"Science, Faculty of"@en . "Mathematics, Department of"@en . "DSpace"@en . "UBCV"@en . "Wong, Erick Bryce"@en . "2012-08-09T17:47:19Z"@en . "2012"@en . "Doctor of Philosophy - PhD"@en . "University of British Columbia"@en . "We study questions in three arithmetic settings, each of which carries aspects of random-like behaviour.\n\nIn the setting of arithmetic functions, we establish mild conditions under which the tuple of multiplicative functions [f\u00E2\u0082\u0081, f\u00E2\u0082\u0082, \u00E2\u0080\u00A6, f_d ], evaluated at d consecutive integers n+1, \u00E2\u0080\u00A6, n+d, closely approximates points in R^d for a positive proportion of n; we obtain a further generalization which allows these functions to be composed with various arithmetic progressions.\n\nSecondly, we examine the eigenvalues of random integer matrices, showing that most matrices have no rational eigenvalues; we also identify the precise distributions of both real and rational eigenvalues in the 2 \u00C3\u0097 2 case.\n \nFinally, we consider the set S(k) of numbers represented by the quadratic form x\u00C2\u00B2 + ky\u00C2\u00B2, showing that it contains infinitely many strings of five consecutive integers under many choices of k; we also characterize exactly which numbers can appear as the difference of two consecutive values in S(k)."@en . "https://circle.library.ubc.ca/rest/handle/2429/42887?expand=metadata"@en . "Structure and Randomness in Arithmetic Settings by Erick Bryce Wong B.Sc., Simon Fraser University, 1994 M.Sc., Simon Fraser University, 1997 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY in The Faculty of Graduate Studies (Mathematics) THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver) August 2012 c\u00C2\u00A9 Erick Bryce Wong 2012 Abstract We study questions in three arithmetic settings, each of which carries as- pects of random-like behaviour. In the setting of arithmetic functions, we establish mild conditions under which the tuple of multiplicative functions [f1, f2, . . . , fd], evaluated at d consecutive integers n + 1, . . . , n + d, closely approximates points in Rd for a positive proportion of n; we obtain a fur- ther generalization which allows these functions to be composed with various arithmetic progressions. Secondly, we examine the eigenvalues of random integer matrices, show- ing that most matrices have no rational eigenvalues; we also identify the precise distributions of both real and rational eigenvalues in the 2\u00C3\u0097 2 case. Finally, we consider the set S(k) of numbers represented by the quadratic form x2 +ky2, showing that it contains infinitely many strings of five consec- utive integers under many choices of k; we also characterize exactly which numbers can appear as the difference of two consecutive values in S(k). ii Preface A version of Chapter 2 was published as: \u00E2\u0080\u00A2 Wong, E. B. (2008) Simultaneous approximation of reals by values of arithmetic functions. Anatomy of Integers, volume 46 of CRM Proc. Lecture Notes, pp. 289\u00E2\u0080\u0093297. Amer. Math. Soc., Providence, RI. [86] Chapter 3 is based on two published papers: \u00E2\u0080\u00A2 Martin, G. and Wong, E. (2008) The number of 2 \u00C3\u0097 2 integer matri- ces having a prescribed integer eigenvalue. Algebra Number Theory, 2(8):979\u00E2\u0080\u00931000. [55] \u00E2\u0080\u00A2 Martin, G. and Wong, E. B. (2009) Almost all integer matrices have no integer eigenvalues. Amer. Math. Monthly, 116(7):588\u00E2\u0080\u0093597. [54] The earliest manuscripts were drafted by Dr. Martin, based on my original research; thereafter, we shared equal responsibility for both writing and subsequent research. In addition, a version of Proposition 3.6 in Section 3.2 \u00E2\u0080\u009CBasic bounds on singularity\u00E2\u0080\u009D was published as Lemma 1.1 in: \u00E2\u0080\u00A2 Pataki, G., Tural M. and Wong, E. B. (2010) Basis reduction and the complexity of branch-and-bound. In Proceedings of the Twenty- First Annual ACM-SIAM Symposium on Discrete Algorithms, pages 1254\u00E2\u0080\u00931261, Philadelphia, PA. SIAM. [66] My primary contribution to this paper was the formulation and proof of this lemma. A version of this proof was submitted to SIAM J. Computing in an extended form of the above paper. iii Table of Contents Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . ix Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1 Setting the stage . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Modelling the primes . . . . . . . . . . . . . . . . . . . . . . 2 1.3 Other arithmetic structures . . . . . . . . . . . . . . . . . . . 4 1.4 Density and the primes . . . . . . . . . . . . . . . . . . . . . 5 1.5 Notation and outline . . . . . . . . . . . . . . . . . . . . . . 7 1.6 Arithmetic functions . . . . . . . . . . . . . . . . . . . . . . . 8 1.7 Distribution of an arithmetic function . . . . . . . . . . . . . 9 1.8 Approximation by arithmetic functions . . . . . . . . . . . . 11 1.9 Comparing arithmetic progressions . . . . . . . . . . . . . . . 12 1.10 Probability in number theory . . . . . . . . . . . . . . . . . . 13 1.11 Random matrices and eigenvalues . . . . . . . . . . . . . . . 15 1.12 Random integer matrices . . . . . . . . . . . . . . . . . . . . 16 iv Table of Contents 1.13 Singularity and computation . . . . . . . . . . . . . . . . . . 18 1.14 Rationality versus reality . . . . . . . . . . . . . . . . . . . . 19 1.15 Sums of two squares . . . . . . . . . . . . . . . . . . . . . . . 20 1.16 Patterns in sums of two squares . . . . . . . . . . . . . . . . 22 1.17 Patterns in quadratic forms . . . . . . . . . . . . . . . . . . . 24 2 Approximation by Arithmetic Functions . . . . . . . . . . . 26 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.2 Main theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 29 2.3 Generalization to linear functions . . . . . . . . . . . . . . . 36 3 Eigenvalues of Random Integer Matrices . . . . . . . . . . . 40 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 3.2 Basic bounds on singularity . . . . . . . . . . . . . . . . . . . 45 3.3 Eigenvalues and determinants . . . . . . . . . . . . . . . . . 48 3.4 Proof of main results for n \u00E2\u0089\u00A5 2 . . . . . . . . . . . . . . . . . 51 3.5 Preliminaries for 2\u00C3\u0097 2 matrices . . . . . . . . . . . . . . . . 53 3.6 Enumeration theorems for integer eigenvalues . . . . . . . . . 56 3.7 Proof of key proposition for enumeration . . . . . . . . . . . 62 3.8 Distribution of real eigenvalues . . . . . . . . . . . . . . . . . 68 3.9 The derivative of the distribution . . . . . . . . . . . . . . . 72 3.10 Further remarks . . . . . . . . . . . . . . . . . . . . . . . . . 75 4 Patterns in x2 + ky2 . . . . . . . . . . . . . . . . . . . . . . . . . 76 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 4.2 Local feasibility of runs . . . . . . . . . . . . . . . . . . . . . 78 4.3 Construction for k = 2 . . . . . . . . . . . . . . . . . . . . . 81 4.4 Construction for k > 2 . . . . . . . . . . . . . . . . . . . . . 82 4.5 How many values of k? . . . . . . . . . . . . . . . . . . . . . 84 4.6 Preliminaries for gaps . . . . . . . . . . . . . . . . . . . . . . 86 4.7 Proof of existence of gaps . . . . . . . . . . . . . . . . . . . . 88 5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 5.1 Arithmetic functions . . . . . . . . . . . . . . . . . . . . . . . 91 v Table of Contents 5.2 Random matrices . . . . . . . . . . . . . . . . . . . . . . . . 92 5.3 Patterns in quadratic forms . . . . . . . . . . . . . . . . . . . 94 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 vi List of Tables 4.1 Distribution of x2 + ky2 modulo 8 . . . . . . . . . . . . . . . 79 vii List of Figures 3.1 Distribution of real and rational eigenvalues . . . . . . . . . . 45 3.2 Cases 0 \u00E2\u0089\u00A4 \u00CE\u00B4 \u00E2\u0089\u00A4 1, 1 \u00E2\u0089\u00A4 \u00CE\u00B4 \u00E2\u0089\u00A4 \u00E2\u0088\u009A2, and \u00E2\u0088\u009A2 \u00E2\u0089\u00A4 \u00CE\u00B4 \u00E2\u0089\u00A4 2 . . . . . . . . . 72 viii Acknowledgements My immediate thanks go to my supervisor Greg Martin, for his careful read- ings of many drafts, and helpful responses to my panicked e-mails. Through many different turns he has been a wise mentor, an inspiring teacher, an affable colleague, and a tireless supporter. I also thank my committee members, Izabella Laba and Jo\u00CC\u0081zsef Solymosi, and my examiners, Igor Shparlinski, Mark Greenstreet, Michael Bennett and Gordon Semenoff, for the valuable contribution of their time and insights. It has been an unusually long journey from my Master\u00E2\u0080\u0099s degree to this point, shaped by countless individuals. I am truly grateful to all, though space permits me to mention but a few. I thank my family for their ever-present love and support. I feel quite fortunate to have always had the freedom to pursue my interests, however esoteric (let me also thank my nephew Paris for sharing in those interests). I recognize the generous assistance of the Natural Sciences and Engineer- ing Research Council of Canada for a portion of this work. I am grateful for the support of the MS/MRI Research Group, especially Andrew Riddehough for his personal encouragement. I owe many thanks to all the staff of the mathematics department, partic- ularly Lee Yupitun for her boundless knowledge of administrative vagaries. I am fortunate to have met people here who fuelled my mathematical curiosity in memorable ways; I am doubly fortunate to count them among my friends. In order of appearance: Vishaal Kapoor, Mike Tsiang, Florina Halasan, Nishant Chandgotia, Vasu Tewari and Paul Pollack. Finally, I wish to thank my wife Eva Koo for companionship, support and encouragement through good times and bad; for believing in me when I could not; for putting some structure in a random life. ix Dedication To my father, who almost made it to 90 years. Happy Birthday, Dad. x Chapter 1 Introduction 1.1 Setting the stage Within mathematics, the field of number theory is notorious for its ability to generate problems which are eminently plausible and well-supported by evidence, and yet for which a rigorous proof appears far beyond the reach of current technology. We would like to begin this chapter by attempting to convey to the reader why such a disparity should exist. One often finds in analytic number theory that the basic statistics of a particular object of interest can be understood quite well. At the same time, such an object seems to possess almost no tractable structure which serves to distinguish it from any other random process with the same statistics. Roughly speaking, so many of the things we are interested in behave as though they are random even though they are decidedly not so (for a concrete analogy, the digits of pi seem to vary randomly, but the 763rd digit of pi will always be 9 no matter how often one repeats the experiment). While it is entirely reasonable to make predictions about such objects, and these predictions are easily seen to be true for the vast majority of their random counterparts, this lack of structure leaves us with no easy way to certify that our particular object is not one of the unlucky few. A prime example to illustrate this dilemma is given by the familiar se- quence of prime numbers: 2, 3, 5, 7, 11, 13, 17, 19, and so on. The primes are in some sense defined by their lack of structure, and it is indeed a non- trivial task to predict the spacings between each prime number and the next. Yet the statistics of this sequence have been well-understood for more than a century. If we use the conventional notation pi(n) to denote the number of prime numbers between 1 and n, then the Prime Number Theorem (a 1 1.2. Modelling the primes triumph of late 19th century mathematics) states that pi(n) is asymptotic to n/ log n, that is lim n\u00E2\u0086\u0092\u00E2\u0088\u009Epi(n) / n log n = 1. As n gets progressively larger, n/ log n becomes a better approximation to pi(n), in the sense of relative error. Thus, a number chosen at random from the interval [1, n] bears a roughly 1 in log n chance of being prime (here, as in the entirety of this thesis, log n denotes the natural logarithm of n). 1.2 Modelling the primes The preceding observation gives rise naturally to the Crame\u00CC\u0081r model, which has been rather successful at making quantitative predictions about primes (though far from perfectly so: see [28, 52]). We describe a slightly simplified version of this model: fix a number n > 2, and choose a random subset P (n) of the numbers {1, 2, . . . , n} by selecting each element independently with probability 1/ log n, so that P (n) has n/ log n elements on average. The Crame\u00CC\u0081r model predicts that if a given pattern is highly likely to occur within P (n), then the same pattern should also occur with similar frequency in the primes up to n. The advantage of such a simple random model lies in the ease with which one can analyze the properties of P (n). The simplest type of pattern one could reasonably apply this model to is that of twin primes: pairs of primes like 71 and 73 which differ by exactly 2. In other words, we wish to understand how often it is that both m and m+2 are simultaneously prime. In the random model, the probability that m and m+ 2 are both included in P (n) is exactly 1/ log2 n for any 1 \u00E2\u0089\u00A4 m \u00E2\u0089\u00A4 n\u00E2\u0088\u0092 2, and so the expected number of \u00E2\u0080\u009Ctwins\u00E2\u0080\u009D in P (n) is (n\u00E2\u0088\u00922)/ log2 n, or roughly n/ log2 n. According to the Crame\u00CC\u0081r model, we should expect the true number of twin primes up to n to be roughly n/ log2 n. This prediction turns out to be quite good, with one significant caveat: notice that P (n) also contains consecutive pairs of the form {m,m+1} with the same frequency n/ log2 n. On the other hand, this pattern is nigh im- possible for primes, because primes are always odd (with the sole exception 2 1.2. Modelling the primes of the prime 2). The primes do possess a structure after all, at least in a negative sense: their lack of divisibility by 2 makes the pattern {m,m+ 1} impossible in the primes (aside from {2, 3}), however common it is in the random model. The same reasoning has a more subtle influence on the pat- tern {m,m + 2}: the frequency should be better than the random model predicts, because if m is prime then m+ 2 is certainly odd, and this makes it more likely to be prime than some random number which may equally be odd or even. This simple idea of examining a problem through the lens of \u00E2\u0080\u009Ceven\u00E2\u0080\u009D and \u00E2\u0080\u009Codd\u00E2\u0080\u009D (more generally, the remainders modulo a particular prime p or prime power pr) is common in number theory, and is referred to as a local argument. By adjusting the na\u00CC\u0088\u00C4\u00B1ve estimate n/ log2 n with a certain correction factor for each prime p (called a local factor), and combining all such factors into one constant, Hardy and Littlewood (in 1923, long before much computational evidence was available) obtained an asymptotic conjecture (called a heuristic) for the number of twin primes up to n: #{m \u00E2\u0089\u00A4 n : m,m+ 2 are prime} \u00E2\u0088\u00BC C2 \u00C2\u00B7 n log2 n , (1.1) where C2 has the approximate value 1.32032363. The Hardy\u00E2\u0080\u0093Littlewood conjecture for twin primes has shown strikingly accurate agreement1 with computational evidence. In [63], Nicely computed the exact number of twin primes up to 1014, and found the prediction to match to within one part in 2 million! (Incidentally, it was this computation which led to the discovery of the infamous \u00E2\u0080\u009CFDIV\u00E2\u0080\u009D division bug in the Intel Pentium processor.) However, we are very far from understanding how to prove that this estimate remains accurate for all larger values of n. In fact, it remains an open problem whether there are infinitely many twin primes, let alone that they occur with the exact frequency predicted by Hardy and Littlewood. 1To achieve this striking accuracy, one should first replace n/ log2 n by the almost identical integral \u00E2\u0088\u00AB n 2 (log t)\u00E2\u0088\u00922dt; this relates to the simplification we referred to earlier. 3 1.3. Other arithmetic structures 1.3 Other arithmetic structures The twin primes problem of the previous section can be viewed as a special case of the study of the gaps between consecutive primes. If we denote the sequence of primes in increasing order by p1, p2, p3, . . ., we may restate the twin primes conjecture as \u00E2\u0080\u009Chow often is pn+1\u00E2\u0088\u0092pn = 2?\u00E2\u0080\u009D. More generally, one could naturally ask what is known about the spacing dn = pn+1\u00E2\u0088\u0092 pn? (The desire to understand gaps between primes was the motivation for Crame\u00CC\u0081r introducing his random model.) The Prime Number Theorem can easily be used to deduce that, on average, dn has size roughly equal to log n. While the state of our knowledge is insufficient to conclude that dn = 2 infinitely often, it is a remarkable achievement of Goldston, Pintz and Y\u00C4\u00B1ld\u00C4\u00B1r\u00C4\u00B1m [26] that dn can be arbitrarily small compared to its average value: for any \u000F > 0, there are infinitely many n for which dn < \u000F log n. One might also ask for longer patterns than twin primes. Note that it is unreasonable to expect {m,m + 2,m + 4} to all be prime for local reasons modulo 3 (there is a sole exception at m = 3). But there is no obvious obstruction that would prevent {m,m + 6,m + 12} from all being prime simultaneously. Here again, Hardy and Littlewood have a more general con- jecture which encompasses all such patterns, and it has been generalized further by various authors [5, 34, 71]. These all reflect a broad principle: if there is no simple reason why a given pattern of primes should not oc- cur, then it should occur infinitely often, with an asymptotically predictable frequency. However, given our inability to prove this for a mere pattern of length 2, there is little hope of establishing such a general claim. However, there is one other type of pattern for which great success has been achieved, particularly in recent years. Notice that the previous pattern {m,m + 6,m + 12} consists of three equally-spaced values in arithmetic progression. If we relax the demand for a particular spacing and allow for any pattern of the form {m,m + d,m + 2d} with some integer d \u00E2\u0089\u00A5 1, then the question becomes: do the primes contain infinitely many arithmetic progressions of length 3? We will shortly address this question and its analogues to longer arith- 4 1.4. Density and the primes metic progressions, but first we take a brief excursion to describe some closely related and beautiful work from the field of additive combinatorics. 1.4 Density and the primes Given a set A \u00E2\u008A\u0086 N of natural numbers, one often describes how well- populated A is, relative to N, by the quantity \u00CE\u00B4 := lim x\u00E2\u0086\u0092\u00E2\u0088\u009E #{n \u00E2\u0089\u00A4 x : n \u00E2\u0088\u0088 A} x . We say that A has asymptotic density \u00CE\u00B4, provided this limit exists (if it does not converge, one can still define lower and upper densities by taking the lim inf or lim sup). For example, the set of all even numbers has asymptotic density 12 , and this could be loosely interpreted as the \u00E2\u0080\u009Cprobability\u00E2\u0080\u009D that a natural number, chosen at random, is even. However, some care is warranted in this interpretation, as we discuss further in Section 1.10. Roth proved in 1953 [68] that every set A having a positive density \u00CE\u00B4 > 0 must contain infinitely many arithmetic progressions of length 3. So long as A is \u00E2\u0080\u009Cdense\u00E2\u0080\u009D, this pattern \u00E2\u0080\u0094 which we abbreviate as 3-AP \u00E2\u0080\u0094 is inherently unavoidable, no matter how one tries to arrange the elements of A to prevent it from recurring. We shall not describe Roth\u00E2\u0080\u0099s elegant proof in complete detail, but we feel compelled to give a loose account of its clever use of a dichotomy between randomness and structure. If A is distributed \u00E2\u0080\u009Crandomly\u00E2\u0080\u009D enough (in a sense that can be made precise in terms of discrete Fourier transforms), then it can be shown that A contains just as many 3-APs as a typical set whose elements are chosen at random with probability \u00CE\u00B4. How many is that? In this truly random model, every given subset of size 3 occurs with probability \u00CE\u00B43 > 0, so a random set will typically contain many 3-APs. On the other hand, if A is not \u00E2\u0080\u009Crandom\u00E2\u0080\u009D in this specific sense, Roth showed that there must be some structure within A that allows one to pass to a new set A\u00E2\u0080\u00B2 having two key properties: A\u00E2\u0080\u00B2 contains fewer 3-APs than A, and the density of A\u00E2\u0080\u00B2 is appreciably greater than \u00CE\u00B4. 5 1.4. Density and the primes Suppose now that A doesn\u00E2\u0080\u0099t contain any 3-APs. Then it does not fall into the \u00E2\u0080\u009Crandom\u00E2\u0080\u009D case, and so we may replace it by A\u00E2\u0080\u00B2 which has higher density. Since A\u00E2\u0080\u00B2 also lacks 3-APs, it too can be replaced by A\u00E2\u0080\u00B2\u00E2\u0080\u00B2, and so on. Eventually we arrive at a set whose density is so high (namely, \u00CE\u00B4 > 23) that it cannot avoid any pattern of length 3, a contradiction. Szemere\u00CC\u0081di later extended Roth\u00E2\u0080\u0099s theorem to arithmetic progressions of any length k, through an extraordinarily intricate argument that introduced entirely new notions of structure and randomness [80]. Since then, much sharper versions of both Roth\u00E2\u0080\u0099s and Szemere\u00CC\u0081di\u00E2\u0080\u0099s theorems have been found: the current best results are due to Sanders [70] and Gowers [27]. Let us now return to the question from the end of the previous section. Recall that the Prime Number Theorem predicts the density of primes up to x to be about 1/ log x, so the asymptotic density of the primes is 0. Unfortunately, Roth\u00E2\u0080\u0099s theorem does not apply to sets as sparse as the primes. There aren\u00E2\u0080\u0099t enough primes to justify the existence of 3-APs on the sole basis that it is unavoidable for any set with the same density (although the refinement by Sanders comes breathtakingly close to doing so). Nevertheless, the primes do contain infinitely many 3-APs (with the same frequency as predicted by random models); this was proven by van der Corput [84] in 1939! Essentially, one can show that the primes are distributed sufficiently nicely that they already fall into the \u00E2\u0080\u009Crandom\u00E2\u0080\u009D case of Roth\u00E2\u0080\u0099s proof: the necessary estimates were present two years earlier in Vinogradov\u00E2\u0080\u0099s pioneering work on exponential sums [85]. It took many more decades to establish the existence of longer k-APs in the primes. In part this is because the precise sense of \u00E2\u0080\u009Crandom\u00E2\u0080\u009D that is needed to enumerate k-APs did not begin to emerge until the seminal work of Host and Kra [40], and Gowers [27]. Building on these insights and much more, Green and Tao [29] proved a landmark achievement: the primes contain arbitrarily long arithmetic progressions. More recently they have shown, together with Ziegler [30\u00E2\u0080\u009333], that such progressions occur with the 6 1.5. Notation and outline precise frequency predicted by random models: for any k \u00E2\u0089\u00A5 2, #{k-APs within primes up to n} \u00E2\u0088\u00BC Dk 2(k \u00E2\u0088\u0092 1) \u00C2\u00B7 n2 logk n , (1.2) where the constant Dk accounts for all local factors (the other terms arising purely from the na\u00CC\u0088\u00C4\u00B1ve model). 1.5 Notation and outline With the necessary background in place, we now go into some detail about the specific questions considered in this dissertation, which cross through all of the themes of the previous sections. Starting in Section 1.6, we discuss arithmetic functions, which constitute the main focus of Chapter 2. This discussion continues until Section 1.11, where we turn to the topic of random matrices from Chapter 3. From Section 1.15 onward, we discuss the combi- natorial patterns central to Chapter 4. Let us take a moment to introduce some common notation that will appear throughout this work. We use the standard convention that N refers to the natural numbers beginning with 1, not 0. Most Roman-lettered variables, unless otherwise specified, are assumed to take values in N, especially n, m, k, d; we shall reserve p to indicate a prime value. The variable \u000F merits some attention: it most often refers to a (small) positive real number, but we use \u000F in some parts for a quantity that is \u00C2\u00B11 (the distinction will be obvious in context). If g(n) is a positive function, the Vinogradov notation f(n)\u001C g(n) indi- cates that there is a constant C > 0 such that |f(n)| \u00E2\u0089\u00A4 Cg(n) regardless of the choice of n. This naturally generalizes to functions of multiple variables, and we may use a subscript to denote that the implied constant may depend on some of these variables. For example, it is well-known that log n \u001C\u000F n\u000F for any \u000F > 0. When working with error terms it is convenient to use the Landau nota- tion O(g(n)) to abbreviate a quantity that is \u001C g(n), and likewise O\u000F(g(n)) if said quantity is \u001C\u000F g(n). We also use o(g(n)) to denote a quantity which is asymptotically much smaller than g(n); that is, f(n) = o(g(n)) if and 7 1.6. Arithmetic functions only if limn\u00E2\u0086\u0092\u00E2\u0088\u009E f(n)/g(n) = 0. In this case, a subscript as in o\u000F(g(n)) allows the rate of convergence of this ratio to depend non-uniformly on \u000F. The notation d | n means that d divides n (or, n is divisible by d). When p is prime, pr \u00E2\u0080\u0096 n more specifically means that pr exactly divides n, so that pr | n but pr+1 - n. The notation (m,n) is often used for the GCD or greatest common (positive) divisor of m and n. We say that m and n are relatively prime or coprime if (m,n) = 1. At times we may use (\u00C2\u00B7, \u00C2\u00B7) to instead denote an ordered pair; again, the context should make this clear. When a set is written in comprehension notation such as {n \u00E2\u0088\u0088 N : n \u00E2\u0089\u00A4 x}, we will often prepend # to denote the cardinality, as seen in (1.1) or (1.2). 