}:\mathcal{X}\to\mathcal{Y}$7, and more generally, for Lipschitz $\mathcal{G}:\mathcal{X}\to\mathcal{Y}$8,

$\mathcal{G}:\mathcal{X}\to\mathcal{Y}$9

The stated conclusion is that the maximal estimation error achieved using known operators is lower or equal to the error achieved when training the full composition from scratch, and that knowledge on $\mathcal{X}$0 and $\mathcal{X}$1 reduces the maximal error in an additive fashion (Maier et al., 2017). This is a worst-case approximation bound rather than a PAC-style sample-complexity theorem, but it formalizes the operator-prior intuition: each learned stage contributes approximation error, so exact insertion of a known stage tightens the bound.

In the X-ray material decomposition experiment, the known operators are a $\mathcal{X}$2 transform $\mathcal{X}$3 and a polynomial expansion $\mathcal{X}$4, followed by a three-layer perceptron $\mathcal{X}$5. The four tested variants are

$\mathcal{X}$6

The empirical trend is monotonic in SSIM, with $\mathcal{X}$7, $\mathcal{X}$8, $\mathcal{X}$9, and $\mathcal{Y}$0, respectively, while Pearson’s $\mathcal{Y}$1 varies only from $\mathcal{Y}$2 to $\mathcal{Y}$3 (Maier et al., 2017). The experiment therefore demonstrates that operator priors can matter structurally even when simple correlation metrics change little.

This formulation is “hard” in a strict sense. The inserted transform is not regularization, data augmentation, or a soft penalty; it is a fixed block in the forward pass. The only stated implementation requirement is that the operator admit gradient or sub-gradient computation for backpropagation (Maier et al., 2017). The paper explicitly mentions filtering, backprojection, eigenvalue computations, nonlinear feature transforms, and sorting or quantile operations as admissible operator types. That breadth makes the precision-learning argument an early and explicit statement of operator priors as architectural constraints.

3. Architectural priors over continuum mappings

A second line of work treats the prior as a consequence of lifting neural-network layers from index space to coordinate or function space. The defining principles stated for neural operators are discretization-agnostic, fixed number of parameters, and universal approximation (Berner et al., 12 Jun 2025). On this view, the prior is that the target mapping is a continuum operator, not a mesh-specific regression rule.

The construction recipe is systematic. One identifies the continuous analogue of a finite-dimensional layer, replaces discrete sums by integrals, inserts quadrature weights in the discretization, defines locality by neighborhoods in the underlying domain rather than by index count, and replaces resolution-dependent tensors by learnable kernel or bias functions (Berner et al., 12 Jun 2025). Thus a finite graph-style layer

$\mathcal{Y}$4

becomes

$\mathcal{Y}$5

The same principle yields integral transforms for MLPs, convolutional operators with fixed physical receptive field, quadrature-aware attention operators, and encoder-decoder operators whose encoders are discretization-agnostic and whose decoders are queryable.

The formal consistency notion is discretization convergence: $\mathcal{Y}$6 for any discrete refinement $\mathcal{Y}$7 and compact $\mathcal{Y}$8 (Berner et al., 12 Jun 2025). This is not merely a numerical convenience. It is the paper’s strongest architectural statement of prior structure: aggregation should approximate integration, outputs should correspond to functions rather than array slots, and the same $\mathcal{Y}$9 should remain meaningful across resolutions.

Different neural-operator architectures then encode different priors. Convolutional operators impose translation-structured mixing and fixed physical locality; FNO-style spectral operators impose a low-frequency or band-limited bias; graph neural operators impose geometry-aware locality on irregular domains; transformer neural operators impose adaptive global nonlocality, but only after inserting quadrature weights into both normalization and aggregation (Berner et al., 12 Jun 2025). Encoder-decoder neural operators impose latent-basis or manifold priors. Composite models such as FNO + LocalConv and FNO + DiffConv combine global spectral and local operator biases.

The reported experiments on 2D Navier–Stokes support the claim that these are genuinely architectural priors rather than mere training heuristics. Fixed physical receptive fields matter, multi-resolution training alone does not induce robust unseen-resolution generalization, and input/output interpolation yields a neural operator but discards fine-scale information (Berner et al., 12 Jun 2025). This suggests that operator priors are realized most effectively when encoded in the layer construction itself rather than appended as data augmentation.

4. Explicit probabilistic priors on operators and structured parameters

A more literal use of “prior” arises when one asks what stochastic process a neural operator induces before training. For arbitrary-depth neural operators with Gaussian-distributed convolution kernels and appropriate infinite-width limits, the limiting object is a function-valued Gaussian process: $x\mapsto f(x)$0 with covariance indexed jointly by two input functions and two spatial coordinates, $x\mapsto f(x)$1 (Souza et al., 19 Oct 2025). This extends the NNGP correspondence from vector-valued networks to operator learning.

The neural-operator layer is written as

$x\mapsto f(x)$2

and the covariance recursion propagates through linear operator layers and nonlinearities via

$x\mapsto f(x)$3

For FNO parameterizations, the prior is especially explicit: the induced operator covariance is a finite spectral sum over retained modes, so the infinite-width FNO prior is translation-equivariant in output coordinates and imposes a hard band-limit (Souza et al., 19 Oct 2025). The paper states that the synthetic experiment confirms sensitivity to this band-limit: models only predict accurately once their band-limit exceeds that of the ground-truth operator.

This result makes operator-space inductive bias mathematically legible. The random kernel parameterization $x\mapsto f(x)$4, the pointwise variance $x\mapsto f(x)$5, the nonlinearity $x\mapsto f(x)$6, and the depth together determine the prior predictive law over output functions conditioned on input functions (Souza et al., 19 Oct 2025). Regression under this prior reduces to exact GP inference on the index set $x\mapsto f(x)$7, with kernel entries

$x\mapsto f(x)$8

A related but more indirect line appears in hierarchical GP priors for Bayesian neural-network weights. That work does not explicitly discuss neural operators in the modern function-to-function sense, but it introduces a GP prior over a weight-generating function $x\mapsto f(x)$9, where each weight code is built from latent unit embeddings, $f\mapsto g$0, and in the local version from augmented codes such as $f\mapsto g$1 or $f\mapsto g$2 (Karaletsos et al., 2020). The resulting prior is correlated, hierarchical, and context-dependent. The paper explicitly notes that this is not yet a full neural-operator prior, but it is relevant because it offers a structured, kernel-controlled way to move beyond i.i.d. parameter priors. A plausible implication is that such kernelized parameter priors could be transferred to branch/trunk units, Fourier modes, graph edge types, or operator blocks in Bayesian neural operators.

5. Physics priors, projections, and generative priors in inverse problems

In inverse problems, neural-operator priors often appear not as a probability law over the operator itself, but as a structured interaction between a neural operator and an external physical or generative prior. A clear example is spectral super-resolution with an atmospheric radiative transfer prior. The method defines a three-stage pipeline: $f\mapsto g$3 where the first stage uses an atmospheric radiative transfer guidance matrix $f\mapsto g$4, the second stage is a coordinate-based neural operator over wavelength, and the third stage is a hard refinement projection enforcing $f\mapsto g$5 (Zhang et al., 22 Nov 2025).

The key prior mechanism is the guidance matrix projection (GMP), defined by the constrained optimization problem

$f\mapsto g$6

The stated theorem gives the optimal pixelwise solution

$f\mapsto g$7

or in matrix form,

$f\mapsto g$8

This is not merely a soft regularizer. It uses the prior to fill the null-space component of the underdetermined system $f\mapsto g$9, first with atmospheric guidance and then with the operator prediction itself (Zhang et al., 22 Nov 2025). The ablation reported in the paper distinguishes the roles clearly: using both ART prior and refinement gives PSNR $f(\cdot)=g(u(\cdot)),$0 and SAM $f(\cdot)=g(u(\cdot)),$1; removing the ART prior but keeping refinement gives $f(\cdot)=g(u(\cdot)),$2 and $f(\cdot)=g(u(\cdot)),$3; removing refinement but keeping ART prior gives $f(\cdot)=g(u(\cdot)),$4 and $f(\cdot)=g(u(\cdot)),$5; removing both gives $f(\cdot)=g(u(\cdot)),$6 and $f(\cdot)=g(u(\cdot)),$7 (Zhang et al., 22 Nov 2025). The paper also reports that replacing hard refinement by a soft consistency regularizer degrades both PSNR and SAM significantly.

