PaperlayerView exact revision on arXiv

Source-linked AI summary

FeatureZ: A General Framework for Feature-Preserving Compression via Pointwise Bounds and Star Classification

Nathaniel Gorski, Xin Liang, Hanqi Guo, Bei Wang

arXiv:2610.12371v1cs.GRcs.ITmath.IT

TL;DR

Most lossy compressors guarantee pointwise error but do not preserve derived features, and existing feature-preserving methods are typically specialized. FeatureZ augments lossy compressors using pointwise bounds and star classifications; experiments report diverse feature preservation with competitive, often superior, compression ratios and runtimes. Its guarantees apply to features expressible within this framework, and specifying such features can be nontrivial.

  • Problem

    Most lossy compressors control pointwise error without preserving derived features, while existing feature-preserving compressors are difficult to develop and typically target one feature.

    Source: §1 · p. 1

  • Method

    FeatureZ augments an existing compressor by enforcing pointwise upper and lower bounds through quantization and consistent star classifications through an iterative procedure.

    Source: §1 · p. 2, §4.2 · p. 4

  • Results

    FeatureZ preserves diverse feature combinations across multiple base compressors, with compression ratios and runtimes competitive with, and often superior to, existing feature-preserving approaches.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Users can incorporate user-defined features by adding a few functions without redesigning the compression pipeline.

    Source: §1 · p. 2

    Feature specification requires sufficient conditions within the framework, can be unintuitive, and does not encompass all features of interest.

    Source: §7 · p. 9

Abstract

from arXiv · show

Geometric and topological features, such as isosurfaces, quantiles, merge trees, and Morse-Smale complexes, are central to the analysis and visualization of scientific data across diverse domains, including medical imaging, climate science, materials science, and astronomy. However, most lossy compressors for scientific data provide only pointwise error guarantees and do not preserve derived features. Existing feature-preserving compressors are often difficult to develop and typically tailored to a single feature. In this paper, we introduce FeatureZ, a lossy compression framework for structured volumetric scalar fields that can preserve a wide class of geometric and topological features. In particular, FeatureZ frames feature-preserving compression as preserving pointwise upper and lower bounds together with a consistent classification of each point's star (i.e., its incident cells) in the underlying structured mesh. Although not all notions of feature preservation fit this framework, many features and topological descriptors targeted by existing compressors do. FeatureZ provides an efficient implementation of feature-preserving compression under this formulation. In particular, FeatureZ operates as an augmentation layer that refines the output of an existing lossy compressor. It first applies quantization to enforce pointwise bounds, and then employs an iterative procedure to ensure consistent star classification. We demonstrate that FeatureZ preserves diverse features during compression with minimal overhead, achieving compression ratios comparable to or better than methods specialized for individual features.

1 INTRODUCTION

Small pointwise error guarantees do not ensure that derived geometric and topological features survive compression, while existing feature-preserving compressors are costly to develop and feature-specific. FeatureZ offers a general, extensible framework that augments existing compressors to preserve a broad class of features under user-specified error bounds.

1 INTRODUCTION

  • Geometric and topological features describe the shape, structure, and connectivity of scientific data and support analysis and visualization across domains.Examples include isocontours in medical imaging, contour trees in astronomy, and Morse–Smale complexes in quantum chemistry.Source: §1 · p. 1
  • Most lossy compressors control pointwise reconstruction error but do not preserve derived features, which can be distorted even when global error bounds are small.Such distortions can compromise downstream analysis and visualization.Source: §1 · p. 1
  • Existing feature-preserving compressors are difficult to develop and typically target one feature, leaving users to manage a fragmented collection of specialized tools.The paper identifies a need for a unified framework that supports user-defined feature preservation.Source: §1 · pp. 1–2
  • FeatureZ formulates preservation through pointwise bounds and consistent star classification, covering many geometric and topological features, including those targeted by existing topology-preserving compressors.The framework also supports several feature classes not previously supported by dedicated compressors.Source: §1 · p. 2
  • FeatureZ augments an existing lossy compressor with quantization-based bound correction and an iterative scan that ensures consistent star classification.The procedure uses local information from each point and its star, avoiding repeated global feature computation.Source: §1 · p. 2

2 RELATED WORK

Scientific data compression balances storage reduction against controlled reconstruction error, but derived features require preservation beyond pointwise error bounds. Prior feature-preserving methods address particular structures, while compression methods span prediction, neural, and transform-based approaches.

