Jerzy Neyman

Life and career

For Berkeley students there is a particular pleasure in this bio: Jerzy Neyman is the reason the department you are sitting in exists.

He was born in 1894 in Bendery, then in the Russian Empire, to a Polish Catholic family that had been barred from living in Russian Poland proper as punishment for participation in earlier uprisings. He studied at Kharkov, where he encountered Sergei Bernstein and, through him, the Russian probability tradition descending from Chebyshev and Markov. The years after 1917 were chaotic and dangerous — he was imprisoned briefly during the Polish-Soviet conflict — and in 1921 he was repatriated to Poland under a population exchange. He worked at an agricultural institute in Bydgoszcz, then in Warsaw, on problems of agricultural experimentation, and in 1923 wrote a master's thesis on agricultural trials that contained something remarkable: a potential-outcomes formalization of causal effects, with a randomization-based analysis, decades before Donald Rubin's independent development of the same framework. It was published in Polish and was essentially unknown outside it until translated in 1990.

A fellowship took him to London in 1925 to work in Karl Pearson's laboratory. He found Pearson's mathematics dated and was somewhat lost, but he formed a close friendship and collaboration with Pearson's son **Egon**, and the two began a correspondence and joint program that ran for years across a continent. Between 1928 and 1938 Neyman and Egon Pearson built the framework that bears their names. Neyman was meanwhile in Warsaw, running a small statistical laboratory under difficult conditions, doing consulting work on Polish census and agricultural problems, and in 1934 he presented to the Royal Statistical Society a paper on survey sampling that established stratified random sampling with proper inference as superior to purposive selection — a foundational document for the whole survey-sampling field.

He moved to University College London in 1934. Relations with Fisher, initially cordial, deteriorated into open hostility; Fisher regarded the Neyman-Pearson decision-theoretic reframing of testing as a philosophical corruption of scientific inference, and said so in print for the rest of his life. In 1938 Neyman accepted an offer from Berkeley — a mathematics department with almost no statistics — and emigrated. He built the Statistical Laboratory, then, in 1955, an independent Department of Statistics, and made it arguably the best in the world. He recruited relentlessly and with unusual openness: David Blackwell's appointment, made over the resistance of a system that had excluded him elsewhere, is part of the department's founding history. Neyman also created the Berkeley Symposia on Mathematical Statistics and Probability, which from 1945 through 1970 were the field's central gathering.

He did wartime work on bombing accuracy, and later in life applied statistics to weather modification experiments, carcinogenesis, and astronomy — including work with Elizabeth Scott on the clustering of galaxies. He was an active liberal on questions of academic freedom, opposing California's loyalty oath. He died in Oakland in 1981, having taught nearly to the end.

Key contributions

**The Neyman–Pearson lemma and the testing framework (1933).** Before Neyman and Egon Pearson, a significance test was a somewhat ad hoc device: compute a statistic, see if it is surprising under the null. Their reframing made testing a *decision problem with two kinds of error*. Fix the probability of rejecting a true null — the Type I error rate, or size α. Among all tests of that size, choose the one maximizing the probability of rejecting a false null — the power, or one minus the Type II error rate β. The lemma proves that for a simple null against a simple alternative, the likelihood-ratio test is uniformly most powerful, by a direct exchange argument showing that any deviation from thresholding the likelihood ratio can only cost power at fixed size.

This is the intellectual source of an enormous amount of downstream machinery: the ROC curve and the whole signal-detection framework, matched filters in communications, uniformly most powerful tests, the concept of power itself, and — through the observation that the likelihood ratio is the optimal discriminant — a direct line to modern classification and detection theory. It also introduced the vocabulary of Type I and Type II errors that every applied field now uses.

**Confidence intervals (1937).** Neyman's construction is subtler than students usually appreciate and is worth stating precisely. A confidence interval is a *random set* constructed by a procedure with the property that, whatever the true fixed parameter value, the set covers it with probability at least 1−α under repeated sampling. The probability statement attaches to the procedure, not to the parameter — the parameter is not random, and a realized interval either contains it or does not. Neyman built the general theory by inverting families of hypothesis tests: the confidence set is the collection of parameter values that would not be rejected. This is the frequentist answer to interval estimation and it remains the default idiom of applied science, notwithstanding the near-universal tendency to interpret it Bayesianly.

