Skip to content

Measuring min-entropy in a physical source

Measuring min-entropy in a physical source means collecting long raw sample sequences under controlled conditions, applying the SP 800-90B estimators, and taking the minimum they return. The result is conditional on the model of the source, so the conditions and assumptions are part of the measurement rather than context for it.

Raw samples, not conditioned output

Measurement happens on raw digitised output, before conditioning. Conditioning is designed to destroy the structure the estimators need, so a measurement taken after it reports close to full entropy regardless of what went in.

This means a source has to expose a raw output path for evaluation. A device that cannot be observed before conditioning cannot be measured by anyone, including its manufacturer — worth asking about when assessing a source you did not build.

The estimator suite

SP 800-90B specifies a family of estimators rather than one. Some assume independent identically distributed samples; others do not and are used when the IID assumption fails, which for a physical source it usually does. The standard takes the minimum across the applicable estimators.

Taking the minimum is deliberately pessimistic. It is the right posture: an entropy estimate that is too high is a security failure, while one that is too low costs throughput.

Conditions are part of the result

A measurement is taken under stated conditions — supply voltage, ambient temperature, device age, drive settings. Those conditions belong in the result, because the estimate does not automatically carry to a device at the corners of its operating envelope.

Serious characterisation therefore measures across the envelope rather than at nominal, which is how you learn which parameters the entropy actually depends on. Those parameters are then the ones worth monitoring in service, and the drift study is where they are identified.

What honest reporting looks like

  • Min-entropy per raw sample, with the estimator that produced the minimum named.
  • Sample count and collection conditions.
  • The model the estimate assumes, and which measurable quantities it depends on.
  • Behaviour across the operating envelope, not only at nominal.
  • What proportion of the signal was attributed to classical noise and excluded.

A datasheet number with none of that is a number quoted rather than produced. Why an estimate is not a guarantee is the conceptual companion to this page.