Operator Learning for Schr\"{o}dinger Equation: Unitarity, Error Bounds, and Time Generalization

Ambuj Tewari; Unique Subedi; Yash Patel

arxiv: 2505.18288 · v2 · submitted 2025-05-23 · 📊 stat.ML · cs.LG

Operator Learning for Schr\"{o}dinger Equation: Unitarity, Error Bounds, and Time Generalization

Yash Patel , Unique Subedi , Ambuj Tewari This is my paper

Pith reviewed 2026-05-19 13:00 UTC · model grok-4.3

classification 📊 stat.ML cs.LG

keywords operator learningSchrödinger equationunitarityerror boundstime generalizationlinear estimatorquantum dynamicsevolution operator

0 comments

The pith

A linear estimator for the Schrödinger evolution operator preserves weak unitarity while delivering uniform error bounds over smooth wave functions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a linear estimator for the evolution operator of the time-dependent Schrödinger equation that maintains a weak form of unitarity. It derives both upper and lower bounds on the estimator's prediction error, with the bounds holding uniformly across classes of sufficiently smooth initial wave functions. Time generalization bounds are also provided to measure how well the estimator performs at time points not seen in training. Experiments on Hamiltonians from hydrogen atoms, ion traps, and optical lattices show relative errors up to two orders of magnitude smaller than neural operator baselines.

Core claim

A linear estimator for the evolution operator is introduced that preserves a weak form of unitarity. Upper and lower bounds on its prediction error are established that hold uniformly over classes of sufficiently smooth initial wave functions. Time generalization bounds quantify extrapolation performance beyond the training times.

What carries the argument

Linear estimator for the evolution operator that preserves a weak form of unitarity.

If this is right

The estimator produces relative prediction errors up to two orders of magnitude smaller than the Fourier Neural Operator or DeepONet on real Hamiltonians including hydrogen atoms and ion traps.
Both upper and lower bounds on prediction error apply uniformly to sufficiently smooth initial wave functions.
Time generalization bounds quantify how prediction performance degrades or holds when extrapolating past the training time points.
The linear structure combined with weak unitarity preservation avoids the property violations common in neural surrogates.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the weak unitarity property survives fitting on noisy or discretized data, the method could support longer stable quantum simulations than non-unitary alternatives.
The uniformity over smoothness classes suggests the estimator may transfer across different physical systems whose wave functions share similar regularity.
Testing the bounds on wave functions near the boundary of the smoothness class would clarify how conservative the uniform guarantees are in practice.

Load-bearing premise

The uniform error bounds and time generalization results depend on initial wave functions belonging to the assumed smoothness classes and on the linear estimator remaining weakly unitary after fitting.

What would settle it

Finding a smooth initial wave function for which the observed prediction error exceeds the derived upper bound on a real-world Hamiltonian would falsify the uniform error claim.

Figures

Figures reproduced from arXiv: 2505.18288 by Ambuj Tewari, Unique Subedi, Yash Patel.

**Figure 1.** Figure 1: Squared amplitude |ψ(x)| 2 of the initial wave, the true wave at T = 0.1, and the estimator’s prediction for the barrier potential with double slits on [0, 2π) 2 . evolution, such as linearity and unitarity. In related fields, it has been demonstrated that incorporating known physical priors is often crucial for effective surrogate learning in data-scarce settings [Batzner et al., 2022, Merchant et al., … view at source ↗

read the original abstract

We consider the problem of learning the evolution operator for the time-dependent Schr\"{o}dinger equation, where the Hamiltonian may vary with time. Existing neural network-based surrogates often ignore fundamental properties of the Schr\"{o}dinger equation, such as linearity and unitarity, and lack theoretical guarantees on prediction error or time generalization. To address this, we introduce a linear estimator for the evolution operator that preserves a weak form of unitarity. We establish both upper bounds and lower bounds on the prediction error of the proposed estimator that hold uniformly over classes of sufficiently smooth initial wave functions. Additionally, we derive time generalization bounds that quantify how the estimator extrapolates beyond the time points seen during training. Experiments across real-world Hamiltonians -- including hydrogen atoms, ion traps for qubit design, and optical lattices -- show that our estimator achieves relative errors up to two orders of magnitude smaller than state-of-the-art methods such as the Fourier Neural Operator and DeepONet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The paper introduces a linear estimator for the evolution operator of the time-dependent Schrödinger equation that preserves a weak form of unitarity by construction. It establishes uniform upper and lower bounds on the prediction error over classes of sufficiently smooth initial wave functions, along with time generalization bounds quantifying extrapolation beyond training times. Experiments on Hamiltonians from hydrogen atoms, ion traps for qubit design, and optical lattices report relative errors up to two orders of magnitude smaller than baselines such as the Fourier Neural Operator and DeepONet.

Significance. If the derivations hold, the work is significant for providing theoretical guarantees on error and time generalization in operator learning while enforcing a physical constraint (weak unitarity) that neural surrogates often ignore. The uniform bounds over Sobolev-type classes and the empirical gains across real-world quantum systems could support more reliable long-horizon simulations in quantum physics and computing, distinguishing this approach from purely data-driven methods.

minor comments (3)

[§2.1] §2.1: The precise definition of the linear estimator (including how the weak unitarity constraint is imposed during fitting) should be stated explicitly with the relevant matrix or operator form to allow direct verification of the subsequent bounds.
[§3] Theorem 3.1 and Theorem 3.2: The constants appearing in the upper and lower error bounds should be compared explicitly to clarify whether the lower bound is informative or reduces to a trivial quantity under the same smoothness assumptions.
[Table 2] Table 2: Include standard deviations across multiple random initial conditions or runs to support the claim of consistent two-order-of-magnitude improvement.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of our work, the recognition of its potential significance for reliable long-horizon quantum simulations, and the recommendation of minor revision. We appreciate the emphasis placed on the theoretical guarantees and the empirical improvements over existing neural operators.

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper introduces a linear estimator explicitly constructed to preserve a weak form of unitarity for the time-dependent Schrödinger evolution operator, then derives uniform upper and lower error bounds over Sobolev-type classes of smooth initial data together with time-generalization bounds. These steps consist of mathematical analysis of the constructed estimator rather than any reduction of a claimed prediction back to fitted parameters, self-definitional loops, or load-bearing self-citations. The unitarity property is an enforced design feature, not a quantity that is both input and output of the same fit. No equations or sections in the provided material show a target result being recovered by construction from the estimator's own definition or from prior author work invoked as an external theorem. The smoothness-class assumptions are stated explicitly as the domain of the guarantees, rendering the overall argument independent and non-circular.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the domain assumption that the Schrödinger evolution operator is linear and that a weak form of unitarity can be preserved by a linear estimator; the error bounds further rest on the assumption that initial wave functions belong to sufficiently smooth function classes. No free parameters or invented entities are mentioned in the abstract.

axioms (2)

domain assumption The evolution operator of the time-dependent Schrödinger equation is linear and satisfies a weak form of unitarity that can be preserved by a suitably constructed linear estimator.
This property is invoked to motivate the estimator design and is the feature the bounds are proved to respect.
domain assumption Initial wave functions belong to classes of sufficiently smooth functions for which uniform error bounds can be derived.
The abstract states that the upper and lower bounds hold uniformly over these classes.

pith-pipeline@v0.9.0 · 5706 in / 1560 out tokens · 54254 ms · 2026-05-19T13:00:44.293001+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages · 2 internal anchors

[1]

Simulating decoherence of coupled two spin qubits using generalized cluster correlation expansion.arXiv preprint arXiv:2402.18722,

Xiao Chen, Silas Hoffman, James N Fry, and Hai-Ping Cheng. Simulating decoherence of coupled two spin qubits using generalized cluster correlation expansion.arXiv preprint arXiv:2402.18722,

work page arXiv
[2]

Growth of sobolev norms of solutions of linear schrödinger equations on some compact manifolds.International Mathematics Research Notices, 2010(12):2305–2328,

16 Jean-Marc Delort. Growth of sobolev norms of solutions of linear schrödinger equations on some compact manifolds.International Mathematics Research Notices, 2010(12):2305–2328,

work page 2010
[3]

Adam: A Method for Stochastic Optimization

Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

work page internal anchor Pith review Pith/arXiv arXiv
[4]

DeepONet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators

Lu Lu, Pengzhan Jin, and George Em Karniadakis. Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators. arXiv preprint arXiv:1910.03193,

work page internal anchor Pith review Pith/arXiv arXiv 1910
[5]

doi: 10.1007/b137836

ISBN 978-3-540-40710-5. doi: 10.1007/b137836. URLhttps://link.springer.com/book/10.1007/b137836. Amil Merchant, Simon Batzner, Samuel S Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk. Scaling deep learning for materials discovery.Nature, 624(7990):80–85,

work page doi:10.1007/b137836
[6]

arXiv preprint arXiv:2211.08875 , year=

Mattes Mollenhauer, Nicole Mücke, and TJ Sullivan. Learning linear operators: Infinite- dimensional regression as a well-behaved non-compact inverse problem.arXiv preprint arXiv:2211.08875,

work page arXiv
[7]

