dev-notes: v0.10 roadmap for covariance and information

[devin-griff]

Rendered (formatted) view: https://github.com/devin-griff/pounce/blob/covariance-information-roadmap/dev-notes/covariance-information-roadmap.md

A dev-notes design proposal for the pyomo-pounce post-solve second-order surface, targeting v0.10. Companion to #255 (the active-set sensitivity roadmap): that note extends estimate(), this one extends covariance().

What it proposes

All additive to the v0.9 covariance() surface, nothing changes an existing signature:

• information(), the un-inverted sibling of covariance(): the reduced Hessian formed directly from the held factor, avoiding the invert-then-reinvert round trip, in the same natural-units convention as the core solve_with_sens.
• wrt= block selection on covariance() and information(): reduce onto any free-variable block post-solve off one retained factor, so a single solve serves multiple blocks. declare_fitted becomes the default block when wrt= is omitted, preserving v0.9 behavior.
• retain_kkt(), a factor-retention switch for the declaration-free case.

Motivation

The moving-horizon-estimation information-form arrival cost needs the Lagrangian reduced Hessian marginalized onto the arrival state, plus the parameter covariance from the same solve. The note works through the marginal-versus-conditional semantics and the one-solve-two-blocks flow.

It also acknowledges that pounce already exposes the reduced Hessian in its core (solve_with_sens), convex-QP (QpSensitivity), and curve_fit interfaces; this brings that to the pyomo-pounce interface.

Docs-only (one dev note). Opening for design discussion before any implementation.

🤖 Generated with Claude Code

[devin-griff]

Two new items at the front. The previous version delegated bound handling to covariance()'s convention in a single line. That convention classifies on slack alone, and embeds a pinned parameter as a zero row, which is right for a variance and wrong for an information matrix, where the same zeros claim the data says nothing.

Item 0 is an activity classifier keyed on barrier curvature over objective curvature, written as pseudocode, and is Rust core work rather than pyomo surface. Item 1 is the dispositions table.

Items 1 to 3 are now 2 to 4, largely intact.

The rest has been broadly revised for clarity, with a new Preconditions section and constraint rows for #362.

224 to 354 lines.

[devin-griff]

Refreshed. A read-only pass over the solver internals ahead of implementing item 0 found the converged state already retains everything the classifier needs — bound multipliers, slacks, curr_mu, curr_sigma_x, and the exact Lagrangian Hessian — so H is never recovered by subtraction and the earlier claim that the pinned block loses precision as μ tightens is retired: S forms densely from the exact Hessian at full precision. The unidentified floor is pinned, and a review pass tightened items 0 and 1 (warning scope, row handling in the guard, notation stated forward).

Implementation of item 0 starts now against this text; the PR will reference back here, and anything implementation teaches lands as one more refresh.