PaperlayerView exact revision on arXiv

Source-linked AI summary

Slow--Fast Response Correction for a Compact Fourth-Order IMEX Two-Derivative Method

Zhixin Huo

arXiv:2609.27761v1math.NA

TL;DR

Compact fourth-order IMEX methods may lose high-order accuracy when relaxation and time-step scales are comparable, raising the question of whether uniform accuracy can be restored without changing the core. The paper introduces a slow–fast response correction and proves uniform fourth-order accuracy for the stated linear settings while preserving key properties of the base method.

  • Problem

    The paper asks whether uniform fourth-order accuracy can be recovered when relaxation time is comparable to the time step without redesigning the compact two-stage, two-solve core.

    Source: §1 · pp. 1–2

  • Method

    SFRC advances the exact slow projection with the unchanged compact IMEX step, projects it back to the slow subspace, and advances the fast complement by its exact semigroup.

    Source: §1 · p. 2

  • Results

    SFRC preserves the base mixed expansion through degree four and pure-fast L-stability while yielding a uniform O(dt^4) global bound for Jin–Xin modes; fixed-grid three-variable models show the same mechanism.

    Source: §1 · pp. 2–3

  • Takeaways & Limitations

    Within the stated linear spectral setting, uniform fourth-order accuracy is compatible with the unchanged compact core, its mixed expansion, and L-stability.

    Source: §9 · p. 26

    SFRC relies on exact slow spectral projectors and fast semigroup actions, limiting its practicality for nonlinear, large, or strongly space- and time-varying problems.

    Source: §8 · p. 24

Abstract

from arXiv · show

A compact two-stage fourth-order two-derivative IMEX method may retain its classical non-commuting fourth-order expansion and L-stable stiff damping while losing high-order accuracy when the relaxation time is comparable with the time step. We ask whether uniform fourth-order accuracy can be recovered without redesigning the original two-stage, two-implicit-solve core. For linear relaxation systems with a separated slow spectral subspace, we introduce a slow--fast response correction (SFRC): the exact slow projection is advanced by the unchanged compact IMEX step and projected back to the slow subspace, while the fast complement is advanced by its exact semigroup. For fixed nonzero relaxation time, the correction is $O(\dt^5)$, so the base mixed expansion through degree four is unchanged, and the pure-fast response remains L-stable. An abstract slow--fast argument reduces the full-state error to a slow-block defect and converts a uniform $O(\dt^5)$ local estimate into a uniform $O(\dt^4)$ global estimate. For each Jin--Xin Fourier mode, the required local estimate is proved by an exact projected-amplification factorization and a uniform bound in $x=\eps/\dt$. In particular, the unchanged compact core already has a uniform fifth-order one-step defect after restriction to the exact finite-$\eps$ slow eigenspace. A two-dimensional three-variable model yields the same mechanism on fixed Fourier grids. Symbolic verification and one- and two-dimensional experiments corroborate the analytical predictions. SFRC is therefore a rigorous linear benchmark identifying the finite-$\eps$ slow--fast information sufficient to reconcile compactness, mixed fourth-order accuracy, stiff damping, and uniform fourth-order time accuracy.

1 Introduction

The paper introduces SFRC, which combines the unchanged compact two-stage IMEX core on an exact slow subspace with exact fast evolution. For linear systems with a separated slow spectral subspace, it proves that this construction preserves fixed-ε fourth-order structure and yields uniform fourth-order accuracy, with Jin–Xin and two-dimensional results establishing the needed slow-block bounds.

