Result · PROXY-TRUE-RETAINED-FI

Proxy-built labels retain the true score's information, not the proxy's

Derived from the literaturepublication: derived from published

Read the proof and evidence →

Meaning

The true score, conditioned on proxy-derived labels, determines achieved Fisher information; proxy moments measure a different quantity.

Scientific context

Every theorem in this area is a statement about a table of numbers; it is silent about where those numbers came from. When the score is estimated — from a classifier, a learned ratio, a simulator — the information actually retained is governed by the true score conditioned on the estimated rule, not by the estimated score conditioned on itself. The two coincide only when the proxy is exact in the relevant sense; this is why estimated-score provenance never claims exact Fisher semantics.

Statement

If q is built from estimated score s_hat, the actual retained information is Var(E[s|q(s_hat)]), not Var(E[s_hat|q(s_hat)]) unless the proxy is exact in the relevant sense.

Relationships

Rests on

Enables

Raises

Open next door

In the library

Carried by RatioProvenance, ScoreProvenance. Reference (opens in a new tab).

Every object and refusal

Local graph
Local map of True retained Fisher information under a score proxy1 prerequisites to the left, 7 results it enables or questions it raises to the right, 0 boundary cases below.Quantized Fisher-information identityQuantized Fisher-info…Reported versus true retention of a frozen label map under a score error: an exact two-term expansion in the Fisher-whitened score errorReported versus true…Applying a frozen rule to the proxy instead of the truth: mislabel mass under a margin condition and an exact Loewner sandwich on the retained informationApplying a frozen rul…Classifier calibration error to Fisher-information lossClassifier calibratio…Finite-sample uncertainty for retention estimatesFinite-sample uncerta…Perturbation theory for estimated scoresPerturbation theory f…Frozen-rule scalar retention: plug-in CLT with a consistent influence-function varianceFrozen-rule scalar re…Frozen-rule vector retention: plug-in CLT for the geometric-mean D-retention with a consistent influence-function varianceFrozen-rule vector re…True retained Fisher information under a score proxyTrue retained Fisher…
Prerequisites to the left, what it enables or raises to the right, boundary cases below. Open in the map.

Proof

O4. True retained FI under an estimated score

If the quantizer uses s^\hat s, the actual retained Fisher information is

Var⁡(E[s∣q(s^)]),\boxed{ \operatorname{Var}(E[s\mid q(\hat s)]), }

not Var⁡(E[s^∣q(s^)])\operatorname{Var}(E[\hat s\mid q(\hat s)]) unless s^=s\hat s=s in the relevant sense.

From Score, ratio and classifier access · chapter contents