Property-Preserving Data Compression for Incompressible Flows

1. Introduction.

• Compression errors can disrupt divergence, periodic-boundary agreement, and velocity-fluctuation statistics, even when velocity errors are small.The introduction notes that differentiation can amplify errors and independently compressed boundary values can disagree.Source: §1 · p. 1
• Fitting velocities with divergence-free functions makes every decompressed coefficient combination divergence-free, without a post-decompression projection solve.The same fitted representation can be paired with different compressors while retaining this guarantee.Source: §1 · p. 1
• Differentiated radial-basis kernels provide zero-divergence functions; periodic coordinate mapping and mixed length scales extend the construction to periodic and multiscale flows.The paper compares velocity statistics before and after compression.Source: §1 · p. 2
• Cavity, Taylor–Green, and species-transport flows test localized wall layers, periodic smooth structure, and multiscale variation, respectively.Together, these examples assess when a small approximation space suffices and what additional storage is needed when it does not.Source: §1 · p. 2
• The fitting options range from global and reduced Nyström fits to residual, multilevel Wendland, and simultaneous redundant-frame interpolation.Residual and all-node corrections can recover nodal accuracy when reduced fits leave error; the comparisons pair representations with four compressors and include nodal error bounds.Source: §1 · p. 2
• A compact periodic landmark space gives substantially higher compression than direct nodal storage for the Taylor–Green example.Source: §1 · p. 3

2.2. Coefficient representations.

• The five coefficient representations vary the fit and the coefficients supplied to the compressor, while sharing the same property-preserving kernel construction.The methods include global interpolation, two-level Nyström, residual interpolation, multilevel Wendland interpolation, and a redundant kernel frame.Source: §2.1 · p. 3, §2.2 · p. 6

2.1.1. Property-preserving kernels.

• Differentiating Matérn or Wendland scalar kernels in both arguments yields matrix-valued basis functions whose columns are divergence-free; periodic kernels use a torus embedding, while physical-scale mixtures target variation across multiple scales.Wendland kernels have compact support; the periodic construction is positive definite at distinct sites modulo the periods and produces fields that are both divergence-free and periodic.Source: §2.1.1 · pp. 3, 5

2.1.2. Fitting, compression, and evaluation.

• A reconstruction satisfies the chosen homogeneous constraints after compression when its evaluation map has range in the constraint operator’s kernel; compression error adds to any fitting error already present.The fitting, compression/decompression, and continuous-field evaluation maps separate how coefficients are formed, altered, and reconstructed.Source: §2.1.2 · pp. 5–6

2.2.1. Global interpolation and reduced Nyström approximation.

• Global Matérn interpolation uses dN coefficients for an exact nodal fit, whereas two-level Nyström reduces the fit to landmark spaces and may leave approximation error before compression.The second landmark level targets residual errors, and its space is orthogonal to the first; components outside the combined spaces remain approximation error.Source: §2.2.1 · pp. 6–8

2.2.2. Multilevel residual interpolation.

• The Nyström–Wendland hybrid interpolates the residual left by a reduced Matérn fit, removing its projection error before compression while preserving the constraints in both terms.A Wendland interpolation system supplies the correction; its local supports yield a sparse positive-definite matrix, while the species-transport scale mixture can make the matrix nearly dense.Source: §2.2.1 · p. 8, §2.2.2 · p. 8
• Multilevel Wendland interpolation fits successive residuals on nested center sets, ending with all nodes so the uncompressed reconstruction is exact to solver tolerance.Decreasing support radii keep cavity and Taylor–Green systems sparse; species transport uses its physical-scale mixture at each level and dense block Cholesky.Source: §2.2.2 · p. 9