1.6 Arithmetic functions The study of arithmetic functions is of central interest in analytic number theory. Strictly speaking, any real-valued function f : N \u00E2\u0086\u0092 R (or, more generally, a complex-valued function) defined on the natural numbers could be called an \u00E2\u0080\u009Carithmetic function\u00E2\u0080\u009D, but in actual usage the term carries a connotation that the value of f(n) has some relation to the divisibility properties of n. The most commonly studied class of arithmetic functions are the multiplicative functions, which are defined to satisfy f(1) = 1 and f(mn) = f(m)f(n), whenever (m,n) = 1. (1.3) (The condition f(1) = 1 is nearly redundant, and could be deduced from (1.3) if we ignore the constant function f \u00E2\u0089\u00A1 0.) It is not hard to see that if the function f is multiplicative then it is uniquely determined by the values f(pr) taken at all prime powers pr. It therefore suffices to specify a given multiplicative function only on prime powers, and allow the definition to extend to all natural numbers via (1.3). We give below a brief list of special functions, being already of natural interest, which also enjoy this multiplicative property. \u00E2\u0080\u00A2 The Euler totient function \u00CF\u0086(n) counts the number of integers be- tween 1 and n that are relatively prime to n (also the number of 8 1.7. Distribution of an arithmetic function invertible elements in the ring Z/nZ). This is multiplicative, with \u00CF\u0086(pr) = pr\u00E2\u0088\u00921(p\u00E2\u0088\u0092 1) for every r \u00E2\u0089\u00A5 1. \u00E2\u0080\u00A2 The divisor function \u00CF\u0084(n) counts the number of positive divisors of n (including n itself). This is multiplicative, with \u00CF\u0084(pr) = r + 1. \u00E2\u0080\u00A2 The sum-of-divisors function \u00CF\u0083(n) is the sum of all positive divisors of n. This is multiplicative, with \u00CF\u0083(pr) = 1 + p+ p2 + \u00C2\u00B7 \u00C2\u00B7 \u00C2\u00B7+ pr = pr+1\u00E2\u0088\u00921p\u00E2\u0088\u00921 . The above three functions are among the most well-studied arithmetic func- tions in number theory, yet much about them is still unknown. To give a sobering example, mathematicians in ancient Greece were interested in so- called \u00E2\u0080\u009Cperfect\u00E2\u0080\u009D numbers such as 6 = 1 + 2 + 3 and 28 = 1 + 2 + 4 + 7 + 14, which are exactly equal to the sum of their respective divisors (not counting the original number itself). In modern notation, the number n is perfect if and only if \u00CF\u0083(n)/n = 2. The simple question \u00E2\u0080\u009Cdoes there exist an odd number that is perfect?\u00E2\u0080\u009D remains unsolved despite being posed over two millennia ago (it may thus stake a reasonable claim as the world\u00E2\u0080\u0099s oldest unsolved problem). Note that the function \u00CF\u0083(n)/n is itself multiplicative (more generally, this is true of any product or quotient of multiplicative functions), and like most such functions, its value depends quite crucially on the prime factors of n. As witnessed by the previous sections, we certainly do not have perfect knowledge of the primes, and multiplicative functions likewise remain an area of intense study. While the previous paragraph highlights the difficulty of understanding when \u00CF\u0083(n)/n is exactly equal to 2, there has been much more success in establishing the approximate behaviour of \u00CF\u0083(n)/n: how often does it fall between, say, 1.999 and 2.001? 1.7 Distribution of an arithmetic function The previous question was settled by Davenport [15], who showed that for any fixed real numbers 1 \u00E2\u0089\u00A4 s < t, the set of positive integers n for which s < \u00CF\u0083(n)/n < t has a positive asymptotic density \u00CE\u00B4(s, t) > 0. (Note that 9 1.7. Distribution of an arithmetic function \u00CF\u0083(n) is always larger than n, so there is no reason to consider values below 1.) Furthermore, this density varies continuously with s and t, so the density associated to any single value of \u00CF\u0083(n)/n is zero. An analogous result was established earlier by Schoenberg [73] for the multiplicative function n/\u00CF\u0086(n). The two functions are quite similar in that they both increase as the number of prime factors of n grows, with smaller primes having significantly greater effect. A celebrated theorem of Erdo\u00CC\u008Bs and Wintner [19] shows that both results fall under a general phenomenon common to many positive-valued multiplicative functions f(n): under mild convergence conditions on the logarithm2 g(n) := log f(n), there is a con- tinuous function F (t) such that, for any t \u00E2\u0088\u0088 R+, the asymptotic density of the set {n : f(n) \u00E2\u0089\u00A4 t} exists and is equal to F (t). We reproduce these conditions below, which are in fact both necessary and sufficient (in each case the summation index p is restricted to prime values): \u00E2\u0088\u0091 |g(p)|>1 1 p , \u00E2\u0088\u0091 |g(p)|\u00E2\u0089\u00A41 g(p) p , and \u00E2\u0088\u0091 |g(p)|\u00E2\u0089\u00A41 g2(p) p converge; \u00E2\u0088\u0091 g(p)6=0 1 p diverges. (1.4) The function F (t) closely recalls the (cumulative) distribution function of a random variable in probability, and it too is called the distribution function of f(n). However, we remind the reader that there is no well-defined random variable that concretely represents all values of f(n); rather, we are taking the limiting distribution of f(n) taken over n \u00E2\u0089\u00A4 x as x \u00E2\u0086\u0092 \u00E2\u0088\u009E. The reader familiar with measure-theoretic probability will recognize this as a weak limit of measures. While one can write down an explicit formula for the distribution of \u00CF\u0083(n)/n in terms of its characteristic function or Fourier transform (first established in a later paper of Schoenberg [74]), such an expression does not lend itself to high-precision computation. It was only very recently that Kobayashi in his doctoral thesis [45] computed F (2) to four decimal places of accuracy, showing that the density 1 \u00E2\u0088\u0092 F (2) of the abundant numbers 2The resulting function g(n) is, not surprisingly, called an additive function, since it satisfies g(mn) = g(m) + g(n) for (m,n) = 1. 10 1.8. Approximation by arithmetic functions {n : \u00CF\u0083(n)/n > 2} is approximately 0.2476. 1.8 Approximation by arithmetic functions The Erdo\u00CC\u008Bs\u00E2\u0080\u0093Wintner theorem can be interpreted as saying that many positive multiplicative functions f(n) can be modelled (subject to the above caveat) by a continuous random variable. Furthermore, for the two cases considered by Davenport and Schoenberg, this distribution function is strictly increas- ing, so that f(n) approximates, within any desired accuracy \u000F, every real number s \u00E2\u0089\u00A5 1 for a positive frequency of n (F (s + \u000F) \u00E2\u0088\u0092 F (s \u00E2\u0088\u0092 \u000F), to be ex- act). In other words, the set of values {f(n) : n \u00E2\u0088\u0088 N} is topologically dense in [1,\u00E2\u0088\u009E), and in a rather strong sense (every point has a large number of approximants). Let us now consider a stronger question that may be reminiscent of the twin primes problem: since n and n+1 share no common prime factors, it is reasonable to believe that f(n) and f(n+1) should behave \u00E2\u0080\u009Cindependently\u00E2\u0080\u009D in some sense. Is it possible for f(n) to approximate one number s1 while f(n + 1) simultaneously approximates a different number s2? Given that the values of f(n) are dense in [1,\u00E2\u0088\u009E), is it true that the pairs of values (f(n), f(n+ 1)) are dense in [1,\u00E2\u0088\u009E)\u00C3\u0097 [1,\u00E2\u0088\u009E)? Schinzel and Wang [72] proved a result in this general direction for the specific function \u00CF\u0086(n): for any dimension d \u00E2\u0089\u00A5 1, the vector of ratios( \u00CF\u0086(n+ 1) \u00CF\u0086(n) , \u00CF\u0086(n+ 2) \u00CF\u0086(n) , . . . , \u00CF\u0086(n+ d) \u00CF\u0086(n) ) (1.5) can approximate any point in (0,\u00E2\u0088\u009E)d arbitrarily well. We remark that while n and n + k may have some prime factors in common, their GCD cannot exceed k, so it is fair to expect \u00CF\u0086(n) and \u00CF\u0086(n + k) to retain some weaker level of independence (when n is large and k is fixed). This is, however, true in a slightly weaker sense than previously: the set of n for which (1.5) approximates a given point to within \u000F (in every coordinate) might have asymptotic density 0. Shortly afterward, Erdo\u00CC\u008Bs and Schinzel [20] established a wide generaliza- 11 1.9. Comparing arithmetic progressions tion for an arbitrary multiplicative f(n) with certain mild hypotheses: for any dimension d \u00E2\u0089\u00A5 1 there is a constant cd such that the vectors (f(n+ 1), f(n+ 2), . . . , f(n+ d)) (1.6) are dense in [cd,\u00E2\u0088\u009E)d. Note that our initial speculation that c2 = 1, in the classical case f(n) = \u00CF\u0083(n)/n, is actually false: it is impossible for f(n) and f(n+1) to both be very close to 1, since one of n or n+1 must be even, and for any even number m, f(m) \u00E2\u0089\u00A5 32 . So it is genuinely necessary to prescribe a different cd for each dimension: the influence of local factors is felt here once more. Moreover, the Erdo\u00CC\u008Bs\u00E2\u0080\u0093Schinzel proof shows that the set of n which ap- proximate any given point in [cd,\u00E2\u0088\u009E)d (within some fixed tolerance \u000F) has a positive asymptotic lower density. The hypotheses on f(n) are similar to the Erdo\u00CC\u008Bs\u00E2\u0080\u0093Wintner conditions, but they do not necessarily imply that f has a distribution function, so it would be unreasonable to expect the aforemen- tioned set to have a precise (nonzero) asymptotic density. However, as an approximation property this is arguably just as strong: we still have that a positive proportion of n \u00E2\u0089\u00A4 x are good approximants, and for some values of x this proportion may be higher still. 1.9 Comparing arithmetic progressions A related problem was more recently considered in a brief note of New- man [62], which is motivated by a simple observation: while \u00CF\u0086(n)/n is well- modelled by a random variable (whose average value works out to exactly 6/pi2), the function \u00CF\u0086(6n) contains an inherent bias: it is easy to verify that if m = 6n, then \u00CF\u0086(m)/m \u00E2\u0089\u00A4 13 , which is considerably smaller than average. Therefore, we might expect that the inequality \u00CF\u0086(6n) > \u00CF\u0086(6n + 1) should occur very rarely3, and rarer still for \u00CF\u0086(30n) > \u00CF\u0086(30n + 1). Indeed, the former does not occur until n \u00E2\u0089\u0088 4.1\u00C3\u0097 1055, and the smallest solution to the 3In fact, one can show that the average value of \u00CF\u0086(m)/m for m = 6n+ 1 is 9/pi2 (much higher than average), which should lower our expectations even further. 12 1.10. Probability in number theory latter, explicitly computed by Martin [53], is 1116 digits long! Nevertheless, Newman showed that such unusual reversals do occur at infinitely many n, and more generally this is true for any inequality of the form \u00CF\u0086(an + b) > \u00CF\u0086(cn + d) where a, b, c, d are fixed integers satisfying ad \u00E2\u0088\u0092 bc 6= 0. (This latter constraint is necessary to prevent cn + d from being a scalar multiple of an + b, which would force them to have almost exactly the same prime factors: for instance, it is not hard to see that \u00CF\u0086(n) > \u00CF\u0086(2n) can never occur.) In Chapter 2, we will show that Newman\u00E2\u0080\u0099s result may be subsumed into a more general version of the Erdo\u00CC\u008Bs\u00E2\u0080\u0093Schinzel theorem that applies to such arithmetic progressions. Specifically, if a1n+b1, a2n+b2, . . . , adn+bd are any choice of non-constant linear functions (none of which is a scalar multiple of another), then the vectors (f(a1n+ b1), f(a2n+ b2), . . . , f(adn+ bd)) (1.7) are dense in [c,\u00E2\u0088\u009E)d for some constant c (which depends on the specific choice of linear functions and not just d), in the same strong sense as before (there are many are good approximants to each point). Our result is more general in one additional sense: each component of the vector may have its own individual multiplicative function in place of f (provided these functions all satisfy similar conditions). As a consequence of our theorem, not only does the inequality \u00CF\u0086(6n) > \u00CF\u0086(6n + 1) occur with at least some positive (albeit very close to 0) fre- quency, but this statement remains true even if we specify that the ratio \u00CF\u0086(6n)/\u00CF\u0086(6n+ 1) should agree with pi to 100 decimal places! 1.10 Probability in number theory It seems appropriate at this point to elaborate on the loose interpretation of probability which recurred in Sections 1.4 and 1.7. Many results in this area can be paraphrased in probabilistic language: already we have witnessed the terms \u00E2\u0080\u009Cprobability\u00E2\u0080\u009D, \u00E2\u0080\u009Caverage value\u00E2\u0080\u009D and \u00E2\u0080\u009Cdistribution function\u00E2\u0080\u009D. Certainly, 13 1.10. Probability in number theory the concept of a \u00E2\u0080\u009Cpositive integer chosen at random\u00E2\u0080\u009D has a natural intuitive appeal, perhaps even more so than that of a real number chosen randomly from the interval [0, 1]. But while the latter concept is readily formalized as a legitimate well-defined random variable, the former is at odds with the measure-theoretic foundations of probability theory. There is no uniform measure that can be placed on a countably infinite set like the naturals (although asymptotic density is uniform, it is only finitely additive). Ultimately, any (non-trivial) statement that refers to \u00E2\u0080\u009Crandom integers\u00E2\u0080\u009D is actually describing a limiting process (where at each step of the limit, the probability space remains finite and well-defined): density as a limit of finite probabilities, average value as a limit of finite averages, and distribution function as a (weak) limit of discrete distributions. But this interpretation, while appealing, tends to obscure the very real error terms underlying the convergence to such limits. To give a simple concrete example, the (limiting) probability that a number is even is 12 , but the true proportion of even numbers in {1, . . . , n} is 12 + O( 1 n), the implied constant depending on whether n itself is odd or even. Similarly, the probability that a number is divisible by 3 is 13 +O( 1 n). The events that a random integer is \u00E2\u0080\u009Cdivisible by 2\u00E2\u0080\u009D and \u00E2\u0080\u009Cdivisible by 3\u00E2\u0080\u009D may be independent in the sense of limits (that is, 12 \u00C2\u00B7 13 = 16), but this ignores the error terms that appear along the way. When working with large numbers of such not-quite-independent events, the accumulation of these error terms can be very significant. For instance, as mentioned in [28], one cannot use this simple approach to accurately count the number of primes between \u00E2\u0088\u009A n and n (by excluding all numbers divisible by some prime p \u00E2\u0089\u00A4 \u00E2\u0088\u009An). Nevertheless, Brun [12] found clever ways to handle these errors to show that the number of twin primes, while still poorly understood, cannot be vastly higher than predicted by Hardy and Littlewood\u00E2\u0080\u0099s heuristic (1.1). These techniques gave rise to the modern field of sieve theory. Having discussed some of the inherent complications of thinking about integers in probabilistic terms, we feel that it is justified by the fact that many beautiful theorems could not be described so elegantly in any other 14 1.11. Random matrices and eigenvalues form. We give two such examples below (the latter is a celebrated result of Erdo\u00CC\u008Bs and Kac [17]; together with the Erdo\u00CC\u008Bs\u00E2\u0080\u0093Wintner theorem, these form the cornerstone of probabilistic number theory). \u00E2\u0080\u00A2 The probability that two random integers are relatively prime is 6/pi2. \u00E2\u0080\u00A2 The number of prime factors of a random n is normally distributed with mean and variance both equal to log log n. 1.11 Random matrices and eigenvalues Let us now turn to a different type of random structure (still arithmetical in nature), for which it is possible to determine the distribution with a much greater precision than most of our previous examples. We first recall some terminology from undergraduate linear algebra: given a square matrix A, we say that A has an eigenvalue \u00CE\u00BB \u00E2\u0088\u0088 C if there exists a nonzero vector x (called an eigenvector) such that Ax = \u00CE\u00BBx. Any given matrix possesses only finitely many eigenvalues: to be specific, if A is an n \u00C3\u0097 n matrix, it has exactly n eigenvalues given by the roots of a cer- tain nth degree polynomial (possibly with some repetitions). These values yield significant information about the structure of A, permitting numer- ous applications in diverse areas including quantum physics, statistics and epidemiology. Perhaps the most prominent usage of all is the PageRank algorithm pioneered by Google Inc. [11]: if A is suitably built to represent all links between web pages, then the principal eigenvector of A (one that is associated with the largest eigenvalue) yields a direct measure of each page\u00E2\u0080\u0099s relative importance. This establishes the set of eigenvalues of a matrix (colourfully known as its spectrum) as a central object of study: just as prime factors are intrinsic to integers, so too are eigenvalues intrinsic4 to matrices. Given a large, or even infinite, family of matrices, we are naturally led to wonder about the statistical distribution of eigenvalues within. In keeping with the arithmetic 4The prefix \u00E2\u0080\u009Ceigen-\u00E2\u0080\u009D derives from a Germanic word meaning \u00E2\u0080\u009Cself\u00E2\u0080\u009D or \u00E2\u0080\u009Cinnate\u00E2\u0080\u009D. 15 1.12. Random integer matrices theme of this thesis, we study in Chapter 3 the (initially vague) question: what can one say about the eigenvalues of a random matrix of integers? Before we expand on this question in the next section, we would be remiss not to mention some classical work on random matrices, much of which has been devoted to continuous (and often symmetric) random models originating from quantum physics. Even in number theory, one of the most frequently discussed matrix models is based not on integers but on complex numbers: it is a probability distribution defined on n\u00C3\u0097n Hermitian matrices5 known as the Gaussian unitary ensemble (GUE). The significance of this model to number theory is that the eigenvalues of random GUE matrices, when n is very large, appear to have spacing that is statistically identical to that of the zeros of the Riemann zeta function; this surprising connection was first made at a (now-famous) chance meeting between Montgomery [58] and physicist Freeman Dyson. (Though it does not occupy a major role in this thesis, the Riemann zeta function may well be the most important object in all of analytic number theory; a greater understanding of its zeros would answer a wide range of outstanding questions about the distribution of primes \u00E2\u0080\u0094 see [9] for a nice summary.) A beautiful result of Edelman [16] is also worth mentioning here: if we populate an n\u00C3\u0097 n matrix with independent entries randomly drawn from a standard normal distribution, the probability that it is diagonalizable over R (meaning essentially6 that its eigenvalues are all real) is exactly 2\u00E2\u0088\u0092n(n\u00E2\u0088\u00921)/4. 1.12 Random integer matrices Our own investigation (jointly with Greg Martin) begins with a question similar to Edelman\u00E2\u0080\u0099s, which was posed by Hetzel, Liew, and Morrison in [37]: what is the probability that a random n\u00C3\u0097 n integer matrix is diagonalizable over the field of rational numbers? As we have previously seen, this question does not refer to any single 5Hermitian matrices are a complex-valued analogue of real symmetric matrices. 6These are not precisely equivalent, should an eigenvalue be repeated; however, this complication occurs with probability 0 so we may safely ignore the distinction. 16 1.12. Random integer matrices distribution of integer matrices; it must be understood as a limiting process \u00E2\u0080\u0094 see the introduction of [37] for a lively discussion justifying the use of the term \u00E2\u0080\u009Cprobability\u00E2\u0080\u009D for this purpose. Just as before, we will restrict the integers to a finite set: for any k \u00E2\u0089\u00A5 1, we can certainly populate an n \u00C3\u0097 n matrix M(k) with entries chosen independently at random from the integers between \u00E2\u0088\u0092k and k. The previous question then becomes: as k \u00E2\u0086\u0092 \u00E2\u0088\u009E, what happens to the probability that M(k) is diagonalizable over Q? We expect that this is equal to the probability that every eigenvalue of M(k) is rational, and indeed Hetzel et al. rigorously justified this equivalence (the two probabilities are not exactly equal but they do have the same limit as k \u00E2\u0086\u0092 \u00E2\u0088\u009E). Computational experiments led them to conjecture that both probabilities converge to 0 (except in the case n = 1, since 1 \u00C3\u0097 1 matrices are already diagonal in a vacuous sense). We confirm in Theorem 3.1 that this is indeed true. In fact we prove a stronger result: the probability that M(k) has any rational eigenvalues tends to 0. In other words, a typical matrix does not have any rational eigenvalues, which generalizes earlier results of [47] and [37] for the case n = 2. Quantitatively, we show that this probability is On,\u000F(k \u00E2\u0088\u00921+\u000F) for any \u000F > 0, which certainly goes to 0 for k large (we could take, say, \u000F = 12). In answering the question of Hetzel et al., we are fixing the matrix size n and allowing the size of the entries k to grow. One might equally ask the converse problem: what happens if we instead fix the size of the entries, but allow the matrix to grow? This has been extensively studied in the more general formulation that the entries are chosen from any fixed probability distribution X. The case where X is a standard normal distribution was studied by Ginibre [24], who proved that as n\u00E2\u0086\u0092\u00E2\u0088\u009E the eigenvalues become uniformly distributed \u00E2\u0080\u0094 in a limiting sense \u00E2\u0080\u0094 throughout the unit disk of radius \u00E2\u0088\u009A n (this random matrix model, now called the Ginibre ensemble, was also the setting of Edelman\u00E2\u0080\u0099s result). Girko [25] showed that this limit occurs for a wider class of distributions X, and proposed that it should be universal, meaning that any random model of this type should converge to this same distribution (provided X has the same mean and variance). Quite recently, a series of papers by Tao and Vu [82, 83] have successfully 17 1.13. Singularity and computation established Girko\u00E2\u0080\u0099s circular law in full generality (in fact, their most recent result holds for a slightly wider class of random matrices than originally conjectured). 1.13 Singularity and computation A closely related, and perhaps simpler, variation of the eigenvalue problem is to find the probability that the random matrix M(k) is singular (equiva- lently, that it has an eigenvalue of 0). In fact, our proof of Theorem 3.1 relies strongly on obtaining a quantitative upper bound of the form On,\u000F(k \u00E2\u0088\u00922+\u000F) for this singularity probability. Whereas quantum mechanics was a key motivation for the study of Gaus- sian models, questions about discrete distributions arise more naturally in the field of theoretical computer science. A well-known problem in combina- torial optimization is integer programming: given a matrix A and a vector b, one wishes to find an integer vector x such that Ax \u00E2\u0089\u00A4 b (here the notation a \u00E2\u0089\u00A4 b means that ai \u00E2\u0089\u00A4 bi for every index i). Were it not for the restriction to integer values, this could be done efficiently by standard techniques for linear programming. With this restriction, the problem is classically known to be NP-complete: in a very precise sense, it is just as hard to solve as several thousand other such problems, widely regarded as intractable. While it is justifiably hard to solve integer programs in generality, the typical case is quite different (say, if A is randomly chosen as before). Our bounds for the singularity probability were adapted by Pataki, Tural and this author [66] to show that most instances can be solved very easily, essen- tially using a na\u00CC\u0088\u00C4\u00B1ve algorithm known as branch-and-bound. This may seem counterintuitive, but it is not uncommon: many NP-complete problems (not all, though; see [49]) are known to have good typical-case behaviour with very few difficult instances. In contrast, the famous problem of factoring an integer into primes is believed to be quite the opposite, with very few easy cases7 (it is still unknown whether factoring is NP-complete). 7Seeing that the security of Internet commerce currently relies on this fact, we sincerely hope this belief is not overturned in the immediate future! 18 1.14. Rationality versus reality 1.14 Rationality versus reality For the special case of 2 \u00C3\u0097 2 random matrices, we are able to give a very sharp answer to Hetzel et al.\u00E2\u0080\u0099s original question: Pr(M(k) is diagonalizable over Q) = C log k k +O (1 k ) , where C \u00E2\u0089\u0088 0.55873957 is an explicit constant. In fact, we devote a significant part of Chapter 3 to calculating the exact limiting distribution of rational eigenvalues of M(k) in the case n = 2. Note that we use the term \u00E2\u0080\u009Cdistribution\u00E2\u0080\u009D in a broader sense than distribution functions: the concept is actually much closer to that of a probability density function, except that it has total area 2 rather than 1, since a matrix has two eigenvalues. To describe this distribution, it is convenient to consider the eigenval- ues of the scaled-down matrix 1kM(k) (which are simply the eigenvalues of M(k), divided by k). If one plots a histogram of the eigenvalues of 1kM(k), restricting to only those that are rational, these histograms will converge as k \u00E2\u0086\u0092\u00E2\u0088\u009E to the distribution depicted by the solid line in Figure 3.1, which is defined explicitly as a piecewise smooth function.8 On the other hand, this limiting process 1kM(k) converges (in the weak sense) to a random matrix with real entries uniformly distributed on the interval [\u00E2\u0088\u00921, 1]. For such a continuous random matrices, there is no obvi- ous structure to the set of rational eigenvalues: any small perturbation of a matrix with a rational eigenvalue could easily shift it to some nearby real value.9 One might therefore ask if the limit distribution of rational eigen- values could be explained simply in terms of the real eigenvalue distribution in the continuous case? To answer this question, we also compute the real eigenvalue distribution for n = 2: this is a genuine distribution, though it is also equal to the limiting 8By contrast, the distribution of \u00CF\u0083(n)/n is known to be singular, having infinitely many points of non-differentiability [18]. 9It is not clear whether one can formalize this intuition by placing a natural uniform measure on the restriction to rational eigenvalues. 