PaperlayerView exact revision on arXiv

Source-linked AI summary

Layer-wise discrete certificates in multicomponent reacting-flow discretizations

Jingchao Zhang

arXiv:2609.27272v1math.NAcs.NA

TL;DR

The paper asks whether conservation, positivity, and free-energy dissipation remain compatible across the different layers of reacting-flow discretizations. It analyzes face diffusion, non-orthogonal reconstruction, and Patankar time integration separately. Tested examples show that guarantees at one layer need not transfer to the next.

  • Problem

    Conservation, positive mass fractions, and dissipation of a convex mixing potential are controlled at different solver stages, and their compatibility when those components are assembled is unclear.

    Source: §1 · pp. 2–3

  • Method

    The paper analyzes mixture-averaged diffusion, weighted least-squares non-orthogonal reconstruction, and positivity-preserving Patankar integration as separate layers, then examines their fully discrete interaction.

    Source: §2.2 · p. 4, §5.1 · p. 12, §6.1 · p. 15

  • Results

    40.0–56.7% of tested detailed-balance networks show positive free-energy change for all six examined third-order members; the tested WLS correction changes entropy rate from −0.4083 to +0.1436.

    Source: §9 · p. 22

  • Takeaways & Limitations

    Conservation, positivity, and free-energy decay require distinct discrete conditions, and a face-level certificate alone does not certify the reconstructed or fully discrete update.

    Source: §9 · p. 22

    The WLS result concerns the tested multipoint construction and the Patankar result covers the named members; a general invariant-domain condition for coupled diffusion–reaction remains open.

    Source: §8.2 · p. 22

Abstract

from arXiv · show

Multicomponent reacting-flow discretizations must conserve total species mass, preserve positive mass fractions, and dissipate a convex mixing potential. These properties are controlled at different stages of a solver, however, and a certificate established at one stage need not survive the next. We study this compatibility problem for mixture-averaged diffusion, non-orthogonal finite-volume reconstruction, and positivity-preserving Patankar time integration. The analysis is isothermal, uses prescribed density, and uses synthetic detailed-balance reaction networks for the chemistry tests. At the face level, the entropy-variable jump is shown to equal a log-mean Gibbs Hessian applied exactly to the mass-fraction jump. Requiring conservation and entropy neutrality of the mass-correction term on the simplex tangent space then yields a unique componentwise-positive face composition and an exact two-point entropy certificate. The arithmetic-weight face criterion is also satisfied by production neutral-air/carbon data and a quasi-neutral ambipolar-air control, but a face-level certificate does not automatically extend to subsequent discretization layers. A weighted least-squares non-orthogonal correction can, in the tested prototype, convert a dissipative two-point operator into an anti-dissipative multipoint operator. Likewise, Patankar rescaling can preserve positivity and conservation while increasing the free energy; an explicit finite-step rational counterexample establishes this failure independently of ensemble statistics. A fully discrete two-cell example further shows that a positive face entropy margin does not guarantee positivity of the fully discrete update. Thus conservation, positivity, and free-energy dissipation are layer-specific properties and require distinct discrete certificates.

1. Introduction

The introduction asks whether conservation, positivity, and thermodynamic dissipation remain compatible when independently designed spatial and temporal components are assembled. It frames the paper’s contribution as a layer-wise analysis of mixture-averaged diffusion, non-orthogonal reconstruction, and Patankar integration.

1. Introduction

  • A guarantee established at one solver stage may not survive the next, even when conservation, positivity, and dissipation are each enforced separately.The introduction links this compatibility question to possible composition drift, nonphysical species states, and artificial free-energy growth.Source: §1 · pp. 2–3
  • Mixture-averaged diffusion, non-orthogonal reconstruction, and Patankar integration pose distinct compatibility questions for a solver assembled from independently designed components.The mass-correction term affects diffusion entropy behavior, tangential reconstruction widens the stencil, and positivity rescaling does not itself constrain free energy.Source: §1 · pp. 2–3
  • The work complements studies of order, stability, and fixed points by evaluating a specified convex detailed-balance free energy, constructing a rational finite-step counterexample, and examining six representative third-order members.These are the paper’s stated analytical and numerical emphases.Source: §1 · p. 3
  • The analysis separates the face, reconstruction, and temporal layers, then examines which guarantees can be combined at the fully discrete level.It assesses conservation, positivity, and dissipation independently for detailed-balance production–destruction systems.Source: §1 · p. 3

2. Governing structure and discrete observables

The paper fixes an isothermal composition-space model and evaluates conservation, positivity, and free-energy dissipation as separate discrete observables. Its setup distinguishes face closure, non-orthogonal reconstruction, and source time integration as separate design choices.

