1: \documentclass[11pt]{crmp-l}
2: \usepackage{amsmath,amsfonts,amsthm,amssymb,amscd,url}
3: \def\classification#1{\def\@class{#1}}
4: \classification{\null}
5: \newcommand{\ie}{i.e.,}
6: \newcommand{\st}{s.t.}
7: \newcommand{\quadrec}[2]{\left(\frac{#1}{#2}\right)}
8: \newcommand{\njac}[2]{\lbrack {#1},{#2}\rbrack}
9: \renewcommand{\labelenumi}{(\alph{enumi})}
10: \renewcommand{\theenumi}{\alph{enumi}}
11: \textwidth 15cm
12: \setlength{\leftmargin}{-.5in}
13: %\setlength{\textheight}{9in}
14: \DeclareFontFamily{OT1}{rsfs}{}
15: \DeclareFontShape{OT1}{rsfs}{n}{it}{<-> rsfs10}{}
16: \DeclareMathAlphabet{\mathscr}{OT1}{rsfs}{n}{it}
17: \DeclareMathOperator{\sgn}{sgn}
18: \DeclareMathOperator{\card}{card}
19: \DeclareMathOperator{\ord}{ord}
20: \DeclareMathOperator{\av}{av}
21: \newcommand{\W}{\mathcal{W}}
22: \DeclareMathOperator{\ar}{ar}
23: \newcommand{\X}{\mathbf{P}}
24: \DeclareMathOperator{\num}{num}
25: \DeclareMathOperator{\den}{den}
26: \DeclareMathOperator{\lcm}{lcm}
27: \DeclareMathOperator{\Lcm}{Lcm}
28: \DeclareMathOperator{\rad}{rad}
29: \DeclareMathOperator{\mo}{\,mod}
30: \DeclareMathOperator{\sq}{sq}
31: \DeclareMathOperator{\sm}{sm}
32: \DeclareMathOperator{\vol}{vol}
33: \DeclareMathOperator{\GL}{GL}
34: \DeclareMathOperator{\diam}{diam}
35: \DeclareMathOperator{\SL}{SL}
36: \DeclareMathOperator{\PGL}{PGL}
37: \DeclareMathOperator{\im}{im}
38: \DeclareMathOperator{\mix}{mix}
39: \DeclareMathOperator{\Area}{Area}
40: \DeclareMathOperator{\Dom}{Dom}
41: \DeclareMathOperator{\Vol}{Vol}
42: \DeclareMathOperator{\Res}{Res}
43: \DeclareMathOperator{\Disc}{Disc}
44: \DeclareMathOperator{\Var}{Var}
45: \DeclareMathOperator{\Sp}{Sp}
46: \newcommand{\Proj}{\mathbb{P}}
47: \DeclareMathOperator{\charac}{char}
48: \DeclareMathOperator{\irr}{irr}
49: \DeclareMathOperator{\rank}{rank}
50: \DeclareMathOperator{\BSD}{\mathit{BSD}}
51: \DeclareMathOperator{\err}{err}
52: \DeclareMathOperator{\Frob}{Frob}
53: \DeclareMathOperator{\Prob}{Prob}
54: \DeclareMathOperator{\Gal}{Gal}
55: \DeclareMathOperator{\sqf}{sqf}
56: \DeclareMathOperator{\Tr}{Tr}
57: \DeclareMathOperator{\tr}{tr}
58: \DeclareMathOperator{\cl}{cl}
59: \DeclareMathOperator{\Cl}{Cl}
60: \DeclareMathOperator{\disc}{Disc}
61: \DeclareMathOperator{\id}{id}
62: \DeclareMathOperator{\entropy}{entropy}
63: \DeclareMathOperator{\len}{len}
64: \DeclareMathOperator{\comm}{comm}
65: \newcommand{\Ann}{\mathscr{A}}
66: \newcommand{\Q}{\mathbb{Q}}
67: \newcommand{\pr}{\mathfrak{p}}
68: \newcommand{\Z}{\mathbb{Z}}
69: \newcommand{\C}{\mathbb{C}}
70: \newcommand{\R}{\mathbb{R}}
71: %\newcommand{\big}{\mathrm{big}}
72: \newcommand{\Hom}{\mathrm{Hom}}
73: \newcommand{\ideal}{\mathscr{I}}
74: \newcommand{\Norm}{N}
75: \newcommand{\Db}{D_{\mathrm{big}}}
76: \newcommand{\Ds}{D_{\mathrm{small}}}
77: \newcommand{\order}{\mathscr{O}}
78: \newcommand{\pK}{\mathfrak{p}}
79: \newcommand{\pL}{\mathfrak{q}}
80: %\newtheorem{cor}{Corollary}
81: \newtheorem{prop}{Proposition}[section]
82: \newtheorem{thm}[prop]{Theorem}
83: \newtheorem*{main}{Main Theorem}
84: \newtheorem*{kprop}{Key Proposition}
85: \newtheorem*{conj}{Conjecture}
86: \newtheorem{cor}[prop]{Corollary}
87: \newtheorem{lem}[prop]{Lemma}
88: \newtheorem{defn}{Definition}
89: \newcommand{\diff}{\mathfrak{d}}
90: \newenvironment{Rem}{{\bf Remark.}}{}
91: \title[Power-free values]{Power-free values, repulsion between points, differing beliefs and the
92: existence of error}
93: \author{Harald Andr\'es Helfgott}
94: \address{H. A. Helfgott, Mathematics Department, University of Bristol,
95: Bristol, BS8 1TW, United Kingdom}
96: \email{h.andres.helfgott@bristol.ac.uk}
97: \begin{document}
98: \begin{abstract}
99: Let $f$ be a cubic polynomial. Then there are infinitely many primes $p$
100: such that $f(p)$ is square-free.
