Peter Bickel

Life and career

Peter Bickel was born in Bucharest in 1940, and his early childhood coincided with the worst years in Europe to be a Jewish child in Romania. His family eventually reached the United States and settled in California. He entered the University of California, Berkeley, took his bachelor's degree there, and completed his PhD in statistics in 1963 under Erich Lehmann — the central figure of Berkeley's postwar statistics department and the author of the books on testing and point estimation that defined mathematical statistics for decades. Bickel joined the Berkeley faculty and, apart from visiting positions, stayed for a career now spanning six decades. He has served as department chair and as president of both the Institute of Mathematical Statistics and the Bernoulli Society, and he received a MacArthur Fellowship in the mid-1980s and election to the National Academy of Sciences.

What makes Bickel unusual among great mathematical statisticians is that he kept moving. Many of his contemporaries found a theoretical home in the 1970s and stayed in it. Bickel's career reads instead as a sequence of arrivals at whatever the field's hard frontier happened to be: robustness and adaptive estimation in the 1970s, the asymptotic theory of the bootstrap around 1980, semiparametric efficiency through the 1980s and early 1990s, hidden Markov models, then — well past the age at which most theoreticians consolidate — high-dimensional covariance estimation, network models, and computational genomics in the 2000s and 2010s, including sustained involvement in the ENCODE and modENCODE genome-annotation consortia.

He is also, by reputation, one of the great mentors of the discipline; a substantial fraction of the theoretical statisticians working on high-dimensional problems today trace an advising line back to him. His textbook with Kjell Doksum, *Mathematical Statistics: Basic Ideas and Selected Topics*, has trained generations of PhD students, and the 1993 monograph *Efficient and Adaptive Estimation for Semiparametric Models*, with Klaassen, Ritov, and Wellner, is the standard reference for the theory he did most to build.

One paper from 1975 has given him a kind of fame outside statistics entirely. With Eugene Hammel and J. William O'Connell, he published "Sex Bias in Graduate Admissions: Data from Berkeley" in *Science*, analyzing his own campus's admissions numbers. That paper is now the canonical real-world instance of Simpson's paradox, and it is taught in more introductory courses than all his theoretical work combined.

Key contributions

**Robust and adaptive estimation.** Bickel's early work sits in the robustness program of Huber, Tukey, and Hampel: build estimators that remain efficient under a nominal model but do not collapse when the model is slightly wrong. His contributions include one-step estimators — take a robust preliminary estimate, apply a single Newton–Raphson correction toward the likelihood, and inherit full asymptotic efficiency without the instability of full maximization — and, more ambitiously, the theory of **adaptive estimation**. The question there is startling: in a symmetric location model where the error density f is *unknown*, can you estimate the location parameter as efficiently as if you knew f? Bickel's 1982 *Annals* paper showed that, remarkably, you can — the unknown nuisance function costs nothing asymptotically. That result, and the conditions under which adaptation is and is not possible, opened the door to the general theory.

**Semiparametric efficiency.** This is the body of work he is most identified with, and it answers a question every applied statistician implicitly faces. Suppose your model has a finite-dimensional parameter θ you care about and an infinite-dimensional nuisance component η you do not. How well can θ possibly be estimated, given that η is unrestricted? The answer is geometric. Consider all one-dimensional parametric submodels passing through the truth; their score functions span a *tangent space* in L². The **efficient influence function** is the projection of the ordinary score for θ onto the orthogonal complement of the nuisance tangent space, and its variance is the **semiparametric efficiency bound** — a Cramér–Rao analogue for infinite-dimensional models. An estimator achieving it is semiparametrically efficient. This framework tells you exactly what price a model's flexibility costs in precision, and it is the theoretical engine underneath modern doubly-robust and targeted-learning estimators for causal effects: the AIPW estimator is not a clever trick, it is the efficient influence function for the average treatment effect written down.

