Explicit Low-Regularity Exponential Integrators

• Explicit low-regularity exponential integrators are time discretizations that integrate the linear propagator exactly while approximating nonlinearities via Duhamel-based expansions to reduce derivative loss.
• They leverage innovative mechanisms such as twisted variables, exact oscillatory phase treatment, and commutator organization to achieve first-order convergence under minimal regularity assumptions.
• These integrators apply to a range of evolution equations—including NLS, DNLS, Klein–Gordon, and Dirac equations—preserving structure and improving computational efficiency.

Explicit low-regularity exponential integrators are time discretizations for evolution equations that combine exact treatment of the linear propagator with Duhamel-based approximation of the nonlinear contribution, while proving convergence under solution regularity close to the PDE well-posedness threshold rather than the higher Sobolev assumptions typically required by classical splitting or classical exponential integrators. In this literature, “low regularity” is not a single universal threshold: for polynomial nonlinear Schrödinger equations it means first-order convergence in $H^r$ for solutions in $H^{r+1}$ with $r>d/2$, for one-dimensional quadratic Schrödinger equations it can mean first-order convergence in $H^r$ with solutions in $H^r$, for the derivative nonlinear Schrödinger equation it means first-order convergence in $H^s$ for $s>\tfrac12$ with initial data in $H^{s+1}$, and for strongly continuous semigroups it means high-order contour approximations with error controlled by graph norms $\|(2\delta-A)^m x\|$ for data in $D(A^m)$ (Ostermann et al., 2016, Ji et al., 28 Jan 2026, Horning et al., 2024).

1. Conceptual scope and historical placement

The central problem addressed by explicit low-regularity exponential integrators is that standard Taylor-based, splitting-based, or classical exponential Runge–Kutta discretizations often incur derivative loss in local error estimates. For semilinear Schrödinger equations, classical splitting and standard exponential integrators typically require two additional spatial derivatives for first-order convergence, whereas the low-regularity exponential-type integrators introduced by Ostermann and Schratz prove first-order convergence in $H^{r+1}$0 for solutions in $H^{r+1}$1, and for one-dimensional quadratic Schrödinger equations even without any loss of regularity (Ostermann et al., 2016). For the derivative nonlinear Schrödinger equation, the difficulty is stronger because the equation is genuinely quasi-linear in the sense that the nonlinearity contains a derivative; there the new contribution is a first-order unfiltered exponential integrator converging in $H^{r+1}$2 for any $H^{r+1}$3 with $H^{r+1}$4, together with a symmetrized version that performs better in terms of both global error and conservation behavior (Ji et al., 28 Jan 2026).

The term “explicit” also varies with context. In the dispersive PDE literature it denotes one-step schemes in which $H^{r+1}$5 is obtained by applying fixed operators to already known quantities, with no equation to solve for $H^{r+1}$6; typical nonlocal operators such as $H^{r+1}$7, $H^{r+1}$8, and $H^{r+1}$9 are diagonal in Fourier space and are efficiently computable via FFTs (Ji et al., 28 Jan 2026). In semigroup-based contour methods for strongly continuous semigroups, “explicit” is used in the exponential-integrator sense at the linear level: resolvent samples are precomputed, and propagation on a finite time horizon is then evaluated by scalar-weighted linear combinations, with no implicit solves involving $r>d/2$0 and no nonlinear solves (Horning et al., 2024).

A broader generalization is provided by the decorated-tree framework, which replaces Taylor expansions in powers of $r>d/2$1 by commutator-based expansions and thereby constructs low-regularity integrators up to arbitrary high order on general domains. In that setting, the local error is driven by nested commutators $r>d/2$2 rather than powers $r>d/2$3, and this generally requires much lower regularity assumptions than classical methods (Bronsard et al., 2022). This suggests that explicit low-regularity exponential integrators are best understood not as a single formula, but as a design paradigm centered on exact linear propagation, oscillatory integral analysis, and commutator or resonance structure.

