The semantics of deletion

“Remove component $i$” means different things in different fields — and the two cavity identities answer different removal questions.

The algebraic similarity between the multiplicative cavity (divide out a survival function) and the rank-one Schur cavity (subtract a mediated interaction) is what started this project. But similarity of formulas can obscure the physical question: the interventions they price are not the same. This page fixes the vocabulary used across the site.

Marginal vs conditional

Let the latent performances be jointly Gaussian, $X \sim N(\mu, \Sigma)$.

InterventionSurviving distributionWhich cavity
Scratch competitor $i$ the marginal: $X_{-i} \sim N(\mu_{-i},\, \Sigma_{-i,-i})$ — the covariance is simply the submatrix multiplicative (races)
Pin / condition $X_i = c$ the conditional: $\Sigma_{-i,-i} - \Sigma_{-i,i}\,\Sigma_{ii}^{-1}\,\Sigma_{i,-i}$ Schur / rank-one
Delete a coordinate of the precision $A = \Sigma^{-1}$ $\left(A_{-i,-i}\right)^{-1}$ — equal to the conditional covariance above Schur / rank-one

Scratching uses a marginal; Schur uses a conditional. A correlated race does not automatically invoke a Schur complement when a competitor is scratched — the field simply loses a coordinate, correlations among survivors unchanged. The Schur complement enters when the intervention pins, conditions, or re-equilibrates the field. The fast correlated transform combines the identities in the legitimate way: conditioning on latent factors (Schur side), then scratching within the conditionally independent race (multiplicative side).

The zoo of removals

Each row of this list is a different physical question with a different cheap identity (submatrix, Schur complement, Woodbury, block downdate) or none at all. Claims on this site name the intervention explicitly.

Back to the introduction, or on to the research program, where the flagship question is now the shared-state race whose homogenized limit is softmax.