2.2.3. Redundant kernel frames.

• The redundant kernel frame combines three nested center sets into an overcomplete representation, and selects the minimum-norm coefficient vector among those satisfying the nodal interpolation conditions.Its finest block contains every active node, giving the collocation matrix full row rank and allowing exact nodal reconstruction.Source: §2.2.3 · pp. 9–10
• For redundant kernel frames, the same kernel at every level gives A = KS and diagonal H = SST, whose entries count repeated kernel columns; minimum-norm fitting divides coefficients among copies.The resulting coefficient formula is c⋆ = ST H−1K−1uX.Source: §2.2.3 · p. 10
• Fitting and evaluation matrices are rebuilt from shared fixed-geometry data rather than stored per field, and coefficient blocks are compressed separately for cavity and Taylor–Green but concatenated for species transport.Table 1 summarizes fitting systems and uncompressed coefficient counts; coefficient counts do not determine compressed byte counts.Source: Table 1 · p. 10, §2.2.3 · p. 11
• Compression leaves the linear constraints satisfied across all five representations; nodal error depends on coefficient error and the evaluation matrix, with a sharper orthogonality-based estimate for two-level Nyström.The paper derives nodal error bounds for all five representations and uses orthogonality to sharpen the two-level Nyström estimate.Source: §2.2.3 · p. 11

3.1. Preservation under coefficient perturbations.

• For divergence and paired periodic-boundary differences, every decompressed field remains constraint-preserving regardless of compressor tolerance.The evaluation map sends coefficient vectors into the kernel of the constraint operator.Source: §3.1 · p. 11

3.2. Error induced in the nodal values.

• The nodal error bound adds the existing fit error to a compression-error term controlled by coefficient tolerances, block sizes, and evaluation-matrix norms.Maximum-absolute and RMS coefficient-error assumptions are covered; choosing tolerances appropriately limits the additional nodal error to τ.Source: §3.2 · pp. 11–12

3.3. Bounds for the coefficient representations.

• Representation-specific bounds use orthogonality for two-level Nyström and norm scaling for interpolation or frame blocks; floating-point identities hold only up to fitting, orthogonality, and evaluation errors.The bounds concern stored nodes, and increasing landmark count can reduce the fit error.Source: §3.3 · pp. 12–13

• The study tests lid-driven cavity, Taylor–Green, and species-transport flows, using five velocity fields per flow.The examples respectively probe localized wall layers, periodicity with smooth structure, and multiscale variation.Source: §4 · p. 13
• The experiments first verify nodal error bounds, then compare compression at matched velocity accuracy and reconstructed-field properties.All methods use the same five velocity fields for each flow.Source: §4 · p. 13
• The three flows test whether reduced approximation spaces suffice and what additional storage is needed when they do not.Source: §4 · p. 13

• Rate–distortion curves show mean compression ratio against the largest relative field error, selecting the highest ratio that meets each limit across all five fields.The selected compressor may change along a curve; fixed-compressor comparisons appear separately.Source: §4.1 · pp. 13–14
• Taylor–Green compression stores one velocity per distinct periodic node, while a separate test compresses duplicate boundary values independently to probe loss of endpoint equality.Source: §4.1 · p. 14
• Incompressibility is assessed by divergence change for direct compression, while kernel reconstructions are divergence-free analytically; Taylor–Green periodicity is measured by boundary-pair velocity mismatch.The property plots average measures over five fields and include only settings that meet the velocity-error limit on every field.Source: §4.1 · p. 14
• For species transport, changes in velocity fluctuations are measured with the second-order structure function and a fitted exponent p over specified separation bins.The exponent is compared with the uncompressed value and the Kolmogorov prediction 2/3; the turbulence-closure source does not establish a resolved inertial cascade.Source: §4.1 · p. 14

• Across tested flows and landmark counts, nodal compression errors stayed below the applicable estimate when the compressor met its coefficient bound.Table 2 reports observed nodal compression error divided by the predicted bound; values below 1 satisfy it.Source: §4.2 · p. 15, Table 2 · p. 15
• The estimate was closest for two-level Nyström because its orthonormal evaluation preserves coefficient-perturbation norm; its total error also includes projection error.The hybrid’s full-node Wendland correction removes approximation error up to interpolation solve accuracy.Source: §4.2 · p. 15
• Independent coefficient perturbations verified the Nyström total-error identity to floating-point accuracy.Source: §4.2 · p. 15
• Storage comparisons use a common velocity-error accuracy of 10^-3, while the kernel guarantees apply at every compression tolerance.Source: §4.2 · p. 15