101: \end{abstract}
102: \maketitle
103: An integer $n$ is said to be {\em square-free} if it is not divisible by
104: any squares other than $1$. More generally, $n$ is {\em free of $k$th powers}
105: if $d\in \mathbb{Z}, d^k|n \Rightarrow d = \pm 1$; square-freeness is what
106: we get in the case $k=2$. Being square-free -- or
107: at least free of $k$th powers for some $k$ --
108: is a desirable property: many things are
109: easier to prove for square-free numbers. Thus, given a set of integers,
110: it is good to know whether infinitely many of them -- or a positive proportion
111: of them -- are square-free\footnote{If a technique is strong enough to
112: prove that infinitely many numbers in the bag are square-free,
113: it is generally also strong enough to show that a positive proportion are,
114: and even to show which proportion are divisible by certain specific squares
115: and no others. This is certainly the case for all the techniques discussed
116: here. Results of this strength are necessary
117: for applications in which, for example, we need to go from the
118: discriminant of an
119: elliptic curve to its conductor, which is essentially the product of the
120: prime factors of the discriminant. See \cite{Hsq} for some general
121: machinery.}.
122:
123: Let $f$ be a polynomial with integer coefficients. Is there an infinite
124: number of integers $n$ such that $f(n)$ is free of $k$th powers?
125: There are some polynomials for which the answer is clearly ``no'': say
126: $k=2$ and $f(n) = 4 n$ or $f(n) = n^2$. Slightly more subtly, consider
127: $f(n) = n (n+1) (n+2) (n+3) + 4$, which is always divisible by $4$.
128: Assume, then, that $f$ has no factors repeated $k$ times and that
129: the following local condition holds:
130: \begin{quotation}
131: (*)\;\;\; for every prime $p$, $f(x) \not\equiv 0 \mo p^k$ has at least one solution in
132: $\mathbb{Z}/p^k \mathbb{Z}$.
133: \end{quotation}
134: (Both conditions are obviously necessary, and
135: both can be checked easily in bounded time.) Then, it is believed,
136: there must be an infinite number of integers $n$ such that $f(n)$ is
137: free of $k$th powers.
138:
139: If $\deg(f) \leq k$, it is not hard to prove as much. If $\deg(f)>k+1$,
140: the problem is too hard, at least when $k$ is small. For $\deg(f)=k+1$,
141: the statement was proven by Erd\H{o}s in 1953. In particular, if
142: $f$ is a cubic polynomial without repeated factors and $f$ satisfies
143: the local condition (*), then there are infinitely many integers $n$ such
144: that $f(n)$ is square-free.
145:
146: Like many proofs in analytic number theory, Erd\H{o}s's proof is rather
147: tricky, in that it uses the fact that most integers are not prime in order
148: to avoid certain essential issues. Perhaps because of this, Erd\H{o}s asked:
149: for $f$ cubic,
150: are there infinitely many primes $p$ such that $f(p)$ is square-free?
151: (More generally, for $f$ of degree $k+1$, are there infinitely many
152: primes $p$ such that $f(p)$ is free of $k$th powers?) He conjectured that
153: there are, but, as might be expected, most tricks used before break down.
154:
155: Hooley (\cite{Ho}, \cite{Ho2}) proved Erd\H{o}s's conjecture
156: for $k\geq 51$. At about the same time, Nair \cite{Na} proved
157: the conjecture for $k\geq 7$, using rather different methods. (He was also
158: the first to treat polynomials with $\deg(f) \geq k+2$, $k$ large;
159: Heath-Brown (\cite{HB}) has attained further progress in this front.)
160: In both approaches, $k$ small is harder than $k$ large; in particular,
161: the case $k\leq 6$ remained open.
162:
163: In \cite{He}, I proved Erd\H{o}s's conjecture for all polynomials $f$
164: with {\em high entropy}; in particular, the proof works when $k=2$,
165: $\deg(f)=3$ and $\Gal(f) = A_3$. However, most cubic polynomials have
166: Galois group $S_3$, and their case remained open until now.
167:
168: I have now managed to prove Erd\H{o}s's conjecture for general cubics.
169: \begin{main}
170: Let $f\in \mathbb{Z}\lbrack x\rbrack$ be a cubic polynomial without repeated
171: roots. Assume that, for every prime $q$, there is a solution
172: $x\in (\mathbb{Z}/q^2 \mathbb{Z})^*$ to $f(x) \not\equiv 0 \mo q^2$.