2.1. Composition state and mixing potential

  • The mixing potential is an isothermal ideal-mixture Gibbs composition potential, with prescribed density and a Hessian positive definite on the simplex tangent space.The ln Z term gives ker H = span{Y}; entropy dissipation means non-increase of this same mixing potential, also called free energy.Source: §2.1 · pp. 3–4
  • The model fixes temperature and treats density as a prescribed positive, time-independent coefficient, while excluding barodiffusion, thermal diffusion, explicit ∇M terms, and gas–surface chemistry.It retains mixture-averaged diffusion, a stiff detailed-balance source, and potentially non-orthogonal meshes.Source: §2.1 · p. 4

2.2. The semi-discrete system and its three discrete choices

  • The semi-discrete scheme has three choices: face matrix and composition, reconstructed normal gradient, and time update; its observables are simplex residual, minimum mass fraction, and free energy.The face mass-correction term enforces zero total diffusive mass flux; the paper calls a construction certified only when an identity, inequality, or invariant proves the property under stated hypotheses.Source: §2.2 · pp. 4–6
  • The non-orthogonal reconstruction scales a tangential-gradient correction with θ: θ = 0 leaves the projected two-point operator, while θ = 1 uses the full correction.Weighted least-squares cell gradients supply the reconstruction, and the geometric schematic distinguishes the normal jump from the tangential contribution.Source: §2.2 · p. 5, Figure 2 · p. 6

2.3. Detailed-balance sources and Patankar updates

  • The source is a conservative production–destruction system built from detailed balance, for which continuous entropy production holds term by term.The reaction-based splitting distributes mass released by losing species proportionally among gaining species.Source: §2.3 · pp. 6–7
  • The first-order Patankar update targets unconditional positivity and exact conservation, but it does not constrain free energy.The construction costs one linear solve, and the missing free-energy inequality does not follow from the other two properties.Source: §2.3 · p. 7

3. Analytical results and numerical evidence

The paper distinguishes analytical results from finite numerical evidence and summarizes proved results alongside their hypotheses. Symbolic and numerical checks are verification controls rather than analytical assumptions.

3. Analytical results and numerical evidence

  • Exact identities and propositions are distinguished from numerical counterexamples and finite-population observations, which receive independent checks where noted.Population definitions, tolerances, and resampling procedures are given in Appendix B.Source: §3 · p. 7
  • Table 1 organizes analytical results by their additional hypotheses, with all rows also assuming the standing conditions of Section 2.Its propositions and corollaries are proved from the stated hypotheses.Source: Table 1 · p. 8

4. Layer 1: the two-point face certificate

The face entropy certificate is exact: entropy-variable jumps are represented by a log-mean Gibbs Hessian, and tangent-space neutrality fixes a unique conservative face weight. Tests show weight-dependent failures for synthetic diffusivities, while the tested transport controls satisfy the arithmetic-weight criterion.

4.1. An exact identity and an exact criterion

  • The entropy-variable jump equals the log-mean Gibbs Hessian times the mass-fraction jump exactly, making the closed-mesh two-point entropy criterion algebraic rather than asymptotic.The Hessian is evaluated at the log-mean composition, and the resulting certificate applies on closed meshes with periodic or zero-flux boundaries.Source: §4.1 · p. 8

4.2. A conservative tangent-neutral face weight

  • Conservation and tangent-space entropy neutrality uniquely determine a componentwise-positive face weight for ζ < 1, with uniqueness tied to the ideal-mixture Gibbs Hessian's rank-one structure.At ζ = 1, the weight has the continuous extension Y_log/s_log and c_f = 0.Source: §4.2 · p. 9
  • The tangent-neutral correction removes its rank-one entropy contribution without changing the Fickian face margin; other conservative weights can have an indefinite entropy contribution.For the normalized log-mean weight, the correction contribution's sign is indefinite.Source: §4.2 · p. 10