2. Core analytical mechanisms

A recurring mechanism is the introduction of a twisted variable. For semilinear Schrödinger equations, one sets $r>d/2$4, so that the variation over one time step is measured in the twisted frame rather than in the original variable. The basic estimate is then that $r>d/2$5 is small in $r>d/2$6 without requiring extra spatial derivatives of $r>d/2$7, in contrast to a direct approximation of $r>d/2$8 by $r>d/2$9, which reintroduces terms of the form $H^r$0 and therefore costs derivatives (Ostermann et al., 2016). The same principle reappears in the time-relaxation schemes for NLS, where the relaxation parameter is introduced in the twisted variable $H^r$1, because relaxing directly in the physical variable causes order reduction under low regularity (Li et al., 3 Oct 2025).

A second mechanism is exact or near-exact treatment of the dominant oscillatory phase. In the polynomial NLS setting, the numerical approximation of the Duhamel term is designed so that the component associated with $H^r$2 is integrated exactly through $H^r$3, yielding the scheme

$H^r$4

for $H^r$5 (Ostermann et al., 2016). For the derivative nonlinear Schrödinger equation, the dominant derivative interaction is isolated after a gauge transformation and twisting by the linear Schrödinger propagator, and the delicate term $H^r$6 is approximated through the explicitly computable operator $H^r$7, which is linear in the first argument, quadratic in the second, and designed to capture the dominant derivative contribution (Ji et al., 28 Jan 2026).

A third mechanism is integration by parts in time, or normal-form reduction in discrete form. In the DNLS analysis, an “integration by parts” lemma in Fourier space expresses an integral with $H^r$8 inside in terms of boundary terms and time derivatives that are one order more regular, thereby reducing derivative loss (Ji et al., 28 Jan 2026). In the singularly perturbed linear dispersive problem with concentrated potential, iterated Duhamel expansions and a transformation that exploits cancellations in oscillatory phases are essential because error bounds based only on regularity estimates of the exact solution are too crude to obtain the optimal $H^r$9-dependence; the optimal bound is $H^r$0, up to logarithmic factors, with no restriction on $H^r$1 in terms of $H^r$2 (Bal et al., 25 Feb 2026). This suggests that low-regularity analysis is inseparable from phase analysis: the regularity threshold alone does not explain the performance of the method.

A fourth mechanism is commutator organization. In the decorated-tree framework, the local error is expressed through nested commutators

$H^r$3

and the trees encode how Duhamel iterations, dominant operators, and lower-order operators combine. The resulting approximations preserve the exponential structure of the linear flow while replacing high powers of $H^r$4 by commutator expressions that require fewer derivatives and weaker boundary compatibility (Bronsard et al., 2022).

3. Representative constructions across equations

For semilinear Schrödinger equations with polynomial nonlinearity, the prototype scheme is the low-regularity exponential-type integrator of Ostermann and Schratz. For general $H^r$5, it is explicit, uses the twisted variable, and proves first-order convergence in $H^r$6 for solutions in $H^r$7 with $H^r$8. For one-dimensional quadratic Schrödinger equations, the paper derives exact Fourier identities involving $H^r$9, so that first-order convergence in $H^s$0 is achieved assuming only $H^s$1, $H^s$2, with no extra derivative (Ostermann et al., 2016). The later parareal work uses precisely such exponential low-regularity integrators as coarse propagators and proves linearly convergent parareal algorithms for quadratic and cubic NLS, with contraction factor proportional to the coarse time-step size even for solutions of limited regularity (Lin et al., 1 Jul 2026).

For the derivative nonlinear Schrödinger equation on $H^s$3,

$H^s$4

the decisive step is a gauge transformation $H^s$5 that turns the original DNLS into a transformed equation for $H^s$6 with

$H^s$7

where the derivative now hits $H^s$8 only. The first-order exponential-type integrator is then designed for the gauged equation, not for DNLS directly, and the principal scheme is fully explicit, unfiltered, and naturally implemented by a Fourier pseudo-spectral method (Ji et al., 28 Jan 2026). This paper also constructs a symmetric two-step version

$H^s$9

with $s>\tfrac12$0, and numerically observes significantly smaller mass and energy drift over long times (Ji et al., 28 Jan 2026).