173: Then there are infinitely many primes $p$ such that $f(p)$ is square-free.
174: \end{main}
175: In fact, $f(p)$ is square-free for a positive proportion $C_f$ of all primes $p$
176: -- where $C_f$ is exactly what we would expect (an infinite
177: product of local densities).
178:
179: The tools used, developed and sharpened in the proof are mostly from
180: diophantine geometry and probabilistic number theory; there is a key use of
181: the modularity of elliptic curves. Let us take a quick walk through the proof.
182: (A full account shall appear elsewhere.)
183:
184: It is not hard to show that
185: \begin{equation}\label{eq:ogto}\begin{aligned}
186: |\{p\leq N:\; &\text{$f(p)$ is square-free}\}| = C_f \frac{N}{\
187: \log N} \cdot (1 + o(1)) \\ &+ O(|\{x,y\leq N, d\leq N (\log N)^{\epsilon} : \text{$x$, $y$ prime},\;
188: d y^2 = f(x)\}|),\end{aligned}\end{equation}
189: where $|S|$ is the number of elements of a set $S$ and
190: $C_f$ is the (non-zero) constant we would expect. This (well-known) initial
191: step goes roughly as follows.
192: Small square factors
193: can be sieved out because we know how many primes there are in arithmetic
194: progressions to small moduli; medium-sized square factors amount to a small
195: error term, since $\sum_{d>m} N/d^2$ is quite small with respect to $N$
196: as soon as $m$ is
197: moderately large. Large square factors cannot be brushed aside by the same
198: argument as medium-sized square factors
199: simply because there are so many of them:
200: an additional term that is overshadowed by $N/d^2$ in the medium range
201: comes to the fore here. This is why the contribution of the
202: large square factors figures in (\ref{eq:ogto}) as the error
203: term within $O(\cdot)$.
204: The sole problem from now on, then, is to show that the expression within $O(\cdot)$
205: is $o(N/\log N)$.
206:
207: As you can tell from the notation, we intend to see this as a problem of
208: bounding the number of integer points $(x,y)$ on curves $d y^2 = f(x)$,
209: $f$ a fixed cubic polynomial. The issues are two. First, we need very good bounds -- almost
210: as good as $O(1)$ for the number of points for each $d$, or at least for
211: every typical $d$. Second, even a bound of $O(1)$ would not be enough!
212: There are $N (\log N)^{\epsilon}$ curves to consider, and a bound of $O(1)$
213: per curve would amount to a total bound of $O(N (\log N)^{\epsilon})$, whereas
214: we need $o(N/\log N)$.
215:
216: Let us begin with the first issue: we want good bounds on the number of
217: integer points $(x,y)$, $x,y\leq N$, on the curve $C_d$ described by
218: $d y^2 = f(x)$, $d$ fixed. Most techniques for bounding the number of
219: points on curves are based on some kind of {\em repulsion}: if there
220: are bees in a room, and each bee stays a yard or more away from every other
221: bee, there cannot be too many bees in the room. Repulsion may happen in
222: the visible geometry of the curve, viz., its graph, as in \cite{BP}; such
223: a perspective, unfortunately, would not be nearly enough in our case.
224: Alternatively, we may look at repulsion in the Mordell-Weil lattice
225: corresponding to the curve.
226:
227: {\em Rational and integer points on curves.}
228: Let $C$ be a curve of genus $g>0$
229: over $\mathbb{Q}$ (or a number field $K$). The curve
230: $C$ can be embedded in its Jacobian $J_C$. The rational points
231: $J_C(K)$ in the Jacobian form a finitely generated abelian group
232: under the group law of the Jacobian; they are, furthermore, endowed with
233: a natural norm given by the square root of the canonical height. Hence
234: $J_C(K)$ can be naturally embedded in $\mathbb{R}^r$, where
235: $r$ is the rank of $J_C(K)$. We thus have an (almost) injective map
236: \[\iota:C(K) \to L \subset \mathbb{R}^r,\]
237: where $C(K)$ is the set of rational points on $C$ and $L$
238: is a lattice in $\mathbb{R}^r$. What can be said about
239: the image $\iota(C(K))$?
240:
241: If the genus $g$ is $1$, $\iota(C(K))$ is all of $L$. However,
242: if $g>1$, then $\iota(C(K))$ looks quite sparse within $L$.
243: Mumford \cite{Mu} proved that the points of $\iota(C(K))$ repel each
244: other: for any $P_1,P_2\in \iota(C(K))$ at about the same distance
245: from the origin $O$, the angle $\angle P_1 O P_2$ separating $P_1$ from $P_2$
246: is at least $60^{\circ}$ (for $g=2$), $70.5^{\circ}$ (for $g=3$),
247: $75.5^{\circ}$ (for $g=4$), \dots -- in general,
248: $\angle P_1 O P_2 \geq \arccos \frac{1}{g}$.
