betaFeedback

WorksComputer Science

Exponential-Nemytskii Neural Operators: Learning Field Evolution Across Step Sizes

Abstract

In variable-step operator learning, agreement at observed step sizes does not by itself determine predictions at other temporal resolutions. We propose an operator whose step-size dependence is prescribed by a first-order Duhamel approximation of semilinear dynamics. It combines exact linear propagation with a local nonlinear function of the field and its derivatives, allowing evaluation below the observation interval without supervision densely sampled across step sizes. Under suitable conditions and at fixed spatial resolution, we prove that agreement between models at two step sizes controls their prediction differences throughout a prescribed range. We also give conditions for local identifiability of both components from two step sizes, supporting interpretation of the learned structure. A latent extension applies the same construction to learned fields. We evaluate the base model on two synthetic problems and the extension on two real datasets against several baselines.

Introduction

Many quantities of scientific and practical interest are continuous fields that evolve in space and time: the elevation of a water surface, the concentration of a diffusing species, the liquid-water content of a storm seen by radar [Veillette et al., 2020], or the radiative power of a spreading wildfire [Wooster et al., 2005]. The continuous field itself is rarely observed directly. Instead, we typically rely on a finite set of noisy samples taken on a spatial grid at discrete observation times. The task is to learn the underlying spatiotemporal evolution from these samples and to predict the field’s future states. Operator learning [8] (Lu et al. 2021) has emerged as a popular approach to this task. Rather than fitting a discrete mapping between fixed-size arrays, it learns the solution operator that advances the field in time as a mapping between function spaces. Because the learned operator acts on the underlying functions rather than their discrete samples, it can be evaluated on spatial grids other than those used during training, and it can be applied to complex systems even when their governing equations remain unknown. The other family of methods are equation learning [1], which focus on recovering governing equations from the data. Our work focuses on operator learning; equation learning is discussed further in § 2.4.

While operator learning treats space as continuous, many autoregressive models, including FNO-2D [Li et al., 2021], the autoregressive version of DeepONet [Lu et al., 2022] and KNO [Xiong et al., 2024], advance the field at a fixed time interval in both training and prediction. The step size is thereby built into the learned map, which does not directly support prediction at arbitrary temporal intervals. Considerable effort has been devoted to developing methods for variable-step prediction, and we review these approaches in § 2.4. In this work, we propose a new operator-learning method for variable-step prediction, based on a first-order Duhamel approximation of semilinear dynamics. We combine this construction with multi-step-size training and establish conditions under which prediction maps at two distinct step sizes locally identify the linear and nonlinear components. Our main contributions are as follows.

  • Variable-step prediction with a stability guarantee. Our model’s dependence on the step size is prescribed by its mathematical construction, rather than learned entirely from data. This structure allows us to establish a theoretical stability guarantee for prediction across temporal resolutions, without requiring training data densely sampled across step sizes. We prove that, under stated conditions, close agreement with a reference model at just two distinct training step sizes controls prediction differences throughout a prescribed range of step sizes, including below the observation interval (Theorem 1).

  • Identifiability and interpretability. When applied directly to the observed fields, our model explicitly learns a linear differential operator and a local nonlinear function of the field and its derivatives. These components can be inspected directly to understand how linear propagation and nonlinear effects contribute to the prediction. This interpretation is ambiguous when different choices of components produce identical predictions. We prove that, under suitable conditions, the linear and nonlinear components of our model are locally identifiable when trained at two distinct step sizes (Theorem 3).

We test the method on two synthetic problems and on two real datasets, and compare it against a variety of baselines from operator learning and equation learning.

Method

Setup and model derivation

We observe a spatio–temporal field u(x,t)u(x,t), x∈Ω⊂RdΩx \in\Omega\subset\mathbb{R}^{d_{\Omega}}, t∈[0,T]t \in[0,T], only through discrete, possibly noisy samples

{u~(xℓ,tj)}ℓ=1,…,Nx; j=1,…,Nt(1)\left\{\widetilde{u}(x_{\ell},t_{j})\right\}_{\ell=1,\ldots,N_{x};\ j=1,\ldots,N_{t}} \tag*{(1)}

taken at NxN_{x} spatial nodes xℓ∈Ωx_{\ell} \in\Omega and NtN_{t} times tj∈[0,T]t_{j} \in[0,T]. Our goal is to learn the underlying model of the field and use it to make predictions.

We assume that u(x,t)u(x,t) is described by a semilinear evolution equation,

∂tu=Au+F(u),u(⋅,0)=u0,(2)\partial_{t}u = Au + F(u), \qquad u(\cdot,0)=u_{0}, \tag*{(2)}

where AA is an unknown linear operator generating a strongly continuous semigroup (etA)t>0(e^{tA})_{t>0} on a Banach space X\mathcal{X}, and F:X→XF:\mathcal{X}\to\mathcal{X} is an unknown, locally Lipschitz nonlinear operator. Our aim is to learn both from Eq. (1) alone, in the form of a neural operator that predicts uu at later times.

We construct our learning model using the Duhamel representation of Eq. (2), which does not merely split the dynamics into a linear part and a correction, but prescribes how the two combine. Write

v(t):=etAu0(3)v(t):=e^{tA}u_{0} \tag*{(3)}

for the purely linear evolution of the initial state. A mild solution of Eq. (2) satisfies the Duhamel formula

u(t)=v(t)+∫0te(t−s)AF(u(s)) ds,(4)u(t)=v(t)+\int_{0}^{t}e^{(t-s)A}F(u(s))\,\mathrm{d}s, \tag*{(4)}

and is the fixed point of w↦v+N[w]w\mapsto v+\mathcal{N}[w], where N[w](t):=∫0te(t−s)AF(w(s)) ds\mathcal{N}[w](t):=\int_{0}^{t}e^{(t-s)A}F(w(s))\,\mathrm{d}s, constructed by Picard iteration from w(0):=vw^{(0)}:=v. Successive Picard iterates w(k+1):=v+N[w(k)]w^{(k+1)}:=v+\mathcal{N}[w^{(k)}] differ by N[w(k)]−N[w(k−1)]\mathcal{N}[w^{(k)}]-\mathcal{N}[w^{(k-1)}], which gains one order in tt at each step since FF is Lipschitz, so only the first correction N[v]\mathcal{N}[v] enters at first order. Keeping it alone, u≈v+N[v]u\approx v+\mathcal{N}[v], and expanding N[v]\mathcal{N}[v] near t=0t=0 (substitute r=t−sr=t-s, then use erA≈Ie^{rA}\approx I and v(t−r)≈v(t)v(t-r)\approx v(t)) gives, at t=τt=\tau,

u(τ)≈v(τ)+τF(v(τ)).(5)u(\tau)\approx v(\tau)+\tau F(v(\tau)). \tag*{(5)}

Over a short horizon the solution is the exactly propagated linear flow plus a single local correction evaluated at the current linear state. Eq. (5) is the model we train to predict uu forward in time.

Parametrisation and training

To use Eq. (5) we must learn AA and FF from data. We parametrise each by a structured class matching its role, rather than by a generic neural network.

Spectral linear propagator. We assume AA is a constant-coefficient differential operator of order at most dd on a periodic domain, A=∑∣n∣≤dan∂nA=\sum_{|\mathbf{n}|\le d}a_{\mathbf{n}}\partial^{\mathbf{n}} with an∈Ra_{\mathbf{n}}\in\mathbb{R} and ∂n=∂1n1⋯∂dΩndΩ\partial^{\mathbf{n}}=\partial_1^{n_1}\cdots\partial_{d_\Omega}^{n_{d_\Omega}}. The Fourier transform F\mathcal{F} replaces ∂j\partial_j by ikjik_j, so AA is a Fourier multiplier with symbol

g(k)=∑∣n∣≤dcn(ik)n,cn=an∈R,(6)g(k)=\sum_{|\mathbf{n}|\le d}c_{\mathbf{n}}(ik)^{\mathbf{n}},\qquad c_{\mathbf{n}}=a_{\mathbf{n}}\in\mathbb{R}, \tag*{(6)}

a polynomial in ikik whose coefficients are those of AA itself; in one dimension g(k)=∑n≤dcn(ik)ng(k)=\sum_{n\le d}c_n(ik)^n. The linear flow Eq. (3) becomes v(x,τ)=F−1[eτg(k)u^0(k)](x)v(x,\tau)=\mathcal{F}^{-1}[e^{\tau g(k)}\widehat{u}_0(k)](x) with u^0=F[u0]\widehat{u}_0=\mathcal{F}[u_0], and the coefficients {cn}\{c_{\mathbf{n}}\} are the trainable parameters. The propagator is the exact solution of the linear part, so the O(τ2)O(\tau^2) truncation in Eq. (5) is attributable to the nonlinear term alone. In the implementation we truncate the frequency modes to ∣k∣≤K|k|\le K, where KK is a hyperparameter (values in Appendix § 7.§ 2.1).

For a bound uniform in kk, we enforce Re⁡g(k)≤α\operatorname{Re}g(k)\le\alpha for all k∈RdΩk\in\mathbb{R}^{d_\Omega}, so that ∣eτg(k)∣≤eατ|e^{\tau g(k)}|\le e^{\alpha\tau}: the linear flow is bounded on L2(Ω)L^2(\Omega) at every step size and every discretisation. This is imposed by the following construction. Only terms of even total order contribute to Re⁡g\operatorname{Re}g; the odd-order terms, which carry transport and dispersion, are purely imaginary and are left free, as is the constant term c0=:αc_0=:\alpha. Among the even-order terms, those in which every kjk_j appears to an even power (k12k_1^2, k12k22k_1^2k_2^2, and so on) are nonnegative everywhere and receive a nonpositive coefficient, so they can only push Re⁡g\operatorname{Re}g down; the rest, such as k1k2k_1k_2, change sign with direction and are dropped.

Nemytskii operator. For the nonlinear part of Eq. (2) we do not consider nonlocal nonlinearities, and assume instead that FF is local of some order ρ∈N0\rho\in\mathbb{N}_0: the value of F(u)F(u) at xx depends on uu only through u(x)u(x) and its spatial derivatives at xx up to order ρ\rho (the ρ\rho-jet of uu at xx). Write

Φ[w](x):=(w(x),∇w(x),…,∇ρw(x),η(x))∈Rm(7)\Phi[w](x):=(w(x),\nabla w(x),\ldots,\nabla^\rho w(x),\eta(x))\in\mathbb{R}^m \tag*{(7)}

for the feature map, which collects the field, its spatial derivatives up to order ρ\rho, and, optionally, externally given fields η(x)\eta(x); its number of entries is mm. We then assume

F(u)(x)=f(Φ[u](x)),f:Rm→R,(8)F(u)(x)=f\bigl(\Phi[u](x)\bigr),\qquad f:\mathbb{R}^m\to\mathbb{R}, \tag*{(8)}

so that FF is a generalised Nemytskii (superposition) operator [9] induced by the finite-dimensional map ff. Accordingly, the goal becomes to learn the finite-dimensional function f:Rm→Rf:\mathbb{R}^m\to\mathbb{R} and we represent it by a multilayer perceptron (MLP) with smooth activations. The derivatives in Φ\Phi are evaluated spectrally, ∇jv=F−1[(ik)jv^]\nabla^jv=\mathcal{F}^{-1}[(ik)^j\widehat{v}].

Exponential-Nemytskii neural operator. Combining the two components using Eq. (5),

Sτ[u0](x)=F−1[eτgFu0](x)⏟v(x,τ)+τf(Φ[v(⋅,τ)](x)),(9)S_\tau[u_0](x)=\underbrace{\mathcal{F}^{-1}\left[e^{\tau g}\mathcal{F}u_0\right](x)}_{v(x,\tau)}+\tau f\bigl(\Phi[v(\cdot,\tau)](x)\bigr), \tag*{(9)}

with trainable objects gg and ff. We call the resulting model the Exponential–Nemytskii Neural Operator (ENNO), reflecting its combination of exponential linear propagation and a learned Nemytskii map acting on local field features. Predictions over horizons longer than a single first-order step are obtained by applying Eq. (9) autoregressively: starting from u0u_0,

uj+1=Sτ[uj],j=0,…,n−1,so thatun=Sτn[u0](10)u_{j+1}=S_\tau[u_j],\qquad j=0,\ldots,n-1,\qquad\text{so that}\qquad u_n=S_\tau^n[u_0] \tag*{(10)}

is the prediction at time t=nτt=n\tau.