4.3. Face weighting and when the criterion is activated

  • On 39,999 random ternary face states, violations were 1.558 × 10^-2 for arithmetic and 1.200 × 10^-3 for normalized log-mean weights, versus none for tangent-neutral and ambient-minimizing weights.The tangent-neutral weight's zero-failure result is specific to this ternary mixture and diffusivity population; the upwind weight failed on 3.72% of the same states.Source: §4.3 · pp. 10–11
  • All 5,210 production neutral-air/carbon faces and 36,000 quasi-neutral ambipolar-air samples pass the arithmetic-weight criterion, with minimum margins +0.4888 and +0.161.Source: §4.3 · p. 11
  • Refinement activates violations only above a per-species log jump of 2.097, while the kinetic-order model keeps a +0.05187 worst margin through jumps of 18.4.Source: §4.3 · p. 11

5. Layer 2: multipoint non-orthogonal reconstruction

The tested weighted least-squares correction preserves conservation and improves geometric accuracy, but its multipoint stencil can lose the two-point entropy certificate and become anti-dissipative. Accuracy and consistency also depend on correction strength and mesh geometry; the paired-refinement tests separate temporal from spatial error.

5.1. The per-face certificate does not transfer

  • The WLS reconstruction reproduces constant and linear fields to 2.0 × 10−15, while conservation residuals remain near machine precision despite 540 positive local entropy contributions in 55 400 samples.The face mass residual is 1.3 × 10−15 and global species residual is 2.1 × 10−19; sampled global dissipation is not a certificate.Source: §5.1 · p. 12
  • With equal diffusivities, the WLS operator’s global entropy rate is +0.1436 versus −0.4083 for the projected two-point member at 𝜃= 0; five of 36 faces are anti-dissipative.The search used 600 states and four optimised starts; the independently reassembled 16-cell counterexample reproduces the rates.Source: §5.1 · p. 12, Figure 4 · p. 12

5.2. Accuracy against entropy

  • At the counterexample state, the global entropy rate crosses zero at 𝜃∗= 0.7398, where the linear-field normal-gradient error is 0.193.The entropy rate is affine in 𝜃, while the error falls from 0.740 at 𝜃= 0 to machine zero at 𝜃= 1.Source: Figure 4 · p. 12, §5.2 · p. 13
  • An entropy-only cap collapsed the explicit-Euler step to 8.3×10−27 after 80 halvings; a common 𝜃 satisfying entropy and one-step positivity advanced 100 steps with potential decrease 1.77 × 10−3.The construction is a global, state-dependent diagnostic, not a local limiter or invariant-domain result; its manufactured linear-gradient error was 0.219 to 0.432.Source: §5.2 · p. 13

5.3. Consistency depends on mesh geometry

  • On random Voronoi geometry, the curved cosine field’s largest spatial residual, 2.7358 × 10−4, exceeds the 5 × 10−5 reference tolerance; orthogonal controls reach observed 𝐿2 and 𝐿∞ orders of 1.986 and 1.965.The paired-grid decomposition attributes part of the residual to centroid point values versus polygonal cell averages.Source: §5.3 · pp. 13–14
  • Observed 𝐿2 orders are 1.98 for the orthogonal control, −0.01 for fixed relative displacement, 0.92 for 𝑂(ℎ), and 1.81 for 𝑂(ℎ2).The fixed-fraction displacement amplifies finest-mesh residuals by 15.1 in 𝐿2 and 25.0 in 𝐿∞.Source: §5.3 · p. 14, Figure 5 · p. 14
  • With unequal state-dependent diffusion (diffusivity ratio 19.95 and mass-correction flux 58.0% of Fickian flux), the four finest-interval 𝐿2 orders are 1.983, −0.006, 0.926, and 1.814.The classification is unchanged, and the state-dependent matrix introduces no additional 𝑂(ℎ) term at these settings.Source: §5.3 · pp. 14–15, Figure 5 · p. 14

