What's new:
What was new in 2004? 2003?
2002? 2001?2000?1999?
Dec 6, 2005
 - Uploaded:
     Sample chapters for the second edition of my book “Solving
     mathematical problems: a personal perspective”.  Due to “popular demand”,
     this book is now being reprinted by Oxford University Press, and should be
     available roughly at the time of the 2006 ICM in August.
Dec 5, 2005
 - Uploaded:
     “The dichotomy
     between structure and randomness, arithmetic progressions, and the primes”.  This is a full-length version of a paper
     submitted for the 2006 ICM proceedings; due to length considerations, the
     actual ICM paper will be a substantially abridged version of this
     paper.  Here I survey a class of
     structure theorems – that roughly speaking obtain a decomposition of
     an arbitrary object into a “structured” (low-complexity,
     self-correlated, measurable) component and a “pseudorandom”
     (high-complexity, discorrelated, anti-measurable) component – and
     how they can be used to prove results such as Szemeredi’s theorem
     and the Green-Tao theorem on arithmetic progressions (in dense subsets of
     integers, and dense subsets of primes, respectively).  The structure theorems exist in many
     contexts (ergodic theory, Fourier analysis, graph theory, hypergraph
     theory, probability theory, information theory) but the point of this
     paper is to emphasize the common themes between these theorems even though
     their settings vary widely.
Nov 8, 2005
 - Uploaded:
     “Inverse
     Littlewood-Offord theorems and the condition number of random discrete
     matrices”, joint with Van
     Vu.  Here we return to random
     Bernoulli matrices, but instead of trying to control the determinant or
     the singularity probability of these matrices, we now consider the
     condition number, which is basically the ratio between the largest and
     smallest singular value.  The
     largest value is quite stable and is known to be comparable to n^{1/2}
     with very high probability, so it is only a matter of understanding the
     lowest singular value.  This is conjectured
     to be of size roughly n^{-1/2} (in analogy with random Gaussian matrices),
     and there is a result of Rudelson showing that this value is at least
     n^{-3/2} with reasonable probability. 
     The main result of this paper is a tail estimate, namely for any A
     there exists a B such that the lowest singular value is at least n^{-B}
     with probability 1 – O(n^{-A}). 
     Equivalently, the condition number is O(n^{-B}) with probability 1
     – O(n^{-A}).  There are three
     major ingredients in the proof, which may have some independent
     interest.  The first ingredient is a
     new inverse Littlewood-Offord theorem which almost completely
     characterizes which hyperplanes intersect the discrete unit cube {-1,1}^n
     with polynomial density (the coefficients of the normal vector have to lie
     almost completely in a generalized arithmetic progression of bounded
     size).  This theorem, which is
     somewhat similar in spirit to Freiman’s inverse sumset theorem, is
     proven mostly by elementary methods (with a touch of Fourier analysis),
     the main difficulty being a technical issue of “clearing
     denominators” with only polynomial losses.  The second ingredient, which is also
     new, is a discretization of progressions, which allows one to split any
     progression as the sum of a sparse progression (wide separation between
     elements) and a small progression. 
     This is a discrete analogue of the fact that a linear map can be
     written as the direct sum of an injective map and a zero map.  Combining these ingredients allows us to
     convert the smallest singular value problem to the singularity problem,
     which we handle by a generalization of the Kahn-Komlos-Szemeredi theorem
     on random Bernoulli matrices, allowing a small number of the rows to be
     replaced by deterministic vectors.
Nov 3, 2005
 - Uploaded:
     “The nonlinear
     Schr\”odinger equation with combined power-type nonlinearities”,
     joint with Monica Visan
     and Xiaoyi Zhang.  Here we
     systematically study the global well-posedness and scattering theory for
     an NLS with two power nonlinearites, either of which can be focusing or
     defocusing, in three and higher dimensions, in the energy space H^1 and
     also in the pseudoconformal space Sigma, without any symmetry assumptions
     on the data.  If the larger power
     was defocusing, it was already known that as long as both powers were
     H^1-subcritical, then one had global well-posedness, and it was also known
     that the single power defocusing H^1-critical equation had global
     well-posedness.  We combine these
     results by extending the former to the endpoint case when the larger power
     is H^1-critical; this is by a perturbation argument using the H^1-critical
     pure power equation as the approximating equation.  Next, we consider the scattering theory
     in H^1, and show that one has scattering in the doubly defocusing case as
     long as both powers lie between the L^2-critical and H^1-critical
     exponents inclusive.  In the case
     that the lower power is L^2-critical, we need an additional hypothesis,
     namely that the pure power L^2-critical equation enjoys scattering and
     spacetime estimates (an “L^2-critical gopher”, if you
     will).   This is especially delicate
     for the “doubly critical” case involving both the L^2-critical
     power and the H^1-critical power; here we need to approximate the
     low-frequency evolution by the L^2-critical equation and then the
     high-frequency evolution by the H^1-critical equation, keeping the medium
     frequencies under control via a Morawetz inequality.  We also exploit the virial identity to
     force blowup in some negative-energy, inward-momentum cases, and the
     pseudoconformal identity to obtain scattering in Sigma in other cases.