Training. In training we use only the one-step model Eq. (9); the multi-step composition Eq. (10) is used only for prediction. A key feature of our approach is that the model is trained with several step sizes τ\tau, which is required for the identifiability of (g,f)(g,f) (§ 2.3.2). Concretely, fix step sizes τ1<⋯<τS\tau_1<\cdots<\tau_S, S≥2S\ge2, chosen among the differences of observation times, and for each τs\tau_s let Ds\mathcal{D}_s be the set of observation times tjt_j for which tj+τst_j+\tau_s is also an observation time. For each tj∈Dst_j\in\mathcal{D}_s, we apply the one step of size τs\tau_s and compare with the observation at tj+τst_j+\tau_s; averaging the squared error over the grid and over Ds\mathcal{D}_s gives

Ls=1∣Ds∣∑tj∈Ds1Nx∑ℓ=1Nx∣Sτs[u~(⋅,tj)](xℓ)−u~(xℓ,tj+τs)∣2,(11)\mathcal{L}_s=\frac{1}{|\mathcal{D}_s|}\sum_{t_j\in\mathcal{D}_s}\frac{1}{N_x}\sum_{\ell=1}^{N_x}\left|S_{\tau_s}[\widetilde{u}(\cdot,t_j)](x_\ell)-\widetilde{u}(x_\ell,t_j+\tau_s)\right|^2, \tag*{(11)}

and the training objective is their sum, L=∑s=1SLs\mathcal{L} = \sum_{s=1}^{S} \mathcal{L}_s. Since τ\tau enters SτS_\tau explicitly, the same (g,f)(g, f) is used for every ss. With a single step size the constant term of gg and the linear part of ff cannot be separated (Lemma 2); a second step size removes this ambiguity (Theorem 3(ii)).

Latent fields. An extension of the method handles the case where u(x,t)u(x,t) does not follow the semilinear model Eq. (2): we map u(x,t)u(x,t) into a latent field, or several latent fields, using an encoding–decoding structure. Namely, we use an encoder E\mathcal{E} to map the observed history into the latent fields w=(w1,…,wM)\boldsymbol{w} = (w^1,\ldots,w^M), advance all of them with the same gg and ff, and use a decoder R\mathcal{R} to map the latent state back:

w0=E[u−T+1,…,u0],wj+1=(Sτ[wj1],…,Sτ[wjM]),uj=R[wj].(12)w_0 = \mathcal{E}[u_{-T+1},\ldots,u_0], \qquad w_{j+1} = \left(S_\tau[w_j^1],\ldots,S_\tau[w_j^M]\right), \qquad u_j = \mathcal{R}[w_j]. \tag*{(12)}

Here we restrict both E\mathcal{E} and R\mathcal{R} to be Nemytskii operators, generated by pointwise maps RT→RM\mathbb{R}^{T} \to\mathbb{R}^{M} and RM→R\mathbb{R}^{M} \to\mathbb{R}, so every learnable object in the model is generated by a map between finite-dimensional spaces. The encoder and decoder are trained jointly with gg and ff. Taking T>1T > 1 acts as a delay embedding: when uu is a partial observation of a higher-dimensional state, a single frame need not determine the subsequent evolution. Only the initial latent state uses the history; the rollout remains Markovian in ww. Note that gg and ff are shared across the latent fields, so the parameter count does not grow with MM: the MM fields evolve independently under the same dynamics but from different initial channels produced by E\mathcal{E}, and are recombined nonlinearly by R\mathcal{R}. The price is interpretability: in the base model gg and ff describe the mechanisms governing uu itself, whereas here they describe the latent fields and no longer correspond directly to the equation for uu.

Theory

Stability across step sizes

We work at fixed spatial resolution KK on a periodic box Ω\Omega. Let VKV_K be the real Fourier space with integer mode indices ∣k∣≤K\lvert k\rvert\le K, let PKP_K be its orthogonal projection, and write UR:={u∈VK:∥u∥L2≤R}U_R := \{u \in V_K : \lVert u\rVert_{L^2} \le R\}. All operator norms below are induced by the L2L^2 norm. For ϑ=(g,f)\vartheta= (g,f), let AA have symbol gg as in Eq. (6), and set FK(v)=PK[f(Φst[v])]F_K(v) = P_K[f(\Phi_{\mathrm{st}}[v])] and StK,ϑu=etAu+tFK(etAu)S_t^{K,\vartheta}u = e^{tA}u + tF_K(e^{tA}u). Here ff is a spatially homogeneous, locally Lipschitz scalar function, and Φst[v]\Phi_{\mathrm{st}}[v] collects the field and its distinct partial derivatives through order ρ≥0\rho\ge0. For a second pair ϑ~=(g~,f~)\widetilde{\vartheta} = (\widetilde{g},\widetilde{f}) from the same class, define A~\widetilde{A}, F~K\widetilde{F}_K, and StK,ϑ~S_t^{K,\widetilde{\vartheta}} analogously. Let MK\mathcal{M}_K denote the real-preserving Fourier multipliers on VKV_K, and define the nonlinear contrast NF(v):=FK(v)−2FK(v/2)+FK(0)N_F(v) := F_K(v) - 2F_K(v/2) + F_K(0). We have the following theorem where the proof is given in Appendix § 7.§ 1.

Theorem 1 (Cross-step stability). Fix R,Δ,M>0R,\Delta,M > 0, and suppose ∥A∥op,∥A~∥op≤M\lVert A\rVert_{\mathrm{op}},\lVert\widetilde{A}\rVert_{\mathrm{op}} \le M. Assume that there exist 0<r0<R0 < r_0 < R and σ,ζ>0\sigma,\zeta> 0 such that

sup⁡v∈Ur0R∥NF(eHv)−NF(v)∥L2≥σ∥H∥op,H∈MK,∥H∥op≤ζ.(13)\sup_{v\in U_{r_0}^{R}} \lVert N_F(e^{H}v)-N_F(v)\rVert_{L^2} \ge\sigma\lVert H\rVert_{\mathrm{op}}, \qquad H\in\mathcal{M}_K,\quad\lVert H\rVert_{\mathrm{op}}\le\zeta. \tag*{(13)}

Let LΔL_\Delta be a Lipschitz constant for FKF_K on Uexp⁡(2MΔ)RU_{\exp(2M\Delta)R}. There exist 0<τ∗≤Δ0 < \tau_* \le\Delta, 0<rΔ<R0 < r_\Delta< R, and CΔ>0C_\Delta> 0, depending only on Ω,K,M,R,Δ,r0,σ,ζ,LΔ\Omega,K,M,R,\Delta,r_0,\sigma,\zeta,L_\Delta, such that for any 0<τ0<τ1≤τ∗0 < \tau_0 < \tau_1 \le\tau_*, setting

ϵobs:=max⁡j=0,1sup⁡u∈UR∥SτjK,ϑu−SτjK,ϑ~u∥L2τj,\epsilon_{\mathrm{obs}} := \max_{j=0,1}\sup_{u\in U_R}\frac{\lVert S_{\tau_j}^{K,\vartheta}u-S_{\tau_j}^{K,\widetilde{\vartheta}}u\rVert_{L^2}}{\tau_j},

we have, for every 0<δ≤Δ0 < \delta\le\Delta,

sup⁡u∈UrΔ∥SδK,ϑu−SδK,ϑ~u∥L2δ≤CΔϵobs(1+2min⁡τ∈[τ0,τ1]∣δ−τ∣τ1−τ0).(14)\sup_{u\in U_{r_\Delta}}\frac{\lVert S_\delta^{K,\vartheta}u-S_\delta^{K,\widetilde{\vartheta}}u\rVert_{L^2}}{\delta} \le C_\Delta\epsilon_{\mathrm{obs}}\left(1+\frac{2\min_{\tau\in[\tau_0,\tau_1]}\lvert\delta-\tau\rvert}{\tau_1-\tau_0}\right). \tag*{(14)}

The linear dependence on this ratio is sharp in general.

Interpretation. The theorem can be understood as follows. If a model’s predictions are close to those of a reference model at two training step sizes, their prediction difference per unit time also has an explicit upper bound at other step sizes. At intermediate step sizes, the same theoretical bound applies as at the training steps. Outside this interval, the bound grows linearly with the extrapolation distance relative to the separation between the training steps. This provides a stability guarantee for using the same model at different temporal resolutions.

Identifiability

We study structural identifiability of (g,f)(g,f) from the one-step prediction maps. Write Φ[ϕ]=(Φst[ϕ],η)\Phi[\phi]=(\Phi_{\mathrm{st}}[\phi],\eta), where Φst\Phi_{\mathrm{st}} lists each distinct partial derivative through order ρ\rho once, including the field value. Accordingly, write z=(zst,zη)z=(z_{\mathrm{st}},z_\eta), with z0z_0 the field value. Let η:Ω→Rdη\eta:\Omega\to\mathbb{R}^{d_\eta} be continuous and E:=η(Ω)E:=\eta(\Omega). Proofs of the two results below are in Appendix A.1.2.

Lemma 2 (Single-step degeneracy). Fix τ0>0\tau_0>0. For γ∈R\gamma\in\mathbb{R}, set ϱγ:=eγτ0\varrho_\gamma:=e^{\gamma\tau_0} and χγ:=(1−ϱγ)/(ϱγτ0)\chi_\gamma:=(1-\varrho_\gamma)/(\varrho_\gamma\tau_0), and define

gγ:=g+γ,fγ(zst,zη):=f(zst/ϱγ,zη)+χγz0.(15)g_\gamma:=g+\gamma,\qquad f_\gamma(z_{\mathrm{st}},z_\eta):=f(z_{\mathrm{st}}/\varrho_\gamma,z_\eta)+\chi_\gamma z_0. \tag*{(15)}

Then

Sτ0[ϕ;gγ,fγ]=Sτ0[ϕ;g,f](16)S_{\tau_0}[\phi;g_\gamma,f_\gamma]=S_{\tau_0}[\phi;g,f] \tag*{(16)}

whenever the maps are defined.

Theorem 3 (Identifiability). Let (g,f)(g,f) and (g′,f′)(g',f') have symbols in the real-coefficient polynomial class of Eq. (6) and nonlinearities that are C1C^1 in zstz_{\mathrm{st}} and continuous in zηz_\eta. Fix r>0r>0 and K≥K0(d,ρ,dΩ)K\ge K_0(d,\rho,d_\Omega), with the sufficient threshold K0K_0 specified in Appendix A.1.2. At every step size τ\tau considered, assume τ∣Im⁡g~(k)∣<π/4\tau\lvert\operatorname{Im}\tilde{g}(k)\rvert<\pi/4 for g~∈{g,g′}\tilde{g}\in\{g,g'\} and all k∈ZdΩk\in\mathbb{Z}^{d_\Omega} with ∣k∣≤K\lvert k\rvert\le K.

(i) One step size. Let τ0>0\tau_0>0. Suppose Sτ0[ϕ;g,f]=Sτ0[ϕ;g′,f′]S_{\tau_0}[\phi;g,f]=S_{\tau_0}[\phi;g',f'] on Ω\Omega for every ϕ∈VK\phi\in\mathcal{V}_K with ∥ϕ∥L2<r\|\phi\|_{L^2}<r, and 1+τ0∂z0f(0,η(x0))≠01+\tau_0\partial_{z_0}f(0,\eta(x_0))\ne0 for some x0∈Ωx_0\in\Omega. Then, for some γ∈R\gamma\in\mathbb{R} and r∗>0r_*>0, g′=g+γg'=g+\gamma and f′=fγf'=f_\gamma on Br∗(0)×EB_{r_*}(0)\times E with Br∗(0)⊂RmstB_{r_*}(0)\subset\mathbb{R}^{m_{\mathrm{st}}}, where fγf_\gamma is defined by Eq. (15). In particular, cn′=cnc'_n=c_n for 1≤∣n∣≤d1\le\lvert n\rvert\le d.*

(ii) Two step sizes. Under the hypotheses of (i), suppose also Sτ1[ϕ;g,f]=Sτ1[ϕ;g′,f′]S_{\tau_1}[\phi;g,f]=S_{\tau_1}[\phi;g',f'] on Ω\Omega for the same ϕ\phi, where τ1>0\tau_1>0, τ1≠τ0\tau_1\ne\tau_0. If, for some zη∗∈Ez_\eta^*\in E, on no neighbourhood of the origin is f(⋅,zη∗)f(\cdot,z_\eta^*) of the form a+bz0a+bz_0 for constants a,b∈Ra,b\in\mathbb{R}, then γ=0\gamma=0. Thus g′=gg'=g, and f′=ff'=f on Br∗(0)×EB_{r_*}(0)\times E.

