Rescaled tolerances by maxpaik16 · Pull Request #93 · polyfem/polysolve

maxpaik16 · 2026-02-26T23:38:53Z

Added functions to nonlinear Problem class to allow user-specified norms with appropriate rescaling functions

Updated nonlinear solver and line search logic to use these norms and custom rescalings

update 11/8

update 5/30

codecov · 2026-02-27T02:10:52Z

Codecov Report

❌ Patch coverage is 76.99115% with 26 lines in your changes missing coverage. Please review.
✅ Project coverage is 81.37%. Comparing base (5d62b80) to head (b216524).

Files with missing lines	Patch %	Lines
src/polysolve/nonlinear/Solver.cpp	73.21%	15 Missing ⚠️
src/polysolve/nonlinear/Criteria.cpp	61.11%	7 Missing ⚠️
src/polysolve/linear/AMGCL.hpp	0.00%	1 Missing ⚠️
src/polysolve/linear/EigenSolver.hpp	0.00%	1 Missing ⚠️
src/polysolve/linear/HypreSolver.hpp	0.00%	1 Missing ⚠️
src/polysolve/linear/Solver.hpp	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #93      +/-   ##
==========================================
- Coverage   81.67%   81.37%   -0.31%     
==========================================
  Files          49       49              
  Lines        2025     2083      +58     
  Branches      269      272       +3     
==========================================
+ Hits         1654     1695      +41     
- Misses        371      388      +17

Flag	Coverage Δ
polysolve	`81.37% <76.99%> (-0.31%)`	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ 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.

teseoch · 2026-03-11T00:19:08Z

src/polysolve/nonlinear/Solver.hpp


+        virtual double compute_grad_norm(const Problem &objFunc, const TVector &x, const TVector &grad) const
+        {
+            return objFunc.grad_norm(grad, norm_type_);


are you sure?

I believe so. Is there an issue I am overlooking?

teseoch · 2026-03-11T00:20:43Z

src/polysolve/nonlinear/line_search/LineSearch.cpp

        default_init_step_size = params["line_search"]["default_init_step_size"];
        step_ratio = params["line_search"]["step_ratio"];
+
+        try_interpolating_step = params["line_search"]["try_interpolating_step"];


why adding if not implemented?

teseoch · 2026-03-11T00:22:47Z

src/polysolve/nonlinear/Problem.hpp

+        virtual double step_norm_rescaling(const std::string &norm_type) const {return 1;}
+        virtual double energy_norm_rescaling(const std::string &norm_type) const {return 1;}
+
+        virtual double grad_norm(const TVector &x, const std::string &norm_type) const {return x.norm();}


this is super misleading. x is the solution, now it is the grad. which one is it?

teseoch · 2026-03-11T00:22:56Z

src/polysolve/nonlinear/Problem.hpp

        /// @return True if the solver should stop, false otherwise.
        virtual bool stop(const TVector &x) { return false; }

+        virtual double grad_norm_rescaling(const std::string &norm_type) const {return 1;}


string? enum?

fixed. I use an enum now. It feels a bit awkward as it seems like something that would be more natural to define in the code defining the problem completely (e.g., PolyFEM), but I don't see a better solution at the moment.

teseoch · 2026-03-11T00:24:17Z

src/polysolve/nonlinear/line_search/LineSearch.hpp

+        double grad_norm_rescaling = 1;
+        bool try_interpolating_step;
+        double rel_interpolation_accuracy_tol = 0;
+        std::string norm_type = "Euclidean";


where are these set?

Solver::set_line_search

src/polysolve/nonlinear/Criteria.cpp

teseoch · 2026-03-11T00:26:00Z

src/polysolve/nonlinear/Solver.cpp


        m_descent_strategy = 0;

+        norm_type_ = solver_params["norm_type"];


the style here is m_

nonlinear-solver-spec.json

teseoch · 2026-03-11T00:28:15Z

src/polysolve/nonlinear/descent_strategies/Newton.cpp

        {
            m_logger.debug("[{}] large (or nan) linear solve residual {}>{} (‖∇f‖={})",
-                           name(), residual, residual_tolerance * characteristic_length, grad.norm());
+                           name(), residual, current_residual_tolerance, objFunc.grad_norm(grad, "L2"));


L2 hardcoded

teseoch · 2026-03-11T00:29:04Z

src/polysolve/nonlinear/Problem.hpp

+        virtual double step_norm_rescaling(const std::string &norm_type) const {return 1;}
+        virtual double energy_norm_rescaling(const std::string &norm_type) const {return 1;}
+
+        virtual double grad_norm(const TVector &x, const std::string &norm_type) const {return x.norm();}


also why is the problem resposible for this?

update 3/13

…earch

teseoch · 2026-03-17T16:21:33Z

src/polysolve/linear/AMGCL.hpp

        // Solve the linear system Ax = b
        virtual void solve(const Ref<const VectorXd> b, Ref<VectorXd> x) override;

+        virtual void set_tolerance(const double tol) override {residual_error_ = tol;}


teseoch · 2026-03-17T16:24:04Z

src/polysolve/nonlinear/Problem.hpp

+    enum class NormType
+    {
+        EUCLIDEAN = 0,
+        L2 = 1,


maybe but is this energy norm, right?

teseoch · 2026-03-17T16:25:52Z

src/polysolve/nonlinear/Solver.cpp


+            try
+            {
+                THessian hessian;


This is really expensive, we compute the Hessian twice every time?

and what if there is not hessian?

maxpaik16 added 2 commits November 8, 2024 17:05

Merge pull request #1 from polyfem/main

d7ac83a

update 11/8

Merge pull request #4 from polyfem/main

1a81319

update 5/30

teseoch reviewed Mar 11, 2026

View reviewed changes

maxpaik16 and others added 20 commits March 13, 2026 15:56

Merge pull request #6 from polyfem/main

f6a8b39

update 3/13

implement rescaled and custom norms in nonlinear solve

8faac49

print out all criteria

cce2e67

debugging relative norms

97e7bc5

fix typo

cc1866f

rename tols

f4bf904

use proper norms for linear solve residual tolerance in Newton

2e54add

add new criteria to pre-update-direction-computation exemption

fb0608d

line search printouts

3341c1f

fix scaling for newton dec

a08693f

fix newton dec logic

00467f2

fix flipped sign in newton dec

9eba0de

add option gradient-interpolation-based step size selection to line s…

a867a48

…earch

improve interp check

aa14ba7

rescale interpolation tolerance (hack)

37d35f1

add not yet tested/implemented messages

99ce90c

axe adaptive newton tol for now

dd06af1

add proper norms to line search

4516d2b

pass all tests

a24aa73

rescale grad norm tol in line search

f842fd4

maxpaik16 force-pushed the rescaled-tolerances branch from cbedb97 to f842fd4 Compare March 14, 2026 19:19

maxpaik16 added 4 commits March 14, 2026 15:20

get rid of interpolating step

75d441d

use NormType enum

07690ae

set line search norm type

841d481

put newton decrement in try/catch

b1e9a40

teseoch reviewed Mar 17, 2026

View reviewed changes

only compute newton dec when activated

b216524


		m_descent_strategy = 0;

		norm_type_ = solver_params["norm_type"];

Conversation

maxpaik16 commented Feb 26, 2026

Uh oh!

codecov bot commented Feb 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

codecov bot commented Feb 27, 2026 •

edited

Loading