Comparison across compressors

• Curves are Pareto envelopes across four compressors, and successive points may use different compressors because the best compressor varies with coefficient distribution.The reported comparisons retain settings not dominated in compression and error.Source: §4.3 · pp. 15–16

Lid-driven cavity

• The tested reduced landmark spaces leave cavity fitting error above 0.2 even at the largest size, while the full Wendland correction removes that approximation error.Tightening coefficient tolerance cannot recover velocity components outside the retained space.Source: §4.3 · p. 16, Figure 1 · p. 16
• At relative error 10^-3, the frame has the highest compression ratio among tested divergence-free cavity methods, followed by the hybrid; both remain below direct compression.Despite its larger coefficient count, the frame requires the least compressed storage among kernel representations at this accuracy.Source: §4.3 · pp. 16–17, Figure 1 · p. 16

Taylor–Green flow

• At relative error 10^-3, two-level periodic Nyström with M = 32 gives the largest tested Taylor–Green compression ratio, 2.92 × 10^3.The reduced space resolves all five fields below the limit; the hybrid requires more storage.Source: §4.3 · p. 17, Figure 2 · p. 17

Species transport

• At relative error 10^-3, global Matérn gives the best species-transport kernel compression at 3.01×, compared with 4.53× for direct compression.The hybrid becomes marginally more economical at 10^-2, so the preferred interpolant can depend on the error limit.Source: Figure 1 · p. 16, §4.3 · p. 17
• For species transport, none of the tested Nyström spaces meets maximum relative error 10^-1, whereas the full Wendland correction removes their approximation error.The hybrid interpolates nodal values to solver tolerance before compression; its plotted error reflects compression and finite-precision fitting and evaluation.Source: §4.3 · pp. 16–17

4.4. Property preservation after compression

• Kernel fields remain divergence-free and periodic at every compression tolerance; the plotted divergence zero band denotes analytic freedom, not measured PHS error.The study selects the most compressed setting meeting each velocity-error limit on all five fields.Source: §4.4 · p. 17, Figure 3 · p. 18

4.4.1. Incompressibility

• In the cavity at relative velocity error 10^-3, direct compression achieves a higher ratio but introduces nonzero divergence change; the frame leads tested divergence-free methods.The storage advantage of direct compression comes without the incompressibility guarantee.Source: §4.4.1 · p. 18, Figure 4 · p. 19
• At relative velocity error 10^-3, Taylor–Green two-level Nyström and the hybrid exceed direct nodal storage's compression ratio while retaining zero analytic divergence.Direct compression introduces nonzero divergence change at this limit.Source: §4.4.1 · p. 18
• For species transport at relative velocity error 10^-3, global Matérn is the most economical representation guaranteeing zero divergence, while direct compression uses less storage but changes divergence.Smaller spaces may exceed the tested velocity-error limits even before compression.Source: §4.4.1 · p. 18, §4.4 · p. 18

4.4.2. Periodicity

• Direct compression's periodic-boundary mismatch can drop to zero with PFPL and rise with MGARD; the increase reflects compressor switching under storage-focused selection.Periodic kernels define continuously periodic reconstructed fields, whereas endpoint copying enforces equality only on paired nodes.Source: §4.4.2 · pp. 18–19, Figure 4 · p. 19

4.4.3. Velocity fluctuations across scales

• At relative velocity error 10^-3, species-transport exponent changes for direct points, global Matérn, multilevel Wendland, and the frame are of order 10^-5.The uncompressed mean fitted exponent is approximately 0.685, near 2/3; agreement with 2/3 is specific to the chosen fitting interval.Source: §4.4.3 · p. 19

Residual levels

• For Taylor–Green at relative error 10^-3, two Wendland corrections reduced median compression ratios by approximately 2–18% for CPU codecs and 1–20% for GPU codecs.The cavity comparison also favored one correction, so the study retained one level.Source: §4.5 · pp. 19–20

