Allow slightly asymmetric square Hessians, and revert to using `HighsHashTable` (mainly) In `assessMatrix` by jajhall · Pull Request #2984 · ERGO-Code/HiGHS

jajhall · 2026-04-20T23:47:30Z

Added a test on the difference between corresponding off-diagonal entries in a square Hessian, allowing them to differ by 1e-10. This should close #2979

Reverted to using HighsHashTable (mainly) In assessMatrix to identify duplicate entries in a matrix column/row, unless a Hessian is being assessed (so duplicate entries are summed) when an unordered_map is used. This closes #2977

…sian asymmetry; modified HighsHessianUtils.cpp to permit it

…ent to std::vector<HighsInt> el_in_vec;

… unordered_map similar for neos; vector 3x slower

…es, when unordered_map is used; silenced unit tests; formatted

codecov · 2026-04-20T23:55:54Z

Codecov Report

❌ Patch coverage is 95.89041% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 81.31%. Comparing base (0a9e7ca) to head (4816a03).
⚠️ Report is 12 commits behind head on latest.

Files with missing lines	Patch %	Lines
highs/model/HighsHessianUtils.cpp	86.95%	3 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##           latest    #2984   +/-   ##
=======================================
  Coverage   81.30%   81.31%           
=======================================
  Files         359      359           
  Lines       88792    88852   +60     
=======================================
+ Hits        72196    72253   +57     
- Misses      16596    16599    +3

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

odow · 2026-04-21T03:29:01Z

  HighsInt num_summation = 0;
  HighsInt num_upper_triangle = 0;
  HighsInt num_hessian_el = 0;
+  const double kSquareHessianAsymmetryTolerance = 1e-10;


Open question whether to hard-code this or make it an option.

The argument in favour of this hard-coding is that anyone wanting to change it needs a really good reason, and they should almost certainly just use the triangular input.

I'd rather hard code it, as you do in JuMP.

If the difference is larger than 1e-10 due to inexact arithmetic, then users should be aware and handling it themselves

odow · 2026-04-21T03:29:18Z

+  }
+  if (dev_run) {
+    tt += getWallTime();
+    printf("Added %d rows individually in %.2gs\n", int(lp.num_row_), tt);


Did you benchmark the before/after or should I do it with the Julia example?

When loading the LP neos row-by-row, It's about 3x faster with the two current ways of tracking duplicates than with 1.14. Using HighsHashTable (as before 1.14) is slightly faster than unordered_map (unsurprisingly) so I left that as the default - unless duplicates need to be summed

