Continued Fractions New and Old Results

Transcription

1 Continued Fractions New and Old Results Jeffrey Shallit School of Computer Science University of Waterloo Waterloo, Ontario N2L 3G1 Canada Joint work with Dzmitry Badziahin 1/43

2 Continued fractions By a continued fraction we mean an expression of the form a 0 + a a a n or 1 a a a a n + where a 1,a 2,..., are positive integers and a 0 is an integer. To save space, as usual, we write [a 0,a 1,...,a n ] for the first expression and [a 0,a 1,...,] for the second. 2/43

3 What s known Euler, Lagrange: a real number has an ultimately periodic continued fraction expansion if and only if it is a quadratic irrational Other interesting numbers: e = [2,(1,2n,1) n=1] = [2,1,2,1,1,4,1,1,6,1,1,8,1,...] e 2 = [7,(3n 1,1,1,3n,12n+6) n=1] = [7,2,1,1,3,18,5,1,1,6,30,...] tan1 = [(1,2n 1) n=1] = [1,1,1,3,1,5,1,7,...] These three are examples of Hurwitz continued fractions, where there is a quasiperiod of terms that grow linearly. No simple pattern is currently known for the expansions of π or e 3 or e 4. 3/43

4 Irrationality measure The irrationality measure µ(x) of a real number x is the supremum of the real numbers µ such that the inequality x p q < 1 q µ has infinitely many solutions in rational numbers p/q. So it measures how well x is approximated by rationals. Liouville s theorem: if x is algebraic of degree n then µ(x) n. Roth s theorem: if x is an algebraic irrational number then µ(x) = 2. If the continued fraction for x has bounded partial quotients then µ(x) = 2. 4/43

5 Irrationality measures of some famous numbers µ(e) = 2 µ(π) (Salikhov) µ(π 2 ) (Rhin & Viola) by studying the integral ( x h (1 x) i y k (1 y) j (1 xy) i+j l ) n dxdy 1 xy = a n b n ζ(2), in particular for (h,i,j,k,l) = (12,12,14,14,13). µ(log 2) (Marcovecchio) µ(ζ(3)) (Rhin & Viola) 5/43

6 The usual conventions Given a real irrational number x with partial quotients x = [a 0,a 1,a 2,...] we define the sequence of convergents by p 2 = 0 p 1 = 1 p n = a n p n 1 +p n 2 (n 0) q 2 = 1 q 1 = 0 q n = a n q n 1 +q n 2 (n 0) and then [a 0,a 1,...,a n ] = p n q n. Furthermore, we use the Hurwitz-Kolden-Frame representation of continued fractions via 2 2 matrices, as follows: [ ][ ] [ ] [ a0 1 a1 1 an 1 pn p M(a 0,...,a n ) := = n q n q n 1 ]. 6/43

7 The matrix approach By taking determinants we immediately deduce the classical identity p n q n 1 p n 1 q n = ( 1) n+1 for n 0. Given a finite sequence z = (a 0,...,a n ) we let z R denote the reversed sequence (a n,...,a 0 ). A sequence is a palindrome if z = z R. By taking the transpose of the equation above, we get M(a n,...,a 0 ) := [ an ] [ a ][ a ] [ pn q = n p n 1 q n 1 ]. Hence if [a 0,a 1,...,a n ] = p n /q n then [a n,...,a 1,a 0 ] = p n /p n 1. 7/43

8 Ultimately periodic continued fractions By an expression of the form [x,w], where x and w are finite strings, we mean the continued fraction [x,w,w,w,...], where the overbar or vinculum denotes the repeating portion. Thus, for example, 7 = [2,1,1,1,4] = [2,1,1,1,4,1,1,1,4,1,1,1,4,...]. 8/43

9 Numbers with easy to describe continued fractions We want to understand the characteristics and Diophantine properties of numbers whose continued fraction expansion is generated by a simple computational model, such as a finite automaton. Famous example: the Thue-Morse sequence on the symbols {a,b} where a = b and b = a, given by t = t 0 t 1 t 2 = abbabaab and defined by a, if n = 0; t n = t n/2, if n even; t n 1, if n odd. 9/43

10 Automatic sequences A continued fraction like [t] = [a,b,b,a,b,a,a,b,...] is called automatic because its terms are generated by a finite automaton. Basic idea: input to the automaton is n expressed in base 2 each symbol read causes a transition to the appropriate state output is associated with the last state reached /43

