{ "id": "terence-tao", "title": "Terence Tao", "lang": "en", "source_lang": "en", "summary": "Australian-American mathematician whose research spans harmonic analysis, partial differential equations, combinatorics, and number theory.", "category": "mathematics", "subcategory": "analysis", "type": "person", "aliases": [ "Terence Chi-Shen Tao", "Terry Tao", "陶哲轩", "陶哲軒" ], "generated_at": "2026-10-22", "model": "pro/gpt-6.1-sol", "url": "https://aiwiki.page/en/terence-tao/", "translations": { "zh-Hans": { "url": "https://aiwiki.page/zh/terence-tao/", "json": "https://aiwiki.page/zh/terence-tao.json" } }, "sections": [ { "heading": "", "text": "Terence Tao (born 17 July 1975 in Adelaide, Australia) is an Australian-American mathematician and a Distinguished Professor at the University of California, Los Angeles (UCLA), where he holds the James and Carol Collins Chair in Mathematics. His research encompasses harmonic analysis, partial differential equations, combinatorics, and number theory, alongside contributions to random matrix theory, representation theory, and compressed sensing. He received the Fields Medal in 2006 for contributions to partial differential equations, combinatorics, harmonic analysis, and additive number theory. (math.ucla.edu)\n\nTao’s work connects methods from analysis with questions about discrete structures and arithmetic. With Ben Green, he proved that the prime numbers contain arithmetic progressions of every finite length; his other research includes the Erdős discrepancy problem and, with collaborators, the recovery of sparse signals from incomplete measurements. His mathematical activities also include textbooks, research exposition, online collaborative projects, and the formal verification of proofs. (arxiv.org)" }, { "heading": "Early life and mathematical education", "text": "Terence Tao was born in Adelaide, Australia, on 17 July 1975 and grew up there. His early education involved acceleration across school levels rather than simply completing a conventional curriculum more quickly. In a 2009 interview with the Australian Mathematical Society’s Gazette, Tao recalled skipping five primary-school grades and beginning high school at eight. The acceleration was staggered by subject: at that age he attended Year 8 classes in subjects such as English and physical education while studying mathematics at Year 11 or Year 12 level. His parents arranged access to advanced classes with school and university staff, and he subsequently took university mathematics while still attending high school. These documented arrangements provide a concrete account of his early development without relying on anecdotes about infant abilities. (math.ucla.edu)\n\nTao represented Australia in the International Mathematical Olympiad in three consecutive years, winning bronze in 1986, silver in 1987, and gold in 1988. The official competition records give his totals as 19, 40, and 34 points respectively, with overall ranks of 87th, 28th, and 13th. His individual problem scores also show the distinction between solving selected problems completely and performing consistently across the paper: in 1986 he received the maximum seven points on two problems, whereas in 1987 he received seven points on five problems and five on the sixth. The progression in medals therefore should not be read as a steadily increasing raw score; his silver-medal total in 1987 exceeded his gold-medal total the following year. (m.imo-official.org)\n\nOlympiad participation overlapped with, rather than preceded, his exposure to higher education. Flinders University records that he was studying mathematics there at ten under the supervision of Garth Gaudry. Competition problem-solving was thus one strand of an education that also included sustained contact with university mathematicians. Recalling his childhood interests in the Gazette interview, Tao distinguished his enjoyment of puzzles and arithmetic from his later understanding of mathematical research: as a child, he had not yet understood that researchers must formulate their own problems rather than merely solve assigned ones. (news.flinders.edu.au)\n\nAt Flinders University, Tao completed a Bachelor of Science with honours in mathematics in December 1991, aged sixteen, followed by a Master of Science in August 1992, aged seventeen. His