U-no: U-shaped neu- ral operators.arXiv preprint arXiv:2204.11127,

Md Ashiqur Rahman, Zachary E Ross, and Kamyar Azizzadenesheli. U-no: U-shaped neural operators.arXiv preprint arXiv:2204.11127,

work page arXiv
[8]

Fourier neural operators for learning dynamics in quantum spin systems.arXiv preprint arXiv:2409.03302,

Freya Shah, Taylor L Patti, Julius Berner, Bahareh Tolooshams, Jean Kossaifi, and Anima Anand- kumar. Fourier neural operators for learning dynamics in quantum spin systems.arXiv preprint arXiv:2409.03302,

work page arXiv
[9]

Physics-informed neural networks as solvers for the time-dependent schrodinger equation.arXiv preprint arXiv:2210.12522,

Karan Shah, Patrick Stiller, Nico Hoffmann, and Attila Cangi. Physics-informed neural networks as solvers for the time-dependent schrodinger equation.arXiv preprint arXiv:2210.12522,

work page arXiv
[10]

These methods typically parametrize the ground state wave functionψθusing a neural network and optimize the parameters by minimizing the energy functional⟨ψθ,Hψθ⟩L2

for an overview. These methods typically parametrize the ground state wave functionψθusing a neural network and optimize the parameters by minimizing the energy functional⟨ψθ,Hψθ⟩L2. This framework has also been extended to the time-dependent Schrödinger equation for many-electron systems by Nys et al. [2024]. This line of work is closely related to Physi...

work page 2024
[11]

A similar strategy was studied by Boullé et al

proposed an operator learning approach that models the solution operator mapping potentials to ground state wave func- tions by learning the associated Green’s functions in a reproducing kernel Hilbert space (RKHS). A similar strategy was studied by Boullé et al. [2022], who used rotational neural networks to learn Green’s functions for static Schrödinger...

work page 2022
[12]

also consid- ered learning the Green’s functions associated with time dependent propagator for1-dimensional Harmonic oscillator. A slightly more general framework was studied by Mizera [2023], who used Fourier Neural Operators (FNOs) [Li et al., 2021] to estimate the time evolution operator for simple quantum systems, such as random potentials and the dou...

work page 2023
[13]

B Extensions to Non-Periodic Domains Extending the results from Sections 3 and 5 to generalboundeddomainΩ⊂Rd is straightforward

trained 20 FNOs to learn the evolution operator for relatively larger quantum spin systems (up to 8-qubit systems), studying both single-step and multi-step time extrapolation. B Extensions to Non-Periodic Domains Extending the results from Sections 3 and 5 to generalboundeddomainΩ⊂Rd is straightforward. This requires choosing an orthonormal basis ofL2(Ω)...

work page 2014
[14]

Let(λj,ϕj)∞ j=1 be the eigenpairs of−∆inΩwith the given boundary conditions

and has since been implemented in works such as [Li et al., 2021, Kovachki et al., 2023]. Let(λj,ϕj)∞ j=1 be the eigenpairs of−∆inΩwith the given boundary conditions. By the Spectral Mapping Theorem, the eigenvalues of the covariance operator(−∆ +I)−βare(λj + 1)−β, while the eigenfunctions remainϕj’s. Applying the Karhunen-Loève Theorem [Hsing and Eubank,...

work page 2021
[15]

Applying this iteratively forjsteps, we obtain∥Fj(ψ)∥Hs =∥ψ∥Hs for allj∈N. H.2 Proof Part (ii) Proof.Our result follows directly from the bound in [Delort, 2010, Theorem 1], originally estab- lished by Bourgain [1999], which states that Fjψ  Hs =∥ψ(·,jT)∥Hs≤c(1 +jT)∥ψ∥Hs. This can be further refined using [Delort, 2010, Equation 1.3], yielding the b...

work page 2010
[16]