Frame ordering and separate block compression

• For Taylor–Green frames, Morton ordering improved median compression by approximately 9% for ZFP, 23% for PFPL, 34% for SZ3, and 17% for MGARD.GPU outcomes varied: PFPL gained 23%, CUDA ZFP changed little, and cuSZ and MGARD-X lost approximately 20% and 22%.Source: §4.5 · p. 20
• For species frames at 10^-2, separate-level compression changed the best ratio from 3.464 to 3.462 on both CPU and GPU, so single-array compression was retained.Error allocation among levels remains to be optimized.Source: §4.5 · p. 20

4.6. GPU compressor comparison.

• PFPL produces the same rate–distortion curve on CPU and GPU, consistent with its design to produce identical compressed files.Absolute-error mode is used on both devices.Source: §4.6 · p. 23, Figure 10 · p. 23
• CUDA ZFP reaches at most approximately 16× compression in the tested range because it prescribes bits per value rather than an error tolerance.CPU ZFP instead prescribes an absolute error tolerance.Source: §4.6 · p. 23, Figure 10 · p. 23
• MGARD-X satisfies its RMS bounds but achieves lower compression ratios than CPU MGARD at comparable field errors on one-dimensional coefficient vectors.Source: §4.6 · p. 23, Figure 10 · p. 23
• Accepted cuSZ streams may exceed the requested maximum coefficient error by up to 1.1×, so coefficient-error estimates must use that acceptance threshold.Small arrays are stored uncompressed; larger arrays also revert to uncompressed storage when they exceed the threshold.Source: §4.6 · p. 23
• CPU-to-GPU speedups are not reported because fitting, compression, decompression, and evaluation times were not measured consistently.Source: §4.6 · p. 23

4.7. Limitations.

• Five velocity fields per dataset underpin the comparisons, and the tested properties are incompressibility, periodicity, and selected second-order velocity statistics.Source: §4.7 · p. 23
• Full trajectories, long-time restart stability, and downstream solver accuracy remain unstudied.Source: §4.7 · p. 23
• Compression ratios exclude geometry, center indices, and kernel parameters shared on a fixed mesh.Source: §4.7 · p. 23
• The CPU and GPU experiments establish no speed advantage because they lack consistent execution-time measurements.Source: §4.7 · p. 23
• Automatic selection of landmark counts, center hierarchies, and kernel scales from accuracy and storage budgets remains open.Source: §4.7 · p. 23

5. Conclusions.

• The methods preserve incompressibility and periodicity analytically and provide nodal error bounds for lossy compression.Source: §5 · p. 24
• A sufficiently accurate uncompressed two-level Nyström fit can yield substantial storage savings through its small coefficient count.If its fitting error exceeds the target, tighter compression tolerances cannot help; residual interpolation or another interpolating representation is needed.Source: §5 · p. 24
• Compressed size depends on the flow and compressor, not just coefficient count, so representations should be compared at the intended accuracy on representative fields.The paper recommends comparing a few representations, perhaps one from each broad class.Source: §5 · p. 24
• GPU compressors avoid host transfers of coefficients, but their different compression modes require comparisons at matched field errors.Source: §5 · p. 24
• Periodic kernels enforce equality across paired boundaries throughout the domain and preserve zero divergence, unlike periodic-node storage, which enforces equality only at paired boundary nodes.Source: §5 · p. 24
• Neither periodic-node storage nor periodic kernels guarantee preservation of turbulent velocity statistics, which must be checked over the spatial scales of interest.Source: §5 · p. 24

Declaration of AI assistance.

