Emmanuel Candes

Life and career

The result that made Emmanuel Candès famous sounds, on first hearing, like a violation of something. You have a signal with a million components. You take a few thousand random linear measurements of it — nowhere near enough to determine it, by any linear-algebra accounting. And yet, if the signal is sparse in some basis, you can recover it **exactly**, with probability approaching one, by solving a convex program. Not approximately. Exactly.

Candès is French, trained at the École Polytechnique, and came to Stanford for a doctorate in statistics under David Donoho — a lineage that placed him at the center of the harmonic-analysis-meets-statistics tradition that Donoho had been building. Donoho's program was about sparsity: the observation that natural signals, expressed in the right basis, have most of their energy in a few coefficients, and that this fact could be turned into estimators with provably optimal behavior. Candès's early work with Donoho on **curvelets** is squarely in that vein — a multiscale representation adapted to curved edges in images, with near-optimal approximation rates for a natural class of images that wavelets handle badly because wavelets are isotropic and edges are not.

He spent a productive period at Caltech in applied and computational mathematics, where the compressed sensing work was done, and then returned to Stanford, where he holds chairs in mathematics and statistics and is affiliated with electrical engineering. The three-department footprint is accurate rather than decorative: his results are theorems in probability and convex geometry, they are estimators with statistical risk properties, and they are algorithms that get implemented in scanners.

He has been recognized about as thoroughly as a mathematician can be — the NSF's Waterman Award, a MacArthur Fellowship, election to the National Academy of Sciences, and major prizes from the applied mathematics societies, several shared with Terence Tao for their joint work.

What is most notable about his trajectory, though, is that he has restarted twice. Having essentially finished compressed sensing, he moved to matrix completion and low-rank recovery, then to phase retrieval, and then — in a genuine change of field — to the statistics of selective inference and distribution-free uncertainty quantification, where his knockoffs framework and his conformal-prediction work have made him a central figure in a literature quite far from harmonic analysis. Few people manage one such transition.

Key contributions

**Compressed sensing.** The precise statement is the achievement. Suppose x ∈ ℝⁿ is s-sparse and you observe y = Ax with A an m × n matrix, m ≪ n. Solve

minimize ‖x̂‖₁ subject to Ax̂ = y.

Candès and Tao introduced the **restricted isometry property** (RIP): A satisfies RIP of order 2s if it acts almost like an isometry on all 2s-sparse vectors, preserving their norms within a factor 1 ± δ. Their theorem is that if δ is small enough, ℓ₁ minimization recovers x exactly — and moreover the recovery is stable: with noisy measurements and merely approximately sparse signals, the error is bounded by the noise level plus the best s-term approximation error. Random matrices (Gaussian, or randomly subsampled Fourier) satisfy RIP with m on the order of s log(n/s), which is within a logarithmic factor of the information-theoretic minimum.

Why this matters conceptually: it says the number of measurements you need scales with the signal's *information content*, not with its ambient dimension, and the recovery is by a tractable convex program rather than an intractable combinatorial search. Why it matters practically: MRI acquisition time is proportional to the number of measurements taken, and a scan that takes a quarter as long is a different clinical object. Compressed sensing reconstruction shipped in commercial MRI systems.

**The Dantzig selector.** With Tao, a companion estimator for the noisy sparse regression problem: minimize ‖β̂‖₁ subject to the residual correlations ‖Aᵀ(y − Aβ̂)‖∞ being small. Its risk is within a logarithmic factor of what an oracle knowing the true support could achieve — an early and influential statement of what high-dimensional regression can and cannot deliver.

**Matrix completion.** The same philosophy, one structure up. You observe a small random subset of the entries of a large matrix that is (approximately) low rank. Candès and Recht proved that minimizing the **nuclear norm** — the sum of singular values, the convex relaxation of rank — subject to matching the observed entries recovers the matrix exactly, provided the number of observed entries exceeds roughly nr·polylog(n) and the matrix satisfies an **incoherence** condition ensuring its singular vectors are not concentrated on a few coordinates. That incoherence caveat is not a technicality: a matrix whose entire signal sits in one entry cannot be recovered from a random sample that misses it. **Robust PCA**, with Li, Ma, and Wright, extended this to decomposing a matrix into a low-rank part plus a sparse gross-corruption part, again by convex relaxation, again with exact-recovery guarantees.

