RSS

Con­ser­va­tion of Ex­pected Evidence

TagLast edit: 20 Nov 2021 23:16 UTC by Vladimir_Nesov

Conservation of Expected Evidence is a consequence of probability theory which states that for every expectation of evidence, there is an equal and opposite expectation of counterevidence [1]. Conservation of Expected Evidence is about both the direction of the update and its magnitude: a low probability of seeing strong evidence in one direction must be balanced by a high probability of observing weak counterevidence in the other direction [2]. The mere expectation of encountering evidence–before you’ve actually seen it–should not shift your prior beliefs. It also goes by other names, including the law of total expectation and the law of iterated expectations.

A consequence of this principle is that absence of evidence is evidence of absence.

Consider a hypothesis H and evidence (observation) E. Prior probability of the hypothesis is P(H); posterior probability is either P(H|E) or P(H|¬E), depending on whether you observe E or not-E (evidence or counterevidence). The probability of observing E is P(E), and probability of observing not-E is P(¬E). Thus, expected value of the posterior probability of the hypothesis is:

P(H|E) ⋅ P(E) + P(HE) ⋅ PE)

But the prior probability of the hypothesis itself can be trivially broken up the same way:

Thus, expectation of posterior probability is equal to the prior probability.

In other way, if you expect the probability of a hypothesis to change as a result of observing some evidence, the amount of this change if the evidence is positive is

D1 = P(H|E) − P(H).

If the evidence is negative, the change is

D2 = P(HE) − P(H).

Expectation of the change given positive evidence is equal to negated expectation of the change given counterevidence:

D1 ⋅ P(E) =  − D2 ⋅ PE).

If you can anticipate in advance updating your belief in a particular direction, then you should just go ahead and update now. Once you know your destination, you are already there.

Notable Posts

See Also

Con­ser­va­tion of Ex­pected Evidence

Eliezer Yudkowsky13 Aug 2007 15:55 UTC
181 points
80 comments2 min readLW link

Mis­takes with Con­ser­va­tion of Ex­pected Evidence

abramdemski8 Jun 2019 23:07 UTC
173 points
22 comments12 min readLW link1 review

Yes Re­quires the Pos­si­bil­ity of No

Scott Garrabrant17 May 2019 22:39 UTC
180 points
50 comments2 min readLW link2 reviews

Ab­sence of Ev­i­dence Is Ev­i­dence of Absence

Eliezer Yudkowsky12 Aug 2007 20:34 UTC
112 points
112 comments3 min readLW link

Con­ser­va­tion of Ex­pected Jury Probability

jefftk22 Aug 2014 15:25 UTC
15 points
5 comments1 min readLW link

Con­ser­va­tion of ex­pected moral ev­i­dence, clarified

Stuart_Armstrong20 Jun 2014 10:28 UTC
20 points
10 comments2 min readLW link

Rad­i­cal Probabilism

abramdemski18 Aug 2020 21:14 UTC
154 points
47 comments35 min readLW link1 review

Ex­pla­na­tion vs Rationalization

abramdemski22 Feb 2018 23:46 UTC
15 points
11 comments4 min readLW link

Up­date Your­self Incrementally

Eliezer Yudkowsky14 Aug 2007 14:56 UTC
75 points
29 comments3 min readLW link

Co-Proofs

abramdemski21 May 2018 21:10 UTC
39 points
13 comments1 min readLW link

Bayesian Judo

Eliezer Yudkowsky31 Jul 2007 5:53 UTC
79 points
107 comments1 min readLW link

The Prin­ci­ple of Pre­dicted Improvement

Ronny23 Apr 2019 21:21 UTC
66 points
17 comments3 min readLW link