Leaf-to-Leaf path distances in Complete Tree Graphs Andrew M. Goldsborough , S. Alex Rautu , Rudolf A. Römer Department of Physics and Centre for Scientific Computing The University of Warwick A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 1 / 28 Motivation - Correlation Functions In tensor networks, correlation functions are related to path length through the network. A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 2 / 28 What about in a tree? A regular tree tensor network (TTN) has the geometry of a complete binary tree Can we find an analogue to the correlation functions that can be solved analytically? A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 3 / 28 Definitions Level n Root node Vertex 1 2 3 4 Separation r Leaf Binary, as each vertex has two daughters. Complete, as all of the leaves are the same depth (number of vertices to the root). Let the path length d be the number of vertices that connects two leaves. Let the system Length L be the number of leaves. L = 2n . A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 4 / 28 The problem We want to find the average path length An (r) for a given separation r in a tree with n levels. Let Pn (r) be the sum of the path lengths connecting leaves of separation r in a tree with n levels. For separation r there are 2n − r possible paths. Thus: An (r) = A. M. Goldsborough (Warwick) Pn (r) 2n − r Paths and Trees 23/01/2014 5 / 28 n = 4, r = 1 r=1 1 3 1 A. M. Goldsborough (Warwick) 5 1 3 1 7 1 Paths and Trees 3 1 5 1 3 1 23/01/2014 Pn(r) 37 6 / 28 n = 4, higher r r=1 2 3 4 5 6 7 8 1 3 3 3 3 5 1 5 5 5 5 5 A. M. Goldsborough (Warwick) 5 5 5 5 5 5 5 7 1 5 5 7 3 5 7 7 3 3 7 7 1 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 1 7 7 7 Paths and Trees 7 3 7 7 7 3 3 7 7 3 5 7 7 1 5 5 7 5 5 5 7 5 5 5 5 5 5 5 1 5 5 3 5 3 3 3 1 23/01/2014 Pn(r) 37 58 63 68 65 62 59 56 7 / 28 n = 4, higher r r=1 2 3 4 5 6 7 8 1 3 3 3 3 5 1 5 5 5 5 5 A. M. Goldsborough (Warwick) 5 5 5 5 5 5 5 7 1 5 5 7 3 5 7 7 3 3 7 7 1 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 1 7 7 7 Paths and Trees 7 3 7 7 7 3 3 7 7 3 5 7 7 1 5 5 7 5 5 5 7 5 5 5 5 5 5 5 1 5 5 3 5 3 3 3 1 23/01/2014 Pn(r) 37 58 63 68 65 62 59 56 8 / 28 Look at just the maximal paths n-1 A. M. Goldsborough (Warwick) n-1 Paths and Trees 23/01/2014 9 / 28 Look at just the maximal paths n-1 Pn (r) = A. M. Goldsborough (Warwick) n-1 2Pn−1 (r) + (2n − 1)r (2n − 1)(2n − r) Paths and Trees r < 2n−1 r ≥ 2n−1 23/01/2014 10 / 28 Recurrence relation n-1 Pn (r) = A. M. Goldsborough (Warwick) n-1 2Pn−1 (r) + (2n − 1)r (2n − 1)(2n − r) Paths and Trees r < 2n−1 r ≥ 2n−1 23/01/2014 11 / 28 Solving the recursion relation Pn (r) = 2Pn−1 (r) + (2n − 1)r 2Pn−1 (r) = 22 Pn−2 (r) + 2(2n − 3)r 22 Pn−2 (r) = 23 Pn−3 (r) + 22 (2n − 5)r .. . 2ν Pn−ν (r) = 2ν+1 Pn−ν−1 (r) + 2ν (2(n − ν) − 1)r .. . 2n−nc Pnc (r) = 2n−nc +1 Pnc −1 (r) + 2n−nc (2nc − 1)r summed gives Pn (r) = 2n−nc +1 Pnc −1 (r) + n−n Xc 2i (2(n − i) − 1)r i=0 A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 12 / 28 Solving the recursion relation The first part is from the non-recursive part of the equation for Pn (r): Pnc −1 (r) = (2nc − 3)(2nc −1 − r) When the final bracket is expanded, it is simply the sum of three geometric series: n−n Xc 2i (2(n − i) − 1)r = rn i=0 n−n c +1 X 2i − r i=1 n−n Xc i=1 i2i+1 − r n−n Xc 2i i=0 = r(2n−nc +1 (2nc + 1) − (3 + 2n)) A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 13 / 28 Finding nc From the limits r < 2n−1 : n > log2 r + 1 The critical value for n is the smallest integer that satisfies this condition or: {nc ∈ Z : log2 r + 1 < nc ≤ log2 