History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary How to Not Count Primes An Analytic Prime Counting Function D.J. Platt Department of Mathematics University of Bristol 28 February 2011 / TCC Number Theory Day D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Outline 1 A Brief History 2 Riemann and π(x) 3 Lagarias, Odlyzko and Galway 4 Our Approach 5 So What Do We Need? D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Sieve of Eratostenes Enumerate and Count for(i=2;i<x;i++) prime[i]=true; D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Sieve of Eratostenes Enumerate and Count for(i=2;i<x;i++) prime[i]=true; for(i=2;i<sqrt(x);i++) if(prime[i]) for(j=i*i;j<x;j+=i) prime[i]=false; D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Sieve of Eratostenes Enumerate and Count for(i=2;i<x;i++) prime[i]=true; for(i=2;i<sqrt(x);i++) if(prime[i]) for(j=i*i;j<x;j+=i) prime[i]=false; O(x log(x) log(log(x))) in time D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Sieve of Eratostenes Enumerate and Count for(i=2;i<x;i++) prime[i]=true; for(i=2;i<sqrt(x);i++) if(prime[i]) for(j=i*i;j<x;j+=i) prime[i]=false; O(x log(x) log(log(x))) in time O(x) in space D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Other Sieves? “Better” sieves are known D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Other Sieves? “Better” sieves are known BUT D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Other Sieves? “Better” sieves are known BUT By the Prime Number Theorem.... D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Other Sieves? “Better” sieves are known BUT By the Prime Number Theorem.... x All sieves must be at least O log(x) in time D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Combinatoric Methods √ π(x) = x − 1 + π( x) . . . D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Combinatoric Methods √ π(x) = x − 1 + π( x) . . . − x2 . . . D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Combinatoric Methods √ π(x) = x − 1 + π( x) . . . − x2 . . . − x3 + x6 . . . D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Combinatoric Methods √ π(x) = x − 1 + π( x) . . . − x2 . . . − x3 + x6 . . . x x x − x5 + 10 + 15 − 30 . . . D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Combinatoric/Sieve Methods Meissel/Lehmer/Lagarias/Miller/Odlyzko/Deléglise/Rivat Use various sieve “tricks” to reduce search tree D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Combinatoric/Sieve Methods Meissel/Lehmer/Lagarias/Miller/Odlyzko/Deléglise/Rivat Use varioussieve “tricks” to reduce search tree 2 Is about O x3 log2 (x) in time D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Combinatoric/Sieve Methods Meissel/Lehmer/Lagarias/Miller/Odlyzko/Deléglise/Rivat Use varioussieve “tricks” to reduce search tree 2 Is about O x3 log2 (x) in time Has been used to compute π 1023 D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Combinatoric/Sieve Methods Meissel/Lehmer/Lagarias/Miller/Odlyzko/Deléglise/Rivat Use varioussieve “tricks” to reduce search tree 2 Is about O x3 log2 (x) in time Has been used to compute π 1023 BUT... D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary The End Of The Road? Parallelised version of M/L/L/M/O/D/R method by Gourdon ran for 360 days D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary The End Of The Road? Parallelised version of M/L/L/M/O/D/R method by Gourdon ran for 360 days Was ±1 out in global checks D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary The End Of The Road? Parallelised version of M/L/L/M/O/D/R method by Gourdon ran for 360 days Was ±1 out in global checks Project was abandoned D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary The End Of The Road? Parallelised version of M/L/L/M/O/D/R method by Gourdon ran for 360 days Was ±1 out in global checks Project was abandoned 6 years later in 2007 Oliveira e Silva recomputed it correctly. D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary The End Of The Road? Parallelised version of M/L/L/M/O/D/R method by Gourdon ran for 360 days Was ±1 out in global checks Project was abandoned 6 years later in 2007 Oliveira e Silva recomputed it correctly. Have such methods reached their limit? D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary π ∗ (x) " π ∗ (x) := 1 2 # P pn <x 1 n + P pn ≤x D.J. Platt 1 n π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary π ∗ (x) " π ∗ (x) := 1 2 # P pn <x 1 n + P pn ≤x 1 n π(x) is cheap to recover from π ∗ (x) (Möbius inversion) D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary π ∗ (x) " π ∗ (x) := 1 2 # P pn <x 1 n + P pn ≤x 1 n π(x) is cheap to recover from π ∗ (x) (Möbius inversion) R∞ Riemann showed given <s > 1, log sζ(s) = π ∗ (x)x −s dx x 0 D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary π ∗ (x) " π ∗ (x) := 1 2 # P pn <x 1 n + P pn ≤x 1 n π(x) is cheap to recover from π ∗ (x) (Möbius inversion) R∞ Riemann showed given <s > 1, log sζ(s) = π ∗ (x)x −s dx x 0 By Mellin inversion π ∗ (x) = D.J. Platt 1 2πi σ+i∞ R σ−i∞ π(x) log ζ(s)x s ds s (σ > 1) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary π ∗ (x) " π ∗ (x) := 1 2 # P pn <x 1 n + P pn ≤x 1 n π(x) is cheap to recover from π ∗ (x) (Möbius inversion) R∞ Riemann showed given <s > 1, log sζ(s) = π ∗ (x)x −s dx x 0 By Mellin inversion π ∗ (x) = 1 2πi (after some analysis) = Li(x) − σ+i∞ R σ−i∞ P ρ D.J. Platt π(x) log ζ(s)x s ds s (σ > 1) Li (x ρ ) − log 2 + O x −1 History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Forcing Convergence L.-O. suggested introducing a suitable Mellin transform pair ˆ φ(t) and φ(s) so that D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Forcing Convergence L.-O. suggested introducing a suitable Mellin transform pair ˆ φ(t) and φ(s) so that σ+i∞ R 1 ˆ log ζ(s)ds + P 1 [χx (pm ) − φ (pm )] π ∗ (x) = 2πi φ(s) m pm σ−i∞ D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Forcing Convergence L.-O. suggested introducing a suitable Mellin transform pair ˆ φ(t) and φ(s) so that σ+i∞ R 1 ˆ log ζ(s)ds + P 1 [χx (pm ) − φ (pm )] π ∗ (x) = 2πi φ(s) m pm σ−i∞ t <x 1 1/2 t = x where χx (t) := 0 t >x D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Forcing Convergence L.-O. suggested introducing a suitable Mellin transform pair ˆ φ(t) and φ(s) so that σ+i∞ R 1 ˆ log ζ(s)ds + P 1 [χx (pm ) − φ (pm )] π ∗ (x) = 2πi φ(s) m pm σ−i∞ t <x 1 1/2 t = x where χx (t) := 0 t >x 1 This algorithm is potentially O x 2 in time. D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Forcing Convergence L.-O. suggested introducing a suitable Mellin transform pair ˆ φ(t) and φ(s) so that σ+i∞ R 1 ˆ log ζ(s)ds + P 1 [χx (pm ) − φ (pm )] π ∗ (x) = 2πi φ(s) m pm σ−i∞ t <x 1 1/2 t = x where χx (t) := 0 t >x 1 This algorithm is potentially O x 2 in time. ˆ Note φ(s) = Riemann. xs s gives φ(t) = χx (t) and we are back to D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary J. Buethe, J. Franke, A. Jost, T. Kleinjung They have developed a version of this algorithm, but based on Weil’s explict formula. D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary J. Buethe, J. Franke, A. Jost, T. Kleinjung They have developed a version of this algorithm, but based on Weil’s explict formula. It is contingent on R.H. (and no I don’t know why). D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary J. Buethe, J. Franke, A. Jost, T. Kleinjung They have developed a version of this algorithm, but based on Weil’s explict formula. It is contingent on R.H. (and no I don’t know why). They have computed a value for π 1024 . D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Galway ˆ G. suggested φ(s) = exp D.J. Platt λ2 s2 2 π(x) xs s , φ(t) = 12 erfc log( xt ) √ (2)λ History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Galway ˆ G. suggested φ(s) = exp λ2 s2 2 xs s , φ(t) = 12 erfc log( xt ) √ Self-duality of Gaussian implies these are “optimal” D.J. Platt π(x) (2)λ History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Galway ˆ G. suggested φ(s) = exp λ2 s2 2 xs s , φ(t) = 12 erfc log( xt ) √ (2)λ Self-duality of Gaussian implies these are “optimal” Now everything converges absolutely so just compute D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Galway ˆ G. suggested φ(s) = exp λ2 s2 2 xs s , φ(t) = 12 erfc log( xt ) √ (2)λ Self-duality of Gaussian implies these are “optimal” Now everything converges absolutely so just compute σ+i∞ R P1 1 m m ˆ log ζ(s)ds + π ∗ (x) = 2πi φ(s) m [χx (p ) − φ (p )] pm σ−i∞ D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Computing The Sum The sum is a prime sieve, centred on x. D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Computing The Sum The sum is a prime sieve, centred on x. At (e.g.) x = 1024 it needs to be 1016 or so wide. D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Computing The Sum The sum is a prime sieve, centred on x. At (e.g.) x = 1024 it needs to be 1016 or so wide. This is (just) achievable. D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Computing The Integral Numerical integration is problematic. D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Computing The Integral Numerical integration is problematic. So let G(s) be a primitive of φˆ such that G(2 + i∞) = −G(2 − i∞) D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Computing The Integral Numerical integration is problematic. So let G(s) be a primitive of φˆ such that G(2 + i∞) = −G(2 − i∞) Then the Pintegral becomes G(1) − G(ρ) − log2 + O x −1 ρ D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Computing The Integral Numerical integration is problematic. So let G(s) be a primitive of φˆ such that G(2 + i∞) = −G(2 − i∞) Then the Pintegral becomes G(1) − G(ρ) − log2 + O x −1 ρ (Note similarity to Riemann’s explicit formula.) D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Computing G Start with ˆ 0 + ih) = φ(s ˆ 0 ) exp(ih(s0 λ2 + log(x))) φ(s exp −λ2 h2 2 1+ sih 0 D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Computing G Start with ˆ 0 + ih) = φ(s ˆ 0 ) exp(ih(s0 λ2 + log(x))) φ(s exp −λ2 h2 2 1+ sih 0 Write everything as a Taylor series D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Computing G Start with ˆ 0 + ih) = φ(s ˆ 0 ) exp(ih(s0 λ2 + log(x))) φ(s exp −λ2 h2 2 1+ sih 0 Write everything as a Taylor series Integrate term by term D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary We Need... A way of computing approximations to real numbers rigorously. D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary We Need... A way of computing approximations to real numbers rigorously. An efficient prime sieve D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary We Need... A way of computing approximations to real numbers rigorously. An efficient prime sieve Accurate and precise zeros of ζ (maybe 1011 of them) D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Rigorous Computations Computers are finite, the real number system isn’t. D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Rigorous Computations Computers are finite, the real number system isn’t. Rounding and truncation must be managed. D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Rigorous Computations Computers are finite, the real number system isn’t. Rounding and truncation must be managed. We use multiple precision interval arithmetic (MPFI by Revol and Rouillier) D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Rigorous Computations Computers are finite, the real number system isn’t. Rounding and truncation must be managed. We use multiple precision interval arithmetic (MPFI by Revol and Rouillier) This is one or two orders of magnitude slower than hardware. D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary An Efficient Prime Sieve Our sieve contains too many primes to compute φ(p) rigorously each time, so D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary An Efficient Prime Sieve Our sieve contains too many primes to compute φ(p) rigorously each time, so We compute a quadratic approximation to φ around some t0 . D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary An Efficient Prime Sieve Our sieve contains too many primes to compute φ(p) rigorously each time, so We compute a quadratic approximation to φ around some t0 . 1, t and t 2 form a basis for this 3 term approximation D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary An Efficient Prime Sieve Our sieve contains too many primes to compute φ(p) rigorously each time, so We compute a quadratic approximation to φ around some t0 . 1, t and t 2 form a basis for this 3 term approximation We sieve intervals I about 232 wide centred on t0 and compute the integers D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary An Efficient Prime Sieve Our sieve contains too many primes to compute φ(p) rigorously each time, so We compute a quadratic approximation to φ around some t0 . 1, t and t 2 form a basis for this 3 term approximation We sieve intervals I about 232 wide centred on t0 and compute the integers P 1 p∈I D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary An Efficient Prime Sieve Our sieve contains too many primes to compute φ(p) rigorously each time, so We compute a quadratic approximation to φ around some t0 . 1, t and t 2 form a basis for this 3 term approximation We sieve intervals I about 232 wide centred on t0 and compute the integers P p∈I P 1 (p − t0 ) p∈I D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary An Efficient Prime Sieve Our sieve contains too many primes to compute φ(p) rigorously each time, so We compute a quadratic approximation to φ around some t0 . 1, t and t 2 form a basis for this 3 term approximation We sieve intervals I about 232 wide centred on t0 and compute the integers P p∈I P 1 (p − t0 ) p∈I P (p − t0 )2 p∈I D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Zeros of ζ We have developed a rigorous, FFT based algorithm for computing ζ on the 12 line. D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Zeros of ζ We have developed a rigorous, FFT based algorithm for computing ζ on the 12 line. It computes many evenly spaced values of ζ in average time O (t ). D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Zeros of ζ We have developed a rigorous, FFT based algorithm for computing ζ on the 12 line. It computes many evenly spaced values of ζ in average time O (t ). We can interpolate using these values to locate the zeros of ζ. D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Zeros of ζ We have developed a rigorous, FFT based algorithm for computing ζ on the 12 line. It computes many evenly spaced values of ζ in average time O (t ). We can interpolate using these values to locate the zeros of ζ. We use Turing’s method to confirm that no zeros have been missed. D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Results We have tested the basic algorithm using Odlyzko’s first 100, 000 zeros to compute π 1011 . D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Results We have tested the basic algorithm using Odlyzko’s first 100, 000 zeros to compute π 1011 . Using UoB’s Bluecrystal cluster, we are computing enough 22 zeros to reach π 10 . D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Summary We have a working version of the Lagarias and Odlyzko analytic π(x) algorithm. D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Summary We have a working version of the Lagarias and Odlyzko analytic π(x) algorithm. It might be able to reach π 1024 . D.J. Platt π(x) History Riemann and π(x) Lagarias, Odlyzko and Galway Our Approach So What Do We Need? Summary Summary We have a working version of the Lagarias and Odlyzko analytic π(x) algorithm. It might be able to reach π 1024 . As a spin off, we will have lots of accurate and precise zeros of ζ to give away. D.J. Platt π(x)
© Copyright 2024