A smaller record can be sufficient for one question and useless for another. The important issue is not just how much has been discarded, but whether the discarded distinctions change the answer.
Four labels, two questions
Let X take the values 0, 1, 2 and 3. Two possible questions are: is the label odd, and is it in the upper half {2,3}? With equally likely labels, each question has two equally likely answers. Recording the parity gives a perfect answer to the first question and no information about the second. Recording the half reverses those roles.
Keeping the full label answers both questions. Keeping nothing answers neither with certainty. These four encodings merge different sets of inputs. The record is sufficient for a chosen deterministic question precisely when every merged set contains inputs with the same answer.
Choose what the record should preserve
The demonstration allows a biased source: the upper half occurs with probability p, while the two labels inside each half remain equally likely. Change the encoding, the question and p. The cards show the actual groups retained by the chosen encoding.
Record: even
Labels 0, 2 · probability 0.50
Answer: no · conditional probability 1.00
Record: odd
Labels 1, 3 · probability 0.50
Answer: yes · conditional probability 1.00
Best answer accuracy100.0%
Record entropy H(Z)1.000 bits
Uncertainty left H(Y|Z)0.000 bits
The retained record determines the answer. Best accuracy: 100.0%; without the record: 50.0%.
An exact four-state probability model, with no recording noise. Entropy is an average information measure, not the number of hardware bits or an energy cost. The best answer is the most probable answer within each retained group.
A high accuracy can conceal an uninformative record
Set p to 0.9, keep parity and ask whether X is in the upper half. Both parity groups contain 90% upper-half cases. Always answering yes achieves 90% accuracy, but the parity record has not improved the answer at all. The same accuracy was possible before reading it.
For this question the residual uncertainty is h₂(0.9), approximately 0.469 bits. Keeping the half reduces it to zero and gives 100% accuracy. Keeping the full label also gives 100%, but retains distinctions the current question does not require.
The calculation uses P(success) = Σz maxy P(Y=y,Z=z). It assumes equal costs for wrong answers. If missing one kind of event is more costly than another, the best decision rule may change. A record's practical value is tied to a task and a loss function, not only to its entropy.
Today's sufficient summary may fail tomorrow
Suppose parity is all we retain. Later an investigator asks whether X was 0 or 2. That distinction was lost when the group was formed; further deterministic processing of the parity bit cannot restore it. An additional observation or assumption is needed. The image example in What an image records exhibits the same issue through a different physical readout.
For prediction, preserving today's answer is not enough. Two histories can give identical current readings yet different future distributions. A summary becomes predictively sufficient only relative to a specified set of future tests. This stronger question motivates predictive states and retained history.
Why this matters for observable quotients
An observational quotient groups descriptions that the allowed readouts cannot distinguish. Our four encodings are finite examples of that operation. Changing the allowed question changes which quotient is adequate. The example exposes the importance of specifying the observation family before identifying histories.
The published completion-principles paper provides a broader mathematical setting for histories, quotients and subsequent structure. The corresponding book section is a useful entry point. The interactive example establishes only its explicit finite probability calculation; it does not derive a physical state space or a unique fundamental observation rule.
Sources and further reading
- What is information about? introduces entropy, conditional uncertainty and the distinction between information and meaning.
- M. Haj Yousef, Completion principles, real commutants, and positive-frequency geometry for orthogonal Hilbert representations, Open Journal of Mathematical Sciences 10 (2026), 996–1012.