- Uploaded:
     “Entropy sum set
     estimates”, a deleted chapter from my upcoming book with Van
     Vu.  Here we take the
     Plunnecke-Ruzsa theory of sumset estimates, but replace finite sets by
     random variables, and cardinality by entropy.  The graph-theoretical Plunnecke
     inequalities no longer seem to hold in this setting, but almost all the
     remaining theory works, and is in fact somewhat cleaner (all constant
     factors in the estimates automatically get replaced by 1, thanks to an
     elegant “tensor power trick” of Ruzsa).  In fact the theory falls out neatly from
     standard Shannon entropy inequalities,
     most notably the submodularity inequality. 
     The same techniques also give some of the sharpest known results
     (due to Katz and Tardos) on the Kakeya conjecture and the Erdos distance
     problem, which we also discuss here.
- Uploaded:
     “Compressions, convex
     geometry, and the Freiman-Bilu theorem”, joint with Ben Green.  This short paper more or less fell out
     of a day’s worth of discussion at MIT, much to the surprise of both
     of us.  Here what we do is we
     observe that the Brunn-Minkowski theorem, combined with the standard
     technique of compressions (as used for instance by Bollobas and Leader) show
     that any suitably large subset of a d-dimensional discrete box must have
     doubling constant at least 2^d (minus epsilon).  This, when combined with Chang’s
     version of Freiman’s theorem, gives a reasonable quantitative
     version of the Freiman-Bilu theorem. 
     To oversimplify a little, we show that any set |A| of integers of
     doubling constant K can be covered by exp( C K^3 log^3 K ) translates of a
     progression of dimension at most log K and size at most |A|.  For comparison, the original arguments
     of Freiman and Bilu require more like exp(exp(K^C)) translates, while
     Chang’s theorem uses only one translate but the progression has
     dimension K and size exp( C K^2 log^3 K ) |A|.  We also observe a variant, that any
     subset of R^d which contains a d-dimensional nondegenerate parallelepiped
     must have doubling constant at least 2^{d/2} (which is pretty sharp, as
     one cannot do better than (3/2)^d).
- Uploaded:
     “Velocity Averaging,
     Kinetic Formulations and Regularizing Effects in Quasilinear PDEs”,
     joint with Eitan
     Tadmor.   In this project (which
     we started six years ago!) we revisit the averaging lemma which provides a
     smoothing effect for the velocity averages of (entropy) solutions to
     kinetic formulations of quasilinear equations.  The main novelty here is an abstract
     approach which allows us for the first time to obtain the smoothing effect
     for second order (convection-diffusion or convection-elliptic) systems,
     going beyond the first order conservation law smoothing effects in
     existing versions of the lemma.  The
     key is that for each fixed “velocity” v, the multiplier
     associated to v, being a combination of a real elliptic quadratic and an
     imaginary linear term obeys a certain “truncation property”
     uniformly in v which allows one to obtain a good Littlewood-Paley theory
     relative to that multiplier. Combined with the standard non-degeneracy
     condition (namely, that for any fixed frequency, the multiplier associated
     to v is large for most velocities v) we obtain the regularizing
     effect.  In the case of higher order
     vanishing of the symbol we obtain slightly better exponents than in
     preceding results even for the first-order case (pure transport).
Oct 26, 2005
 - Uploaded:
     “Maximal multilinear
     operators”, joint with Ciprian Demeter and Christoph Thiele.  In this paper we generalize a result of
     Michael Lacey concerning the bilinear maximal operator, to now obtain
     similar L^p bounds for arbitrary multilinear maximal operators, both in
     the context of functions on R and (via a simple transference principle)
     functions on a measure-preserving system, though the strength of the
     estimates depends greatly on the rank of the multilinear averaging (the
     main results are in the rank one case, the remaining ones being obtained
     by interpolation).  We follow
     Lacey’s approach with some additional simplifications.  Firstly we use now-standard multilinear
     interpolation techniques and time-frequency discretization techniques to
     reduce to a certain restricted weak type estimate for a model
     operator.  Using standard tree
     selection algorithms, the problem reduces to that of establishing a
     maximal Bessel inequality for a forest of wave packets, with a certain
     logarithmic loss.  By using a
     certain “good lambda” type argument one can make this
     logarithmic loss depend only on the tree multiplicity.  By spatial localization arguments (and a
     nonmaximal Bessel inequality, which has some independent interest) and the
     Radamacher-Menshov inequality we can normalize the trees to have common
     spatial interval, at which point the claim follows from an inequality of
     Bourgain.
Sep 23, 2005
 - Uploaded:
     “New bounds for
     Szemeredi's Theorem, I: Progressions of length 4 in finite field
     geometries”, joint with Ben Green.  This paper was in fact my first joint
     project with Ben, and our first approach to this problem eventually
     revolved around finding versions of Szemeredi’s theorem relative to
     certain quadratic varieties, which we treated as being rather pseudorandom
     objects (at least from a linear perspective).  Then Ben observed that having a version
     of Szemeredi’s theorem relative to pseudorandom sets would be very
     useful for obtaining progressions in the primes, and so we dropped this
     project to study
     primes.  Two years later, we
     finally returned to the problem. 