11 Previous results Queffélec proved that if a, b are distinct positive integers, then the real number [t] = [t 0,t 1,t 2,...] is transcendental. Later, a simpler proof was found by Adamczewski and Bugeaud. Queffélec also proved the transcendence of a much wider class of automatic continued fractions. Many new recent results about properties of automatic, morphic, and Sturmian continued fractions (Allouche, Davis, Quéffelec, Zamboni, Adamczewski, Bugeaud,...) Finally, Bugeaud proved (2012) that any automatic continued fraction is either a quadratic number or is transcendental. 11/43

12 Another automatic continued fraction Consider the number α k = n 0k 2n = k 1 +k 2 +k 4 +k 8 + for an integer k 2. These numbers, sometimes mistakenly called the Fredholm numbers, are all known to be transcendental. Their continued fraction is (for k 3): α k = [0,k 1,k+2,k,k,k 2,k,k+2,k,k 2,k+2,k,k 2,...] where the partial quotients all lie in the set {0,k 2,k 1,k,k +2}. (JOS and Kmošek, independently). 12/43

13 Another automatic continued fraction There is an automaton that generates this sequence: k k ,1 k ,1 0 k 2 0,1 1 k k k k 1 k Figure: Automaton generating the continued fraction for α k 13/43

14 k-regular sequences Drawback: all automatic sequences (and the more general class of morphic sequences) take only finitely many distinct values A more general class, allowing unbounded terms, is the k-regular sequences of integers, for integer k 2. These are sequences (a n ) n 0 where the k-kernel, defined by {(a k e n+i) n 0 : e 0, 0 i < k e }, is contained in a finitely generated module. Example: s k (n), the function counting the sum of the base-k digits of n. It satisfies s k (k e n+i) = s k (n)+s k (i) and hence each sequence (s k (k e n+i)) n 0 can be written as a linear combination of the sequence (s k (n)) n 0 and the constant sequence 1. 14/43

15 A conjecture Conjecture Every continued fraction where the terms form a k-regular sequence of positive integers is either transcendental or quadratic. 15/43

16 A particular 2-regular non-automatic sequence Let s = s 0 s 1 s 2 = (1,2,1,4,1,2,1,8,...) where s i = 2 ν 2(i+1) and ν p (x) is the p-adic valuation of x (the exponent of the largest power of p dividing x). To see that s is 2-regular, notice that every sequence in the 2-kernel is a linear combination of s itself and the constant sequence (1,1,1,...). Define the real number σ to have continued fraction expansion σ = [s] = [s 0,s 1,s 2,...] = [1,2,1,4,1,2,1,8,...] = The sequence s is sometimes called the ruler sequence. 16/43

17 Properties of σ σ has slowly growing partial quotients (indeed, s i i +1 for all i) But empirical calculation for σ 2 = demonstrates the appearance of some exceptionally large partial quotients! Here are the first few terms: σ 2 = [1,1,4,1,74,1,8457,1, ,1,1,1,2,1, , 11, 37, 1, 4, 2, ,11,5,3,12,2,34,2,9,8,1,1,2,7, ,. First noticed by Paul D. Hanna and Robert G. Wilson in November /43

18 The basic idea To prove that the partial quotients of σ 2 truly do grow doubly exponentially, we will rely on the following classical estimate: 1 (a n+1 +2)qn 2 < x p n q n < 1 a n+1 qn 2. So to show that some partial quotients of x are huge, it is sufficient to find convergents p n /q n of x such that x p n /q n is much smaller than q 2 n. 18/43

19 An Old Result We now recall a classical result: Lemma Let a 0 be a positive integer and w denote a finite palindrome of positive integers. Then p [a 0,w,2a 0 ] = q, where M(w) = [ α β γ δ ] p = a 2 0α+2a 0 β +δ q = α. 19/43

20 Two sequences We now define two related sequences for n 2: u(n) = (s 1,s 2,...,s 2 n 3) v(n) = (s 1,s 2,...,s 2 n 2) = (u(n),1) Recall that s n = 2 v 2(n+1) is defined in the introduction. The following table gives the first few values of these quantities: n u(n) v(n) /43

21 A proposition The following proposition, which is easily proved by induction, gives the relationship between these sequences, for n 2: Proposition (a) u(n+1) = (v(n),2 n,v(n) R ); (b) u(n) is a palindrome; (c) v(n+1) = (v(n),2 n,1,v(n)). 21/43

