Order prefixes: the full block and position sums #
The Alice question marginal μ_X.
Equations
- CommutingRepetition.TracialStrategy.margX μ x = ∑ y : Y, μ x y
Instances For
The Bob-block conditional bound (eqs bob-block-conditioning-budget, #
bob-block-chain-rule)
The total prior mass over core tuples is the number of core words.
The block conditional bound: at a fixed Alice set SX and a Bob
background SY covering the rest, the core-posterior expectation of the
log-ratio of the class masses at the full Bob word against the background,
minus the product-prior conditional, is at most t₀ + s₀ — the chain rule
D(ℚ⁰_{class} ‖ P_{class}) − D(ℚ⁰_{background} ‖ P_{background}), the
first term bounded pointwise by log(1/p) and the second by Gibbs.
The second chain term (Alice side): the Bob reverse experiment #
The integrand of the second chain term at datum r:
log ℚ⁰(y_i ∣ X_{{i}∪C_X}, Y_{C_Y}, Z) − log μ(y_i ∣ x_i).
Equations
- One or more equations did not get rendered due to their size.
Instances For
Per-base telescoped bound (eq bob-block-chain-rule): at a fixed Bob
base the cut sum of the conditional log-ratios telescopes to the full-block
conditional, bounded by t₀ + s₀.
The second chain term is at most 2(t₀ + s₀)/m (eqs
bob-block-conditioning-budget through second-history-chain-term): the reveal
datum re-read as (Bob base, interior cut), the size-biased law (2/m)·β, the
cut sum telescoped per base.
The flattened laws: histories, live marginals, and the first chain #
term
ℚ(h, x) = ∑_y ℚ(h, x, y).
Instances For
ℚ(i, X_i = x) = ∑_{h : h.i = i} ℚ(h, x).
Equations
Instances For
Live-coordinate mass vanishes on the core.
Sums over the flattened space grouped by the live coordinate and the live Alice question.
The Alice question marginal of the core posterior.
Instances For
ℚ⁰_X ≤ p⁻¹ · μ_X^{⊗n} (the conditioning budget for the question
marginal).
The core marginal at a coordinate, as a coordinate marginal of ℚ⁰_X.
The first chain term is at most t₀/m (eq first-history-chain-term):
the live marginal is m⁻¹ times the core question marginal, and the
marginals tensorize against μ_X.
The second chain term in core form: the ℚ-expectation of
log(ℚ(h,x,y) μ_X(x) / (ℚ(h,x) μ(x,y))) is the reveal-law average of the
core integrand histLogA.
The Alice conjunct: absolute continuity, subnormalization, log split #
A vanishing live-question factor kills the flattened posterior.
J_A is subnormalized: its total mass is m⁻¹ ∑_x μ_X(x) · #{i ∉ D : ℚ(i, x) > 0} ≤ 1.
The pointwise log split (eq JA-chain-rule): log(ℚ/J_A) is the
conditional live-answer term plus the live-question term.
The log-sum form of the Alice conjunct: the ℚ-weighted log-ratio
against the defaultless J_A is at most (3t₀ + 2s₀)/m. Consumed at assembly
(node 1.2.11) where the sampler's law dominates J_A up to a rounding factor.
The Alice conjunct of history_relative_entropy:
D(ℚ ‖ J_A) ≤ (3t₀ + 2s₀)/m.