Motivation and scope

  • Unlike schemes designed for uniform order, this work retains a fixed two-stage core; the strict-limit defect cannot be repaired by endpoint correction alone when ε/∆t is finite.The companion analysis reports an O(∆t^2) one-step defect for the unmodified strict-limit macro map and coefficient rigidity once the midpoint stage and two-solve topology are fixed.Source: §1 · p. 2

SFRC construction

  • SFRC advances the exact slow projection with the original IMEX formula and the fast complement with its exact semigroup, without changing stage coefficients or adding implicit solves.The method requires the exact slow projection and fast semigroup action, which are inexpensive for the small constant-coefficient Fourier blocks studied but not generally negligible.Source: §1 · p. 2

Preserved properties

  • For fixed ε > 0, the O(h^5) correction preserves the base method’s non-commuting mixed expansion through degree four, while the pure-fast response remains L-stable.On the pure-fast scalar limit SFRC reduces to exp(z), preserving L-stability but not the base rational stability function.Source: §1 · p. 2

Jin–Xin analysis

  • For each Jin–Xin Fourier mode, an exact slow-projected amplification factorization and a uniform bound for x = ε/h ≥ 0 yield an ε-independent O(h^5) local defect and uniform O(h^4) global error.The bound requires non-coalescing slow and fast eigenvalues; exact fast evolution makes the result applicable to arbitrary initial data.Source: §1 · p. 3

Two-dimensional extension

  • For the periodic two-dimensional three-variable model, a closed 2 × 2 relaxation block and bounded 3 × 3 slow projector establish modewise uniform fourth-order accuracy on fixed Fourier grids satisfying a non-coalescence condition.The reduction lets the same slow-amplification factorization apply after replacing the one-dimensional coefficients with their two-dimensional counterparts.Source: §1 · p. 3

Jin–Xin analysis

  • The unchanged compact core already has a uniform fifth-order one-step response after restriction and projection to the exact finite-ε slow eigenspace, locating the order-loss mechanism in the slow–fast mismatch.The obstruction is the mismatch between strict-equilibrium information and the finite-ε slow–fast decomposition, not a hidden low-order defect in the slow-projected core.Source: §1 · p. 3

2 The compact fourth-order base method and the properties to preserve

This section establishes the compact base method and identifies the properties a correction must preserve: its mixed expansion through order four for non-commuting operators and the implicit substructure’s L-stability. It also distinguishes fixed-operator accuracy from estimates uniform in a singular relaxation limit.

2 The compact fourth-order base method and the properties to preserve

  • The fixed-operator estimate does not assert uniformity when B = B0/eps becomes singular.Its constant is not claimed to remain bounded in that regime.Source: §2 · p. 4
  • The implicit substructure is L-stable, as established by the original analysis [1].The passage also notes that the original analysis proves A-stability.Source: §2 · p. 4
  • A correction preserves the base mixed expansion through order four when its operator coefficients agree with the base method’s for every fixed non-commuting pair A, B.Because the base expansion contains distinct AB and BA words, this operator-level criterion is stronger than checking a few commuting scalar order conditions; an O(h^5) perturbation leaves degree-four coefficients unchanged.Source: §2 · p. 5
  • The companion study’s asymptotic results motivate the analysis but are not assumptions; this paper derives its SFRC properties independently.The passage specifically identifies the uniform slow-block estimate, full-state convergence theorem, and strict SFRC limit as results proved here.Source: §2 · p. 5

3 Slow–fast response correction

With uniformly bounded complementary slow and fast spectral projectors, SFRC advances the slow component with the original method and the fast component with its exact semigroup, without adding an implicit stage or solve. It preserves the base method’s fixed-ε fourth-order expansion and retains pure-fast L-stability, though it does not preserve the original rational stability function.

3 Slow–fast response correction

  • SFRC advances the slow component with the original two-stage formula followed by slow projection, and advances the fast component with its exact semigroup, adding no implicit stage or solve.The method assumes complementary slow and fast spectral projectors with uniformly bounded norms; exact projection and fast semigroup operations are additional costs that depend on the problem.Source: §3 · p. 5
  • For fixed ε > 0, the correction is O(h^5), so SFRC preserves ordinary fourth order and the base method’s non-commuting mixed expansion through degree four.The evaluated mixed combinations containing distinct AB and BA contributions remain unchanged.Source: §3 · p. 6
  • The fixed-ε order result does not establish uniform accuracy, because its O(h^5) constant may blow up as ε → 0.The passage identifies closing this gap as the purpose of Sections 4–6.Source: §3 · p. 6
  • On a purely fast scalar block, SFRC is L-stable but replaces the base method’s rational stability function and algebraic stiff decay with exact exponential response.For ℜλ ≤ 0, the fast response is e^(hλ).Source: §3 · p. 6

