Distribution of the Longest Gap in Positive Linear Recurrence Sequences

Transcription

1 Distribution of the Longest Gap in Positive Linear Recurrence Sequences Shiyu Li 1, Philip Tosteson 2 Advisor: Steven J. Miller 2 1 University of California, Berkeley 2 Williams College Young Mathematicians Conference Columbus, Ohio, July 27-29, 2012

2 Introduction

3 Zeckendorf Decompositions Fibonacci Numbers: F n+1 = F n + F n 1 ;

4 Zeckendorf Decompositions Fibonacci Numbers: F n+1 = F n + F n 1 ; F 1 = 1, F 2 = 2, F 3 = 3, F 4 = 5,.

5 Zeckendorf Decompositions Fibonacci Numbers: F n+1 = F n + F n 1 ; F 1 = 1, F 2 = 2, F 3 = 3, F 4 = 5,. Zeckendorf s Theorem Every positive integer can be written uniquely as a sum of non-consecutive Fibonacci numbers.

6 Zeckendorf Decompositions Fibonacci Numbers: F n+1 = F n + F n 1 ; F 1 = 1, F 2 = 2, F 3 = 3, F 4 = 5,. Zeckendorf s Theorem Every positive integer can be written uniquely as a sum of non-consecutive Fibonacci numbers. Example: 2012 = = F 16 + F 13 + F 8 + F 3 + F 1.

7 Zeckendorf Decompositions Fibonacci Numbers: F n+1 = F n + F n 1 ; F 1 = 1, F 2 = 2, F 3 = 3, F 4 = 5,. Zeckendorf s Theorem Every positive integer can be written uniquely as a sum of non-consecutive Fibonacci numbers. Example: 2012 = = F 16 + F 13 + F 8 + F 3 + F 1. Lekkerkerker s Theorem (1952) The average number of summands in the Zeckendorf decomposition for integers in [F n, F n+1 ) tends to n ϕ n, where ϕ = is the golden mean.

8 Gaps: Our object of study Instead of looking at the number of summands, we study the spacings between them, as follows: Definition: Gaps If x [F n, F n+1 ) has Zeckendorf decomposition x = F n + F n g1 + F n g2 + + F n gk, we define the gaps in its decomposition to be {g 1, g 1 g 2,, g k 1 g k }. Example: 2012 = F 16 + F 13 + F 8 + F 3 + F 1.

9 Gaps: Our object of study Instead of looking at the number of summands, we study the spacings between them, as follows: Definition: Gaps If x [F n, F n+1 ) has Zeckendorf decomposition x = F n + F n g1 + F n g2 + + F n gk, we define the gaps in its decomposition to be {g 1, g 1 g 2,, g k 1 g k }. Example: 2012 = F 16 + F 13 + F 8 + F 3 + F 1. Gaps of length 3, 5, 5, and 2.

10 Previous Results (Miller-Beckwith) Theorem (Zeckendorf Gap Distribution) For Zeckendorf decompositions, P(k) = 1 φ k φ = the golden mean. for k 2, with

11 Previous Results (Miller-Beckwith) Theorem (Zeckendorf Gap Distribution) For Zeckendorf decompositions, P(k) = 1 φ k φ = the golden mean. for k 2, with Figure: Distribution of gaps in [F 1000, F 1001 ); F

12 Our Problem: Given a random number x in the interval [F n, F n+1 ), what is the probability that x has longest gap equal to r?

13 Why is this exciting?

14 Why is this exciting? In probability theory, the longest run of heads in a sequence of n coin tosses has generated much work. The longest run grows logarithmically, and has finite variance, independent of n.