     Actually things turned out rather well, because in the meantime Ben
     came up with a better proof of the main result (being a hybrid of energy
     incrementation and density increment arguments, rather than a pure density
     increment argument, inspired by some arguments of Heath-Brown and
     Szemeredi, and avoiding the need to relativize to quadratic
     varieties).  Ultimately, the goal is
     to show that given any finite abelian group G, any subset of density at
     least 1/(log |G|)^c for some small absolute constant c > 0 contains
     nontrivial progressions of length 4 (the result for 3 is already known,
     for instance by Bourgain’s proof of Roth’s theorem).  Note that c=1 would essentially yield
     the Erdos-Turan conjecture for length 4 progressions (this is unknown even
     for length 3 progressions, where one essentially only has c=1/2, except in
     the finite field geometry case where one does have c=1).  In a previous paper we
     established 1/(log log |G|)^c by developing an inverse U^3 theorem.    In this paper we establish the 1/(log
     |G|)^c bound in the special case of a finite field geometry, i.e. G is a
     vector space over a small finite field (e.g. F_5).  The idea is to first combine the inverse
     U^3 theorem with the energy incrementation argument from our  primes paper to
     obtain a Koopman-von Neumann structure theorem that decomposes any
     function into a structured component of low quadratic complexity, and an
     error which is quadratically uniform. 
     Discarding the error, the task is thus to count progressions of
     length 4 in the structured component, which is a function depending only
     on a bounded number of linear and quadratic phase functions.  This can be done by linear Fourier
     analysis (and Gauss sum computations), and eventually leads to a
     substantial density increment on the function if it has an anomalous
     number of progressions of length 4. 
     This is then fed back into the density increment argument to obtain
     the 1/(log |G|)^c bound.  The final
     computation (which is reminiscent of Fourier analysis on 2-step nilmanifolds)
     is a little messy, but by being more low-tech one can obtain a “cheap”
     bound of the form exp( - c sqrt(log log |G|) ) without much difficulty, by
     ignoring almost all of the quadratic structure (slicing the quadratic
     variety right away into linear spaces).
- Updated
     again: “Sharp
     well-posedness and ill-posedness results for a quadratic non-linear
     Schr\"odinger equation”, joint with Ioan Bejenaru.  A rather confusing typo (the weight w
     was reported as min(1, -tau)^10 instead of max(1,-tau)^10) has been
     fixed.  Thanks to Hideo Takaoka for
     pointing out the error.
Sep 14, 2005
Sep 12, 2005
 - Uploaded:
     “On the Multilinear
     Restriction and Kakeya conjectures”, joint with Jon Bennett and Tony
     Carbery.  This marks my return
     to restriction and Kakeya problems after an absence of several years.  Here we investigate the n-linear
     analogues of the restriction and Kakeya conjectures in n dimensions, where
     there is a surprising amount of self-similarity in the problem, and where
     the role of curvature completely vanishes, being replaced instead by the
     concept of transversality.  Here,
     the induction on scale machinery of Bourgain and Wolff can be used to show
     that the multilinear Kakeya and restriction conjectures are equivalent up to epsilons, and that
     the Kakeya conjecture implies itself in a certain scale-inductive
     sense.  This latter observation can
     be refined using the heat flow monotonicity methods that were already
     known to work for the Brascamp-Lieb problem, thus yielding a heat-flow
     proof of the multilinear Kakeya conjecture and thence the multilinear
     restriction conjecture, up to epsilons. 
     Sadly, this does not resolve the linear versions of these
     conjectures, though it does eliminate a certain “non-plany”
     scenario and also has an application to a combinatorial analogue of the
     Kakeya problem, namely the joints problem of Sharir.  We also discuss some variable
     coefficient and k-plane analogues that can be treated by the method.  One unusual feature of the approach is
     the reliance of identities involving a general exponent p which are
     difficult to prove directly, but which we establish by first proving the
     identities when p is a positive integer and then extrapolating (via
     analytic continuation and approximation theory) to general p (in fact, we
     need these identities for p < 1, usually).
Aug 20, 2005
Aug 18, 2005
 - Uploaded:
     “Progress on nonlinear wave
     equations”.  These are
     some rather informal slides on progress on critical geometric nonlinear
     wave equations for the upcoming Cambridge
     conference.
- Updated:
     “A quantitative proof of
     Wiener’s theorem”. 
     As pointed out by Brian Davies, there is a very short quantitative
     proof of Wiener’s theorem due to Newman which is somewhat along the
     lines of the arguments in this note.
Aug 11, 2005
 - Uploaded:
     “Sharp well-posedness
     and ill-posedness results for a quadratic non-linear Schr\"odinger
     equation”, joint with Ioan Bejenaru.  Here we revisit a simple nonlinear
     Schrodinger equation, namely the 1D quadratic NLS i u_t + u_xx = u^2.  This equation was shown to be locally
     well-posed in H^s(R) for s > -3/4 by Kenig, Ponce and Vega using the X^{s,b} spaces,
     and it was later shown by Nakanishi, Takaoka, and Tsutsumi that this was the
     limit of the method (except for a Besov endpoint obtained by Muramutu and
     Taoka).  However, an inspection of
     iterates suggested that it was possible to attain and even go below -3/4
     by using other spaces than the X^{s,b} ones.  After some trial and error, Ioan and I
     found a suitable (though somewhat complicated) choice of spaces which
     enabled us to obtain local well-posedness down to s > -1.  Here, the iterates suggested that this
     was the sharp limit, and in fact there is an abstract argument (which works
     for any nonlinear equation containing a suitable “high-to-low
     cascade”, and may be of independent interest) that shows that
     illposedness occurs at s=-1 or below. 
     
- Updated:
     “Error Correction
     via Linear Programming”. 
     Due to a merge, this article is now somewhat longer and is joint
     with Mark Rudelson
     and Roman Vershynin
     as well as Emmanuel
     Candes.