4 Abstract full-state uniform fourth-order theorem

The abstract theorem converts a uniform fifth-order slow-block local defect into full-state uniform fourth-order accuracy, because SFRC advances the fast component exactly. Its guarantee depends on verifying that defect for the model; it does not establish it for every compact IMEX method or provide a nonlinear projector construction.

Assumption 4.1

  • The required slow one-step defect is O(h^5) with a constant independent of ε, for 0 < h ≤ h0 and 0 < ε ≤ ε0.This uniform local estimate is the model-specific assumption needed by the abstract theorem.Source: §4 · p. 6

Lemma 4.2

  • The corrected slow block is uniformly stable because iteration stays on the invariant slow range, yielding a stability factor e^{CT} rather than e^{CT/ε}.Restricting to the slow range avoids accumulating a factor involving the slow projector norm at every step.Source: §4 · p. 7

Theorem 4.3

  • Under the spectral-splitting assumptions and a uniform slow local defect, the corrected method achieves full-state fourth-order time accuracy uniformly in ε and h.The theorem applies for every fixed T and arbitrary initial data.Source: §4 · p. 7
  • Exact fast evolution cancels the fast numerical error, so arbitrary initial data need no slow-manifold preparation in this linear benchmark.Approximating the fast evolution instead would require a separate initial-layer error analysis.Source: §4 · pp. 7–8

Remark 4.4

  • The theorem is conditional: it neither guarantees the slow defect for every compact IMEX method nor supplies a nonlinear projector construction.The paper verifies the defect analytically for the linear Jin–Xin Fourier block in the next section.Source: §4 · p. 8

5 Rigorous verification for the linear Jin–Xin Fourier block

For the linear Jin–Xin Fourier block, SFRC achieves uniform fourth-order time accuracy across relaxation scales under a non-coalescence condition on resolved modes. The proof combines uniformly bounded spectral projectors with a uniform fifth-order defect for the slow-projected compact core.

5.1 Uniformly bounded spectral projector

  • Under the non-coalescence condition on the resolved parameter set, the slow eigenvalue and slow and fast projectors are uniformly bounded over the prescribed ε and wavenumber ranges.The condition excludes a slow–fast eigenvalue collision.Source: §5.1 · pp. 8–9, §5 · p. 8

5.2 Exact projected amplification factorization

  • An exact projected-amplification factorization is certified by a finite polynomial identity, with no floating-point fitting, asymptotic truncation, or numerically inferred cancellation.The two-stage solves reduce the projected response to a rational expression whose denominator is explicitly identified.Source: §5.2 · pp. 9–10

5.3 Uniform bound across the distinguished scale

  • The unchanged compact core has a uniform O(h5) one-step defect when restricted to the exact finite-ε slow eigenspace and projected back, not for arbitrary data.The estimate resolves ε≪h, ε=O(h), and ε≫h in one bound, despite ε/h becoming unbounded as h→0 at fixed ε.Source: §5.3 · pp. 11–12
  • SFRC is uniformly fourth-order accurate for arbitrary initial data on Jin–Xin modes with |κ|≤K, uniformly for 0<ε≤ε0, and the result extends to fixed finite Fourier grids.The constants may depend on the fixed wavenumber bound or grid; no uniform-increasing-cutoff PDE estimate or nonlinear relaxation result is claimed.Source: §5.3 · p. 12

5.4 Strict relaxation limit

  • At the strict relaxation endpoint, the limiting slow map is the fourth-order Taylor method for u_t=−cd u, so SFRC is AA4 without a separate endpoint switch.This endpoint conclusion follows from the same finite-ε slow response used in the uniform proof.Source: §5.4 · p. 13

6 Two-dimensional modewise extension

On a fixed Fourier grid satisfying an explicit non-coalescence condition, the two-dimensional modewise extension proves uniform fourth-order-in-time accuracy for SFRC across 0 < ε ≤ ε0.