15 Why is this exciting? In probability theory, the longest run of heads in a sequence of n coin tosses has generated much work. The longest run grows logarithmically, and has finite variance, independent of n. After modifications, our problem looks similar

16 Why is this exciting? In probability theory, the longest run of heads in a sequence of n coin tosses has generated much work. The longest run grows logarithmically, and has finite variance, independent of n. After modifications, our problem looks similar The study of these behaviors has many applications. For example:

17 Why is this exciting? In probability theory, the longest run of heads in a sequence of n coin tosses has generated much work. The longest run grows logarithmically, and has finite variance, independent of n. After modifications, our problem looks similar The study of these behaviors has many applications. For example: Fraud detection (IRS)

18 Why is this exciting? In probability theory, the longest run of heads in a sequence of n coin tosses has generated much work. The longest run grows logarithmically, and has finite variance, independent of n. After modifications, our problem looks similar The study of these behaviors has many applications. For example: Fraud detection (IRS) Machine running times and Algorithms

19 Why is this exciting? In probability theory, the longest run of heads in a sequence of n coin tosses has generated much work. The longest run grows logarithmically, and has finite variance, independent of n. After modifications, our problem looks similar The study of these behaviors has many applications. For example: Fraud detection (IRS) Machine running times and Algorithms Philadelphia International Airport

20

21 Method Our plan of attack is as follows:

22 Method Our plan of attack is as follows: Recast the problem in a combinatorial perspective

23 Method Our plan of attack is as follows: Recast the problem in a combinatorial perspective Use generating functions!

24 Method Our plan of attack is as follows: Recast the problem in a combinatorial perspective Use generating functions! Discover relationships between important quantities

25 Method Our plan of attack is as follows: Recast the problem in a combinatorial perspective Use generating functions! Discover relationships between important quantities Analyze limiting behavior

26 Method Our plan of attack is as follows: Recast the problem in a combinatorial perspective Use generating functions! Discover relationships between important quantities Analyze limiting behavior First we ll review what we need to know.

27 Review of Generating Functions Given a sequence a n, its generating function is A(x) = n=0 a nx n.

28 Review of Generating Functions Given a sequence a n, its generating function is A(x) = n=0 a nx n. Let s find the number of ways to have k numbers sum to n (where order matters), all the numbers are less than b.

29 Review of Generating Functions Given a sequence a n, its generating function is A(x) = n=0 a nx n. Let s find the number of ways to have k numbers sum to n (where order matters), all the numbers are less than b. So we have a a k = n, a 1,..., a k < b.

30 Review of Generating Functions Given a sequence a n, its generating function is A(x) = n=0 a nx n. Let s find the number of ways to have k numbers sum to n (where order matters), all the numbers are less than b. So we have a a k = n, a 1,..., a k < b. We can represent this by the nth coefficient of the function (1 + x x b 1 ) k.

31 Review of Generating Functions Given a sequence a n, its generating function is A(x) = n=0 a nx n. Let s find the number of ways to have k numbers sum to n (where order matters), all the numbers are less than b. So we have a a k = n, a 1,..., a k < b. We can represent this by the nth coefficient of the function (1 + x x b 1 ) k. Using geometric series expansion, this is equal to ( ) 1 x n k. 1 x

32 Partial Fraction Expansion Now we will introduce a basic idea of our derivations. Consider: f (x) = 1 1 x x 2.

33 Partial Fraction Expansion Now we will introduce a basic idea of our derivations. Consider: f (x) = 1 1 x x 2. Using the method of partial fractions, we can write this as: f (x) = c 1 x + φ + c 2 x 1/φ, where φ and 1/φ are the roots of 1 x x 2, and c 1 and c 2 are determined by algebra to be c 1 = c 2 = 5.

34 Partial Fraction Expansion So we have: f (x) = [ 5 φ 1 + (1/φ)x + (1/φ) 1 φx ].

35 Partial Fraction Expansion So we have: f (x) = [ 5 φ 1 + (1/φ)x + (1/φ) 1 φx ]. Which by geometric series expansion gives: f (x) = i=0 1 5 [( 1/φ) i + φ i] x i.

36 Partial Fraction Expansion So we have: f (x) = [ 5 φ 1 + (1/φ)x + (1/φ) 1 φx ]. Which by geometric series expansion gives: f (x) = i=0 1 5 [( 1/φ) i + φ i] x i. And by Binet s well known formula for the Fibonacci numbers, we see that: f (x) = F i x i. i=0

37 Why will this be useful?

38 Why will this be useful? If we can find the generating function for our desired quantity in closed form: 1 f (x) = 1 x x 2,

39 Why will this be useful? If we can find the generating function for our desired quantity in closed form: 1 f (x) = 1 x x 2, the partial fraction expansion reduces its limiting behavior to the study of the roots of a polynomial. f (x) = i=0 1 5 [( 1/φ) i + φ i] x i.