Other inverse-problem papers in the corpus use a different decomposition: the neural operator is the forward surrogate, while the prior is separate and generative. In diffusion-prior-enhanced velocity model building, a neural operator approximates the map

$f(\cdot)=g(u(\cdot)),$8

and a separate unconditional DDPM regularizes the recovered high-resolution velocity model (Ma et al., 29 Dec 2025). The paper is explicit that this is not a case where the neural operator itself is the prior; instead it is a neural-operator-based inverse problem with a separate generative prior. Likewise, in reservoir history matching, a PINO/FNO forward surrogate is coupled with VCAE and DDIM priors over permeability and porosity fields, plus a CCR mixture-of-experts closure model (Etienam et al., 2024). In both cases, the operator supports efficient data-consistent updates, while the prior enforces plausibility of the latent field.

This suggests a useful distinction. Some neural-operator priors constrain the hypothesis class of the operator directly; others constrain the effective domain or codomain of an operator-centered inverse problem through projection, latent reparameterization, or generative denoising.

6. Empirical consequences, limitations, and open questions

Across the surveyed papers, operator priors are associated with several recurring empirical effects. Embedded known transforms reduce background artifacts and improve structural fidelity in X-ray material decomposition, with SSIM increasing from $f(\cdot)=g(u(\cdot)),$9 to $u$0 as more known-transform structure is inserted (Maier et al., 2017). Architectures designed in function space generalize more robustly across resolutions than fixed-resolution CNNs or ViTs, and multi-resolution training alone does not substitute for a discretization-aware operator prior (Berner et al., 12 Jun 2025). In spectral super-resolution, combining a neural operator with an atmospheric prior yields continuous spectral reconstruction and zero-shot extrapolation, with reported gains such as continuous PSNR $u$1 versus $u$2 and $u$3, and zero-shot PSNR $u$4 versus $u$5 and $u$6 (Zhang et al., 22 Nov 2025). In operator-GP theory, the main consequence is not a benchmark score but an explicit covariance law that reveals the prior bias of FNO and related architectures (Souza et al., 19 Oct 2025).

The literature also states important limitations. Precision learning provides maximal approximation-error bounds, not expected generalization bounds, does not prove tightness, and assumes that the inserted operator actually matches the true operator sequence (Maier et al., 2017). The architecture-conversion framework emphasizes discretization convergence but leaves open probabilistic uncertainty over operators, data efficiency, geometry dependence, and stronger practical approximation guarantees (Berner et al., 12 Jun 2025). Infinite-width operator-GP theory concerns the model at initialization, assumes Gaussian priors with $u$7 scaling and iterated width limits, and does not show that SGD-trained finite neural operators converge to the corresponding GP posterior (Souza et al., 19 Oct 2025). Projection-based inverse methods rely on domain-specific forward models and assumptions such as full-row-rank spectral response matrices or physically adequate atmospheric priors (Zhang et al., 22 Nov 2025). Hybrid inverse frameworks with diffusion or VCAE priors depend strongly on the realism of the training distribution and do not by themselves provide full Bayesian posterior inference or calibrated uncertainty quantification (Ma et al., 29 Dec 2025).

Several open directions follow directly from these limitations. One is the unification of architectural and probabilistic views: current work separately studies discretization-aware operator classes and Gaussian-process limits, but does not yet provide a general Bayesian theory for discretization-convergent neural operators (Berner et al., 12 Jun 2025). Another is the design of genuine stochastic neural-operator priors over coefficient fields or latent random functions, rather than the common pattern of “operator surrogate + separate generative prior,” which is explicitly identified as distinct in reservoir and seismic inversion studies (Etienam et al., 2024). A third is selective prior placement: the structured GP-weight literature suggests that priors need not cover all parameters uniformly, but can target semantically meaningful blocks such as branch/trunk fusion layers, spectral filters, or output projections (Karaletsos et al., 2020).

In synthesis, neural-operator priors are best understood as a hierarchy of increasingly explicit constraints on operator learning. At one end are hard architectural insertions of known transforms. In the middle are continuum-consistent operator architectures whose locality, equivariance, basis structure, and quadrature rules encode an implicit prior over admissible mappings. At the other end are explicit stochastic priors, whether induced by infinite-width limits or imposed through structured generative models and analytic projection operators. The unifying principle is that operator learning becomes more effective when prior knowledge is expressed at the level of operators, function spaces, and physical constraints, rather than only through generic parameter regularization.