Imagine a sensor producing a restless stream of values. The trace looks irregular, so we call it noise. Now suppose a nearby electrical signal nudges every second sample in a predictable direction. The picture can still look busy while the uncertainty available to a computer is smaller than it appears.
Physical randomness is a story about what remains unpredictable under a stated model. Collecting fluctuations, digitizing them, and producing useful bits are different jobs. Their separation explains why a random-bit generator needs more reasoning than a source that merely looks chaotic.
A balanced count can hide a completely predictable rule
Write 010101010101 on a piece of paper. It has equal numbers of zeros and ones, but the next symbol is easy to guess. Now write six zeros followed by six ones. The total is still balanced. A count has preserved one property while overlooking the order.
This makes a good first task: invent three twelve-bit strings with exactly six ones. Give each a different rule. How many samples would a reader need to recognize that rule? You have constructed examples in which an attractive summary statistic says little about predictability.
NIST's statistical-test publication explicitly distinguishes testing from proof that a generator is appropriate for an application. Statistical tests can reveal problems, but they cannot replace a sound account of the generator and its security properties. A few reassuring charts are not a complete explanation.
Processing can concentrate uncertainty without inventing it
NIST SP 800-90B separates a noise source, optional conditioning, and health testing. Conditioning is deterministic processing that can reduce bias or increase entropy per output bit. It cannot repair a source that has stopped providing uncertainty by creating new entropy from nothing.
A small pair rule makes the distinction tangible. Read consecutive, nonoverlapping input pairs. Discard 00 and 11. Turn 01 into 0, and 10 into 1. The output is shorter, and the discarded pairs make the lost quantity visible.
Watch a pair filter keep and discard bits
Deterministic step table exposes bias/dependence assumptions, discarded pairs, and an unfinished trailing bit. No random numbers or security keys are generated.
Open watch a pair filter keep and discard bits in its own window.
Under a specific mathematical assumption—independent input bits with the same fixed probability p of being 1—the probabilities of 01 and 10 are equal: (1−p)p and p(1−p). Among the retained unequal pairs, the two outcomes are equally likely. That argument relies on the assumption. It does not prove arbitrary physical samples satisfy it.
Use the preset repeating input 010101010101. The pair rule emits only zeros. The input was balanced, yet dependence defeated the informal promise that a filter would make everything fair. Next use 00111001. Its pairs are 00, 11, 10, and 01; only the last two survive, giving 10.
The same assumption also predicts an efficiency cost. Unequal pairs occur with probability 2p(1−p). At p = 0.9, only 18 percent of pairs survive on average, producing one output bit per retained pair. That is an expected output length of 9 percent of the input bit count, not a guaranteed result for a short sequence.
Keep the raw evidence available
The filter demonstration is entirely deterministic: it processes the text you enter and generates no physical randomness. That limitation is the point. A clean-looking output cannot tell you everything about the source that produced the input. Reviewing raw behavior and its physical origin can expose failures that later processing obscures.
Imagine a sensor that is stuck at one value. A monitoring system should notice the failure rather than keep issuing a reassuring status because another part of the device is still running. In a real design, the monitoring rules must suit the source and the failures it can experience.
NIST and BSI's discussion of tailored health tests addresses that source-specific work. The presentation distinguishes observing a source's operation from merely assessing a finished bitstream. Different physical mechanisms can have different ways of going wrong.
Make a claim small enough to defend
Take the sentence “These bits are random.” Replace it with four questions: What physical process supplies uncertainty? How are samples taken? What assumptions support the estimate? What happens when the source fails? Each question asks for a different kind of evidence.
A useful pair-filter challenge
Predict the output before pressing Process. Then count input bits, discarded pairs, retained pairs, and any trailing unpaired bit. Try an odd-length input. The unfinished bit must remain unused rather than silently receiving a fabricated partner. Try all zeros, then the alternating preset. Both produce warnings for different visible reasons.
The most interesting part of a useful random source is often the discipline around it. Engineers do not merely collect disorder. They decide which uncertainty can be credited, what processing preserves it, and how an apparently lively machine will reveal that it has become predictable. The question behind the noise is always: predictable to whom, with which information?
Sources and further exploration: NIST SP 800-22: statistical testing and its limits; NIST SP 800-90B: entropy-source design; NIST and BSI: tailored health tests.
Explore all 50 new articles · Browse the complete illustrated collection
Explore more art, ideas, and independent projects at Incoprea.