249:
250: Assume now that the points $P_1$ and $P_2$ come from {\em integer} points
251: on $C$. Then, as I pointed out in my thesis (\cite[Ch.\ 4]{Heth} or
252: \cite[Lem.\ 4.16]{Hesq}; see also the earlier work of Silverman
253: \cite{Si}, \cite[Prop.\ 5]{GS}, which seems to have originated from
254: the same observation) the angle $\angle P_1 O P_2$ is larger
255: than if $P_1$ and $P_2$ were merely rational: the angle is at least
256: $60^{\circ}$ for $g=1$, $75.5^{\circ}$ for $g=2$, \dots -- in general, at least
257: $\arccos \frac{1}{2 g}$. We obtain better bounds as a consequence. (If
258: $g=1$, we obtain bounds where Mumford's work does not by itself give any.)
259: The bounds are obtained by means of results on sphere-packing; indeed, the number of points fitting at a certain distance from
260: the origin and
261: at a separation of $\geq 60^{\circ}$ from each other is precisely the
262: number of solid spheres that can fit around a sphere of the same size
263: in the given dimension.
264:
265: In the case of the curve $E_d:d y^2 = f(x)$, the bounds we obtain are of
266: the form $c_1^{\omega(d)}$, $c_1>1$ fairly small ($<2$). This is still not good
267: enough, as, on the average,
268: it amounts to about $(\log N)^{c_2}$, $c_2$ a small
269: but fixed non-zero constant,
270: for $d\sim N$; what we would like is a bound of the form $(\log N)^{\epsilon}$.
271:
272: {\em Visible vs.\ Mordell-Weil geometry.} In \cite{HV}, we remark upon the
273: following phenomenon. Let $P_1$, $P_2$ be two integer (or rational) points
274: on $C(K)$ at about the same distance from the origin. Suppose that
275: their coordinates $(x_1,y_1)$, $(x_2,y_2)$ are close to each other, either
276: in the real place (that is, $|x_1-y_1|$ and $|x_2-y_2|$ are small) or
277: $p$-adically (that is, $x_1 \equiv x_2 \mo d$ and $y_1 \equiv y_2 \mo d$
278: for a large integer $d$). Then the angle $\angle P_1 O P_2$ in the
279: Mordell-Weil lattice is even larger
280: than it would already have to be. In other words: if two points are close
281: to each other in the graph of the curve in $\mathbb{R}^2$, they must
282: be especially far from each other in the Mordell-Weil lattice.
283: Thus, if
284: we partition the set of all rational points into sets of points close
285: to each other in the graph of the curve, we shall obtain an especially
286: good bound on the number of elements of each such set. The main concern
287: is then to keep the number of such sets small.
288:
289: In the case of the curve $E_d : d y^2 = f(x)$, we have that any two
290: points $(x_1,y_1)$, $(x_2,y_2)$ on $E_d$
291: induce points $P_1 = (x_1, d^{1/2} y_1)$,
292: $P_2 = (x_2, d^{1/2} y_2)$ on $E : y^2 = f(x)$. The $y$-coordinates of
293: $P_1$ and $P_2$ are already close to each other modulo $d$ (that is,
294: modulo the prime ideals in $\mathbb{Q}(d^{1/2})$ dividing $d$). The
295: congruence classes mod $d$ into which $x_1$ and $x_2$ may fall are
296: rather restricted, as $f(x_1) \equiv 0 \mo d$ and $f(x_2) \equiv 0 \mo d$;
297: the total number of congruence classes $x$ modulo $d$ for which
298: $f(x) \equiv 0 \mo d$ is at most $3^{\omega(d)}$. If $P_1$ and $P_2$
299: have $x$-coordinates in the same congruence class modulo $d$, then the
300: angle $\angle P_1 O P_2$ turns out to be at least $90^{\circ}$, or
301: $90^{\circ} - \epsilon$. Very few points can fit in $\mathbb{R}^r$ lying
302: at about the same distance from the origin and subtending angles of
303: $90^{\circ}-\epsilon$ or more from each other.
304:
305: There is the problem that $3^{\omega(d)}$ is too large -- a power of
306: $(\log N)$, on the average, since $\omega(d)$ is usually about $\log \log N$.
307: However, with probability $1$, an integer $d\leq N$ has a large divisor $d_0$
308: ($>N^{1 -
309: \epsilon}$) with few prime divisors ($< \epsilon \log \log N$). We may
310: thus consider points $P_1$, $P_2$ congruent to each other modulo $d_0$,
311: rather than $d$, and obtain angles $\angle P_1 O P_2$ of size at least
312: $90^{\circ}-\epsilon$ while considering at most $3^{\omega(d_0)}$ possible
313: congruence classes. The total bound on the number of integer points
314: $(x,y)$ on $E_d:d y^2 = f(x)$ with $x,y\leq N$ is $O((\log N)^{\epsilon'})$
315: for every typical $d$, that is, for each $d\leq N$ outside a set
316: of size at most $N/(\log N)^{1000}$.