**Potential outcomes and randomization inference.** The 1923 thesis defines, for each experimental unit, the outcome it would exhibit under each treatment, and takes the causal effect to be a contrast between these — only one of which is ever observed. He derived properties of the randomization-based estimator of the average treatment effect under this model. The framework is the foundation of modern causal inference, and Neyman got there first by a wide margin, in a language no one outside Poland read.

**Survey sampling theory.** His 1934 RSS paper established the theory of stratified sampling with optimal allocation, showed how to obtain valid confidence intervals for population quantities from a probability sample, and argued decisively against purposive selection. Modern official statistics — censuses, labor force surveys, election polling — runs on this.

**Errors in variables, BAN estimators, and applied work.** He contributed the theory of best asymptotically normal estimators, the minimum chi-squared approach, work on contagious distributions for modeling clustered counts (the Neyman Type A distribution for larval counts), and a long series of applied collaborations in astronomy, meteorology, and medicine.

**Berkeley.** The institutional contribution is real and should be counted. The department, the symposia, and the students constitute a permanent change in the field's shape.

In battle

Neyman is a heavyweight: mean 53.0, median 58, twenty-one problems at 80 or above, thirty-six at 70 or above, and only twenty weak. He is the roster's specialist in *guaranteed error rates*, and that specialty turns out to reach further than one might guess.

His best problem is "Signal or just noise?" at 99 — the radar-detection framing that is the Neyman–Pearson lemma in its natural setting. Fix the false-alarm rate, maximize detection probability, and the likelihood-ratio threshold is provably optimal. Around it sit the problems his framework was built for: "How big must the study be?" (96) is power analysis and sample-size determination, which literally does not exist as a question without his Type II error concept. "Does the extra parameter earn its keep?" (91) is the likelihood-ratio test for nested models. "One test or twenty?" (88) is multiple comparisons — the error-rate framing makes the multiplicity problem statable and solvable. "Estimate the tank total" (90) is the German tank problem, estimation with a coverage guarantee. "Where to place the measurements" (90) is optimal design. "When you can't randomize" (90) reaches directly to his 1923 potential-outcomes work. "Counting accidents" (88) is count-data modeling with interval estimates. His category profile is exactly what you would predict: testing 78.1 across sixteen problems — the highest testing average on the roster — experimental design 69.6, small-sample 69.1.

The unexpected number is **fairness at 60.5 across six problems**, and it repays thought. Modern algorithmic-fairness criteria are, structurally, statements about *conditional error rates across groups* — equalized odds is a Type I/Type II error condition, calibration is a coverage condition. A student who understands that fairness auditing is Neyman–Pearson accounting applied to subgroups will find his showing here obvious rather than surprising.

His losses divide into two clean categories. **Discrete computation and program semantics**: "The compiler that beats the coder" (5) is his floor — data-flow analysis and register allocation, where the profile notes that nothing about hypothesis testing bears on whether a program transformation preserves semantics. "The wall around the data structure" (8) is abstract data types and encapsulation; "The optimal codebook" (8) is source coding. His systems average is 6.5 and computability 12.5. **Adversarial and perceptual problems**: "Beat the world champion" (5) is game-playing search; "A hundred robots, no collisions" (8) is multi-agent motion planning; "The picture that isn't there" (8) is image reconstruction. Games 15.0, perception 10.0, search 11.0, RL 18.0.

There is also a weakness worth flagging that is philosophical rather than numeric, and the game's summary states it: the accept/reject framing that gives Neyman his guarantees also flattens scientific nuance into a binary. His war with Fisher was about exactly this, and the replication crisis is in part a story about what happens when a decision procedure designed for industrial acceptance sampling is used as a criterion of scientific truth. Play Neyman when the problem asks for a procedure with a provable error guarantee, a power calculation, a sample size, an interval with coverage, or a detection threshold. Avoid him on computation, adversaries, and any question where the answer needed is a magnitude rather than a verdict.