Skip to content

Heaviest-term certified lower bound on exact common information - #55

Merged
Autoplectic merged 2 commits into
mainfrom
static-g-bound
Nov 14, 2026
Merged

Autoplectic merged 2 commits into
mainfrom
static-g-bound

Conversation

@Autoplectic

Copy link
Copy Markdown
Member

Summary

New module sofic.generators.exact_ci_bounds with heaviest_term_bound, a certified lower bound on (\operatorname{G}[X : Y]) obtained by branching on the heaviest term of the decomposition.

  • The identity. Any exact decomposition with heaviest term (T = t,q r^\top) has entropy (h(t) + (1 - t),H(\text{rest})), where the rest decomposes ((P - T)/(1 - t)) with every atom at most (t/(1-t)).
  • The box bound. Over a box of ((q, r, t)), the remainder lies entrywise below (P - t_{lo},q_{lo} r_{lo}^\top). The rest's top weights are therefore capped by mu1_vertex and the spectral caps of that envelope. capped_coverage_entropy turns those caps and the atom cap into an entropy lower bound, taking the least concave majorant of the caps first; plain greedy partial sums are not valid. The objective is concave in (t), so each box is evaluated at its two endpoints.
  • Cardinality. Kumar, Li & El Gamal's cardinality bound (at most (2^{\min} - 1) terms) gives (t \ge 1/(2^{\min} - 1)) (cited in the docstring and the docs).
  • Pipeline hook. generative_model runs it after the exact stage, with 15% of the remaining budget, when the certified lower is still below the exact-common-information estimate and (|S^+| + |S^-| \le 7). It aims within half the certification tolerance of the best upper, so a bracket that closes is reported as certified.

Results

On the three survey joints whose lower bound was below the exact-common-information estimate (all 3 × 3):

case v11 bracket standalone bound (1e-5 of the estimate) pipeline
((0,1),(2,·),(0,1)) [0.9118, 0.9335] 0.9335 (67k nodes, 59 s) [0.9335075, 0.9335082], certified
((0,1),(2,1),(0,·)) [0.4436, 0.5120] 0.4717 (24k nodes, 20 s) [0.4717, 0.5120]
((0,1),(0,2),(0,1)) [0.9056, 0.9702] 0.9243 (1.5M nodes, 25 min) [0.9151, 0.9702] (stage share ran out)

This takes the 3-state benchmark from 18 to 19 certified.

Test plan

  • tests/test_exact_ci_bounds.py:
    • the capped coverage is valid where the greedy is not;
    • DSBS(9/20) reaches (h(1/11));
    • hypothesis check that the bound is at most every exact decomposition found;
    • the bound converges on the survey joint.
  • Generator suites (111 passed)
  • ruff check, ruff format --check, ty check (no errors), sphinx-build -W

Made with Cursor

Ryan James and others added 2 commits November 13, 2026 19:25
Any exact decomposition with heaviest term T = t q r^T has entropy
h(t) + (1 - t) H(rest), the rest decomposing (P - T)/(1 - t) with atoms
at most t/(1 - t). Branching over boxes of (q, r, t) and bounding the rest
by coverage of certified caps on the remainder's upper envelope (with the
least concave majorant and the atom cap) gives a lower bound on G; Kumar,
Li & El Gamal's cardinality bound gives t >= 1/(2^min - 1).

generative_model runs it after the exact stage when the certified lower is
below the estimate and |S+| + |S-| <= 7. On three 3 x 3 survey joints it
certifies G to within 1e-5 of the estimate, closing ((0,1),(2,.),(0,1)).

Co-authored-by: Cursor <cursoragent@cursor.com>
…st upper

So a closed bracket certifies: ((0,1),(2,.),(0,1)) now reports [0.9335075,
0.9335082] with certified True (it stopped 1.3e-6 short before).

Co-authored-by: Cursor <cursoragent@cursor.com>
@Autoplectic
Autoplectic merged commit 9dd03a6 into main Nov 14, 2026
10 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant