Skip to content

What is min-entropy, and why an estimate is not a guarantee

Min-entropy measures unpredictability by the most likely outcome: if the single likeliest value has probability p, the min-entropy is minus log2 of p. Standards use it because it is the worst case, and every estimate of it is conditional on a model of the source — which is why an estimate is not a guarantee.

Why the worst case rather than the average

Shannon entropy measures average uncertainty, which is the right tool for compression. For security the average is the wrong question, because an attacker does not guess averagely. They guess the most likely value first.

Min-entropy answers the question an attacker actually poses: how likely is the single best guess? A source that produces one value 25% of the time and spreads the rest evenly may have respectable Shannon entropy and only two bits of min-entropy per sample. The attacker uses the 25%.

What the estimate is conditional on

This is the part that separates a defensible number from a decorative one. Estimating min-entropy requires assumptions: what the underlying distribution looks like, whether samples are independent, whether the physical process is behaving as modelled, and how much of the digitised signal is classical noise rather than the quantum process you are counting.

Change any of those and the number changes. So a min-entropy figure without its model is not a measurement, it is a decoration. When we state a figure, the model it depends on is stated with it — and the methodology used to produce it is described rather than summarised.

Why an estimate is never a guarantee

Two independent reasons, and both matter.

The first is that estimation is inference from finite data. SP 800-90B specifies estimators that deliberately take a conservative view, and a conservative estimate is still an estimate. The second is that the estimate describes the source under the conditions it was measured. A device at a different temperature, at end of life, or under deliberate interference may not match the model any more.

Neither is a reason to distrust the number. Both are reasons to keep measuring, which is what runtime health testing is for and why drift is the failure mode worth designing against.

Using the number correctly

Min-entropy per raw sample is the figure to ask for and to compare. Per output bit is post-conditioning and close to 1 by construction for any competent design, which makes it useless for comparison. The compression ratio between them is where the design actually lives.