For the nonlinear Klein–Gordon equation in the classical and non-relativistic regimes, the uniformly accurate schemes are again explicit exponential-type integrators, but now the goal is simultaneously low regularity and robustness with respect to the parameter $s>\tfrac12$1. The first-order scheme converges with order $s>\tfrac12$2 under $s>\tfrac12$3, the second-order scheme with order $s>\tfrac12$4 under $s>\tfrac12$5, with constants independent of $s>\tfrac12$6, and in the limit $s>\tfrac12$7 the schemes converge to low-regularity integrators for the NLS limit system (Calvo et al., 2021). An earlier uniformly accurate Klein–Gordon construction follows the same general philosophy—exact treatment of the oscillatory linear part, no step-size restriction, and asymptotic convergence to Lie or Strang splitting in the nonlinear Schrödinger limit—but is formulated directly in terms of exponential-type integrators for the twisted variables $s>\tfrac12$8 (Baumstark et al., 2016).

For nonlinear Dirac equations and the Dirac–Poisson system, the ultra low-regularity integrator uses $s>\tfrac12$9 rather than the full free Dirac propagator as the basic exponential operator. Thanks to the projector decomposition

$H^{s+1}$0

the Duhamel terms can be integrated explicitly and produce only bounded Fourier multipliers $H^{s+1}$1. The resulting ULI is explicit, exponential-type, and proves first-order convergence in $H^{s+1}$2 for solutions in $H^{s+1}$3, $H^{s+1}$4, i.e. without any additional regularity on the solution; the second-order extension is also explicit and converges in $H^{s+1}$5 under $H^{s+1}$6 and $H^{s+1}$7 (Schratz et al., 2019).

For strongly continuous semigroups, the contour-integral methods are not time-stepping methods in the usual sense, but they provide a high-order explicit propagation mechanism for $H^{s+1}$8 under only $H^{s+1}$9-semigroup assumptions. With the rational regularizer

$\|(2\delta-A)^m x\|$0

the paper derives the quadrature

$\|(2\delta-A)^m x\|$1

proves error bounds with truncation error $\|(2\delta-A)^m x\|$2 for $\|(2\delta-A)^m x\|$3, and emphasizes that the resulting scheme is global in time and embarrassingly parallelizable (Horning et al., 2024). A plausible implication is that this semigroup-level construction broadens the meaning of “exponential low-regularity integrator” beyond dispersive Fourier settings.

4. Error theory and regularity thresholds

The low-regularity error theory is equation-specific but structurally similar. For polynomial NLS, the main theorem states that if $\|(2\delta-A)^m x\|$4, $\|(2\delta-A)^m x\|$5, $\|(2\delta-A)^m x\|$6, then the exponential-type integrator converges in $\|(2\delta-A)^m x\|$7 with order $\|(2\delta-A)^m x\|$8, and in particular with first order when $\|(2\delta-A)^m x\|$9 (Ostermann et al., 2016). For one-dimensional quadratic Schrödinger equations, the same framework yields first-order convergence in $D(A^m)$0 assuming only $D(A^m)$1, $D(A^m)$2, because the central oscillatory integrals are computed exactly in terms of $D(A^m)$3 and exponentials (Ostermann et al., 2016).

For DNLS, the main convergence theorem states that for $D(A^m)$4 and $D(A^m)$5, the numerical solution $D(A^m)$6 of the transformed scheme satisfies

$D(A^m)$7

with constants independent of $D(A^m)$8 and $D(A^m)$9. The proof combines a local error bound

$H^{r+1}$00

with a quasi-isometric stability estimate for the operator

$H^{r+1}$01

and then a discrete Gronwall argument (Ji et al., 28 Jan 2026). The threshold $H^{r+1}$02 is significant because $H^{r+1}$03 is an algebra there, and because periodic DNLS well-posedness is naturally built around $H^{r+1}$04 (Ji et al., 28 Jan 2026).

For uniformly accurate Klein–Gordon integrators, the relevant low-regularity thresholds are $H^{r+1}$05 for first order and $H^{r+1}$06 for second order, with $H^{r+1}$07. The commutator structure is essential: the local error involves exactly one derivative for order 1 and two derivatives for order 2, uniformly in the parameter $H^{r+1}$08, instead of high powers of $H^{r+1}$09 (Calvo et al., 2021). The earlier Klein–Gordon paper proves uniform first- and second-order convergence under regularity assumptions that are not more restrictive than those required for integrating the NLS limit system, and explicitly avoids any asymptotic or multiscale expansion of the solution (Baumstark et al., 2016).

For nonlinear Dirac equations, the first-order ULI is genuinely ultra low-regularity: it converges in $H^{r+1}$10 for solutions in $H^{r+1}$11, $H^{r+1}$12, and the second-order extension converges in $H^{r+1}$13 for solutions in $H^{r+1}$14 with $H^{r+1}$15. The paper also proves a second-order $H^{r+1}$16 bound under $H^{r+1}$17 and $H^{r+1}$18, whereas classical second-order schemes require $H^{r+1}$19-type regularity for the same $H^{r+1}$20 order (Schratz et al., 2019).

For the time-relaxation mass-preserving schemes, the general theorem proves that the relaxed integrator preserves the order $H^{r+1}$21 of the base low-regularity integrator: $H^{r+1}$22 provided the base scheme satisfies stability and local truncation estimates and the relaxation parameter satisfies $H^{r+1}$23 together with a consistency estimate comparing $H^{r+1}$24 and $H^{r+1}$25 (Li et al., 3 Oct 2025). By contrast, relaxing directly in the physical variable $H^{r+1}$26 is proved to reduce the order from two to one for the cubic NLS example (Li et al., 3 Oct 2025).

5. Explicitness, implementation, and structure preservation

The computational architecture of these methods is typically Fourier pseudo-spectral. In the DNLS work, implementation is naturally done by a Fourier pseudo-spectral method: represent $H^{r+1}$27 on a grid, compute FFTs, apply diagonal multipliers for $H^{r+1}$28, $H^{r+1}$29, and $H^{r+1}$30, and evaluate nonlinear products in physical space (Ji et al., 28 Jan 2026). The polynomial NLS, Klein–Gordon, Dirac, and time-relaxation papers follow the same broad pattern: linear exponentials and $H^{r+1}$31-functions are diagonal in Fourier space, while nonlinearities are evaluated pointwise in physical space (Ostermann et al., 2016, Calvo et al., 2021, Schratz et al., 2019, Li et al., 3 Oct 2025). In this sense, the adjective “explicit” refers simultaneously to the absence of implicit nonlinear solves and to the diagonalizability of the relevant linear operators.

Structure preservation is an active axis of development. The DNLS symmetrized method is time-reversible and numerically shows significantly better long-time conservation of mass and energy than the basic first-order method, although under low-regularity assumptions only first-order convergence is proved (Ji et al., 28 Jan 2026). The time-relaxation NLS framework goes further and enforces exact discrete mass conservation by choosing a scalar parameter $H^{r+1}$32 so that

$H^{r+1}$33

and therefore $H^{r+1}$34 for all $H^{r+1}$35 (Li et al., 3 Oct 2025). The resulting RLRI-v schemes are fully explicit, mass-conserving, and numerically display long-time $H^{r+1}$36-error at machine precision over very long times (Li et al., 3 Oct 2025).

In stochastic Maxwell equations, the explicit exponential integrator

$H^{r+1}$37

inherits the unitary group structure of the Maxwell operator. For additive noise, the method has strong order $H^{r+1}$38; for multiplicative noise, strong order $H^{r+1}$39; and in the linear additive-noise case it preserves exactly the symplectic structure, the evolution of the energy, and the evolution of the divergence in the sense of expectation (Cohen et al., 2019). Although this is an SPDE setting rather than a dispersive low-regularity PDE in the deterministic sense, it illustrates the same principle: exact treatment of the linear group can deliver stability, structure preservation, and favorable regularity requirements simultaneously.