6 Two-dimensional modewise extension

  • Under the checkable non-coalescence condition, the full slow projector and slow generator remain uniformly bounded as ε varies.A zero of ζ2 merges the coupled (u, p) eigenvalues and makes the rank-one slow projector ill-conditioned.Source: §6 · pp. 13, 15
  • On any fixed finite Fourier grid satisfying condition (71), SFRC is uniformly fourth-order accurate in time for 0 < ε ≤ ε0, with no uniformity in the Fourier cutoff asserted.This is a time-discretization result at fixed spatial resolution.Source: §6 · p. 14
  • Intertwining identifies the projected base amplification with that of a reduced 2 × 2 block; its purely imaginary parameter yields a uniform O(h^5) slow local defect.The reduced block permits the same x = ε/h estimate as in the Jin–Xin case, and the bounded full projector and slow generator then give the full-state estimate.Source: §6 · pp. 14–15

7 Numerical verification

The numerical experiments test distinct quantitative consequences of the analysis, rather than proving its theorems. Across modal and spatial tests, SFRC exhibits fourth-order accuracy over relaxation scans for prepared and unprepared data, while diagnostics probe the correction, implementation, and stiff response.

7 Numerical verification

  • The experiments are designed to test distinct consequences of the proofs, including algebraic and implementation checks, and report complete four-level refinement histories.For 1D modes, cancellation-resistant formulas and an independent exact-symbolic check of Lemma 5.3 are used; spatial examples use Fourier spectral differentiation to isolate temporal effects.Source: §7 · p. 15
  • On the active 1D and 2D modes, minimum sampled scaled separation factors are about 9.10 × 10^-1 and 4.48 × 10^-1, respectively, confirming the tests remain within the non-coalescence regime.The corresponding maximum projector norms are about 1.51 and 2.96; these diagnostics check the test cases’ spectral regime but do not replace analytical bounds.Source: §7 · p. 15, Table 2 · p. 15

7.1 Example 1: structural preservation, local defect, implementation cross-check, and stiff response

  • For one Jin–Xin mode, the SFRC scaled slow defect is about 1.71 × 10^-3 at ε = 10^-12 and 7.1 × 10^-2 at ε = 1, without growth across stiff and nonstiff regimes.At fixed ε = 0.2, the correction rates move toward five, consistent with the correction being invisible through degree four; the defect diagnostic uses the unchanged base step between exact slow projections.Source: §7.1 · p. 16, Table 3 · p. 16, Figure 1 · p. 16
  • The factorized implementation reproduces the literal SFRC matrix to near floating-point precision, while its exact-exponential fast response retains L-stability and damps more strongly for large negative z.At z = −1 the base and SFRC responses are comparable; at z = −10 and beyond, the exact exponential is many orders of magnitude smaller.Source: §7.1 · pp. 16–17, Figure 2 · p. 17, Table 6 · p. 17

7.2 Example 2: modal uniform-accuracy scan and distinguished scales

  • For equilibrium data, SFRC rates progress from 3.991 to 3.998 while base and EDC approach one; for unprepared data, SFRC stays near fourth order versus roughly second-order base and EDC decay.The full (h, ε) grid shows no diagonal ridge near ε ≈ h, so fourth-order behavior is not limited to avoiding that distinguished scale.Source: §7.2 · p. 19

7.3 Example 3: one-dimensional periodic Jin–Xin problem

  • On the 1D periodic Jin–Xin grid, SFRC’s final observed rate is 3.998 for both preparations; the base method’s equilibrium rate is about one, and unprepared data yield much larger errors.At ε = h = 0.04, the SFRC profile is visually indistinguishable from the exact field at plotting resolution, while the base pointwise error remains visible.Source: §7.3 · p. 21, Table 11 · p. 21, Table 12 · p. 21