40 Why will this be useful? If we can find the generating function for our desired quantity in closed form: 1 f (x) = 1 x x 2, the partial fraction expansion reduces its limiting behavior to the study of the roots of a polynomial. f (x) = i=0 1 5 [( 1/φ) i + φ i] x i.

41 Results

42 Cumulative Distribution Function Pick x randomly from the interval [F n, F n+1 ). We prove explicitly the cumulative distribution of x s longest gap.

43 Cumulative Distribution Function Pick x randomly from the interval [F n, F n+1 ). We prove explicitly the cumulative distribution of x s longest gap. Theorem Let r = φ 2 /(φ 2 + 1). Define f as f (n, u) = log rn/ log φ + u for some fixed u R. Then, as n, the probability that x [F n, F n+1 ) has longest gap less than or equal to f (n) converges to e(1 u) log φ+{f (n)} P(L(x) f (n, u)) = e

44 Cumulative Distribution Function Pick x randomly from the interval [F n, F n+1 ). We prove explicitly the cumulative distribution of x s longest gap. Theorem Let r = φ 2 /(φ 2 + 1). Define f as f (n, u) = log rn/ log φ + u for some fixed u R. Then, as n, the probability that x [F n, F n+1 ) has longest gap less than or equal to f (n) converges to e(1 u) log φ+{f (n)} P(L(x) f (n, u)) = e Immediate Corollary: If f (n, u) grows any slower or faster than log n/ log φ, then P(L(x) f (n)) goes to 0 or 1 respectively.

45 Mean and Variance We can use the CDF to determine the regular distribution function, and particularly the mean and variance.

46 Mean and Variance We can use the CDF to determine the regular distribution function, and particularly the mean and variance. Let φ2 log( φ P(u) = P (L(x) 2 +1 n) ) + u, log φ

47 Mean and Variance We can use the CDF to determine the regular distribution function, and particularly the mean and variance. Let φ2 log( φ P(u) = P (L(x) 2 +1 n) ) + u, log φ then the distribution of the longest gap is approximately d du P(u).

48 Mean and Variance We can use the CDF to determine the regular distribution function, and particularly the mean and variance. Let φ2 log( φ P(u) = P (L(x) 2 +1 n) ) + u, log φ then the distribution of the longest gap is approximately d du P(u). The mean is given by µ = The variance follows similarly. u d du P(u)du.

49 Mean and Variance So the mean is about ( ) log φ 2 φ µ = e e(1 u) log φ e (1 u) log φ log φ du. log φ

50 Mean and Variance So the mean is about ( ) log φ 2 φ µ = e e(1 u) log φ e (1 u) log φ log φ du. log φ ( ( ) φ 2 = 1/ log φ log φ ) log we w dw

51 Mean and Variance So the mean is about ( ) log φ 2 φ µ = e e(1 u) log φ e (1 u) log φ log φ du. log φ ( ( ) φ 2 = 1/ log φ log φ Theorem In the continuous approximation, the mean is ( ) log φ 2 φ 2 +1 n γ, log φ where γ is the Euler- Mascheroni constant. 0 ) log we w dw

52 Fibonacci Case Generating Function We let G n,k,f be the number of ways to have a Zeckendorf decomposition with k nonzero summands and all gaps less than f (n).

53 Fibonacci Case Generating Function We let G n,k,f be the number of ways to have a Zeckendorf decomposition with k nonzero summands and all gaps less than f (n). G n,k,f is the coefficient of x n+1 for the generating function 1 1 x j=2 k 1 f (n) 1 x j

54 The Combinatorics 1 Why the n + 1st coefficient of 1 x f (n) 1 x j j=2 k 1?

55 The Combinatorics 1 Why the n + 1st coefficient of 1 x f (n) 1 x j j=2 k 1? Let y = F n + F n g1 + F n g1 g F n g1 g n 1, then we have the following:

56 The Combinatorics 1 Why the n + 1st coefficient of 1 x f (n) 1 x j j=2 k 1? Let y = F n + F n g1 + F n g1 g F n g1 g n 1, then we have the following: The sum of the gaps of x is n.