**Phase retrieval.** In X-ray crystallography and diffraction imaging, detectors record intensities and lose phase. Recovering the signal from magnitude-only measurements is a nonconvex problem with a long history of heuristics. Candès and collaborators developed **PhaseLift**, lifting to a semidefinite program over rank-one matrices, and later **Wirtinger flow**, a nonconvex gradient method with a careful spectral initialization and a proof that it converges to the global optimum — one of the early demonstrations that a well-initialized nonconvex method can be *analyzed*, not merely tuned.

**Knockoffs.** A genuine break with his earlier work, and possibly his most important statistical contribution. The problem: you run a variable-selection procedure on high-dimensional data, and you want to control the false discovery rate among the variables you report — but the selection is data-dependent and adaptive, so classical FDR procedures do not apply. The knockoff construction, with Rina Barber and later extended to **model-X knockoffs** with Fan, Janson, and Lv, manufactures for each real variable a synthetic "knockoff" copy that mimics its correlation structure with the other variables but is conditionally independent of the response. The knockoffs act as internal negative controls: run any selection procedure on the augmented set, and the number of knockoffs that sneak in gives an estimate of the number of false real discoveries, which supports a data-dependent threshold with provable finite-sample FDR control — no p-values, no asymptotics, and in the model-X version no assumption about how the response depends on the covariates.

**Conformal prediction.** More recently, Candès has been a driver of distribution-free predictive inference: conformalized quantile regression, which yields prediction intervals that adapt their width to local difficulty; extensions to covariate shift via weighted exchangeability; and conformal methods for survival and counterfactual settings.

In battle

Candès has one of the most decisive high-dimensional profiles in the game: mean 38.6 overall, but `high-dim` at 86.2 and `optimization` at 75.2 — the two highest category scores anyone here carries in those areas — and ten problems above 80.

**P094 — Reconstruct from too few measurements** at 98 is his career restated as a prompt: RIP conditions on the measurement matrix, ℓ₁ relaxation, exact recovery with high probability, medical imaging as motivation. Nowak scores 70 on the same problem and will lose to him. **P280 — The matrix with 99% holes** at 98 is nuclear-norm matrix completion, again his own theorem with its own incoherence caveat.

**P281 — Inference after the search** at 97 is knockoffs and selective inference — and it is worth noticing how different this problem is from the two above. A player who thinks of Candès as "the compressed sensing guy" will underplay him on statistical inference problems where he is in fact first-rate. **P018 — p = 20,000, n = 200** at 95 beats Wainwright's 93 and Nowak's 81 in the most contested high-dimensional problem in the pool. **P124 — Twenty thousand tests at once** (90) is multiple testing and FDR. **P285 — Test the many with the blood of few** (87) is group testing as sparse recovery, where Nowak's adaptive designs edge him at 90. **P044 — Sketch the enormous matrix** (86) is randomized low-rank approximation, where Mahoney's 97 wins. **P288 — Let the images choose the basis** (85) reaches back to curvelets and adaptive representations.

The category table shows a wide competent middle: `information` 62.5, `networks` 60.0, `regression` 40.7, `testing` 35.4, `small-sample` 33.8. He is a serious statistician as well as a serious mathematician, and the profile credits both.

The losses are instructive because they are not where you would guess. `classification` at 12.5 is his **lowest** category — remarkable for someone this central to machine learning, and correct: Candès has essentially no record in supervised classification, decision boundaries, or pattern recognition. On **P010**-style or margin problems he is simply absent.

**P083 — The grammar beneath the sentence** at 8 is his floor: formal language theory and the Chomsky hierarchy, five decades before his career and sharing no machinery with convex recovery, notwithstanding his genuine theorem-proving pedigree. **P214 — The paradox in the admissions data** (10) is Simpson's paradox and **P215 — The therapy the trial reversed** (11) is a clinical-trial reversal — `causality` at 26.2 across seventeen problems is the single largest weakness in his profile, and it is a real one: nothing in sparse recovery gives you an account of confounding. **P146 — The tree you can read** (10) is interpretable models, **P193 — Search deep on a shoestring of memory** (10) is bounded-memory search, and **P216 — Calculus for a jagged path** (11) is stochastic calculus.

His battle identity names the boundary exactly: he wins undersampled recovery with convex relaxation and proof, and loses when sparsity or incoherence collapses. That is not a rhetorical hedge — it is the assumption every one of his theorems is conditioned on, and he has always been explicit about it. Give him a problem where the signal has low-dimensional structure and the measurement is linear, and he is close to unbeatable. Take the structure away and the guarantees go with it.