home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!ogicse!das-news.harvard.edu!husc-news.harvard.edu!ramanujan!elkies
- From: elkies@ramanujan.harvard.edu (Noam Elkies)
- Newsgroups: sci.math
- Subject: Re: Distribution of primes mod 4
- Message-ID: <1993Jan24.123639.19744@husc3.harvard.edu>
- Date: 24 Jan 93 17:36:37 GMT
- Article-I.D.: husc3.1993Jan24.123639.19744
- References: <ARA.93Jan21081239@camelot.ai.mit.edu> <1993Jan21.141800.17997@linus.mitre.org> <1993Jan21.212120.251@leland.Stanford.EDU>
- Organization: Harvard Math Department
- Lines: 55
- Nntp-Posting-Host: ramanujan.harvard.edu
-
- In article <1993Jan21.212120.251@leland.Stanford.EDU>
- ilan@leland.Stanford.EDU (ilan vardi) writes:
- :In article <1993Jan21.141800.17997@linus.mitre.org>
- :bs@gauss.mitre.org (Robert D. Silverman) writes:
- :>
- :>let u(x) = #{n <= x; pi(n,1,4) < pi(n,3,4)}
- :>
- :>Then one would expect that u(x) = x/2 for almost all x. That is to say,
- :>for large x, about 1/2 the integers less than x have pi(n,1,4) < pi(n,3,4)
- :>and for about 1/2 the integers the inequality is reversed. This can be
- :>made more precise;
- :>
- :>u(x) = x/2 + O(x^{1-epsilon}) for any fixed epsilon.
- :
- :
- :Yo! This is clearly wrong, and you can't do better than epsilon = 1/2
- :(which is the generalized Riemann Hypothesis). In other words
- :epsilon <= 1/2 in the above term.
-
- Careful here; this is different from the usual question of the number
- of primes <x in a given congruence class. In fact, even under the
- optimal assumption that the extended Riemann Hypothesis holds for
- the L-function L(s) = 1 - 1/3^s + 1/5^s - 1/7^s - + ... and that
- the imaginary parts of the zeros in the upper half of the critical
- strip 0<Re(s)<1 are linearly independent over Q, one does not expect
- the set { n : pi(n;1,4) < pi(n;3,4) } to have a natural density, only
- a logarithmic one; and, more remarkably, that density is strictly
- between 1/2 and 1! (Thanks to Hugh Montgomery for clearing up
- some of my confusion in this regard.)
-
- To see how this happens, let \chi be the character mod 4 (i.e.
- \chi(m)=0 if m even, +-1 if m=4k+-1), and let \Lambda be the
- von Mangoldt function: \Lambda(p^k)=\log(p) for p prime and k>=1,
- \Lambda(m)=0 if m is not a prime power. Then one finds that
- sum(\chi(m)\Lambda(m), m=1..x) is approximated by the sum of
- x^rho/rho with rho running over the nontrivial zeros of L(s);
- under our assumption on these rhos, this is sqrt(x) times an
- almost periodic function of log(x) with zero mean, finite mean square
- (this since the sum of 1/|rho|^2 converges) but attaining arbitrarily
- large values (because sum(1/|rho|) diverges). Now in the Prime
- Number Theorem, when we pass from the sum of \Lambda(m) to the
- count of primes <x the first thing we do is ignore the contribution
- of all prime powers higher than the first, which contribute
- "negligibly" to the sum. However in our case this "negligible"
- contribution is asymptotically sqrt(x) [note that chi(p^2) is
- always 1 for p>2!]. Thus we find that pi(x;1,4)-pi(x;3,4)
- is approximated by sqrt(x)/log(x) times an almost periodic
- function of log(x) whose mean is -1! Hence the estimate on
- the set of x for which it is negative.
-
- --Noam D. Elkies (elkies@zariski.harvard.edu)
- Dept. of Mathematics, Harvard University
-
- P.S. Thanks to Paul C. Leyland for the references to Hans Riesel's
- book and to Bays and Hudson's computational work.
-