57 The Combinatorics 1 Why the n + 1st coefficient of 1 x f (n) 1 x j j=2 k 1? Let y = F n + F n g1 + F n g1 g F n g1 g n 1, then we have the following: The sum of the gaps of x is n. Each gap is 2.

58 The Combinatorics 1 Why the n + 1st coefficient of 1 x f (n) 1 x j j=2 k 1? Let y = F n + F n g1 + F n g1 g F n g1 g n 1, then we have the following: The sum of the gaps of x is n. Each gap is 2. Each gap is < f (n).

59 The Combinatorics 1 Why the n + 1st coefficient of 1 x f (n) 1 x j j=2 k 1? Let y = F n + F n g1 + F n g1 g F n g1 g n 1, then we have the following: The sum of the gaps of x is n. Each gap is 2. Each gap is < f (n). The gaps uniquely identify y because of Zeckendorf s Theorem!

60 The Combinatorics Then G n,k,f is the number of ways to choose k gaps between 2 and f (n) 1, that add up to n.

61 The Combinatorics Then G n,k,f is the number of ways to choose k gaps between 2 and f (n) 1, that add up to n. So its the n th coefficient of = 1 1 x x 2(k 1) 1 [ x x f (n) 1] k 1. 1 x ( 1 x f 2 1 x ) k 1

62 The Combinatorics Then G n,k,f is the number of ways to choose k gaps between 2 and f (n) 1, that add up to n. So its the n th coefficient of = 1 1 x x 2(k 1) 1 [ x x f (n) 1] k 1. 1 x ( 1 x f 2 1 x ) k 1 For fixed k, this is surprisingly hard to analyze. We only care about the sum over all k.

63 The Generating Function pt 2 If we sum over k we get the total number of x [F n, F n+1 ) with longest gap < f (n) call it G k,f.

64 The Generating Function pt 2 If we sum over k we get the total number of x [F n, F n+1 ) with longest gap < f (n) call it G k,f. It s the n + 1st coefficient of F(x) = 1 1 x ( ) 1 x f 2 = 1 x k=1 x 1 x x 2 + x f (n)

65 The Generating Function pt 2 If we sum over k we get the total number of x [F n, F n+1 ) with longest gap < f (n) call it G k,f. It s the n + 1st coefficient of F(x) = 1 1 x ( ) 1 x f 2 = 1 x k=1 x 1 x x 2 + x f (n) Let s work with this (use partial fractions and Rouché) to find the CDF.

66 Partial Fractions pt 2 Write the roots of x f x 2 x 1 as {α i } f i=1. We can write our generating function F(x) = x 1 x x 2 + x f (n) =

67 Partial Fractions pt 2 Write the roots of x f x 2 x 1 as {α i } f i=1. We can write our generating function F(x) = x 1 x x 2 + x f (n) = f (n) i=1 f (n)α f (n) i α i 2α 2 i α i ( ) x j. α i j=1

68 Partial Fractions pt 2 Write the roots of x f x 2 x 1 as {α i } f i=1. We can write our generating function F(x) = x 1 x x 2 + x f (n) = f (n) i=1 f (n)α f (n) i α i 2α 2 i α i ( ) x j. α i j=1 We can take the n + 1st coefficient of this expansion to find the number of y with gaps less than f (n).

69 Partial Fractions pt 3 Divide the number of y [F n, F n+1 ) with longest gap < f (n), by the total number of y, which is F n+1 F n 1 5 φ n.

70 Partial Fractions pt 3 Divide the number of y [F n, F n+1 ) with longest gap < f (n), by the total number of y, which is F n+1 F n 1 5 φ n. Theorem The proportion of y [F n, F n+1 ) with L(x) < f (n) is exactly f (n) i=1 f (n)α f (n) i 5(α i ) 2α 2 i α i ( 1 α i ) n+1 1 (φ n ( 1/φ) n )

71 Partial Fractions pt 3 Divide the number of y [F n, F n+1 ) with longest gap < f (n), by the total number of y, which is F n+1 F n 1 5 φ n. Theorem The proportion of y [F n, F n+1 ) with L(x) < f (n) is exactly f (n) i=1 f (n)α f (n) i 5(α i ) 2α 2 i α i ( 1 α i ) n+1 1 (φ n ( 1/φ) n ) Now, we find out about the roots of x f x 2 x + 1.

72 Rouché and Roots When f (n) is large z f (n) is very small, for z < 1. Thus, by Rouché s theorem from complex analysis:

73 Rouché and Roots When f (n) is large z f (n) is very small, for z < 1. Thus, by Rouché s theorem from complex analysis: Lemma For f N and f 4, the polynomial p f (z) = z f z 2 z + 1 has exactly one root z f with z f <.9. Further, z f R and z f = 1 φ + zf f, so as f, z f converges to 1 φ. z f +φ

74 Rouché and Roots When f (n) is large z f (n) is very small, for z < 1. Thus, by Rouché s theorem from complex analysis: Lemma For f N and f 4, the polynomial p f (z) = z f z 2 z + 1 has exactly one root z f with z f <.9. Further, z f R and z f = 1 φ + zf f, so as f, z f converges to 1 φ. z f +φ And we only care about the small roots.

75 Getting the CDF As f grows, only one root goes to 1/φ. The other roots don t matter. So,

76 Getting the CDF As f grows, only one root goes to 1/φ. The other roots don t matter. So, Theorem If lim n f (n) =, the proportion of y with L(y) < f (n) is, as n ( lim (φz f ) n φz f (n) ) n f = lim 1 +. n n φ + z f And if f (n) is bounded, then P f = 0.

77 Getting the CDF Theorem In the limit, if lim n f (n) = then ( P f = lim P n,f = lim (φz f ) n φz f (n) ) n f = lim 1 +. n n n φ + z f And if f (n) is bounded, then P f = 0.

78 Getting the CDF Theorem In the limit, if lim n f (n) = then ( P f = lim P n,f = lim (φz f ) n φz f (n) ) n f = lim 1 +. n n n φ + z f And if f (n) is bounded, then P f = 0. We re almost there!:

79 Getting the CDF Theorem In the limit, if lim n f (n) = then ( P f = lim P n,f = lim (φz f ) n φz f (n) ) n f = lim 1 +. n n n φ + z f And if f (n) is bounded, then P f = 0. We re almost there!: Take logarithms

80 Getting the CDF Theorem In the limit, if lim n f (n) = then ( P f = lim P n,f = lim (φz f ) n φz f (n) ) n f = lim 1 +. n n n φ + z f And if f (n) is bounded, then P f = 0. We re almost there!: Take logarithms Taylor expand

81 Getting the CDF Theorem In the limit, if lim n f (n) = then ( P f = lim P n,f = lim (φz f ) n φz f (n) ) n f = lim 1 +. n n n φ + z f And if f (n) is bounded, then P f = 0. We re almost there!: Take logarithms Taylor expand Use that z f 1/φ as n

82 Getting the CDF Theorem In the limit, if lim n f (n) = then ( P f = lim P n,f = lim (φz f ) n φz f (n) ) n f = lim 1 +. n n n φ + z f And if f (n) is bounded, then P f = 0. We re almost there!: Take logarithms Taylor expand Use that z f 1/φ as n Re-exponentiate

83 Cumulative Distribution Function After some technical details, we get our long awaited theorem! Theorem Let r = φ 2 /(φ + 1). Define f : as f (n) = log rn/ log φ + u for some fixed u R. Then, as n, the probability that x [F n, F n+1 ) has longest gap less than f (n) is P(L(x) < f (n)) = e eu log φ +{f }

84 Conclusions and Future Work

85 Positive Linear Recurrence Sequences This method can be greatly generalized to Positive Linear Recurrence Sequences ie: linear recurrences with non-negative coefficients. WLOG: H n+1 = c 1 H n (t1 =0) + c 2 H n t2 + + c L H n tl.