5.4. Paired refinement and cross-seed sufficiency

  • Paired refinement gives a smallest temporal self-convergence order of 1.99991 and a finest-diagonal time-contamination ratio at most 2.04 × 10−6, separating temporal from spatial error.The reported 2.676 spatial order is a finite three-grid estimate, not an asymptotic-order claim; the multi-seed population does not meet the interval-stability target.Source: §5.4 · p. 15

6. Layer 3: positivity-preserving time integration

Patankar time integrators can preserve positivity and conservation without dissipating free energy. The tested remedy—rejecting and recomputing offending steps—recovers the three properties but advances only a fraction of proposed physical time.

6.1. Positivity and conservation do not imply dissipation

  • Across 150 networks and 30 step sizes, MP–Euler violates dissipation on 14.76% of steps, while none of 4 408 converged positive implicit-Euler roots does.On one network, Patankar schemes remain positive but raise Φ above Δt≈3τchem; classical schemes lose positivity before raising it.Source: §6.1 · pp. 16–17, Figure 6 · p. 16

6.2. An explicit counterexample

  • A rational finite-step counterexample shows that first-order Patankar integration can raise free energy despite unconditional positivity and conservation.The increase is certified independently of floating-point arithmetic; MPRK22 also has a strictly positive increase along an ascent direction.Source: §6.2 · p. 17

6.3. Coverage of the third-order members

  • All six tested third-order members remain positive and conservative yet raise free energy on 40.0–56.7% of 120 detailed-balance networks.The results cover the named members, not arbitrary Patankar schemes.Source: §6.3 · p. 18, Figure 7 · p. 18

6.4. Step rejection and adaptive recomputation

  • Dissipation-enforcing rejection and recomputation advances median fractions of 0.165–0.293 of proposed time across six members, with a smallest fraction of 0.109.The trajectories satisfy positivity, conservation, and monotone decrease of Φ; enforcing dissipation converts the unconditional method into an adaptive one.Source: §6.4 · p. 18, Figure 7 · p. 18
  • Observed convergence order under rejection and recomputation is not uniform; two MPRK43I variants fall below the reference tolerance of 2.8 on the tested refinements.High-precision checks identify a terminal error floor in one sample and a pre-asymptotic transition in another; these diagnoses do not reverse the observed failure.Source: §6.4 · pp. 18–19

7. What can and cannot be composed

A positive face entropy margin does not by itself ensure positivity after a fully discrete update. Spatial reconstruction and time integration require separate conditions, while tested repairs and guarantees remain state- or setting-dependent.

Fully discrete positivity

  • With a positive face margin of 0.8562, the two-cell backward-Euler example still produces a negative left-cell component of −5.995 × 10^-7 at Δt=2 × 10^-5.The step-size and composition scans show state-dependent positivity behavior, not a transferable timestep bound.Source: §7 · pp. 19–20

Source-only time integration

  • For source-only implicit Euler, a computed positive root dissipates free energy under convexity and detailed balance, but this does not establish root existence or cover the coupled spatial operator.The argument also requires mass conservation from the source balance.Source: §7 · p. 20

Spatial reconstruction

  • In the tested WLS reconstruction, increasing θ shifts the projected two-point rate from −0.4083 at θ=0 to a global rate of +0.1436 at θ=1, while linear-field error falls from 0.740 to machine zero.The curves cross at θ*=0.7398 with linear-field error 0.193; a common θ is a global state-dependent control, not a local limiter.Source: §7 · p. 20, Table 2 · p. 21

Time integration

  • Patankar rescaling preserves positivity and conservation for the examined members but does not ensure free-energy decay; rejection and recomputation recovers all three only as an observed adaptive control.The tested control advances 0.11–0.29 of proposed physical time, rather than providing an unconditional certificate for the original update.Source: §7 · p. 20

Composition conditions

  • Operator-split composition inherits the properties only when each admissible substep separately proves dissipation, conservation, and positivity for the same potential.The two-cell example shows why the face entropy inequality alone cannot supply the needed invariant-domain condition.Source: §7 · pp. 19–20

8. Discussion

The tested WLS reconstruction loses the face-by-face entropy decomposition through multipoint coupling, while Patankar rescaling constrains the update destination rather than its direction. The conclusions are bounded by the tested discretizations and model setting.