Interpretation. Lemma 2 shows that the linear and nonlinear components cannot be uniquely separated at a single step size, and longer rollouts at that same step size do not resolve this ambiguity. Theorem 3 states that a second distinct step size removes this ambiguity under the stated conditions. Thus, using two distinct step sizes resolves the ambiguity between the linear and nonlinear components, establishing structural identifiability of the model.

Related work

We review existing methods that support variable-step prediction. Equation-learning methods can recover continuous-time governing equations from data [2] (Rudy et al. 2017; Long et al. 2018). These equations permit flexible time stepping, although prediction still relies on numerical integration of the learned dynamics. Here we focus on operator-learning methods, which learn evolution maps directly. One such approach treats time as a query coordinate, as in DeepONet with a space–time trunk [Lu et al., 2021] and nonlinear coordinate decoders [Seidman et al., 2022]. These models directly learn spatiotemporal neural representations and may require substantial paired training data [Wang et al., 2021]. The second approach, exemplified by Poseidon [6] and OSG-FNO [3], takes the step size τ\tau as an input. Such models can in principle be evaluated at any step size, but their dependence on τ\tau is entirely learned through training. Outside the range of observed step sizes—in particular below the finest step the data provide—fitting the available data alone cannot guarantee reliable prediction. A third approach learns a vector field in function space and predicts by numerical integration: TI-DeepONet [Nayak and Goswami, 2026] fits it through a numerical integrator, while CFO [7] fits it to estimated time derivatives. The learned dynamics may consequently inherit errors from the numerical integration or derivative estimation used during training.

In our numerical experiments we compare against one representative of each of the three families above: NOMAD [Seidman et al., 2022] for temporal-spatial representation, OSG-FNO [3] for step-size learning, and TI-DeepONet [Nayak and Goswami, 2026] for learned vector fields. We further compare against E-WSINDy [4] from the equation-learning literature, KNO [Xiong et al., 2024] as a fixed-step method, and persistence.

Numerical experiments

We present four numerical experiments: two synthetic examples and two applications to real-world data. We compare our method with 5 benchmarks discussed in § 2.4. We use the recommended or default settings from the original papers and implementations whenever available and applicable. Otherwise, we select suitable hyperparameters based on validation performance. The architecture, spectral cutoff, symbol degree and data settings of ENNO in each experiment are listed in Appendix § 7.§ 2.1. All neural models are trained for at most 3,000 epochs. All baselines are configured to a parameter count comparable to ours, except OSG-FNO, which we could not reduce below three times ours without seriously degrading its training.

The KdV equation

The first synthetic example is the Korteweg–de Vries equation ut+6uux+uxxx=0u_t + 6u u_x + u_{xxx} = 0 on the periodic domain [0,10][0,10]. Its symbol g(k)=ik3g(k) = i k^3 lies in the polynomial class of Eq. (6) once d≥3d \ge3, with c3=−1c_3 = -1 and all other coefficients zero, so the learned symbol can be compared with the truth coefficient by coefficient. Training uses 2030 trajectories on an Nx=128N_x = 128 grid, corrupted by Gaussian noise at 5% of each trajectory’s own standard deviation; 50 and 200 further trajectories with unseen initial conditions serve for validation and test. The model is trained with two step sizes, τ1=10−3\tau_1 = 10^{-3} and τ2=2×10−3\tau_2 = 2 \times10^{-3}. KNO needs ten frames of context, so all methods are scored on the same targets, u(t1),u(t5),u(t10)u(t_1), u(t_5), u(t_{10}); the single-frame methods are conditioned on u(t0)u(t_0) and KNO on the ten frames up to and including u(t0)u(t_0).

The degree dd decides whether the true model can be recovered, so we first ablate it in Table 1, with everything else held fixed. The error falls by a factor of 2.7 from d=2d = 2 to d=3d = 3 and does not improve clearly at d=4d = 4. We use d=3d = 3 in what follows. We also ablate the latent-field structure, which brings no improvement here, as expected since the true model lies in our class at this degree (Appendix § 7.§ 2.2). To test the identifiability of the two components, we show the recovered coefficients in Table 2, where all of them agree well with the truth. Table 3 compares against the baselines, reporting the mean over the 200 test trajectories with its standard error of the mean (SE); ENNO is lowest at all four target times. The t1/2t_{1/2} column evaluates each model at half the smallest training step, which none of them has seen: KNO cannot be evaluated there at all, and OSG-FNO and TI-DeepONet are worse at t1/2t_{1/2} than at t1t_1 despite the shorter horizon.

t1t_1t5t_5t10t_{10}
d=1d = 12.8953.6764.200
d=2d = 22.1792.6952.989
d=3d = 30.3130.7211.115
d=4d = 40.3190.7431.151

Table 1. KdV: relative L2L^2 error (%) at three target frames, averaged over the 200 test trajectories, for symbol degrees d=1,…,4d = 1, \ldots, 4.

c0c_0c1c_1c2c_2c3c_3
truth000−1-1
recovered−0.061-0.061−0.081-0.0814.9×10−54.9 \times10^{-5}−0.979-0.979

Table 2. KdV: recovered symbol coefficients at d=3d = 3 against the truth.

Methodt1/2t_{1/2}t1t_1t5t_5t10t_{10}
ENNO (d=3d = 3)0.197±0.0080.197 \pm0.0080.313±0.0150.313 \pm0.0150.721±0.0320.721 \pm0.0321.115±0.0491.115 \pm0.049
TI-DeepONet2.374±0.0662.374 \pm0.0661.369±0.0351.369 \pm0.0352.556±0.0782.556 \pm0.0783.329±0.1053.329 \pm0.105
E-WSINDy0.830±0.0610.830 \pm0.0611.459±0.0791.459 \pm0.0792.436±0.1242.436 \pm0.1242.793±0.1432.793 \pm0.143
OSG-FNO3.264±0.0833.264 \pm0.0832.045±0.0892.045 \pm0.0892.008±0.0882.008 \pm0.0882.112±0.0802.112 \pm0.080
KNO—2.049±0.0962.049 \pm0.0961.824±0.0831.824 \pm0.0831.912±0.0871.912 \pm0.087
NOMAD4.800±0.1474.800 \pm0.1474.815±0.1524.815 \pm0.1524.804±0.1524.804 \pm0.1524.832±0.1494.832 \pm0.149

Table 3. KdV: relative L2L^2 error (%) at four target times, reported as mean ± SE over the 200 test trajectories. t1/2t_{1/2} is half the smallest training step and is not an observation time; methods that cannot be evaluated below the observation interval are marked “—”.

The advection-diffusion equation

The second synthetic example is the advection-diffusion equation ut=ν(uxx+uyy)−u(ux+uy)u_t = \nu(u_{xx} + u_{yy}) - u(u_x + u_y) with ν=0.05\nu= 0.05 on the periodic domain [0,1]2[0,1]^2. Its symbol g(k)=−ν∣k∣2g(k) = -\nu\lvert k \rvert^2 lies in the polynomial class of Eq. (6) once d≥2d \ge2, with c(2,0)=c(0,2)=νc_{(2,0)} = c_{(0,2)} = \nu and all other coefficients zero, so the learned symbol can again be compared with the truth coefficient by coefficient. Training uses 100 trajectories on a 128×128128 \times128 grid, corrupted by Gaussian noise at 5% of each trajectory’s own standard deviation; 10 and 25 further trajectories serve for validation and test. The encoder and decoder are identity maps, and the model is trained with two step sizes, τ1=10−3\tau_1 = 10^{-3} and τ2=2×10−3\tau_2 = 2 \times10^{-3}. As on KdV, all methods use the last input frame only, except KNO which uses the preceding ten.

The ablation in Table 4 selects d=2d = 2, the true order, with no clear improvement at d=3d = 3. We first estimate the symbol coefficients: the diffusion coefficient is recovered as 0.0488 against the true 0.05, and the remaining coefficients, whose true values are zero, stay below 0.07 in magnitude (Table 5). Table 6 compares against the baselines at comparable parameter counts. ENNO is lowest at all four target times, with E-WSINDy the closest competitor, unlike on KdV, where its candidate library must cover a third-order dispersive term. At t1/2t_{1/2} it remains lowest, while KNO cannot be evaluated there at all. Neither branch–trunk method tracks the dynamics here: the error of NOMAD is already at the scale of the solution at t1t_1 and barely changes over the horizon, while that of TI-DeepONet grows from about 15% at t1t_1 to 51% at t10t_{10}. Since TI-DeepONet is competitive on KdV (second-best at t1t_1), the failure here is not one of training budget.

t1t_1t5t_5t10t_{10}
d=1d = 12.4394.0114.647
d=2d = 20.0210.0740.129
d=3d = 30.0210.0740.128

Table 4. Advection–diffusion: relative L2L^2 error (%) at three target frames, averaged over the 25 test trajectories, for symbol degrees d=1,2,3d = 1, 2, 3.

c(0,0)c_{(0,0)}c(1,0)/c(0,1)c_{(1,0)}/c_{(0,1)}c(2,0)=c(0,2)c_{(2,0)} = c_{(0,2)}
truth00/00.05
recovered0.0290.054/0.0610.0488

Table 5. Advection–diffusion: recovered symbol coefficients at d=2d = 2 against the truth. The two first-order coefficients are listed as x/yx/y.

Methodt1/2t_{1/2}t1t_1t5t_5t10t_{10}
ENNO (d=2d = 2)0.011±0.0010.011 \pm0.0010.021±0.0020.021 \pm0.0020.074±0.0080.074 \pm0.0080.129±0.0130.129 \pm0.013
E-WSINDy0.030±0.0030.030 \pm0.0030.030±0.0030.030 \pm0.0030.145±0.0160.145 \pm0.0160.283±0.0300.283 \pm0.030
KNO—0.269±0.0370.269 \pm0.0370.376±0.0490.376 \pm0.0490.562±0.0720.562 \pm0.072
OSG-FNO0.262±0.0370.262 \pm0.0370.524±0.0600.524 \pm0.0602.617±0.3032.617 \pm0.3035.223±0.6075.223 \pm0.607
TI-DeepONet14.597±2.12214.597 \pm2.12214.551±2.06814.551 \pm2.06829.970±3.33029.970 \pm3.33050.831±5.84350.831 \pm5.843
NOMAD79.330±5.49179.330 \pm5.49160.124±4.97160.124 \pm4.97159.634±4.90959.634 \pm4.90958.969±4.89858.969 \pm4.898

Table 6. Advection–diffusion: relative L2L^2 error (%) at four target times, reported as mean ± SE over the 25 test trajectories. t1/2t_{1/2} is half the smallest training step and is not an observation time; methods that cannot be evaluated below the observation interval are marked “—”.

Radar nowcasting

Radar nowcasting aims to predict the precipitation field 0–2 hours ahead from recent radar observations, filling the gap that numerical weather prediction leaves at these lead times. We take the vertically integrated liquid (VIL) product of SEVIR [Veillette et al., 2020] as the observed field u(x,t)u(x,t), sampled on a uniform 128×128128 \times128 grid over a square domain Ω⊂R2\Omega\subset\mathbb{R}^{2} of 384×384384 \times384 km2^{2}, with consecutive frames five minutes apart. Given the ten most recent frames, the task is to predict the next ten frames at five-minute intervals, covering lead times of 5–50 minutes. We use 11,159 training events, 200 validation events and 500 test events; the test events come from separate files covering November–December 2019. Our model encodes the ten input frames into 20 latent fields and is trained with two step sizes, τ1=5\tau_{1}=5 and τ2=10\tau_{2}=10 minutes. All methods predict the same final ten frames of each test event.

Our model achieves the lowest MSE at all three lead times in Table 7. The comparison with KNO is particularly informative because both models receive the same ten-frame history. Relative to KNO, our model reduces MSE by about 22%, 19% and 11% at 5, 25 and 50 minutes, respectively. The improvement is therefore present in the first prediction and persists at all three lead times, demonstrating a sustained advantage over the tested forecast horizon. Figure 1 illustrates the rollout on a test event. ENNO tracks the motion and coarse structure of the storm cells throughout, while the high-intensity cores are progressively smoothed.

Figure 1. Qualitative SEVIR forecasts for one test event at lead times of 5, 25 and 50 minutes, matching Table 7. Top row: ground truth; bottom row: ENNO’s autoregressive predictions.

Method5 minutes25 minutes50 minutes
ENNO1.4 ± 0.055.4 ± 0.159.2 ± 0.25
KNO1.8 ± 0.056.7 ± 0.1710.3 ± 0.28
OSG-FNO1.8 ± 0.048.6 ± 0.1513.7 ± 0.26
NOMAD5.6 ± 0.148.9 ± 0.2413.4 ± 0.40
TI-DeepONet6.4 ± 0.1611.4 ± 0.2919.5 ± 0.50
E-WSINDy12.5 ± 0.30FailedFailed
Persistence2.6 ± 0.0915.3 ± 0.4224.3 ± 0.66

Table 7. SEVIR test MSE (×10−3\times10^{-3}) at three forecast lead times, reported as mean ±\pm SE over the 500 test events. Bold marks the best result. “Failed” denotes nonfinite predictions during rollout.

Wildfire prediction

Wildfire management relies on knowing where a fire will burn next and how intensely. The goal of this example is to predict the next day’s fire radiative power (FRP) field from the FRP observed on the preceding days together with environmental drivers—wind, terrain, vegetation and moisture—recorded on the most recent day. We use WildfireSpreadTS [5], which records 607 wildfire events across the western United States between 2018 and 2021 as daily multi-channel images at 375 m resolution, one channel holding the observed FRP and the rest holding environmental drivers for the same day. We take Tin∈{1,5}T_{\mathrm{in}} \in\{1,5\} daily FRP fields as the state, transformed by log⁡(1+FRP)\log(1+\mathrm{FRP}) since FRP spans about four orders of magnitude, and 25 environmental channels from the most recent input day. The latter enter the model as the externally given fields η\eta of Eq. (7), so that ff can modulate the predicted intensity pointwise according to local wind, slope and fuel conditions. Our model is trained with two step sizes, τ1=1\tau_1 = 1 and τ2=2\tau_2 = 2 days; evaluation is always at one day ahead. All neural predictors are trained with the same weighted squared loss on log-FRP, giving weight 50 to pixels with positive FRP and weight 1 to the remaining valid pixels: fire pixels make up under 0.15% of the valid pixels, and without this weighting training can favour near-zero predictions. Pixels introduced by padding events to a common grid size are excluded from both training and evaluation.

We use four folds over 2018–2021, one per test year. Each fold trains on two years, selects a checkpoint on a third, and evaluates on the remaining year. Model selection and evaluation use the area under the precision–recall curve (AUC-PR), with predicted log-FRP serving as a continuous score for detecting pixels with positive FRP. Table 8 shows that our model achieves the highest mean AUC-PR at both history lengths. With one input day, it leads in every held-out year, attaining a mean of 0.3210 compared with 0.2895 for KNO. With five input days, our model and OSG-FNO achieve similar means (0.3122 versus 0.3098), each leading in two of the four held-out years; OSG-FNO uses more than ten times as many trainable parameters (8.4M versus 0.75M). Longer histories do not improve our model’s mean AUC-PR here, and persistence exceeds three of the trained baselines with one input day.

MethodTin=1T_{\mathrm{in}} = 1Tin=5T_{\mathrm{in}} = 5
ENNO0.3210 ± 0.08480.3122 ± 0.0835
KNO0.2895 ± 0.08670.2889 ± 0.0744
OSG-FNO0.2664 ± 0.08110.3098 ± 0.0887
TI-DeepONet0.1435 ± 0.0373Unsupported
NOMAD0.1290 ± 0.0394Unsupported
E-WSINDy0.0337 ± 0.06350.2533 ± 0.0954
Persistence0.2473 ± 0.09840.2530 ± 0.0972

Table 8. WildfireSpreadTS test AUC-PR, reported as mean ± sample standard deviation across four held-out years. Bold marks the highest mean. “Unsupported” indicates that the evaluated implementation does not support five-frame inputs.

Conclusions

We presented ENNO, an operator-learning method that combines exact linear propagation with a local nonlinear correction and prescribes its step-size dependence through a first-order Duhamel approximation. Under suitable conditions, prediction maps at two distinct step sizes locally identify these components, and agreement between models at those steps controls prediction differences across a prescribed range. Experiments on two synthetic PDEs and two real-world datasets demonstrate strong predictive performance, with a latent extension handling the real-data applications. The synthetic examples also show accurate recovery of the principal differential coefficients. The current guarantees concern the base model at fixed spatial resolution. Extending the analysis to the latent construction and obtaining stability bounds uniform under spatial refinement are left to future work.

AI use statement

In this work, we used generative AI tools for tasks that require disclosure as follows. The theoretical results of this paper—the model construction, the statements of Theorem 1, Lemma 2 and Theorem 3, and the proof strategies—were conceived and developed in full by the authors, and complete proofs were written before any AI tool was involved. We then used generative AI tools to check these proofs. The tools identified a number of errors and proposed corrections, some of which we adopted; they also suggested improvements to the arguments, some of which we adopted. Every adopted correction and improvement was verified by the authors before being incorporated, and none of them changed the statements of the results. We disclose this use under both assisting in the writing of proofs and providing critical ingredients for proving mathematical claims.

We have reviewed all AI-assisted work. All AI assistance was carried out under the direction of the authors, and the resulting material was subsequently checked and verified by two authors: every proof, including every AI-suggested correction, was verified line by line independently of the tool’s output; AI-edited text was checked against the claims the authors intended to make. We take responsibility for the final content of this work, including text, claims or artifacts produced with the aid of generative AI.

Reproducibility statement

The model architecture, Fourier cutoff, symbol degree, nonlinear-net inputs and the constraint parametrisation used in each experiment are listed in Appendix § 7.§ 2.1, together with the data-generation procedure for the two synthetic problems and the splits for all four experiments. SEVIR and WildfireSpreadTS are publicly available; the preprocessing applied to each is described in § 3.3 and § 3.4 and Appendix § 7.§ 2.1. Code for the model, the baselines and the evaluation pipeline will be made publicly available.

Appendix

Proofs

Proofs of § 2.3.1

Throughout this appendix, write

X:=VK,F:=FK,F~:=F~K,St:=StK,ϑ,S~t:=S~tK,ϑ~.X := \mathcal{V}_{K}, \qquad F := F_{K}, \qquad\widetilde{F} := \widetilde{F}_{K}, \qquad S_{t} := S_{t}^{K,\vartheta}, \qquad\widetilde{S}_{t} := \widetilde{S}_{t}^{K,\widetilde{\vartheta}}.

For vectors, ∥⋅∥\lVert\cdot\rVert denotes the L2(Ω)L^{2}(\Omega) norm; for linear operators, it denotes the induced operator norm. Since XX is finite-dimensional, the locally Lipschitz maps FF and F~\widetilde{F} are Lipschitz on every bounded ball. All Fourier multipliers considered below preserve real-valued fields and commute with one another.

Proof of the estimate in Theorem 1. Set

B:=A−A~,h:=τ1−τ0,ε:=εobs.B := A - \widetilde{A}, \qquad h := \tau_{1} - \tau_{0}, \qquad\varepsilon:= \varepsilon_{\mathrm{obs}}.

Then ∥B∥≤2M\lVert B\rVert\le2M. We may choose

τ∗:=min⁡{Δ,ζ2M,log⁡(R/r0)3M},rΔ:=r0e−MΔ.(17)\tau_* := \min\left\{\Delta,\frac{\zeta}{2M},\frac{\log(R/r_0)}{3M}\right\}, \qquad r_\Delta:= r_0e^{-M\Delta}. \tag*{(17)}

We first control the difference between the linear operators. For t>0t>0 and v∈Xv\in X, define

Zt(v):=S~t(e−tA~v)−S~t(e−tA~v)t.(18)Z_t(v):=\frac{\widetilde{S}_t(e^{-t\widetilde{A}}v)-\widetilde{S}_t(e^{-t\widetilde{A}}v)}{t}. \tag*{(18)}

Commutativity of AA and A~\widetilde{A} gives the exact identity

Zt(v)=etB−Itv+F(etBv)−F~(v).(19)Z_t(v)=\frac{e^{tB}-I}{t}v+F(e^{tB}v)-\widetilde{F}(v). \tag*{(19)}

Consequently,

Zt(v)−2Zt(v/2)+Zt(0)=NF(etBv)−NF~(v),(20)Z_t(v)-2Z_t(v/2)+Z_t(0)=\mathcal{N}_F(e^{tB}v)-\mathcal{N}_{\widetilde{F}}(v), \tag*{(20)}

where NF~(v):=F~(v)−2F~(v/2)+F~(0)\mathcal{N}_{\widetilde{F}}(v):=\widetilde{F}(v)-2\widetilde{F}(v/2)+\widetilde{F}(0).

Fix w∈Ur0w\in U_{r_0} and set v=e−τ0Bwv=e^{-\tau_0B}w. For j=0,1j=0,1,

∥e−τjA~v∥≤eMτj+2Mτ0r0≤e3Mτ∗r0≤R.\lVert e^{-\tau_j\widetilde{A}}v\rVert\le e^{M\tau_j+2M\tau_0}r_0\le e^{3M\tau_*}r_0\le R.

The same bound holds with vv replaced by v/2v/2 or 00. The definition of ε\varepsilon therefore implies

∥Zτj(v)∥, ∥Zτj(v/2)∥, ∥Zτj(0)∥≤ε,j=0,1.\lVert Z_{\tau_j}(v)\rVert,\ \lVert Z_{\tau_j}(v/2)\rVert,\ \lVert Z_{\tau_j}(0)\rVert\le\varepsilon,\qquad j=0,1.

Subtracting Eq. (20) at the two observed step sizes eliminates NF~(v)\mathcal{N}_{\widetilde{F}}(v) and yields

∥NF(ehBw)−NF(w)∥≤8ε.\lVert\mathcal{N}_F(e^{hB}w)-\mathcal{N}_F(w)\rVert\le8\varepsilon.

Since hB∈MKhB\in\mathcal{M}_K and ∥hB∥≤2Mτ∗≤ζ\lVert hB\rVert\le2M\tau_*\le\zeta, Eq. (13) gives

∥B∥≤8εσh.(21)\lVert B\rVert\le\frac{8\varepsilon}{\sigma h}. \tag*{(21)}

To control the variation with the step size, define

Qt(B):=∫01BestB ds,0≤t≤Δ.Q_t(B):=\int_0^1 Be^{stB}\,ds,\qquad0\le t\le\Delta.

For t>0t>0, this equals (etB−I)/t(e^{tB}-I)/t. For s,t∈[0,Δ]s,t\in[0,\Delta], the exponential integral identity gives

∥etB−esB∥≤∥B∥e2MΔ∣t−s∣,(22)\lVert e^{tB}-e^{sB}\rVert\le\lVert B\rVert e^{2M\Delta}|t-s|, \tag*{(22)}
∥Qt(B)−Qs(B)∥≤12∥B∥2e2MΔ∣t−s∣≤M∥B∥e2MΔ∣t−s∣.(23)\lVert Q_t(B)-Q_s(B)\rVert\le\frac{1}{2}\lVert B\rVert^2e^{2M\Delta}|t-s|\le M\lVert B\rVert e^{2M\Delta}|t-s|. \tag*{(23)}

For v∈Ur0v\in U_{r_0} and 0≤t≤Δ0\le t\le\Delta,

∥etBv∥≤e2MΔr0≤e2MΔR.\lVert e^{tB}v\rVert\le e^{2M\Delta}r_0\le e^{2M\Delta}R.

Thus the Lipschitz bound LΔL_\Delta applies to the nonlinear terms in Eq. (19). It follows that, for s,t>0s,t>0,

∥Zt(v)−Zs(v)∥≤DΔ∥B∥∣t−s∣,DΔ:=r0(M+LΔ)e2MΔ.(24)\lVert Z_t(v)-Z_s(v)\rVert\le D_\Delta\lVert B\rVert|t-s|,\qquad D_\Delta:=r_0(M+L_\Delta)e^{2M\Delta}. \tag*{(24)}

The term F~(v)\widetilde{F}(v) cancels in this difference.