22 Two matrices Furthermore, we can define the sequence of associated matrices with u(n) and v(n): [ ] cn e M(u(n)) := n d n f n [ ] wn y M(v(n)) := n. x n z n 22/43

23 A calculation Now let σ n = [1,u(n),2] Then by the previous lemma with a 0 = 1 and w = u(n) we get σ n = cn +2e n +f n c n. Write σ = [s 0,s 1,...] and [s 0,s 1,...,s n ] = pn q n. Furthermore define ˆσ n = [1,u(n)]. Notice that σ, σ n, and ˆσ n all agree on the first 2 n 2 partial quotients. 23/43

24 An inequality We have σ ˆσ n < 1 q 2 n 3q 2 n 2 by a classical theorem on continued fractions. Furthermore, since s 2 n 3 = 2, s 2 n 2 = 1, we have, for n 3, that σ < σ n < ˆσ n. Hence σ σ n < 1 q 2 n 3q 2 n 2. 24/43

25 A good approximation Now by considering M(1)M(u(n))M(1) = = [ ][ ][ 1 1 cn e n d n f n 1 0 [ ] 2cn +d n e n +f n, c n +d n c n ] we see that q 2 n 3 = c n and q 2 n 2 = c n +d n. 25/43

26 A good approximation (continued) For simplicity write g n = c n +2e n +f n. Then σ σ n = σ gn < 1, and so σ 2 g n = σ c n gn c n c n c 2 n σ + gn < 3. So we have already found good approximations of σ 2 by rational numbers. Next we show g n and c n have a large common factor, which will improve the quality of the approximation. c n c 2 n 26/43

27 Improving the approximation From previous results we get that the matrix [ ] cn+1 e n+1, d n+1 f n+1 associated with u(n+1), is equal to the matrix associated with (v(n),2 n,v(n) R ), which is [ ][ ][ ] wn y n 2 n 1 wn x n 1 0 [ = x n z n 2 n wn 2 +2w n y n 2 n w n x n +x n y n +w n z n 2 n w n x n +x n y n +w n z n 2 n xn 2 +2x n z n y n z n ]. 27/43

28 Improving the approximation Notice that c n+1 = (2 n w n +2y n )w n. (1) On the other hand, we have g n+1 = c n+1 +2d n+1 +f n+1 = c n+1 +2(2 n w n x n +x n y n +w n z n )+2 n xn 2 +2x n z n = c n+1 +2(2 n w n x n +x n y n +w n z n )+2 n xn 2 +2x n z n +2(x n y n w n z n +1) = c n+1 +(2 n w n +2y n )2x n +2 n xn 2 +2x n z n +2 = (2 n w n +2y n )(2x n +w n )+2 n xn 2 +2x n z n +2, By Euclidean division, we get gcd(g n+1,2 n w n +2y n ) = gcd(2 n w n +2y n,2 n xn 2 +2x n z n +2). 28/43

29 The other matrix Next we work with v(n+1). We get that the matrix [ ] wn+1 y n+1 x n+1 z n+1 associated with v(n+1) is equal to the matrix associated with (v(n),2 n,1,v(n)), which is [ wn y n x n z n ][ 2 n ][ ][ wn y n x n z n [ (2 = n +1)wn 2 +2 n w n x n +y n (w n +x n ) (2 n +1)w n y n +2 n ] w n z n +y n (y n +z n ) (2 n +1)x n w n +2 n xn 2 +z n (w n +x n ) (2 n +1)x n y n +2 n. x n z n +z n (y n +z n ) ] 29/43

30 The other matrix Letting r n := 2 n (w n +x n )+w n +y n +z n, we see that 2 n+1 w n+1 +2y n+1 = 2 n+1 ((2 n +1)wn 2 +2 n w n x n +y n (w n +x n ))+ 2((2 n +1)w n y n +2 n w n z n +y n (y n +z n )) = 2(2 n w n +y n )(2 n (w n +x n )+w n +y n +z n ) = 2(2 n w n +y n )r n. (2) 30/43

31 A simplification Now x n+1 = (2 n +1)x n w n +2 n xn 2 +z n (w n +x n ) = x n (2 n (w n +x n )+w n +y n +z n )+w n z n x n y n = x n r n +1 and z n+1 = (2 n +1)x n y n +2 n x n z n +z n (y n +z n ) = z n (2 n (w n +x n )+w n +y n +z n )+(2 n +1)(x n y n w n z n ) = z n r n 2 n 1. 31/43