86 Positive Linear Recurrence Sequences This method can be greatly generalized to Positive Linear Recurrence Sequences ie: linear recurrences with non-negative coefficients. WLOG: H n+1 = c 1 H n (t1 =0) + c 2 H n t2 + + c L H n tl. Theorem (Zeckendorf s Theorem for PLRS recurrences) Any b N has a unique legal decomposition into sums of H n, b = a 1 H i1 + + a ik H ik. Here legal reduces to non-adjacency of summands in the Fibonacci case.

87 Generating Function, pt n The number of b [H n, H n+1 ), with longest gap < f is the coefficient of x n s in the generating function:

88 Generating Function, pt n The number of b [H n, H n+1 ), with longest gap < f is the coefficient of x n s in the generating function: [ ((c1 1)x t (c L 1)x t ) ( x s+1 x f L k 0 1 x ) +

89 Generating Function, pt n The number of b [H n, H n+1 ), with longest gap < f is the coefficient of x n s in the generating function: x t 1 [ ((c1 1)x t (c L 1)x t ) ( x s+1 x f L k 0 ( x s+t 2 t 1 +1 x f 1 x 1 x ) + ) ( + + x t L 1 x s+t L t L x f ) ] k 1 x

90 Generating Function, pt n The number of b [H n, H n+1 ), with longest gap < f is the coefficient of x n s in the generating function: x t 1 [ ((c1 1)x t (c L 1)x t ) ( x s+1 x f L k 0 ( x s+t 2 t 1 +1 x f 1 x 1 x ) + ) ( + + x t L 1 x s+t L t L x f ) ] k 1 x 1 ( c1 1 + c 2 x t c L x t ) L 1 x

91 Generating Function, pt n The number of b [H n, H n+1 ), with longest gap < f is the coefficient of x n s in the generating function: x t 1 [ ((c1 1)x t (c L 1)x t ) ( x s+1 x f L k 0 ( x s+t 2 t 1 +1 x f 1 x A geometric series! 1 x ) + ) ( + + x t L 1 x s+t L t L x f ) ] k 1 x 1 ( c1 1 + c 2 x t c L x t ) L 1 x

92 Some More Definitions Let λ i be the eigenvalues of the recurrence, and p i their coefficients. Define G(x) = L ) (x 1λi i=2 P(x) = (c 1 1)x t 1 + c 2 x t c L x t L, R(x) = c 1 x t 1 + c 2 x t (c L 1)x t L. and M(x) = 1 c 1 x c 2 x t 2+1 c L x t L+1

93 Exact Cumulative Distribution There is again a critical root, z f 1/λ 1 exponentially as f.

94 Exact Cumulative Distribution There is again a critical root, z f 1/λ 1 exponentially as f. Theorem PLRS Cumulative Distribution The cumulative distribution of the longest gap in [H n, H n+1 ) is: P(L(x) < f ) = P(z f ) / (p 1 λ 1 p 1 ) ( 1 ) n z f M (z f ) + f zf f R(z f ) + z f +1 f R (z f ) z f λ 1 +H(n, f ) where there exists ɛ with 1/λ 1 < ɛ < 1, such that H n,f f ɛ n

95 Exact Cumulative Distribution There is again a critical root, z f 1/λ 1 exponentially as f. Theorem PLRS Cumulative Distribution The cumulative distribution of the longest gap in [H n, H n+1 ) is: P(L(x) < f ) = P(z f ) / (p 1 λ 1 p 1 ) ( 1 ) n z f M (z f ) + f zf f R(z f ) + z f +1 f R (z f ) z f λ 1 +H(n, f ) where there exists ɛ with 1/λ 1 < ɛ < 1, such that H n,f f ɛ n Our techniques handle this!

96 References References Kologlu, Kopp, Miller and Wang: Fibonacci case. Miller - Wang: Main paper. Miller - Wang: Survey paper.

97 Thank you!!! Our thanks go out to: The hardworking fundraisers and organizers for The Ohio State Young Mathematicians Conference! The National Science Foundation Professor Steven J. Miller, and Williams College