Now fix 0<δ≤Δ0<\delta\le\Delta and u∈UrΔu\in U_{r_\Delta}. Set v=eδA~uv=e^{\delta\widetilde{A}}u. By Eq. (17),

∥v∥≤eMΔrΔ=r0.\lVert v\rVert\le e^{M\Delta}r_\Delta=r_0.

Moreover,

∥e−τjA~v∥≤eMτ∗r0≤R,j=0,1,\left\|e^{-\tau_j\widetilde{A}}v\right\| \le e^{M\tau_*}r_0 \le R,\qquad j=0,1,

so ∥Zτj(v)∥≤ε\left\|Z_{\tau_j}(v)\right\| \le\varepsilon. Choose τj\tau_j to be the nearer of the two observed step sizes. Equations (21) and (24) imply

∥Sδu−S~δu∥δ=∥Zδ(v)∥≤ε+DΔ∥B∥∣δ−τj∣≤ε(1+8DΔσ∣δ−τj∣h).(25)\frac{\left\|S_\delta u-\widetilde{S}_\delta u\right\|}{\delta} =\left\|Z_\delta(v)\right\| \le\varepsilon+D_\Delta\|B\||\delta-\tau_j| \le\varepsilon\left(1+\frac{8D_\Delta}{\sigma}\frac{|\delta-\tau_j|}{h}\right). \tag*{(25)}

Write

d(δ):=min⁡τ∈[τ0,τ1]∣δ−τ∣.d(\delta):=\min_{\tau\in[\tau_0,\tau_1]}|\delta-\tau|.

The distance to the nearer endpoint satisfies

∣δ−τj∣≤h2+d(δ).|\delta-\tau_j|\le\frac{h}{2}+d(\delta).

Therefore,

1+8DΔσ∣δ−τj∣h≤(1+4DΔσ)(1+2d(δ)h).1+\frac{8D_\Delta}{\sigma}\frac{|\delta-\tau_j|}{h} \le\left(1+\frac{4D_\Delta}{\sigma}\right)\left(1+\frac{2d(\delta)}{h}\right).

Taking the supremum over u∈UrΔu\in\mathcal{U}_{r_\Delta} proves Eq. (14), with

CΔ:=1+4DΔσ.C_\Delta:=1+\frac{4D_\Delta}{\sigma}.

All constants have the dependence stated in the theorem. □

We next verify the nonlinear separation condition. The following lemma includes the scalar condition without derivative features.

Lemma 4 (Derivative-free nonlinearities). Suppose F(v)=PK[f(v)]F(v)=P_K[f(v)], where f:R→Rf:\mathbb{R}\to\mathbb{R} is locally Lipschitz. Let

Cf(s):=f(s)−2f(s/2)+f(0).C_f(s):=f(s)-2f(s/2)+f(0).

Assume that, for some ν>0\nu>0 and I=[s−,s+]I=[s_-,s_+] with 0<s−<s+<R/∣Ω∣1/20<s_-<s_+<R/|\Omega|^{1/2},

∣Cf(s)−Cf(t)∣≥ν∣s−t∣,s,t∈I.|C_f(s)-C_f(t)|\ge\nu|s-t|,\qquad s,t\in I.

Then Eq. (13) holds for some 0<r0<R0<r_0<R and ζ,ζ>0\zeta,\zeta>0. These constants can be chosen depending only on Ω,K,R,I,ν\Omega,K,R,I,\nu.

Proof. Since CfC_f is continuous and injective on II, it is strictly monotone there. Hence there exists ϵf∈{−1,1}\epsilon_f\in\{-1,1\} such that

ϵf(Cf(s)−Cf(t))(s−t)≥ν∣s−t∣2,s,t∈I.(26)\epsilon_f\bigl(C_f(s)-C_f(t)\bigr)(s-t)\ge\nu|s-t|^2,\qquad s,t\in I. \tag*{(26)}

Let

c:=s−+s+2,dI:=s+−s−2.c:=\frac{s_-+s_+}{2},\qquad d_I:=\frac{s_+-s_-}{2}.

Finite dimensionality gives a constant CK>0C_K>0 such that ∥v∥L∞≤CK∥v∥\|v\|_{L^\infty}\le C_K\|v\| for v∈Xv\in X. Choose η>0\eta>0 sufficiently small that

r0:=c∣Ω∣1/2+η<R,CKη≤dI/4.r_0:=c|\Omega|^{1/2}+\eta<R,\qquad C_K\eta\le d_I/4.

Next choose 0<ζ≤1/20<\zeta\le1/2 such that

CK(eζ−1)r0≤dI/4.C_K(e^\zeta-1)r_0\le d_I/4.

Fix H∈MKH\in\mathcal{M}_K with ∥H∥≤ζ\|H\|\le\zeta. For every real q∈Xq\in X with ∥q∥=1\|q\|=1, the fields

v±:=c1±ηqv_\pm:=c\mathbf{1}\pm\eta q

belong to Ur0U_{r_0} and take values in II. Also,

∥eHv±−v±∥L∞≤CK(eζ−1)r0≤dI/4,\lVert e^H v_{\pm}-v_{\pm}\rVert_{L^\infty}\le C_K(e^\zeta-1)r_0\le d_I/4,

so eHv±e^H v_{\pm} take values in II.

For either of these fields, put w=(eH−I)vw=(e^H-I)v. Because w∈Xw\in X and NF(v)=PK[Cf(v)]\mathcal{N}_F(v)=P_K[\mathcal{C}_f(v)], orthogonality of PKP_K and Eq. (26) imply

ϵf⟨NF(eHv)−NF(v),w⟩=ϵf∫Ω(Cf(eHv)−Cf(v))(eHv−v) dx≥ν∥w∥2.\epsilon_f\langle\mathcal{N}_F(e^H v)-\mathcal{N}_F(v),w\rangle =\epsilon_f\int_\Omega(\mathcal{C}_f(e^H v)-\mathcal{C}_f(v))(e^H v-v)\,dx \ge\nu\lVert w\rVert^2.

Consequently,

∥NF(eHv)−NF(v)∥≥ν∥(eH−I)v∥.\lVert\mathcal{N}_F(e^H v)-\mathcal{N}_F(v)\rVert\ge\nu\lVert(e^H-I)v\rVert.

Let T:=eH−IT:=e^H-I. Choose a real unit vector q∈Xq\in X attaining ∥Tq∥=∥T∥\lVert Tq\rVert=\lVert T\rVert. Since Tv+−Tv−=2ηTqTv_+-Tv_-=2\eta Tq,

max⁡{∥Tv+∥,∥Tv−∥}≥η∥T∥.\max\{\lVert Tv_+\rVert,\lVert Tv_-\rVert\}\ge\eta\lVert T\rVert.

Finally, ∥H∥≤1/2\lVert H\rVert\le1/2 gives

∥eH−I−H∥≤12e1/2∥H∥2≤12∥H∥,\lVert e^H-I-H\rVert\le\frac{1}{2}e^{1/2}\lVert H\rVert^2\le\frac{1}{2}\lVert H\rVert,

and hence ∥T∥≥∥H∥/2\lVert T\rVert\ge\lVert H\rVert/2. Combining these estimates yields

sup⁡v∈Ur0∥NF(eHv)−NF(v)∥≥νη2∥H∥.\sup_{v\in U_{r_0}}\lVert\mathcal{N}_F(e^H v)-\mathcal{N}_F(v)\rVert\ge\frac{\nu\eta}{2}\lVert H\rVert.

Thus Eq. (13) holds with σ=νη/2\sigma=\nu\eta/2.

The next lemma covers the quadratic transport nonlinearities in the KdV and advection–diffusion examples.

Lemma 5 (Quadratic transport nonlinearities). Let Ω=∏j=1dΩ(R/ℓjZ)\Omega=\prod_{j=1}^{d_\Omega}(\mathbb{R}/\ell_j\mathbb{Z}), with the Euclidean Fourier cutoff ∣k∣≤K\lvert k\rvert\le K and K≥2K\ge2. Suppose

F(v)=−PK[v(a⋅∇v)],a∈RdΩ∖{0}.F(v)=-P_K[v(a\cdot\nabla v)],\qquad a\in\mathbb{R}^{d_\Omega}\setminus\{0\}.

Then, for every r0>0r_0>0, there exist σ,ζ>0\sigma,\zeta>0 such that Eq. (13) holds.

Proof. Write D:=a⋅∇D:=a\cdot\nabla. Since FF is homogeneous of degree two,

NF(v)=12F(v)=−14PKD(v2).\mathcal{N}_F(v)=\frac{1}{2}F(v)=-\frac{1}{4}P_KD(v^2).

Its derivative satisfies

DNF(v)[Hv]=−12PKD(vHv).(27)D\mathcal{N}_F(v)[Hv]=-\frac{1}{2}P_KD(vHv). \tag*{(27)}

We first show that

DNF(v)[Hv]=0for all real v∈X⟹H=0.D\mathcal{N}_F(v)[Hv]=0\quad\text{for all real }v\in X \qquad\Longrightarrow\qquad H=0.

Suppose the left-hand side holds. Polarization of Eq. (27) gives

PKD(uHv+vHu)=0,u,v∈X.P_KD(uHv+vHu)=0,\qquad u,v\in X.

This bilinear identity extends to the complexification of XX. Let b(k)b(k) denote the Fourier symbol of HH, and set αj:=2πaj/ℓj\alpha_j:=2\pi a_j/\ell_j. Testing the identity on the Fourier modes indexed by pp and qq gives

(α⋅(p+q))(b(p)+b(q))=0,∣p∣,∣q∣,∣p+q∣≤K.(28)(\alpha\cdot(p+q))(b(p)+b(q))=0,\qquad\lvert p\rvert,\lvert q\rvert,\lvert p+q\rvert\le K. \tag*{(28)}

Choose ii with αi≠0\alpha_i\ne0. Taking p=q=eip=q=e_i is allowed because K≥2K\ge2, and gives b(ei)=0b(e_i)=0. Taking p=0p=0, q=eiq=e_i then gives b(0)=0b(0)=0. For every retained mode kk with α⋅k≠0\alpha\cdot k\ne0, taking p=0p=0, q=kq=k gives

b(k)=0.b(k)=0.

It remains to consider a nonzero retained mode kk satisfying α⋅k=0\alpha\cdot k = 0. If αiki≠0\alpha_i k_i \ne0 for some ii, choose p=−sgn⁡(ki)eip = -\operatorname{sgn}(k_i)e_i. Then

α⋅p≠0,∣k+p∣2=∣k∣2−2∣ki∣+1≤∣k∣2.\alpha\cdot p \ne0,\qquad\lvert k+p\rvert^2=\lvert k\rvert^2-2\lvert k_i\rvert+1\leq\lvert k\rvert^2.

Thus pp and k+pk+p are retained and have nonzero α\alpha-frequency. Equation (28) with q=kq=k therefore implies b(k)=0b(k)=0.

Otherwise, αiki=0\alpha_i k_i=0 for every ii. Choose i,ji,j with αi≠0\alpha_i\ne0 and kj≠0k_j\ne0. Then ki=0k_i=0, αj=0\alpha_j=0, and i≠ji\ne j. Set p=ei−sgn⁡(kj)ejp=e_i-\operatorname{sgn}(k_j)e_j. Now

∣p∣=2≤K,α⋅p=αi≠0,∣k+p∣2=∣k∣2−2∣kj∣+2≤∣k∣2.\lvert p\rvert=\sqrt{2}\leq K,\qquad\alpha\cdot p=\alpha_i\ne0,\qquad\lvert k+p\rvert^2=\lvert k\rvert^2-2\lvert k_j\rvert+2\leq\lvert k\rvert^2.

The same use of Eq. (28) gives b(k)=0b(k)=0. Hence b(k)=0b(k)=0 for every retained mode, proving H=0H=0.

By finite dimensionality and compactness, this injectivity implies

κK:=min⁡H∈MK∥H∥=1 sup⁡∥v∥≤1∥DNF(v)[Hv]∥>0.\kappa_K:=\min_{\substack{H\in M_K\\ \lVert H\rVert=1}}\ \sup_{\lVert v\rVert\leq1}\lVert DN_F(v)[Hv]\rVert>0.