317:
318: This is almost as good as $O(1)$. The problem, as said before, is that
319: this is not good enough; since there are $N$ integers $d=1,2,\dotsc,N$ to consider,
320: the total bound would be $O(N)$. The issue, then, is how to eliminate most
321: $d$. Probabilistic number theory has just made its first appearance;
322: it shall play a crucial role in what follows.
323:
324: {\em The perspective of the value and the perspective of the argument.}
325: Our task is still to show that
326: \begin{equation}\label{eq:malato}
327: |\{x,y\leq N, d\leq N (\log N)^{\epsilon} : \text{$x$, $y$ prime},\;
328: d y^2 = f(x)\}|\end{equation}
329: is at most $o(N/\log N)$. We have a good bound for each $d$, namely,
330: $O((\log N)^{\epsilon})$ for each $d$ outside a small set,
331: and a reasonable bound for each $d$ inside that small set.
332: The idea will now be to consider, in a solution to $d y^2 = f(x)$, what
333: kind of integer $d = f(x)/y^2$ typically is, and whether it looks much like
334: a typical integer $d$. We will show that, for most $x$, the integer
335: $d = f(x)/y^2$ must look rather strange, and that thus there can be
336: few such $d$.
337: Stated otherwise: we shall prove that
338: every prime $x\leq N$ must either lie within a fixed set of size $o(N/\log N)$
339: or be such that, if $d y^2 = f(x)$ for some prime $y$ and some integer
340: $d\leq N (\log N)^{\epsilon}$, then $d$ must lie within a fixed set of size
341: $O(N/(\log N)^{1 + 10 \epsilon})$. Combined with our bound for each $d$,
342: this will yield immediately that (\ref{eq:malato}) is indeed
343: $o(N/\log N)$.
344:
345: What are, then, the ways in which $f(x)$ will tend to be strange for
346: a random prime $x$? And which of those ways of strangeness will carry
347: over to $d$, if $f(x)$ can be written in the form $d q^2$, where $q$
348: is a large prime?
349:
350: As far as the second question is concerned: since $q$ is prime,
351: $f(x)$ and $d$ have almost
352: the same number of prime divisors. Thus, if we can show that
353: the number of prime divisors of $f(x)$ is strange for $x$ random, we will
354: have shown that the number of prime divisors of $d$ is strange for
355: $x$ random.
356:
357: Now, for $x$ random, the number of prime divisors $w(f(x))$ will be
358: about $\log \log N$. Thus, $w(d)$ will also be about $\log \log N$.
359: Unfortunately, this is typical, not strange, for an integer of the size of
360: $d$.
361:
362: Consider, however, primes of different kinds.
363: Let $K = \mathbb{Q}(\alpha)$, where $\alpha$ is a root of $f(x) = 0$.
364: Then some primes $p$ will split completely in $K/\mathbb{Q}$,
365: some primes will not split at all, and some primes may split yet not
366: split completely. We can write $w_1(n)$, $w_2(n)$ and $w_3(n)$ for
367: the number of prime divisors of $n$ of each kind. Then, as we shall
368: see, $w_j(f(x))$ (and thus $w_j(d)$) will tend to be strange for
369: $x$ random.
370:
371: Suppose first that $K/\mathbb{Q}$ has Galois group $A_3$. Then
372: every prime $p$ must either split completely or not split at all.
373: If $p$ does not split at all, then $f(x) \equiv 0 \mo p$ has no
374: solutions. Hence $w_2(f(x)) = 0$, and so $w_2(d)=0$. This is certainly
375: atypical for an integer $d\leq N$. (Usually $w_2(d) \sim \frac{2}{3}
376: \log \log N$.) Now suppose $p$ splits completely. Then
377: $f(x) \equiv 0 \mo p$ has three solutions mod $p$. Thus, for a random
378: prime $x$, we shall have $f(x) \equiv 0 \mo p$ with probability $3/p$.
379: Hence $w_2(f(x))$ will most likely be about $\sum_{\text{$p$ splits
380: completely}} \frac{3}{p} \sim \log \log N$. Thus
381: $w_2(d) \sim \log \log N$, whereas an integer $d\leq N$ usually
382: has $w_2(d) \sim \frac{1}{3} \log \log N$.
383:
384: It is not enough, however, to show that $d$ is strange (i.e., in a set
385: of size $o(N)$); we must show that $d$ is strange enough (i.e., in a set
386: of size $O(N/(\log N)^{1 + \epsilon})$). How odd is it for a random integer
387: $d\leq N$ to have $w_1(d)=0$ and $w_2(d)\sim \log \log N$? The
388: number
389: $w_1(d)$ equals $\sum_{\text{$p$ does not split}} X_p$,
390: where $X_p$ is a random variable
391: taking the value $1$ when $p|d$ and $0$ otherwise.
392: Similarly, $w_2(d) = \sum_{\text{$p$ splits completely}} X_p$.
393: Now $X_p$ is $1$ with probability $1/p$ and $0$ with probability $1-1/p$.