2 RELATED WORK

  • Lossy compression permits controlled distortion in exchange for higher compression ratios, while lossless methods typically achieve limited reduction, often under 2×.Scientific instruments and simulations can generate terabytes of data.Source: §2 · p. 2
  • Prediction-based compressors estimate values and encode residuals, with methods ranging from B-spline interpolation to SZ3's combination of predictors.Deep-learning approaches include coordinate-based implicit neural representations and encoder-decoder autoencoders; INRs are computationally expensive to train and evaluate.Source: §2 · p. 2
  • Transform-based compressors decorrelate data using methods such as ZFP's block transforms, tensor decomposition, and SPERR's wavelet transforms with SPECK coding.Derived features such as isocontours and quantiles are used in scientific visualization and medical image analysis.Source: §2 · p. 2
  • Prior feature-preserving compressors target specific structures, including persistence diagrams, contour trees, Morse–Smale segmentations, extrema, quantities of interest, and local order.Related feature-preserving methods have also been developed for vector and tensor fields.Source: §2 · p. 2

3 TECHNICAL BACKGROUND

The background defines piecewise-linear scalar fields and local mesh neighborhoods, reviews key topological descriptors, and explains quantization for error-bounded compression. These concepts underpin FeatureZ's later use of pointwise bounds and star classifications.

3.1 Piecewise-Linear Scalar Fields

  • A piecewise-linear scalar field assigns values to mesh vertices and extends them linearly over simplices, with simulation of simplicity resolving equal vertex values by coordinate order.A vertex's star comprises its incident simplices and faces; its link excludes the vertex, and upper and lower links classify neighboring simplices by value.Source: §3.1 · pp. 2–3

3.2 Topology of Piecewise-Linear Scalar Fields

  • Topological descriptors encode critical points and the evolution of connected components and level sets, including merge trees, persistence diagrams, contour trees, and Morse–Smale segmentations.Persistence simplification removes features below a threshold; the figure illustrates these descriptors and simplified trees.Source: §3.2 · p. 3, Figure 2 · p. 3

3.3 Quantization in Lossy Compression

  • Linear-scaling quantization shifts a predicted value into an interval around the ground truth to satisfy error bound ξ, while logarithmic scaling refines intervals to meet additional pointwise bounds.Logarithmic scaling increases precision until bounds are satisfied or stores the value losslessly; quantization indices are entropy coded.Source: §3.3 · pp. 3–4, Figure 3 · p. 3

4 PRESERVATION OF FEATURES

FeatureZ preserves features by enforcing sufficient conditions expressed through pointwise bounds and consistent local star classifications. The framework applies these conditions to diverse features and allows multiple feature types to be preserved together.

4 PRESERVATION OF FEATURES

  • FeatureZ is an augmentation framework that enforces a global error bound while preserving user-defined features characterized by pointwise bounds and consistent star classification.Its supported features share sufficient conditions for preservation.Source: §4 · p. 4, §4.1 · p. 4

4.1 Overview of Feature Preservation

  • For each feature, FeatureZ assigns points a star-based classification and prescribes pointwise upper and lower bounds, requiring reconstructed values to satisfy the bounds or equal the original.Features are supported when their invariance follows from these conditions; multiple features can be preserved by enforcing each set of conditions.Source: §4.1 · p. 4

4.2 Preservation of Diverse Features

  • Critical points are preserved by matching criticality types, while simplified contour and merge trees use bounds that preserve thresholded branches and limit extra-branch persistence to η + 2ξ.Choosing η = ε − 2ξ yields matching ε-simplified trees when ε ≥ 2ξ; unsimplified trees use η = 0 and may contain extra branches of persistence at most 2ξ.Source: §4.2 · p. 4
  • Morse–Smale segmentation is preserved by classifying extrema and the highest and lowest vertices in each link, while local order is encoded by the partition into upper and lower links.Matching these classifications preserves the neighbors followed by integral lines and therefore the segmentation.Source: §4.2 · pp. 4–5
  • Isocontours and their sublevel and superlevel sets are preserved with pointwise bounds, quantiles with membership-preserving thresholds, and feature combinations by enforcing all associated bounds and classifications.The paper also describes preservation capabilities of prior compressors and identifies quantiles and simultaneous preservation as unsupported targets among the listed prior methods.Source: §4.2 · p. 5

5 CORE COMPRESSION ALGORITHM

FeatureZ augments an existing compressor with corrections that preserve a user-defined feature while meeting pointwise error bounds. Its pipeline combines bound-aware quantization with iterative checks and corrections of star classifications.

5 CORE COMPRESSION ALGORITHM

  • FeatureZ augments a base compressor with compact corrections to preserve a user-defined feature Φ while enforcing Φ(f)=Φ(f′) and pointwise error ≤ ξ.The corrections are stored alongside the base compressor output.Source: §5 · p. 5

5.0.1 Pipeline

  • The pipeline compresses and decompresses first, computes feature bounds, quantizes within them, iteratively corrects stars, then losslessly encodes quantization data with the base output.The initial quantization maintains both the feature bounds and the pointwise error bound.Source: §5.0.1 · p. 5

5.0.2 Iterative Star Correction

  • When star checks find bad points, FeatureZ increments the precision of selected vertices, tightens their quantization bounds, and re-quantizes them until no bad points remain.If precision exceeds the allowed maximum, the vertex is stored losslessly; the procedure ensures at least one bad point differs from its reconstruction to ensure termination.Source: §5.0.2 · p. 5
  • To reduce repeated work, FeatureZ processes only dirty vertices, marking updated vertices and neighbors for the next iteration and switching to sparse storage below 20% dirty.The dirty set starts as a bit array and switches to storing indices when fewer than 20% of vertices remain dirty.Source: §5.0.2 · p. 5

5.0.3 Identification of Bad Points

  • Feature-specific checks mark mismatched critical-point types or link memberships bad; Morse–Smale checks also flag vertices when extremal link neighbors differ.For the Morse–Smale condition, the compared neighbors are the largest in the upper link and the smallest in the lower link.Source: §5.0.3 · pp. 5–6

6 EXPERIMENTAL RESULTS

The experiments assess FeatureZ across seven features, datasets, and compressor configurations. FeatureZ generally matches or exceeds specialized methods in reconstruction quality, while runtime varies with preservation strategy and includes overhead from correction and lossless compression.

6.1 Overview of Experiments

  • The evaluation covers seven features, including critical points and percentiles without dedicated compressors, plus a combined contour-tree and Morse–Smale task.FeatureZ is tested with four base compressors across nine datasets; comparisons include feature-specific methods where supported.Source: §6.1 · p. 6

6.2 Reconstruction Quality

  • FeatureZ achieves reconstruction quality comparable to or better than specialized compressors while supporting a broader range of features.Across features, it generally outperforms pMSz and LOPC and is competitive with TopoA and QPET.Source: §6.2 · pp. 6, 8, Figure 4 · p. 7
  • At comparable compression ratios on Earthquake, more than 96% of points receive incorrect Morse–Smale segmentation regions with unaugmented SZ3.The comparison highlights feature errors that can persist even when compression ratios match FeatureZ.Source: §6.2 · p. 8

6.3 Analysis of Compression Time

  • The base compressor accounts for less than 5% of runtime; bound computation, star correction, and lossless compression contribute substantial overhead.Bound computation is often dominated by simplified contour-tree construction, while star correction incurs neighborhood-access and classification costs.Source: §6.3 · p. 8
  • Percentile and isocontour preservation achieve higher throughput without star correction, while critical-point and Morse–Smale preservation run slower due to iterative correction.Simplified contour-tree preservation also avoids star correction but requires costly contour-tree computation.Source: §6.3 · pp. 8–9
  • FeatureZ’s runtime is O(n log n) for contour-tree preservation and O(n) for other implemented features.These bounds assume the implemented features’ star-classification cost is O(1) and their k, S, and M parameters are small constants.Source: §6.3 · p. 9

7 LIMITATIONS AND DISCUSSIONS

FeatureZ’s limitations span overhead, feature and parameter constraints, visual artifacts, and design trade-offs. Some choices improve efficiency but can limit flexibility or compression quality.

Storage and runtime overhead

  • Storage overhead reaches ~2×, and runtime remains slower than some specialized approaches, including GPU-based methods.Throughput is comparable to some feature-preserving methods; star correction and several bound computations offer parallelization opportunities.Source: §7 · p. 9

Visual artifacts

  • Isocontour artifacts can be noticeable because points may shift by up to one grid-cell length, though the artifacts diminish at higher resolutions.Source: §7 · p. 9

Design choices

  • Local guarantees and greedy vertex-by-vertex correction favor neighborhood operations but may limit flexibility and introduce more corrections than necessary.Joint correction strategies could potentially preserve the same features with lower storage overhead.Source: §7 · p. 9

Feature specification

  • The current framework does not cover every feature of interest, and defining sufficient preservation conditions can be unintuitive for novice users.Source: §7 · p. 9

Parameter selection

  • Without a systematic strategy for selecting error parameters, initial settings may be suboptimal; decreasing the error bound can sometimes reduce file size and improve reconstruction quality.Source: §7 · p. 9

Degenerate cases

  • In adversarial cases, preserving an exceptionally complex unsimplified contour tree may require near-lossless storage.Source: §7 · p. 9

Preserving arbitrary features

  • A variant for arbitrary categorical features without star classification requires repeated feature recomputation, causing substantial overhead.Source: §7 · p. 9

8 CONCLUSION

FeatureZ preserves a broad class of features under a user-specified error bound by enforcing pointwise bounds and correct star classification. The supplement documents datasets, implementation and proofs, and provides additional experimental results and parameter values.

8 CONCLUSION

  • FeatureZ combines pointwise upper and lower bounds with iterative star classification to preserve a broad class of features while maintaining the user-specified error bound ξ.Source: §8 · p. 9

A SUPPLEMENTARY OVERVIEW

  • The supplement covers datasets, critical-point extraction, adding features, correctness proofs, additional experiments, and experimental parameter values.Source: A SUPPLEMENTARY OVERVIEW · p. 12

B DESCRIPTIONS OF DATASETS

The supplement describes nine evaluation datasets and details critical-point extraction, feature integration, and proofs of pipeline correctness. Under its stated assumptions, the pipeline terminates with every point in-bounds and correctly classified.

B DESCRIPTIONS OF DATASETS

  • The experimental study evaluates nine datasets, including QMCPACK, Viscous Fingering, Tangaroa, Earthquake, Ionization, Isabel, Miranda, S3D, and SCALE-LETKF.Source: B DESCRIPTIONS OF DATASETS · p. 12

C EXTRACTION OF CRITICAL POINTS

  • Interior criticality is determined from a 14-bit encoding of upper-link membership; boundary points use an unambiguous subset encoding to avoid absent-neighbor ambiguity.Source: C EXTRACTION OF CRITICAL POINTS · p. 12

D DEFINING AND INCORPORATING NEW FEATURES

E CORRECTNESS PROOFS

  • The supplement provides correctness proofs for the main pipeline, individual features, and contour-tree preservation.Source: E CORRECTNESS PROOFS · p. 12

E.1 Definitions

  • The correctness setup defines upper and lower bounds, a star-classification function, and conditions for points to be in-bounds, correctly classified, and nondegenerate.Source: §E.1 · pp. 12–13

E.2 Correctness of Main Pipeline

  • Assuming nondegenerate classification, the main pipeline terminates with every point in-bounds and correctly classified; these guarantees alone do not ensure preservation of a particular feature.Source: §E.2 · p. 13

E.3 Correctness of Individual Feature Preservation

The proofs establish preservation of critical points, Morse–Smale segmentation, local order, isocontours, and quantiles under the framework’s specified conditions. The contour-tree correctness section begins with merge- and contour-tree preliminaries and persistence simplification.

E.3 Correctness of Individual Feature Preservation

  • FeatureZ correctly preserves critical points when all points are correctly classified, since this ensures each point retains its criticality type.Source: §E.3 · p. 14
  • Correct classification preserves Morse–Smale segmentation by keeping forward and backward integral lines identical between the original and reconstructed fields.Source: §E.3 · p. 14
  • When no vertices are marked bad, FeatureZ preserves the less-than relationships that define local order.Source: §E.3 · p. 14
  • For a preserved isovalue, every simplex is intersected by the original isocontour if and only if it is intersected by the reconstructed one, provided all points are in-bounds.Source: §E.3 · p. 14
  • Quantile preservation keeps the threshold separating the m smallest points from the n −m largest consistent across the original and reconstructed fields.Source: §E.3 · p. 14

E.4 Correctness of Contour Tree Preservation

  • The contour-tree treatment introduces merge- and contour-tree notation and defines persistence simplification by pruning low-persistence branches before combining simplified merge trees.Source: §E.4 · p. 14, §E.4.1 · p. 15

E.4.2 Supporting Lemmas

The supporting lemmas establish how ε-bounds preserve contour boundaries and merge-tree structure under a global pointwise error bound. They also bound the persistence of additional minima in the reconstructed function.

Supporting lemmas

  • A contour at a saddle’s isovalue that bounds segmentation regions remains intersected by the same simplex after reconstruction.The proof treats vertices, edges, and higher-dimensional simplices.Source: §E.4.2 · p. 16
  • If two points lie in different sublevel-set components at a retained saddle, they remain disconnected below that saddle’s value after reconstruction.The lemma supports preservation of merge-tree ancestor relationships.Source: §E.4.2 · p. 16
  • An ε-persistent split saddle retains its pairing with the same minimum in the reconstructed merge tree.The result follows by preserving the relevant connected-component relationships and the minima in those components.Source: §E.4.2 · p. 16
  • A reconstructed minimum that is not a node of the original persistence-simplified tree has persistence at most ε + 2ξ.Here ξ is the global error bound assumed for f′.Source: §E.4.2 · pp. 16–17

E.4.3 Proof of Main Result

The proof establishes FeatureZ’s contour-tree preservation guarantees, while the additional experiments cover reconstruction quality and runtime behavior. The reported throughput response to changing the error bound varies by configuration.

E.4.3 Proof of Main Result

  • FeatureZ preserves simplified contour trees and unsimplified contour trees under the respective constructions described in the proof.For simplified trees, the construction uses (ε − 2ξ)-boundedness; for unsimplified trees, it uses 0-boundedness.Source: §E.4.3 · p. 17

F ADDITIONAL EXPERIMENTS

  • The additional experiments include rate-distortion curves for evaluated compressors and features, runtime breakdowns, pipeline ablations, parameter sensitivity, feature-error comparisons, and evaluations with Neurcomp and TopoSZp.The overview lists these as appendix experiments.Source: F ADDITIONAL EXPERIMENTS · pp. 17–18

F.1 Reconstruction Quality

  • The reconstruction-quality curves compare FeatureZ variants using ZFP, SZ3, SPERR, and HPEZ with specialized feature-preserving compressors where applicable, plus the unaugmented base compressors.Dominated rate-distortion points are removed.Source: §F.1 · p. 18

F.2.1 Runtime for FeatureZ (ZFP), FeatureZ (SPERR), and FeatureZ (HPEZ)

  • The runtime results cover pipeline-stage timings and compression or decompression throughput for FeatureZ variants and other compressors, with pMSz reported only for compression.The text notes pMSz is available only as a GPU implementation.Source: §F.2.1 · p. 18

F.2.2 Error-Bound Versus Compression Time

  • Changing the error bound can raise, lower, or leave compression throughput largely unchanged, depending on the configuration.The comparison covers four base compressors, three datasets, and the preserved features.Source: §F.2.2 · p. 18

F.3.1 Linear- versus Logarithmic-Scaling Quantization

This subsection compares linear- and logarithmic-scaling quantization within FeatureZ. Logarithmic scaling generally yields better rate-distortion performance, with feature- and compressor-specific exceptions.

F.3.1 Linear- versus Logarithmic-Scaling Quantization

  • Logarithmic-scaling quantization generally provides better rate-distortion performance, but the approaches perform nearly identically for Morse–Smale segmentation, local order, and critical-point preservation.Those three features rely solely on star correction.Source: §F.3.1 · p. 19
  • For simplified contour-tree and isosurface preservation, feature-specific bounds let logarithmic-scaling quantization improve compression efficiency and remain competitive with TopoA and QPET.These features depend on bounds rather than star correction.Source: §F.3.1 · p. 19
  • With ZFP, linear-scaling quantization often performs slightly better than logarithmic-scaling quantization, while the two SZ3 variants have generally comparable throughput.The throughput comparison covers compression and decompression.Source: §F.3.1 · p. 19

F.3.2 Contribution of Individual Pipeline Components

The component ablation finds that bound enforcement and star correction have non-additive storage overheads, and that their feature-error effects differ. The parameter study finds no clear relationship between tested parameter values and performance metrics.

F.3.2 Contribution of Individual Pipeline Components

  • For unsimplified contour trees, upper and lower bounds have greater standalone storage overhead than star correction, but their combined bit-rate is substantially below the sum of their individual overheads.Lossless compression consistently reduces final bit-rate, with the reduction varying across datasets.Source: §F.3.2 · p. 19
  • Using only feature-specific bounds can produce higher feature error than the base compressor, whereas star correction yields substantially lower error and lossless compression does not change feature error.The authors attribute the higher error under bounds alone to small jagged artifacts from logarithmic-scaling quantization.Source: §F.3.2 · p. 19

F.4 Varying Feature-Specific Parameters

  • Varying the persistence threshold or number of isovalues shows no clear correlation with compression ratio, compression throughput, or decompression throughput.The experiments use the largest standard error bound for each feature-dataset pair.Source: §F.4 · pp. 19–20

F.5 Errors Introduced by Base Compressors

Unaugmented compressors often leave substantial feature errors, especially for neighborhood-level and global structures. FeatureZ improves rate–distortion performance over TopoSZp, while Neurcomp offers little benefit as its base compressor.

F.5 Errors Introduced by Base Compressors

  • At low bit-rates, base compressors often produce high feature error; neighborhood-level and global structures are generally less preserved than percentiles and isocontours.Feature error decreases rapidly with increasing bit-rate but often remains substantial even at relatively high bit-rates.Source: §F.5 · p. 20
  • ZFP consistently achieves higher bit-rates and correspondingly lower feature error than other base compressors, often reconstructing well below its requested error bound.The comparison adjusts error bounds to achieve compression ratios similar to FeatureZ and omits some ZFP trials when no suitable ratio is available.Source: §F.5 · p. 20

F.6 Augmenting Neurcomp with FeatureZ

  • Neurcomp experiments use a target compression ratio of 100, but actual ratios vary with network-weight compression and FeatureZ’s error bound and preserved feature.The target was selected by an initial grid search.Source: §F.6 · p. 21
  • Neurcomp has the worst rate–distortion performance by a substantial margin among the evaluated bases, and FeatureZ compression throughput never exceeds 0.3 MB/s.Its decompression throughput is typically about an order of magnitude lower than that of other base compressors; INR training likely dominates compression time.Source: §F.6 · p. 21, Table 18 · p. 21, Table 19 · p. 21

F.7 Comparison with TopoSZp

  • The TopoSZp comparison evaluates bit-rate, PSNR, compression time, and decompression time, and also reports feature error because TopoSZp preserves only a subset of critical points.TopoSZp supports only 2D single-precision scalar fields.Source: §F.7 · p. 21
  • Every augmented FeatureZ variant consistently outperforms TopoSZp in rate–distortion performance.The comparison is limited to three additional 2D single-precision datasets; TopoSZp’s feature-error measure is not directly comparable to the paper’s elsewhere.Source: Figure 9 · p. 21, §F.7 · p. 21
  • TopoSZp achieves substantially higher compression and decompression throughput than FeatureZ, while performing much of its topological correction during decompression.Its feature-error values are generally low but vary across datasets and error bounds, and use a different critical-point definition.Source: §F.7 · p. 21, Figure 10 · p. 21, Table 21 · p. 22

G PARAMETER VALUES FOR EXPERIMENTS

The experiments select error bounds to avoid settings that worsen both bit-rate and PSNR, and compare compressors using bounds chosen to be optimal for each. Parameter tables and figures document the tested settings and trade-offs.

G PARAMETER VALUES FOR EXPERIMENTS

  • The appendix lists the isovalues for isocontour preservation and the error bounds used in the main FeatureZ experiments.The bounds are generally shared across tasks, with exceptions specified for particular task–dataset pairs.Source: G PARAMETER VALUES FOR EXPERIMENTS · p. 22, Table 24 · p. 25
  • For feature-preserving compressors, larger error bounds can sometimes decrease bit-rate while improving PSNR, likely because stronger feature corrections increase storage overhead.Such settings are described as suboptimal because a smaller bound yields both lower bit-rate and higher reconstruction quality.Source: G PARAMETER VALUES FOR EXPERIMENTS · p. 22
  • For fair comparisons, the experiments choose bounds that are optimal for each compressor and feature, beginning with five evenly spaced bounds between 1 × 10^-n and 10 × 10^-n.They start at n = 4, increase n when needed to eliminate suboptimal bounds, and omit limited remaining suboptimal configurations for a compressor rather than increasing n globally.Source: G PARAMETER VALUES FOR EXPERIMENTS · p. 22
  • Base-compressor comparisons use bounds selected to produce compression ratios comparable to FeatureZ, with FeatureZ run at the largest bound listed for each task and dataset.The corresponding base-compressor bounds are reported in a table; the teaser image instead uses SZ3 with an error bound of 2.067 × 10^-5.Source: G PARAMETER VALUES FOR EXPERIMENTS · p. 22, Table 23 · p. 24
Loading 2610.12371v1…