- Uploaded:
     The short story “Quadratic
     reciprocity and the theta function”.  This is a completely classical
     derivation of Gauss’s reciprocity formula via the functional
     equation for the Jacobi theta function. 
     It is hardly new, and was written mostly for my own benefit, but
     this cute little proof seems to be not as widely known (except to number
     theorists) as it should be, so I’m putting it on-line anyway.
June 30, 2005
 - Uploaded:
     “Stability of
     energy-critical nonlinear Schr\"odinger equations in high dimensions”,
     joint with Monica Visan.  This paper arose from Monica’s
     investigation of the global well-posedness theory for defocusing energy
     critical NLS for large energy data in high dimensions.  It turns out in order to complete the
     global theory successfully, one had to go back and revisit the local
     well-posedness theory, and in particular obtain a stability result (that
     near-solutions to NLS can always be approximated by genuine solutions
     provided a certain spacetime norm is small) for the equation.  This is a fairly routine application of
     Strichartz estimates in six and fewer dimensions, but in seven and higher
     dimensions the nonlinearity becomes subquadratic, and so the derivative of
     the nonlinearity becomes non-Lipschitz; this causes a severe difficulty in
     iterating in the usual H^1 Strichartz spaces (and it is likely that the
     solution map is not even Lipschitz in these spaces for sufficiently high
     dimension).  However, it is possible
     to recover Holder continuity of the solution map, as well as a stability
     result with similar Holder continuous dependence on the perturbation.  The trick is to iterate not in the usual
     Strichartz spaces (which cost a full derivative and thus sacrifice the
     Lipschitz nature of the nonlinearity), but instead to iterate in exotic
     (non-admissible) Strichartz spaces, which have the same scaling as the
     usual spaces but require only a fraction of a derivative to be applied to
     the nonlinearity.  One in fact
     obtains Lipschitz bounds in the exotic spaces, which after further
     iteration then recovers estimates of Holder continuity type in the
     original Strichartz spaces.  In
     particular, we obtain for the first time uniform local well-posedness for the energy-critical NLS in
     higher dimensions (previously, the continuous dependence on the data was
     either not known to be uniform, or was only known to be uniform in rougher
     topologies than the energy topology). 
     As a (rather standard) consequence we also obtain scattering
     results (with bicontinuous wave operators) provided that an a priori spacetime bound is established
     on finite energy solutions.  (In the
     defocusing energy-critical NLS, this bound, which also establishes global
     well-posedness and regularity for this equation, will be forthcoming in
     Monica’s thesis.)
June 8, 2005
 - Uploaded:
     The expository note “A
     trilinear maximal inequality via arithmetic combinatorics”,
     joint with Christoph Thiele
     and Ciprian Demeter.  Here we use (of all things) arithmetic
     combinatorics in order to obtain a non-trivial estimate on a trilinear
     maximal estimate (reminiscent of the bilinear maximal function estimate of
     Michael Lacey).  It only gains an
     epsilon over the trivial estimate, however it has the unusual feature of
     being completely combinatorial in nature, using no Fourier analysis.  There is perhaps a hope that this type
     of technique can help give an estimate on the trilinear Hilbert transform,
     which appears to be inaccessible to a purely time-frequency approach,
     though we emphasize that the maximal operator here is somewhat simpler
     than that transform.
June 6, 2005
 - Created:
     A book page, showing the status of my various
     book projects.  Most of it is
     links to other things already on my web page, but the one new thing is
     some sample chapters from my upcoming
     CBMS lecture notes on nonlinear dispersive and wave equations.
June 4, 2005
 - Uploaded:
     “The Dantzig
     selector: statistical estimation when $p$ is much larger than $n$”,
     joint with Emmanuel Candes.  This continues my sequence of papers
     with Emmanuel exploring the applications of the l^1 minimization method,
     in this case to linear regression models in multivariant statistics.  The setup is that there are p unknowns
     (represented as a vector x in R^p), and a (known) linear measurement n x p
     matrix A.  We observe n measurements
     y = Ax + z, where z (in R^n) is a noise term with a known power (in
     particular, one has a bound on A^t z). 
     The objective is then to obtain an estimator for x from y which has
     as small error (as measured in the mean square sense) as possible.  The novelty here is that we treat the
     case p >> n, when there are many more unknowns than
     measurements.  Of course, this
     problem is in general unsolvable, but with the additional assumptions that
     x is sparse (supported on
     somewhat fewer than n values) or compressible
     (concentrated in the span of fewer than n vectors in a fixed basis) and
     that A is suitably generic
     – or more precisely, that it obeys the “uniform uncertainty
     principle” introduced in our earlier papers – then we do have
     a very efficient estimator, which we dub the “Dantzig
     selector” in honour of the late George Dantzig, as it is a linear
     programming method.  In fact the
     selector is very simple: we choose the estimator x^ to minimize the l^1
     norm of y-Ax^ subject to the constraint that the residual A^t z is less
     than the expected noise level in l^infty. 