r + 2} hence: nc = blog2 rc + 2 ≡ q + 2 A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 14 / 28 The result Putting it all together: 2r Pn (r) = 2n 2q + 1 + q − (3 + 2n)r 2 1 2r n 2 2q + 1 + q − (3 + 2n)r An (r) = n 2 −r 2 In the limit of n → ∞: lim An (r) = 2q + 1 + n→∞ A. M. Goldsborough (Warwick) Paths and Trees 2r 2q 23/01/2014 15 / 28 Results 50 n = 20 n→∞ 40 An(r) 30 20 10 0 1 10 100 1000 10000 1e+05 1e+06 |r| A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 16 / 28 What about ternary trees? Ternary trees have three daughters at each vertex P3,n (r) = A. M. Goldsborough (Warwick) 3P3,n−1 (r) + 2(2n − 1)r (2n − 1)(3n − r) Paths and Trees r < 3n−1 r ≥ 3n−1 23/01/2014 17 / 28 What about m-ary trees? Pm,n (r) = r < mn−1 r ≥ mn−1 mPm,n−1 (r) + (m − 1)(2n − 1)r (2n − 1)(mn − r) Pm,n (r) = m 2qm + 1 + n 2r q m m (m − 1) 1+m − 2n − 1−m r 1 2r 1+m n Am,n (r) = n m 2qm + 1 + qm − 2n − r m −r m (m − 1) 1−m lim Am,n (r) = 2qm + 1 + n→∞ 2r mqm (m − 1) where qm = blogm rc. A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 18 / 28 Results for n → ∞ 50 m=2 m=3 m=5 m = 10 m = 50 m = 100 40 An(r) 30 20 10 0 1 10 100 1000 10000 1e+05 1e+06 |r| A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 19 / 28 Variance Want a measure of uncertainty for the averages. Variance can be defined as Qm,n (r) Pm,n (r) 2 Vm,n (r) = n − m −r mn − r Where Qm,n (r) is the sum of the squares of the path lengths: mQm,n−1 (r) + (m − 1)(2n − 1)2 r r < mn−1 Qm,n (r) = 2 n (2n − 1) (m − r) r ≥ mn−1 A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 20 / 28 Variance In the end: Vm,n (r) = h 4r 2n qm m m (m + 1) − r m2qm (mn − r)2 (m − 1)2 − mn m2qm +2 (n − qm )2 − m2qm +1 (2n2 − n(4qm + 2) + 2qm (qm + 1) − 1) + m2qm (qm − n + 1)2 i − mqm (2n − 2qm − 1)(m − 1)r + m2qm +1 r and lim Vm,n (r) = n→∞ A. M. Goldsborough (Warwick) 4r (mqm (m + 1) − r) m2qm (m − 1)2 Paths and Trees 23/01/2014 21 / 28 Variance for binary trees 8 Vn(r) 6 4 2 n = 20 n→∞ 0 1 100 10000 1e+06 |r| A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 22 / 28 Various m with n → ∞ 10 9 8 m=2 m=3 m=5 m = 10 m = 50 m = 100 7 Vn(r) 6 5 4 3 2 1 0 1 100 10000 1e+06 |r| A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 23 / 28 General raw moments The k-th raw statistical moment is: E[X k ] = Am,k,n (r) = 0 Pm,k,n (r) mn − r 0 Pm,k,n (r) = mn−nc (2n0c − 1)k (mnc − r) 1 1 0 + r(m − 1)(−2)k Φ m, −k, − n − mn−nc Φ m, −k, − n0c 2 2 h 0 0 lim Am,k,n (r) = m−nc (2n0c − 1)k (mnc − r) n→∞ 1 0 k r(m − 1)(−2) Φ m, −k, − nc 2 where n0c = nc − 1 = q + 1 and Φ is the Hurwitz-Lerch transcendent or Hurwitz-Lerch zeta-function. A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 24 / 28 Periodic Systems A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 25 / 28 General raw moments for periodic systems P BC OBC OBC Pm,k,n (r) = Pm,k,n (r) + Pm,k,n (mn − r) 0 0 0 0 = mn−nc (2n0c − 1)k (mnc − r) + mn−ñc (2ñ0c − 1)k (mñc − (mn − r)) 1 1 n−n0c 0 n Φ m, −k, − nc + (m − 1)(−2) m Φ m, −k, − n − rm 2 2 k 1 0 − (mn − r)mn−ñc Φ m, −k, − ñ0c 2 where n0c = blogm rc + 1 as before and ñ0c = blogm (2n − r)c + 1. A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 26 / 28 Conclusions We can find an analytic expression for average path length for a given separation for binary trees. This can be generalised to m-ary trees, periodic boundaries and general raw statistical moments. The general equation can be concisely written in terms of the Hurwitz-Lerch Transcendent. Maths is fun! A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 27 / 28 Thank you! A. M. Goldsborough (Warwick) Paths and Trees 23/01/2014 28 / 28