curriculum vitae identifies Gaudry as his adviser for both degrees. The master’s thesis was titled Convolution operators generated by right-monogenic and harmonic kernels, documenting a specialized interest in mathematical analysis before his doctoral studies. Gaudry’s role extended beyond supervising this work: Tao later recalled that, after his bachelor’s degree, Gaudry strongly encouraged him to study abroad. His subsequent move to Princeton University followed this period of Australian schooling, competition participation, and university training. (math.ucla.edu)" }, { "heading": "Doctoral training and academic career", "text": "Tao completed his Ph.D. in mathematics at Princeton University in June 1996 under the supervision of Elias Stein. His dissertation, Three regularity results in harmonic analysis, brought together three investigations that appeared as journal articles: “Weak-type endpoint bounds for Riesz means” and “On the almost-everywhere convergence of wavelet summation methods,” both published in 1996, and “Weak-type endpoint bounds for homogeneous convolution operators,” published in 1999. These studies placed his doctoral work within harmonic analysis, with an emphasis on bounds for operators and convergence of approximation procedures. Tao’s description of the dissertation explicitly credits Stein’s guidance and identifies the thesis as essentially the union of these three papers. (math.ucla.edu)\n\nIn 1996, Tao joined the University of California, Los Angeles (UCLA), as a Hedrick Assistant Professor, an appointment his curriculum vitae records for 1996–1998. His subsequent progression should be distinguished from this initial position: his professional biography dates his professorship at UCLA to 1999, while the detailed appointments table lists an acting assistant professorship in 1999 and the full professorship from 2000. UCLA’s institutional account, published on August 22, 2006, states that he joined the faculty in 1996 and was promoted to full professor at age 24. These records establish the rapid transition from his initial Hedrick appointment to the professorial rank, while preserving the differing dates used in the biographical summary and appointment chronology. (teorth.github.io)\n\nTao’s later appointments include Distinguished Professor, dated from 2009 in his curriculum vitae, and the James and Carol Collins Chair in UCLA’s College of Letters and Science. A UCLA announcement dated January 11, 2007, identified him as the chair’s first holder and reported that James and Carol Collins had endowed it with a $1 million gift. Unlike chairs restricted to a particular discipline, this chair could be awarded to a professor in any of the College’s more than 30 departments. UCLA’s mathematics faculty directory lists Tao with both the Distinguished Professor designation and the Collins Chair. (teorth.github.io)\n\nUCLA’s departmental profile describes his research interests as analysis and oscillatory integrals; his professional homepage gives a broader disciplinary list: harmonic analysis, partial differential equations, geometric and arithmetic combinatorics, analytic number theory, compressed sensing, and algebraic combinatorics. His detailed research profile specifies oscillatory integrals and Kakeya maximal functions, nonlinear dispersive equations, multilinear operators such as the bilinear Hilbert transform, and connections between arithmetic combinatorics, number theory, and ergodic theory. It also lists secondary interests in representation theory, geometric partial differential equations, and inverse scattering for rough data, documenting an academic research programme extending beyond the subjects of his dissertation. (math.ucla.edu)" }, { "heading": "Harmonic analysis and geometric questions", "text": "The Fourier restriction problem asks when the Fourier transform of a function can be meaningfully restricted to a lower-dimensional surface, such as a sphere, with bounds in suitable L^p spaces. Its dual formulation estimates oscillatory integrals assembled from frequencies on that surface. The Kakeya problem concerns the size of sets containing a unit line segment in every direction. Their connection becomes visible through wave packets: localized oscillatory contributions concentrate along thin tubes, making estimates for functions depend on