Section 12 pasting: H-A consistency #
The vertical-line to point-consistency transport and the completed-measurement statement of
cor:h-a-consistency: the counterpart of
MIPRE/Background/LIDT/MIPStarRE/LDT/Pasting/ComparisonLemmas/HAConsistency.lean in the port of
planning/c6b-plan.md (milestone M11, section "Port conventions").
A strategy is a SymStrat params.next 𝔓 K, a slice family an IdxPolyFamily params 𝔓 and a
candidate polynomial submeasurement H : SubMeas (Polynomial params.next) 𝔓; the consistency
statements are ConsRels on strategy.state, and the completeness hypothesis is
strategy.state.CompletenessAtLeast (H.liftLeft strategy.state) …. Three uses of the vendored
strategy fields become theorems of the model:
- the bounds of the two failure probabilities by
1(vendored throughstrategy.isNormalized) areS.bipartiteConsError_uniform_le_oneandPasting.bipartiteSSCError_uniform_le_one; - the swap of the consistency relation (vendored through
strategy.densityFixed) isS.consRel_symm_of_density_fixed, and the triangle stepPreliminaries.triangleSub_righttakes no normalization argument; - the residual mass
ev(R(1 − H.total)) = 1 − ev(L H.total)(vendored throughstrategy.permInvState.swap_ev, an entrywise Kronecker identity andev_one_of_isNormalized) isS.ev_L_eq_ev_R,S.leftTensor_subandS.ev_one_of_isNormalized. No statement carries a swap or normalization hypothesis, the vendored statements having none.
Not ported #
Nothing: every vendored declaration has a counterpart of the same name.
References #
In LionSR/MIPStarRE at commit 507e8122, not in this repository:
references/ldt-paper/ld-pasting.texblueprint/src/chapter/ch09_pasting.tex
Convert source-style vertical-line consistency to point consistency using only the axis-parallel and self-consistency estimates.
This is the main estimate in cor:h-a-consistency, stated without the
intermediate HBConsistencyStatement type. It takes only the line-consistency estimate
for a candidate polynomial submeasurement H, restricts that estimate to the
point on each vertical line, and then applies the good-strategy
point-to-vertical-line comparison. The diagonal-line estimate in
strategy.IsGood is not used in this transport step; it enters earlier in the
construction of the line-consistency estimate.
Paper reference: cor:h-a-consistency proof in ld-pasting.tex
lines 1098–1117.
Steps:
- Restrict the line-consistency hypothesis to a single point on the line
- Apply
triangleSubwith theA-Bconsistency bound fromhgood - Error bound:
ν₆ + √(8mε + 4δ) ≤ 47k²m(...) ≤ 100k²m(...).
The completion and large-k hypotheses are carried by the downstream
completed-measurement theorem; this submeasurement argument only uses the positive
k regime and the displayed line-consistency estimate.
Convert source-style vertical-line consistency to point consistency.
This source-style restatement specializes
hAConsistency_submeas_from_lineConsistency_of_axis_self to the two estimates
contained in strategy.IsGood.
Specialization of hAConsistency_submeas_from_lineConsistency to the
constructed pasted submeasurement.
Internal form of cor:h-a-consistency from the one-point sandwich estimates.
This theorem separates the genuinely earlier pasting input from the diagonal
test. Once the estimates of lem:ld-sandwich-line-one-point are known for all
positions, the passage from H-B consistency to H-A consistency uses only the
axis-parallel and self-consistency estimates of the ambient strategy.
Internal form of cor:h-a-consistency from cor:G-hat-facts.
This is the same proof as
hAConsistency_submeas_ofLinePointBounds_of_axis_self, with the one-point
sandwich estimates constructed from GHatFactsStatement.
Internal form of cor:h-a-consistency from the Section 11 commutativity
conclusion.
This exposes the precise upstream mathematical input needed for the G-hat
construction. The diagonal-line estimate is not used in the H-B to H-A
transport; it is used only insofar as it is needed to prove the commutativity
conclusion supplied here.
cor:h-a-consistency.
This is the point-consistency part of the pasted-submeasurement chain. The
completed-measurement consistency is deliberately separated as
hAConsistency_completed, since the paper proves it only after
cor:ld-pasting-N-completeness.
Complete a polynomial submeasurement after its point consistency and mass lower bound have been proved.
This is the completion step in cor:h-a-consistency, stated for an arbitrary
submeasurement H. The source argument first proves point consistency for a
submeasurement and then completes it by adding the missing mass to a fixed
fallback polynomial.
Completed-measurement version of cor:h-a-consistency.
This theorem is intentionally downstream of cor:ld-pasting-N-completeness:
it may use the submeasurement consistency together with the completeness bound
for the constructed pasted submeasurement to control the added completion mass.