394: Suppose the variables $X_p$ to be mutually independent. Then the
395: theory of large deviations (Cramer's theorem, or, more appropriately,
396: Sanoff's theorem) offers the upper bound
397: \[\Prob\left(\sum_{\text{$p$ does not split}} X_p = 0\; \wedge
398: \sum_{\text{$p$ splits completely}} X_p > (1- \epsilon) \log \log N\right)
399: \ll \frac{1}{(\log N)^{\log 3 - \epsilon'}}.\]
400:
401: Now, of course, the variables $X_p$ are not actually mutually independent;
402: the variables $X_{p_1},X_{p_2},\dotsc,X_{p_k}$ can be assumed to be
403: (approximately) mutually independent only when $p_1 p_2 \dotsb p_k < N$.
404: However, the fact that $X_p$ has only a small probability of being non-zero
405: allows us to use the main
406: technique from Erd\"os and Kac's Gaussian paper \cite{EK}
407: to show that we may treat the variables
408: $X_p$, for our purposes, as if they were mutually independent. Thus we do
409: obtain
410: \[\Prob\left( w_1(d) = 0\; \wedge w_2(d) > (1 - \epsilon) \log \log N\right)
411: \frac{1}{(\log N)^{\log 3 - \epsilon'}} = O(N/(\log N)^{1 + \epsilon''}),\]
412: as desired.
413:
414: We are done proving the main theorem when $\Gal_f = A_3$. What happens
415: when $\Gal_f = S_3$? While our analysis is in the main still valid,
416: the exponent that we obtain instead of $\log 3$ is $\frac{1}{2} \log 3$,
417: which is less than $1$, and thus insufficient. (In general, for
418: $d y^k = f(x)$, $\deg(f) = k+1$, the exponent we get may be expressed
419: as an {\em entropy}, which will depend on $\Gal(f)$ alone. Sometimes
420: $\entropy(\Gal(f))>1$, and we are done, and sometimes, as in the case of
421: $\Gal(f)=S_3$, the entropy is $<1$.)
422:
423: The reason is that, when $\Gal_f = S_3$, half of the primes split in
424: $K/\mathbb{Q}$. These primes divide $f(x)$ with exactly the same probability
425: that they would divide a random integer, and thus they are useless.
426: What is to be done, then? How can one bridge a gap of size
427: $1/(\log N)^{1 - \frac{1}{2} \log 3}$?
428:
429: {\em The existence of error. Modularity.}
430: Again: what is a way of strangeness such that, if $f(x)$ is strange and
431: $d = f(x)/q^2$, $q$ a prime, then $d$ must be strange as well? Having
432: too few or too many prime factors of some kind is one way. Is there
433: another one?
434:
435: We have used the fact that $q^2$ has only one prime factor; let us now
436: use the fact that $q^2$ is a square. For any prime modulus $p$, the
437: integer $d$ will be a square mod $p$ if and only if $f(x)$ is a square
438: mod $p$. Now, a random integer is as likely to be a square mod $p$
439: as a non-square
440: mod $p$. How likely is $f(x)$ to be a square mod $p$ for a random integer $x$
441: (or a random prime $x$)?
442:
443: By the Weil bounds, there are $p + O(\sqrt{p})$ points on the curve
444: $y^2 = f(x) \mod p$. Hence the probability that $f(x)$ will be a square
445: mod $p$
446: is $\frac{1}{2} + O(p^{-1/2})$. This is not good, as $\frac{1}{2}$
447: would be the probability if there were nothing amiss to be exploited.
448: Let us show that an error of size about
449: $p^{-1/2}$ is in fact present a positive proportion of the time.
450:
451: Write the number of points on the curve
452: $y^2 = f(x) \mod p$ as $p + 1 - a_p$. Then the probability that
453: $f(x)$ will be a square $\mo p$ is precisely
454: $\frac{1}{2} - \frac{a_p}{2 p} + O(1/p)$; we have to give a lower bound,
455: on the average, for the size of $|a_p|/2p$ (or, rather,
456: $a_p^2/p^2$, since we shall later use a variance bound).
457: Now, the $L$-function of $E:y^2 = f(x)$ is
458: $\sum a_n n^{-s}$. By the modularity of elliptic curves (proven by Wiles
459: et al.), there is a modular form $g$ associated to $L$. We may, in turn,
460: define a Rankin-Selberg $L$-function $L_{g\otimes g}$
461: associated to $g$, and use the standard facts that
462: $L_{g\otimes g} = \sum a_n^2 n^{-s}$ and that $L_{g\otimes g}$ has a simple
463: pole at
464: $s=2$. By some Tauberian work (or proceeding as in the proof of
465: the prime number theorem)
466: we deduce that
467: $\sum_{p\leq z} |a_p|^2/p^2$ is asymptotic to $\log \log z$; in other words,
468: $a_p^2$ is of size about $p$ on the average.
469:
470: It is somewhat unpleasant to have to use modularity here, as we need not
471: know the behaviour of $L_E$ (or $L_{g\otimes g}$) inside the critical strip.
472: Still, it is hard to see how to do without modularity or some strong kindred
473: result.