To see why, observe that we can rewrite ⟨Λs(∆ψ),Λsψ⟩=⟨(Λs∆Λ −s)Λsψ,Λsψ⟩

32 This follows because⟨Λs(∆ψ),Λsψ⟩is a real number. To see why, observe that we can rewrite ⟨Λs(∆ψ),Λsψ⟩=⟨(Λs∆Λ −s)Λsψ,Λsψ⟩. Since(Λ s∆Λ −s)is a self-adjoint operator onL 2, the inner product must be real. So, the only contribution comes from −i ℏΛs(Vψ). Thus, we obtain d dtEs(t) =−2 ℏ Im⟨Λs(Vψ),Λsψ⟩L2. Applying the Cauchy–Schwarz inequality, ⏐⏐⏐⏐ d dtEs...

work page 2021
[17]

shaking,

Coloumb PotentialFor a particle exposed to a radially symmetric electric field, such as in a Hydrogen atom, the potential is given byV(x) =−ke2 r2 . We specifically focus on the case of a fixed radius ofr= 1, for which the system can modeled as a uniform field in spherical coordinates. As discussed, both the pseudospectral solver and estimator were comput...

work page 2005
[18]

34 Table 5: Parameter values used in the implementation of each potential. Potential Name Parameter Values Free Particle — BarrierV 0 = 50.0,w= 0.2 Harmonic Oscillatorm= 1.0,ω= 2.0 Random Field (GRF)α= 1,β= 1,γ= 4 Paul TrapU 0 = 10.0,V 0 = 15.0,ω= 3.0,r0 = 2.0 Shaken LatticeV 0 = 4.0,k lat = 4π,A= 0.08,ωsh = 15.0 Gaussian PulseV= 100.0,x 0 = 0.0,y 0 = 0.0...

work page 2080

[1] [1]

Simulating decoherence of coupled two spin qubits using generalized cluster correlation expansion.arXiv preprint arXiv:2402.18722,

Xiao Chen, Silas Hoffman, James N Fry, and Hai-Ping Cheng. Simulating decoherence of coupled two spin qubits using generalized cluster correlation expansion.arXiv preprint arXiv:2402.18722,

work page arXiv

[2] [2]

Growth of sobolev norms of solutions of linear schrödinger equations on some compact manifolds.International Mathematics Research Notices, 2010(12):2305–2328,

16 Jean-Marc Delort. Growth of sobolev norms of solutions of linear schrödinger equations on some compact manifolds.International Mathematics Research Notices, 2010(12):2305–2328,

work page 2010

[3] [3]

Adam: A Method for Stochastic Optimization

Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

work page internal anchor Pith review Pith/arXiv arXiv

[4] [4]

DeepONet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators

Lu Lu, Pengzhan Jin, and George Em Karniadakis. Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators. arXiv preprint arXiv:1910.03193,

work page internal anchor Pith review Pith/arXiv arXiv 1910

[5] [5]

doi: 10.1007/b137836

ISBN 978-3-540-40710-5. doi: 10.1007/b137836. URLhttps://link.springer.com/book/10.1007/b137836. Amil Merchant, Simon Batzner, Samuel S Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk. Scaling deep learning for materials discovery.Nature, 624(7990):80–85,

work page doi:10.1007/b137836

[6] [6]

arXiv preprint arXiv:2211.08875 , year=

Mattes Mollenhauer, Nicole Mücke, and TJ Sullivan. Learning linear operators: Infinite- dimensional regression as a well-behaved non-compact inverse problem.arXiv preprint arXiv:2211.08875,

work page arXiv

[7] [7]

U-no: U-shaped neu- ral operators.arXiv preprint arXiv:2204.11127,

Md Ashiqur Rahman, Zachary E Ross, and Kamyar Azizzadenesheli. U-no: U-shaped neural operators.arXiv preprint arXiv:2204.11127,

work page arXiv

[8] [8]

Fourier neural operators for learning dynamics in quantum spin systems.arXiv preprint arXiv:2409.03302,

Freya Shah, Taylor L Patti, Julius Berner, Bahareh Tolooshams, Jean Kossaifi, and Anima Anand- kumar. Fourier neural operators for learning dynamics in quantum spin systems.arXiv preprint arXiv:2409.03302,

work page arXiv

[9] [9]

Physics-informed neural networks as solvers for the time-dependent schrodinger equation.arXiv preprint arXiv:2210.12522,

Karan Shah, Patrick Stiller, Nico Hoffmann, and Attila Cangi. Physics-informed neural networks as solvers for the time-dependent schrodinger equation.arXiv preprint arXiv:2210.12522,

work page arXiv

[10] [10]