highs 1.15.0 Created-by: HarmonybrewBot Commit-by: HarmonybrewBot Merged-by: HarmonybrewBot Description: Created by `brew bump` --- Created with `brew bump-formula-pr`.<details> <summary>release notes</summary> <pre>The highlights of v1.15 are the first variant of the parallel MIP solver and the addition of HiPO in Python. ## Code changes Following [PR #2886](ERGO-Code/HiGHS#2886), our prototype multithreaded MIP solver is available, refactoring the worker/node-search logic of the branch-and-cut solver to run across multiple threads. This is the first release of the parallel MIP solver. A fix to the parallel scheduler was merged in [PR #3087](ERGO-Code/HiGHS#3087). Following [PR #2994](ERGO-Code/HiGHS#2994), HiPO can now be used to solve LPs and QPs from `highspy`, with documentation added in [PR #3024](ERGO-Code/HiGHS#3024) and [PR #3060](ERGO-Code/HiGHS#3060). HiPO performance was further improved by faster handling of free variables ([PR #3013](ERGO-Code/HiGHS#3013)) and faster triangular solves ([PR #3014](ERGO-Code/HiGHS#3014)). Prompted by [#2957](ERGO-Code/HiGHS#2957), square Hessians that are only slightly asymmetric are now accepted by `Highs::qFormatOk`, rather than being rejected outright ([PR #2965](ERGO-Code/HiGHS#2965), [PR #2984](ERGO-Code/HiGHS#2984) and [PR #3003](ERGO-Code/HiGHS#3003)). Identification of non-convexity in the QP solvers was also improved, with extended documentation added in [PR #3056](ERGO-Code/HiGHS#3056), and a further bug affecting some QPs was fixed in [PR #3069](ERGO-Code/HiGHS#3069). QP hot start was added in [PR #3089](ERGO-Code/HiGHS#3089). Following [PR #2982](ERGO-Code/HiGHS#2982), several bugs in the computation of an Irreducible Infeasible Subsystem (IIS) via `Highs::getIis` were fixed. Following [PR #3046](ERGO-Code/HiGHS#3046), zero-cost singleton columns are no longer fixed to an infinite bound during `HPresolve::dualFixing`, and the size of shifts applied during reduced-cost fixing was reduced in [PR #2986](ERGO-Code/HiGHS#2986). Strengthened variable bounds found during presolve are now retained ([PR #3010](ERGO-Code/HiGHS#3010)), a bug in the computation of the fractional value of a variable with finite lower and upper bounds was fixed ([PR #3005](ERGO-Code/HiGHS#3005)), and the handling of infeasibilities for semi-continuous and semi-integer variables was corrected ([PR #3018](ERGO-Code/HiGHS#3018)). Following [PR #3042](ERGO-Code/HiGHS#3042), parallel simplex is no longer used when solving the LP relaxations that arise within the MIP solver, since it isn't safe to do so when the MIP solver is itself running in parallel. Following [PR #2971](ERGO-Code/HiGHS#2971), `highspy` now supports scalar division, and significant improvements were made to `highspy`'s static typing in [PR #2983](ERGO-Code/HiGHS#2983). Context manager support was added to `highspy` in [PR #3093](ERGO-Code/HiGHS#3093). Following [PR #3039](ERGO-Code/HiGHS#3039), `Highs::setBasis` and `Highs::setLogicalBasis` have been implemented, and additional methods were added to the C# wrapper in [PR #3041](ERGO-Code/HiGHS#3041). Prompted by [#3044](ERGO-Code/HiGHS#3044), a further bug was fixed in [PR #3092](ERGO-Code/HiGHS#3092), and [PR #3091](ERGO-Code/HiGHS#3091) corrects the re-checking of an implied bound, fixing [#3090](ERGO-Code/HiGHS#3090). Following [PR #2981](ERGO-Code/HiGHS#2981), C-heap memory leaks in solver state cleanup were fixed, and [PR #3081](ERGO-Code/HiGHS#3081) fixes a bug in column stuffing. Following [PR #3016](ERGO-Code/HiGHS#3016), `Highs::presolve()` now logs and returns failure appropriately when called inconsistently with the state of the incumbent model. Prompted by [#3007](ERGO-Code/HiGHS#3007), a potential typo in `FactorHiGHSSolver::chooseNla()` was corrected, and several other minor fixes were collected in [PR #3017](ERGO-Code/HiGHS#3017) and [PR #3004](ERGO-Code/HiGHS#3004) (the latter for `HighsCliqueTable`). Following [PR #3057](ERGO-Code/HiGHS#3057), lines in HiGHS options files that contain only spaces are now ignored, and the text of the first line containing an error is printed in the resulting message. ## Build changes The Python build has been updated and an additional python package is available. The HiPO dependencies are linked via the optional `highspy-extras`, e.g. Metis, for all platforms, and OpenBLAS, for Windows and Linux. The `highspy-extras` package is automatically consumed by `highspy` and does not need to be imported manually. Note, that `highspy-extras` is distributed under the Apache 2.0 license, due to the dependencies' licensing. On Linux, `libblas` is no longer supported, in favour of OpenBLAS. MKL support would be considered at a later stage. Prompted by [#3000](ERGO-Code/HiGHS#3000), [PR #3027](ERGO-Code/HiGHS#3027) fixes a recent `alpine:edge` compilation issue. Following [PR #3025](ERGO-Code/HiGHS#3025), all remaining thread sanitizer data race warnings are cleared. GPU support was added for local Python pip installs in [PR #3071](ERGO-Code/HiGHS#3071) and the BLAS library used by HiPO is now checked at runtime ([PR #3080](ERGO-Code/HiGHS#3080)). The `cibuildwheel` workflow and `pyproject` configuration were updated to `cibuildwheel` 4.0.0 ([PR #3061](ERGO-Code/HiGHS#3061)), and the NuGet publishing workflow now uses trusted publishing rather than an API key ([PR #3062](ERGO-Code/HiGHS#3062), following [PR #2959](ERGO-Code/HiGHS#2959)). The macOS GitHub Actions build times were significantly improved in [PR #3033](ERGO-Code/HiGHS#3033), and the C# wrapper build was fixed to statically link the MSVC STL and correctly install on win32 ([PR #3070](ERGO-Code/HiGHS#3070)). Following [PR #3001](ERGO-Code/HiGHS#3001), the install location of the readme and license files was corrected, and explicit `-fexceptions` copts were added to the Bazel config ([PR #3072](ERGO-Code/HiGHS#3072) and [PR #3075](ERGO-Code/HiGHS#3075)). </pre> <p>View the full release notes at <a href="https://github.com/ERGO-Code/HiGHS/releases/tag/v1.15.0">https://github.com/ERGO-Code/HiGHS/releases/tag/v1.15.0</a>.</p> </details> <hr> See merge request: Harmonybrew/homebrew-core!12881