     This estimator turns out to be a sort of nonlinear soft
     thresholding for x, and we obtain mean square error estimates that are
     with an absolute constant (256, to be precise) of the ideal mean square
     error.  (In numerical simulations,
     the ratio between actual and ideal MSE is more like 20).  Our arguments are deterministic and
     elementary, based on the deterministic geometry of UUP matrices.  We also have a refined version of the
     Dantzig selector, which we call the “Gauss-Dantzig selector”,
     which compensates for the underestimation bias of the Dantzig selector and
     can be thought of as a nonlinear hard thresholding for x rather than a
     soft thresholding for x.  In
     numerical experiments, this refined selector gives an MSE which attains
     the ideal MSE almost exactly; basically, it uses the ordinary Dantzig
     selector to locate the significant unknowns (usually it locates the exact
     support of the true unknowns x with near-perfect accuracy) and then uses a
     standard Gauss least-squares method to obtain the ideal selector based on
     those unknowns.  The method seems to
     be quite robust and general; the only catch is that it requires the noise
     to be uniformly distributed with respect to the measurement matrix, in the
     sense that good l^infty control is assumed on A^t z; this is generically
     true up to a logarithmic loss in the number of unknowns p.  (Our previous paper,
     with Justin Romberg, explored
     the case of adversarial noise, in which the only assumption on the noise
     is l^2 bounds.  There, the ideal MSE
     is of course significantly worse, but a slight variant of the Dantzig
     selector still attains this MSE up to an absolute constant.)   We also give a number of demonstrates
     of the method to recovery of compressible signals (images, spike trains,
     piecewise polynomial signals) from noisy and incomplete measurements.
May 31, 2005
 - Uploaded:
     “Finiteness bounds
     for Holder-Brascamp-Lieb multilinear inequalities”, joint with
     Jon Bennett, Tony
     Carbery, and Michael
     Christ, submitted, Math. Res.
     Letters.  Here, we give an
     elementary proof of one of the results in our larger paper, namely a
     necessary and sufficient condition for a multilinear inequality of
     Holder-Brascamp-Lieb type (possibly with localization or discretization)
     to hold with a finite constant.  As
     it turns out, such an inequality can be deduced purely from an iterative
     argument using nothing more than Holder’s inequality (as the base
     case), and either the Fubini-Tonelli theorem, Minkowski’s inequality
     or the multilinear interpolation theorem (depending on whether a critical
     space exists or not, and whether one of the exponents is equal to 1).  This generalizes the familiar proof of
     Young’s inequality (or the Loomis-Whitney inequality, including its
     generalizations to arbitrary measure spaces and co-ordinate planes by
     Finner) using exactly the same tools. 
     
May 18, 2005
 - Uploaded:
     “Obstructions to
     uniformity, and arithmetic patterns in the primes”, submitted, Quarterly J. Pure Appl. Math.
     (special issue in honour of John Coates). 
     In this survey I discuss the approach Ben Green and I have to
     finding arithmetic progressions (and hopefully other patterns too) in the
     prime numbers and similar such sets. 
     In particular, I discuss how we use enveloping sieves, such as those
     constructed from truncated divisor sums by Goldston and Yildirim, to view
     the primes as a set of positive relative density, as opposed to one of
     zero absolute density, and then how one can hope to count patterns in the
     primes by first quantifying the obstructions there are to uniformity (in
     the sense of Gowers), and then averaging out the effect of those
     obstructions.  There is a
     distinction between the “hard” obstructions to uniformity
     coming from Fourier analysis, and more recently from the quadratic Fourier
     analytic techniques of Gowers (as well as the closely related work from
     ergodic theory from Bergelson, Host, Kra, Ziegler, and others), and the
     “soft” obstructions coming from the earlier work of
     Furstenberg, Katznelson, Weiss, and others; the situation is still not
     completely clarified yet but there are many intriguing interconnections
     between harmonic analysis, combinatorics, ergodic theory, group theory,
     and number theory being unearthed.
May 14, 2005
 - Updated:
     the short story “An
     ergodic transference theorem”. 
     I have completely rewritten and expanded this article, using a more
     abstract viewpoint and also going into greater detail into the Furstenberg
     correspondence principle.  In fact
     we manage to derive this principle without any appeal to the axiom of
     choice.
May 9, 2005
 - Uploaded:
     “Random symmetric
     matrices are almost surely non-singular”, joint with Kevin
     Costello and Van Vu,
     submitted, Duke Math. J..  Here we adapt an argument of Komlos to
     show that a random symmetric n x
     n matrix with entries +-1, distributed independently on the upper
     triangular half of the matrix (and then fixed by symmetry on the strict
     lower triangular component), is invertible with probability 1 – O(
     n^{-1/8+eps} ) for any eps > 0. 
     The main difficulty here is that the symmetry couples different
     entries together, making the standard incremental approach of exposing one
     row of the matrix at a time rather intractable.  However if one adopts Komlos’
     original approach of exposing a row and column simultaneously then one can proceed.  The main difficulty that needs to be
     overcome is that the exposing of both a row and a column changes the
     determinant by a random quadratic form, and one needs a quadratic
     Littlewood-Offord inequality to show that this form is usually
     non-zero.  We establish such an
     inequality using a general Cauchy-Schwarz approach (reminiscent of
     Weyl’s approach to exponential sums), which may be of interest for
     other applications.
- Uploaded:
     The expository note “Santalo’s
     inequality”.  This is another
     “deleted scene” from my additive combinatorics book with Van
     Vu, which we are still in the process of editing (in particular, removing
     a certain amount of tangential material). 
     Here we give the proof of Santalo’s inequality giving an
     upper bound for the product of a volume of a symmetric convex body and its
     polar.  We also discuss the known
     lower bounds, giving in particular Kuperberg’s simple (though
     non-optimal) result. 
