Skip to content

QPA: finite module algebra

QPA, short for Quivers and Path Algebras, is a GAP package for computing with vector spaces and maps arranged along a directed graph, together with relations between compositions. Its tools extend from module maps to complexes and resolutions, which express a module through simpler modules. This breadth makes QPA a natural comparison for the algebra we perform on a finite encoding: both programs can answer questions about its maps and extension classes.

Results

TamerOp was 135.7× as fast across the 200 completed comparisons, measuring the time to build each program's input objects and compute the complete answer. The weighted ratio has an approximate 95% interval of 134.6–136.8×. QPA's other 20 requests failed while evaluating elementary tensors, so those requests have no speed ratio. TamerOp completed them.

The size of the difference depends strongly on the operation. A Hom basis, which describes all compatible maps between two modules, has a 3.60× aggregate advantage for TamerOp. Resolutions and operations on extension classes show larger differences. Across all completed requests, the smallest measured ratio was 1.76×; every round of measurements favored TamerOp.

QPA time divided by TamerOp time for the completed comparison and nine kinds of mathematical request. All ratios exceed one; Hom bases have the smallest aggregate difference.

The horizontal axis is QPA time / TamerOp time: a value of 10 means that TamerOp took one tenth of QPA's time for that comparison. Points summarize paired measurements, and the bars show approximate 95% intervals calculated from five rounds of measurements. The logarithmic scale keeps the smaller and larger ratios visible together.

Requested answerCompleted pairsQPA time / TamerOp time
Complete Hom basis as module maps123.60×
Dimension of Ext¹12292.9×
Complete Hom maps, Ext representatives and their coordinates12308.1×
Kernel, image and cokernel, with their structural maps2078.4×
Projective or injective resolution, with its maps40739.1×
Pushout or pullback, with canonical maps4023.5×
Homology modules or a mapping cone, with maps40132.9×
Yoneda product table: Ext¹ × Ext¹ → Ext²12601.7×
Induced Ext maps in degrees 1 and 2121388×
Balanced tensor product and elementary-tensor coordinates0 of 20QPA evaluation errors

Ext describes extension classes, and the Yoneda product composes such classes; the product request returns their coordinates in a basis of Ext². These requests compare actual maps, representatives and coordinates wherever they are part of the requested answer. The request for the dimension of Ext¹ is listed separately because it asks for less output.

Actual runtimes and input sizes

For many of these inputs, TamerOp's complete request takes around a millisecond. QPA's median time ranges from less than a millisecond for a Hom basis to about 2.34 seconds for induced Ext maps. The following plot gives the median case time in eight request groups. Resolutions, pushouts/pullbacks and homology/mapping cones are grouped here and separated in the table below; the table includes the combined Hom/Ext and tensor requests as well.

Median complete-request times for eight kinds of finite module computation, with TamerOp in blue and QPA in orange.

All table times are milliseconds (1,000 ms = 1 second). Each entry gives the median across the indicated inputs, followed by their minimum and maximum in parentheses. For each input, we take the middle time from each round of measurements, then the middle of those five times. These ranges describe different inputs. They are not confidence intervals. “Input dimension sum” adds the vector space dimensions in every supplied module record; several modules contribute to a map or diagram request.

RequestInputsInput dimension sumTamerOp ms: median (min–max)QPA ms: median (min–max)
Hom basis1212–740.391 (0.143–2.266)0.863 (0.570–7.239)
Ext¹ dimension1212–810.757 (0.542–2.154)164.583 (56.066–1,335.089)
Complete Hom/Ext122–810.962 (0.653–3.384)240.087 (78.012–2,180.459)
Kernel, image, cokernel2017–990.803 (0.413–2.610)30.026 (13.452–441.323)
Projective resolution201–370.694 (0.465–1.602)101.116 (78.825–2,166.712)
Injective resolution201–421.008 (0.621–2.366)2,123.944 (171.368–6,637.038)
Pushout203–1430.525 (0.240–1.935)20.953 (3.603–429.239)
Pullback2019–1230.509 (0.345–1.253)4.414 (2.255–7.638)
Homology204–1801.085 (0.526–3.620)445.181 (19.458–16,045.105)
Mapping cone2022–2530.806 (0.552–2.211)15.550 (7.547–41.173)
Yoneda products128–201.269 (0.883–3.421)776.278 (129.790–11,717.391)
Induced Ext maps1220–561.392 (1.147–2.150)2,339.180 (338.540–16,328.116)
Balanced tensor and evaluations2025–731.109 (0.631–1.646)No verified completion

The ratios compare the same input in each paired pass before aggregation. Dividing the two medians in this table would instead compare summaries across different inputs, so it need not reproduce the reported ratio. The complete CSV table gives every input's times, dimensions and field.

What the comparison computes

A module on a finite poset assigns a vector space to each label and a linear map to each order relation. QPA expresses the same data using a quiver, a directed graph, whose relations require the same compositions. Both programs therefore receive the same mathematical problem while using their own data structures and algorithms. The QPA manual describes its module, map and resolution operations.

The inputs use the Boolean three-cube, a 4×4 grid and capped crowns, with 8, 16, 9 or 17 poset vertices. Modules include intervals, projective and injective controls, nonsplit extensions, and modules specified by projective presentations. Coefficients are exact rational numbers or arithmetic modulo the primes 2, 3 and 101. Changing the coefficient field can change the answer, so correctness is checked in each field.

For the complete Hom/Ext and resolution requests, the maximum degree is three on the Boolean cube and two on the grid and crowns. Product requests compute all products of the chosen Ext bases. The tensor request computes the quotient by the balancing relations and the coordinates of every elementary tensor; this is where QPA's evaluations failed.

The comparison starts from finite module data. Constructing a finite encoding from geometry is a separate computation, and the derived answers here belong to the specified finite poset category. The accounts of finite encodings and mathematical categories explain that relationship.

Answers are checked using independent exact linear algebra. Hom checks establish that the returned maps span the whole solution space. Resolution checks verify the differentials and exactness; diagram checks verify the defining commuting maps and universal properties. Independent bar resolutions put Ext classes, products and induced maps into a common model where their coordinates can be compared. Tensor checks verify the balancing relations. These checks establish the requested mathematical answers without requiring the tools to choose identical bases.

How the measurements were taken

The main timer starts with the finite category and the matrix entries already prepared. It measures building each program's modules and computing the requested answer, including its maps and coordinates. Loading and compiling the software, reading the input files, checking the answer and writing it to disk happen outside the timer.

Separate measurements locate the difference between construction and query:

PhaseQPA time / TamerOp time
Building the input objects3.94×
Complete query201.6×
Construction plus query, measured together135.7×

All three rows use the same 200 completed comparisons. Construction alone can favor QPA on some inputs; the complete query accounts for most of TamerOp's advantage. The combined time is measured directly, rather than obtained by adding independently measured medians.

We ran five rounds of measurements, taking three timings for each program, input and phase in every round. A preliminary run compiles the code before timing begins. Each measured request then computes its answer afresh, and the programs take turns going first between rounds.

The original weighting gives equal weight to five mathematical areas, then to request variants, fields and cases within them. Completed pairs retain 90% of that original weight. Their geometric mean gives the 135.7× result; giving each completed case equal weight instead gives 129.1×. The 95% interval uses variation between the five paired passes on this fixed suite and computer. QPA's unsuccessful tensor requests remain in the completion record and never receive a numerical time ratio. The measurement methods explain the aggregation and missing-result conventions.

Computer, versions and memory

The study ran on the benchmark computer on 3 October 2026. Each program used one computation thread, with one program running at a time.

ComponentRecorded version or setting
TamerOp snapshotqpa-v1-529681519b5d1b40
Julia1.12.1
GAP / QPA4.16.1 / 1.37

TamerOp's faster computations used more RAM: the largest recorded amounts were 1.57 GB for TamerOp and 145 MB for QPA, measured for each program as a whole. The data dictionary explains the memory measurements, and the software record gives the full versions and build settings.

Reproduce the results

The reproduction recipe walks through checking the downloads and recomputing the reported comparisons. Start with the summary JSON, whose SHA-256 is:

3bfe85fd092d11c176b1b0961b1f5aac127028378c061f81131c11e2764cf302
DownloadWhat it contains
Per-case CSVTimes, ratios, input sizes, fields and weights
Per-pass observationsThe paired process medians used in the aggregates
Provenance JSONMachine, software and source identities
SHA-256 manifestChecksums of the original data and figures
Data dictionaryUnits, phases and field meanings
Figure record and checksumsSources and integrity of the figures displayed here

The QPA study provides the frozen TamerOp source, exact module inputs and original algebra workers. The common recipe walks through checking the evidence, preparing the tools and collecting a new measurement on your computer.