highs 1.15.1 Created-by: HarmonybrewBot Commit-by: HarmonybrewBot Merged-by: HarmonybrewBot Description: Created by `brew bump` --- Created with `brew bump-formula-pr`.<details> <summary>release notes</summary> <pre>Patch release v1.15.1 fixes: Apache release binaries not including HiPO fixed in [PR #3112](ERGO-Code/HiGHS#3112). Build for users with C++20 fixed in [PR #3114](ERGO-Code/HiGHS#3114). OpenBLAS setup issue affecting performance fixed in [PR #3122](ERGO-Code/HiGHS#3122). Values passed for string option values (except for file names) are converted to lower case [PR #3108](ERGO-Code/HiGHS#3108). --- The highlights of v1.15 are the first variant of the parallel MIP solver and the addition of HiPO in Python. ## Code changes Following [PR #2886](ERGO-Code/HiGHS#2886), our prototype multithreaded MIP solver is available, refactoring the worker/node-search logic of the branch-and-cut solver to run across multiple threads. This is the first release of the parallel MIP solver. A fix to the parallel scheduler was merged in [PR #3087](ERGO-Code/HiGHS#3087). Following [PR #2994](ERGO-Code/HiGHS#2994), HiPO can now be used to solve LPs and QPs from `highspy`, with documentation added in [PR #3024](ERGO-Code/HiGHS#3024) and [PR #3060](ERGO-Code/HiGHS#3060). HiPO performance was further improved by faster handling of free variables ([PR #3013](ERGO-Code/HiGHS#3013)) and faster triangular solves ([PR #3014](ERGO-Code/HiGHS#3014)). Prompted by [#2957](ERGO-Code/HiGHS#2957), square Hessians that are only slightly asymmetric are now accepted by `Highs::qFormatOk`, rather than being rejected outright ([PR #2965](ERGO-Code/HiGHS#2965), [PR #2984](ERGO-Code/HiGHS#2984) and [PR #3003](ERGO-Code/HiGHS#3003)). Identification of non-convexity in the QP solvers was also improved, with extended documentation added in [PR #3056](ERGO-Code/HiGHS#3056), and a further bug affecting some QPs was fixed in [PR #3069](ERGO-Code/HiGHS#3069). QP hot start was added in [PR #3089](ERGO-Code/HiGHS#3089). Following [PR #2982](ERGO-Code/HiGHS#2982), several bugs in the computation of an Irreducible Infeasible Subsystem (IIS) via `Highs::getIis` were fixed. Following [PR #3046](ERGO-Code/HiGHS#3046), zero-cost singleton columns are no longer fixed to an infinite bound during `HPresolve::dualFixing`, and the size of shifts applied during reduced-cost fixing was reduced in [PR #2986](ERGO-Code/HiGHS#2986). Strengthened variable bounds found during presolve are now retained ([PR #3010](ERGO-Code/HiGHS#3010)), a bug in the computation of the fractional value of a variable with finite lower and upper bounds was fixed ([PR #3005](ERGO-Code/HiGHS#3005)), and the handling of infeasibilities for semi-continuous and semi-integer variables was corrected ([PR #3018](ERGO-Code/HiGHS#3018)). Following [PR #3042](ERGO-Code/HiGHS#3042), parallel simplex is no longer used when solving the LP relaxations that arise within the MIP solver, since it isn't safe to do so when the MIP solver is itself running in parallel. Following [PR #2971](ERGO-Code/HiGHS#2971), `highspy` now supports scalar division, and significant improvements were made to `highspy`'s static typing in [PR #2983](ERGO-Code/HiGHS#2983). Context manager support was added to `highspy` in [PR #3093](ERGO-Code/HiGHS#3093). Following [PR #3039](ERGO-Code/HiGHS#3039), `Highs::setBasis` and `Highs::setLogicalBasis` have been implemented, and additional methods were added to the C# wrapper in [PR #3041](ERGO-Code/HiGHS#3041). Prompted by [#3044](ERGO-Code/HiGHS#3044), a further bug was fixed in [PR #3092](ERGO-Code/HiGHS#3092), and [PR #3091](ERGO-Code/HiGHS#3091) corrects the re-checking of an implied bound, fixing [#3090](ERGO-Code/HiGHS#3090). Following [PR #2981](ERGO-Code/HiGHS#2981), C-heap memory leaks in solver state cleanup were fixed, and [PR #3081](ERGO-Code/HiGHS#3081) fixes a bug in column stuffing. Following [PR #3016](ERGO-Code/HiGHS#3016), `Highs::presolve()` now logs and returns failure appropriately when called inconsistently with the state of the incumbent model. Prompted by [#3007](ERGO-Code/HiGHS#3007), a potential typo in `FactorHiGHSSolver::chooseNla()` was corrected, and several other minor fixes were collected in [PR #3017](ERGO-Code/HiGHS#3017) and [PR #3004](ERGO-Code/HiGHS#3004) (the latter for `HighsCliqueTable`). Following [PR #3057](ERGO-Code/HiGHS#3057), lines in HiGHS options files that contain only spaces are now ignored, and the text of the first line containing an error is printed in the resulting message. ## Build changes The Python build has been updated and an additional python package is available. The HiPO dependencies are linked via the optional `highspy-extras`, e.g. Metis, for all platforms, and OpenBLAS, for Windows and Linux. The `highspy-extras` package is automatically consumed by `highspy` and does not need to be imported manually. Note, that `highspy-extras` is distributed under the Apache 2.0 license, due to the dependencies' licensing. On Linux, `libblas` is no longer supported, in favour of OpenBLAS. MKL support would be considered at a later stage. Prompted by [#3000](ERGO-Code/HiGHS#3000), [PR #3027](ERGO-Code/HiGHS#3027) fixes a recent `alpine:edge` compilation issue. Following [PR #3025](ERGO-Code/HiGHS#3025), all remaining thread sanitizer data race warnings are cleared. GPU support was added for local Python pip installs in [PR #3071](ERGO-Code/HiGHS#3071) and the BLAS library used by HiPO is now checked at runtime ([PR #3080](ERGO-Code/HiGHS#3080)). The `cibuildwheel` workflow and `pyproject` configuration were updated to `cibuildwheel` 4.0.0 ([PR #3061](ERGO-Code/HiGHS#3061)), and the NuGet publishing workflow now uses trusted publishing rather than an API key ([PR #3062](ERGO-Code/HiGHS#3062), following [PR #2959](ERGO-Code/HiGHS#2959)). The macOS GitHub Actions build times were significantly improved in [PR #3033](ERGO-Code/HiGHS#3033), and the C# wrapper build was fixed to statically link the MSVC STL and correctly install on win32 ([PR #3070](ERGO-Code/HiGHS#3070)). Following [PR #3001](ERGO-Code/HiGHS#3001), the install location of the readme and license files was corrected, and explicit `-fexceptions` copts were added to the Bazel config ([PR #3072](ERGO-Code/HiGHS#3072) and [PR #3075](ERGO-Code/HiGHS#3075)). </pre> <p>View the full release notes at <a href="https://github.com/ERGO-Code/HiGHS/releases/tag/v1.15.1">https://github.com/ERGO-Code/HiGHS/releases/tag/v1.15.1</a>.</p> </details> <hr> See merge request: Harmonybrew/homebrew-core!13385