May 2, 2005
 - Uploaded:
     “The Brascamp--Lieb
     Inequalities: Finiteness, Structure and Extremals”, joint with
     Jon Bennett, Tony
     Carbery, and Michael
     Christ, submitted, Inventiones Math..  In this paper we revisit the
     general-rank, higher-dimensional Brascamp-Lieb inequalities, which
     generalize several useful inequalities in analysis such as Young’s
     inequality, the Loomis-Whitney inequality, and Holder’s
     inequality.  We present a new,
     elementary, and completely self-contained proof of these inequalities with
     the sharp constant using a multilinear heat flow approach (which we
     discovered while working on the Kakeya problem, but that’s another
     story, to be told in a subsequent paper), and also give necessary and
     sufficient conditions for the constant to be finite, and for the constant
     to be attainable.  It turns out that
     there is a lot of geometric structure in these inequalities, involving
     quiver representations, of all things, and a factorization of
     Brascamp-Lieb “data” into indecomposable or simple components,
     as well as a normal form, turns out to be essential to understanding them.
     We also obtain some curious monotonicity formulae for the heat equation
     (as well as some other convolution operators) as a by-product.  This is also my first paper which
     involves commutative diagrams (five of them, to be exact).  Correction
     May 19: As pointed out to me by a co-author who shall remain
     anonymous, this is in fact my second paper involving commutative
     diagrams.  The first one can be
     found here.
Apr 23, 2005
Apr 8, 2005
 - Uploaded:
     “The
     Gaussian primes contain arbitrarily shaped constellations”,
     submitted, J. d’Analyse Mathematique.  This is a totally reworked version of my
     preceding paper (the one with the impossibly long title).  Here we take the argument of Ben Green
     and myself that establishes long arithmetic
     progressions in the primes, and extends it to the Gaussian integers,
     thus showing that the Gaussian primes contain constellations of any given
     shape and orientation.  The main new
     difficulty is that Szemeredi’s theorem, which was a major input in
     my paper with Ben, must now be replaced by another result.  The obvious candidate here is the
     Furstenberg-Katznelson multidimensional Szemeredi theorem.  However, I do not know how to run the
     rest of the argument with this input, mainly because I do not understand
     the characteristic factor for multidimensional Szemeredi recurrence (this
     is in fact an active area of research in ergodic theory at present).  Instead, I use the hypergraph removal
     lemma of Gowers and Nagle-Rodl-Schacht-Skokan (or more precisely the
     variant introduced in my
     preceding paper) as the tool used to control the “Gowers
     anti-uniform” component of the Gaussian primes.  The remaining parts of the proof closely
     follow the paper with Ben, thus we develop a generalized von Neumann
     theorem to deal with the “Gowers uniform” component of the
     primes, and a “Furstenberg structure theorem” to split the
     primes into the Gowers anti-uniform and Gowers uniform components.  A large portion of the paper is devoted
     to establishing a pseudorandom majorant for the Gaussian primes, which
     requires a Goldston-Yildirim type analysis of Gaussian truncated divisor
     sums.  Here we encounter some new
     technical difficulties, arising from some correlations in the Gaussian
     primes that are not present in the rational setting. 
- Uploaded:
     “Error Correction via Linear
     Programming”, joint with Emmanuel Candes,
     submitted, FOCS ’05 proceedings. 
     This is a condensed version of our linear coding paper,
     giving an elementary proof of the main result, namely that a system of
     random overdetermined measurements can correct a fixed fraction of errors
     perfectly using the basis pursuit method.
Mar 28, 2005
 - Uploaded:
     the short story “An
     ergodic transference theorem”. 
     This is an attempt to redo the transference theorem of Ben Green
     and myself, which can be used (among other things), to derive long
     arithmetic progressions in the primes, by “soft” ergodic
     methods, and in particular by reworking the Furstenberg correspondence
     principle.  Interestingly, the
     ergodic methods can partially achieve this, but the catch is that one
     requires arbitrarily many linear forms conditions, instead of finitely
     many.  As such, one does not recover
     our result about long arithmetic progressions in the primes, although one
     does get some weaker results, such as a Szemeredi theorem for subsets of
     positive _relative_ density of a random set of {1,…,N} of density
     1/log N.  I do not know whether this
     need for infinitely many linear forms condition is an inherent feature of
     the soft ergodic method, or if it can be somehow overcome.
Mar 24, 2005
 - Uploaded:
     “A variant of the
     hypergraph removal lemma”, submitted, J.
     Combin. Thy..  This is the first
     half of the new version of my Gaussian constellations paper, which will
     now appear in two separate papers; the second paper will appear
     shortly.   Here we reprove the
     hypergraph removal lemma of Gowers and Nagle-Rodl-Schacht-Skokan, with a
     slight extra twist (which is needed in order for a transference result in
     the second paper to work properly). 
     We take an “ergodic” perspective to this lemma, viewing
     the problem as one of taking an arbitrary set and decomposing it into a
     “compact” part which is of “bounded complexity”,
     and a “weakly mixing part” which is pseudorandom (epsilon
     regular).  The version of the
     Szemeredi regularity lemma that we need is thus similar in spirit to the
     Furstenberg structure theorem.  As
     usual we need a counting lemma, which is the usual careful application of
     Cauchy-Schwarz.  One feature of the
     regularity lemma (which is shared in common with a recent paper of Rodl
     and Schacht on the same issue) is that we do not obtain regularity on the
     original hypergraph directly, but instead work with another hypergraph
     which is a small modification of the original graph (in Rodl-Schacht, they
     add or remove a small number of edges; here we allow the weight function
     of the hypergraph to be perturbed in L^2). 
     This allows us to obtain excellent regularity control on the new
     hypergraph, which greatly facilitates the counting lemma.  At 23 pages, this paper is then one of
     the shortest elementary proofs available of the hypergraph regularity
     lemma and its consequences such as Szemeredi’s theorem and the
     Furstenberg-Katznelson theorem.
Mar 17, 2005
 - Uploaded:
     “Some analysis at UCLA”.  This is a very short (10 min)
     presentation to incoming UCLA grad students.  I try to sketch out the proof of long
     arithmetic progressions in the primes, but clearly with the time
     constraints it is only the briefest of sketches.
Mar 3, 2005
 - Uploaded:
     “Stable Signal Recovery
     from Incomplete and Inaccurate Measurements”, with Emmanuel Candes and Justin Romberg.  This is the latest in our sequence of
     papers exploring the basis pursuit method (l^1 minimization) to
     reconstruct sparse or compressible data from a sufficiently
     “skew” measurement ensemble. 
     Here we address an issue not treated in the earlier papers, namely
     that of robustness – what happens to the reconstruction method when
     non-sparse noise is introduced into the measurements (e.g. roundoff
     error).  As it turns out, the most
     naïve relaxation method in the basis pursuit method – allowing the
     minimizer to approximate the measurements rather than match it exactly
     – gives a surprisingly robust algorithm, indeed the error incurred
     at the end is at most a constant times the best possible error one could
     achieve with any algorithm, assuming that the measurement matrix obeys an
     restricted orthonormality property which we have called the uniform
     uncertainty principle.  Our
     arguments are simple and in fact give some slight improvements over our
     earlier results in the error-free case, for instance increasing the
     corruption rate we can tolerate in linear coding by a modest amount.
Mar 1, 2005
 - Uploaded:
     “An inverse
     theorem for the Gowers $U^3(G)$ norm”, joint with Ben Green, submitted, Proc.
     Edin. Math. Soc..  This paper
     systematically studies the quadratic (d=3) case of the Gowers uniformity
     norm U^d(G), which can be defined for any finite abelian group G but was
     introduced by Gowers for the cyclic groups G = Z/NZ.  (A closely related norm U^d(X) for
     measure-preserving systems X was studied by Host and Kra).  This norm is given by an explicit
     formula (an average over d-dimensional cubes), and informally measures the
     extent to which the phase of a function on G behaves like a polynomial of
     degree d-1 or less (or in other words the d’th
     “derivative” of the phase behaves like zero).  Thus for instance the U^3(G) norm of f
     is large when the phase of f behaves quadratically.  These norms are of fundamental
     importance in Gowers’ approach to obtaining arithmetic progressions
     in dense sets, and also in my paper with Ben
     establishing long arithmetic progressions in the primes.  When the U^d norm are small, one can
     accurately compute the number of progressions of length d+1 in an
     associated set; thus a small U^3(G) norm is useful for counting
     progressions of length 4.  However,
     it is not always easy to determine when a set has small U^d norm.  In a paper of Gowers it was shown that
     in order for the U^3(Z_N) norm of a bounded function to be small, it
     suffices for that function to have no “local quadratic bias”
     – that on most short progressions the function does not correlate
     strongly with a quadratic phase function; this is the essential step in
     Gowers’ proof of Szemeredi’s theorem for progressions of
     length 4.  Here we extend and
     globalize this analysis, showing that in order for the U^3(G) norm of a
     bounded function to be small, it is necessary and sufficient that the
     function have no “global quadratic bias” – that it does
     not correlate with a globally quadratic phase function.  (Actually, this is an oversimplification;
     one also has to check a certain class of “almost global quadratic
     phase functions”, which are quadratic on a large Bohr set).  This gives a new proof of
     Szemeredi’s theorem for progressions of length 4 and arbitrary
     groups G of order coprime to 2 and 3, with a reasonably good bound (any
     subset of G with density 1 / (log log |G|)^c will have a progression of
     length 4).  In a future paper we
     shall also apply this inverse theorem (or more precisely a generalization
     of it to pseudorandom measures) to establish that the primes contain the
     expected number of progressions of length 4.  Our main result can be viewed as
     presenting a complete list of obstructions in order for a bounded function
     to be quadratically uniform (i.e. have small U^3(G)) norm.  There are some other lists of
     obstructions known – namely generalized quadratic phase functions,
     as well as 2-step nilsequences; in the second half of the paper we show
     how all of these lists of obstructions are related, and give some new
     quantitative estimates on the amount of obstruction presented by these
     lists.