474: Marc Hindry and Mladen Dimitrov have pointed out to me that, if one wants
475: to give a (conditional) statement on $k$th-power-free values of polynomials
476: of degree $k+1$, $k>2$, it may be simpler and more proper to work assuming
477: Tate's conjecture on $L_{C\times C}$ rather than automorphicity.
478:
479: {\em Using many small differences. Exponential moments and high moments.}
480: Now, how may we use these small differences between the probability of
481: $d$ being a square (for $d$ a random integer) and the probability of
482: $d$ being a square (for $d = f(x)/q^2$, $x$ a random prime)?
483:
484: Suppose I am throwing a fair coin in the air. A gentle wind blows;
485: it may change directions very often, but becomes gradually milder.
486: I know that the wind has a slight effect on the way the coin lands: if the
487: wind blows from the east, then, I posit, heads are more likely, whereas, if
488: the wind blow from the west, tails are more likely. You, however, will not
489: believe me. How shall I make my point?
490:
491: Let us assume I can measure the
492: strength and direction of the wind before every coin throw. I shall throw the
493: coin in the air many times, betting on heads or tails according to what I
494: reckon to be more likely, given the wind. If, at the end, I have collected
495: statistically significant winnings, you will have to acknowledge that I am in
496: the right.
497:
498: Our situation is analogous. Instead of wind, we have $a_p$; instead of a coin,
499: we have whether or not $f(x)$ is a square mod $p$ for a random prime $x$.
500: (The prime $x$ stays fixed as $p$ varies.) If $f(x)$ (and thus $d$) lands on the more likely
501: side of squareness or non-squareness for significantly more than one-half of
502: all primes $p$, then $d$ will be sufficiently strange.
503:
504: We can let $X_p$ be a random variable taking the value
505: $\frac{-a_p}{p}$ when $f(x)$ is a
506: square mod $p$, and the value $\frac{a_p}{p}$ when $f(x)$ is a non-square mod
507: $p$. Then $X_p$ is $\frac{-a_p}{p}$ with probability $\frac{1}{2} -
508: \frac{a_p}{2 p}$, and $\frac{a_p}{p}$ with probability
509: $\frac{1}{2} + \frac{a_p}{2 p}$. It can be seen easily that
510: the expected value of $\sum_{p\leq z} X_p$ is
511: $\sum_{p\leq z} a_p^2/p^2 \sim \log \log z$.
512: We may assume pairwise independence and obtain that $\Var(\sum_{p\leq z} X_p)
513: = \sum_{p\leq z} \left(\frac{a_p^2}{p^2} - \frac{a_p^4}{p^4}\right)
514: \sim \log \log z - O(1) \sim \log \log z$.
515: Thus, Chebyshev gives us that
516: $\sum_{p\leq z} X_p$ is $> (1- o(1)) \log \log z$ a
517: proportion
518: $1$ of the time. Now let $Y_p$ be $\frac{-a_p}{p}$
519: when a random integer $d\leq x$ is
520: a square residue modulo $p$; let $Y_p$ be $\frac{a_p}{p}$ otherwise. What is the probability that
521: $\sum_{p\leq z} Y_p$ be larger than $(1 - o(1)) \log \log z$?
522:
523: The probability that $Y_p$ take either of its two possible values is $1/2$.
524: Suppose that the variables $Y_p$ were mutually independent. Then the
525: expected value $\mathbb{E}(e^{\sum Y_p})$
526: of $e^{\sum Y_p}$ would be the product of the expected values of
527: $e^{Y_p}$. We can use this as follows. First of all,
528: \[
529: \Prob(\sum_{p\leq z} Y_p > (1 - o(1)) \log \log z) \leq
530: \Prob(e^{\sum_{p\leq z} Y_p} > (\log z)^{1 - o(1)}) \leq
531: \frac{\mathbb{E}(e^{\sum_{p\leq z} Y_p})}{(\log z)^{1 - o(1)}}.\]
532: Now, as we were saying,
533: \[\begin{aligned}
534: \mathbb{E}(e^{\sum_{p\leq z} Y_p}) &=
535: \mathbb{E}(\prod_{p\leq z} e^{Y_p}) =
536: \prod_{p\leq z} \mathbb{E}(e^{Y_p}) =
537: \prod_{p\leq z} \left(\frac{1}{2} e^{\frac{a_p}{p}} + \frac{1}{2}
538: e^{-\frac{a_p}{p}}\right)\\
539: &= \prod_{p\leq z} \left(1 + \frac{1}{2} \frac{a_p^2}{p^2} +
540: \frac{1}{4!} \frac{a_p^2}{p^4} + \dotsc\right) \ll
541: \prod_{p\leq z} e^{\frac{1}{2} \frac{a_p^2}{p^2}}\\
542: &=e^{\sum_{p\leq z} \frac{a_p^2}{p^2}}
543: = e^{(1 + o(1)) \log \log z} = (\log z)^{1 + o(1)} .
544: \end{aligned}
545: \]
546: Hence
547: \begin{equation}\label{eq:homero}
548: \Prob(\sum_{p\leq z} Y_p > (1 - o(1)) \log \log z)
549: \ll
550: \frac{1}{(\log z)^{1/2 - o(1)}} ,\end{equation}
551: which is the bound we desire.