jajhall added 6 commits April 19, 2026 23:51

Added square-near-symmetric-hessian unit test with epsilon square Hes…

5e052e9

…sian asymmetry; modified HighsHessianUtils.cpp to permit it

Now using std::unordered_map<HighsInt, HighsInt> index_set as equival…

adbbfc6

…ent to std::vector<HighsInt> el_in_vec;

Reworked so one or both of el_in_vec and index_set can be used: both OK

0c202f8

unordered_map seems little better than vector

4ec0e8b

Code for HighsHashTable, unordered_map and vector: HighsHashTable and…

25e2bf1

… unordered_map similar for neos; vector 3x slower

Cleaned up assessMatrix: using HighsHashTable unless summing duplicat…

4816a03

…es, when unordered_map is used; silenced unit tests; formatted

jajhall requested a review from odow April 20, 2026 23:47

odow reviewed Apr 21, 2026

View reviewed changes

This was referenced Apr 21, 2026

Native heap corruption in v1.14.0 when passHessian receives a square Hessian with epsilon-level asymmetry (via CVXPY/highspy) #2979

Closed

Performance regression building LP in v1.14.0 #2977

Closed

jajhall merged commit 052d6ab into latest Apr 21, 2026
391 of 392 checks passed

jajhall deleted the fix-2979 branch April 21, 2026 08:55

This was referenced Jun 28, 2026

highs 1.15.0 Homebrew/homebrew-core#290283

Merged

highs 1.15.1 Homebrew/homebrew-core#290997

Merged

This was referenced Oct 2, 2026

Bump Highs.Native from 1.14.0 to 1.15.1 KonkovDV/OpenRebar#100

Closed

Bump Highs.Native from 1.14.0 to 1.15.1 KonkovDV/OpenRebar#141

Closed

Bump Highs.Native from 1.14.0 to 1.15.1 KonkovDV/OpenRebar#144

Closed

This was referenced Oct 11, 2026

Bump Highs.Native from 1.14.0 to 1.15.1 KonkovDV/OpenRebar#160

Open

Bump Highs.Native from 1.14.0 to 1.15.1 KonkovDV/OpenRebar#178

Open

Navigation Menu

Repository navigation

Allow slightly asymmetric square Hessians, and revert to using `HighsHashTable` (mainly) In `assessMatrix` - #2984

Allow slightly asymmetric square Hessians, and revert to using `HighsHashTable` (mainly) In `assessMatrix`#2984
jajhall merged 6 commits into
latestfrom
fix-2979

jajhall commented Apr 20, 2026

Uh oh!

codecov Bot commented Apr 20, 2026 •

edited

Loading

Uh oh!

odow Apr 21, 2026

Uh oh!

jajhall Apr 21, 2026

Uh oh!

odow Apr 21, 2026

Uh oh!

jajhall Apr 21, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

jajhall commented Apr 20, 2026

Uh oh!

codecov Bot commented Apr 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

odow Apr 21, 2026

Choose a reason for hiding this comment

Uh oh!

jajhall Apr 21, 2026

Choose a reason for hiding this comment

Uh oh!

odow Apr 21, 2026

Choose a reason for hiding this comment

Uh oh!

jajhall Apr 21, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

codecov Bot commented Apr 20, 2026 •

edited

Loading