Feb 15, 2005
 - Uploaded:
     “Decoding by Linear
     Programming”, joint with Emmanuel Candes,
     submitted.  This is an application of
     the uniform uncertainty principle (URP) developed in our previous paper.  Here we show that any ensemble of linear
     measurements (i.e. a matrix) which obeys the UUP (which basically asserts
     that any sparse set of rows in this matrix behaves like an almost
     orthonormal system), is a good matrix for linear coding and decoding.  In particular we can use such a matrix
     as a linear code (but over the reals or integers, not over a finite
     field), and can reconstruct the plaintext from the ciphertext even if a
     constant fraction of entries of the ciphertext are deleted.  Also this shows that the UUP implies the
     exact reconstruction property (ERP) also discussed in our previous papers,
     and in fact implies a deterministic and uniform version of this property,
     in that all sparse signals are
     reconstructed exactly.  In our
     previous paper we showed that various standard measurement ensembles
     (Gaussian, Bernoulli, randomly selected Fourier) obeyed the UUP with high
     probability, and thus for instance Gaussian linear codes can now be
     decoded via a simple linear programming algorithm even if a constant
     fraction (we obtain 0.3% in theory, but it seems closer to 30% in
     practice) of the ciphertext is corrupted.
Jan 27, 2005
 - Temporarily withdrawn: The paper
     “Information
     theory, relative versions of the hypergraph regularity and removal lemmas,
     the Szemer\'edi-Furstenberg-Katznelson theorem, and prime constellations
     in number fields”.  There
     are a number of minor problems with this paper, and one moderately serious
     one, which I need to address. 
     Normally I would simply correct these errors and post an updated
     version, but these issues will require a bit of time on my part, and the
     current version contains a number of steps which are phrased in an
     awkward, misleading, slightly wrong, or (in two places) extremely wrong
     manner.  I anticipate to replace the
     paper with a cleaner, corrected version in a few weeks.
Jan 19, 2005
 - Uploaded:
     “On the
     singularity probability of random Bernoulli matrices”, with Van Vu, submitted to  J. Amer.
     Math. Soc..  In this paper we
     consider the probability that a random n x n matrix whose entries are +-1
     is singular (has determinant zero). 
     This probability is conjectured to be (1/2 + o(1))^n; a famous
     paper of Kahn, Komlos, and Szemeredi showed 0.999^n, which we refined in an earlier paper
     to 0.939^n.  Here we push things a
     bit further, to (3/4 + o(1))^n, which seems to be the limit of the
     method.  Basically the point is to
     replace the discrete unit cube (where the original problem is set) with a
     weighted variant on which there is some further concentration on
     hyperplanes; this weighted variant stops being better than the original
     cube at ¾, which is why the method stops. 
     A new innovation is to use some additive combinatorics (Fourier
     analysis and Freiman’s theorem) to identify which hyperplanes are
     exceptional with respect to this concentration, and count these
     hyperplanes directly.
- Uploaded:
     “Information
     theory, relative versions of the hypergraph regularity and removal lemmas,
     the Szemer\'edi-Furstenberg-Katznelson theorem, and prime constellations
     in number fields”, submitted to Journal d'Analyse
     Mathématique.  As the title
     suggests, this paper may take a little time to explain.  In my paper with Ben
     Green establishing arbitrarily long progressions in the primes, we
     started with Szemeredi’s theorem on arithmetic progressions, and
     used quantitative ergodic methods to then deduce a relative version of
     that theorem, working on a pseudorandom set.  This then implied progressions in the
     primes.  In this paper we take a
     somewhat stronger theorem than the Szemeredi theorem, namely the
     Szemeredi-Furstenberg-Katznelson theorem on constellations in
     multidimensional sets (or equivalently, multiple recurrence for commuting
     shifts on a measure-preserving system). 
     This in turn is implied by an even stronger theorem, namely the
     hypergraph removal lemma of Gowers and Rodl-Skokan.  This theorem in turn is deduced
     primarily by means of a hypergraph regularity lemma (also due to Gowers and
     Rodl-Skokan).  What we do here is give
     an information-theoretic proof of the hypergraph regularity lemma, by
     means of the Shannon entropy
     inequalities.  This then leads to a
     new proof of the hypergraph removal lemma, hence the
     Szemeredi-Furstenberg-Katznelson theorem, and hence of the Szemeredi
     theorem.  Furthermore, in this
     setting it is almost trivial to perturb the setup from a non-relative
     theorem to a relative theorem, in fact this amounts to making certain
     random variables “almost independent” instead of genuinely
     independent.  Because of this, we
     can establish a relative Szemeredi-Furstenberg-Katznelson theorem on a
     pseudorandom measure (and also reprove the relative Szemeredi theorem of
     Green and myself).  By invoking the
     Goldston-Yildirim analysis from our previous paper (with some simplifications
     introduced in my short story on
     this topic), we can then obtain multidimensional constellations in the
     primes (with one annoying caveat, which is too technical to describe
     here).  We also obtain a similar
     result for the Gaussian primes (the primes in the Gaussian integers); this
     time we do not have this technical caveat because the Gaussian almost
     primes are more pseudorandom than the Cartesian product of the rational
     almost primes.  We also speculate on
     whether there is some ergodic theoretic or measure theoretic
     interpretation of these results; they seem to suggest that the presence of
     the shift operator is in fact not necessary for recurrence theorems!
- Updated:
     The short story “An
     information-theoretic proof of Szemeredi’s regularity lemma”.  We now give two information-theoretic proofs of the regularity lemma, one
     based on an “energy increment argument” (and closer in spirit
     to Szemeredi’s original proof) and one based on an “entropy
     increment argument” which is broadly similar but fits much better
     into the framework of Shannon’s
     entropy inequalities.  This second
     proof served as a simplified model to the hypergraph version in the paper
     above, so it may well be worth reading this shorter and simpler paper
     first to motivate the more complicated paper above.