Homogeneity in vv and HH therefore gives

sup⁡∥v∥≤r0∥DNF(v)[Hv]∥≥κKr02∥H∥.(29)\sup_{\lVert v\rVert\leq r_0}\lVert DN_F(v)[Hv]\rVert\geq\kappa_K r_0^2\lVert H\rVert. \tag*{(29)}

To pass to finite perturbations, introduce the continuous symmetric bilinear map

B(v,w):=−14PKD(vw),NF(v)=B(v,v).B(v,w):=-\frac{1}{4}P_KD(vw),\qquad N_F(v)=B(v,v).

Choose CB>0C_B>0 such that ∥B(v,w)∥≤CB∥v∥∥w∥\lVert B(v,w)\rVert\leq C_B\lVert v\rVert\lVert w\rVert. For ∥H∥≤1\lVert H\rVert\leq1, set T=eH−IT=e^H-I. The bounds

∥T∥≤e∥H∥,∥T−H∥≤e2∥H∥2\lVert T\rVert\leq e\lVert H\rVert,\qquad\lVert T-H\rVert\leq\frac{e}{2}\lVert H\rVert^2

and the quadratic identity give

NF(eHv)−NF(v)−DNF(v)[Hv]=2B(v,(T−H)v)+B(Tv,Tv).N_F(e^H v)-N_F(v)-DN_F(v)[Hv] =2B(v,(T-H)v)+B(Tv,Tv).

Consequently,

∥NF(eHv)−NF(v)−DNF(v)[Hv]∥≤C0∥v∥2∥H∥2,C0:=CB(e+e2).\lVert N_F(e^H v)-N_F(v)-DN_F(v)[Hv]\rVert\leq C_0\lVert v\rVert^2\lVert H\rVert^2,\qquad C_0:=C_B(e+e^2).

Together with Eq. (29), this yields

sup⁡∥v∥≤r0∥NF(eHv)−NF(v)∥≥r02(κK−C0∥H∥)∥H∥.\sup_{\lVert v\rVert\leq r_0}\lVert N_F(e^H v)-N_F(v)\rVert\geq r_0^2(\kappa_K-C_0\lVert H\rVert)\lVert H\rVert.

Taking

ζ:=min⁡{1,κK/(2C0)},σ:=κKr02/2\zeta:=\min\{1,\kappa_K/(2C_0)\},\qquad\sigma:=\kappa_Kr_0^2/2

proves the claim.

For the KdV equation

ut+auux+buxxx=0,a≠0,u_t+auu_x+bu_{xxx}=0,\qquad a\ne0,

take

A=−b∂x3,F(v)=−aPK[vvx].A=-b\partial_x^3,\qquad F(v)=-aP_K[vv_x].

For the two-dimensional equation

ut=ν(uxx+uyy)−u(ux+uy),ν=0.05,u_t=\nu(u_{xx}+u_{yy})-u(u_x+u_y),\qquad\nu=0.05,

on the periodic domain [0,1]2[0,1]^2, take

A=ν(∂x2+∂y2),F(v)=−PK[v(vx+vy)].A=\nu(\partial_x^2+\partial_y^2),\qquad F(v)=-P_K[v(v_x+v_y)].

In both cases, Lemma 5 verifies Eq. (13) for K≥2K \ge2. The linear operators are bounded on XX, and the projected quadratic nonlinearities are locally Lipschitz. Thus both equations provide reference models satisfying the hypotheses of Theorem 1. The constants may depend on the fixed spatial resolution KK.

Finally, we establish the sharpness assertion in Theorem 1. We show that the linear growth in

min⁡τ∈[τ0,τ1]∣δ−τ∣τ1−τ0\frac{\min_{\tau\in[\tau_0,\tau_1]} \lvert\delta-\tau\rvert}{\tau_1-\tau_0}

cannot, in general, be replaced by sublinear growth with constants having the dependence stated in the theorem. Take the reference model

A=0,F(v)=Q(v):=PK[v2].A=0,\qquad F(v)=Q(v):=P_K[v^2].

For f(s)=s2f(s)=s^2, the scalar contrast is Cf(s)=s2/2C_f(s)=s^2/2. On every positive interval I=[s−,s+]I=[s_-,s_+],

∣Cf(s)−Cf(t)∣=s+t2∣s−t∣≥s−∣s−t∣.\lvert C_f(s)-C_f(t)\rvert=\frac{s+t}{2}\lvert s-t\rvert\ge s_-\lvert s-t\rvert.

Lemma 4 therefore verifies the nonlinear separation condition.

Fix the reference model, the parameters in the theorem, and the resulting constants τ∗\tau_\ast and rΔr_\Delta. Choose

0<a<τ∗/3,0<∣γ∣≤M,0<a<\tau_\ast/3,\qquad0<\lvert\gamma\rvert\le M,

and define a fixed comparison model by

A~=γI,F~(v)=e−2γaQ(v)+χv,χ:=e−γa−1a.\widetilde{A}=\gamma I,\qquad\widetilde{F}(v)=e^{-2\gamma a}Q(v)+\chi v,\qquad\chi:=\frac{e^{-\gamma a}-1}{a}.

This is a model of the same form, with a locally Lipschitz scalar nonlinearity. A direct calculation gives

Dt(v):=S~tv−Stvt=α(t)v+β(t)Q(v),(30)D_t(v):=\frac{\widetilde{S}_t v-S_t v}{t}=\alpha(t)v+\beta(t)Q(v), \tag*{(30)}

where

α(t):=(1+tχ)eγt−1t,β(t):=e2γ(t−a)−1.\alpha(t):=\frac{(1+t\chi)e^{\gamma t}-1}{t},\qquad\beta(t):=e^{2\gamma(t-a)}-1.

Both coefficients vanish at t=at=a, so the two models agree exactly at that step size.

Now take observed steps

τ0=a,τ1=a+h,0<h<a/2,\tau_0=a,\qquad\tau_1=a+h,\qquad0<h<a/2,

and let h↓0h\downarrow0, keeping both models fixed. The functions α\alpha and β\beta are continuously differentiable on [a,3a/2][a,3a/2], and QQ is bounded on URU_R. Consequently, for a constant C1C_1 independent of hh,

εobs=sup⁡v∈UR∥Da+h(v)∥≤C1h.(31)\varepsilon_{\mathrm{obs}}=\sup_{v\in U_R}\lVert D_{a+h}(v)\rVert\le C_1h. \tag*{(31)}

Consider the fixed inference step δ=2a≤Δ\delta=2a\le\Delta. Since Q(−v)=Q(v)Q(-v)=Q(v),

Dδ(v)+Dδ(−v)2=β(δ)Q(v).\frac{D_\delta(v)+D_\delta(-v)}{2}=\beta(\delta)Q(v).

Testing the constant fields v=±rΔ/∣Ω∣1/2v=\pm r_\Delta/|\Omega|^{1/2} yields

sup⁡v∈UrΔ∥Dδ(v)∥≥∣e2γa−1∣rΔ2∣Ω∣1/2=:c1>0.(32)\sup_{v\in U_{r_\Delta}}\lVert D_\delta(v)\rVert\ge\lvert e^{2\gamma a}-1\rvert\frac{r_\Delta^2}{|\Omega|^{1/2}}=:c_1>0. \tag*{(32)}

On the other hand,

qh:=min⁡τ∈[a,a+h]∣2a−τ∣h=a−hh⟶∞.q_h:=\frac{\min_{\tau\in[a,a+h]}\lvert2a-\tau\rvert}{h}=\frac{a-h}{h}\longrightarrow\infty.

If the amplification factor in the theorem could be replaced by 1+ψ(qh)1+\psi(q_h) with ψ(q)=o(q)\psi(q)=o(q), then Eq. (31) would make its right-hand side tend to zero:

Cεobs(1+ψ(qh))≤CC1h(1+ψ(qh))⟶0.C\varepsilon_{\mathrm{obs}}(1+\psi(q_h))\le CC_1h(1+\psi(q_h))\longrightarrow0.

This contradicts Eq. (32). All model parameters and permitted constant dependencies remain fixed as h↓0h\downarrow0.

The same argument with the fixed inference step δ=a/2\delta=a/2 establishes the corresponding sharpness for extrapolation below the observed interval. Thus the linear growth order is optimal in general; no optimality of the numerical coefficient 2 is asserted.

Proofs of Section 2.3.2

Write n:=dΩn := d_{\Omega} and Jxϕ:=Φst[ϕ](x)J_x\phi:= \Phi_{\mathrm{st}}[\phi](x). Throughout this section, StS_t denotes the original one-step map

St[ϕ;g,f](x)=(etAϕ)(x)+tf(JxetAϕ,η(x)).(33)S_t[\phi;g,f](x) = (e^{tA}\phi)(x) + tf(J_xe^{tA}\phi,\eta(x)). \tag*{(33)}

In particular, there is no projection of the final nonlinear output in this definition. We write the Fourier modes as eik⋅xe^{\mathrm{i}k\cdot x} on the normalised torus. On a periodic box with side lengths ℓj\ell_j, replace kjk_j by 2πkj/ℓj2\pi k_j/\ell_j in the differential symbols. This invertible rescaling preserves the polynomial degrees and the arguments below; the bandwidth bounds refer to integer mode indices.

Proof of Lemma 2. Let v=eτ0Aϕv=e^{\tau_0A}\phi. Since gγ=g+γg_\gamma=g+\gamma, the corresponding linear operator is Aγ=A+γIA_\gamma=A+\gamma I, and hence

eτ0Aγϕ=ϱγv,Jx(ϱγv)=ϱγJxv.e^{\tau_0A_\gamma}\phi=\varrho_\gamma v,\qquad J_x(\varrho_\gamma v)=\varrho_\gamma J_xv.

Substitution into (33) gives

Sτ0[ϕ;gγ,fγ](x)=ϱγv(x)+τ0f(Jxv,η(x))+τ0χγϱγv(x)= v(x)+τ0f(Jxv,η(x)),S_{\tau_0}[\phi;g_\gamma,f_\gamma](x)=\varrho_\gamma v(x)+\tau_0f(J_xv,\eta(x))+\tau_0\chi_\gamma\varrho_\gamma v(x) =\ v(x)+\tau_0f(J_xv,\eta(x)),

where the last equality uses ϱγ(1+τ0χγ)=1\varrho_\gamma(1+\tau_0\chi_\gamma)=1. This proves (16). Repeated composition also gives equality of all rollouts for which both iterates are defined.

For the identifiability theorem, the following explicit choice of bandwidth suffices:

D:=max⁡{1,d},D:=\max\{1,d\},
N∗:=max⁡{(2ρ+1)D+2ρ+1, ρ+(ρ+2)(D+2ρ+1)},N_*:=\max\{(2\rho+1)D+2\rho+1,\ \rho+(\rho+2)(D+2\rho+1)\},
T∗:=⌊N∗−12⌋,T_*:=\left\lfloor\frac{N_*-1}{2}\right\rfloor,
K0:=⌈nmax⁡{ρ,DT∗}⌉.(34)K_0:=\left\lceil\sqrt{n}\max\{\rho,DT_*\}\right\rceil. \tag*{(34)}

We first establish the two properties supplied by this threshold.

Lemma 6 (Uniform realisation of local features). If K≥ρnK\geq\rho\sqrt{n}, then Jx:VK→RmstJ_x:\mathcal{V}_K\to\mathbb{R}^{m_{\mathrm{st}}} is surjective for every x∈Ωx\in\Omega. Moreover, there are linear right inverses Rx:Rmst→VKR_x:\mathbb{R}^{m_{\mathrm{st}}}\to\mathcal{V}_K satisfying

JxRx=I,CJ:=sup⁡x∈Ω∥Rx∥<∞,J_xR_x=I,\qquad C_J:=\sup_{x\in\Omega}\|R_x\|<\infty,

where the operator norms use the Euclidean norm on the feature space and the L2L^2 norm on VK\mathcal{V}_K.

Proof. Suppose a real linear functional annihilates the range of JxJ_x. Writing its coefficients as aαa_\alpha, this means

∑∣α∣≤ρaα∂αϕ(x)=0,ϕ∈VK.\sum_{|\alpha|\leq\rho}a_\alpha\partial^\alpha\phi(x)=0,\qquad\phi\in\mathcal{V}_K.

Complexifying this real linear identity and testing Fourier modes gives

∑∣α∣≤ρaα(ik)α=0,k∈Zn,∣k∣≤K.\sum_{|\alpha|\leq\rho}a_\alpha(\mathrm{i}k)^\alpha=0,\qquad k\in\mathbb{Z}^n,\quad|k|\leq K.

The retained frequencies contain the tensor grid {0,…,ρ}n\{0,\ldots,\rho\}^n. A polynomial of degree at most ρ\rho in each coordinate that vanishes on this grid is identically zero, by repeated univariate interpolation. Thus all aαa_\alpha vanish, proving surjectivity. Choose a linear right inverse at one point. Translating its output produces right inverses at all other points, because differentiation commutes with translations. Translations preserve the L2L^2 norm, so these right inverses have the same operator norm.

The next lemma supplies the finite-frequency rigidity needed to identify the differential symbol.

Lemma 7 (Finite-frequency rigidity). Let HH, PP, QQ be complex polynomials in nn real variables, with degrees at most dd, ρ\rho, ρ\rho, respectively, and suppose P(0)≠0P(0) \ne0. Assume that, for every k∈Znk \in\mathbb{Z}^n with ∣k∣≤K\lvert k\rvert\le K,

eH(k)P(k)=Q(k),∣Im⁡H(k)∣<π/2.e^{H(k)}P(k)=Q(k), \qquad\lvert\operatorname{Im} H(k)\rvert< \pi/2.

If K≥K0K \ge K_0 with K0K_0 given by Eq. (34), then HH is constant.

Proof. We first prove a univariate statement. Let hh, pp, qq be complex polynomials of degrees at most DD, ρ\rho, ρ\rho, with pp and qq nonzero. Suppose that at least N∗N_* distinct real points tjt_j satisfy

eh(tj)p(tj)=q(tj),∣Im⁡h(tj)∣<π/2.(35)e^{h(t_j)}p(t_j)=q(t_j), \qquad\lvert\operatorname{Im} h(t_j)\rvert< \pi/2. \tag*{(35)}

We show that hh is constant.

Write h(t)=a(t)+ib(t)h(t)=a(t)+ib(t) for real tt, where aa and bb are real polynomials. Taking squared moduli in Eq. (35) gives zeros of

R(t):=e2a(t)∣p(t)∣2−∣q(t)∣2.R(t):=e^{2a(t)}\lvert p(t)\rvert^2-\lvert q(t)\rvert^2.

Here ∣p(t)∣2\lvert p(t)\rvert^2 and ∣q(t)∣2\lvert q(t)\rvert^2 are real polynomials in tt of degrees at most m:=2ρm:=2\rho.

Suppose first that aa is nonconstant. Differentiating RR exactly m+1m+1 times eliminates the second polynomial and gives

R(m+1)(t)=e2a(t)W(t),deg⁡W≤m+(m+1)(D−1).R^{(m+1)}(t)=e^{2a(t)}W(t), \qquad\deg W \le m+(m+1)(D-1).

The polynomial WW is nonzero: at each differentiation, the polynomial factor is transformed by U↦U′+2a′UU\mapsto U'+2a'U, which maps every nonzero polynomial to a nonzero polynomial when a′a' is a nonzero polynomial. Repeated Rolle’s theorem therefore bounds the number of distinct real zeros of RR by

m+(m+1)(D−1)+(m+1)=m+(m+1)D.m+(m+1)(D-1)+(m+1)=m+(m+1)D.

This contradicts the definition of N∗N_*. Hence a(t)=a0a(t)=a_0 is constant. The polynomial ∣q(t)∣2−e2a0∣p(t)∣2\lvert q(t)\rvert^2-e^{2a_0}\lvert p(t)\rvert^2 then has more zeros than its degree, so

∣q(t)∣2=e2a0∣p(t)∣2for all real t.(36)\lvert q(t)\rvert^2=e^{2a_0}\lvert p(t)\rvert^2 \qquad\text{for all real }t. \tag*{(36)}

Cancel the common polynomial factors in q/pq/p and write the result as U/VU/V, with UU and VV coprime. For a polynomial WW, let W#(z):=W(z‾)‾W^\#(z):=\overline{W(\overline{z})}. Equation (36) implies the polynomial identity

UU#=e2a0VV#.UU^\#=e^{2a_0}VV^\#.

In particular, deg⁡U=deg⁡V\deg U=\deg V. Coprimality implies that UU divides V#V^\#, and therefore

U=CV#,∣C∣=ea0.U=CV^\#, \qquad\lvert C\rvert=e^{a_0}.

The polynomial VV has no real roots, since any such root would also be a root of UU. Consequently, the reduced rational function admits a smooth argument on the real line:

e−a0U(t)V(t)=eiθ(t).e^{-a_0}\frac{U(t)}{V(t)}=e^{i\theta(t)}.

If the roots of VV are z1,…,zℓz_1,\ldots,z_\ell, counted with multiplicity, then ℓ≤ρ\ell\le\rho, Im⁡zj≠0\operatorname{Im}z_j\ne0, and

θ′(t)=−2∑j=1ℓIm⁡zj(t−Re⁡zj)2+(Im⁡zj)2.(37)\theta'(t)=-2\sum_{j=1}^{\ell}\frac{\operatorname{Im}z_j}{(t-\operatorname{Re}z_j)^2+(\operatorname{Im}z_j)^2}. \tag*{(37)}

In particular,

∫R∣θ′(t)∣ dt≤2πρ.\int_{\mathbb{R}}\lvert\theta'(t)\rvert\,dt\le2\pi\rho.

At most ρ\rho of the sample points are zeros of pp. At each remaining point, Eq. (35) gives

θ(tj)−b(tj)∈2πZ.\theta(t_j)-b(t_j)\in2\pi\mathbb{Z}.

The range of θ\theta has width at most 2πρ2\pi\rho, while b(tj)∈(−π/2,π/2)b(t_j) \in(-\pi/2,\pi/2). Thus at most ρ+2\rho+2 distinct integers can occur in the preceding identity. By the definition of N∗N_*, one integer ss occurs at least D+2ρ+1D+2\rho+1 times. The function

θ(t)−b(t)−2πs\theta(t)-b(t)-2\pi s

therefore has at least that many distinct zeros.

By Eq. (37), θ′−b′\theta'-b' is a rational function with a strictly positive real denominator of degree at most 2ρ2\rho and a numerator of degree at most D+2ρ−1D+2\rho-1. Rolle’s theorem supplies at least D+2ρD+2\rho distinct zeros of this numerator, so it vanishes identically. Hence b′=θ′b'=\theta' on R\mathbb{R}. Since θ′(t)→0\theta'(t)\to0 as ∣t∣→∞\lvert t\rvert\to\infty, the polynomial b′b' must be zero. Both aa and bb are constant, proving the univariate statement.

We now return to the multivariate polynomials. The identity at k=0k=0 implies Q(0)≠0Q(0)\ne0. For each nonzero direction v∈{0,…,D}nv\in\{0,\ldots,D\}^{n}, restrict the polynomials to

hv(t):=H(tv),pv(t):=P(tv),qv(t):=Q(tv).h_v(t):=H(tv),\qquad p_v(t):=P(tv),\qquad q_v(t):=Q(tv).

The polynomials pvp_v and qvq_v are nonzero because their values at zero are nonzero. For every integer tt with ∣t∣≤T∗\lvert t\rvert\le T_*,

∣tv∣≤T∗Dn≤K.\lvert tv\rvert\le T_*D\sqrt{n}\le K.

Thus these restrictions satisfy the univariate hypotheses at 2T∗+1≥N∗2T_*+1\ge N_* distinct points. It follows that H(tv)H(tv) is constant in tt, and in particular H(v)=H(0)H(v)=H(0).

Therefore H−H(0)H-H(0) vanishes on the tensor grid {0,…,D}n\{0,\ldots,D\}^{n}. Its degree in each coordinate is at most DD, so repeated univariate interpolation gives H≡H(0)H\equiv H(0).

Proof of Theorem 3. We first prove part (i). Choose x0x_0 as in its nondegeneracy assumption and set η0:=η(x0)\eta_0:=\eta(x_0). Differentiate the equality of the prediction maps at the zero input and evaluate at x0x_0. For every real ψ∈VK\psi\in V_K, this gives

(eτ0Aψ)(x0)+τ0∑∣α∣≤ρ∂zαf(0,η0) ∂α(eτ0Aψ)(x0)=(eτ0A′ψ)(x0)+τ0∑∣α∣≤ρ∂zαf′(0,η0) ∂α(eτ0A′ψ)(x0).\left(e^{\tau_0A}\psi\right)(x_0)+\tau_0\sum_{\lvert\alpha\rvert\le\rho}\partial_{z_\alpha}f(0,\eta_0)\,\partial^\alpha\left(e^{\tau_0A}\psi\right)(x_0) = \left(e^{\tau_0A'}\psi\right)(x_0)+\tau_0\sum_{\lvert\alpha\rvert\le\rho}\partial_{z_\alpha}f'(0,\eta_0)\,\partial^\alpha\left(e^{\tau_0A'}\psi\right)(x_0).

Define the polynomials

P(k):=1+τ0∑∣α∣≤ρ∂zαf(0,η0)(ik)α,Q(k):=1+τ0∑∣α∣≤ρ∂zαf′(0,η0)(ik)α.P(k):=1+\tau_0\sum_{\lvert\alpha\rvert\le\rho}\partial_{z_\alpha}f(0,\eta_0)(\mathrm{i}k)^\alpha,\qquad Q(k):=1+\tau_0\sum_{\lvert\alpha\rvert\le\rho}\partial_{z_\alpha}f'(0,\eta_0)(\mathrm{i}k)^\alpha.

Complexifying the linear identity and testing Fourier modes yields

eτ0g(k)P(k)=eτ0g′(k)Q(k),∣k∣≤K.(38)e^{\tau_0g(k)}P(k)=e^{\tau_0g'(k)}Q(k),\qquad\lvert k\rvert\le K. \tag*{(38)}

The nondegeneracy assumption states precisely that

P(0)=1+τ0∂z0f(0,η0)≠0.P(0)=1+\tau_0\partial_{z_0}f(0,\eta_0)\ne0.

Apply Lemma 7 to

H(k):=τ0(g(k)−g′(k)).H(k):=\tau_0\left(g(k)-g'(k)\right).

Indeed, Eq. (38) gives eH(k)P(k)=Q(k)e^{H(k)}P(k)=Q(k), and the two anti-aliasing bounds imply

∣Im⁡H(k)∣≤τ0∣Im⁡g(k)∣+τ0∣Im⁡g′(k)∣<π/2.\lvert\operatorname{Im}H(k)\rvert\le\tau_0\lvert\operatorname{Im}g(k)\rvert+\tau_0\lvert\operatorname{Im}g'(k)\rvert<\pi/2.

Hence g−g′g-g' is constant. Since both differential symbols have real coefficients, their values at zero are real. Thus, for some γ∈R\gamma\in\mathbb{R},

g′=g+γ,A′=A+γI.g'=g+\gamma,\qquad A'=A+\gamma I.

Equality of the positive-order polynomial coefficients follows immediately.

It remains to identify the nonlinear function on a common feature neighbourhood. Let Rx\mathcal{R}_x and CJC_J be as in Lemma 6. For any step size tt at which the prediction maps agree on the stated input ball, define

rt:=r2CJ∥e−tA′∥VK→VK.r_t := \frac{r}{2C_J\lVert e^{-tA'}\rVert_{V_K\to V_K}}.

This is positive because the exponential is invertible on the finite-dimensional space VKV_K. For ∣zst∣<rt\lvert z_{\mathrm{st}}\rvert< r_t, take

ϕ:=e−tA′Rxzst.\phi:= e^{-tA'}\mathcal{R}_x z_{\mathrm{st}}.

Then ∥ϕ∥L2<r\lVert\phi\rVert_{L^2}<r, and

JxetA′ϕ=zst,JxetAϕ=e−γtzst.J_xe^{tA'}\phi=z_{\mathrm{st}},\qquad J_xe^{tA}\phi=e^{-\gamma t}z_{\mathrm{st}}.

Equality of the two prediction maps at xx therefore gives

f′(zst,η(x))=f(e−γtzst,η(x))+e−γt−1tz0,∣zst∣<rt.(39)f'(z_{\mathrm{st}},\eta(x))=f(e^{-\gamma t}z_{\mathrm{st}},\eta(x))+\frac{e^{-\gamma t}-1}{t}z_0,\qquad\lvert z_{\mathrm{st}}\rvert<r_t. \tag*{(39)}

The radius rtr_t is independent of xx. Since E=η(Ω)E=\eta(\Omega), this identity holds for every zη∈Ez_\eta\in E. Taking t=τ0t=\tau_0 and r∗=rτ0r_*=r_{\tau_0} proves f′=fγf'=f_\gamma on Br∗(0)×EB_{r_*}(0)\times E, completing part (i).

For part (ii), put

aj:=e−γτj,χj:=aj−1τj,j=0,1.a_j:=e^{-\gamma\tau_j},\qquad\chi_j:=\frac{a_j-1}{\tau_j},\qquad j=0,1.

Equation (39) applies at both observed step sizes. Thus, on a common neighbourhood of the origin and for every zη∈Ez_\eta\in E,

f(a0zst,zη)+χ0z0=f(a1zst,zη)+χ1z0.(40)f(a_0z_{\mathrm{st}},z_\eta)+\chi_0z_0=f(a_1z_{\mathrm{st}},z_\eta)+\chi_1z_0. \tag*{(40)}

Suppose γ≠0\gamma\ne0, and choose zη∗z_\eta^* as in the non-affinity assumption. Since τ0≠τ1\tau_0\ne\tau_1,

λ:=a0a1>0,λ≠1.\lambda:=\frac{a_0}{a_1}>0,\qquad\lambda\ne1.

Setting w=a1zstw=a_1z_{\mathrm{st}} in Eq. (40) gives

f(λw,zη∗)−f(w,zη∗)=bw0,b:=χ1−χ0a1,f(\lambda w,z_\eta^*)-f(w,z_\eta^*)=bw_0,\qquad b:=\frac{\chi_1-\chi_0}{a_1},

for all ww sufficiently close to zero. Define

H∗(w):=f(w,zη∗)−bλ−1w0.H_*(w):=f(w,z_\eta^*)-\frac{b}{\lambda-1}w_0.

Then H∗(λw)=H∗(w)H_*(\lambda w)=H_*(w) locally. Writing μ:=min⁡{λ,λ−1}∈(0,1)\mu:=\min\{\lambda,\lambda^{-1}\}\in(0,1), this implies, on a sufficiently small ball,

H∗(μw)=H∗(w).H_*(\mu w)=H_*(w).

Iteration and continuity at zero yield

H∗(w)=H∗(μmw)⟶H∗(0)(m→∞).H_*(w)=H_*(\mu^m w)\longrightarrow H_*(0)\qquad(m\to\infty).

Consequently,

f(w,zη∗)=f(0,zη∗)+bλ−1w0f(w,z_\eta^*)=f(0,z_\eta^*)+\frac{b}{\lambda-1}w_0

on a neighbourhood of the origin. This contradicts the assumed non-affinity. Therefore γ=0\gamma=0. Part (i) now gives g′=gg'=g and f′=ff'=f on its original feature domain Br∗(0)×EB_{r_*}(0)\times E.

Numerical Experiments

Main ENNO hyperparameters

Tables 9 and 10 list, for each of the four experiments, the architecture of ENNO and the data it was trained and evaluated on. Throughout, KK is the Euclidean cutoff on integer mode indices, ∣k∣≤K|k| \le K, as in § 2.3.1; dd is the degree of the symbol gg in Eq. (6); ρ\rho is the order of the jet entering the nonlinear map ff in Eq. (8); and vv denotes the linearly propagated field to which ff is applied in Eq. (9). The coefficients of the even monomials in Re⁡g\operatorname{Re} g (§ 2.2) are parametrised as −softplus⁡(θ)-\operatorname{softplus}(\theta), so that Re⁡g≤c0\operatorname{Re} g \le c_0 holds by construction; the odd-order coefficients and c0c_0 are unconstrained. One consequence is that a constrained coefficient whose true value is zero cannot reach zero exactly and settles at the floor of the parametrisation; this is the case for c2c_2 in Table 2. The unconstrained coefficients are initialised from N(0,0.012)\mathcal{N}(0, 0.01^2). The nonlinear map ff and, where they are not the identity, the encoder EE and the decoder RR are multilayer perceptrons with Tanh activations; the output layer of ff is zero-initialised, so that training starts from the purely linear flow.

ENNO is trained with Adam for at most 3,000 epochs (the cap applied to all neural models in § 3), and its checkpoint is selected by early stopping on the validation set (patience 15 for the synthetic problems, evaluated on the noised validation trajectories, and 200 for the real data; for WildfireSpreadTS the criterion is validation AUC-PR, § 3.4). All reported results come from a single training seed per configuration (per fold for WildfireSpreadTS). Relative L2L^2 errors on the synthetic problems are computed against the clean, noise-free test trajectories; the targets at t1/2t_{1/2} come from the same solver runs at the finer output interval (Table 10).

KdVAdvection–diffusionSEVIRWildfireSpreadTS
Domain and grid[0,10][0, 10] periodic, 128 points[0,1]2[0, 1]^2 periodic, 128×128128 \times128384×384 km2384 \times384\ \mathrm{km}^2; VIL on 3842384^2 pixels, average-pooled to 1282128^2daily fire maps at 375 m, padded to 360×320360 \times320
SourceFourier pseudospectral in space, DOP853 in time (rtol 10−810^{-8}); 10,001 frames over T=1T = 1, every 10th keptexplicit Euler, Δt=10−5\Delta t = 10^{-5}, central differences in space; 5,001 frames over T=0.5T = 0.5, every 10th keptreal radar, 49 frames per eventreal observations, 2018–2021
Initial conditions3 sine modes (amplitude U(−1,1)/m2\mathcal{U}(-1,1)/m^2 for mode mm) plus 1 or 2 solitons (amplitude U(1.5,3.5)\mathcal{U}(1.5,3.5))sums of 1–3 Gaussian bumps (amplitude U(0.5,1.5)\mathcal{U}(0.5,1.5), σ∼U(0.05,0.25)\sigma\sim\mathcal{U}(0.05,0.25)) and low-mode Fourier fields——
Observation interval10−310^{-3}10−310^{-3}5 min1 day
Training step sizes τ1,τ2\tau_1, \tau_210−310^{-3}, 2×10−32 \times10^{-3}10−310^{-3}, 2×10−32 \times10^{-3}5, 10 min1, 2 days
Targets at t1/2t_{1/2}every 5th frame of the same runsevery 5th frame of the same runs——
Train / val / test2,030 / 50 / 200 trajectories100 / 10 / 25 trajectories11,159 / 200 / 500 eventsfour leave-one-year-out folds
Noise (train and val, all methods)i.i.d. Gaussian, 5% of each trajectory’s std, fixed seed; test trajectories cleansamenonenone

Table 10. Data used in the four experiments.

KdVAdvection–diffusionSEVIR (VIL)WildfireSpreadTS
Latent channels MM11204 (Tin=1T_{\mathrm{in}} = 1) / 20 (Tin=5T_{\mathrm{in}} = 5)
History length (frames)1110Tin∈{1,5}T_{\mathrm{in}} \in\{1, 5\}
Encoder EE / decoder RRidentity (learnable variants in Table 11)identitypointwise MLPs R10→R20\mathbb{R}^{10} \to\mathbb{R}^{20} and R20→R\mathbb{R}^{20} \to\mathbb{R}, 5 hidden layers ×\times 256same design, RTin→RM\mathbb{R}^{T_{\mathrm{in}}} \to\mathbb{R}^{M} and RM→R\mathbb{R}^{M} \to\mathbb{R}, 5 hidden layers ×\times 256
Cutoff KK (∣k∣≤K|k| \le K)164646432
Symbol degree dd3 (ablation over 1–4, Table 1)2 (ablation over 1–3, Table 4)66
Constant term c0c_0trainedtrainedfixed at 0fixed at 0
Jet order ρ\rho2211
Inputs of ffvv, vxv_x, vxxv_{xx}vv, vxv_x, vyv_y, vxxv_{xx}, vyyv_{yy}, vxyv_{xy}vv, vxv_x, vyv_yvv, vxv_x, vyv_y and 25 environmental channels η\eta
Size of ff4 hidden layers ×\times 1284×1284 \times1284×2564 \times2564×2564 \times256

Table 9. Architecture of ENNO in the four experiments.

Ablation results of the KdV example

The encoder and decoder of Eq. (12) are identity maps in the configuration reported above, so the model acts directly on uu. Table 11 makes them learnable, with MM latent fields, at each symbol degree. Where the degree is below the true order, the learnable variant with M≥4M \ge4 is lower at every target time, and the gain saturates between M=4M = 4 and M=16M = 16: at d=1d = 1 it reaches 2.449% at t10t_{10} against 4.189% for the identity. At d≥3d \ge3, where the true symbol is already in the class, the pattern splits: the identity is lower at t1t_1 and t5t_5, by a factor of 2.7 at t1t_1, while the learnable variant grows more slowly and ends slightly lower at t10t_{10}. We report the identity configuration, which is the more accurate of the two over most of the horizon.

Variantt1t_1t5t_5t10t_{10}
d=1d = 1, identity2.8983.6834.189
d=1d = 1, M=1M = 13.0692.6293.276
d=1d = 1, M=4M = 41.8662.3172.468
d=1d = 1, M=16M = 161.8772.2892.449
d=2d = 2, identity2.1812.7002.994
d=2d = 2, M=1M = 11.9902.4542.856
d=2d = 2, M=4M = 41.5322.2332.529
d=2d = 2, M=16M = 161.5352.2352.529
d=3d = 3, identity0.3130.7211.115
d=3d = 3, M=1M = 10.8520.9150.994
d=3d = 3, M=4M = 40.8460.9391.028
d=3d = 3, M=16M = 160.8420.9351.014
d=4d = 4, identity0.3180.7381.152
d=4d = 4, M=1M = 10.8460.9201.007
d=4d = 4, M=4M = 40.8400.9681.086
d=4d = 4, M=16M = 160.8320.9631.098

Table 11. KdV: encoder/decoder ablation. Relative L2L^2 error (%) at three target frames. “Identity” is the parameter-free configuration reported in the main text; the other rows make them learnable, with MM latent fields. Best in each block in bold.

References

  1. [1]Steven L. Brunton and J. Nathan Kutz. Data-Driven Science and Engineering: Machine Learning, Dynamical Systems, and Control. Cambridge University Press, 2019.DOI
  2. [2]Steven L. Brunton, Joshua L. Proctor, and J. Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the National Academy of Sciences, 113(15):3932–3937, 2016.DOI
  3. [3]Junfeng Chen, Kailiang Wu, and Dongbin Xiu. DUE: A deep learning framework and library for modeling unknown equations. SIAM Review, 67(4):873–902, 2025.arxiv.org/abs/2504.10373
  4. [4]Urban Fasel, J. Nathan Kutz, Bingni W. Brunton, and Steven L. Brunton. Ensemble-SINDy: Robust sparse model discovery in the low-data, high-noise limit, with active learning and control. Proceedings of the Royal Society A, 478(2260):20210904, 2022.DOI
  5. [5]Sebastian Gerard, Yu Zhao, and Josephine Sullivan. WildfireSpreadTS: A dataset of multi-modal time series for wildfire spread prediction. In Advances in Neural Information Processing Systems, volume 36, 2023.DOI
  6. [6]Maximilian Herde, Bogdan Raonić, Tobias Rohner, Roger Käppeli, Roberto Molinaro, Emmanuel de Bézenac, and Siddhartha Mishra. Poseidon: Efficient foundation models for PDEs. In Advances in Neural Information Processing Systems, volume 37, pages 72525–72624, 2024.arxiv.org/abs/2405.19101
  7. [7]Xianglong Hou, Xinquan Huang, and Paris Perdikaris. CFO: Learning continuous-time PDE dynamics via flow-matched neural operators. In International Conference on Learning Representations, 2026.arxiv.org/abs/2512.05297
  8. [8]Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizdazenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to PDEs. Journal of Machine Learning Research, 24(89):1–97, 2023.
  9. [9]Jingye Li, Alex Bespalov, and Jinglai Li. Nemytskii neural operator: A nonlinear model reduction method for parametrized partial differential equations. arXiv preprint, 2025. arXiv:2511.07684.arxiv.org/abs/2511.07684

Paper details

vlnc:W_4dv9vmg5
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

generative AI tools

View statement
Looking for related papers…
Contents