**Bootstrap theory.** Efron proposed the bootstrap in 1979; Bickel and David Freedman, in 1981, proved when it works. Their paper established consistency of the bootstrap for a broad class of smooth statistics via weak-convergence arguments, and — equally important — exhibited cases where it *fails*: extreme order statistics, parameters on a boundary, U-statistics in degenerate cases. The remedies, including the m-out-of-n bootstrap and subsampling, grow directly from that analysis. Students should know this pairing: Efron gave the method, Bickel and Freedman gave the conditions.

**High-dimensional covariance estimation.** With Elizaveta Levina in 2008, Bickel produced two of the most cited papers in modern high-dimensional statistics: regularization of large covariance matrices by **banding** (when variables have a natural ordering, shrink entries far from the diagonal to zero) and by **thresholding** (when they do not, zero out small entries). They proved consistency in operator norm under sparsity assumptions, with explicit rates depending on log p / n — establishing that covariance estimation is tractable when p ≫ n *provided* you can state a structural assumption and are willing to be judged against it. He and Levina also gave the theoretical account of why naive Bayes and independence-rule classifiers, which are obviously wrong, nevertheless beat the "correct" Fisher discriminant in high dimensions: estimating p²/2 covariance entries from n samples introduces more error than the independence assumption removes.

**Networks and genomics.** With Aiyou Chen he gave an asymptotic framework for community detection, analyzing Newman–Girvan modularity and likelihood-based methods under block models and establishing consistency conditions — bringing statistical rigor to a literature that had been largely algorithmic. His genomics work brought careful inference to consortium-scale annotation data.

In battle

Bickel's computed profile is one of the most *even* strong profiles on the roster: mean 46.6 over 100 problems, median 45, eight dominant scores, and only thirteen problems at 20 or below. He does not have Pearl's or Rubin's towering causal peaks, but he almost never bottoms out on anything statistical — the mark of someone whose tools are foundational rather than specialized.

His top score is **P214 — The paradox in the admissions data** (98), and the judge's note is unusually direct: he is not analogous to this problem, he is one of its actual authors. The 1975 *Science* paper is the definitive analysis of exactly this dataset, and his department-by-department decomposition — showing the campus-wide gap reversed within departments because women applied disproportionately to more selective ones — is precisely the resolution the problem demands.

His high-dimensional category average of **90** is the highest single category number in his profile, and it is earned by two specific problems: **P282 — The eigenvalues are lying** (90), where the sample covariance eigenvalues of a p-comparable-to-n matrix are systematically distorted and his regularization theory says what to do about it, and **P283 — Five hundred stocks, one year of days** (90), the singular-covariance problem that banding and thresholding were built for. **P019 — Regression with lies in it** (87) and **P126 — When the bell curve won't hold** (87) are his robustness program directly: contaminated data and non-Gaussian errors are exactly what M-estimators and adaptive procedures exist to survive. **P004 — The bootstrap when theory fails** (85) puts him just below Efron on Efron's own problem, appropriately — Efron proposed it, Bickel proved when it holds. **P106 — Squeeze the estimator dry** (82) is the efficiency-bound question, his home ground. **P059 — The confounder you can't measure** (83) reflects semiparametric efficiency theory's role in modern causal estimation, and it is worth noting that his causality average of 47.1 is strong for someone who is not primarily a causal-inference figure.

The losses share a single feature: they are all problems with no probability model at all. He scores 5 on **P201 — The dice make it learnable**, the TD-Gammon problem, where temporal-difference learning through self-play is a reinforcement-learning mechanism his toolkit has no contact with. **P079 — The language for the job** (6) is 1959 programming-language design, **P163 — The fastest-growing function** (8) is computability, **P193 — Search deep on a shoestring of memory** (8) is memory-bounded heuristic search, and **P084 — Find the needle documents** (8) is information retrieval. **P083 — The grammar beneath the sentence** (8) pulls his NLP average down to 8, his worst category.

Draft Bickel when the problem is a genuine statistical question and the naive answer is about to fail — contaminated data, p near or above n, a nuisance function you cannot avoid, an estimator whose efficiency you need to defend. He is the roster's best pick for "is this procedure actually optimal, and how would you know?" Keep him away from anything algorithmic, symbolic, or engineered.