how tubes with different directions overlap. (arxiv.org)\n\nIn their 2005 paper, Jonathan Bennett, Anthony Carbery, and Tao established near-optimal multilinear restriction and Kakeya estimates. In \\(d\\) dimensions, they considered products involving \\(d\\) families whose normal or tube directions are uniformly transverse: the determinant formed by these directions stays bounded away from zero. This spanning condition, rather than curvature, controls their interaction. Their proof used monotonicity formulae for evolving Gaussian functions, related to heat flow. The restriction estimate allowed an arbitrarily small loss \\(R^\\varepsilon\\) on a ball of radius \\(R\\); the Kakeya estimate excluded the endpoint exponent. These results did not establish the full linear conjectures. (arxiv.org)\n\nWith Gerd Mockenhaupt, Tao developed analogous questions over a finite field in a paper first submitted in 2002 and published in Duke Mathematical Journal in 2004. Here, continuous Euclidean space is replaced by the finite set \\(\\mathbb{F}_q^n\\), and Fourier integrals become finite sums involving additive characters—the finite-field counterparts of complex exponentials. Restriction estimates therefore become bounds for exponential sums associated with algebraic surfaces, including paraboloids. The Kakeya question asks how small a set can be if it contains an entire affine line in every direction. This setting removes limiting and small-scale analytic difficulties while retaining the interaction between oscillation and incidence geometry. Their paper transferred several Euclidean arguments to this discrete setting while identifying additional phenomena arising from finite-field algebra. (arxiv.org)\n\nA different connection between Fourier analysis and geometry appears in Tao’s 2003 counterexample to Fuglede’s conjecture. A measurable set \\(\\Omega\\) of finite positive volume is spectral if \\(L^2(\\Omega)\\) has an orthonormal basis consisting of normalized exponentials\n\n\\[|\\Omega|^{-1/2}e^{2\\pi i\\lambda\\cdot x}, \\qquad \\lambda\\in\\Lambda,\\]\n\nfor a suitable frequency set \\(\\Lambda\\). A translational tile, by contrast, admits translates that cover Euclidean space without overlaps, apart from sets of measure zero. Fuglede’s conjecture asserted that these analytic and geometric properties were equivalent. (arxiv.org)\n\nTao constructed a finite union of unit cubes in \\(\\mathbb{R}^5\\) that was spectral but could not tile by translations, disproving the implication from spectrality to tiling in dimension five and, through extensions of the construction, every higher dimension. The construction began with a six-point spectral configuration in \\(\\mathbb{F}_3^5\\). Its cardinality prevented a translational tiling of that finite group, whose order is \\(3^5=243\\). Tao transferred this finite algebraic obstruction into a Euclidean construction, producing a set with an exponential basis but no translational tiling. (arxiv.org)" }, { "heading": "Nonlinear partial differential equations", "text": "Tao’s research on nonlinear partial differential equations examines whether prescribed initial data generate solutions, whether those solutions are unique, and whether their regularity persists as time evolves. Local well-posedness combines existence and uniqueness over a time interval with continuous dependence on the initial data; global well-posedness extends this control to arbitrarily large times. Regularity concerns the preservation of smoothness or bounds on derivatives, while scattering asks whether a nonlinear solution approaches a solution of the corresponding linear equation at large positive or negative times. These are distinct questions: a solution can exist globally without scattering. Much of Tao’s work addresses how dispersion and conservation laws prevent concentration, or why they sometimes fail to do so. (math.ucla.edu)\n\nIn a 2005 paper with Monica Visan and Xiaoyi Zhang, Tao studied the nonlinear Schrödinger equation with two power nonlinearities,\n\n\\[i\\partial_tu+\\Delta u =\\lambda_1|u|^{p_1}u+\\lambda_2|u|^{p_2}u, \\qquad x\\in\\mathbb R^n,\\]\n\nwhere \\(n\\geq3\\), the coefficients are nonzero real numbers, and \\(00\\), even if the lower-power term is focusing. They also proved global well-posedness for either choice of signs when \\(p_2<4/n\\). Thus both the dimension-dependent exponents and the signs of the nonlinearities determine the available conclusions. The thresholds \\(4/n\\) and \\(4/(n-2)\\) are respectively mass-critical and energy-critical. (arxiv.org)\n\nTheir scattering theorem imposed stronger restrictions. When both coefficients are positive and \\(4/n0,\\]\n\nwhere \\(\\pi(N)\\) counts primes up to \\(N\\), then \\(A\\) also contains arbitrarily long progressions. Thus the conclusion survives restriction to any subset occupying a positive proportion of the primes along an unbounded sequence of intervals. (arxiv.org)\n\nTheir proof combines three ingredients: Szemerédi’s theorem, a transference principle, and estimates of Daniel Goldston and Cem Yıldırım. The transference principle carries a dense-set conclusion into a sparse setting whose surrounding measure satisfies suitable pseudorandomness conditions. Here “pseudorandom” means that specified averages and correlations behave approximately as they would for a uniform measure—not that the primes are independent random choices. The proof controls the structured and uniform components of a function representing the primes, allowing the positive-density theorem to operate despite their vanishing density among all integers. (arxiv.org)\n\nThe number-theoretic ingredient constructs a nonnegative majorizing weight from truncated divisor sums, concentrated on “almost primes.” Goldston–Yıldırım correlation estimates establish the required pseudorandomness of this larger weighted environment, within which suitably normalized primes have positive relative density. A preliminary \\(W\\)-trick removes systematic divisibility biases: \\(W\\) is a product of small primes, and attention is restricted to numbers \\(Wn+b\\) with \\(b\\) coprime to \\(W\\). This separates unavoidable local arithmetic restrictions from the uniformity needed for transference. The argument establishes existence through density and correlation estimates rather than by computing particular long progressions. (arxiv.org)\n\nTao and Tamar Ziegler extended this framework to polynomial patterns in work submitted in 2006 and published in Acta Mathematica in 2008. For any finite collection \\(P_1,\\ldots,P_k\\in\\mathbb Z[m]\\) satisfying \\(P_i(0)=0\\), and any \\(\\varepsilon>0\\), they proved that infinitely many integer pairs \\(x,m\\), with \\(1\\leq m\\leq x^\\varepsilon\\), make every \\(x+P_i(m)\\) prime simultaneously. Taking \\(P_i(m)=(i-1)m\\) recovers arithmetic progressions; choosing \\(0,m^2,2m^2\\) gives three-term prime progressions with square common difference. The zero-constant-term hypothesis is essential to the stated theorem: arbitrary fixed offsets can introduce divisibility obstructions. Their proof uses polynomial extensions of the transference machinery, localized uniformity estimates, and an induction procedure that reduces polynomial averaging problems to linear ones. (arxiv.org)" }, { "heading": "Discrepancy, multiplicative functions, and the Collatz map", "text": "In September 2015, Tao proved the Erdős discrepancy problem: every sequence \\(a:\\mathbb N\\to\\{-1,+1\\}\\) satisfies\n\n\\[\\sup_{d,m\\geq1}\\left|\\sum_{j=1}^{m}a(jd)\\right|=\\infty.\\]\n\nThe sums run along homogeneous arithmetic progressions \\(d,2d,\\ldots,md\\). Thus, however the two signs are assigned, some such progression has an imbalance exceeding any prescribed bound. The quantification over both \\(d\\) and \\(m\\) matters: the theorem does not assert that ordinary partial sums, corresponding to \\(d=1\\), must be unbounded. His paper, published in Discrete Analysis in 2016, also established a vector-valued extension for sequences of unit vectors in a real or complex Hilbert space. (arxiv.org)\n\nThe proof connected this arbitrary-sign statement to multiplicative functions. A Fourier-analytic reduction developed through Polymath5 converted the vector-valued problem into one about random completely multiplicative functions \\(g\\), for which \\(g(rs)=g(r)g(s)\\) for all positive integers \\(r,s\\). The required conclusion became unboundedness of the expected squared magnitude of their partial sums. These functions were auxiliary objects in the proof; the original discrepancy theorem imposed no multiplicativity assumption on the sign sequence. (arxiv.org)\n\nTao then used his logarithmically averaged form of an Elliott-type correlation theorem. Under suitable hypotheses, it controls correlations between multiplicative functions evaluated at different affine-linear arguments, weighting the summands by \\(1/n\\). Together with a van der Corput argument, this forced any hypothetical bounded-discrepancy example to resemble a modulated Dirichlet character, \\(n\\mapsto\\chi(n)n^{it}\\), in a precise sense measured through its values at primes. An extension of another Polymath5 argument excluded this remaining structured case. The proof therefore combined a collaborative reduction, Tao’s correlation estimates, and a final analysis of structured multiplicative functions. (arxiv.org)\n\nIn September 2019, Tao announced a separate result concerning the Collatz map, defined by \\(C(N)=N/2\\) for even \\(N\\) and \\(C(N)=3N+1\\) for odd \\(N\\). Writing \\(C_{\\min}(N)=\\min_{k\\geq0}C^k(N)\\), he proved that, for every function \\(F(N)\\to+\\infty\\),\n\n\\[C_{\\min}(N)0 \\quad\\Longleftrightarrow\\quad c_{\\lambda\\mu}^{\\nu}>0,\\]\n\nwhere scaling multiplies every part by \\(N\\). This is a statement about nonvanishing, not equality of the coefficients before and after scaling. (arxiv.org)\n\nA honeycomb consists of edges in three prescribed directions in the plane \\(x+y+z=0\\), with balancing conditions at vertices and semi-infinite edges recording boundary data. For fixed boundary data, the configurations form a convex polytope; counting its integral configurations gives the corresponding representation-theoretic multiplicity. Thus a discrete counting problem becomes a geometric problem involving continuously variable edge lengths, with integral honeycombs selecting the configurations relevant to representations. (arxiv.org)\n\nThe proof’s crucial step was to pass from real feasibility to integral feasibility. An integral honeycomb for the scaled partitions can be divided by \\(N\\), producing a real honeycomb for the original partitions, but generally not an integral one. Knutson and Tao showed that a nonempty honeycomb polytope with integral boundary data contains an integral honeycomb. They selected a special extremal configuration by maximizing a weighted-perimeter functional and established its integrality. This required more than ordinary convexity: some vertices of honeycomb polytopes with integral boundaries are themselves nonintegral. (arxiv.org)\n\nHoneycombs also encode possible eigenvalues of sums of Hermitian matrices. Given decreasing eigenvalue lists \\(\\lambda,\\mu,\\nu\\), the existence of matrices \\(A,B\\) with these spectra for \\(A,B,A+B\\) is equivalent to the existence of an appropriately bounded real honeycomb. Combined with Alexander Klyachko’s work, saturation completed the proof of Alfred Horn’s 1962 conjecture describing the possible spectra through linear inequalities. (arxiv.org)\n\nIn subsequent work with Christopher Woodward, published in 2004, Knutson and Tao introduced triangular “puzzles” to determine which inequalities define essential facets of the spectral cone. They also proved William Fulton’s conjecture that multiplicity one persists under scaling:\n\n\\[c_{\\lambda\\mu}^{\\nu}=1 \\quad\\Longrightarrow\\quad c_{N\\lambda,N\\mu}^{N\\nu}=1 \\qquad(N\\geq1).\\]\n\nThis strengthened the scaling picture beyond the positivity assertion of saturation. (arxiv.org)" }, { "heading": "Random matrices and spectral universality", "text": "Tao’s collaboration with Van Vu on random matrix theory developed methods for comparing spectral statistics of matrices whose entries have different probability distributions. Their work distinguishes global questions—such as the proportion of eigenvalues in a fixed interval—from local questions about individual eigenvalues, neighbouring gaps, and correlations at the mean-spacing scale. For a Wigner matrix \\(M_n\\), with independent upper-triangular entries subject to Hermitian symmetry, centred entries, and off-diagonal variance one, the normalization \\(W_n=M_n/\\sqrt n\\) gives the global semicircle density\n\n\\[\\rho_{\\mathrm{sc}}(x)=\\frac{1}{2\\pi}\\sqrt{4-x^2},\\qquad |x|\\leq 2.