8.1. Why the certificates do not transfer

  • The tested cell-centred WLS correction couples faces, removing the two-point quadratic-form decomposition; global dissipation on smooth interior states still coexists with 166 positive local contributions across twelve cases.This is specific to the tested geometry, state family, and WLS construction, and does not rule out other structure-preserving spatial designs.Source: §8.1 · p. 21
  • Patankar rescaling constrains where an update lands, not its direction, so positive scalar relaxation cannot repair the same fixed ascent direction.The tested repair recomputes a different update and therefore incurs a physical-time cost.Source: §8.1 · p. 21

8.2. Scope and limitations

  • The transport criterion remained nonnegative for production neutral-air/carbon data and the quasi-neutral ambipolar-air control; its activation was confined to synthetic diffusivity populations..Source: §8.2 · p. 21
  • The WLS result concerns a tested multipoint construction, the Patankar result covers named members, and the common-θ control is global and state-dependent.The analysis is isothermal with prescribed density, excludes gas–surface and thermal coupling, and treats WLS and Patankar separately rather than in one reacting-flow simulation.Source: §8.2 · p. 22

9. Conclusions

The conclusions establish that each discretization layer needs its own certificate: a face-level entropy guarantee does not ensure stability after spatial reconstruction or time integration. The examples show that conservation, positivity, and free-energy dissipation are distinct constraints.

Conclusions

  • The face-level log-mean Gibbs Hessian identity yields a unique componentwise-positive face composition and an exact two-point entropy certificate under conservation and tangent-space entropy neutrality.Arithmetic-weight controls are also nonnegative for production neutral-air/carbon data and a quasi-neutral ambipolar-air model.Source: §9 · p. 22
  • The projected two-point operator has entropy rate −0.4083, but the fully corrected weighted least-squares operator has rate +0.1436 in the counterexample.The correction retains conservation and linear exactness despite losing the face-wise certificate.Source: §9 · p. 22
  • Patankar rescaling preserves positivity and conservation but can increase free energy: all six tested third-order members show this on 40.0–56.7% of tested detailed-balance networks.The finite-step counterexample is explicit; when an update is an ascent direction, positive scalar rescaling cannot restore dissipation.Source: §9 · p. 22
  • A positive face entropy margin does not ensure an inward-pointing species transport, and the two-cell backward-Euler example produces a negative mass fraction.The example separates the face-level entropy condition from positivity of the fully discrete update.Source: §9 · pp. 22–23
  • Conservation, positivity, and thermodynamic dissipation arise from different discrete algebraic structures and are not automatically inherited across solver stages.The paper’s layer-wise framework therefore requires compatible certificates at each stage.Source: §9 · p. 23

Appendix A. Proofs

The proofs connect entropy-variable jumps to mass-fraction jumps through a log-mean Gibbs Hessian, then use that identity to derive face entropy criteria and the unique tangent-neutral composition. They also establish how the criteria behave under quadratic forms and convex-potential rescaling.

Appendix A. Proofs

  • Applying the logarithmic-mean identity to both logarithms and using [[Z]] = b^T[[Y]] gives [[w]] = Hlog[[Y]].This expresses the entropy-variable jump exactly through the mass-fraction jump and the log-mean Hessian.Source: Appendix A · p. 23
  • On a closed mesh, interior-face cancellation makes each face contribute |f|[[w]]^T J_f to the summed semi-discrete entropy balance.Proposition 1 converts this face contribution into a condition on the diffusion flux.Source: Appendix A · p. 23
  • Only the symmetric part of Hlog A_f contributes to the quadratic form used in the face entropy criterion.The antisymmetric part does not affect that quadratic form.Source: Appendix A · p. 23
  • The log-mean Hessian has rank N−1 when ζ = 1, with Ylog spanning its kernel; when ζ ≠ 1, it is invertible.These cases underpin the separate derivations of the normalized face composition.Source: Appendix A · p. 23
  • Tangent neutrality requires Hlog Y_f to be a scalar multiple of 1, and conservation then fixes Y_f uniquely as the normalized (Hlog)^−1 1.For ζ = 1, the solution is Ylog/slog, also the limit of the general expression.Source: Appendix A · p. 24
  • For z in the simplex tangent space, replacing d by d + δ1 adds δ z^T Hlog z to the numerator of the Rayleigh quotient and leaves its denominator unchanged.This is the stated effect of the change in d in Proposition 6.Source: Appendix A · p. 24
  • Convexity ensures Φ(y + γΔy) ≥ Φ(y) + γ∇Φ(y)·Δy ≥ Φ(y) for γ ≥ 0 whenever the point lies in dom Φ.Thus a nonnegative directional derivative cannot be made dissipative merely by positive scalar rescaling.Source: Appendix A · p. 24