32 More simplification Tedious simplification now gives 2 n xn+1 2 +x n+1z n+1 +1 = (2 n +1)w n x n z n +2 n (2 n +1)w n xn 2 +z n r n +(2 n 1)x n +2 n xny 2 n +2 n+1 xnz 2 n +x n y n z n +2 2n xn 3 +x n zn. 2 Reindexing, we get c n+2 = w n+1 (2 n+1 w n+1 +2y n+1 ) = 2w n+1 (2 n w n +y n )r n, 32/43

33 Refining the approximation Also, from the argument above about gcd s, we see that 2r n g n+2. Hence for n 2 we have for integers g n+2 c n+2 = P n+2 Q n+2 P n+2 := g n+2 2r n Q n+2 := w n+1 (2 n w n +y n ). It remains to see that P n+2 /Q n+2 are particularly good rational approximations to σ 2. 33/43

34 Since w n /x n and y n /z n denote successive convergents to a continued fraction, we have w n x n, w n y n, and w n z n. It follows that Q n+2 = w n+1 (2 n w n +y n ) = ((2 n +1)w 2 n +2 n w n x n +y n (w n +x n ))(2 n w n +y n ) (2 n+1 +3)w 2 n (2 n +1)w n = (2 n+1 +3)(2 n +1)w 3 n. 34/43

35 The approximation is better than we hoped On the other hand, c n+2 = 2Q n+2 r n > 2(2 n +1)wn 2 2 n w n (2 n +1)w n = 2 n+1 (2 n +1) 2 wn 4 Q 4/3 2 n+1 (2 n +1) 2 n+2 ((2 n+1 +3)(2 n +1)) 4/3 > Q 4/3 n+2. 35/43

36 Finally, the result about σ 2 Putting this all together, we get Theorem σ2 P n+2 < Q 8/3 n+2 for all integers n 2. Q n+2 36/43

37 Irrationality measure The result we have just shown can be nicely formulated in terms of the irrationality measure. Theorem The irrationality measure of σ 2 is at least 8/3. It would be interesting to find the precise value of the irrationality measure of σ 2. Numerical experiments suggest that the bound 8/3 is sharp; however, we do not currently know how to prove that. 37/43

38 Remarks Note that the classical Khintchine theorem states that for almost all real numbers (in terms of Lebesgue measure), the irrationality measure equals two. Hence our theorem says that the number σ 2 belongs to a very tiny set of zero Lebesgue measure. From Roth s theorem we conclude that σ 2 (and hence σ) are transcendental numbers. However... By a 2007 result of Adamczewski and Bugeaud, we already know that if a continued fraction of a real number x starts with arbitrarily long palindromes, then x is transcendental. Clearly σ satisfies this property. 38/43

39 Some partial quotients are double exponential in size We now provide a lower bound for some very large partial quotients of σ 2. For each n 2 we have σ2 P n+2 < Q 8/3 n+2 < 1 Q n+2 2Q 2 n+2. In particular this implies that the rational number P n+2 /Q n+2 is a convergent of σ 2. 39/43

40 Observations Notice that P n+2 and Q n+2 are not necessarily relatively prime. Let P n+2 / Q n+2 denote the reduced fraction of P n+2 /Q n+2. If P n+2 / Q n+2 is the m th convergent of σ 2, then define A n+2 to be the (m+1) th partial quotient of σ 2. 40/43

41 An inequality Then the estimate above implies 1 < (A n+2 +2) Q n+2 2 σ2 P n+2 Q n+2 < 3 c 2 n+2 where we used the fact (established earlier) that c n+2 = 2r n Q n+2 2r n Q n+2. Hence A n r2 n rn 2 Q, n /43

42 The last result From the formula for r n and the inequalities w n x n,w n y n,w n z n one can easily derive (2 n +1)w n r n (2 n+1 +3)w n. This, together with the formula for w n+1, gives the estimate r n+1 (2 n+1 +1)w n+1 (2 n+1 +1)(2 n +1)w 2 n > r 2 n +1. Therefore we get Theorem The sequence 4 3 r2 n 2, and in turn the sequence A n+2, both grow doubly exponentially. This explains the observation of Hanna and Wilson. 42/43

43 Open problems 1. Say something interesting about the continued fraction expansion of e What are the diophantine properties of the powers of the number [0,1,1,2,1,2,2,3,...] where the n th partial quotient is s 2 (n)? 43/43