7.4 Example 4: two-dimensional periodic relaxation system

  • In two dimensions, SFRC rates reach 3.986 for prepared and deliberately unprepared data (v=q=0); on the prepared scan, Ebase/ESFRC rises from about 2.6 × 10^3 at h=0.04 to 5.9 × 10^5 at h=0.005.The ratio compares errors, not computational efficiency; on the unprepared scan, the base maximizer stays at 10−4 and exhibits much larger unresolved-fast error.Source: §7.4 · p. 22

8 Discussion, limitations, and implications for method design

SFRC preserves the compact method’s fourth-order mixed expansion and L-stable pure-fast response while proving uniform fourth-order accuracy on fixed Fourier grids. The result is a rigorous linear benchmark, not a general-purpose computational correction.

Discussion, limitations, and implications for method design

  • The correction leaves the compact two-stage, two-implicit-solve core unchanged, is fifth order for fixed nonzero relaxation time, and yields uniform fourth-order accuracy on stated fixed Fourier grids.The pure-fast response is the exact exponential, and the strict relaxation limit is fourth order; uniformity is not claimed as the spatial cutoff tends to infinity.Source: §8 · pp. 23–24
  • SFRC is a rigorous benchmark rather than a broadly inexpensive correction: exact spectral evolution is practical for the studied Fourier blocks but can be costly or ill-defined for nonlinear, space-varying, or large unstructured problems.Its large base/SFRC error ratios are accuracy comparisons, not CPU-time or work–precision speedups. The fast error is identically zero, unlike preparation-dependent controls in other IMEX analyses; this stronger guarantee comes with narrower applicability. The unchanged core’s slow-projected response already has a uniform O(h5) local defect, so SFRC preserves that response and controls the fast complement rather than repairing a low-order slow-stage formula.Source: §8 · p. 24
  • Practical variants could approximate the slow projector and fast semigroup, but must preserve uniform O(h5) slow defects, avoid O(1) initial-layer error, and leave the fixed-ε mixed expansion unchanged.Suggested replacements include Chapman–Enskog or micro–macro projectors and source-exponential, rational, or Krylov fast evolution.Source: §8 · pp. 24–25
  • A nonlinear extension is not established; it would require a smooth slow manifold and uniformly accurate slow and fast steps, while practical approximations must still be shown to satisfy the needed bounds.The sufficient-condition program requires a uniformly Lipschitz local retraction, an O(h5) slow defect uniform in ε/h, stable fourth-order fast evolution through the initial layer, and an O(h5) fixed-ε correction; a discrete Gronwall argument could then extend the global-error mechanism.Source: §8 · p. 25
  • The earlier equilibrium-defect correction remains a minimal strict-limit repair, whereas SFRC controls both slow response and the fast spectral component across the intermediate scale.The earlier correction needs only equilibrium information and no spectral decomposition; SFRC provides one rigorous linear approach to uniform correction.Source: §8 · p. 25
  • The exact factorization explains how apparent singular coefficients cancel on the finite-ε slow eigenspace, enabling high order despite the unprojected base scheme’s first-order limiting-equilibrium map.The factor cd + λ_s = O(ε) compensates implicit denominators for small ε/h, while x^4 denominator growth controls the fixed-ε regime.Source: §8 · p. 25

9 Conclusions

The paper constructs a slow–fast response correction that preserves the compact IMEX core’s fourth-order mixed expansion and L-stable fast damping while proving uniform fourth-order time accuracy in the stated linear setting. Its central result is a compatibility theorem supported by a uniform slow-defect proof; nonlinear uniform accuracy and practical approximate projectors or fast solvers remain open.