Appendix B. Verification populations and supporting controls

The appendix records verification procedures and the populations used for supporting controls. It reports benchmark checks, sample construction, and the stated refinement and interval-stability outcomes.

Appendix B. Verification populations and supporting controls

  • Three independent checks address implementation error: exact symbolic re-derivation of face-closure propositions, an external-package rerun of the time counterexample, and tests on three production–destruction benchmarks.The benchmarks do not satisfy detailed balance and therefore are excluded from free-energy claims.Source: Appendix B · p. 24
  • The six third-order members reproduce the Robertson reference to 8.7 × 10−11–1.5 × 10−9 over short times and remain positive and conservative over long times.The Brusselator reference error decreases with step size; the nonlinear algal-bloom problem has finest observed orders of 2.932–2.969.Source: Appendix B · p. 24
  • The multi-seed population contains 1,440 entries from 1,089 distinct computed grid–state cases in 480 paired refinement sequences.Networks are constructed to satisfy detailed balance, and conservation is exact by construction through molar masses solved from the mass-balance null space.Source: Appendix B · p. 24
  • The face-weight population has 39,999 admissible samples after excluding one draw with indistinguishable left and right states.The population definition fixed seeds, modes, perturbation families, grids, tolerances, and extension rule before evaluation; confidence intervals use 4,000 seed-cluster bootstrap replicates.Source: Appendix B · p. 25
  • Table B.1 describes numerical evidence populations, noting that evaluation counts need not equal independent statistical units and that its final column records outcomes under stated criteria.The caption gives the table’s scope but not its cell values.Source: Table B.1 · p. 25
  • Figure B.1 reports terminal L2 error along Δt = h/40 and bootstrap intervals for the finest L2 order; the fixed-η family fails the interval-stability target.The orthogonal control is a constant reference without seed variation, and fixed-η has no convergence order to estimate.Source: Figure B.1 · p. 26

Appendix C. Reproduction

The reproduction appendix documents checks for figures, reported values, and critical counterexamples, alongside a self-contained default reproduction command. It also records selected supporting values and the optional external dependency.

Appendix C. Reproduction

  • The default reproduction check verifies quoted numbers, runs the regression suite, and compiles the PDF; the scripts use frozen data files and record checksums.Each manuscript figure command is annotated with its generating script and data path, and a verification script rereads numeric values from authoritative data files.Source: Appendix C · pp. 26–27
  • The independent Julia reproduction gives relative state difference 1.23 × 10−15 and minimum component 1.13 × 10−2.These are supporting reproduction values, not additional claims.Source: Appendix C · p. 26
  • Run `bash manuscript/reproduce.sh` from the project root for the complete manuscript reproduction check.The submitted supplementary archive instead uses `bash reproduce.sh` from its root.Source: Appendix C · p. 26
  • The `--refresh-critical` command regenerates the tangent-neutral identity, frozen face-weight population, spatial positivity counterexample, and other critical checks.It also runs an independent NumPy-only reassembly of the Layer-2 counterexample.Source: Appendix C · p. 27
  • Only the optional external Patankar cross-check requires Julia and PositiveIntegrators.jl; no upstream solver source is modified by the archive scripts.The code, frozen input data, and checksum manifest accompany the manuscript as supplementary material.Source: Appendix C · p. 27
Loading 2609.27272v1…