Large-scale implementation of exponential actions is itself a research topic. For explicit exponential Runge–Kutta integrators on stiff semilinear problems, adaptive rational Krylov methods reduce linear combinations of $H^{r+1}$40-functions to a small number of matrix exponentials on enlarged systems, derive a posteriori error estimates for rational Krylov approximations at single time points, and numerically show approximately constant numbers of rational Krylov iterations as stiffness grows, enabling near-linear runtime scaling with problem size (Bergermann et al., 2023). This is not a low-regularity convergence theory by itself, but it functions as enabling technology for explicit exponential methods whose analytical design already incorporates low-regularity structure.

6. Comparisons, extensions, and open directions

A persistent comparison class is classical Lie and Strang splitting and classical exponential integrators. For polynomial NLS, the low-regularity exponential-type integrator outperforms Lie splitting, Strang splitting, and the classical first-order exponential integrator under rough data, while the classical schemes recover their nominal orders only for smoother solutions (Ostermann et al., 2016). For quadratic and cubic NLS inside parareal, coarse propagators based on Lie, Strang, and first- or third-order exponential Runge–Kutta schemes do not enjoy the same iteration-wise contraction behavior as coarse propagators based on exponential low-regularity integrators (Lin et al., 1 Jul 2026). For singularly perturbed linear dispersive equations with concentrated potential, Lie and centered splitting schemes and low-regularity integrators fail to display the optimal $H^{r+1}$41 error rate achieved by the tailored exponential integrator (Bal et al., 25 Feb 2026).

The class of equations covered is expanding. Besides NLS and DNLS, the data include low-regularity or ultra low-regularity integrators for KdV and mKdV in the cited background of several papers, nonlinear Dirac equations (Schratz et al., 2019), stochastic Maxwell equations (Cohen et al., 2019), Klein–Gordon equations across oscillatory regimes (Calvo et al., 2021, Baumstark et al., 2016), and abstract semilinear problems generated by strongly continuous contraction semigroups (Li et al., 3 Oct 2025). The contour-integral work extends semigroup propagation beyond analytic semigroups to general $H^{r+1}$42-semigroups, with high-order error bounds governed by domain regularity $H^{r+1}$43 rather than smoothing (Horning et al., 2024). The decorated-tree framework goes further and explicitly proposes arbitrary-order low-regularity integrators for parabolic, hyperbolic, and dispersive equations on general domains (Bronsard et al., 2022).

Several limitations remain stable across the literature. Many rigorous results are one-dimensional and periodic. The DNLS analysis is restricted to one space dimension and periodic boundary conditions (Ji et al., 28 Jan 2026). The parareal convergence theory is proved only in $H^{r+1}$44 for quadratic and cubic NLS, while two-dimensional cubic and one-dimensional quintic NLS are treated numerically only (Lin et al., 1 Jul 2026). The symmetric DNLS method is formally second order in the classical smooth setting, but only first-order convergence is proved under low regularity (Ji et al., 28 Jan 2026). The mass-preserving relaxation framework preserves mass exactly but not energy (Li et al., 3 Oct 2025). A plausible implication is that explicit structure-preserving low-regularity methods are presently strongest for quadratic invariants and for equations whose dominant oscillations admit exact or nearly exact phase analysis.

Open directions are stated explicitly in the sources. They include pushing regularity thresholds to critical or supercritical regimes for DNLS (Ji et al., 28 Jan 2026), extending low-regularity parareal theory to higher dimensions and more general nonlinearities (Lin et al., 1 Jul 2026), constructing higher-order explicit relaxed low-regularity schemes (Li et al., 3 Oct 2025), adapting rational-Krylov exponential-action engines to more general operators and to Rosenbrock or EPIRK settings (Bergermann et al., 2023), and extending contour-integral high-order semigroup approximations to nonlinear exponential integrators in the $H^{r+1}$45-semigroup setting (Horning et al., 2024). Across these directions, the unifying principle remains unchanged: embed the central oscillations of the PDE into the discretization, express local defects through resonance structure or nested commutators rather than raw derivatives, and preserve explicitness by keeping the linear flow exact.