[gvegayon]

Summary

Adds Model::sample_approx(), which draws arrays with Metropolis–Hastings and the tie/no-tie (TNT) proposal. It never computes the support and does not need store_psets(). It is designed as the basis for simulation-based estimation (MCMC-MLE and stochastic approximation) in a later release.

auto state = model.sample_approx_init(y0);              // array + stats + free cells
model.sample_approx(state, theta, 1000u);               // burn-in, nothing recorded
model.sample_approx(state, theta, n, thin, &stats);     // appends n x nterms() stats
Network y = model.sample_approx(y0, theta, nsteps);     // single draw

• Persistent state: SampleApproxState holds the current array, its running raw statistics, the free cells and the current ties. Each call resumes the chain, and θ can change between calls.
• Statistics: updated only through change statistics, never recounted. When stats_out is not null, they are appended every thin proposals, on the scale of get_stats_target() (after any transform).
• TNT step: as in ergm's MH_TNT. With probability ½ a random tie is proposed for removal; otherwise a random free cell is toggled. Acceptance includes the proposal ratio q(y′→y)/q(y→y′). Without it, the chain converges to the wrong distribution.
• Rules: free cells come from the rules. Proposals that fail the dynamic rules are rejected; these rules are checked against the statistics, so rules that inspect the array itself are not supported.
• Method: chosen with SampleApproxMethod (only tnt for now).
• DEFM: new DEFM::simulate_approx(). simulate() now shares the same loop and behaves as before.
• Docs: doc/sample-approx.md, published under Implementation.
• barry.hpp and defm.hpp are regenerated.

Tests

• Exact comparison: chain means match the exact E[s(Y); θ] within 1.4% for sparse and dense θ, Network and NetworkDense, and a model with a nonlinear transform. Removing the proposal correction makes these checks fail (errors of 36–96%).
• Persistence: recounting the array matches the running statistics, and two calls give the same output as one longer call with the same seed.
• Rules and errors: dynamic rules are enforced, and an invalid start array or a wrong params length throws.
• DEFM: simulate_approx reproduces the deterministic torus walk (test 20) and matches the analytic marginal rate.

make test-portable passes (194 assertions in 26 test cases). Valgrind and the MkDocs build were not run locally, because valgrind and Doxygen are not installed.

Out of scope

• MCMC-MLE and stochastic-approximation fitting
• Other sampling methods
• Making conditional_prob handle nonlinear transforms

🤖 Generated with Claude Code

[github-actions]

Version bump check ✅

group	base	head	bump
barry	0.3.0	0.4.0	✅ minor update

_{Checked 1 group(s) against 53002710023a613a6c80709a0337c73bfecf75e8 · please-bump}

[github-actions]

Thank you for your contribution @gvegayon 🚀! Your documentation-site-pr-33 is ready for download 👉 here 👈!
_{(The artifact expires on 2027-01-26T00:21:46Z. You can re-generate it by re-running the workflow here.)}

[codecov]

Codecov Report

❌ Patch coverage is 96.15385% with 5 lines in your changes missing coverage. Please review.
✅ Project coverage is 90.39%. Comparing base (5300271) to head (3f25170).

Files with missing lines	Patch %	Lines
include/barry/model-meat.hpp	95.09%	5 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #33      +/-   ##
==========================================
+ Coverage   89.54%   90.39%   +0.85%     
==========================================
  Files          46       46              
  Lines        3262     3395     +133     
==========================================
+ Hits         2921     3069     +148     
+ Misses        341      326      -15     

☔ View full report in Codecov by Harness.
📢 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.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.