552:
553: Now, the variables $Y_p$ are not in fact mutually independent, and, since the
554: probabilities we are dealing with are close to $1/2$ rather than to $0$, we
555: cannot apply the tricks in Erd\"os-Kac. A simpler approach will in fact do.
556: Let $z = N^{\frac{1}{3 \log \log N}}$ and $k = \frac{1}{2} \log \log z$.
557: Then, while the variables $Y_p$ are not mutually independent, they are more
558: than pairwise independent: any $2k$ of them are mutually independent (with a
559: small error term). We can thus proceed as in the proof of Chebyshev's theorem,
560: taking a $(2k)$th power instead of a square. The bound thus obtained is
561: essentially as good as (\ref{eq:homero}): we obtain that
562: the probability that
563: $\sum_{p\leq z} Y_p/\sqrt{p}$ be larger than $(1 - o(1)) \log \log z$
564: is $O(1/(\log z)^{1/2 - o(1)}) = O(1/(\log N)^{1/2 - o(1)})$.
565:
566: We conclude that, if $d$ is strange in the two ways
567: we have considered -- having
568: numbers of prime divisors differing from the norm, and ``agreeing with
569: the wind'' for considerably more than half of all $p$'s -- then it lies
570: in a set of cardinality at most
571: \[O\left(N \cdot \frac{1}{(\log N)^{\frac{1}{2} \log 3 + \frac{1}{2} -
572: \epsilon}}\right) = O\left(\frac{N}{(\log N)^{1.0493\dotsc- \epsilon}}\right) .\]
573: This is smaller than $o\left(\frac{N}{\log N}\right)$,
574: which was the goal we set ourselves in the discussion after (\ref{eq:malato}).
575: Thus (\ref{eq:malato}) is indeed at most $o(N/\log N)$, and we are done
576: proving the main theorem.
577: \begin{thebibliography}{88}
578: \bibitem{BP}
579: Bombieri, E., and J. Pila, The number of integral points on arcs and ovals,
580: {\em Duke Math. J.} {\bf 59} (1989), no.\ 2, 337--357.
581: \bibitem{EK}
582: Erd\H{o}s, P., and M. Kac, The Gaussian law of errors in the theory
583: of additive number theoretic functions, {\em Amer.\ J.\ Math.} {\bf 62} (1940),
584: 738--742.
585: \bibitem{GS}
586: R. Gross and J. Silverman, $S$-integer points on elliptic curves, {\em
587: Pacific J.\ Math.} {\bf 167} (1995), 263--288.
588: \bibitem{HB}
589: Heath-Brown, R., Counting rational points on algebraic varieties,
590: {\em C. I. M. E. lecture notes,} to appear.
591: \bibitem{He}
592: Helfgott, H. A., Power-free values, large deviations, and integer points
593: on irrational curves, to appear in {\em J. Th\'eor. Nombres Bordeaux}.
594: \bibitem{Heot}
595: Helfgott, H. A., On the behaviour of root numbers in families of elliptic
596: curves, submitted, math.NT/0408141.
597: \bibitem{Hesq}
598: Helfgott, H. A., On the square-free sieve, {\em Acta Arith.} {\bf 115}
599: (2004) 349--402.
600: \bibitem{Heth}
601: Helfgott, H. A., {\em Root numbers and the parity problem}, Ph.D.\ thesis,
602: Princeton Univ., \url{http://www.arxiv.org/abs/math.NT/0305435}.
603: \bibitem{HV}
604: Helfgott, H. A., and A. Venkatesh, Integral points on elliptic curves
605: and $3$-torsion in class groups, {\em J. Amer.\ Math.\ Soc.} {\bf 19}
606: (2006), 527--550.
607: \bibitem{Ho}
608: Hooley, C., On power-free numbers and polynomials. I, {\em J.\ reine angew.\
609: Math.} {\bf 293}/{\bf 294} (1977), 67--85.
610: \bibitem{Ho2}
611: Hooley, C., On power-free numbers and polynomials. II, {\em J.\ reine angew.\
612: Math.} {\bf 295} (1977), 1--21.
613: \bibitem{Hsq}
614: Helfgott, H. A., On the square-free sieve, {\em Acta Arith.} {\bf 115}
615: (2004) 349--402.
616: \bibitem{Na}
617: Nair, M., Power-free values of polynomials, II, {\em Proc.\ London Math.\
618: Soc.\ (3)} {\bf 38} (1979), no.\ 2, 353--368.
619: \bibitem{Mu}
620: Mumford, D., A remark on Mordell's conjecure, {\em Amer. J. Math.} {\bf 87}
621: (1965), 1007--1016.
622: \bibitem{Si}
623: Silverman, J. H., A quantitative version of Siegel's theorem: integral points
624: on elliptic curves and Catalan curves, {\em J. Reine Angew.\ Math.}
625: {\bf 378} (1987), 60--100.
626: \end{thebibliography}
627: \end{document}
628: