Susceptibility of the Two-Dimensional Ising Model

Transcription

1 Susceptibility of the Two-Dimensional Ising Model Craig A. Tracy UC Davis September 2014

2 1. Definition of 2D Ising Model Outline

3 Outline 1. Definition of 2D Ising Model 2. Why is the nearest neighbor zero-field 2D Ising model exactly solvable? (Lars Onsager and Bruria Kaufman, )

4 Outline 1. Definition of 2D Ising Model 2. Why is the nearest neighbor zero-field 2D Ising model exactly solvable? (Lars Onsager and Bruria Kaufman, ) 3. Spontaneous magnetization some interesting history of mathematics

5 Outline 1. Definition of 2D Ising Model 2. Why is the nearest neighbor zero-field 2D Ising model exactly solvable? (Lars Onsager and Bruria Kaufman, ) 3. Spontaneous magnetization some interesting history of mathematics 4. Toeplitz determinants and spin-spin correlation functions

6 Outline 1. Definition of 2D Ising Model 2. Why is the nearest neighbor zero-field 2D Ising model exactly solvable? (Lars Onsager and Bruria Kaufman, ) 3. Spontaneous magnetization some interesting history of mathematics 4. Toeplitz determinants and spin-spin correlation functions 5. Massive scaling limit and connection with Painlevé III

7 Outline 1. Definition of 2D Ising Model 2. Why is the nearest neighbor zero-field 2D Ising model exactly solvable? (Lars Onsager and Bruria Kaufman, ) 3. Spontaneous magnetization some interesting history of mathematics 4. Toeplitz determinants and spin-spin correlation functions 5. Massive scaling limit and connection with Painlevé III 6. The Ising susceptibility and the natural boundary conjecture Only #6 reports on new developments joint with Harold Widom

8 Consider the square lattice Z 2. A spin configuration σ is an assignment of ±1 to each site (i, j) Z 2 ; that is, σ { 1, 1} Z2.

9 Consider the square lattice Z 2. A spin configuration σ is an assignment of ±1 to each site (i, j) Z 2 ; that is, σ { 1, 1} Z2. The energy of a configuration σ in box Λ is E Λ (σ) = J 1 σ i j σ i j+1 J 2 σ i j σ i+1 j h i,j Λ i,j Λ i,j Λ σ i j

10 Consider the square lattice Z 2. A spin configuration σ is an assignment of ±1 to each site (i, j) Z 2 ; that is, σ { 1, 1} Z2. The energy of a configuration σ in box Λ is E Λ (σ) = J 1 σ i j σ i j+1 J 2 σ i j σ i+1 j h i,j Λ i,j Λ i,j Λ The Gibbs measure gives the probability of configuration σ in box Λ at inverse temperature β: P Λ (σ) := exp( βe(σ)), Z Λ is called the partition function Z Λ (β, h) σ i j

11 Consider the square lattice Z 2. A spin configuration σ is an assignment of ±1 to each site (i, j) Z 2 ; that is, σ { 1, 1} Z2. The energy of a configuration σ in box Λ is E Λ (σ) = J 1 σ i j σ i j+1 J 2 σ i j σ i+1 j h i,j Λ i,j Λ i,j Λ The Gibbs measure gives the probability of configuration σ in box Λ at inverse temperature β: Remarks: P Λ (σ) := exp( βe(σ)), Z Λ is called the partition function Z Λ (β, h) We assume J i > 0 so the system is ferromagnetic, i.e. like spins favored. σ i j

12 Consider the square lattice Z 2. A spin configuration σ is an assignment of ±1 to each site (i, j) Z 2 ; that is, σ { 1, 1} Z2. The energy of a configuration σ in box Λ is E Λ (σ) = J 1 σ i j σ i j+1 J 2 σ i j σ i+1 j h i,j Λ i,j Λ i,j Λ The Gibbs measure gives the probability of configuration σ in box Λ at inverse temperature β: Remarks: P Λ (σ) := exp( βe(σ)), Z Λ is called the partition function Z Λ (β, h) We assume J i > 0 so the system is ferromagnetic, i.e. like spins favored. The coefficient h gives the coupling of the system to an external magnetic field. σ i j

13 Consider the square lattice Z 2. A spin configuration σ is an assignment of ±1 to each site (i, j) Z 2 ; that is, σ { 1, 1} Z2. The energy of a configuration σ in box Λ is E Λ (σ) = J 1 σ i j σ i j+1 J 2 σ i j σ i+1 j h i,j Λ i,j Λ i,j Λ The Gibbs measure gives the probability of configuration σ in box Λ at inverse temperature β: Remarks: P Λ (σ) := exp( βe(σ)), Z Λ is called the partition function Z Λ (β, h) We assume J i > 0 so the system is ferromagnetic, i.e. like spins favored. The coefficient h gives the coupling of the system to an external magnetic field. In practice we will take periodic boundary conditions which means the system is defined on a torus with m rows and n columns. σ i j

14 Consider the square lattice Z 2. A spin configuration σ is an assignment of ±1 to each site (i, j) Z 2 ; that is, σ { 1, 1} Z2. The energy of a configuration σ in box Λ is E Λ (σ) = J 1 σ i j σ i j+1 J 2 σ i j σ i+1 j h i,j Λ i,j Λ i,j Λ The Gibbs measure gives the probability of configuration σ in box Λ at inverse temperature β: Remarks: P Λ (σ) := exp( βe(σ)), Z Λ is called the partition function Z Λ (β, h) We assume J i > 0 so the system is ferromagnetic, i.e. like spins favored. The coefficient h gives the coupling of the system to an external magnetic field. In practice we will take periodic boundary conditions which means the system is defined on a torus with m rows and n columns. This defines the 2D Ising model with nearest neighbor interactions on the square lattice in a magnetic field. σ i j

15 Thermodynamic quantities and correlation functions Free energy per lattice site βf (β, h) = lim Λ 1 Λ log Z Λ(β, h)

16 Thermodynamic quantities and correlation functions Free energy per lattice site βf (β, h) = lim Λ 1 Λ log Z Λ(β, h) Magnetization and spontaneous magnetization M(β, h) := f h, M 0(β) := lim M(β, h) h 0 +

17 Thermodynamic quantities and correlation functions Free energy per lattice site βf (β, h) = lim Λ 1 Λ log Z Λ(β, h) Magnetization and spontaneous magnetization M(β, h) := f h, M 0(β) := lim M(β, h) h 0 + Susceptibility χ(β, h) = M (β, h) h

18 Thermodynamic quantities and correlation functions Free energy per lattice site βf (β, h) = lim Λ 1 Λ log Z Λ(β, h) Magnetization and spontaneous magnetization M(β, h) := f h, M 0(β) := lim M(β, h) h 0 + Susceptibility χ(β, h) = M (β, h) h Spin-spin correlation function σ 00 σ MN = lim Λ E Λ (σ 00 σ MN ) = lim Λ σ σ 00σ MN e βe(σ) σ e βe(σ)

19 Why is the zero-field h = 0 2D Ising model exactly solvable? Or stated differently, what s the problem for h 0?

20 Why is the zero-field h = 0 2D Ising model exactly solvable? Or stated differently, what s the problem for h 0? Since we don t know f (β, h) as a function of h, how do we find M 0 and χ(β, 0)?

21 Why is the zero-field h = 0 2D Ising model exactly solvable? Or stated differently, what s the problem for h 0? Since we don t know f (β, h) as a function of h, how do we find M 0 and χ(β, 0)? Second question a bit easier to answer: zero-field susceptibility: χ(β) = M 2 0 (β) = lim N σ 00σ NN M,N Z [ σ00 σ MN M 2 0 ]

22 Method of Transfer Matrices On a torus of m rows and n columns can write where V is a 2 n 2 n matrix. Z mn (β, h) = Tr(V m ) ( )

23 Method of Transfer Matrices On a torus of m rows and n columns can write where V is a 2 n 2 n matrix. Z mn (β, h) = Tr(V m ) To see this let s α represent the configuration of row α, s α = (s 1, s 2,..., s n ), s j = ±1. Define the 2 n 2 n matrix V 1 by incorporating the Boltzmann factors in row α: n V 1 ( s, s ) = δ s, s e βj 1s αs α+1 α=1 For Boltzmann factors on columns we introduce n V 2 ( s, s ) = and for the magnetic field j=1 V 3 ( s, s ) = δ s, s e βj 2s j s j n j=1 e βhs j ( )

24 Defining V = V 1 V 2 V 3 it is easy to check that Z mn (β, h) = Tr(V m ) ( ) Thus the problem reduces to the spectral theory of V ; or more precisely, the largest eigenvalue of V in computing f (β, h) in the thermodynamic limit. For h 0 this is an open problem.

25 Defining V = V 1 V 2 V 3 it is easy to check that Z mn (β, h) = Tr(V m ) ( ) Thus the problem reduces to the spectral theory of V ; or more precisely, the largest eigenvalue of V in computing f (β, h) in the thermodynamic limit. For h 0 this is an open problem. For h = 0; namely, V 3 = I, a diagonalization was first accomplished by Lars Onsager (1944). His analysis was subsequently simplified by Bruria Kaufman (1949). It is Kaufman s point of view which we now summarize.

26 Kaufman s Analysis Stated concisely here is what Kaufman realized: The 2 n 2 n matrices V 1 and V 2 ; and hence V = V 1 V 2, are spin representations of rotations in the orthogonal group O(2n). Furthermore, V is a spin representative of a product of commuting plane rotations. Thus the spectral analysis is reduced to that of a 2n 2n orthogonal matrix. Indeed, due to the translational invariance of the interaction energy, the spectral theory reduces to solving quadratic equations! 1 V is not a spin-representative of a rotation when h 0. 1 Actually, this statement is true for a certain direct sum decomposition V = V + V. The statements apply to V ±.

27 Combinatorial Approach to 2D Ising Model Kasteleyn s theory (1963) of dimers on planar lattices: Partition function is expressible as a Pfaffian. 2 Fisher (1966) building on work of Kac and Ward (1952) showed the 2D Ising model (as defined above) is equivalent to a dimer problem. Figure: A six-site cluster that may be used to convert the Ising problem into a dimer problem. See McCoy & Wu, The Two-Dimensional Ising Model for details. 2 For modern treatment see work of Rick Kenyon.

28 The Spontaneous Magnetization: Some history 3 Onsager, well-known for being cryptic, announced in a discussion section at a conference in Florence (1949) 4 that he and Kaufman had recently obtained an exact formula for the spontaneous magnetization: M 0 = (1 k 2 ) 1/8, k := (sinh 2βJ 1 sinh 2βJ 2 ) 1 ( ) Onsager gave no details to how he and Kaufman obtained ( ). Figure: Spontaneous magnetization. k = 1 defines the critical temperature. 3 See, P. Deift, A. Its and I. Krasovsky, Toeplitz Matrices and Toeplitz Determinants under the Impetus of the Ising Model: Some History and Some Recent Results. 4 And on a blackboard at Cornell on 23 August 1948.

29 What was later revealed was that Onsager and Kaufman first showed that σ 00 σ 0N was expressible as a Toeplitz determinant. Unpublished, they also showed the diagonal correlation is a Toeplitz determinant and the expression is a bit simpler:

30 What was later revealed was that Onsager and Kaufman first showed that σ 00 σ 0N was expressible as a Toeplitz determinant. Unpublished, they also showed the diagonal correlation is a Toeplitz determinant and the expression is a bit simpler: σ 00 σ NN = det (ϕ m n ) m,n=0,...,n 1 with ϕ m = 1 2π e i mθ ϕ(e i θ ) dθ, ϕ(z) = 2π 0 [ ] 1 k/z 1/2 1 kz

31 What was later revealed was that Onsager and Kaufman first showed that σ 00 σ 0N was expressible as a Toeplitz determinant. Unpublished, they also showed the diagonal correlation is a Toeplitz determinant and the expression is a bit simpler: with σ 00 σ NN = det (ϕ m n ) m,n=0,...,n 1 ϕ m = 1 2π e i mθ ϕ(e i θ ) dθ, ϕ(z) = 2π 0 [ ] 1 k/z 1/2 1 kz Today we know the strong Szegö limit theorem (plus some conditions on ϕ) ( ) det(ϕ m n ) lim N µ N = exp k(log ϕ) k (log ϕ) k k=1 and a simple application gives M 2 0.

32 What was later revealed was that Onsager and Kaufman first showed that σ 00 σ 0N was expressible as a Toeplitz determinant. Unpublished, they also showed the diagonal correlation is a Toeplitz determinant and the expression is a bit simpler: with σ 00 σ NN = det (ϕ m n ) m,n=0,...,n 1 ϕ m = 1 2π e i mθ ϕ(e i θ ) dθ, ϕ(z) = 2π 0 [ ] 1 k/z 1/2 1 kz Today we know the strong Szegö limit theorem (plus some conditions on ϕ) ( ) det(ϕ m n ) lim N µ N = exp k(log ϕ) k (log ϕ) k k=1 and a simple application gives M 2 0. But Szegö had not yet proved his strong limit theorem (1952)!

33 Onsager had, in fact, derived (nonrigorously) the limit formula. He communicated this result to Shizuo Kakutani who communicated the result to Szegö. (I verified this story when I met, many years later, Kakutani at Yale.)

34 Onsager had, in fact, derived (nonrigorously) the limit formula. He communicated this result to Shizuo Kakutani who communicated the result to Szegö. (I verified this story when I met, many years later, Kakutani at Yale.) Before all this was revealed(!), C. N. Yang (1952) gave an independent derivation which is a subtle perturbation argument.

35 Onsager had, in fact, derived (nonrigorously) the limit formula. He communicated this result to Shizuo Kakutani who communicated the result to Szegö. (I verified this story when I met, many years later, Kakutani at Yale.) Before all this was revealed(!), C. N. Yang (1952) gave an independent derivation which is a subtle perturbation argument. To quote Yang from his Selected Works : I was thus led to a long calculation, the longest in my career. Full of local, tactical tricks, the calculation proceeded by twists and turns. There were many obstructions. But always, after a few days, a new trick was somehow found that pointed to a new path. The trouble was that I soon felt I was in a maze and was not sure whether in fact, after so many turns, I was anywhere nearer the goal than when I began. This kind of strategic overview was very depressing, and several times I almost gave up. But each time something drew me back, usually a new tactical trick that brightened the scene, even though only locally. Finally, after six months of work off and on, all the pieces suddenly fitted together, producing miraculous cancellations, and I was staring at the amazingly simple final result...

36 Spin-spin correlation functions Though both the row and diagonal correlations are expressible as Toeplitz determinants, there is no Toeplitz representation for the general case σ 00 σ MN See, Planar Ising Correlations by John Palmer, Progress in Mathematical Physics,

37 Spin-spin correlation functions Though both the row and diagonal correlations are expressible as Toeplitz determinants, there is no Toeplitz representation for the general case σ 00 σ MN. If we use the Case-Geronimo-Borodin-Okounkov (CGBO) formula that expresses a Toeplitz determinant as a Fredholm determinant (times a normalization constant), we arrive at new representations for these correlations. It is this type of expression that generalizes. For T < T c it takes the form σ 00 σ MN = M 2 0 det(i K MN ) This was first derived by Wu, McCoy, Barouch & CT (1976) and then put on a rigorous footing by Palmer and CT (1981) See, Planar Ising Correlations by John Palmer, Progress in Mathematical Physics,

38 Massive Scaling Functions & Universality Most interest lies for T, the temperature, close to the critical temperature T c. It is in this limit we expect to see universality. 6 We state this for the case J 1 = J 2.

39 Massive Scaling Functions & Universality Most interest lies for T, the temperature, close to the critical temperature T c. It is in this limit we expect to see universality. Precisely, let ξ(t ) denote the correlation length, which is known to diverge at T T ± c, then the massive scaling limit 6 from below T c is F (r) := where lim scaling is σ 00 σ MN lim scaling M0 2 = det(i K < ) M, N, ξ(t ) such that r := M 2 + N 2 /ξ(t ) is fixed. 6 We state this for the case J 1 = J 2.

40 Massive Scaling Functions & Universality Most interest lies for T, the temperature, close to the critical temperature T c. It is in this limit we expect to see universality. Precisely, let ξ(t ) denote the correlation length, which is known to diverge at T T ± c, then the massive scaling limit 6 from below T c is F (r) := where lim scaling is σ 00 σ MN lim scaling M0 2 = det(i K < ) M, N, ξ(t ) such that r := M 2 + N 2 /ξ(t ) is fixed. Somewhat similar formula exists for F + (r). 6 We state this for the case J 1 = J 2.

41 Massive Scaling Functions & Universality Most interest lies for T, the temperature, close to the critical temperature T c. It is in this limit we expect to see universality. Precisely, let ξ(t ) denote the correlation length, which is known to diverge at T T ± c, then the massive scaling limit 6 from below T c is F (r) := where lim scaling is σ 00 σ MN lim scaling M0 2 = det(i K < ) M, N, ξ(t ) such that r := M 2 + N 2 /ξ(t ) is fixed. Somewhat similar formula exists for F + (r). Note that the scaling functions are rotationally invariant. 6 We state this for the case J 1 = J 2.

42 Massive Scaling Functions & Universality Most interest lies for T, the temperature, close to the critical temperature T c. It is in this limit we expect to see universality. Precisely, let ξ(t ) denote the correlation length, which is known to diverge at T T ± c, then the massive scaling limit 6 from below T c is F (r) := where lim scaling is σ 00 σ MN lim scaling M0 2 = det(i K < ) M, N, ξ(t ) such that r := M 2 + N 2 /ξ(t ) is fixed. Somewhat similar formula exists for F + (r). Note that the scaling functions are rotationally invariant. The scaling functions F ± are expected to be universal for a large class of 2D ferromagnetic systems. There is no proof (as far as I know), but it is generally accepted by physicists using nonrigorous renormalization group arguments. 6 We state this for the case J 1 = J 2.

43 Connection with integrable differential equations, Painlevé III There are alternative expressions for F ± (WMTB, 1976; MTW, 1977):

44 Connection with integrable differential equations, Painlevé III There are alternative expressions for F ± (WMTB, 1976; MTW, 1977): ( [ 1 ( ) ] } dψ 2 F (r) = cosh ψ(r)/2 exp m 2 sinh 2 ψ(x) x dx 4 dx F + (r) = (tanh ψ(r)/2) F (r) r where d 2 ψ dr r dψ dr = 1 2 sinh(2ψ), ψ(r) 2K 0(r), r.

45 Connection with integrable differential equations, Painlevé III There are alternative expressions for F ± (WMTB, 1976; MTW, 1977): ( [ 1 ( ) ] } dψ 2 F (r) = cosh ψ(r)/2 exp m 2 sinh 2 ψ(x) x dx 4 dx F + (r) = (tanh ψ(r)/2) F (r) r where d 2 ψ dr r dψ dr = 1 2 sinh(2ψ), ψ(r) 2K 0(r), r. η(r) = e ψ(r) is a Painlevé III function.

46 Connection with integrable differential equations, Painlevé III There are alternative expressions for F ± (WMTB, 1976; MTW, 1977): ( [ 1 ( ) ] } dψ 2 F (r) = cosh ψ(r)/2 exp m 2 sinh 2 ψ(x) x dx 4 dx F + (r) = (tanh ψ(r)/2) F (r) r where d 2 ψ dr r dψ dr = 1 2 sinh(2ψ), ψ(r) 2K 0(r), r. η(r) = e ψ(r) is a Painlevé III function. M. Sato, T. Miwa & M. Jimbo ( ) gave an isomonodromy deformation analysis interpretation for the appearance of Painlevé III in the 2D Ising model. They also derived a total system of PDEs for the n-point scaling functions. The asymptotic analysis of these PDEs is an open problem.

47 The problem of the Ising susceptibility The zero-field susceptibility χ(t ) is defined by M(T, H) χ(t ) := H H=0 +. To distinguish between T < T c and T > T c we write χ and χ +, respectively.

48 The problem of the Ising susceptibility The zero-field susceptibility χ(t ) is defined by M(T, H) χ(t ) := H H=0 +. To distinguish between T < T c and T > T c we write χ and χ +, respectively. Since and we study β 1 χ(t ) = M,N Z [ σ00 σ MN M 2 ] 0, σ 00 σ MN = M 2 0 det(i K MN ), T < T c, M,N Z But first some history of the problem [det(i K MN ) 1]

49 M. Fisher in 1959 initiated the analysis of the analytic structure of χ near the critical temperature T c by relating it to the long-distance asymptotics of the correlation function at T c (a result known to Kaufmann and Onsager).

50 M. Fisher in 1959 initiated the analysis of the analytic structure of χ near the critical temperature T c by relating it to the long-distance asymptotics of the correlation function at T c (a result known to Kaufmann and Onsager). WMTB ( ) derived the exact form factor expansion of χ and related the the coefficients C ± in the asymptotic expansion ( χ ± (T ) = C ± 1 T /T c 7/4 + O 1 T /T c 3/4), T T c ± to integrals involving a Painlevé III function.

51 M. Fisher in 1959 initiated the analysis of the analytic structure of χ near the critical temperature T c by relating it to the long-distance asymptotics of the correlation function at T c (a result known to Kaufmann and Onsager). WMTB ( ) derived the exact form factor expansion of χ and related the the coefficients C ± in the asymptotic expansion ( χ ± (T ) = C ± 1 T /T c 7/4 + O 1 T /T c 3/4), T T c ± to integrals involving a Painlevé III function. The analysis of χ as a function of the complex variable T was initiated by Guttmann and Enting (1996). By use of high-temperature expansions they conjectured that χ + (T ) has a natural boundary in the complex T -plane.

52 M. Fisher in 1959 initiated the analysis of the analytic structure of χ near the critical temperature T c by relating it to the long-distance asymptotics of the correlation function at T c (a result known to Kaufmann and Onsager). WMTB ( ) derived the exact form factor expansion of χ and related the the coefficients C ± in the asymptotic expansion ( χ ± (T ) = C ± 1 T /T c 7/4 + O 1 T /T c 3/4), T T c ± to integrals involving a Painlevé III function. The analysis of χ as a function of the complex variable T was initiated by Guttmann and Enting (1996). By use of high-temperature expansions they conjectured that χ + (T ) has a natural boundary in the complex T -plane. B. Nickel (1999,2000) analyzed the n-dimensional integrals appearing in the form factor expansion and identified a class of complex singularities, now called Nickel singularities, that lie on a curve and which become ever more dense with increasing n. This provides very strong support for the existence of a natural boundary for χ ± curve is k = 1.

53 Orrick, Nickel, Guttmann & Perk (2001) and Chan, Guttmann, Nickel & Perk (2011) on the basis of high- and low-temperature expansions (300+ terms!) give the following conjecture for the critical point behavior:

54 Orrick, Nickel, Guttmann & Perk (2001) and Chan, Guttmann, Nickel & Perk (2011) on the basis of high- and low-temperature expansions (300+ terms!) give the following conjecture for the critical point behavior: Set τ = 1 2 ( k 1 k) χ ± := k B T χ ± = C 0,± τ 7/4 F ± + B

55 Orrick, Nickel, Guttmann & Perk (2001) and Chan, Guttmann, Nickel & Perk (2011) on the basis of high- and low-temperature expansions (300+ terms!) give the following conjecture for the critical point behavior: Set τ = 1 2 ( k 1 k) χ ± := k B T χ ± = C 0,± τ 7/4 F ± + B where B is of the form ( short-distance terms) B = q q=0 p=0 b (p,q) (log τ ) p τ q with numerical approximations to b (p,q), p 5,

56 Orrick, Nickel, Guttmann & Perk (2001) and Chan, Guttmann, Nickel & Perk (2011) on the basis of high- and low-temperature expansions (300+ terms!) give the following conjecture for the critical point behavior: Set τ = 1 2 ( k 1 k) χ ± := k B T χ ± = C 0,± τ 7/4 F ± + B where B is of the form ( short-distance terms) B = q q=0 p=0 b (p,q) (log τ ) p τ q with numerical approximations to b (p,q), p 5, and F ± =k 1 4 [1 + τ 2 2 τ 4 ( C ) ( 6± τ C ) 6± τ ( C 6± 793C ) ] 10± τ with high-precision decimal estimates for C 6± and C 10±.