These methods typically parametrize the ground state wave functionψθusing a neural network and optimize the parameters by minimizing the energy functional⟨ψθ,Hψθ⟩L2

for an overview. These methods typically parametrize the ground state wave functionψθusing a neural network and optimize the parameters by minimizing the energy functional⟨ψθ,Hψθ⟩L2. This framework has also been extended to the time-dependent Schrödinger equation for many-electron systems by Nys et al. [2024]. This line of work is closely related to Physi...

work page 2024

[11] [11]

A similar strategy was studied by Boullé et al

proposed an operator learning approach that models the solution operator mapping potentials to ground state wave func- tions by learning the associated Green’s functions in a reproducing kernel Hilbert space (RKHS). A similar strategy was studied by Boullé et al. [2022], who used rotational neural networks to learn Green’s functions for static Schrödinger...

work page 2022

[12] [12]

also consid- ered learning the Green’s functions associated with time dependent propagator for1-dimensional Harmonic oscillator. A slightly more general framework was studied by Mizera [2023], who used Fourier Neural Operators (FNOs) [Li et al., 2021] to estimate the time evolution operator for simple quantum systems, such as random potentials and the dou...

work page 2023

[13] [13]

B Extensions to Non-Periodic Domains Extending the results from Sections 3 and 5 to generalboundeddomainΩ⊂Rd is straightforward

trained 20 FNOs to learn the evolution operator for relatively larger quantum spin systems (up to 8-qubit systems), studying both single-step and multi-step time extrapolation. B Extensions to Non-Periodic Domains Extending the results from Sections 3 and 5 to generalboundeddomainΩ⊂Rd is straightforward. This requires choosing an orthonormal basis ofL2(Ω)...

work page 2014

[14] [14]

Let(λj,ϕj)∞ j=1 be the eigenpairs of−∆inΩwith the given boundary conditions

and has since been implemented in works such as [Li et al., 2021, Kovachki et al., 2023]. Let(λj,ϕj)∞ j=1 be the eigenpairs of−∆inΩwith the given boundary conditions. By the Spectral Mapping Theorem, the eigenvalues of the covariance operator(−∆ +I)−βare(λj + 1)−β, while the eigenfunctions remainϕj’s. Applying the Karhunen-Loève Theorem [Hsing and Eubank,...

work page 2021

[15] [15]

Applying this iteratively forjsteps, we obtain∥Fj(ψ)∥Hs =∥ψ∥Hs for allj∈N. H.2 Proof Part (ii) Proof.Our result follows directly from the bound in [Delort, 2010, Theorem 1], originally estab- lished by Bourgain [1999], which states that Fjψ  Hs =∥ψ(·,jT)∥Hs≤c(1 +jT)∥ψ∥Hs. This can be further refined using [Delort, 2010, Equation 1.3], yielding the b...

work page 2010

[16] [16]

To see why, observe that we can rewrite ⟨Λs(∆ψ),Λsψ⟩=⟨(Λs∆Λ −s)Λsψ,Λsψ⟩

32 This follows because⟨Λs(∆ψ),Λsψ⟩is a real number. To see why, observe that we can rewrite ⟨Λs(∆ψ),Λsψ⟩=⟨(Λs∆Λ −s)Λsψ,Λsψ⟩. Since(Λ s∆Λ −s)is a self-adjoint operator onL 2, the inner product must be real. So, the only contribution comes from −i ℏΛs(Vψ). Thus, we obtain d dtEs(t) =−2 ℏ Im⟨Λs(Vψ),Λsψ⟩L2. Applying the Cauchy–Schwarz inequality, ⏐⏐⏐⏐ d dtEs...

work page 2021

[17] [17]

shaking,

Coloumb PotentialFor a particle exposed to a radially symmetric electric field, such as in a Hydrogen atom, the potential is given byV(x) =−ke2 r2 . We specifically focus on the case of a fixed radius ofr= 1, for which the system can modeled as a uniform field in spherical coordinates. As discussed, both the pseudospectral solver and estimator were comput...

work page 2005

[18] [18]

34 Table 5: Parameter values used in the implementation of each potential. Potential Name Parameter Values Free Particle — BarrierV 0 = 50.0,w= 0.2 Harmonic Oscillatorm= 1.0,ω= 2.0 Random Field (GRF)α= 1,β= 1,γ= 4 Paul TrapU 0 = 10.0,V 0 = 15.0,ω= 3.0,r0 = 2.0 Shaken LatticeV 0 = 4.0,k lat = 4π,A= 0.08,ωsh = 15.0 Gaussian PulseV= 100.0,x 0 = 0.0,y 0 = 0.0...

work page 2080