19 1.15. Sums of two squares distribution for the real eigenvalues of 1kM(k). It is depicted by the dashed line in Figure 3.1, which is noticeably different from the previous one. One obvious difference is the fact that the rational distribution is unimodal, while the real distribution is bimodal. In particular, it is more common for M(k) to have a rational eigenvalue near 0 than one near 3k/4; yet the reverse is true for real eigenvalues! This curious disparity suggests that there is some arithmetic structure of 1kM(k) that is lost upon passing to the continuous limit. It would not be fair to attribute such structure to local effects, since the graphs do not display any obvious dependence on the value of k (or on the eigenvalue itself) modulo small primes. However, a powerful result of Katznelson [43] may shed some light on the nature of this structure, as we discuss further in Section 3.10. Indeed, Katznelson\u00E2\u0080\u0099s result could be used, with some effort, to recover the rational eigenvalue distribution of M(k) for n = 2, or any larger value of n. 1.15 Sums of two squares Having explored genuinely random phenomena in the previous four sections, we now turn our attention to a third and final arithmetic setting, which bears many similarities to the primes (and poses many of the same challenges). Let us write S for the set of integers which can be written as x2 + y2, the sum of two square integers. For concreteness we list the first sixteen terms of S below: S = {0, 1, 2, 4, 5, 8, 9, 10, 13, 16, 17, 18, 20, 25, 26, 29, . . .}. It is apparent from this list that not all natural numbers belong to S. Indeed, since x2 is always 0 or 1 modulo 4, no number of the form 4m + 3 will appear. This local argument succinctly explains why 3 and 7 are not in S, but not why 21 is absent (this could be ruled out by a similar process modulo 9, but we will shortly give a more comprehensive reason). Despite the seemingly arbitrary construction of S, it actually enjoys a 20 1.15. Sums of two squares rather rich history, owing to its surprisingly deep structure with roots at the heart of classical number theory. The third-century Greek mathematician Diophantus of Alexandria was the first to record the identity (a2 + b2)(c2 + d2) = (ac\u00E2\u0088\u0092 bd)2 + (ad+ bc)2, (1.8) which establishes that S is closed under multiplication. However, the critical observation indicating the true nature of S was made by Pierre de Fermat10 in 1640: every prime of the form 4n+ 1 belongs to S. A rigorous proof of Fermat\u00E2\u0080\u0099s claim did not appear until a century later by Euler (crediting Fermat for the necessary technique of infinite descent), who used it to characterize members of S by their prime decomposition: the natural number n belongs to S if and only if, in the prime factorization of n, every prime of the form 4m+ 3 appears to an even power. For instance, 74 \u00C2\u00B7 17 \u00C2\u00B7 232 belongs to S (since both 7 and 23 have even exponents), while 74 \u00C2\u00B7 17 \u00C2\u00B7 23 does not. So even though S is constructed through addition, it nonetheless carries a highly multiplicative structure that is not entirely unlike the primes.11 In- deed, Landau [48] used this very structure to establish a theorem analogous to the Prime Number Theorem, giving the precise density of S: #{n \u00E2\u0089\u00A4 x : n \u00E2\u0088\u0088 S} \u00E2\u0088\u00BC C x\u00E2\u0088\u009A log x , (1.9) where the constant C \u00E2\u0089\u0088 0.76422365 is largely composed of local factors as in the Hardy\u00E2\u0080\u0093Littlewood conjecture. From this we see that the asymptotic density of S is equal to 0. The quantitative density O(1/ \u00E2\u0088\u009A log x) situates S (on a logarithmic scale) exactly halfway between the positive density O(1) and the density of primes O(1/ log x). 10Fermat and Diophantus share another connection: Fermat penned his so-called \u00E2\u0080\u009CLast Theorem\u00E2\u0080\u009D (famously proven by Wiles) in his copy of Diophantus\u00E2\u0080\u0099s Arithmetica. 11One could, somewhat facetiously, describe the primes as those numbers n for which every prime less than n appears to a power less than 1. 21 1.16. Patterns in sums of two squares 1.16 Patterns in sums of two squares We have observed some commonalities between S and the primes: they are roughly comparable in density, and their similar multiplicative struc- ture suggests that a Crame\u00CC\u0081r-like model would be successful in formulating predictions about S (tempered by the local factors that arise from Euler\u00E2\u0080\u0099s characterization). It is natural to ask very similar questions about S that were raised about primes in Sections 1.2 and 1.3 (most of which are widely considered hopeless). However, there is one obvious advantage enjoyed by S, which affords further progress on these combinatorial questions. Not only does S have an exclusive12 multiplicative structure, it also has an inclusive, additive struc- ture: by definition, we can show that n \u00E2\u0088\u0088 S simply by finding two integers x and y such that x2 + y2 = n. There is no comparable way to construct a prime out of arbitrary integers.13 As a straightforward example, consider the \u00E2\u0080\u009Ctwin primes\u00E2\u0080\u009D problem for S: it is obvious that n2 and n2 + 1 both belong to S for any n, so S contains infinitely many patterns of type {m,m+1}. Even the longer pattern {m,m+1,m+2} is easy to find: given any three consecutive numbers n\u00E2\u0088\u00921, n, n+ 1 belonging to S with n > 1 (for instance n = 9), it is also true that {n2 \u00E2\u0088\u0092 1, n2, n2 + 1} \u00E2\u008A\u0082 S. The first inclusion follows from the factorization (n \u00E2\u0088\u0092 1)(n + 1) and Dio- phantus\u00E2\u0080\u0099s formula (1.8); the other two are obvious, as noted before. We will further discuss such runs of consecutive elements in Section 1.17, but for S there is no hope of extending this to runs of length 4: every fourth number has the form 4m+ 3, which is excluded. Let us turn now to the question of gaps (spacings between consecutive terms of S). We have just witnessed infinitely many gaps of size 1 in S, so 12That is, we can prove certain integers do not belong to S based solely on, say, the power of 3 dividing them. A similar process of exclusion clearly holds for primes. 13While \u00E2\u0080\u009Cprime-generating\u00E2\u0080\u009D formulae do exist, many are trivial curiosities; even the most interesting one, a polynomial constructed by Jones et al. [41], is not practical here. 22 1.16. Patterns in sums of two squares we know that gaps can be arbitrarily small, and often. Conversely, it is not hard to see that gaps in S can be arbitrarily large: by Landau\u00E2\u0080\u0099s asymptotic (1.9), the average size of the nth gap is roughly 1.3 \u00E2\u0088\u009A log n. This leaves one existence question: does every d \u00E2\u0088\u0088 N occur as a gap somewhere in S? Spiro and Hensley [79] answered this, showing that every gap size does in fact occur, and infinitely often. Note that we have constructed only \u00E2\u0088\u009A n instances of the pattern {m,m+ 1} up to n; however, random models predict that there should be about n/ log n (ignoring the multiplicative constant). For the pattern {m,m + 1,m + 2}, the above construction gives only log log n occurrences; a more efficient construction could improve this to n1/4, but this is still a far cry from the expected n/(log n)3/2. Much stronger results on S were obtained by Hooley [39], who has stud- ied the combinatorics of S extensively. Using analytic techniques (cleverly drawing from sieve theory) he showed that the pattern {m,m + 1} occurs at least \u001D n/ log n times up to n. Since a similar upper bound \u001C n/ log n is also known by standard sieve methods, this establishes the correct order of magnitude for the frequency of twins {m,m + 1} in S, if not a precise asymptotic (with a known constant). Hooley also showed in [38] that the pattern {m,m+ a,m+ b} occurs infinitely often in S, for any a and b. So far, there seems to be greater hope of establishing patterns in S than in the prime case. It is ironic then, that for arithmetic progressions the primes have enjoyed far more success than S. For instance, the existence of arbitrarily long k-APs in S was not known until it became an immediate consequence of the Green\u00E2\u0080\u0093Tao theorem.14 The expected asymptotic for the quantity of 3-APs in S remains (to our best knowledge) unproven, unlike van der Corput\u00E2\u0080\u0099s theorem for primes. In general, precise asymptotics for S seem to be quite difficult \u00E2\u0080\u0094 very recently, though, Matthiesen [56] has given a complete solution to a modified form of the k-APs problem (we discuss this modification in the conclusion). 14Green and Tao actually showed that a set containing any positive proportion of primes contains long progressions, and S contains about half of all primes. 23 1.17. Patterns in quadratic forms 1.17 Patterns in quadratic forms We can more generally view S within a wider context: for k \u00E2\u0089\u00A5 1, let S(k) be the set of numbers of the form x2 +ky2, so that S is equivalent to S(1). This extension is not entirely arbitrary, in part due to an identity generalizing (1.8), found by the 7th century Indian mathematician Brahmagupta: (a2 + kb2)(c2 + kd2) = (ac\u00E2\u0088\u0092 kbd)2 + k(ad+ bc)2. (1.10) In Chapter 4 we devote our attention wholly to S(k), but we pause briefly to mention a truly fascinating related area of number theory. The theory of binary quadratic forms such as x2 + ky2 and more generally ax2 + bxy+ cy2 blossomed in the late 18th century under Euler, Lagrange and Legendre, culminating in Gauss\u00E2\u0080\u0099s 1801 masterwork Disquisitiones Arithmeticae [22]. While not all such forms possess a self-multiplication property15 such as (1.10), Gauss uncovered a beautiful group structure which interconnects them all. Much more recently, Bhargava [7] has shed entirely new light on Gauss\u00E2\u0080\u0099s composition laws, and applied these profound insights to solve some outstanding problems in algebraic number theory! Let us return to the specific topic of S(k). The density of S(k) is very similar to that of S, having an asymptotic formula of the same form as (1.9), albeit with a different constant \u00E2\u0080\u0094 here, the constant is more intricate, containing additional arithmetic factors (related to the work mentioned in the previous paragraph). This asymptotic was established by Bernays (then a student of Landau) in his 1912 doctoral thesis [6], and was considerably sharpened recently by Blomer and Granville [8]. The local considerations for S(k) are rather dependent on the value of k. Let us discuss how this affects the two problems of runs and gaps from the previous section. While S(1) does not contain more than 3 consecutive integers, S(2) contains {0, 1, 2, 3, 4}, as well as {96, 97, 98, 99, 100}. On the other hand, one can show that S(k) never admits any run of length 6; in this sense, such quintuples or 5-runs are of maximal interest for the S(k) 15However, a surprising consequence of Gauss\u00E2\u0080\u0099s work is that the product of any three numbers of the form ax2 + bxy + cy2 can be expressed in the same form! 24 1.17. Patterns in quadratic forms setting.16 This raises the natural question: for which values of k does S(k) admit (infinitely many) consecutive quintuples?. We give a partial answer to this question, showing that S(2) does con- tain infinitely many 5-runs, which answers a question of Moshe Rosenfeld (personal communication). More generally, we can show that S(k) has the same property for several small values of k, as well as for an infinite sequence of distinct values k. However, this technique does not apply to all values of k: we do not know whether S(34) has infinitely many 5-runs (heuristics suggest that it should). For the analogous question of which gaps appear in S(k), the only local obstruction occurs when k is divisible by 4 (in which case S(k) only contains numbers of the form 0 or 1 modulo 4). For such k, it is impossible for S(k) to contain gaps of size 2 or any number of the form 4d+2. In Section 4.7 we fully generalize Spiro and Hensley\u00E2\u0080\u0099s result to each S(k): aside from this one class of exceptions, every d \u00E2\u0088\u0088 N appears infinitely often as a gap in S(k). 16However, the form xy takes all possible values. While this example is rather trivial, it shows that local factors do not preclude, say, 30!(x2 + y2) + xy from having long runs. 25 Chapter 2 Simultaneous Approximation by Arithmetic Functions 2.1 Introduction In this chapter we study the multidimensional approximation of points using values taken by arithmetic functions on arithmetic progressions, establish- ing generalizations of some theorems of Erdo\u00CC\u008Bs and Schinzel [20], as well as Newman\u00E2\u0080\u0099s result from Section 1.9. While it is multiplicative functions that comprise our main interest, these theorems are more naturally formulated in terms of the logarithms of such functions, which are additive. We will adopt the (non-standard) notation \u00E2\u0080\u0096 \u00C2\u00B7 \u00E2\u0080\u0096 from [20], which is defined by: \u00E2\u0080\u0096x\u00E2\u0080\u0096 := \u00EF\u00A3\u00B1\u00EF\u00A3\u00B2\u00EF\u00A3\u00B3x, if |x| \u00E2\u0089\u00A4 1;1, if |x| > 1. We first give the formal statement of the Erdo\u00CC\u008Bs\u00E2\u0080\u0093Schinzel theorem. Define F to be the class of all additive functions f : N\u00E2\u0086\u0092 R satisfying the following: Condition 2.1. f(n) is a real-valued additive function such that (a) \u00E2\u0088\u0091 p \u00E2\u0080\u0096f(p)\u00E2\u0080\u00962/p is convergent; (b) There exists a real c1 = c1(f) such that, for any integer M > 0, the set of values {f(N) : (N,M) = 1} is dense in (c1,\u00E2\u0088\u009E); (c) The partial sums of \u00E2\u0088\u0091 p \u00E2\u0080\u0096f(p)\u00E2\u0080\u0096/p are bounded (note that the summand is not necessarily positive). 26 2.1. Introduction Erdo\u00CC\u008Bs and Schinzel proved [20] that the following approximation theorem holds for any f \u00E2\u0088\u0088 F : Theorem 2.2 (Erdo\u00CC\u008Bs\u00E2\u0080\u0093Schinzel). For any h \u00E2\u0088\u0088 N there exists ch such that, for any \u000F > 0 and every sequence of h numbers \u00CE\u00B11, \u00CE\u00B12, . . . , \u00CE\u00B1h \u00E2\u0089\u00A5 ch, the set of natural numbers n which satisfy the simultaneous inequalities |f(n+ i)\u00E2\u0088\u0092 \u00CE\u00B1i| < \u000F (i = 1, 2, . . . , h) (2.1) has positive lower density. They furthermore showed [20, p. 482] that any additive f for which Theorem 2.2 holds must also satisfy Condition 2.1, so that their result is best possible. It is worth comparing these hypotheses to the Erdo\u00CC\u008Bs\u00E2\u0080\u0093Wintner conditions which determine the existence of a distribution function for f(n). Those conditions, from (1.4), may be rephrased17 in the present notation as \u00E2\u0080\u00A2 \u00E2\u0088\u0091p \u00E2\u0080\u0096f(p)\u00E2\u0080\u00962/p is convergent; \u00E2\u0080\u00A2 \u00E2\u0088\u0091p \u00E2\u0080\u0096f(p)\u00E2\u0080\u0096/p converges. Relative to the above, we see that Condition 2.1 relaxes the requirement that \u00E2\u0088\u0091 p \u00E2\u0080\u0096f(p)\u00E2\u0080\u0096/p converge to merely having bounded partial sums. At the same time, it adds a new requirement that the values of f remain dense after excluding certain congruence classes. One fairly common situation, which is sufficient to imply this denseness requirement, is that f(p) \u00E2\u0089\u00A5 0, f(p)\u00E2\u0086\u0092 0 and \u00E2\u0088\u0091 p f(p) diverges (or that this is true on some subsequence of primes). In such case we may explicitly take c1 = 0. We note that if f is a positive multiplicative function for which log f \u00E2\u0088\u0088 F , then the conclusion of Theorem 2.2 holds equally well for f . The class F includes a wide range of such logarithmic functions, notably log \u00CF\u0083(n)\u00E2\u0088\u0092 log n and log n \u00E2\u0088\u0092 log \u00CF\u0086(n) (we may take c1 = 0 here, for the reason mentioned above). The theorem can therefore be used to approximate any tuple of sufficiently large reals by consecutive values of \u00CF\u0083(n)/n, or sufficiently small 17We have suppressed the divergence requirement from (1.4), which is not needed for existence but only for continuity of the distribution function. 27 2.1. Introduction positive reals by \u00CF\u0086(n)/n. Theorem 2.2 also readily implies a similar ap- proximation result for consecutive values of f(n + 1) \u00E2\u0088\u0092 f(n), and it is easy to see that this yields results for \u00CF\u0083(n + 1)/\u00CF\u0083(n) and \u00CF\u0086(n + 1)/\u00CF\u0086(n); thus, Theorem 2.2 is a powerful generalization of earlier results by Schinzel and Wang [72]. We will return to this \u00E2\u0080\u009Crelative\u00E2\u0080\u009D formulation in Section 2.3. Shao had previously given [77] a generalization of Schinzel and Wang\u00E2\u0080\u0099s results for \u00CF\u0086 and \u00CF\u0083. However, this result placed additional assumptions on the growth rate of f(pr) as p\u00E2\u0086\u0092\u00E2\u0088\u009E for each r \u00E2\u0089\u00A5 1. By contrast, Theorem 2.2 needs only bounds on f(p), and it obtains a positive proportion of n rather than just an infinite set. As a cute example, any multiplicative function f with f(p) = 1 for all primes p, and each of f(p2) = 3 and f(p2) = 12 for infinitely many p necessarily satisfies Condition 2.1, and therefore has the approximation property (2.1) for many values of n. (One could freely replace 3 and 12 with any reals \u00CE\u00B1 and \u00CE\u00B2 such that log\u00CE\u00B1 log \u00CE\u00B2 is a negative irrational.) Our main theorem is a generalization of Theorem 2.2 to the case of mul- tiple arbitrarily chosen functions from F . The method of proof is essentially the same as that in [20], for which we give a self-contained exposition. Theorem 2.3. Let f1, f2, . . . , fh \u00E2\u0088\u0088 F . Then there exists a constant c such that for any sequence of h numbers \u00CE\u00B11, \u00CE\u00B12, . . . , \u00CE\u00B1h \u00E2\u0089\u00A5 c, and \u000F > 0, the set of natural numbers n satisfying the simultaneous inequalities |fi(n+ i)\u00E2\u0088\u0092 \u00CE\u00B1i| < \u000F (i = 1, 2, . . . , h) (2.2) has positive lower density. In Section 2.3, we generalize this result further: rather than consider the values of f at the consecutive integers n + 1, n + 2, . . . , n + h, we may take (essentially) arbitrary linear functions in n. The technique is entirely elementary, and for expository purposes we choose to generalize a slightly different theorem from [20]; the application of the same technique to Theo- rem 2.3 is comparably straightforward. It is natural to ask how the density of n satisfying (2.2), or the smallest such n, might vary as \u000F \u00E2\u0086\u0092 0. While estimates for these could be obtained 28 2.2. Main theorem from our construction, they are exponential in 1/\u000F, and the parameters de- pend on specific convergence properties of f(p), which makes it difficult to give a reasonable bound that is uniform over F . However, Alkan, Ford and Zaharescu [2] have recently proved a very sharp result along these lines, af- ter imposing some reasonable regularity constraints on f(pr). They obtain an analogue of Theorem 2.3 (and its generalization to linear forms, Corol- lary 2.11) for infinitely many n, where \u000F may be replaced by n\u00E2\u0088\u0092c for some positive constant c, depending only on h and the regularity parameters. This gives a growth rate of just n\u001C (1/\u000F)1/c, on some sequence of \u000F\u00E2\u0086\u0092 0. 2.2 Main theorem Let us first outline the basic strategy for Theorem 2.3: by restricting n to be divisible by h!, we have that n + i is necessarily divisible by i (for 1 \u00E2\u0089\u00A4 i \u00E2\u0089\u00A4 h), but we retain the freedom to choose finitely many (small) prime divisors for each (n + i)/i independently. Condition 2.1b then allows us to approximate each target \u00CE\u00B1i by the contributions from these small prime divisors, by further restricting n to a particular congruence class. We then use Conditions 2.1a and 2.1c to show that the effect of the large primes on f(n) is harmlessly small for a positive proportion of n within this class. We make use of the following facts, which are well-known or elementary: Proposition 2.4. (a) The number of positive integers n \u00E2\u0089\u00A4 x which satisfy n \u00E2\u0089\u00A1 r (mod Q) and d | n is x/dQ + O(1), uniformly over all naturals d,Q, r where (d,Q) = 1. (b) The number of positive integers n \u00E2\u0089\u00A4 x with at most two prime factors is O(x log log x/ log x) = o(x). We will also use a topological observation, essentially a restatement of the simple fact that a bounded subset of Rh is totally bounded. Note that the symbol \u00E2\u0080\u0096 \u00C2\u00B7 \u00E2\u0080\u0096\u00E2\u0088\u009E below refers to the usual supremum norm; it is unrelated to the \u00E2\u0080\u0096 \u00C2\u00B7 \u00E2\u0080\u0096 notation appearing in Condition 2.1. 29 2.2. Main theorem Proposition 2.5. Let h \u00E2\u0088\u0088 N and let {sn}n>0 be a bounded sequence in Rh. Then for any \u00CE\u00B7 > 0, there exists a finite index set K \u00E2\u008A\u0082 N, such that for any n \u00E2\u0088\u0088 N there exists k \u00E2\u0088\u0088 K such that k \u00E2\u0089\u00A4 n and \u00E2\u0080\u0096sn \u00E2\u0088\u0092 sk\u00E2\u0080\u0096\u00E2\u0088\u009E < \u00CE\u00B7. Proof. Let M > 0 be an upper bound for \u00E2\u0080\u0096sn\u00E2\u0080\u0096\u00E2\u0088\u009E (uniform over all n). We can cover the set S = {sn : n \u00E2\u0088\u0088 N} with finitely many open h-dimensional cubes B1,B2, . . . ,Bm of side length \u00CE\u00B7, such that Bi \u00E2\u0088\u00A9 S 6= \u00E2\u0088\u0085 for each i = 1, 2, . . . ,m (we need at most (1 + 2M/\u00CE\u00B7)h such cubes). We may then take K = {k(i) : 1 \u00E2\u0089\u00A4 i \u00E2\u0089\u00A4 m}, where k(i) is defined as the least k for which sk \u00E2\u0088\u0088 Bi. For any n \u00E2\u0088\u0088 N, we certainly have sn \u00E2\u0088\u0088 Bi for some i, whence n \u00E2\u0089\u00A5 k(i) and \u00E2\u0080\u0096sn \u00E2\u0088\u0092 sk(i)\u00E2\u0080\u0096\u00E2\u0088\u009E < \u00CE\u00B7 since sn and sk(i) lie in the same cube. Lastly, we require the following lemma, which appears as part of equation (21) in [20]. There is a minor flaw in the argument used there, so we include an independent proof for the sake of completeness. Lemma 2.6. There is an absolute constant C > 0 such that for any x > 0, \u00E2\u0088\u0091 \u00E2\u0088\u009A x 4. Switching the order of summation, the above sum is equal to \u00E2\u0088\u0091 \u00E2\u0088\u009A x 0 be fixed. By iterative application of Condition 2.1b, we can find pairwise co-prime integers N1, N2, . . . , Nh > 1 such that (Ni, h!) = 1 (i = 1, 2, . . . , h), |fi(Ni)\u00E2\u0088\u0092 \u00CE\u00B1i + fi(i)| < \u000F 2 (i = 1, 2, . . . , h). (2.4) Let N = \u00E2\u0088\u008Fh i=1Ni and let k1 > h be the largest prime factor of N . For any k > k1, we can define Pk := \u00E2\u0088\u008F hk fi(p r). (2.6) We wish to choose k so that this quantity has magnitude at most \u000F/2 for a positive proportion of n \u00E2\u0088\u0088 Ak. Then by (2.4) we will have |fi(n+ i)\u00E2\u0088\u0092 \u00CE\u00B1i| < \u000F as desired. We define the following parameters (here C is the constant in Lemma 2.6): \u00C2\u00B5 := min(1, \u000F/ \u00E2\u0088\u009A 96Ch), \u00CE\u00B7 := \u000F/ \u00E2\u0088\u009A 48h. (2.7) Let P denote the set of primes p such that |fi(p)| \u00E2\u0089\u00A4 \u00C2\u00B5 for i = 1, 2, . . . , h. For convenience we use the notation P(k) for the subset {p \u00E2\u0088\u0088 P : p > k}. By Condition 2.1a, the sum \u00E2\u0088\u0091 p/\u00E2\u0088\u0088P 1/p \u00E2\u0089\u00A4 \u00C2\u00B5\u00E2\u0088\u00922 \u00E2\u0088\u0091h i=1 \u00E2\u0088\u0091 p \u00E2\u0080\u0096fi(p)\u00E2\u0080\u00962/p must converge. As \u00E2\u0088\u0091 p 1/p 2 also converges, there is an integer k2 \u00E2\u0089\u00A5 k1 such that \u00E2\u0088\u0091 p>k2, p/\u00E2\u0088\u0088P 1 p + \u00E2\u0088\u0091 p>k2 1 p2 < 1 3h . (2.8) For any k > k2 > h, let Nk denote the subset of Ak consisting all n \u00E2\u0088\u0088 Ak with the following property: (n+ 1)(n+ 2) \u00C2\u00B7 \u00C2\u00B7 \u00C2\u00B7 (n+h) is not divisible by any prime p > k outside of P, nor is it divisible by p2 for any prime p > k (here and below, Pk, Qk and Ak remain as constructed above). We estimate the counting function Nk(x) = #{n \u00E2\u0089\u00A4 x : n \u00E2\u0088\u0088 Nk}. The 32 2.2. Main theorem counting function of Ak is clearly Ak(x) = x/Qk + O(1). If n \u00E2\u0089\u00A4 x and n \u00E2\u0088\u0088 Ak \ Nk, then n + i (for some 1 \u00E2\u0089\u00A4 i \u00E2\u0089\u00A4 h) must be divisible by some prime p /\u00E2\u0088\u0088 P with k < p \u00E2\u0089\u00A4 x+ i, or by p2 for some k < p \u00E2\u0089\u00A4 \u00E2\u0088\u009Ax+ i. For any fixed p and i, Proposition 2.4a gives exactly x/pQk + O(1) values of n \u00E2\u0089\u00A4 x where p | n+ i, and x/p2Qk +O(1) values of n \u00E2\u0089\u00A4 x where p2 | n+ i. Thus Ak(x)\u00E2\u0088\u0092Nk(x) \u00E2\u0089\u00A4 h\u00E2\u0088\u0091 i=1 ( \u00E2\u0088\u0091 kk, p/\u00E2\u0088\u0088P 1 p + \u00E2\u0088\u0091 p>k 1 p2 ) +O ( \u00E2\u0088\u0091 p\u00E2\u0089\u00A4x+h h ) \u00E2\u0089\u00A4 x 3Qk + o(x). So Nk(x) \u00E2\u0089\u00A5 23x/Qk + o(x); in other words, Nk has lower density at least 23 relative to the progression Ak. For any n \u00E2\u0088\u0088 Nk, the integer (n+ i)/iNi is a squarefree product of primes p with p > k and p \u00E2\u0088\u0088 P. Thus equation (2.6) restricted to n \u00E2\u0088\u0088 Nk reduces to fi(n+ i)\u00E2\u0088\u0092 fi(iNi) = \u00E2\u0088\u0091 p|n+i p\u00E2\u0088\u0088P(k) fi(p). (2.9) We will show that for x sufficiently large, we can choose k = k(x) so that E(x) = \u00E2\u0088\u0091 n\u00E2\u0088\u0088Nk, n\u00E2\u0089\u00A4x h\u00E2\u0088\u0091 i=1 (fi(n+ i)\u00E2\u0088\u0092 fi(iNi))2 < \u000F 2 12 x Qk + ok(x), (2.10) and thus there are at most 13x/Qk + ok(x) values of n \u00E2\u0088\u0088 Nk up to x for which |fi(n+ i)\u00E2\u0088\u0092 fi(iNi)| \u00E2\u0089\u00A5 \u000F2 for some i = 1, 2, . . . , h. By the previously estimated density of Nk there are at least 13x/Qk + ok(x) values of n \u00E2\u0089\u00A4 x satisfying inequality (2.2). This will be sufficient to prove Theorem 2.3, provided that we have a uniform upper bound on k (hence also on Qk) independent of x: we are free to vary k with x provided that we restrict the choice of k(x) to within a finite set of possible values, so that the (lower) density of Nk is bounded below by a positive quantity. 33 2.2. Main theorem By Condition 2.1a, there exists an integer k3 \u00E2\u0089\u00A5 k2 such that h\u00E2\u0088\u0091 i=1 \u00E2\u0088\u0091 p\u00E2\u0088\u0088P(k3) fi(p) 2 p \u00E2\u0089\u00A4 h\u00E2\u0088\u0091 i=1 \u00E2\u0088\u0091 p>k3 |fi(p)|<\u00C2\u00B5 fi(p) 2 p < \u000F2 24 . (2.11) We have by Condition 2.1c that the partial sums of \u00E2\u0088\u0091 p \u00E2\u0080\u0096fi(p)\u00E2\u0080\u0096/p are bounded, and since \u00E2\u0088\u0091 p/\u00E2\u0088\u0088P \u00E2\u0080\u0096fi(p)\u00E2\u0080\u0096/p converges absolutely by Condition 2.1a, the partial sums of \u00E2\u0088\u0091 p\u00E2\u0088\u0088P \u00E2\u0080\u0096fi(p)\u00E2\u0080\u0096/p are also bounded. We denote these partial sums by Sk,i, that is Sk,i := \u00E2\u0088\u0091 p\u00E2\u0088\u0088P, p\u00E2\u0089\u00A4k \u00E2\u0080\u0096fi(p)\u00E2\u0080\u0096 p = \u00E2\u0088\u0091 p\u00E2\u0088\u0088P, p\u00E2\u0089\u00A4k fi(p) p . Applying Proposition 2.5 to the sequence of h-tuples {(Sk,i)hi=1}k>k3 , we see that there exists a finite index set K \u00E2\u008A\u0082 P(k3) such that for any x > k3 there exists k = k(x) \u00E2\u0088\u0088 K such that k \u00E2\u0089\u00A4 x+ h and for all i = 1, 2, . . . , h, \u00E2\u0088\u00A3\u00E2\u0088\u00A3Sbx+hc,i \u00E2\u0088\u0092 Sk,i\u00E2\u0088\u00A3\u00E2\u0088\u00A3 = \u00E2\u0088\u00A3\u00E2\u0088\u00A3\u00E2\u0088\u00A3\u00E2\u0088\u00A3\u00E2\u0088\u00A3 \u00E2\u0088\u0091 k k3, as required by the previous discussion. We now estimate the sum in (2.10) for this choice of k. By (2.9) we have E(x) = \u00E2\u0088\u0091 n\u00E2\u0088\u0088Nk n\u00E2\u0089\u00A4x h\u00E2\u0088\u0091 i=1 (fi(n+ i)\u00E2\u0088\u0092 fi(iNi))2 = h\u00E2\u0088\u0091 i=1 \u00E2\u0088\u0091 n\u00E2\u0088\u0088Nk n\u00E2\u0089\u00A4x ( \u00E2\u0088\u0091 p|n+i p\u00E2\u0088\u0088P(k) fi(p) )2 \u00E2\u0089\u00A4 h\u00E2\u0088\u0091 i=1 \u00E2\u0088\u0091 n\u00E2\u0088\u0088Ak n\u00E2\u0089\u00A4x+h\u00E2\u0088\u0092i ( \u00E2\u0088\u0091 p|n+i p\u00E2\u0088\u0088P(k) fi(p) 2 + 2 \u00E2\u0088\u0091 pq|n+i p,q\u00E2\u0088\u0088P(k) p>q fi(p)fi(q) ) . 34 2.2. Main theorem Changing the order of summation and applying Proposition 2.4a gives E(x) \u00E2\u0089\u00A4 h\u00E2\u0088\u0091 i=1 ( \u00E2\u0088\u0091 p\u00E2\u0089\u00A4x+h p\u00E2\u0088\u0088P(k) fi(p) 2 ( x pQk +O(h) ) + 2 \u00E2\u0088\u0091 pq\u00E2\u0089\u00A4x+h p,q\u00E2\u0088\u0088P(k) fi(p)fi(q) ( x pqQk +O(h) )) = x Qk ( h\u00E2\u0088\u0091 i=1 \u00E2\u0088\u0091 p\u00E2\u0089\u00A4x+h p\u00E2\u0088\u0088P(k) fi(p) 2 p + 2 h\u00E2\u0088\u0091 i=1 \u00E2\u0088\u0091 pq\u00E2\u0089\u00A4x+h p,q\u00E2\u0088\u0088P(k) p>q fi(p)fi(q) pq ) +O ( h\u00E2\u0088\u0091 i=1 \u00E2\u0088\u0091 p\u00E2\u0089\u00A4x+h h\u00C2\u00B52 + h\u00E2\u0088\u0091 i=1 \u00E2\u0088\u0091 pq\u00E2\u0089\u00A4x+h h\u00C2\u00B52 ) . We apply (2.11) to the first double sum and Proposition 2.4b to the error term to obtain E(x) < x Qk ( \u000F2 24 + 2h max i\u00E2\u0089\u00A4h \u00E2\u0088\u0091 pq\u00E2\u0089\u00A4x+h p,q\u00E2\u0088\u0088P(k) p>q fi(p)fi(q) pq ) + o(x). To prove (2.10) it suffices to show that for each i = 1, 2, . . . , h, 2 \u00E2\u0088\u0091 pq\u00E2\u0089\u00A4x+h p,q\u00E2\u0088\u0088P(k) p>q fi(p)fi(q) pq < \u000F2 24h . (2.13) By equation (2.12) we have for any 1 \u00E2\u0089\u00A4 i \u00E2\u0089\u00A4 h, 2 \u00E2\u0088\u0091 qq fi(p)fi(q) pq < \u000F2 48h + 2 \u00E2\u0088\u0091 pq>x+h p,q\u00E2\u0088\u0088P(k) q 0, the set of natural numbers n satisfying the simultaneous inequalities |f(n+ i)\u00E2\u0088\u0092 f(n+ i\u00E2\u0088\u0092 1)\u00E2\u0088\u0092 \u00CE\u00B1i| < \u000F (i = 1, 2, . . . , h) (2.14) has positive lower density. For instance, any additive function f with f(p) = log\u00E2\u0088\u00921 p satisfies Theo- rem 2.7 (again taking c1(f) = 0 in Condition 2.1b) but not Condition 2.1c. The elimination of Condition 2.1c relies on a delicate cancellation between values of f , so their proof does not readily adapt to multiple functions fi 36 2.3. Generalization to linear functions satisfying this weaker condition. However, it can be generalized in a differ- ent direction: we may replace the consecutive integers n, n + 1, . . . , n + h, at which f is evaluated, by h+ 1 (non-constant) linear functions of n, say a0n+ b0, a1n+ b1, . . . , ahn+ bh (ai \u00E2\u0088\u0088 N, bi \u00E2\u0088\u0088 Z). (2.15) We need to impose one condition on the coefficients ai and bi, which is that no two forms ain+ bi and ajn+ bj may divide one another as polynomials. This requirement is natural as we cannot expect, say, \u00CF\u0086(3n+6) and \u00CF\u0086(2n+4) to approximate independent reals (indeed, the function \u00CF\u0086(3n+ 6)/\u00CF\u0086(2n+ 4) takes only four distinct values). Thus we add the constraint that aibj 6= ajbi (0 \u00E2\u0089\u00A4 i < j \u00E2\u0089\u00A4 h), (2.16) whereupon the following theorem holds: Theorem 2.8. Let h \u00E2\u0088\u0088 N and let (ain + bi)hi=0 be linear forms satisfying (2.15) and (2.16). Then Theorem 2.7 holds with (2.14) replaced by |f(ain+ bi)\u00E2\u0088\u0092 f(ai\u00E2\u0088\u00921n+ bi\u00E2\u0088\u00921)\u00E2\u0088\u0092 \u00CE\u00B1i| < \u000F (i = 1, 2, . . . , h). (2.17) We will see that Theorem 2.8 actually follows by an elementary argument from Theorem 2.7. We first show that Theorem 2.7 still holds when n is restricted to an arbitrary arithmetic progression. We remark that Erdo\u00CC\u008Bs and Schinzel\u00E2\u0080\u0099s techniques were strong enough to show positive relative density even in the case that n is restricted to the set of primes! We make use of the following observation which is immediate upon in- spection of the proof of Theorem 6 [20, p. 477]. There is a minor flaw in the construction which is immaterial to our proposition, and is easily repaired by replacing (h+ 1)! with (h+ 1)!2, similar to our equation (2.5). Proposition 2.9. The conclusion of Theorem 2.7 holds when n is restricted to the congruence class n \u00E2\u0089\u00A1 1 (mod (h+ 1)!). We note that this specific congruence class, just as in the proof of Theo- rem 2.3, is not essential to the construction and was used only as a means of 37 2.3. Generalization to linear functions simplifying the bookkeeping needed in the argument. Indeed, the preceding proposition readily generalizes to the following lemma. Lemma 2.10. Let f(n) be an additive function satisfying Conditions 2.1a and 2.1b. Then for any a \u00E2\u0088\u0088 N and b \u00E2\u0088\u0088 Z, the set of n \u00E2\u0089\u00A1 b (mod a) for which n satisfies (2.14) also has positive lower density. Proof. Without loss of generality, we may assume that b \u00E2\u0089\u00A5 a, so that a | (h + b)!. Now, let h\u00E2\u0080\u00B2 = h + b \u00E2\u0088\u0092 1 and apply Theorem 2.7 to the h\u00E2\u0080\u00B2-tuple (\u00CE\u00B1\u00E2\u0080\u00B21, \u00CE\u00B1\u00E2\u0080\u00B22, . . . , \u00CE\u00B1\u00E2\u0080\u00B2h\u00E2\u0080\u00B2), where \u00CE\u00B1 \u00E2\u0080\u00B2 i+b\u00E2\u0088\u00921 = \u00CE\u00B1i and we are free to choose \u00CE\u00B1 \u00E2\u0080\u00B2 i arbitrarily for i < b. By Proposition 2.9, any n\u00E2\u0080\u00B2 satisfying (2.14) for the tuple (\u00CE\u00B1\u00E2\u0080\u00B2i) is congruent to 1 mod (h + b)!, hence n = n\u00E2\u0080\u00B2 + b \u00E2\u0088\u0092 1 satisfies (2.14) for the requisite h-tuple (\u00CE\u00B11, . . . , \u00CE\u00B1h) as well as the modular constraint. Our strategy for proving Theorem 2.8 is simple: we first transform the problem into a specific instance of Lemma 2.10, and then we correct for the discrepancy introduced by the transformation. Proof of Theorem 2.8. LetA be the lowest common multiple of a0, a1, . . . , ah. By making a substitution of the form n\u00E2\u0086\u0092 n+ k, we may assume that each bi > 0. We multiply each linear form ain + bi by an appropriate integer to obtain the form An + Bi with Bi > 0. Condition (2.16) implies that the natural numbers B0, . . . , Bh are distinct. By reordering the ai and bi in condition (2.17) and making corresponding adjustments to the values \u00CE\u00B11, \u00CE\u00B12, . . . , \u00CE\u00B1h and \u000F, we may assume that the linear forms (ain+ bi) h i=0 are ordered so that B0 < B1 < \u00C2\u00B7 \u00C2\u00B7 \u00C2\u00B7 < Bh. It follows from Lemma 2.10, using h\u00E2\u0080\u00B2 = Bh and the congruence class n\u00E2\u0080\u00B2 \u00E2\u0089\u00A1 0 (mod A), that for any h-tuple (\u00CE\u00B1\u00E2\u0080\u00B21, . . . , \u00CE\u00B1\u00E2\u0080\u00B2h), the set of n such that |f(An+Bi)\u00E2\u0088\u0092 f(An+Bi\u00E2\u0088\u00921)\u00E2\u0088\u0092 \u00CE\u00B1\u00E2\u0080\u00B2i| < \u000F (i = 1, 2, . . . , h) (2.18) has positive lower density. This is clear when Bi\u00E2\u0088\u00921 and Bi are consecutive integers; if \u00E2\u0088\u0086Bi = Bi\u00E2\u0088\u0092Bi\u00E2\u0088\u00921 > 1, then we are free to assign arbitrary targets to the values of f(An+Bi\u00E2\u0088\u00921 + j)\u00E2\u0088\u0092 f(An+Bi\u00E2\u0088\u00921 + j \u00E2\u0088\u0092 1) for 1 \u00E2\u0089\u00A4 j \u00E2\u0089\u00A4 \u00E2\u0088\u0086Bi, provided that the targets sum to \u00CE\u00B1\u00E2\u0080\u00B2i, while replacing \u000F by \u000F/\u00E2\u0088\u0086Bi. 38 2.3. Generalization to linear functions It remains to account for the discrepancy between f(An+Bi) and f(ain+ bi). Since An+Bi = Mi(ain+bi) for some integer Mi | A, and f is additive, the value of f(An+Bi)\u00E2\u0088\u0092 f(ain+ bi) is precisely determined by the p-adic valuations vp(An+Bi) and vp(ain+ bi) for just the finitely many primes p dividing Mi. In particular, if we restrict n to a suitable congruence class so that vp(An) > vp(Bi) for each prime p dividing A and all 0 \u00E2\u0089\u00A4 i \u00E2\u0089\u00A4 h, then vp(An+Bi) takes the constant value vp(Bi) and likewise vp(ain+ bi) takes the constant value vp(bi), so that f(An+Bi)\u00E2\u0088\u0092 f(ain+ bi) = f(Bi)\u00E2\u0088\u0092 f(bi). A second appeal to Lemma 2.10 allows us to do just that, and so Theorem 2.8 holds by taking \u00CE\u00B1\u00E2\u0080\u00B2i = \u00CE\u00B1i + f(Bi)\u00E2\u0088\u0092 f(bi)\u00E2\u0088\u0092 f(Bi\u00E2\u0088\u00921) + f(bi\u00E2\u0088\u00921) in (2.18). The method used to prove Theorem 2.8 is easily adapted to generalize Theorem 2.3, using congruence (2.5) in place of Proposition 2.9, to obtain the corollary below. Note that the constant c now also depends on the coefficients (ai, bi). Corollary 2.11. Let f1, f2, . . . , fh \u00E2\u0088\u0088 F , and let (ain+bi)hi=1 be linear forms satisfying (2.15) and (2.16) (ignoring a0 and b0). Then Theorem 2.3 holds with (2.2) replaced by |fi(ain+ bi)\u00E2\u0088\u0092 \u00CE\u00B1i| < \u000F (i = 1, 2, . . . , h). (2.19) 39 Chapter 3 Eigenvalues of Random Integer Matrices 3.1 Introduction In this chapter we conduct our investigation of the eigenvalues of random integer matrices, in response to the question posed by Hetzel, Liew and Morrison [37]: what is the probability that a random n\u00C3\u0097 n integer matrix is diagonalizable over the field of rational numbers? Since there is no uniform probability distribution on Z, let us clarify the meaning of this question. For k \u00E2\u0089\u00A5 1, let Ik := {\u00E2\u0088\u0092k,\u00E2\u0088\u0092k + 1, . . . , k \u00E2\u0088\u0092 1, k} be the set of integers of height at most k. Since |Ik| = 2k + 1 is finite, we can construct a random n\u00C3\u0097nmatrixM where each entry is chosen independently and uniformly at random from Ik. The probability that M has any given property (such as being diagonalizable over Q) is then a function of the parameter k. Our aim is to determine how this function behaves as k \u00E2\u0086\u0092\u00E2\u0088\u009E; in particular, does it approach a limiting value as k \u00E2\u0086\u0092\u00E2\u0088\u009E? For any given integers n \u00E2\u0089\u00A5 2 and k \u00E2\u0089\u00A5 1, the set of random n \u00C3\u0097 n matrices with entries in Ik forms a well-defined (finite) probability space; it will be convenient to compute probabilities just by counting matrices, so we introduce some notation for certain sets of matrices. Let Mn(k) be the set of all n \u00C3\u0097 n matrices whose entries are in Ik; we are choosing matrices uniformly from Mn(k), a set with cardinality (2k + 1)n2 . The probability that a random matrix in Mn(k) satisfies a particular property is given simply by the number of matrices in Mn(k) having that property, divided by (2k + 1)n 2 . 40 3.1. Introduction For a given integer \u00CE\u00BB, letM\u00CE\u00BBn(k) denote the set of all matrices inMn(k) that have \u00CE\u00BB as an eigenvalue. In particular, M0n(k) is the subset of singu- lar matrices in Mn(k). Likewise, we denote the set of matrices in Mn(k) having at least one integer eigenvalue byMZn(k) = \u00E2\u008B\u0083 \u00CE\u00BB\u00E2\u0088\u0088ZM\u00CE\u00BBn(k). The prob- ability that a random matrix in Mn(k) has an integer eigenvalue is thus |MZn(k)|/(2k + 1)n 2 . Our first result in this chapter affirms and strengthens the conjecture made by Hetzel et al.: for any n \u00E2\u0089\u00A5 2, the probability that a random n \u00C3\u0097 n integer matrix has at least one integer eigenvalue tends to 0 as k \u00E2\u0086\u0092\u00E2\u0088\u009E. We give a quantitative upper bound on this probability: Theorem 3.1. Given any integer n \u00E2\u0089\u00A5 2 and any real number \u00CE\u00B5 > 0, the probability that a randomly chosen matrix in Mn(k) has any integer eigen- values is \u001Cn,\u00CE\u00B5 1/k1\u00E2\u0088\u0092\u00CE\u00B5. Consequently, the probability that a randomly chosen matrix in Mn(k) is diagonalizable over Q is \u001Cn,\u00CE\u00B5 1/k1\u00E2\u0088\u0092\u00CE\u00B5. Given an integer matrix M \u00E2\u0088\u0088Mn(k), a necessary (and nearly sufficient) condition for it to be diagonalizable over Q is that all of its eigenvalues be rational. Moreover, since the characteristic polynomial det(\u00CE\u00BBI\u00E2\u0088\u0092M) is monic with integer coefficients, the familiar \u00E2\u0080\u009Crational roots theorem\u00E2\u0080\u009D [64, \u00C2\u00A74.3] implies that every rational eigenvalue of M must be an integer. Hence any integer matrix that is diagonalizable over the rationals must be contained in MZn(k), and so the second assertion of the theorem follows immediately from the first. The special case n = 2 of Theorem 3.1 was previously obtained in [37] and also earlier by Kowalsky [47]. Unravelling the \u001C-notation, the theorem states that there exists a constant C, depending on n and \u00CE\u00B5, such that |MZn(k)|/|Mn(k)| \u00E2\u0089\u00A4 C/k1\u00E2\u0088\u0092\u00CE\u00B5 for all k \u00E2\u0089\u00A5 1. Note that |Mn(k)| \u001Cn kn 2 (with the implied constant being highly dependent on n), and so the theorem also gives an upper bound for the number of matrices in Mn(k) having at least one integer eigenvalue, that is |MZn(k)| \u001Cn,\u00CE\u00B5 kn 2\u00E2\u0088\u00921+\u00CE\u00B5. The key tool used to establish Theorem 3.1 is the following related esti- 41 3.1. Introduction mate for |M0n(k)|, the number of singular matrices in Mn(k): Lemma 3.2. Given any integer n \u00E2\u0089\u00A5 2 and any real number \u00CE\u00B5 > 0, the probability that a random matrix in Mn(k) is singular is \u001Cn,\u00CE\u00B5 1/k2\u00E2\u0088\u0092\u00CE\u00B5. In other words, |M0n(k)| \u001Cn,\u00CE\u00B5 kn 2\u00E2\u0088\u00922+\u00CE\u00B5. This bound is far from optimal in general (as we discuss in Section 3.10), but it is essentially best possible when n = 2: in the latter part of this chapter, we conduct a detailed investigation of the distribution of eigenvalues for the 2\u00C3\u0097 2 case, in particular obtaining the precise asymptotic |M02(k)| = 96 pi2 k2 log k +O(k2). More generally, we determine the asymptotic incidence of all integer eigenvalues of M2(k), as given by the following theorem: Theorem 3.3. Define the function V : [\u00E2\u0088\u00922, 2]\u00E2\u0086\u0092 R by V (\u00E2\u0088\u0092\u00CE\u00B4) = V (\u00CE\u00B4) and V (\u00CE\u00B4) = \u00EF\u00A3\u00B1\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B2\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B3 4\u00E2\u0088\u0092 2\u00CE\u00B4 \u00E2\u0088\u0092 \u00CE\u00B42 + \u00CE\u00B42 log(1 + \u00CE\u00B4)\u00E2\u0088\u0092 2(1\u00E2\u0088\u0092 \u00CE\u00B4) log(1\u00E2\u0088\u0092 \u00CE\u00B4), if 0 \u00E2\u0089\u00A4 \u00CE\u00B4 < 1, 1 + log 2, if \u00CE\u00B4 = 1, 4\u00E2\u0088\u0092 2\u00CE\u00B4 \u00E2\u0088\u0092 \u00CE\u00B42 + \u00CE\u00B42 log(\u00CE\u00B4 + 1) + 2(\u00CE\u00B4 \u00E2\u0088\u0092 1) log(\u00CE\u00B4 \u00E2\u0088\u0092 1), if 1 < \u00CE\u00B4 \u00E2\u0089\u00A4 \u00E2\u0088\u009A2, \u00CE\u00B42 \u00E2\u0088\u0092 2\u00CE\u00B4 \u00E2\u0088\u0092 (\u00CE\u00B42 \u00E2\u0088\u0092 2\u00CE\u00B4 + 2) log(\u00CE\u00B4 \u00E2\u0088\u0092 1) if \u00E2\u0088\u009A2 < \u00CE\u00B4 \u00E2\u0089\u00A4 2. (3.1) Then for any integer \u00CE\u00BB between \u00E2\u0088\u00922k and 2k, |M\u00CE\u00BB2(k)| = 24V (\u00CE\u00BB/k) pi2 k2 log k +O(k2), (3.2) where the implied constant is absolute. If |\u00CE\u00BB| > 2k then M\u00CE\u00BB2(k) is empty. We remark that the function V (\u00CE\u00B4) is continuous and, with the exception of the points of infinite slope at \u00CE\u00B4 = \u00C2\u00B11, differentiable everywhere (even at \u00CE\u00B4 = \u00C2\u00B12, if we further define V (\u00CE\u00B4) \u00E2\u0089\u00A1 0 for |\u00CE\u00B4| > 2). Technically, equa- tion (3.2) is not an asymptotic formula when \u00CE\u00BB is extremely close to \u00C2\u00B12k, because then the value of V (\u00CE\u00BB/k) can have order of magnitude 1/ log k or smaller, making the \u00E2\u0080\u009Cmain term\u00E2\u0080\u009D no bigger than the error term. However, it truly is an asymptotic formula for |\u00CE\u00BB| < 2k\u00E2\u0088\u0092\u00CF\u0088(k)k/(log k)1/3, where \u00CF\u0088(k) is any function tending to infinity (the exponent 1/3 arises because V (\u00CE\u00B4) approaches 0 cubically as \u00CE\u00B4 tends to 2 from below). 42 3.1. Introduction By summing (3.2) over all possible values of \u00CE\u00BB, we obtain an asymptotic formula for |M\u00CE\u00BB2(k)|, which gives a sharp quantitative answer to the original question, in the special case n = 2. We defer the details of this calculation to Section 3.6. Corollary 3.4. Let C = ( 7 \u00E2\u0088\u009A 2 + 4 + 3 log( \u00E2\u0088\u009A 2 + 1) ) /3pi2 \u00E2\u0089\u0088 0.55873957. The probability that a random matrix in M2(k) has integer eigenvalues is asymptotically equal to C(log k)/k. More precisely, |MZ2 (k)| = 16Ck3 log k +O(k3). (3.3) If M \u00E2\u0088\u0088M2(k) has eigenvalue \u00CE\u00BB, then the scaled matrix k\u00E2\u0088\u00921M has eigen- value \u00CE\u00BB/k, which is the argument of V that appears on the right-hand side of (3.2). A natural probabilistic interpretation of Theorem 3.3 is that for large k, the rational eigenvalues of k\u00E2\u0088\u00921M are distributed according to the density given by V . Note that the entries of k\u00E2\u0088\u00921M are sampled from a discrete, evenly-spaced subset of [\u00E2\u0088\u00921, 1]. As k \u00E2\u0086\u0092 \u00E2\u0088\u009E the random variable describing each entry converges weakly to the uniform distribution on the interval [\u00E2\u0088\u00921, 1]. Let M2([\u00E2\u0088\u00921, 1]) be the probability space of all 2 \u00C3\u0097 2 matrices whose entries are independent random variables drawn from this distribution. One might ask whether the distribution given by Theorem 3.3 merely approximates the continuous distribution of (real) eigenvalues inM2([\u00E2\u0088\u00921, 1]); the answer, perhaps surprisingly, is no, as we establish by also computing this latter distribution explicitly. Theorem 3.5. If M is chosen uniformly from M2([\u00E2\u0088\u00921, 1]), the expected number of eigenvalues of M in the interval [s, t] is \u00E2\u0088\u00AB t s W (\u00CE\u00B4) d\u00CE\u00B4, where W (\u00CE\u00B4) 43 3.1. Introduction is the density function given by W (\u00E2\u0088\u0092\u00CE\u00B4) = W (\u00CE\u00B4) and W (\u00CE\u00B4) = \u00EF\u00A3\u00B1\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B2\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B3 (80 + 20\u00CE\u00B4 + 90\u00CE\u00B42 + 52\u00CE\u00B43 \u00E2\u0088\u0092 107\u00CE\u00B44)/(144(1 + \u00CE\u00B4)) \u00E2\u0088\u0092 (5\u00E2\u0088\u0092 7\u00CE\u00B4 + 8\u00CE\u00B42)(1\u00E2\u0088\u0092 \u00CE\u00B4) log(1\u00E2\u0088\u0092 \u00CE\u00B4)/12 \u00E2\u0088\u0092 \u00CE\u00B4(1\u00E2\u0088\u0092 \u00CE\u00B42) log(1 + \u00CE\u00B4)/4, if 0 \u00E2\u0089\u00A4 \u00CE\u00B4 \u00E2\u0089\u00A4 1, \u00CE\u00B4(20 + 10\u00CE\u00B4 \u00E2\u0088\u0092 12\u00CE\u00B42 \u00E2\u0088\u0092 3\u00CE\u00B43)/(16(1 + \u00CE\u00B4)) + (3\u00CE\u00B4 \u00E2\u0088\u0092 1)(\u00CE\u00B4 \u00E2\u0088\u0092 1) log(\u00CE\u00B4 \u00E2\u0088\u0092 1)/4 + \u00CE\u00B4(\u00CE\u00B42 \u00E2\u0088\u0092 1) log(\u00CE\u00B4 + 1)/4, if 1 \u00E2\u0089\u00A4 \u00CE\u00B4 \u00E2\u0089\u00A4 \u00E2\u0088\u009A2, \u00CE\u00B4(\u00CE\u00B4 \u00E2\u0088\u0092 2)(2\u00E2\u0088\u0092 6\u00CE\u00B4 + 3\u00CE\u00B42)/(16(\u00CE\u00B4 \u00E2\u0088\u0092 1)) \u00E2\u0088\u0092 (\u00CE\u00B4 \u00E2\u0088\u0092 1)3 log(\u00CE\u00B4 \u00E2\u0088\u0092 1)/4, if \u00E2\u0088\u009A2 \u00E2\u0089\u00A4 \u00CE\u00B4 \u00E2\u0089\u00A4 2, 0, if \u00CE\u00B4 \u00E2\u0089\u00A5 2. (3.4) Just like V (\u00CE\u00B4), the function W (\u00CE\u00B4) is continuous and differentiable every- where, again with the exception of two points of infinite slope at \u00CE\u00B4 = \u00C2\u00B11. (The value W (1) = 1532 makes the function continuous there, although the value of a density function at any one point is irrelevant to the distribution.) It also shares the same cubic decay as \u00CE\u00B4 tends to 2 from below. But there are also striking qualitative differences between the functions V and W . In Figure 3.1 we plot these on the same axes, normalized18 so that the area un- der each curve is 2. In the case ofM2(k), this normalization corresponds to conditioning on having integer eigenvalues: that is, scaling V by the proba- bility C(log k)/k from Corollary 3.4. ForM2([\u00E2\u0088\u00921, 1]) we are conditioning on having real eigenvalues, an event which occurs with probability 4972 (this can be obtained by integrating W (\u00CE\u00B4), analogously to the proof of Corollary 3.4, or by a direct integral computation which was done in [37]). Notice that the distribution W (\u00CE\u00B4) is bimodal, having maxima at \u00CE\u00B4 \u00E2\u0089\u0088 \u00C2\u00B10.75030751. Thus, a random matrix in M2([\u00E2\u0088\u00921, 1]) is more likely to have an eigenvalue of near 34 than one near 0. That the opposite is true for V (\u00CE\u00B4) shows that the eigenvalue distribution of MZ2 (k) is not purely the 18We normalize to area 2 since a 2\u00C3\u0097 2 matrix has two eigenvalues. One could view this as the sum of respective probability densities for the upper and lower eigenvalues. 44 3.2. Basic bounds on singularity Figure 3.1: Distribution of real and rational eigenvalues for n = 2 result of magnitude considerations, but also encodes some of the arithmetic structure of the integers up to k. It is true, however, that if we sample the real eigenvalues of matrices in M2(k), then these do converge to the exact distribution given by W (\u00CE\u00B4): eigenvalues vary continuously with the matrix coefficients, so this sampling limit behaves much like a Riemann sum. We may thus view V (\u00CE\u00B4) and W (\u00CE\u00B4) as the distributions of integer and real eigenvalues of M2(k) for large k. 3.2 Basic bounds on singularity It is not hard to show a simpler version of Lemma 3.2, with a singularity bound that is qualitatively weaker but more explicit and also uniform in n: Proposition 3.6. The probability that a random matrix in Mn(k) is sin- gular is at most 1/2k. To see this, we need a lemma which bounds the size of intersection between a linear subspace and a discrete hypercube in Rn. With minor modifications this previously appeared as problem B4 on the 67th William Lowell Putnam Mathematical Competition [44]. The elegant proof below is due to Catalin Zara. 45 3.2. Basic bounds on singularity Lemma 3.7. Let V be a d-dimensional subspace of Rn, and S be any finite set of m real numbers. Then |Sn \u00E2\u0088\u00A9 V | \u00E2\u0089\u00A4 md. Proof. Form a matrix A whose rows are the points in Sn\u00E2\u0088\u00A9V . This has rank at most d, so we can find d columns of A which span all other columns: each row of A can be uniquely determined by the d entries corresponding to those columns. Since |S| = m, there are at most md distinct choices for those entries, hence at most md distinct rows. Proof of Proposition 3.6. Let M be randomly chosen from Mn(k), and let Ei be the event that the ith row of M is contained in the span of the previous i\u00E2\u0088\u0092 1 rows (E1 is just the probability that the first row is zero). If M is singular, then at least one of E1, . . . , En must occur, so the probability that M is singular is at most \u00E2\u0088\u0091n i=1 Pr(Ei). By Lemma 3.7 there are at most (2k+1)i\u00E2\u0088\u00921 possible choices for row i that are spanned by the preceding rows, so Pr(Ei) \u00E2\u0089\u00A4 (2k+1)i\u00E2\u0088\u00921\u00E2\u0088\u0092n. The result then follows by bounding \u00E2\u0088\u0091n i=1 Pr(Ei) with the geometric series \u00E2\u0088\u0091\u00E2\u0088\u009E j=1(2k + 1) \u00E2\u0088\u0092j = 1/2k. Remark. Proposition 3.6 is used with minor alterations in [66], where it was beneficial for the application to have explicit constants that do not grow with n. Unfortunately, the decay rate of O(1/k) is not quite strong enough to yield an \u00E2\u0080\u009Calmost all\u00E2\u0080\u009D result for the rational eigenvalue problem (we need a bound of o(1/k) to get any qualitative analogue of Theorem 3.1). In order to achieve stronger decay with respect to k, we are willing to sacrifice some uniformity \u00E2\u0080\u0094 indeed, our method yields an implicit constant that grows very rapidly with n (like a factorial). We record without proof a well-known bound (see for instance [59, p. 56]) for the number-of-divisors function \u00CF\u0084(n) that will be useful here and in later sections. Although more explicit bounds are available, the one below is quite sufficient for our purposes. Proposition 3.8. For any real \u00CE\u00B4 > 0 and nonzero integer n, \u00CF\u0084(n)\u001C\u00CE\u00B4 |n|\u00CE\u00B4. We first prove Lemma 3.2 in the easiest case n = 2, but in a slightly more general form that will be useful for induction to larger n. We wish to 46 3.2. Basic bounds on singularity show that the probability that a random 2\u00C3\u0097 2 matrix is singular is equally small (\u001C\u00CE\u00B5 k\u00E2\u0088\u00922+\u00CE\u00B5, albeit with a different implied constant) even if we choose the entries from different arithmetic progressions (each having the same cardinality as Ik). Lemma 3.9. Fix positive real numbers \u00CE\u00B1, B, and \u00CE\u00B5. Let k be a positive integer, and let L1(x), L2(x), L3(x), and L4(x) be nonconstant linear poly- nomials whose coefficients are integers at most Bk\u00CE\u00B1 in absolute value. Then the number of solutions to the equation L1(a)L2(b) = L3(c)L4(d) (3.5) with all of a, b, c, and d in Ik is \u001C\u00CE\u00B1,B,\u00CE\u00B5 k2+\u00CE\u00B5. Proof. First we consider the solutions for which both sides of equation (3.5) equal 0. In this case, at least two of the linear factors L1(a), L2(b), L3(c), and L4(d) equal 0. If, for example, L1(a) = 0 and L3(c) = 0 (the other cases are exactly the same), this completely determines the values of a and c; since there are 2k + 1 choices for each of b and d, the total number of solutions for which both sides of equation (3.5) equal 0 is \u001C k2. Otherwise, fix any values for c and d for which the right-hand side of equation (3.5) is nonzero, a total of at most (2k+1)2 \u001C k2 choices. Then the right-hand side is some nonzero integer that is at most (k \u00C2\u00B7Bk\u00CE\u00B1 +Bk\u00CE\u00B1)2 \u00E2\u0089\u00A4 4B2k2+2\u00CE\u00B1 in absolute value, and L1(a) must be a divisor of that integer. By Proposition 3.8 with \u00CE\u00B4 = \u00CE\u00B5/(2 + 2\u00CE\u00B1), the right-hand side of equation (3.5) has only \u001C\u00CE\u00B1,\u00CE\u00B5 (4B2k2+2\u00CE\u00B1)\u00CE\u00B5/(2+2\u00CE\u00B1) \u001C\u00CE\u00B1,B,\u00CE\u00B5 k\u00CE\u00B5 divisors which may serve as candidates for L1(a); each of these completely determines a value for a (which might not even be an integer). Then the possible values for L2(b) and hence b are determined as well. We conclude that there are a total of \u001C\u00CE\u00B1,B,\u00CE\u00B5 k2+\u00CE\u00B5 solutions to equation (3.5) as claimed. Remark. It is not important that the Li be linear: the above proof works essentially without change for any four nonconstant polynomials of bounded degree. We will not need such a generalization, however, as the determinant of a matrix depends only linearly on each matrix element. 47 3.3. Eigenvalues and determinants Remark. For our application, both \u00CE\u00B1 and B will depend only on the single parameter n, so the dependence\u001C\u00CE\u00B1,B,\u00CE\u00B5 may be rephrased as\u001Cn,\u00CE\u00B5, absorbing the explicit references to \u00CE\u00B1 and B. 3.3 Eigenvalues and determinants We previously noted that if M \u00E2\u0088\u0088Mn(k) has a rational eigenvalue \u00CE\u00BB then \u00CE\u00BB must be an integer. Furthermore, \u00CE\u00BB could be as high as nk (if we take M to be the n\u00C3\u0097n matrix with all entries equal to k), or as low as \u00E2\u0088\u0092nk (if we take M to be the negation of the previous example). In fact it is not possible for \u00CE\u00BB to be any larger (even when it is complex), and an analogous bound holds for the continuous case as shown by the lemma below. Lemma 3.10. Any eigenvalue of a matrix in Mn(k) is bounded in abso- lute value by nk. Any eigenvalue of a matrix in Mn([\u00E2\u0088\u00921, 1]) is bounded in absolute value by n. Proof. We invoke Gershgorin\u00E2\u0080\u0099s \u00E2\u0080\u009Ccircle theorem\u00E2\u0080\u009D [23], a standard result in spectral theory: let M = (mij) be an n \u00C3\u0097 n matrix, and let D(z, r) denote the disk of radius r around the complex number z. Then Gershgorin\u00E2\u0080\u0099s theorem says that all of the eigenvalues of M must lie in the union of the disks D ( m11, \u00E2\u0088\u0091 1\u00E2\u0089\u00A4j\u00E2\u0089\u00A4n j 6=1 |m1j | ) , D ( m22, \u00E2\u0088\u0091 1\u00E2\u0089\u00A4j\u00E2\u0089\u00A4n j 6=2 |m2j | ) , . . . , D ( mnn, \u00E2\u0088\u0091 1\u00E2\u0089\u00A4j\u00E2\u0089\u00A4n j 6=n |mnj | ) . In particular, if all entries of M are bounded in absolute value by B, then all eigenvalues are bounded in absolute value by nB. Remark. The examples preceding the lemma show that the bound is tight; indeed, one could also prove this using Perron\u00E2\u0080\u0093Frobenius theory for the spectral radius of M , which provides some intuition for the fact that the all-k\u00E2\u0080\u0099s matrix is maximal. The next ingredient is a curious determinantal identity, which was clas- sically known but presently seems to have fallen out of common knowledge. 48 3.3. Eigenvalues and determinants Before we state this identity, it will be very helpful to define some prelim- inary notation. For the remainder of this section, we will adhere to the convention that capital letters denote matrices, boldface lowercase letters denote column vectors, and regular lowercase letters denote scalars. Let In denote the n \u00C3\u0097 n identity matrix, whose jth column we denote by ej (the jth standard basis vector). Let M be an n \u00C3\u0097 n matrix, and let mj denote its jth column and mij its ijth entry. Note that Mej = mj by the definition of matrix multiplication. We caution the reader that, in the definitions below, aij does not denote the entries of A, nor those of aj . Let aij denote the ijth cofactor of M , that is, the determinant of the (n \u00E2\u0088\u0092 1) \u00C3\u0097 (n \u00E2\u0088\u0092 1) matrix obtained from M by deleting its ith row and jth column. Let A = Adj(M) denote the adjugate matrix of M , that is, the matrix whose ijth entry is (\u00E2\u0088\u00921)i+jaji. It is a standard consequence of Laplace\u00E2\u0080\u0099s determinant expansion [36, \u00C2\u00A74.III]19 that MA = (detM)In. Finally, let aj denote the jth column of A. Note that Maj = (detM)ej , since both sides are the jth column of (detM)In. Lemma 3.11. Fix an integer n \u00E2\u0089\u00A5 3. Let M be an n \u00C3\u0097 n matrix with cofactors aij. Also let Z denote the (n\u00E2\u0088\u0092 2)\u00C3\u0097 (n\u00E2\u0088\u0092 2) matrix obtained from M by deleting the first two rows and first two columns, so that M = \u00EF\u00A3\u00AB\u00EF\u00A3\u00AC\u00EF\u00A3\u00AD m11 m12 *m21 m22 * \u00E2\u0088\u0097 \u00E2\u0088\u0097 Z \u00EF\u00A3\u00B6\u00EF\u00A3\u00B7\u00EF\u00A3\u00B8 , (3.6) where \u00E2\u0088\u0097 represents irrelevant entries. Then a11a22\u00E2\u0088\u0092a12a21 = (detM)(detZ). It is important to note that when detZ 6= 0, the cofactor a11 is a linear polynomial of m22 whose leading coefficient is detZ, and whose constant term depends only on Z and the starred entries; conversely, the cofactor a22 is a linear polynomial of m11 with leading coefficient detZ (similar corre- spondences hold for the pair a12 and a21). For example, when n = 3 the determinant of the 1\u00C3\u00971 matrix Z is simply the lower-right entry m33 of M ; 19Our definition of cofactor differs slightly from that in [36], which includes the (\u00E2\u0088\u00921)i+j term. We prefer this for simplicity as we work mostly with aij and not the adjugate A. 49 3.3. Eigenvalues and determinants the identity in question thus translates to (m11m33 \u00E2\u0088\u0092m13m31)(m22m33 \u00E2\u0088\u0092m23m32) \u00E2\u0088\u0092 (m12m33 \u00E2\u0088\u0092m13m32)(m21m33 \u00E2\u0088\u0092m23m31) = m33 detM. (3.7) For any fixed dimension n, the assertion of Lemma 3.11 is just a polynomial identity like the above, which is verifiable by direct computation; however, a proof that works for general n requires a bit of cunning. Proof. Define an n\u00C3\u0097 n matrix B = ( a1 a2 e3 \u00C2\u00B7 \u00C2\u00B7 \u00C2\u00B7 en ) = \u00EF\u00A3\u00AB\u00EF\u00A3\u00AC\u00EF\u00A3\u00AD a11 \u00E2\u0088\u0092a21 0\u00E2\u0088\u0092a12 a22 0 \u00E2\u0088\u0097 \u00E2\u0088\u0097 In\u00E2\u0088\u00922 \u00EF\u00A3\u00B6\u00EF\u00A3\u00B7\u00EF\u00A3\u00B8 . Since B is in lower-triangular block form, its determinant is easy to evaluate: detB = det ( a11 \u00E2\u0088\u0092a21 \u00E2\u0088\u0092a12 a22 ) \u00C2\u00B7 det In\u00E2\u0088\u00922 = a11a22 \u00E2\u0088\u0092 a12a21. Moreover, MB = ( Ma1 Ma2 Me3 \u00C2\u00B7 \u00C2\u00B7 \u00C2\u00B7 Men ) = ( (detM)e1 (detM)e2 m3 \u00C2\u00B7 \u00C2\u00B7 \u00C2\u00B7 mn ) = \u00EF\u00A3\u00AB\u00EF\u00A3\u00AC\u00EF\u00A3\u00AD detM 0 \u00E2\u0088\u00970 detM \u00E2\u0088\u0097 0 0 Z \u00EF\u00A3\u00B6\u00EF\u00A3\u00B7\u00EF\u00A3\u00B8 . Since MB is in upper-triangular block form, its determinant det(MB) = (detM)2(detZ) is also easy to evaluate. Using the identity detM \u00C2\u00B7 detB = det(MB), we conclude that (detM)(a11a22 \u00E2\u0088\u0092 a12a21) = (detM)2(detZ). Both sides of this last identity are polynomial functions of the n2 variables mij representing the entries of M . The factor detM on both sides is a 50 3.4. Proof of main results for n \u00E2\u0089\u00A5 2 nonzero polynomial, and hence it can be canceled to obtain (detM)(detZ) = a11a22 \u00E2\u0088\u0092 a12a21 as desired. Remark. This proof generalizes readily to an analogous identity for larger minors of the adjugate matrix A. Indeed, Muir\u00E2\u0080\u0099s classic treatise on deter- minants [60, Ch. VI, \u00C2\u00A7175] includes this general form of Lemma 3.11 in a chapter wholly devoted to compound determinants (that is, determinants of matrices whose elements are themselves determinants). The same result can also be found in Scott\u00E2\u0080\u0099s reference of equally old vintage [75, p. 62], which was made freely available online by the Cornell University Library Historical Math collection.20 3.4 Proof of main results for n \u00E2\u0089\u00A5 2 We are now ready to prove the singularity bound of Lemma 3.2. We proceed by an induction on n, which requires n = 2 and n = 3 as base cases. In the arguments that follow, let M denote a matrix in Mn(k) with entries mij . Base case n = 2: The determinant of ( m11 m21 m12 m22 ) is 0 precisely when m11m22 = m12m21. By Lemma 3.9, there are \u001C\u00CE\u00B5 k2+\u00CE\u00B5 solutions to this equation with the variables m11, m12, m21, and m22 all in Ik. This imme- diately shows that |M02(k)| \u001C\u00CE\u00B5 k2+\u00CE\u00B5 as claimed. Since 1/|M2(k)| \u001C k\u00E2\u0088\u00924, we see that the probability of a randomly chosen matrix from M2(k) being singular is |M02(k)|/|M2(k)| \u001C\u00CE\u00B5 k\u00E2\u0088\u00922+\u00CE\u00B5. Base case n = 3: We first estimate the number of matrices in M03(k) whose lower right-hand entry m33 is nonzero. Fix the five entries in the last row and last column ofM , withm33 6= 0; there are a total of 2k(2k+1)4 \u001C k5 possibilities. Using the earlier identity (3.7), we see that if detM = 0 then 20http://ebooks.library.cornell.edu/cgi/t/text/text-idx?c=math;idno= 01670002 51 3.4. Proof of main results for n \u00E2\u0089\u00A5 2 we must have (m11m33 \u00E2\u0088\u0092m13m31)(m22m33 \u00E2\u0088\u0092m23m32) = (m12m33 \u00E2\u0088\u0092m13m32)(m21m33 \u00E2\u0088\u0092m23m31). This equation is of the form L1(m11)L2(m22) = L3(m12)L4(m21), where the Li are nonconstant linear polynomials whose coefficients are at most k 2 in absolute value. (Note that we have used the fact that m33 6= 0 in asserting that the Li are nonconstant.) Applying Lemma 3.9 with \u00CE\u00B1 = 2, we see that there are \u001C\u00CE\u00B5 k2+\u00CE\u00B5 solutions to this equation with m11, m12, m21, and m22 all in Ik. This shows that there are\u001C\u00CE\u00B5 k7+\u00CE\u00B5 matrices inM03(k) whose lower right-hand entry m33 is nonzero. If any of the entries in the last row of M is nonzero, then we can swap two columns of M , if necessary, to bring that entry into the lower right- hand position; any matrix so formed corresponds to at most three distinct matrices in M03(k), and so there are still \u001C\u00CE\u00B5 k7+\u00CE\u00B5 matrices in M03(k) that have any nonzero entry in the last row. Finally, any matrix whose last row consists of all zeros is certainly inM03(k), but there are only (2k+ 1)6 \u001C k6 such matrices. We conclude that the total number of matrices in M03(k) is \u001C\u00CE\u00B5 k7+\u00CE\u00B5, so that the probability of a randomly chosen matrix from M3(k) being singular is |M03(k)|/|M3(k)| \u001C\u00CE\u00B5 k\u00E2\u0088\u00922+\u00CE\u00B5 as claimed. Inductive step for n \u00E2\u0089\u00A5 4: Write a matrix M \u00E2\u0088\u0088 Mn(k) in the form (3.6). Some such matrices will have detZ = 0; however, by the induction hypothesis for n \u00E2\u0088\u0092 2, the probability that this occurs is \u001Cn,\u00CE\u00B5 k\u00E2\u0088\u00922+\u00CE\u00B5 (inde- pendent of the entries outside Z), which is an allowably small probability. Otherwise, fix values in Ik for the n2 \u00E2\u0088\u0092 4 entries other than m11, m12, m21, and m22 such that detZ 6= 0. It suffices to show that, conditioned on any such fixed values, the probability that M is singular is \u001Cn,\u00CE\u00B5 k\u00E2\u0088\u00922+\u00CE\u00B5, as we let the remaining variables m11, m12, m21, and m22 range over Ik, . By Lemma 3.11, we see that detM = 0 is equivalent to a11a22 = a12a21. Recall that the cofactor a11 is a linear polynomial of m22 with leading coef- ficient detZ, while a22 is a linear polynomial of m11 with leading coefficient detZ (and similarly for the pair a12 and a21). Moreover, the coefficients 52 3.5. Preliminaries for 2\u00C3\u0097 2 matrices of these polynomials are at most (n \u00E2\u0088\u0092 1)! kn\u00E2\u0088\u00921 in magnitude: by Laplace expansion, each coefficient is the sum of at most (n \u00E2\u0088\u0092 1)! products, each composed of at most n \u00E2\u0088\u0092 1 entries of M . We may thus apply Lemma 3.9 with \u00CE\u00B1 = n\u00E2\u0088\u0092 1 and B = (n\u00E2\u0088\u0092 1)! to see that the (conditional) probability of a11a22 = a12a21 is \u001Cn,\u00CE\u00B5 k\u00E2\u0088\u00922+\u00CE\u00B5, as desired. We now have everything we need to prove Theorem 3.1: given any integer n \u00E2\u0089\u00A5 2 and any real number \u00CE\u00B5 > 0, the probability that a randomly chosen matrix inMn(k) has an integer eigenvalue is\u001Cn,\u00CE\u00B5 1/k1\u00E2\u0088\u0092\u00CE\u00B5. By Lemma 3.10, any such eigenvalue \u00CE\u00BB is at most nk in absolute value. For each individual \u00CE\u00BB, we observe that if M \u00E2\u0088\u0088Mn(k) has eigenvalue \u00CE\u00BB, then M \u00E2\u0088\u0092\u00CE\u00BBIn is a singular matrix with integer entries bounded in size by k + |\u00CE\u00BB| \u00E2\u0089\u00A4 (n + 1)k. Thus, every matrix in M\u00CE\u00BBn(k) is also contained in the set{ M + \u00CE\u00BBIn : M \u00E2\u0088\u0088M0n ( (n+ 1)k )} . By Lemma 3.2, the cardinality of this set is \u001Cn,\u00CE\u00B5 ((n + 1)k)n2\u00E2\u0088\u00922+\u00CE\u00B5 \u001Cn,\u00CE\u00B5 kn 2\u00E2\u0088\u00922+\u00CE\u00B5 for any fixed \u00CE\u00BB. Summing over all integer values of \u00CE\u00BB from \u00E2\u0088\u0092nk to nk (admittedly, some matrices are counted multiple times, but the upper bound thus obtained is still valid), we conclude that the total number of matrices in MZn(k) is \u001Cn,\u00CE\u00B5 kn 2\u00E2\u0088\u00921+\u00CE\u00B5. In other words, the probability that a matrix inMn(k) has an integer eigenvalue is |MZn(k)|/|Mn(k)| \u001Cn,\u00CE\u00B5 1/k1\u00E2\u0088\u0092\u00CE\u00B5, as desired. 3.5 Preliminaries for 2\u00C3\u0097 2 matrices Having established Theorem 3.1 and Lemma 3.2 for general n, we can give considerably sharper results in the simplest case n = 2. The remainder of this chapter is devoted to the proofs of Theorem 3.3 and Corollary 3.4 for integer eigenvalues, and also Theorem 3.5 for real eigenvalues. The 2 \u00C3\u0097 2 case is particularly nice: since the trace of an integer matrix is itself an integer, it follows that if one eigenvalue is an integer then both are. Consequently, there is no significant distinction between belonging to 53 3.5. Preliminaries for 2\u00C3\u0097 2 matrices MZ2 (k) and being diagonalizable over Q (if M has rational eigenvalues but is not rationally diagonalizable, then its eigenvalues must be repeated, and we confirm in Lemma 3.14 that this occurs very rarely). We begin with some elementary observations for 2 \u00C3\u0097 2 matrices (both integer-valued and real-valued) that will simplify our computations of the functions V (\u00CE\u00B4) and W (\u00CE\u00B4). The key to the precise enumeration of M\u00CE\u00BB2(k) is the simple structure of singular integer matrices: Lemma 3.12. For any singular matrix M \u00E2\u0088\u0088 M02(k), either at least two entries of M equal zero, or else there exist nonzero integers a, b, c, d with (a, b) = 1 such that M = ( ac bc ad bd ) . (3.8) Moreover, this representation of M is unique up to replacing each of a, b, c, d by its negative. Proof. If one of the entries of M equals zero, then a second one must equal zero as well for the determinant to vanish. Otherwise, given M = ( m11 m12 m21 m22 ) with none of the mij equal to zero, define c = (m11,m12), and set a = m11/c and b = m12/c, so that (a, b) = 1. Since M is singular, the second row of m must be a multiple of the first row\u00E2\u0080\u0094that is, there exists a real number d such that m21 = ad and m22 = bd. Since a and b are relatively prime, moreover, d must in fact be an integer. This shows that every such matrix admits a representation of type (3.8). If there is another representation M = ( a\u00E2\u0080\u00B2c\u00E2\u0080\u00B2 b\u00E2\u0080\u00B2c\u00E2\u0080\u00B2 a\u00E2\u0080\u00B2d\u00E2\u0080\u00B2 b\u00E2\u0080\u00B2d\u00E2\u0080\u00B2 ) , then (a\u00E2\u0080\u00B2, b\u00E2\u0080\u00B2) = 1 implies (a\u00E2\u0080\u00B2c\u00E2\u0080\u00B2, b\u00E2\u0080\u00B2c\u00E2\u0080\u00B2) = |c\u00E2\u0080\u00B2|, which shows that |c\u00E2\u0080\u00B2| = c; the equalities |a\u00E2\u0080\u00B2| = |a|, |b\u00E2\u0080\u00B2| = |b|, and |d\u00E2\u0080\u00B2| = |d| follow quickly. 54 3.5. Preliminaries for 2\u00C3\u0097 2 matrices For a 2\u00C3\u0097 2 matrix M = (ab cd), we define discM = (trM)2 \u00E2\u0088\u0092 4 detM = (a\u00E2\u0088\u0092 d)2 + 4bc, where trM is the trace a+ d and detM is the determinant ad\u00E2\u0088\u0092 bc. It is easily seen that discM is the discriminant of the characteristic polynomial of M . We record the following elementary facts, which will be useful in the proofs of Lemma 3.14 and Proposition 3.20. Lemma 3.13. Let M be a 2\u00C3\u0097 2 matrix with real entries. (a) M has repeated eigenvalues if and only if discM = 0. (b) M has real eigenvalues if and only if discM \u00E2\u0089\u00A5 0. (c) detM < 0 if and only if M has two real eigenvalues of opposite sign. (d) If detM > 0 and discM \u00E2\u0089\u00A5 0, then the eigenvalues of M have the same sign as trM . Proof. Let \u00CE\u00BB1, \u00CE\u00BB2 denote the eigenvalues of M , so that trM = \u00CE\u00BB1 + \u00CE\u00BB2, detM = \u00CE\u00BB1\u00CE\u00BB2, and discM = (\u00CE\u00BB1\u00E2\u0088\u0092\u00CE\u00BB2)2, each of which is real. Parts (a), (b) and (d) follow immediately from these observations, and part (c) from the fact that \u00CE\u00BB2 = \u00CE\u00BB1 if \u00CE\u00BB1, \u00CE\u00BB2 are complex. The next lemma gives a bound for the probability of a matrix having repeated eigenvalues. It is natural to expect this probability to approach 0 as k increases, and indeed such a result was obtained in [37] for matrices of arbitrary size, with an upper bound of On(1/k). We give a simple proof of a stronger bound for the 2 \u00C3\u0097 2 case, as well as an analogous qualitative statement for real matrices which will be helpful in the proof of Theorem 3.5. Lemma 3.14. The number of matrices in M2(k) with a repeated eigen- value is \u001C\u00CE\u00B5 k2+\u00CE\u00B5 for every \u00CE\u00B5 > 0. The probability that a random matrix in M2([\u00E2\u0088\u00921, 1]) has a repeated eigenvalue or is singular is 0. Proof. By Lemma 3.13(a), the 2\u00C3\u0097 2 matrix (ac bd) has a double eigenvalue if and only if 4bc = \u00E2\u0088\u0092(a\u00E2\u0088\u0092 d)2. For matrices in M2([\u00E2\u0088\u00921, 1]) this is easily seen to be a zero-probability event, as is the event that detM = ad\u00E2\u0088\u0092 bc = 0. 55 3.6. Enumeration theorems for integer eigenvalues For matrices inM2(k), we enumerate how many can satisfy 4bc = \u00E2\u0088\u0092(a\u00E2\u0088\u0092 d)2. This follows the same argument as Lemma 3.9: if a = d then one of b or c must be 0, yielding at most \u001C k2 solutions. Otherwise, this there are O(k2) choices for a and d that make (a \u00E2\u0088\u0092 d) nonzero, each yielding \u001C\u00CE\u00B5 k\u00CE\u00B5 choices for b, c. Combining the two yields \u001C\u00CE\u00B5 k2+\u00CE\u00B5 solutions in total. 3.6 Enumeration theorems for integer eigenvalues Let \u00C2\u00B5(n) be the Mo\u00CC\u0088bius function, characterized by the identity \u00E2\u0088\u0091 d|n \u00C2\u00B5(d) = \u00EF\u00A3\u00B1\u00EF\u00A3\u00B2\u00EF\u00A3\u00B31, if n = 1,0, if n > 1. (3.9) The well-known Dirichlet series identity 1/\u00CE\u00B6(s) = \u00E2\u0088\u0091\u00E2\u0088\u009E n=1 \u00C2\u00B5(n)n \u00E2\u0088\u0092s is valid for 1 (see [59, Corollary 1.10], for example). In particular, \u00E2\u0088\u0091 \u00C2\u00B5(n) n2 = 6 pi2 , and we can estimate the tail of this series (using |\u00C2\u00B5(n)| \u00E2\u0089\u00A4 1) to obtain the quantitative estimate \u00E2\u0088\u0091 d\u00E2\u0089\u00A4k \u00C2\u00B5(d) d2 = 6 pi2 +O ( 1 k ) . (3.10) Lemma 3.15. For nonzero integers a, b and parameters k, \u00CE\u00BB, define the function Nk,\u00CE\u00BB(a, b) = #{(c, d) \u00E2\u0088\u0088 Z2, c 6= 0, d 6= 0 : |ac+ \u00CE\u00BB|, |bc|, |ad|, |bd+ \u00CE\u00BB| \u00E2\u0089\u00A4 k}. (3.11) Then \u00E2\u0088\u00A3\u00E2\u0088\u00A3M\u00CE\u00BB2(k)\u00E2\u0088\u00A3\u00E2\u0088\u00A3 = 4\u00E2\u0088\u0091 d\u00E2\u0089\u00A4k \u00C2\u00B5(d) \u00E2\u0088\u0091 1\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2\u00E2\u0089\u00A4k/d Nk,\u00CE\u00BB(d\u00CE\u00B1, d\u00CE\u00B2) +O(k 2), where the implied constant is independent of \u00CE\u00BB and k. Proof. Fix an integer 0 \u00E2\u0089\u00A4 \u00CE\u00BB \u00E2\u0089\u00A4 2k, and let M \u00E2\u0088\u0088 M\u00CE\u00BB2(k), so that M \u00E2\u0088\u0092 \u00CE\u00BBI is 56 3.6. Enumeration theorems for integer eigenvalues singular. By Lemma 3.12, either at least two entries of M \u00E2\u0088\u0092 \u00CE\u00BBI equal zero, or else M \u00E2\u0088\u0092 \u00CE\u00BBI has exactly two representations of the form (3.8). In the former case, there are 2k+ 1 choices for each of the two potentially nonzero entries, hence O(k2) such matrices in total (even taking into account the several different choices of which two entries are nonzero). In the latter case, there are exactly two corresponding quadruples a, b, c, d of integers as in Lemma 3.12. Taking into account that each entry of M must be at most k in absolute value, we deduce that \u00E2\u0088\u00A3\u00E2\u0088\u00A3M\u00CE\u00BB2(k)\u00E2\u0088\u00A3\u00E2\u0088\u00A3 = 12 #{(a, b, c, d) \u00E2\u0088\u0088 Z4 : a, b, c, d 6= 0, (a, b) = 1, |ac+ \u00CE\u00BB|, |bc|, |ad|, |bd+ \u00CE\u00BB| \u00E2\u0089\u00A4 k}+O(k2) = 12 \u00E2\u0088\u0091 1\u00E2\u0089\u00A4|a|,|b|\u00E2\u0089\u00A4k (a,b)=1 Nk,\u00CE\u00BB(a, b) +O(k 2). Because of the symmetries Nk,\u00CE\u00BB(a, b) = Nk,\u00CE\u00BB(\u00E2\u0088\u0092a, b) = Nk,\u00CE\u00BB(a,\u00E2\u0088\u0092b) = Nk,\u00CE\u00BB(\u00E2\u0088\u0092a,\u00E2\u0088\u0092b), we have\u00E2\u0088\u00A3\u00E2\u0088\u00A3M\u00CE\u00BB2(k)\u00E2\u0088\u00A3\u00E2\u0088\u00A3 = 2 \u00E2\u0088\u0091 1\u00E2\u0089\u00A4a,b\u00E2\u0089\u00A4k (a,b)=1 Nk,\u00CE\u00BB(a, b) +O(k 2). The only term in the sum where a = b is the term a = b = 1, and for all other terms we can invoke the additional symmetry Nk,\u00CE\u00BB(a, b) = Nk,\u00CE\u00BB(b, a), seen to be valid by switching the roles of c and d in the definition (3.11) of Nk,\u00CE\u00BB(a, b). We obtain\u00E2\u0088\u00A3\u00E2\u0088\u00A3M\u00CE\u00BB2(k)\u00E2\u0088\u00A3\u00E2\u0088\u00A3 = 4 \u00E2\u0088\u0091 1\u00E2\u0089\u00A4a e and |1 \u00E2\u0088\u0092 \u00CE\u00B4| log |1 \u00E2\u0088\u0092 \u00CE\u00B4|\u00E2\u0088\u00921 \u001C 1 \u001C 1/\u00CE\u00B2 if 1 \u00E2\u0089\u00A4 \u00CE\u00B2 \u00E2\u0089\u00A4 e. In either case, the last error term can be simplified to O((1 + log \u00CE\u00B2)/\u00CE\u00B22), which establishes the lemma in second case. Proof of Proposition 3.17. We consider separately the four cases correspond- ing to the different parts of the definition (3.1) of V (\u00CE\u00B4). For brevity we will suppress the dependence on \u00CE\u00B4, \u00CE\u00B1 and \u00CE\u00B2 from the notation for C and D. \u00E2\u0080\u00A2 Case 1: 0 \u00E2\u0089\u00A4 \u00CE\u00B4 < 1. In this case we have 0 < 1\u00E2\u0088\u0092 \u00CE\u00B4 < (1 + \u00CE\u00B4)\u00E2\u0088\u00921 \u00E2\u0089\u00A4 1 and C = \u00EF\u00A3\u00B1\u00EF\u00A3\u00B2\u00EF\u00A3\u00B3 2\u00CE\u00B2 , if \u00CE\u00B1 \u00E2\u0089\u00A4 (1\u00E2\u0088\u0092 \u00CE\u00B4)\u00CE\u00B2,1\u00E2\u0088\u0092\u00CE\u00B4 \u00CE\u00B1 + 1 \u00CE\u00B2 , if (1\u00E2\u0088\u0092 \u00CE\u00B4)\u00CE\u00B2 \u00E2\u0089\u00A4 \u00CE\u00B1 and D = \u00EF\u00A3\u00B1\u00EF\u00A3\u00B2\u00EF\u00A3\u00B3 2\u00CE\u00B2 , if \u00CE\u00B1 \u00E2\u0089\u00A4 (1 + \u00CE\u00B4)\u00E2\u0088\u00921\u00CE\u00B2,1\u00E2\u0088\u0092\u00CE\u00B4 \u00CE\u00B2 + 1 \u00CE\u00B1 , if (1 + \u00CE\u00B4) \u00E2\u0088\u00921\u00CE\u00B2 \u00E2\u0089\u00A4 \u00CE\u00B1. Therefore \u00E2\u0088\u0091 1\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2 CD = \u00E2\u0088\u0091 1\u00E2\u0089\u00A4\u00CE\u00B1<(1\u00E2\u0088\u0092\u00CE\u00B4)\u00CE\u00B2 2 \u00CE\u00B2 \u00C2\u00B7 2 \u00CE\u00B2 + \u00E2\u0088\u0091 max{1,(1\u00E2\u0088\u0092\u00CE\u00B4)\u00CE\u00B2}\u00E2\u0089\u00A4\u00CE\u00B1<(1+\u00CE\u00B4)\u00E2\u0088\u00921\u00CE\u00B2 ( 1\u00E2\u0088\u0092 \u00CE\u00B4 \u00CE\u00B1 + 1 \u00CE\u00B2 ) 2 \u00CE\u00B2 + \u00E2\u0088\u0091 (1+\u00CE\u00B4)\u00E2\u0088\u00921\u00CE\u00B2\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2 ( 1\u00E2\u0088\u0092 \u00CE\u00B4 \u00CE\u00B1 + 1 \u00CE\u00B2 )( 1\u00E2\u0088\u0092 \u00CE\u00B4 \u00CE\u00B2 + 1 \u00CE\u00B1 ) . (3.15) 64 3.7. Proof of key proposition for enumeration (The first sum might be empty, but this does not invalidate the argument that follows.) The first sum is simply 4 \u00CE\u00B22 \u00E2\u0088\u0091 1\u00E2\u0089\u00A4\u00CE\u00B1<(1\u00E2\u0088\u0092\u00CE\u00B4)\u00CE\u00B2 1 = 4 \u00CE\u00B22 ( (1\u00E2\u0088\u0092 \u00CE\u00B4)\u00CE\u00B2 +O(1)) = 4(1\u00E2\u0088\u0092 \u00CE\u00B4) \u00CE\u00B2 +O ( 1 \u00CE\u00B22 ) . By Lemma 3.18, the second sum is \u00E2\u0088\u0091 max{1,(1\u00E2\u0088\u0092\u00CE\u00B4)\u00CE\u00B2}\u00E2\u0089\u00A4\u00CE\u00B1<(1+\u00CE\u00B4)\u00E2\u0088\u00921\u00CE\u00B2 ( 1\u00E2\u0088\u0092 \u00CE\u00B4 \u00CE\u00B1 + 1 \u00CE\u00B2 ) 2 \u00CE\u00B2 = 2 \u00CE\u00B2 ( 1 1 + \u00CE\u00B4 + \u00CE\u00B4 \u00E2\u0088\u0092 1\u00E2\u0088\u0092 (1\u00E2\u0088\u0092 \u00CE\u00B4) log(1\u00E2\u0088\u0092 \u00CE\u00B42) ) +O ( 1 + log \u00CE\u00B2 \u00CE\u00B22 ) , while the third sum is\u00E2\u0088\u0091 (1+\u00CE\u00B4)\u00E2\u0088\u00921\u00CE\u00B2\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2 ( 1\u00E2\u0088\u0092 \u00CE\u00B4 \u00CE\u00B1 + 1 \u00CE\u00B2 )( 1\u00E2\u0088\u0092 \u00CE\u00B4 \u00CE\u00B2 + 1 \u00CE\u00B1 ) = (1\u00E2\u0088\u0092 \u00CE\u00B4) \u00E2\u0088\u0091 (1+\u00CE\u00B4)\u00E2\u0088\u00921\u00CE\u00B2\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2 1 \u00CE\u00B12 + \u00CE\u00B42 \u00E2\u0088\u0092 2\u00CE\u00B4 + 2 \u00CE\u00B2 \u00E2\u0088\u0091 (1+\u00CE\u00B4)\u00E2\u0088\u00921\u00CE\u00B2\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2 1 \u00CE\u00B1 + 1\u00E2\u0088\u0092 \u00CE\u00B4 \u00CE\u00B22 \u00E2\u0088\u0091 (1+\u00CE\u00B4)\u00E2\u0088\u00921\u00CE\u00B2\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2 1 = (1\u00E2\u0088\u0092 \u00CE\u00B4) ( 1 + \u00CE\u00B4 \u00CE\u00B2 \u00E2\u0088\u0092 1 \u00CE\u00B2 +O ( 1 \u00CE\u00B22 )) + \u00CE\u00B42 \u00E2\u0088\u0092 2\u00CE\u00B4 + 2 \u00CE\u00B2 ( log \u00CE\u00B2 (1 + \u00CE\u00B4)\u00E2\u0088\u00921\u00CE\u00B2 +O ( 1 \u00CE\u00B2 )) + 1\u00E2\u0088\u0092 \u00CE\u00B4 \u00CE\u00B22 ( \u00CE\u00B2 \u00E2\u0088\u0092 \u00CE\u00B2 1 + \u00CE\u00B4 +O(1) ) = 1 \u00CE\u00B2 ( 2\u00E2\u0088\u0092 \u00CE\u00B42 \u00E2\u0088\u0092 2 1 + \u00CE\u00B4 + (\u00CE\u00B42 \u00E2\u0088\u0092 2\u00CE\u00B4 + 2) log(1 + \u00CE\u00B4) ) +O ( 1 \u00CE\u00B22 ) . (3.16) 65 3.7. Proof of key proposition for enumeration This case of the proposition then follows from (3.15) upon noting that 4\u00E2\u0088\u00924\u00CE\u00B4+ 2 1 + \u00CE\u00B4 +2\u00CE\u00B4\u00E2\u0088\u00922\u00E2\u0088\u00922(1\u00E2\u0088\u0092\u00CE\u00B4) log(1\u00E2\u0088\u0092\u00CE\u00B42)+2\u00E2\u0088\u0092\u00CE\u00B42\u00E2\u0088\u0092 2 1 + \u00CE\u00B4 +(\u00CE\u00B42\u00E2\u0088\u00922\u00CE\u00B4+2) log(1+\u00CE\u00B4) = 4\u00E2\u0088\u0092 2\u00CE\u00B4 \u00E2\u0088\u0092 \u00CE\u00B42 + \u00CE\u00B42 log(1 + \u00CE\u00B4)\u00E2\u0088\u0092 2(1\u00E2\u0088\u0092 \u00CE\u00B4) log(1\u00E2\u0088\u0092 \u00CE\u00B4). \u00E2\u0080\u00A2 Case 2: \u00CE\u00B4 = 1. In this case we have C = 1 \u00CE\u00B2 and D = \u00EF\u00A3\u00B1\u00EF\u00A3\u00B2\u00EF\u00A3\u00B32/\u00CE\u00B2, if \u00CE\u00B1 \u00E2\u0089\u00A4 \u00CE\u00B2/2,1/\u00CE\u00B1, if \u00CE\u00B2/2 \u00E2\u0089\u00A4 \u00CE\u00B1. Therefore \u00E2\u0088\u0091 1\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2 CD = \u00E2\u0088\u0091 1\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2/2 1 \u00CE\u00B2 \u00C2\u00B7 2 \u00CE\u00B2 + \u00E2\u0088\u0091 \u00CE\u00B2/2\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2 1 \u00CE\u00B2 \u00C2\u00B7 1 \u00CE\u00B1 = 2 \u00CE\u00B22 ( \u00CE\u00B2 2 +O(1) ) + 1 \u00CE\u00B2 ( log \u00CE\u00B2 \u00CE\u00B2/2 +O ( 1 \u00CE\u00B2 )) = 1 + log 2 \u00CE\u00B2 +O ( 1 \u00CE\u00B22 ) , as desired. \u00E2\u0080\u00A2 Case 3: 1 < \u00CE\u00B4 \u00E2\u0089\u00A4 \u00E2\u0088\u009A2. In this case we have C = \u00EF\u00A3\u00B1\u00EF\u00A3\u00B2\u00EF\u00A3\u00B30, if \u00CE\u00B1 \u00E2\u0089\u00A4 (\u00CE\u00B4 \u00E2\u0088\u0092 1)\u00CE\u00B2,1\u00E2\u0088\u0092\u00CE\u00B4 \u00CE\u00B1 + 1 \u00CE\u00B2 , if (\u00CE\u00B4 \u00E2\u0088\u0092 1)\u00CE\u00B2 \u00E2\u0089\u00A4 \u00CE\u00B1 and D = \u00EF\u00A3\u00B1\u00EF\u00A3\u00B2\u00EF\u00A3\u00B3 2\u00CE\u00B2 , if \u00CE\u00B1 \u00E2\u0089\u00A4 (\u00CE\u00B4 + 1)\u00E2\u0088\u00921\u00CE\u00B2,1\u00E2\u0088\u0092\u00CE\u00B4 \u00CE\u00B2 + 1 \u00CE\u00B1 , if (\u00CE\u00B4 + 1) \u00E2\u0088\u00921\u00CE\u00B2 \u00E2\u0089\u00A4 \u00CE\u00B1. Therefore \u00E2\u0088\u0091 1\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2 CD = \u00E2\u0088\u0091 max{1,(\u00CE\u00B4\u00E2\u0088\u00921)\u00CE\u00B2}\u00E2\u0089\u00A4\u00CE\u00B1<(\u00CE\u00B4+1)\u00E2\u0088\u00921\u00CE\u00B2 ( 1\u00E2\u0088\u0092 \u00CE\u00B4 \u00CE\u00B1 + 1 \u00CE\u00B2 ) 2 \u00CE\u00B2 + \u00E2\u0088\u0091 (\u00CE\u00B4+1)\u00E2\u0088\u00921\u00CE\u00B2\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2 ( 1\u00E2\u0088\u0092 \u00CE\u00B4 \u00CE\u00B1 + 1 \u00CE\u00B2 )( 1\u00E2\u0088\u0092 \u00CE\u00B4 \u00CE\u00B2 + 1 \u00CE\u00B1 ) . (3.17) (We note that (\u00CE\u00B4\u00E2\u0088\u00921)\u00CE\u00B2 \u00E2\u0089\u00A4 (\u00CE\u00B4+1)\u00E2\u0088\u00921\u00CE\u00B2 for \u00CE\u00B4 between 1 and \u00E2\u0088\u009A2. For very small \u00CE\u00B2 we might have 1 > (\u00CE\u00B4 + 1)\u00E2\u0088\u00921\u00CE\u00B2, in which case the first sum is empty, but 66 3.7. Proof of key proposition for enumeration that does not invalidate the argument that follows.) By Lemma 3.18, the first sum is \u00E2\u0088\u0091 max{1,(\u00CE\u00B4\u00E2\u0088\u00921)\u00CE\u00B2}\u00E2\u0089\u00A4\u00CE\u00B1<(1+\u00CE\u00B4)\u00E2\u0088\u00921\u00CE\u00B2 ( 1\u00E2\u0088\u0092 \u00CE\u00B4 \u00CE\u00B1 + 1 \u00CE\u00B2 ) 2 \u00CE\u00B2 = 2 \u00CE\u00B2 ( 1 1 + \u00CE\u00B4 + 1\u00E2\u0088\u0092 \u00CE\u00B4 \u00E2\u0088\u0092 (1\u00E2\u0088\u0092 \u00CE\u00B4) log(\u00CE\u00B42 \u00E2\u0088\u0092 1) ) +O ( 1 + log \u00CE\u00B2 \u00CE\u00B22 ) , while the second sum has already been evaluated in equation (3.16) above. This case of the proposition then follows from (3.17) by noting that 2 1 + \u00CE\u00B4 +2\u00E2\u0088\u00922\u00CE\u00B4\u00E2\u0088\u00922(1\u00E2\u0088\u0092\u00CE\u00B4) log(\u00CE\u00B42\u00E2\u0088\u00921)+2\u00E2\u0088\u0092\u00CE\u00B42\u00E2\u0088\u0092 2 \u00CE\u00B4 + 1 +(\u00CE\u00B42\u00E2\u0088\u00922\u00CE\u00B4+2) log(\u00CE\u00B4+1) = 4\u00E2\u0088\u0092 2\u00CE\u00B4 \u00E2\u0088\u0092 \u00CE\u00B42 + \u00CE\u00B42 log(\u00CE\u00B4 + 1) + 2(\u00CE\u00B4 \u00E2\u0088\u0092 1) log(\u00CE\u00B4 \u00E2\u0088\u0092 1). \u00E2\u0080\u00A2 Case 4: \u00E2\u0088\u009A2 < \u00CE\u00B4 \u00E2\u0089\u00A4 2. Just as in Case 3, we have C = \u00EF\u00A3\u00B1\u00EF\u00A3\u00B2\u00EF\u00A3\u00B30, if \u00CE\u00B1 \u00E2\u0089\u00A4 (\u00CE\u00B4 \u00E2\u0088\u0092 1)\u00CE\u00B2,1\u00E2\u0088\u0092\u00CE\u00B4 \u00CE\u00B1 + 1 \u00CE\u00B2 , if (\u00CE\u00B4 \u00E2\u0088\u0092 1)\u00CE\u00B2 \u00E2\u0089\u00A4 \u00CE\u00B1 and D = \u00EF\u00A3\u00B1\u00EF\u00A3\u00B2\u00EF\u00A3\u00B3 2\u00CE\u00B2 , if \u00CE\u00B1 \u00E2\u0089\u00A4 (\u00CE\u00B4 + 1)\u00E2\u0088\u00921\u00CE\u00B2,1\u00E2\u0088\u0092\u00CE\u00B4 \u00CE\u00B2 + 1 \u00CE\u00B1 , if (\u00CE\u00B4 + 1) \u00E2\u0088\u00921\u00CE\u00B2 \u00E2\u0089\u00A4 \u00CE\u00B1. However, the inequality (\u00CE\u00B4\u00E2\u0088\u00921)\u00CE\u00B2 \u00E2\u0089\u00A4 \u00CE\u00B1 automatically implies that (\u00CE\u00B4+1)\u00E2\u0088\u00921\u00CE\u00B2 \u00E2\u0089\u00A4 \u00CE\u00B1 when \u00CE\u00B4 \u00E2\u0089\u00A5 \u00E2\u0088\u009A2. Therefore \u00E2\u0088\u0091 1\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2 CD = \u00E2\u0088\u0091 (\u00CE\u00B4\u00E2\u0088\u00921)\u00CE\u00B2\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2 ( 1\u00E2\u0088\u0092 \u00CE\u00B4 \u00CE\u00B1 + 1 \u00CE\u00B2 )( 1\u00E2\u0088\u0092 \u00CE\u00B4 \u00CE\u00B2 + 1 \u00CE\u00B1 ) . (In this case we will not need to use the precise lower bound max{1, (\u00CE\u00B4 \u00E2\u0088\u0092 67 3.8. Distribution of real eigenvalues 1)\u00CE\u00B2} \u00E2\u0089\u00A4 \u00CE\u00B1 for the summation over \u00CE\u00B1.) This yields \u00E2\u0088\u0091 1\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2 CD = (1\u00E2\u0088\u0092 \u00CE\u00B4) \u00E2\u0088\u0091 (\u00CE\u00B4\u00E2\u0088\u00921)\u00CE\u00B2\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2 1 \u00CE\u00B12 + \u00CE\u00B42 \u00E2\u0088\u0092 2\u00CE\u00B4 + 2 \u00CE\u00B2 \u00E2\u0088\u0091 (\u00CE\u00B4\u00E2\u0088\u00921)\u00CE\u00B2\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2 1 \u00CE\u00B1 + 1\u00E2\u0088\u0092 \u00CE\u00B4 \u00CE\u00B22 \u00E2\u0088\u0091 (\u00CE\u00B4\u00E2\u0088\u00921)\u00CE\u00B2\u00E2\u0089\u00A4\u00CE\u00B1<\u00CE\u00B2 1 = (1\u00E2\u0088\u0092 \u00CE\u00B4) ( 1 (\u00CE\u00B4 \u00E2\u0088\u0092 1)\u00CE\u00B2 \u00E2\u0088\u0092 1 \u00CE\u00B2 +O ( 1 (\u00CE\u00B4 \u00E2\u0088\u0092 1)2\u00CE\u00B22 )) + \u00CE\u00B42 \u00E2\u0088\u0092 2\u00CE\u00B4 + 2 \u00CE\u00B2 ( log \u00CE\u00B2 (\u00CE\u00B4 \u00E2\u0088\u0092 1)\u00CE\u00B2 +O ( 1 (\u00CE\u00B4 \u00E2\u0088\u0092 1)\u00CE\u00B2 )) + 1\u00E2\u0088\u0092 \u00CE\u00B4 \u00CE\u00B22 ( \u00CE\u00B2 \u00E2\u0088\u0092 (\u00CE\u00B4 \u00E2\u0088\u0092 1)\u00CE\u00B2 +O(1) ) = 1 \u00CE\u00B2 ( \u00CE\u00B42 \u00E2\u0088\u0092 2\u00CE\u00B4 \u00E2\u0088\u0092 (\u00CE\u00B42 \u00E2\u0088\u0092 2\u00CE\u00B4 + 2) log(\u00CE\u00B4 \u00E2\u0088\u0092 1) ) +O ( 1 \u00CE\u00B22 ) , where the error terms have been simplified since \u00CE\u00B4\u00E2\u0088\u00921 is bounded away from 0. This concludes the proof of the key proposition which yields Theorem 3.3. 3.8 Distribution of real eigenvalues In proving Theorem 3.5, it will be convenient to define the odd function G(z) = \u00E2\u0088\u00AB z 0 \u00E2\u0088\u0092 log |t| dt = z(1\u00E2\u0088\u0092 log |z|), (3.18) whose relevance is demonstrated by the following lemma. Lemma 3.19. If B and C are independent random variables uniformly distributed on [\u00E2\u0088\u00921, 1], then the product BC has the distribution function FBC(z) = Pr(BC < z) = 1 2(1 +G(z)) for z \u00E2\u0088\u0088 [\u00E2\u0088\u00921, 1]. (Of course FBC(z) = 0 for any z < \u00E2\u0088\u00921, and FBC(z) = 1 for any z > 1.) Proof. Note that |B| and |C| are uniformly distributed on [0, 1]. For 0 \u00E2\u0089\u00A4 z \u00E2\u0089\u00A4 1, we easily check that Pr(|BC| < z) = \u00E2\u0088\u00AB 10 min{1, z/s} ds = G(z). Thus |BC| is distributed on [0, 1] with density f|BC|(z) = \u00E2\u0088\u0092 log z, and by 68 3.8. Distribution of real eigenvalues symmetry BC has density fBC(z) = \u00E2\u0088\u009212 log |z| on [\u00E2\u0088\u00921, 1]. The lemma follows upon computing FBC(z) = \u00E2\u0088\u00AB z \u00E2\u0088\u00921 fBC(s) ds. It will also be helpful to define the following functions, which are sym- metric in x and y: \u00CE\u00BD1(x, y) = 1/2 +G(xy)/2 +G((x\u00E2\u0088\u0092 y)2/4), (3.19) \u00CE\u00BD2(x, y) = 1/2\u00E2\u0088\u0092G(xy)/2, (3.20) \u00CE\u00BD(x, y) = \u00EF\u00A3\u00B1\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B2\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B4\u00EF\u00A3\u00B3 \u00CE\u00BD1(x, y), if xy < 1 and x+ y < 0, \u00CE\u00BD2(x, y), if xy < 1 and x+ y > 0, 1 +G((x\u00E2\u0088\u0092 y)2/4), if xy > 1 and x+ y < 0, 0, otherwise. (3.21) To prove Theorem 3.5, we first consider the distribution function FW (\u00CE\u00B4) = \u00E2\u0088\u00AB t<\u00CE\u00B4 W (t) dt associated to the density W (\u00CE\u00B4). For a random matrix M inM2([\u00E2\u0088\u00921, 1]) and a real number \u00CE\u00B4, we will derive an expression for the expected number of real eigenvalues of M falling below \u00CE\u00B4, then differentiate it to obtain W (\u00CE\u00B4). It is clear that W (\u00E2\u0088\u0092\u00CE\u00B4) = W (\u00CE\u00B4) since the setM2([\u00E2\u0088\u00921, 1]) is closed under negation, so it suffices to compute W (\u00CE\u00B4) for \u00CE\u00B4 \u00E2\u0088\u0088 [0, 2]. It turns out that our calculations for FW will be somewhat simplified by considering FW (\u00E2\u0088\u0092\u00CE\u00B4) rather than FW (\u00CE\u00B4). Proposition 3.20. We have FW (\u00E2\u0088\u0092\u00CE\u00B4) = 1 4 \u00E2\u0088\u00AB 1+\u00CE\u00B4 \u00E2\u0088\u00921+\u00CE\u00B4 \u00E2\u0088\u00AB 1+\u00CE\u00B4 \u00E2\u0088\u00921+\u00CE\u00B4 \u00CE\u00BD(x, y) dx dy for all 0 \u00E2\u0089\u00A4 \u00CE\u00B4 \u00E2\u0089\u00A4 2, where \u00CE\u00BD is defined in equation (3.21). Proof. We denote the entries of M by A, B, C, D; by assumption these are independent random variables uniformly distributed in [\u00E2\u0088\u00921, 1]. Let \u00CE\u00B4 be 69 3.8. Distribution of real eigenvalues fixed in the range [0, 2], and consider the shifted matrix M \u00E2\u0080\u00B2 = M+\u00CE\u00B4I, which we write as M \u00E2\u0080\u00B2 = ( X B C Y ) , where X, Y range independently and uniformly in [\u00E2\u0088\u00921 + \u00CE\u00B4, 1 + \u00CE\u00B4] and B, C are as before. Clearly the eigenvalues of M lying below \u00E2\u0088\u0092\u00CE\u00B4 correspond to the negative (real) eigenvalues of M \u00E2\u0080\u00B2. By Lemma 3.14, we are free to exclude the null set where M \u00E2\u0080\u00B2 is singular or has repeated eigenvalues. Outside of this null set, M \u00E2\u0080\u00B2 has exactly one negative eigenvalue if and only if detM \u00E2\u0080\u00B2 = XY \u00E2\u0088\u0092BC < 0, by Lemma 3.13(c). Likewise by Lemma 3.13(d), M \u00E2\u0080\u00B2 has exactly two negative eigenvalues if and only if XY \u00E2\u0088\u0092BC > 0, X + Y < 0 and discM \u00E2\u0080\u00B2 = (X \u00E2\u0088\u0092 Y )2 + 4BC > 0. We thus have: FW (\u00E2\u0088\u0092\u00CE\u00B4) = Pr(BC > XY )+2 Pr ( X+Y < 0 and \u00E2\u0088\u0092(X \u00E2\u0088\u0092 Y ) 2 4 < BC < XY ) . By conditioning on the values of X and Y , we may express this probability as the average value FW (\u00E2\u0088\u0092\u00CE\u00B4) = 1 4 \u00E2\u0088\u00AB 1+\u00CE\u00B4 \u00E2\u0088\u00921+\u00CE\u00B4 \u00E2\u0088\u00AB 1+\u00CE\u00B4 \u00E2\u0088\u00921+\u00CE\u00B4 \u00CF\u0081(x, y) dx dy, where for fixed x and y, \u00CF\u0081(x, y) = Pr(BC > xy) + 2 Pr(x+ y < 0 and \u00E2\u0088\u0092(x\u00E2\u0088\u0092 y)2/4 < BC < xy) = Pr(BC > xy) + 2 Pr(\u00E2\u0088\u0092(x\u00E2\u0088\u0092 y)2/4 < BC < xy)1{x+ y < 0} (3.22) (here 1{\u00C2\u00B7} denotes the indicator function of the indicated relation). To complete the proof it suffices to show that \u00CF\u0081 equals the function \u00CE\u00BD defined 70 3.8. Distribution of real eigenvalues in equation (3.21). The probabilities appearing in equation (3.22) are effectively given by Lemma 3.19. However, there is some case-checking involved in applying this lemma, since the value of, say, Pr(BC > xy) = 1 \u00E2\u0088\u0092 FBC(xy) depends on whether xy < \u00E2\u0088\u00921, \u00E2\u0088\u00921 \u00E2\u0089\u00A4 xy \u00E2\u0089\u00A4 1, or xy < \u00E2\u0088\u00921. We make some observations to reduce the number of cases we need to examine. Note that (x \u00E2\u0088\u0092 y)2/4 is bounded between 0 and 1 for any x, y \u00E2\u0088\u0088 [\u00E2\u0088\u00921 + \u00CE\u00B4, 1 + \u00CE\u00B4], so that \u00E2\u0088\u0092(x \u00E2\u0088\u0092 y)2/4 always lies in the interval [\u00E2\u0088\u00921, 1] prescribed by Lemma 3.19. From the identity (x + y)2 \u00E2\u0088\u0092 (x \u00E2\u0088\u0092 y)2 = 4xy we see also that xy \u00E2\u0089\u00A5 \u00E2\u0088\u0092(x\u00E2\u0088\u0092 y)2/4. Thus xy is never lower than \u00E2\u0088\u00921, and we need only consider whether xy > 1 (in which case FBC(xy) = 1). We therefore have Pr(BC > xy) = 1\u00E2\u0088\u0092 FBC(xy) = 12(1\u00E2\u0088\u0092G(xy))1{xy < 1} and 2 Pr(\u00E2\u0088\u0092(x\u00E2\u0088\u0092 y)2/4 < BC < xy) = 2FBC(xy)\u00E2\u0088\u0092 2FBC(\u00E2\u0088\u0092(x\u00E2\u0088\u0092 y)2/4) = 1{xy > 1}+G(xy)1{xy < 1}+G((x\u00E2\u0088\u0092 y)2/4). Inserting these two evaluations into the formula (3.22), we obtain \u00CF\u0081(x, y) = 12(1\u00E2\u0088\u0092G(xy))1{xy < 1} + ( 1{xy > 1}+G(xy)1{xy < 1}+G ((x\u00E2\u0088\u0092 y)2 4 )) 1{x+ y < 0}. It can be verified that this last expression is indeed equal to the right-hand side of the definition (3.21) of \u00CE\u00BD, which establishes the proposition. Since W (\u00CE\u00B4) = W (\u00E2\u0088\u0092\u00CE\u00B4) = \u00E2\u0088\u0092 dd\u00CE\u00B4FW (\u00E2\u0088\u0092\u00CE\u00B4), to finish the proof of Theorem 3.5 it suffices to show that \u00E2\u0088\u0092 dd\u00CE\u00B4FW (\u00E2\u0088\u0092\u00CE\u00B4) equals the expression in formula (3.4). 71 3.9. The derivative of the distribution \u00CE\u009D1 \u00CE\u009D2 0 \u00CE\u009D2 0 \u00CE\u009D2 0 Figure 3.2: The three cases 0 \u00E2\u0089\u00A4 \u00CE\u00B4 \u00E2\u0089\u00A4 1, and 1 \u00E2\u0089\u00A4 \u00CE\u00B4 \u00E2\u0089\u00A4 \u00E2\u0088\u009A2, and \u00E2\u0088\u009A2 \u00E2\u0089\u00A4 \u00CE\u00B4 \u00E2\u0089\u00A4 2 3.9 The derivative of the distribution Proposition 3.20 expresses FW (\u00CE\u00B4) as an integral of a function \u00CE\u00BD (which is independent of \u00CE\u00B4) over the square S\u00CE\u00B4 := [\u00E2\u0088\u00921+\u00CE\u00B4, 1+\u00CE\u00B4]2 \u00E2\u008A\u0082 R2. Since the region S\u00CE\u00B4 varies continuously with \u00CE\u00B4, the derivative \u00E2\u0088\u0092 dd\u00CE\u00B4FW (\u00E2\u0088\u0092\u00CE\u00B4) can be computed by an appropriate line integral around the boundary of S\u00CE\u00B4. Indeed, by the fundamental theorem of calculus, we have \u00E2\u0088\u0092 d d\u00CE\u00B4 FW (\u00E2\u0088\u0092\u00CE\u00B4) = \u00E2\u0088\u00921 4 d d\u00CE\u00B4 (\u00E2\u0088\u00AB 1+\u00CE\u00B4 \u00E2\u0088\u00921+\u00CE\u00B4 \u00E2\u0088\u00AB 1+\u00CE\u00B4 \u00E2\u0088\u00921+\u00CE\u00B4 \u00CE\u00BD(x, y) dx dy ) = \u00E2\u0088\u00921 4 (\u00E2\u0088\u00AB 1+\u00CE\u00B4 \u00E2\u0088\u00921+\u00CE\u00B4 \u00CE\u00BD(1 + \u00CE\u00B4, y) dy \u00E2\u0088\u0092 \u00E2\u0088\u00AB 1+\u00CE\u00B4 \u00E2\u0088\u00921+\u00CE\u00B4 \u00CE\u00BD(\u00E2\u0088\u00921 + \u00CE\u00B4, y) dy + \u00E2\u0088\u00AB 1+\u00CE\u00B4 \u00E2\u0088\u00921+\u00CE\u00B4 \u00CE\u00BD(x, 1 + \u00CE\u00B4) dx\u00E2\u0088\u0092 \u00E2\u0088\u00AB 1+\u00CE\u00B4 \u00E2\u0088\u00921+\u00CE\u00B4 \u00CE\u00BD(x,\u00E2\u0088\u00921 + \u00CE\u00B4) dx ) = 1 2 \u00E2\u0088\u00AB 1+\u00CE\u00B4 \u00E2\u0088\u00921+\u00CE\u00B4 \u00CE\u00BD(x,\u00E2\u0088\u00921 + \u00CE\u00B4) dx\u00E2\u0088\u0092 1 2 \u00E2\u0088\u00AB 1+\u00CE\u00B4 \u00E2\u0088\u00921+\u00CE\u00B4 \u00CE\u00BD(x, 1 + \u00CE\u00B4) dx, (3.23) where we have used the symmetry \u00CE\u00BD(x, y) = \u00CE\u00BD(y, x) to reduce the integral to just the top and bottom edges of S\u00CE\u00B4 (where y = 1 + \u00CE\u00B4 and y = \u00E2\u0088\u00921 + \u00CE\u00B4, respectively). The evaluation of (3.23) divides into three cases depending on the be- haviour of the indicator functions 1{x + y < 0} and 1{xy < 1} on the boundary of S\u00CE\u00B4. Figure 3.2 depicts the possible intersections of S\u00CE\u00B4 with the line x+ y = 0 and the hyperbola xy = 1. \u00E2\u0080\u00A2 Case 1: 0 \u00E2\u0089\u00A4 \u00CE\u00B4 \u00E2\u0089\u00A4 1. For this range of \u00CE\u00B4, the line x + y = 0 intersects 72 3.9. The derivative of the distribution the bottom edge of S\u00CE\u00B4 at x = 1 \u00E2\u0088\u0092 \u00CE\u00B4, while the hyperbola xy = 1 intersects the top edge at x = (1 + \u00CE\u00B4)\u00E2\u0088\u00921. Thus by the definition of \u00CE\u00BD, equation (3.23) becomes \u00E2\u0088\u0092 d d\u00CE\u00B4 FW (\u00E2\u0088\u0092\u00CE\u00B4) = 1 2 (\u00E2\u0088\u00AB 1\u00E2\u0088\u0092\u00CE\u00B4 \u00E2\u0088\u00921+\u00CE\u00B4 \u00CE\u00BD1(x,\u00E2\u0088\u00921 + \u00CE\u00B4) dx + \u00E2\u0088\u00AB 1+\u00CE\u00B4 1\u00E2\u0088\u0092\u00CE\u00B4 \u00CE\u00BD2(x,\u00E2\u0088\u00921 + \u00CE\u00B4) dx\u00E2\u0088\u0092 \u00E2\u0088\u00AB (1+\u00CE\u00B4)\u00E2\u0088\u00921 \u00E2\u0088\u00921+\u00CE\u00B4 \u00CE\u00BD2(x, 1 + \u00CE\u00B4) dx ) . The following elementary antiderivative (easily obtained by substitution and integration by parts) hold for any fixed nonzero real number y using the definitions (3.18), (3.19), and (3.20) of G, \u00CE\u00BD1, and \u00CE\u00BD2:\u00E2\u0088\u00AB \u00CE\u00BD1(x, y) dx = 1 2x+ 1 8x 2y(3\u00E2\u0088\u0092 2 log |xy|) + 136(x\u00E2\u0088\u0092 y)3(5\u00E2\u0088\u0092 6 log |(x\u00E2\u0088\u0092 y)/2|),\u00E2\u0088\u00AB \u00CE\u00BD2(x, y) dx = 1 2x\u00E2\u0088\u0092 18x2y(3\u00E2\u0088\u0092 2 log |xy|). (3.24) Therefore in this case \u00E2\u0088\u0092 d d\u00CE\u00B4 FW (\u00E2\u0088\u0092\u00CE\u00B4) = 1 2 (( 1 2x+ 1 8x 2(\u00E2\u0088\u00921 + \u00CE\u00B4)(3\u00E2\u0088\u0092 2 log |x(\u00E2\u0088\u00921 + \u00CE\u00B4)|) + 136(x+ 1\u00E2\u0088\u0092 \u00CE\u00B4)3(5\u00E2\u0088\u0092 6 log |(x+ 1\u00E2\u0088\u0092 \u00CE\u00B4)/2|) )\u00E2\u0088\u00A3\u00E2\u0088\u00A3\u00E2\u0088\u00A31\u00E2\u0088\u0092\u00CE\u00B4 x=\u00E2\u0088\u00921+\u00CE\u00B4 + ( 1 2x\u00E2\u0088\u0092 18x2(\u00E2\u0088\u00921 + \u00CE\u00B4)(3\u00E2\u0088\u0092 2 log |x(\u00E2\u0088\u00921 + \u00CE\u00B4)|) )\u00E2\u0088\u00A3\u00E2\u0088\u00A3\u00E2\u0088\u00A31+\u00CE\u00B4 x=1\u00E2\u0088\u0092\u00CE\u00B4 \u00E2\u0088\u0092 (12x\u00E2\u0088\u0092 18x2(1 + \u00CE\u00B4)(3\u00E2\u0088\u0092 2 log |x(1 + \u00CE\u00B4)|))\u00E2\u0088\u00A3\u00E2\u0088\u00A3\u00E2\u0088\u00A3(1+\u00CE\u00B4)\u00E2\u0088\u00921x=\u00E2\u0088\u00921+\u00CE\u00B4 ) = 80 + 20\u00CE\u00B4 + 90\u00CE\u00B42 + 52\u00CE\u00B43 \u00E2\u0088\u0092 107\u00CE\u00B44 144(1 + \u00CE\u00B4) \u00E2\u0088\u0092 (5\u00E2\u0088\u0092 7\u00CE\u00B4 + 8\u00CE\u00B4 2)(1\u00E2\u0088\u0092 \u00CE\u00B4) 12 log(1\u00E2\u0088\u0092 \u00CE\u00B4) \u00E2\u0088\u0092 \u00CE\u00B4(1\u00E2\u0088\u0092 \u00CE\u00B4 2) 4 log(1 + \u00CE\u00B4) (after some algebraic simplification), which verifies the first case of The- orem 3.5. (The integrands really are continuous, despite the presence of terms that could be log 0, because the function G is continuous at 0; hence evaluating the integrals by antiderivatives is valid.) 73 3.9. The derivative of the distribution \u00E2\u0080\u00A2 Case 2: 1 \u00E2\u0089\u00A4 \u00CE\u00B4 \u00E2\u0089\u00A4 \u00E2\u0088\u009A2. Now, the line x + y = 0 does not intersect S\u00CE\u00B4, while the hyperbola xy = 1 intersects the top edge at x = (1 + \u00CE\u00B4)\u00E2\u0088\u00921. Thus by the definition of \u00CE\u00BD and the antiderivative (3.24) of \u00CE\u00BD2, equation (3.23) becomes \u00E2\u0088\u0092 d d\u00CE\u00B4 FW (\u00E2\u0088\u0092\u00CE\u00B4) = 1 2 (\u00E2\u0088\u00AB 1+\u00CE\u00B4 \u00E2\u0088\u00921+\u00CE\u00B4 \u00CE\u00BD2(x,\u00E2\u0088\u00921 + \u00CE\u00B4) dx\u00E2\u0088\u0092 \u00E2\u0088\u00AB (1+\u00CE\u00B4)\u00E2\u0088\u00921 \u00E2\u0088\u00921+\u00CE\u00B4 \u00CE\u00BD2(x, 1 + \u00CE\u00B4) dx ) = 1 2 (( 1 2x\u00E2\u0088\u0092 18x2(\u00E2\u0088\u00921 + \u00CE\u00B4)(3\u00E2\u0088\u0092 2 log |x(\u00E2\u0088\u00921 + \u00CE\u00B4)|) )\u00E2\u0088\u00A3\u00E2\u0088\u00A3\u00E2\u0088\u00A3\u00E2\u0088\u00A31+\u00CE\u00B4 x=\u00E2\u0088\u00921+\u00CE\u00B4 \u00E2\u0088\u0092 (12x\u00E2\u0088\u0092 18x2(1 + \u00CE\u00B4)(3\u00E2\u0088\u0092 2 log |x(1 + \u00CE\u00B4)|))\u00E2\u0088\u00A3\u00E2\u0088\u00A3\u00E2\u0088\u00A3\u00E2\u0088\u00A3(1+\u00CE\u00B4)\u00E2\u0088\u00921 x=\u00E2\u0088\u00921+\u00CE\u00B4 ) = \u00CE\u00B4(20 + 10\u00CE\u00B4 \u00E2\u0088\u0092 12\u00CE\u00B42 \u00E2\u0088\u0092 3\u00CE\u00B43) 16(1 + \u00CE\u00B4) + (3\u00CE\u00B4 \u00E2\u0088\u0092 1)(\u00CE\u00B4 \u00E2\u0088\u0092 1) 4 log(\u00CE\u00B4 \u00E2\u0088\u0092 1) + \u00CE\u00B4(\u00CE\u00B42 \u00E2\u0088\u0092 1) 4 log(\u00CE\u00B4 + 1), which verifies the second case of Theorem 3.5. \u00E2\u0080\u00A2 Case 3: \u00E2\u0088\u009A2 < \u00CE\u00B4 \u00E2\u0089\u00A4 2. As before, the line x + y = 0 does not intersect S\u00CE\u00B4, while the hyperbola xy = 1 intersects the bottom edge at x = (\u00CE\u00B4\u00E2\u0088\u0092 1)\u00E2\u0088\u00921. Thus by the definition of \u00CE\u00BD and the antiderivative (3.24) of \u00CE\u00BD2, equation (3.23) becomes \u00E2\u0088\u0092 d d\u00CE\u00B4 FW (\u00E2\u0088\u0092\u00CE\u00B4) = 1 2 \u00E2\u0088\u00AB (\u00E2\u0088\u00921+\u00CE\u00B4)\u00E2\u0088\u00921 \u00E2\u0088\u00921+\u00CE\u00B4 \u00CE\u00BD2(x,\u00E2\u0088\u00921 + \u00CE\u00B4) dx = 1 2 ( 1 2x\u00E2\u0088\u0092 18x2(1 + \u00CE\u00B4)(3\u00E2\u0088\u0092 2 log |x(1 + \u00CE\u00B4)|) )\u00E2\u0088\u00A3\u00E2\u0088\u00A3\u00E2\u0088\u00A3\u00E2\u0088\u00A3(\u00E2\u0088\u00921+\u00CE\u00B4)\u00E2\u0088\u00921 x=\u00E2\u0088\u00921+\u00CE\u00B4 = \u00CE\u00B4(\u00CE\u00B4 \u00E2\u0088\u0092 2)(2\u00E2\u0088\u0092 6\u00CE\u00B4 + 3\u00CE\u00B42) 16(\u00CE\u00B4 \u00E2\u0088\u0092 1) \u00E2\u0088\u0092 (\u00CE\u00B4 \u00E2\u0088\u0092 1)3 4 log(\u00CE\u00B4 \u00E2\u0088\u0092 1), which verifies the third case of Theorem 3.5. Since the last case of Theorem 3.5 is a consequence of Lemma 3.10, the proof of the theorem is complete. Remark. One could also use the same method to extract the individual dis- tributions of the upper and lower eigenvalues of M : for instance, eliminating 74 3.10. Further remarks the factor of 2 from equation (3.22) would yield an expression for the dis- tribution of the lower eigenvalue of M . 3.10 Further remarks Theorem 3.3 can also be obtained from a powerful result of Y. R. Katznelson [43], which gives an asymptotic formula for the number of singular integer matrices contained in the dilate k \u00C2\u00B7 B of any convex body B \u00E2\u008A\u0086 R4 (here we are embedding the 2\u00C3\u00972 matrices into R4 in the obvious way).21 In this more general setting, the count remains proportional to k2 log k, with a constant determined by integrating B with respect to a certain measure. This measure is fairly unusual: it is singular, being supported only on the hypersurface of R4 consisting of singular matrices. The explicit evalua- tion of this measure is roughly analogous to our case-by-case considerations in Section 3.7, modulo the significant complications of carrying error terms in our case. In some sense, this measure must encode the structural differ- ence between rational eigenvalues and real eigenvalues. We expect that real eigenvalues are likewise governed by a measure that is absolutely continuous. Katznelson\u00E2\u0080\u0099s result for larger n shows that the true proportion of singular matrices in Mn(k) decays as (log k)/kn, with a constant depending on n that is precisely computable (at least in principle). This is a considerable sharpening of our Lemma 3.2 for n \u00E2\u0089\u00A5 3, and it is very close to the obvious lower bound of order 1/kn obtained by counting matrices that have a row of all zeros. From this one can deduce a sharper form of Theorem 3.1, that (log k)/kn\u00E2\u0088\u00921 is the correct order of magnitude for the probability of having at least one integer eigenvalue (see [78] for further results along this line). This probability is also an upper bound for the probability of diagonaliz- ability over Q, but when n > 2 we do not know the precise decay rate of the latter. We note that an obvious lower bound of order 1/k(n 2\u00E2\u0088\u0092n)/2 is given by counting upper-triangular matrices having distinct diagonal entries, and it seems reasonable to conjecture that the true order of magnitude only differs from this bound by some power (possibly dependent on n) of log k. 21We thank Martin Widmer for directing our attention to this paper. 75 Chapter 4 Patterns in x2 + ky2 4.1 Introduction In this chapter we investigate the existence of certain combinatorial patterns in the set S(k) := {x2 + ky2 : x, y \u00E2\u0088\u0088 Z} (we will confine our attention to the positive definite case k \u00E2\u0089\u00A5 1). Let us define the two primary patterns of interest, each of which addresses a basic natural question that one might ask of any integer sequence: \u00E2\u0080\u009Chow many consecutive integers can it have?\u00E2\u0080\u009D and \u00E2\u0080\u009Chow far apart can two successive elements be?\u00E2\u0080\u009D For d \u00E2\u0088\u0088 N, we define a run of length d (or d-run) to be any string of d consecutive numbers {m,m+ 1, . . . ,m+ d\u00E2\u0088\u0092 1}. When further specificity is required, we call this the d-run starting at m. Should a given run be entirely contained in another set S \u00E2\u008A\u0086 N, we say that S contains this run (or some intuitively similar turn of phrase). By a gap of size d (or d-gap) we mean an ordered pair (m,m+d), which we also think of as starting at m. We will say that S contains the gap (m,m+ d) when S contains m and m + d but no other integer between m and m + d. (When d > 1, this is equivalent to the complement N \ S having the (d\u00E2\u0088\u0092 1)-run starting at m+ 1.) Our primary concern in this chapter shall be the qualitative question of whether S(k) contains infinitely many d-runs or d-gaps for a given d. We simply wish to know whether a locally feasible pattern must occur in- finitely often in S(k), with comparatively less regard for the much harder problem of quantifying the precise frequency of occurrence (we discuss some quantitative aspects in the conclusion). In the case of S(1), which is the classical sums of two squares, both runs and gaps are fairly well understood. With regard to runs, one sees easily (by local considerations modulo 4) that S(1) contains no run of length 4. 76 4.1. Introduction Conversely, it is well-known (see, for example, [14]) that S(1) has infinitely many runs of length 3, which is made apparent by the relations: 4n4 + 4n2 = (2n2)2 + (2n)2, 4n4 + 4n2 + 1 = (2n2 + 1)2 + 02, 4n4 + 4n2 + 2 = (2n2 + 1)2 + 12. Sharper quantitative results are known, but only for short runs: Hooley [39] proved that the number of 2-runs in S(1) up to x satisfies the lower bound #{n \u00E2\u0089\u00A4 x : {n, n+ 1} \u00E2\u008A\u0082 S} \u001D x log x . (4.1) By standard sieve methods (for instance, Selberg\u00E2\u0080\u0099s sieve as used in [14]), there is an upper bound of exactly the same order of magnitude, meaning that Hooley\u00E2\u0080\u0099s bound is best possible, aside from a constant factor. In general, the length of runs appearing in any S(k) is uniformly bounded: we will see in the next section that S(k) never contains any run of length 6 or higher, which follows from a simple local argument. However, it is possible to find runs of length 5 in certain sets S(k), and our main interest here lies in this extremal case: which sets S(k) admit infinitely many 5-runs? M. Rosenfeld (personal communication) previously observed that S(2) has many such runs, the first of which begin at 0, 96, 800, 2400, 3200 and 3648. This led him to conjecture that there are infinitely many 5-runs in S(2), and we affirm this with the first main result of this chapter: Theorem 4.1. There are infinitely many numbers m for which {18m2, 18m2 + 1, 18m2 + 2, 18m2 + 3, 18m2 + 4} \u00E2\u008A\u0082 S(2). The proof of Theorem 4.1 uses a direct construction, and the same tech- nique can be used (with slight modifications) with certain other values of k, such as k = 46. However, we will see in Section 4.2 that there is a positive density of k for which S(k) \u00E2\u0080\u009Cshould\u00E2\u0080\u009D contain 5-runs (in that there are no local obstructions prohibiting so); our method, unfortunately, only works 77 4.2. Local feasibility of runs for k in a set of asymptotic density 0. (The precise conditions are somewhat tedious, and we defer their discussion to Section 4.3.) However, it does allow us to establish the following general theorem about larger values of k: Theorem 4.2. There are infinitely many k > 2 for which S(k) contains infinitely many 5-runs. The question of gaps in S(1) was answered by Spiro and Hensley [79], who proved the universal existence of gaps: for any d \u00E2\u0088\u0088 N, S(1) contains infinitely many gaps of size d. Just as in our theorems for runs, their proof is explicitly constructive; we greatly generalize this construction to all sets S(k), proving universal existence with a single type of exception: Theorem 4.3. Fix any k, d \u00E2\u0088\u0088 N subject to the restriction that either 4 - k or d 6\u00E2\u0089\u00A1 2 (mod 4). Then there exists an explicitly computable quadratic polynomial P (n) such that S(k) has the gap (P (n), P (n) + d) for all n. Note that the restriction in the hypotheses is entirely necessary: if 4 | k, then S(k) consists solely of integers congruent to 0 or 1 modulo 4, so the difference between two elements of S(k) is never of the form 4m+ 2. 4.2 Local feasibility of runs We first examine all local obstructions to the existence of d-runs in S(k), in particular obtaining necessary conditions for S(k) to admit 5-runs. The underlying principle here is to consider the set of values attained by the quadratic form x2 + ky2 for x, y in the p-adic integers Zp for some p; more directly, we are considering the range of x2 + ky2 modulo pr for every prime p and each r \u00E2\u0089\u00A5 1. If there is any prime p for which this set does not contain d consecutive values, then clearly S(k) does not have any d-runs. Let us first dispense with the 2-adic case, which already reveals a uniform upper bound on d, as shown by the proposition below. Proposition 4.4. The length of any run in S(k) is at most 5. Moreover, if S(k) achieves this length, then k \u00E2\u0089\u00A1 2 (mod 4). 78 4.2. Local feasibility of runs k (mod 8) x2 + ky2 (mod 8) Run length 0, 4 0, 1, 4 2 1 0, 1, 2, 4, 5 3 2 0, 1, 2, 3, 4, 6 5 3, 7 7, 0, 1, 3, 4 3 5 0, 1, 4, 5, 6 3 6 6, 7, 0, 1, 2, 4 5 Table 4.1: Distribution of x2 + ky2 modulo 8 Proof. This follows by straightforward (but tedious) calculation of all possi- ble residues of x2 + ky2 modulo 8. Table 4.1 summarizes the possible values of x2 +ky2 depending on the congruence class of k modulo 8, and the longest resulting string of residues. We have ordered the residues so as to make the maximum number of consecutive values more apparent. We know from the introduction that the upper bound of the proposition is tight; for instance S(2), does contain 5-runs. This motivates us to try to characterize all k for which S(k) has any 5-runs. Let us complete our analysis of local obstructions with an eye toward this central question. We first recall the well-known structure of p-adic squares [76, Ch. II, \u00C2\u00A73.3]: Proposition 4.5. If p is odd and p - x, then x is a square modulo pr if and only if x is a square mod p. If p = 2, r \u00E2\u0089\u00A5 3 and x is odd, then x is a square modulo 2r if and only if x \u00E2\u0089\u00A1 1 (mod 8). In other words (after factoring out powers of p from x), the \u00E2\u0080\u009Csquareness\u00E2\u0080\u009D of x modulo some high prime power pr depends only on its quadratic char- acter modulo a fixed prime power (either p, or 23). The next fact shows that for most p, S(k) covers all congruence classes mod p. It follows from a well-known counting argument [65, Theorem 5.14]: the sets {x2 : x \u00E2\u0088\u0088 Z/pZ} and {b\u00E2\u0088\u0092 ky2 : y \u00E2\u0088\u0088 Z/pZ} are simply too large not to intersect. Proposition 4.6. If p - k, then the equation x2 + ky2 \u00E2\u0089\u00A1 b (mod p) has a solution for any fixed b \u00E2\u0088\u0088 Z/pZ. 79 4.2. Local feasibility of runs We will also need a result of Buell and Hudson [13] on runs of consecutive quadratic residues. Proposition 4.7. For any d, every sufficiently large prime p has d consec- utive quadratic residues (also d consecutive nonresidues). In particular, for d = 5 this holds for all p \u00E2\u0089\u00A5 197. Theorem 4.8. The existence of 5-runs in S(k) is locally feasible if and only if k \u00E2\u0089\u00A1 2 (mod 4) and none of the following primes divide k: 3, 5, 7, 11, 13, 19, 29, 31, 37, 53, 149. Proof. For 2-adic feasibility, we look to Proposition 4.4, which clearly shows k \u00E2\u0089\u00A1 2 (mod 4) is necessary; since this bounds the power of 2 dividing k, Table 4.1 essentially also shows 2-adic sufficiency.22 Let us now consider p-adic feasibility for some odd p - k. By Proposi- tion 4.6, x2 +ky2 attains all possible residues modulo p, giving an \u00E2\u0080\u009Cinfinite\u00E2\u0080\u009D run in Z/pZ. Not all of these necessarily lift to arbitrary p-adic values, but certainly any nonzero residue will (Proposition 4.5), and 0 itself is certainly attained. This yields the run {\u00E2\u0088\u0092p+ 1, . . . , p\u00E2\u0088\u0092 1} of length 2p\u00E2\u0088\u0092 1 \u00E2\u0089\u00A5 5; such primes pose no obstruction to local feasibility. It remains to consider odd primes p which do divide k. Since x2 + ky2 reduces to x2 modulo p, a necessary condition for p-adic feasibility is that there are 5 consecutive squares in Z/pZ. We claim it is also sufficient (note that p > 5 in such case): this is obvious by Proposition 4.5 if the 5 consecutive squares are nonzero, but even when one of the squares is 0, the same proposition allows us to lift the other 4 squares so that they remain consecutive modulo pr for all r. By Proposition 4.7, certainly all p \u00E2\u0089\u00A5 197 have 5 consecutive squares (since we count 0 as a square but it is not a quadratic residue, our require- ment is weaker). The theorem follows by explicitly computing which primes p < 197 fail to have 5 consecutive squares mod p. These are precisely the primes listed in the statement of the theorem. 22To be sure, one could extend the table to modulo 16, but nothing new is revealed. 80 4.3. Construction for k = 2 As an immediate consequence, the set of k for which S(k) locally admits 5-runs has a positive density of about 8%. Corollary 4.9. The set of k \u00E2\u0088\u0088 N satisfying the conditions of Theorem 4.8 has density equal to 215 \u00C2\u00B7 36 \u00C2\u00B7 5 11 \u00C2\u00B7 19 \u00C2\u00B7 29 \u00C2\u00B7 31 \u00C2\u00B7 53 \u00C2\u00B7 149 \u00E2\u0089\u0088 0.0804969. Proof. This is just 14\u00CF\u0086(n)/n, where n is the product of the listed primes. 4.3 Construction for k = 2 Clearly k = 2 is the first integer to satisfy the conditions of Theorem 4.8, and we observed in the first section that S(2) contains several 5-runs. Let us now give the simple construction of infinitely many quintuples in S(2). Proof of Theorem 4.1. Three of the stated inclusions are immediate for any m, from the easy identities 18m2 = (4m)2 + 2(m)2 18m2 + 1 = (1)2 + 2(3m)2 18m2 + 4 = (2)2 + 2(3m)2. For the other two terms, we will show that m may be chosen simultaneously equal to 9a2 + 4a and 3b2 + b, whence m satisfies the additional identities 18m2 + 2 = (4m+ 2a)2 + 2(m\u00E2\u0088\u0092 4a\u00E2\u0088\u0092 1)2, 18m2 + 3 = (4m+ 2b\u00E2\u0088\u0092 1)2 + 2(m\u00E2\u0088\u0092 4b\u00E2\u0088\u0092 1)2. Let us verify that the Diophantine equation 9a2 + 4a = 3b2 + b has infinitely many solutions (it has at least the basic solution m = a = b = 0). Making the substitution c = 18a + 4, d = 6b + 1 yields a Pell-type equation with modular restrictions: c2 \u00E2\u0088\u0092 3d2 = 13, c \u00E2\u0089\u00A1 4 (mod 18), d \u00E2\u0089\u00A1 1 (mod 6). (4.2) 81 4.4. Construction for k > 2 This has an infinite family of solutions with the initial solution (c0, d0) = (4, 1) and recurrence (cn+1, dn+1) = (2cn+3dn, cn+2dn). One easily verifies that the subsequence (c6n, d6n) satisfies the necessary congruences (4.2). Remark. Since the ring Z[ \u00E2\u0088\u009A\u00E2\u0088\u00922] is a unique factorization domain, the exis- tence of rational values x, y with n = x2 + 2y2 is enough to conclude that n \u00E2\u0088\u0088 S(2). Thus, the above construction actually produces runs in S(2) when- ever 18m2 is an integer, regardless of the integrality of c and d. It turns out that this relaxation is satisified for the denser subsequence (c2n, d2n), which yields the run at 800; in contrast, the first run given by (c6n, d6n) occurs at the much larger 1224111751200. 4.4 Construction for k > 2 We now generalize the construction of Theorem 4.1 to certain other values of k. We will see that under certain conditions it is possible to take a single 5-run in S(k), and generate from it an infinite sequence of 5-runs. Note that S(2) is unique in that it is the only S(k) containing the run {0, 1, 2, 3, 4}, which one might call the \u00E2\u0080\u009Cgenerator\u00E2\u0080\u009D of Theorem 4.1 since it corresponds to (c0, d0). While the strategy in this section also works for S(2), it cannot be used with the run {0, 1, 2, 3, 4} and thereby yields a less efficient construction. We first require an elementary proposition regarding generalized Pell equations. It is certainly well-known in various equivalent forms; the version below will be convenient to us. Proposition 4.10. Let A,B,C,D be fixed integers such that BD > 0 is non-square. If the system m = x2 +A B = y2 + C D (4.3) has at least one integer solution (m,x, y) with (x, y) 6= (0, 0), then it has infinitely many. Proof. Let (m0, x0, y0) be an initial solution to (4.3). Rearranging the right- 82 4.4. Construction for k > 2 hand equality and substituting z = By yields the Pell equation z2 \u00E2\u0088\u0092BDx2 = B(AD \u00E2\u0088\u0092BC). (4.4) Define N := BD and let a + b \u00E2\u0088\u009A N be a fundamental unit in Z[ \u00E2\u0088\u009A N ]. Then (4.4) has an infinite family of integer solutions (zn, xn) given by( zn xn ) = ( a Nb b a )n( z0 x0 ) , where z0 = By0. Let M denote the 2 \u00C3\u0097 2 matrix above. The terms of this sequence are necessarily distinct because (z0, x0) 6= (0, 0) and the eigenvalues a \u00C2\u00B1 b\u00E2\u0088\u009AN of M are not roots of unity. If r is the order of M in the finite group SL2(Z/BZ), then the subsequence (zrn, xrn) is congruent to (z0, x0) modulo B. Thus, xrn \u00E2\u0089\u00A1 x0 (mod B) and zrn \u00E2\u0089\u00A1 z0 \u00E2\u0089\u00A1 0 (mod B), which ensures that both yrn = zrn/B and mrn = (x 2 rn +A)/B are integers. We can now give a sufficient condition for the existence of infinitely many 5-runs within a particular S(k). Theorem 4.11. Let k > 2 be fixed, and suppose there exists m > 0 such that km2 + 2 and km2 + 3 belong to S(k), hence km2 + 2 = a2 + kb2 and km2 + 3 = c2 +kd2 for some a, b, c, d \u00E2\u0088\u0088 Z. If (m\u00E2\u0088\u0092 b)(m\u00E2\u0088\u0092d) is not a square, then S(k) contains infinitely many 5-runs. Proof. Just as for Theorem 4.1, the inclusion {km2, km2+1, km2+4} \u00E2\u008A\u0082 S(k) is obvious, so it suffices to produce infinitely many m for which {km2 + 2, km2 + 3} \u00E2\u008A\u0082 S(k). We make the substitutions B = m \u00E2\u0088\u0092 b and D = m \u00E2\u0088\u0092 d, whereupon the conditions km2 + 2 = a2 + kb2 and km2 + 3 = c2 + kd2 are equivalent to m = a2 + (kB2 \u00E2\u0088\u0092 2) 2Bk = c2 + (kD2 \u00E2\u0088\u0092 3) 2Dk . (4.5) In light of Proposition 4.10, we may hold B and D fixed and obtain infinitely many solutions to (4.5) in m, a, c, provided that the initial solution 83 4.5. How many values of k? satisfies (a, c) 6= (0, 0) and BD > 0 (we have assumed by hypothesis that BD is not a square, and so neither is 4BDk2). But any solution to (4.5) for k > 2 satisfies a2 \u00E2\u0089\u00A1 2 (mod k), which means a 6= 0. Furthermore, we must have b 6= m or else a2 = 2; we must have b \u00E2\u0089\u00A4 m or else a2 = 2+k(m2\u00E2\u0088\u0092b2) < 2\u00E2\u0088\u0092 k < 0. Thus B > 0 and likewise D > 0 by an identical argument. The condition on (m\u00E2\u0088\u0092b)(m\u00E2\u0088\u0092d) in Theorem 4.11 can easily be weakened by observing that we may freely substitute b\u00E2\u0086\u0090 \u00E2\u0088\u0092b or d\u00E2\u0086\u0090 \u00E2\u0088\u0092d to modify the quantity (m\u00E2\u0088\u0092 b)(m\u00E2\u0088\u0092 d). The conclusion can only fail if all such values that result are simultaneously squares. We conjecture that such a coincidence cannot occur (in which case Theorem 4.11 would hold without this additional restriction), but we are unable to prove this. In any case, the previous observation immediately gives the following. Corollary 4.12. The conclusion of Theorem 4.11 also holds if any of the following quantities is not a square: (m\u00E2\u0088\u0092 b)(m\u00E2\u0088\u0092 d),m2 \u00E2\u0088\u0092 b2, or m2 \u00E2\u0088\u0092 d2. 4.5 How many values of k? The preceding corollary allows us to rapidly establish the existence of in- finitely many 5-runs in S(k) for a specific value of k, by finding a single one of the appropriate form. We have used this to verify, for 91 different values of k \u00E2\u0089\u00A4 5000, the existence of infinitely many 5-runs \u00E2\u0080\u0094 we have yet to find any generator that fails to satisfy the non-squareness conditions, which are presumably unnecessary. However, there is a rather strict condition on k implicit in this method: in order for km2 + 2 and km2 + 3 to belong to S(k), both 2 and 3 must be squares modulo k, which limits which primes p can divide k. Combining this observation with Theorem 4.8, we see by quadratic reciprocity that the following condition is necessary: Condition 4.13. k is equal to 2 times a product of (not necessarily distinct) primes p \u00E2\u0089\u00A1 \u00C2\u00B11 (mod 24). 84 4.5. How many values of k? Our computations suggest that this is the only restriction: the 91 values of k mentioned above are exactly those values of k \u00E2\u0089\u00A4 5000 which satisfy Condition 4.13. Even so, the set of such k still has asymptotic density 0, in contrast to the positive density given by Corollary 4.9. The method of Wirsing and Odoni (for example, [21, Proposition 4]) can be used to make this quantitatively precise: the number of k \u00E2\u0089\u00A4 x satisfying this condition is asymptotic to x/(log x)3/4 times an explicit constant. While S(34) does contain 5-runs (and computations suggest that they occur with approximately the expected frequency), it does not satisfy Con- dition 4.13, so we have no proof that it has infinitely many. It is natural to ask if there are infinitely many k for which the construction necessarily works: this is answered by Theorem 4.2, which we now prove. Conveniently, the proof uses Proposition 4.10 again; however, the Pellian construction inevitably gives a sequence of k that is far sparser than x/(log x)3/4. Proof of Theorem 4.2. We apply Theorem 4.11, which requires that we find families of solutions to km2 + 2 = a2 + kb2 and km2 + 3 = c2 + kd2. One such solution is given for k = 2 by 2(20)2 + 2 = 802 = 282 + 2(3)2 2(20)2 + 3 = 803 = 92 + 2(19)2. Now, observe that holding (m, b, d) = (20, 3, 19) fixed and solving for k gives k = a2 \u00E2\u0088\u0092 2 m2 \u00E2\u0088\u0092 b2 = c2 \u00E2\u0088\u0092 3 m2 \u00E2\u0088\u0092 d2 . (4.6) We now apply Proposition 4.10 to obtain an infinite sequence of tuples (k, a, c) satisfying (4.6), starting with (2, 28, 9) and noting that m2 \u00E2\u0088\u0092 b2 = 17 \u00C2\u00B723 and m2\u00E2\u0088\u0092d2 = 3 \u00C2\u00B713 satisfy the conditions on B and D. Since m, b and d are held constant and (m\u00E2\u0088\u0092b)(m\u00E2\u0088\u0092d) = 17, the conditions of Theorem 4.11 are satisfied across this entire sequence. The resulting values of k are clearly distinct, since a and c are determined up to sign by k. 85 4.6. Preliminaries for gaps 4.6 Preliminaries for gaps We now turn to the question of gaps in S(k), with the eventual goal of proving Theorem 4.3. The general strategy for the proof is simple: we first construct a polynomial P0(y) such that {P0(y), P0(y) + d} \u00E2\u008A\u0082 S(k) for any y. We then choose an arithmetic progression for y so that P0(y) + j /\u00E2\u0088\u0088 S(k) for 0 < j < d for any y in this progression; passing to this progression immediately yields the desired polynomial P0(an + b) in the conclusion of the theorem. While this basic outline is no different from that used by Spiro and Hensley [79], for general k there are some subtleties in choosing P0 that did not appear there. We will require several propositions which should be considered well- known (the first was essentially due to Euler in his characterization of sums of two squares). We provide the simple proofs here. Proposition 4.14. If a prime p satisfies ( \u00E2\u0088\u0092k p ) = \u00E2\u0088\u00921, and p \u00E2\u0080\u0096 n, then n /\u00E2\u0088\u0088 S(k). Proof. This follows immediately from the fact that x2 + ky2 \u00E2\u0089\u00A1 0 (mod p) has only the trivial solution x \u00E2\u0089\u00A1 y \u00E2\u0089\u00A1 0 (mod p) (else \u00E2\u0088\u0092k would have the square root xy\u00E2\u0088\u00921 modulo p). Proposition 4.15. If q1 < q2 < \u00C2\u00B7 \u00C2\u00B7 \u00C2\u00B7 < qr are primes and q0 = \u00E2\u0088\u00921, then for any choice of \u000F0, \u000F1, \u000F2, . . . , \u000Fr \u00E2\u0088\u0088 {\u00C2\u00B11}, there are infinitely many odd primes p which satisfy the constraints ( qi p ) = \u000Fi for each 0 \u00E2\u0089\u00A4 i \u00E2\u0089\u00A4 r. Proof. Recall that the constraint ( \u00E2\u0088\u00921 p ) = \u000F0 uniquely determines p as either 1 or 3 mod 4. For each odd qi, quadratic reciprocity then gives ( qi p ) = \u000Fi for p in any of 12(qi\u00E2\u0088\u00921) congruence classes mod 4qi (depending on the choice of \u000F0). Similarly, if q1 = 2, we easily check that the equation ( q1 p ) = \u000Fi admits a (unique) solution modulo 8, for any choice of \u000F0 and \u000F1. By Chinese remaindering, we obtain a congruence class for p modulo 4 \u00E2\u0088\u008F i qi, in which all constraints are simultaneously satisfied. The proposition now follows from Dirichlet\u00E2\u0080\u0099s theorem. 86 4.6. Preliminaries for gaps Lemma 4.16. Let A,B be positive integers such that AB is not a square. There are infinitely many primes p such that ( \u00E2\u0088\u0092A p ) = \u00E2\u0088\u00921 and ( \u00E2\u0088\u0092B p ) = 1. Proof. We assume without loss of generality that A and B are squarefree, as A and An2 have identical Legendre symbols, apart from the finitely many primes dividing n. Since AB is non-square (and positive), there exists some prime q dividing exactly one of A and B. If q | A, we may use Proposition 4.15 to choose p such that ( \u00E2\u0088\u00921 p ) = 1, ( q p ) = \u00E2\u0088\u00921, and( q\u00E2\u0080\u00B2 p ) = 1 for all other primes q\u00E2\u0080\u00B2 dividing AB. Likewise, if q | B, we choose( \u00E2\u0088\u00921 p ) = ( q p ) = \u00E2\u0088\u00921, and ( q\u00E2\u0080\u00B2 p ) = 1 for all other q\u00E2\u0080\u00B2 dividing AB. Remark. This lemma is a special case of the more general principle that, if A1, A2, . . . , Ar are multiplicatively independent integers (in the sense that no non-empty subset has a product equal to a square), then for any assignment of \u00C2\u00B11 values to the Legendre symbols ( Ai p ) , there are infinitely many primes p which satisfy that assignment (in fact the set of such p has density exactly 2\u00E2\u0088\u0092r relative to the set of all primes). Proposition 4.17. Let f(x) = ax2 + bx + c be a quadratic with integer coefficients and discriminant \u00E2\u0088\u0086 = b2\u00E2\u0088\u0092 4ac. Suppose p is an odd prime such that p - a and ( \u00E2\u0088\u0086 p ) = 1. Then there exists a number m such that p \u00E2\u0080\u0096 f(m). Proof. Under the given assumptions, the quadratic formula readily yields a solution to f(m) \u00E2\u0089\u00A1 p (mod p), namely m \u00E2\u0089\u00A1 (\u00E2\u0088\u0092b + \u00E2\u0088\u009A\u00E2\u0088\u0086)/2a (mod p). Furthermore, since \u00E2\u0088\u0086 is nonzero mod p, the derivative f \u00E2\u0080\u00B2(m) = 2am + b is non-vanishing, and so Hensel\u00E2\u0080\u0099s lemma ensures that we can lift m to a solution of f(m) \u00E2\u0089\u00A1 p (mod p2), whence p exactly divides f(m). Proposition 4.18. Let be P (n) be a nonconstant polynomial of degree r, and B > 0 a fixed integer. Then the asymptotic density of n for which P (n) is B-friable (has no prime factors exceeding B) is 0. Proof. By an elementary counting argument, the number of distinct B- friable values up to y is \u001C (log y)B. For any n \u00E2\u0089\u00A4 x, P (n) is bounded 87 4.7. Proof of existence of gaps by Cxr (for some constant C) and takes each value at most r times, so there are at most \u001Cr,P r(r log x+C)B \u001Cr,P (log x)B values of n \u00E2\u0089\u00A4 x where P (n) is B-friable. Since B is fixed (one could even allow B to grow slowly with x), this set has asymptotic density 0. 4.7 Proof of existence of gaps Proof of Theorem 4.3. Let P0(y) be the quadratic polynomial x 2 + ky2, where x is a linear function of y with integer coefficients to be chosen shortly. Clearly, we have P0(y) \u00E2\u0088\u0088 S(k) for any y. To ensure that P0(y) + d \u00E2\u0088\u0088 S(k), we write P0(y) + d = (x + \u00CE\u00B4) 2 + k(y \u00E2\u0088\u0092 \u00CE\u00B7)2, where \u00CE\u00B4 and \u00CE\u00B7 are constants to be chosen below. This is equivalent to d = 2\u00CE\u00B4x+ \u00CE\u00B42 \u00E2\u0088\u0092 2k\u00CE\u00B7y + k\u00CE\u00B72, or x = d\u00E2\u0088\u0092 \u00CE\u00B42 \u00E2\u0088\u0092 k\u00CE\u00B72 2\u00CE\u00B4 + k\u00CE\u00B7 \u00CE\u00B4 y. (4.7) For a given k and d, we will fix \u00CE\u00B4 and restrict \u00CE\u00B7 so that the above expression, viewed as a linear function of y, has integer coefficients. This can be done in several cases as follows: \u00E2\u0080\u00A2 If d is odd, we may simply choose \u00CE\u00B4 = 1 and \u00CE\u00B7 = 0 (so x is constant). \u00E2\u0080\u00A2 If d is even and k is odd, we choose \u00CE\u00B4 = 1 and restrict \u00CE\u00B7 to be odd. \u00E2\u0080\u00A2 If d \u00E2\u0089\u00A1 0 (mod 4), we choose \u00CE\u00B4 = 2 and \u00CE\u00B7 even. \u00E2\u0080\u00A2 If d \u00E2\u0089\u00A1 k \u00E2\u0089\u00A1 2 (mod 4), we choose \u00CE\u00B4 = 2 and \u00CE\u00B7 odd. Note that these cases cover all possible choices of k and d satisfying the hypotheses (there is a small amount of overlap between the cases, which is not important). For any \u00CE\u00B7 subject to the above restrictions, we now choose x as determined by (4.7) (which is linear in y), so that x2 + ky2 + d \u00E2\u0088\u0088 S(k) for any y. It remains to choose a specific \u00CE\u00B7 subject to the above restrictions, as well as an arithmetic progression for y that induces a genuine gap of size d at x2 + ky2: we need x2 + ky2 + j /\u00E2\u0088\u0088 S(k) for every 0 < j < d. 88 4.7. Proof of existence of gaps In the first case that d is odd, we have already chosen \u00CE\u00B7 = 0. For all other cases, we fix \u00CE\u00B7 to be any integer (subject to the parity constraints above) such that \u00CE\u00B42 + k\u00CE\u00B72 has a prime factor q which exceeds 2d. The existence of such an \u00CE\u00B7 is guaranteed by Proposition 4.18, as very few choices of \u00CE\u00B42 + k\u00CE\u00B72 are 2d-friable. To exclude x2 + ky2 + j from S(k) for a given j, it suffices by Proposi- tion 4.14 to exhibit a prime pj such that ( \u00E2\u0088\u0092k pj ) = \u00E2\u0088\u00921 and pj \u00E2\u0080\u0096 x2 + ky2 + j for some appropriate choice of y. Recall that x and y are not independently chosen but related by (4.7), which we abbreviate as x = Cy +D by taking C := k\u00CE\u00B7 \u00CE\u00B4 , D := d\u00E2\u0088\u0092 \u00CE\u00BE 2\u00CE\u00B4 , where \u00CE\u00BE := \u00CE\u00B42 + k\u00CE\u00B72. Then x2 + ky2 + j, as a function of y, is the quadratic P0(y) = (C 2 + k)y2 + 2CDy + (D2 + j). In order to apply Proposition 4.17, we must verify that the discriminant of P0 is a nonzero square mod pj (we also need pj to be coprime to C2 + k, but since C and k are fixed, there are only finitely many values of pj to exclude). Specifically, we need( C2D2 \u00E2\u0088\u0092 (C2 + k)(D2 + j) pj ) = (\u00E2\u0088\u0092kD2 \u00E2\u0088\u0092 j(C2 + k) pj ) = 1. (4.8) By Lemma 4.16, there are infinitely many odd primes pj satisfying both( \u00E2\u0088\u0092k pj ) = \u00E2\u0088\u00921 and (4.8), provided that the product k2D2 + jk(C2 + k) is not a square. Noting that C2 + k = k\u00CE\u00BE/\u00CE\u00B42, this is equivalent, up to a square factor of k2/4\u00CE\u00B42, to \u00E2\u0088\u0086j := (d\u00E2\u0088\u0092 \u00CE\u00BE)2 + 4\u00CE\u00BEj. (4.9) We claim that \u00E2\u0088\u0086j is not square for any value of 0 < j < d. Note that this quantity is unavoidably square at the endpoints \u00E2\u0088\u00860 = (d \u00E2\u0088\u0092 \u00CE\u00BE)2 and \u00E2\u0088\u0086d = (d+ \u00CE\u00BE) 2, by virtue of the construction. When d is odd, the fact that \u00E2\u0088\u0086j is not square is immediate: we chose \u00CE\u00B4 = 1 and \u00CE\u00B7 = 0, so \u00CE\u00BE = 1 and there are no even squares between (d \u00E2\u0088\u0092 1)2 and (d+ 1)2. 89 4.7. Proof of existence of gaps When d is even, recall that we chose \u00CE\u00B7 so that \u00CE\u00BE = \u00CE\u00B42 + k\u00CE\u00B72 has a large prime factor q > 2d. In this case \u00CE\u00BE is now larger than d, so it is more natural to write \u00E2\u0088\u0086j as (\u00CE\u00BE \u00E2\u0088\u0092 d)2 + 4\u00CE\u00BEj. We now show that (\u00CE\u00BE \u00E2\u0088\u0092 d)2 + 4\u00CE\u00BEj = A2 admits no integer solutions for 0 < j < d, or equivalently with \u00CE\u00BE \u00E2\u0088\u0092 d < A < \u00CE\u00BE + d. Since q | \u00CE\u00BE, any such solution must have A2 \u00E2\u0089\u00A1 d2 (mod q), whence A \u00E2\u0088\u0092 \u00CE\u00BE \u00E2\u0089\u00A1 A \u00E2\u0089\u00A1 \u00C2\u00B1d (mod q). But d and \u00E2\u0088\u0092d are necessarily the smallest representatives of their respective congruence classes mod q (because q > 2d), so we must have |A\u00E2\u0088\u0092 \u00CE\u00BE| \u00E2\u0089\u00A5 d. Putting this all together, for each 0 < j < d we may now choose a distinct prime pj satisfying ( \u00E2\u0088\u0092k pj ) = \u00E2\u0088\u00921 so that pj \u00E2\u0080\u0096 P0(mj)+j for some mj . By choosing y \u00E2\u0089\u00A1 mj (mod p2j ) simultaneously for all 0 < j < d, we obtain a congruence class for y modulo \u00E2\u0088\u008F j p 2 j such that P0(y) + j /\u00E2\u0088\u0088 S(k), so that (P0(y), P0(y) + d) is indeed a gap of size d. Remark. The above proof yields a quantitative lower bound of\u001Dk,d \u00E2\u0088\u009A x gaps of size d in S(k) up to x, where the implied constant decays rather rapidly with d. In the case of S(1) we can do much better: Hooley\u00E2\u0080\u0099s quantitative lower bounds for the number of pairs {m,m+ 1} in S(1) can be generalized to yield a lower bound of \u001Dd x/ log x for the number of {m,m+ d} pairs in S(1). While not every such pair is necessarily a gap, Selberg\u00E2\u0080\u0099s sieve shows for any 0 < c < d, the triple {m,m+c,m+d} appears only\u001Cc,d x/(log x)3/2 times. Since there are only Od(1) possible choices for c, this shows that the majority of pairs {m,m + d} have no other elements in between, and are genuine gaps. 90 Chapter 5 Conclusion 5.1 Arithmetic functions We studied a broad class of arithmetic functions f(n) in Chapter 2, showing that (with some restrictions) for any \u00CE\u00B5 > 0 and any choices for \u00CE\u00B11, \u00CE\u00B12, . . . , \u00CE\u00B1h, the approximations |fi(ain+ bi)\u00E2\u0088\u0092 \u00CE\u00B1i| < \u000F (i = 1, 2, . . . , h) (5.1) occur simultaneously for a positive proportion of values n. This generalizes the Erdo\u00CC\u008Bs\u00E2\u0080\u0093Schinzel theorem in two directions: one may freely compose the functions with different arithmetic progressions, and also freely choose a function at each coordinate 1 \u00E2\u0089\u00A4 i \u00E2\u0089\u00A4 h. Let us offer an interpretation of this latter aspect that seems intuitively attractive. To paraphrase the earlier results, a single value of n may aim at several targets simultaneously, using the values of a fixed arithmetic func- tion; moreover, each individual function (such as \u00CF\u0083(n)/n or \u00CF\u0086(n)/n) gives rise to a large set of such approximants n. What our Corollary 2.11 adds is that the sets arising from different functions also have large intersection: they cannot conspire to avoid one another \u00E2\u0080\u0094 unless condition (2.16) is vio- lated, making them too interdependent. We might say they are distributed \u00E2\u0080\u009Crandomly\u00E2\u0080\u009D enough that their intersections behave (up to an extremely large local factor) like those of random sets having similar size. We suspect that there is not much room for further generalization if our goal is to retain a positive proportion of approximants n. However, we have already described the contemporary work of Alkan, Ford and Zaharescu which passes up this requirement in exchange for a much more effective 91 5.2. Random matrices approximation (one whose accuracy improves nicely with n). More recently, Luca, Mej\u00CC\u0081\u00C4\u00B1a Huguet and Nicolae [51] have taken such results in a novel direction: they prove a result analogous to Schinzel and Wang\u00E2\u0080\u0099s (1.5), except where \u00CF\u0086(n) is composed with the Fibonacci sequence Fn. In other words, the following sequence of points is also dense in the positive orthant of Rd:( \u00CF\u0086(Fn+1) \u00CF\u0086(Fn) , \u00CF\u0086(Fn+2) \u00CF\u0086(Fn) , . . . , \u00CF\u0086(Fn+d) \u00CF\u0086(Fn) ) . Their proof makes use of the fact that Fibonacci numbers, although ex- ponentially sparse, have a rich divisor structure that enables the sort of constructions we saw in the proof of Theorem 2.3; the same techniques work for similar sequences like 2n\u00E2\u0088\u0092 1 (but not 2n, which is very poor in divisors). Many difficult problems on arithmetic functions can be solidly answered when restricted to these sequences. For instance, while no one has entirely ruled out odd perfect numbers, Luca [50] has shown that no Fibonacci num- ber is perfect, and Pollack [67] proved that there are no perfect numbers composed of the same digit repeatedly (excluding the trivial example 6). Seeing that approximation properties are robust against composing with such exotic sequences rather than linear functions, we think it would be interesting to consider compositions with more natural (but less charming) sequences, such as higher-degree polynomials. 5.2 Random matrices In Chapter 3, we studied questions regarding the number of n \u00C3\u0097 n integer matrices having a particular eigenvalue. In particular, we established upper bounds for the singularity probability in general, as well as the exact distri- bution of real and rational eigenvalues for n = 2. The latter investigation revealed some rather unexpected differences between the two, and the loca- tion of the mode(s) strikes the author as the most surprising aspect worthy of investigation: for instance, might further modes appear as n increases, and is there a natural explanation for their appearance? 92 5.2. Random matrices Our singularity bounds have been considerably improved by Shparlinski [78], who obtained stronger bounds, more generally for matrices shifted by adding a constant matrix (also more generally, having a prescribed determi- nant rather than just 0). As mentioned in the last section of said chapter, the precise number of integer matrices which are diagonalizable over Q, or have all eigenvalues simultaneously rational, remains open for any n \u00E2\u0089\u00A5 3. We know of only one result specific to such matrices, in the special case of symmetric (0, 1)-matrices, also corresponding to integral graphs. These were considered by Ahmadi et al. [1] and shown to occur with exponentially small probability as n increases. Also, Maze [57] has given very precise distributional results on the fine structure23 of random integer matrices. Just as in the eigenvalue problem, there is a wealth of results on the singularity probability in the converse case where we fix the size of entries k (thus fixing the distribution of each entry) and let n increase. Komlo\u00CC\u0081s [46] first showed that large random matrices are rarely singular, regardless of the distribution; that is, as n\u00E2\u0086\u0092\u00E2\u0088\u009E, the singularity probability shrinks to 0 for any fixed k. In fact, it decreases exponentially quickly, as first shown by Kahn, Komlo\u00CC\u0081s and Szemere\u00CC\u0081di [42] for (0, 1)-matrices, and in much greater generality by Rudelson and Vershynin [69]. Recently, Bourgain, Vu and Wood [10] have proved an impressive gen- eral bound for the singularity probability, one that incorporates both the polynomial decay observed in k (for fixed n), and the exponential decay seen in n (when k fixed). For n sufficiently large their bound, in the case of the matrices Mn(k), is essentially (2k)\u00E2\u0088\u0092n/2. It would be interesting to explore how explicit such a bound could be made, which might yield a far more effective version of Proposition 3.6. The author\u00E2\u0080\u0099s recent efforts in this area are concerned with obtaining explicit lower bounds for the magnitude of the determinant of a typical random matrix (the determinant may be viewed as a crude measure of how far away a matrix is from being singular; once again, there is a rich history of bounding determinants away from zero in various other contexts, for instance [69, 81]). 23However, this concerns the Hermite normal form (a division-free version of row echelon form) rather than the spectrum. 93 5.3. Patterns in quadratic forms 5.3 Patterns in quadratic forms We have studied the patterns of runs and gaps in the set S(k) = {x2 + ky2 : x, y \u00E2\u0088\u0088 Z}. We gave a qualitatively complete answer to the existence of gaps for all S(k), but significant quantitative questions remain. For instance, the largest gap in S(k) up to x is certainly at least \u00E2\u0088\u009A log x, and heuristically should be not much larger (a power of log x). However, the best known upper bound for this quantity is O(x1/4), and it comes from an obvious argument: [59, p. 43] describes this state of the art as \u00E2\u0080\u009Csomewhat embarrassing\u00E2\u0080\u009D.24 Even qualitatively, much remains unknown about runs, with S(34) being the first of many examples for which the existence of infinitely many 5-runs is open. One avenue of pursuit considered by the author has been to find new polynomial identities which could take the place of the trivial inclusions in the proofs of Theorem 4.1 and Theorem 4.11, thus extending the reach of this method beyond current limitations. For more general binary quadratic forms, we noted previously that they may (locally) admit extremely long runs, for instance the (positive definite) form xy+k!(x2+y2) for large k. Such forms are certainly worth investigating and are likely amenable to the methods of Theorem 4.3, but clearly new ideas are needed to tackle runs of any significant length. It seems unlikely that analytic methods can prevail for runs of such length, either: even for 3-runs, we know of no result remotely comparable in strength to (4.1). Considering the spectacular success of Green, Tao and Ziegler in enu- merating arithmetic progressions of primes, one naturally expects to obtain better quantitative results for APs. Matthiesen [56] has very recently used those methods to obtain an asymptotic formula for the number of `-APs in S(k), vastly generalizing earlier work of Heath-Brown [35]. The catch is that in both results, the elements of S(k) are weighted with multiplicity: that is, since there are 4 solutions to x2 + y2 = 1, and 4 solutions to x2 + y2 = 2, the progression {0, 1, 2} counts as appearing 16 times in S(1). Such an interpretation is fairly natural, and it does not detract from the 24Upper bounds for prime gaps are also very far from predictions, but they may take pride in actually progressing: the current best is due to Baker, Harman and Pintz [3]. 94 5.3. Patterns in quadratic forms beauty of Matthiesen\u00E2\u0080\u0099s sharp result. However, the multiplicity rk(n) of a given n \u00E2\u0088\u0088 S(k) could be as small as 2, or as large as exp(ck log n/ log log n), depending on n. This makes for a rather wide variation in the number of times each arithmetic progression gets counted, so it is difficult to make any deductions about the number of distinct progressions this way. The primary focus of the author\u00E2\u0080\u0099s current research is to obtain sharper results for this unweighted problem in S(k). Such precision is typically more accessible in the weighted case considered by Heath-Brown and Matthiesen.25 Even though rk(n) fluctuates highly with n, its sum \u00E2\u0088\u0091x n=1 rk(n) is under- stood with reasonable accuracy (at least for k fixed), certainly better than the error term in Landau\u00E2\u0080\u0099s unweighted count (1.9). Still, one might hope to obtain estimates akin to Hooley\u00E2\u0080\u0099s (4.1), estab- lishing at least the correct order of magnitude which is not currently known from existing results. A result which dates back to de la Valle\u00CC\u0081e Poussin\u00E2\u0080\u0099s original 1897 proof of the Prime Number Theorem [61, p. 72] is that S(k) contains a specific positive fraction of all primes. The Green\u00E2\u0080\u0093Tao theo- rem therefore shows that the number of r-APs in S(k) up to x is at least \u001Dk,d x2/(log x)d. It is plausible that a careful unweighting of Matthiesen\u00E2\u0080\u0099s estimate might do better (by some positive power of log x). However, we expect that the powerful Green\u00E2\u0080\u0093Tao machinery can be used to improve this all the way to \u001Dk,d x2/(log x)d/2, as predicted by random heuristics (and matching the upper bounds given by sieves). In fact, it should even be possible to get lower bounds for some higher-order arithmetic structures, such as pairs of arithmetic progressions {m,m + d,m + 2d} \u00E2\u0088\u00AA {m+ 1,m+d+ 1,m+ 2d+ 1}: Zhou [87] has established this for primes, on assumption of the twin primes conjecture (and unconditionally for almost- primes having only 1 or 2 prime factors). 25If the reader will forgive the anthropomorphism, it is as though n \u00E2\u0088\u0088 S(k) wants to be counted with weight rk(n). The primes also want to be counted with the von Mangoldt weight \u00CE\u009B(n), but this varies very smoothly, making it a simple matter to remove. 95 Bibliography [1] O. Ahmadi, N. Alon, I. F. Blake, and I. E. Shparlinski. Graphs with integral spectrum. Linear Algebra Appl., 430(1):547\u00E2\u0080\u0093552, 2009. [2] Emre Alkan, Kevin Ford, and Alexandru Zaharescu. Diophantine ap- proximation with arithmetic functions. I. Trans. Amer. Math. Soc., 361(5):2263\u00E2\u0080\u00932275, 2009. [3] R. C. Baker, G. Harman, and J. Pintz. The difference between consec- utive primes. II. Proc. London Math. Soc. (3), 83(3):532\u00E2\u0080\u0093562, 2001. [4] Paul T. Bateman and Harold G. Diamond. Analytic number theory. World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2004. [5] Paul T. Bateman and Roger A. Horn. A heuristic asymptotic formula concerning the distribution of prime numbers. Math. Comp., 16:363\u00E2\u0080\u0093 367, 1962. [6] P. Bernays. U\u00CC\u0088ber die Darstellung von positiven, ganzen Zahlen durch die primitiven, bina\u00CC\u0088ren quadratischen Formen einer nicht-quadratischen Diskriminante. PhD thesis, Georg-August-Universita\u00CC\u0088t, Go\u00CC\u0088ttingen, 1912. [7] Manjul Bhargava. Higher composition laws. I. A new view on Gauss composition, and quadratic generalizations. Ann. of Math. (2), 159(1):217\u00E2\u0080\u0093250, 2004. [8] Valentin Blomer and Andrew Granville. Estimates for representation numbers of quadratic forms. Duke Math. J., 135(2):261\u00E2\u0080\u0093302, 2006. 96 Bibliography [9] Peter Borwein, Stephen Choi, Brendan Rooney, and Andrea Weirath- mueller, editors. The Riemann Hypothesis: A Resource for the Affi- cionado and Virtuoso Alike. CMS Books in Mathematics/Ouvrages de Mathe\u00CC\u0081matiques de la SMC. Springer, New York, 2008. [10] Jean Bourgain, Van H. Vu, and Philip Matchett Wood. On the singular- ity probability of discrete random matrices. J. Funct. Anal., 258(2):559\u00E2\u0080\u0093 603, 2010. [11] S. Brin and L. Page. The anatomy of a large-scale hypertextual web search engine. In Seventh International World-Wide Web Conference (WWW 1998), 1998. [12] V. Brun. La se\u00CC\u0081rie 15 + 1 7 + 1 11 + 1 13 + 1 17 + 1 19 + 1 29 + 1 31 + 1 41 + 1 43 + 1 59 + 1 61 +\u00C2\u00B7 \u00C2\u00B7 \u00C2\u00B7 ou\u00CC\u0080 les de\u00CC\u0081nominateurs sont \u00E2\u0080\u009Cnombres premiers jumeaux\u00E2\u0080\u009D est convergente ou finie. Darboux Bull. (2) 43, 100-104:124\u00E2\u0080\u0093128, 1919. [13] D. A. Buell and R. H. Hudson. On runs of consecutive quadratic residues and quadratic nonresidues. BIT, 24(2):243\u00E2\u0080\u0093247, 1984. [14] Todd Cochrane and Robert E. Dressler. Consecutive triples of sums of two squares. Arch. Math. (Basel), 49(4):301\u00E2\u0080\u0093304, 1987. [15] Harold Davenport. U\u00CC\u0088ber numeri abundantes. Sitzungsber. Preu\u00C3\u009F. Akad. Wiss., Phys.-Math. Kl., 1933(26-29):830\u00E2\u0080\u0093837, 1933. [16] Alan Edelman. The probability that a random real Gaussian matrix has k real eigenvalues, related distributions, and the circular law. J. Multivariate Anal., 60(2):203\u00E2\u0080\u0093232, 1997. [17] P. Erdo\u00CC\u008Bs and M. Kac. The Gaussian law of errors in the theory of additive number theoretic functions. Amer. J. Math., 62:738\u00E2\u0080\u0093742, 1940. [18] Paul Erdo\u00CC\u008Bs. On the smoothness of the asymptotic distribution of addi- tive arithmetical functions. Amer. J. Math., 61:722\u00E2\u0080\u0093725, 1939. [19] Paul Erdo\u00CC\u008Bs and Aurel Wintner. Additive arithmetical functions and statistical independence. Amer. J. Math., 61:713\u00E2\u0080\u0093721, 1939. 97 Bibliography [20] P. Erdo\u00CC\u008Bs and A. Schinzel. Distributions of the values of some arithmeti- cal functions. Acta Arith., 6:473\u00E2\u0080\u0093485, 1960/1961. [21] Steven Finch, Greg Martin, and Pascal Sebah. Roots of unity and nullity modulo n. Proc. Amer. Math. Soc., 138(8):2729\u00E2\u0080\u00932743, 2010. [22] Carl Friedrich Gauss. Disquisitiones arithmeticae. Springer-Verlag, New York, 1986. Translated and with a preface by Arthur A. Clarke, Re- vised by William C. Waterhouse, Cornelius Greither and A. W. Groo- tendorst and with a preface by Waterhouse. [23] S. Gershgorin. U\u00CC\u0088ber die Abgrenzung der Eigenwerte einer Matrix. Izv. Akad. Nauk SSSR, Otd. Mat. Estest. Nauk, VII. Ser. No., 6:749\u00E2\u0080\u0093754, 1931. [24] Jean Ginibre. Statistical ensembles of complex, quaternion, and real matrices. J. Mathematical Phys., 6:440\u00E2\u0080\u0093449, 1965. [25] V. L. Girko. The circular law. Teor. Veroyatnost. i Primenen., 29(4):669\u00E2\u0080\u0093679, 1984. [26] Daniel A. Goldston, Ja\u00CC\u0081nos Pintz, and Cem Y. Y\u00C4\u00B1ld\u00C4\u00B1r\u00C4\u00B1m. Primes in tuples. I. Ann. of Math. (2), 170(2):819\u00E2\u0080\u0093862, 2009. [27] W. T. Gowers. A new proof of Szemere\u00CC\u0081di\u00E2\u0080\u0099s theorem. Geom. Funct. Anal., 11(3):465\u00E2\u0080\u0093588, 2001. [28] Andrew Granville. Harald Crame\u00CC\u0081r and the distribution of prime num- bers. Scand. Actuar. J., 1:12\u00E2\u0080\u009328, 1995. Harald Crame\u00CC\u0081r Symposium (Stockholm, 1993). [29] Ben Green and Terence Tao. The primes contain arbitrarily long arith- metic progressions. Ann. of Math. (2), 167(2):481\u00E2\u0080\u0093547, 2008. [30] Ben Green and Terence Tao. Linear equations in primes. Ann. of Math. (2), 171(3):1753\u00E2\u0080\u00931850, 2010. [31] Ben Green and Terence Tao. The Mo\u00CC\u0088bius function is strongly orthog- onal to nilsequences. Ann. of Math. (2), 175(2):541\u00E2\u0080\u0093566, 2012. 98 Bibliography [32] Ben Green and Terence Tao. The quantitative behaviour of polynomial orbits on nilmanifolds. Ann. of Math. (2), 175(2):465\u00E2\u0080\u0093540, 2012. [33] Ben Green, Terence Tao, and Tamar Ziegler. An inverse theorem for the Gowers U s+1[N ]-norm. Ann. of Math. (2). To appear. Preprint available at arXiv:1009.3998v2. [34] G. H. Hardy and J. E. Littlewood. Some problems of \u00E2\u0080\u0098Partitio numero- rum\u00E2\u0080\u0099; III: On the expression of a number as a sum of primes. Acta Math., 44(1):1\u00E2\u0080\u009370, 1923. [35] D. R. Heath-Brown. Linear relations amongst sums of two squares. In Number theory and algebraic geometry, volume 303 of London Math. Soc. Lecture Note Ser., pages 133\u00E2\u0080\u0093176. Cambridge Univ. Press, Cam- bridge, 2003. [36] Jim Hefferon. Linear Algebra (online book). http://joshua.smcvt. edu/linearalgebra/, accessed May 17, 2012. [37] Andrew J. Hetzel, Jay S. Liew, and Kent E. Morrison. The probabil- ity that a matrix of integers is diagonalizable. Amer. Math. Monthly, 114(6):491\u00E2\u0080\u0093499, 2007. [38] C. Hooley. On the intervals between numbers that are sums of two squares. II. J. Number Theory, 5:215\u00E2\u0080\u0093217, 1973. [39] Christopher Hooley. On the intervals between numbers that are sums of two squares. III. J. Reine Angew. Math., 267:207\u00E2\u0080\u0093218, 1974. [40] Bernard Host and Bryna Kra. Nonconventional ergodic averages and nilmanifolds. Ann. of Math. (2), 161(1):397\u00E2\u0080\u0093488, 2005. [41] James P. Jones, Daihachiro Sato, Hideo Wada, and Douglas Wiens. Diophantine representation of the set of prime numbers. Amer. Math. Monthly, 83(6):449\u00E2\u0080\u0093464, 1976. [42] J. Kahn, J. Komlo\u00CC\u0081s, and E. Szemere\u00CC\u0081di. On the probability that a random\u00C2\u00B11-matrix is singular. J. Amer. Math. Soc., 8(1):223\u00E2\u0080\u0093240, 1995. 99 Bibliography [43] Y. R. Katznelson. Singular matrices and a uniform bound for congru- ence groups of SLn(Z). Duke Math. J., 69(1):121\u00E2\u0080\u0093136, 1993. [44] K. Kedlaya and L. Ng. Solutions to the 67th William Lowell Putnam Mathematical Competition. http://amc.maa.org/a-activities/ a7-problems/putnam/-pdf/2006s.pdf (online resource), accessed May 23, 2012. [45] Mitsuo Kobayashi. On the Density of Abundant Numbers. PhD thesis, Dartmouth College, 2010. [46] J. Komlo\u00CC\u0081s. On the determinant of random matrices. Studia Sci. Math. Hungar., 3:387\u00E2\u0080\u0093399, 1968. [47] Hans-Joachim Kowalsky. Ganzzahlige Matrizen mit ganzzahligen Eigenwerten. Abh. Braunschweig. Wiss. Ges., 34:15\u00E2\u0080\u009332, 1982. [48] E. Landau. U\u00CC\u0088ber die Einteilung der positiven ganzen Zahlen in vier Klassen nach der Mindestzahl der zu ihrer additiven Zusammensetzung erforderlichen Quadrate. Arch. der Math. u. Phys. (3), 13:305\u00E2\u0080\u0093312, 1908. [49] Leonid A. Levin. Average case complete problems. SIAM J. Comput., 15(1):285\u00E2\u0080\u0093286, 1986. [50] Florian Luca. Perfect Fibonacci and Lucas numbers. Rend. Circ. Mat. Palermo (2), 49(2):313\u00E2\u0080\u0093318, 2000. [51] Florian Luca, V. Janitzio Mej\u00CC\u0081\u00C4\u00B1a Huguet, and Florin Nicolae. On the Euler function of Fibonacci numbers. J. Integer Seq., 12(6):Article 09.6.6, 15pp., 2009. [52] Helmut Maier. Primes in short intervals. Michigan Math. J., 32(2):221\u00E2\u0080\u0093 225, 1985. [53] Greg Martin. The smallest solution of \u00CF\u0086(30n+1) < \u00CF\u0086(30n) is . . .. Amer. Math. Monthly, 106(5):449\u00E2\u0080\u0093451, 1999. 100 Bibliography [54] Greg Martin and Erick Wong. The number of 2 \u00C3\u0097 2 integer matri- ces having a prescribed integer eigenvalue. Algebra Number Theory, 2(8):979\u00E2\u0080\u00931000, 2008. [55] Greg Martin and Erick B. Wong. Almost all integer matrices have no integer eigenvalues. Amer. Math. Monthly, 116(7):588\u00E2\u0080\u0093597, 2009. [56] L. Matthiesen. Linear correlations amongst numbers represented by positive definite binary quadratic forms. Acta Arith. To appear. Preprint available at arXiv:1106.4690. [57] Ge\u00CC\u0081rard Maze. Natural density distribution of Hermite normal forms of integer matrices. J. Number Theory, 131(12):2398\u00E2\u0080\u00932408, 2011. [58] H. L. Montgomery. The pair correlation of zeros of the zeta function. In Analytic number theory (Proc. Sympos. Pure Math., Vol. XXIV, St. Louis Univ., St. Louis, Mo., 1972), pages 181\u00E2\u0080\u0093193. Amer. Math. Soc., Providence, R.I., 1973. [59] Hugh L. Montgomery and Robert C. Vaughan. Multiplicative number theory. I. Classical theory, volume 97 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 2007. [60] T. Muir. A treatise on the theory of determinants. Revised and enlarged by William H. Metzler. Dover Publications Inc., New York, 1960. [61] W ladys law Narkiewicz. The development of prime number theory. Springer Monographs in Mathematics. Springer-Verlag, Berlin, 2000. From Euclid to Hardy and Littlewood. [62] D. J. Newman. Euler\u00E2\u0080\u0099s \u00CF\u0086 function on arithmetic progressions. Amer. Math. Monthly, 104(3):256\u00E2\u0080\u0093257, 1997. [63] Thomas R. Nicely. Enumeration to 1014 of the twin primes and Brun\u00E2\u0080\u0099s constant. Virginia J. Sci., 46(3):195\u00E2\u0080\u0093204, 1995. [64] Ivan Niven. Numbers: rational and irrational, volume 1 of New Math- ematical Library. Random House, New York, 1961. 101 Bibliography [65] Ivan Niven, Herbert S. Zuckerman, and Hugh L. Montgomery. An introduction to the theory of numbers. John Wiley & Sons Inc., New York, fifth edition, 1991. [66] Ga\u00CC\u0081bor Pataki, Mustafa Tural, and Erick B. Wong. Basis reduction and the complexity of branch-and-bound. In Proceedings of the Twenty- First Annual ACM-SIAM Symposium on Discrete Algorithms, pages 1254\u00E2\u0080\u00931261, Philadelphia, PA, 2010. SIAM. [67] Paul Pollack. Perfect numbers with identical digits. Integers, 11A:A18, 11pp., 2011. [68] K. F. Roth. On certain sets of integers. J. London Math. Soc., 28:104\u00E2\u0080\u0093 109, 1953. [69] Mark Rudelson and Roman Vershynin. The Littlewood-Offord problem and invertibility of random matrices. Adv. Math., 218(2):600\u00E2\u0080\u0093633, 2008. [70] Tom Sanders. On Roth\u00E2\u0080\u0099s theorem on progressions. Ann. of Math. (2), 174(1):619\u00E2\u0080\u0093636, 2011. [71] A. Schinzel and W. Sierpin\u00CC\u0081ski. Sur certaines hypothe\u00CC\u0080ses concernant les nombres premiers. Acta Arith. 4 (1958), 185\u00E2\u0080\u0093208; erratum, 5:259, 1958. [72] A. Schinzel and Y. Wang. A note on some properties of the functions \u00CF\u0095(n), \u00CF\u0083(n) and \u00CE\u00B8(n). Ann. Polon. Math., 4:201\u00E2\u0080\u0093213, 1958. [73] I. Schoenberg. U\u00CC\u0088ber die asymptotische Verteilung reeller Zahlen mod 1. M. Z., 28:171\u00E2\u0080\u0093199, 1928. [74] I. J. Schoenberg. On asymptotic distributions of arithmetical functions. Trans. Amer. Math. Soc., 39(2):315\u00E2\u0080\u0093330, 1936. [75] Robert Forsyth Scott. The theory of determinants and their applica- tions. Revised by G. B. Mathews. Cambridge University Press, Cam- bridge, 1904. 102 Bibliography [76] J.-P. Serre. A course in arithmetic. Springer-Verlag, New York, 1973. Translated from the French, Graduate Texts in Mathematics, No. 7. [77] Pin-Tsung Shao. On the distribution of the values of a class of arith- metical functions. Bull. Acad. Pol. Sci., Cl. III, 4:569\u00E2\u0080\u0093572, 1956. [78] I. E. Shparlinski. Some counting questions for matrices with restricted entries. Linear Algebra Appl., 432(1):155\u00E2\u0080\u0093160, 2010. [79] Claudia A. Spiro and Douglas A. Hensley. Problems and Solutions: So- lutions of Advanced Problems: 6539. Amer. Math. Monthly, 97(3):253\u00E2\u0080\u0093 254, 1990. [80] E. Szemere\u00CC\u0081di. On sets of integers containing no k elements in arithmetic progression. Acta Arith., 27:199\u00E2\u0080\u0093245, 1975. Collection of articles in memory of Juri\u00CC\u0086\u00C4\u00B1 Vladimirovic\u00CC\u008C Linnik. [81] Terence Tao and Van Vu. On random \u00C2\u00B11 matrices: singularity and determinant. Random Structures Algorithms, 28(1):1\u00E2\u0080\u009323, 2006. [82] Terence Tao and Van Vu. Random matrices: the circular law. Commun. Contemp. Math., 10(2):261\u00E2\u0080\u0093307, 2008. [83] Terence Tao and Van Vu. Random matrices: universality of ESDs and the circular law. Ann. Probab., 38(5):2023\u00E2\u0080\u00932065, 2010. With an appendix by Manjunath Krishnapur. [84] J. G. van der Corput. U\u00CC\u0088ber Summen von Primzahlen und Primzahlquadraten. Math. Ann., 116(1):1\u00E2\u0080\u009350, 1939. [85] I.M. Vinogradov. Some theorems concerning the theory of primes. Rec. Math. Moscou, n. Ser., 2:179\u00E2\u0080\u0093195, 1937. [86] Erick B. Wong. Simultaneous approximation of reals by values of arith- metic functions. In Anatomy of Integers, volume 46 of CRM Proc. Lecture Notes, pages 289\u00E2\u0080\u0093297. Amer. Math. Soc., Providence, RI, 2008. [87] Binbin Zhou. The Chen primes contain arbitrarily long arithmetic pro- gressions. Acta Arith., 138(4):301\u00E2\u0080\u0093315, 2009. 103"@en . "Thesis/Dissertation"@en . "2012-11"@en . "10.14288/1.0072975"@en . "eng"@en . "Mathematics"@en . "Vancouver : University of British Columbia Library"@en . "University of British Columbia"@en . "Attribution-NonCommercial-NoDerivatives 4.0 International"@en . "http://creativecommons.org/licenses/by-nc-nd/4.0/"@en . "Graduate"@en . "Structure and randomness in arithmetic settings"@en . "Text"@en . "http://hdl.handle.net/2429/42887"@en .