• Codex assisted with manuscript editing.Source: Declaration of AI assistance · p. 24
• Codex assisted with code development and computational experiment workflows.Source: Declaration of AI assistance · p. 24
• The authors assume responsibility for all content.Source: Declaration of AI assistance · p. 24

SM1. Reproducibility and experimental parameters

• Compression curves average per-field compression ratios across five test fields and report worst-field relative ℓ2 and maximum pointwise errors, retaining nondominated settings that satisfy coefficient-error checks.Kernel methods compress coefficient vectors, direct compression acts on nodal velocities, and reconstructed-field error is checked independently.Source: SUPPLEMENTARY MATERIALS:… · p. 29

SM2. Additional nodal error comparisons

• Among settings meeting their coefficient bounds, the largest observed fractions of the allowances b_globalτ were 0.037 and 0.015, respectively; norm calculations used floating-point arithmetic.The computed b_global factors were 1.087 for cavity and 1.277 for Taylor–Green.Source: SUPPLEMENTARY MATERIALS:… · p. 32
• For Taylor–Green, the largest fractions of allocated nodal allowances were 0.051 for multilevel Wendland, 0.055 for the Morton-ordered blockwise frame, and 0.211 for the hybrid.One frame ZFP test at τ = 40000 exceeded its requested maximum coefficient error by about 10% and was excluded from curves and theorem comparisons requiring that bound.Source: SUPPLEMENTARY MATERIALS:… · p. 32
• Two-level Nyström checks found orthogonality errors below 5.8 × 10^-12 for cavity and 2.0 × 10^-13 for Taylor–Green, with squared total-error identity discrepancies below 4.1 × 10^-15∥uX∥2^2.The largest discrepancy relative to squared total error was 3.2 × 10^-7; deterministic coefficient perturbations produced nodal perturbations below 1.2 × 10^-9, including evaluation roundoff.Source: SUPPLEMENTARY MATERIALS:… · p. 32

SM2.1. Sensitivity of the structure-function exponent

• The mean fitted exponent ranged from 0.762 on the shifted-smaller-separation interval to 0.507 on the extended interval, changes much larger than compression-induced shifts.The original interval yielded 0.685; the sensitivity indicates that the fitted exponent cannot identify a robust inertial-range law, and compression-induced changes apply only to the fixed original interval.Source: SUPPLEMENTARY MATERIALS:… · p. 33

SM3. Maximum pointwise errors and comparisons by compressor

• Supplementary comparisons report largest pointwise-error curves and separate compressor curves at fixed landmark counts; direct-point periodicity mismatch rises when the selected compressor changes from PFPL to MGARD.The mismatch increase reflects maximizing compression subject to velocity accuracy without a boundary-mismatch constraint, not a threshold for a single compressor.Source: SUPPLEMENTARY MATERIALS:… · p. 33, Figure SM1 · p. 34, Figure SM2 · p. 34

Compressor curves

• Additional nondominated error–compression curves cover cavity with SZ3 and MGARD, Taylor–Green with ZFP, PFPL, SZ3, and MGARD, and species transport with ZFP, SZ3, and MGARD.The figures report the largest relative ℓ2 error across five velocity fields versus mean compression ratio.Source: Figure SM6 · p. 36, Figure SM7 · p. 37, Figure SM8 · p. 37

Boundary mismatch

• For Taylor–Green, the direct-point boundary-mismatch curve switches from ZFP to PFPL and then MGARD as the velocity-error limit increases.The upward jump occurs when MGARD replaces a PFPL setting with zero boundary mismatch; exact zeros appear in the labeled band.Source: Figure SM12 · p. 39

Compressor curves

• For species transport with PFPL, the loosest settings erase velocity variation: all five fields have unit relative error and compress to 84 bytes each (approximately 3254×).The high-ratio tail reflects loss of the signal, not accurate reconstruction.Source: Figure SM14 · p. 40
• For species-transport PFPL curves, the direct-point result flattens near unit relative error as loose tolerances erase velocity variation.The compressor is fixed throughout the figure.Source: Figure SM14 · p. 40

Device-resident compressor curves

• Device-resident comparisons extend to cavity flow with M = 128 and Taylor–Green with M = 256, plotting compressed size against measured reconstruction error.PFPL (CPU) is the reference; CUDA ZFP uses fixed rate, while the other modes control error.Source: SUPPLEMENTARY MATERIALS:… · p. 42, Figure SM17 · p. 42, Figure SM18 · p. 43