Conclusions

  • The correction applies the unchanged compact IMEX core to the exact slow component and advances the fast complement exactly, adding no implicit solve while preserving the mixed expansion and L-stable damping.For fixed nonzero relaxation time, the correction is a fifth-order perturbation; the pure-fast response is ez.Source: §9 · p. 25
  • An abstract slow–fast decomposition converts a uniform fifth-order local defect into a fourth-order global estimate, verified for Jin–Xin modes through an exact factorization and bounds uniform in x = ε/h.The proof bounds the implicit denominators for Fourier modes and shows the remaining coefficient stays bounded for all x ≥ 0; the two-dimensional three-variable model extends the result modewise on fixed Fourier grids.Source: §9 · p. 25
  • The compatibility result shows the order loss is not intrinsic to the two-stage, two-solve core: on the exact finite-ε slow eigenspace, that core already has a uniform fifth-order one-step defect.The rigorous result is limited to the stated linear spectral setting and fixed Fourier resolutions; experiments corroborate, but do not prove, its quantitative consequences.Source: §9 · p. 26
  • A general nonlinear uniform fourth-order theorem remains open, as do practical replacements for the exact projector and fast semigroup that preserve the uniform slow defect and accurately evolve the fast layer.Potential approximations include Chapman–Enskog, micro–macro, source-exponential, rational, or Krylov methods, but they require additional uniform local and initial-layer estimates.Source: §9 · p. 26

Data and code availability · A Additional algebra behind the factorization

The appendix makes the projected-amplification factorization independently checkable by deriving its scalar ingredients and showing that the defect numerator has no terms below degree five. The supplied passages address the algebra appendix and provide no data or code availability details.

A Additional algebra behind the factorization

  • Intermediate slow-eigenvector identities show that the source action is finite and cancel the singular 1/ε factor.These steps use the inverse of I + αB and the relation Ars = λrs − Brs to reduce the midpoint equation.Source: A Additional algebra behind the… · p. 26
  • In the basis V = (rs, Brs), direct application of A and B gives the representation that proves (40).The appendix uses this basis to reduce the slow-eigenvector calculation to explicit two-dimensional algebra.Source: A Additional algebra behind the… · p. 26
  • In the same basis, the two implicit matrices are explicit and their determinants equal the quantities stated in (42).These determinants contribute to the scalar rational function used in the factorization.Source: A Additional algebra behind the… · p. 27
  • The derivation makes every ingredient of q = ωT bSe1 explicit before polynomial expansion.The scalar rational function follows from the displayed basis identities and determinants.Source: A Additional algebra behind the… · pp. 26–27
  • Reducing every occurrence of a2d2 through (39) collects the remaining denominators into the scalar factors D8 and D6.The scalar left projection also contributes a factor to the factorization.Source: A Additional algebra behind the… · p. 27
  • The spectral non-coalescence condition is exactly the condition that biorthogonal normalization remains nondegenerate after removing the natural 1/ε scaling.The identity 1 + 2ελ = ζ gives this equivalence.Source: A Additional algebra behind the… · p. 27
  • The defect numerator is exactly -h^5λ^3(cd+λ)N, so all terms below total h-degree five cancel and the certificate vanishes identically, proving (35).Clearing 24(1 + 2ελ)D8D6 from q − T4(hλ) yields the certificate; the complete polynomial N is given, with no omitted higher-order term.Source: A Additional algebra behind the… · p. 27

B Why the global proof has no stiff exponential constant

The global error proof avoids a stiff exponential constant because the exact fast response cancels from the error, leaving only the uniformly bounded slow block to propagate. This yields an accumulated error bound with a constant independent of ε.

Why the global proof has no stiff exponential constant

C Implementation remarks

The implementation avoids direct eigendecomposition at tiny ε, instead computing the slow projector and projected scalar response with closed-form expressions.

C Implementation remarks

  • At tiny ε, the one-dimensional code uses a scaled projector formula, a cancellation-free root, and an exact factorization for the projected scalar response.Direct double-precision eigendecomposition of matrices with entries of size 1/ε can let roundoff obscure the uniform limit, despite the projector being mathematically bounded.Source: C Implementation remarks · p. 27
Loading 2609.27761v1…