\\]\n\nThis density describes the overall spectrum but does not determine its microscopic fluctuations. Near an interior point \\(E\\in(-2,2)\\), the mean eigenvalue spacing of \\(W_n\\) is approximately \\(1/(n\\rho_{\\mathrm{sc}}(E))\\). (arxiv.org)\n\nThe Four Moment Theorem, surveyed by Tao and Vu in December 2011, supplies a comparison principle at this finer scale. In one formulation, two Wigner ensembles satisfy a uniform stretched-exponential tail bound, their corresponding off-diagonal entries match moments through order four, and their diagonal entries match through order two. For complex entries, matching means equality of mixed moments of the real and imaginary parts up to the specified total degree. After the fine-scale normalization \\(A_n=\\sqrt n M_n\\), expectations of sufficiently smooth functions of a fixed number of selected eigenvalues differ by a quantity tending to zero. The test functions must satisfy explicit derivative bounds; the theorem is not an assertion that the entire spectra have identical distributions. (arxiv.org)\n\nThe proof uses the Lindeberg replacement strategy: entries are exchanged one at a time, replacing each off-diagonal entry together with its conjugate to preserve Hermitian symmetry. A perturbative expansion shows that matching moments cancels the leading contributions to the expected change. Estimates controlling eigenvalue gaps and eigenvector delocalization bound the remaining errors, allowing them to be summed over the matrix’s entries. This connects assumptions on individual random variables to the collective behaviour of eigenvalues. (arxiv.org)\n\nIn their February 2012 survey, Tao and Vu described how comparison theorems, combined with results for Gaussian-divisible ensembles and moment-matching constructions, yield universality beyond Gaussian matrices. For complex Hermitian ensembles under the relevant hypotheses, bulk correlation statistics converge, in specified weak senses, to those governed by the sine kernel,\n\n\\[K(x,y)=\\frac{\\sin\\!\\bigl(\\pi(x-y)\\bigr)}{\\pi(x-y)},\\]\n\nwith diagonal value \\(K(x,x)=1\\). The rescaling by the local semicircle density is essential. Real symmetric ensembles have a different symmetry class, while spectral edges require a different scaling from the bulk. Moment matching is a hypothesis of the comparison theorem, not a claim that every universality result requires four prescribed moments; their survey also discusses versions using sufficiently high finite-moment bounds rather than stretched-exponential tails. (arxiv.org)\n\nTheir 2011 paper “Random matrices: Universal properties of eigenvectors” extended comparison to eigenvector coordinates and established associated central limit theorems. Because a unit eigenvector is defined only up to sign or complex phase, the proof uses rank-one projections \\(P_i=u_i u_i^*\\), which remove this ambiguity. Under suitable matching and moment hypotheses, selected coordinates, normalized on their natural \\(n^{-1/2}\\) scale and with an appropriate phase convention, exhibit Gaussian limiting behaviour. The paper also proves universality results for the resolvent \\(R(z)=(M_n-zI)^{-1}\\), extending to the real axis under its stated conditions and implying results for inverse matrices. Resolvents encode both eigenvalue locations and eigenvector weights, so these results compare more than the spectrum alone. (arxiv.org)" }, { "heading": "Compressed sensing and sparse recovery", "text": "Tao’s collaboration with Emmanuel Candès and Justin Romberg investigated when incomplete linear measurements determine a sparse signal, a central problem in compressed sensing. Their paper “Robust Uncertainty Principles: Exact Signal Reconstruction from Highly Incomplete Frequency Information,” submitted in September 2004 and published in IEEE Transactions on Information Theory in 2006, considered signals with relatively few nonzero entries observed through a randomly selected subset of their Fourier coefficients. For a fixed signal of length \\(N\\) with \\(s\\) nonzero entries, their results gave exact recovery with high probability when the number of sampled frequencies was sufficiently large relative to \\(s\\log N\\), with constants depending on the desired probability of success. This was a probabilistic recovery guarantee, not a statement that every selection of that many frequencies suffices. (arxiv.org)\n\nThe underlying inverse problem can be written \\(y=Ax\\), where \\(x\\) has \\(N\\) coordinates and the measurement matrix \\(A\\) has only \\(m