57 Diagonal susceptibility: A simpler problem 7 From now on we restrict to T < T c : βχ d = [ σ00 σ NN M 2 ] 0 N Z 7 First introduced by Assis, Boukraa, Hassani, van Hoeiji, Maillard & McCoy (2012).

58 Diagonal susceptibility: A simpler problem 7 From now on we restrict to T < T c : βχ d = [ σ00 σ NN M 2 ] 0 N Z σ 00 σ NN = det(t N (ϕ)) GCBO = M 2 0 det(i K N ) K N = H N (Λ)H N (Λ 1 ), Λ(ξ) = ϕ (ξ)/ϕ + (ξ) = (1 kξ)(1 k/ξ). Here ϕ = ϕ + ϕ is the Wiener-Hopf factorization of ϕ and H N (ψ) is the Hankel operator with entries (ψ N+i+j+1 ) i,j 0. Sum to be analyzed: S := [det(i K N ) 1]. N=1 7 First introduced by Assis, Boukraa, Hassani, van Hoeiji, Maillard & McCoy (2012).

59 Proposition (T Widom): Let H N (du) and H N (dv) be two Hankel matrices acting on l 2 (Z + ) with i, j entries x N+i+j du(x), y N+i+j dv(y) respectively, where u and v are measures supported inside the unit circle. Set K N = H N (du)h N (dv). Then ( 1) n i [det(i K N ) 1] = (n!) 2 x iy i 1 N=1 n=1 i x iy i ( ( )) 1 2 det du(x i )dv(y i ) 1 x i y j i

60 det(φj (x k )) j,k=1,...,n det(ψ j (x k )) j,k=1,...,n dx 1 dx N =N! det( φ j (x)ψ k (x) dx) j,k=1,...,n Proposition (T Widom): Let H N (du) and H N (dv) be two Hankel matrices acting on l 2 (Z + ) with i, j entries x N+i+j du(x), y N+i+j dv(y) respectively, where u and v are measures supported inside the unit circle. Set K N = H N (du)h N (dv). Then ( 1) n i [det(i K N ) 1] = (n!) 2 x iy i 1 N=1 n=1 i x iy i ( ( )) 1 2 det du(x i )dv(y i ) 1 x i y j Ideas in proof: 1. First use Fredholm expansion of det(i K N ). 2. Then use the Andréief identity: i 3. Symmetrization argument

61 Applying this to K N = H N (Λ)H N (Λ 1 ) followed by a deformation of contours gives S = n=1 S n with 1 κ 2n 1 1 i S n = (n!) 2 π 2n x ( ( )) iy i κ n 1 2 i x det iy i 1 κx i y j Λ 1 (x i ) dx i dy i Λ 1 (y i ) i i where κ := k 2 and Λ 1 (x) = (1 x)(1 κx) x

62 Applying this to K N = H N (Λ)H N (Λ 1 ) followed by a deformation of contours gives S = n=1 S n with 1 κ 2n 1 1 i S n = (n!) 2 π 2n x ( ( )) iy i κ n 1 2 i x det iy i 1 κx i y j Λ 1 (x i ) dx i dy i Λ 1 (y i ) where i i κ := k 2 and Λ 1 (x) = (1 x)(1 κx) Alternate representation of S n (Cauchy determinant identity): 1 κ n(n+1) (n!) 2 π 2n i x iy i 1 κ n (x) 2 (y) i x 2 iy i x i,j (1 κx iy j ) 2 i Λ 1 (x i ) Λ 1 (y i ) dx idy i

63 Applying this to K N = H N (Λ)H N (Λ 1 ) followed by a deformation of contours gives S = n=1 S n with 1 κ 2n 1 1 i S n = (n!) 2 π 2n x ( ( )) iy i κ n 1 2 i x det iy i 1 κx i y j Λ 1 (x i ) dx i dy i Λ 1 (y i ) where i i κ := k 2 and Λ 1 (x) = (1 x)(1 κx) Alternate representation of S n (Cauchy determinant identity): 1 κ n(n+1) (n!) 2 π 2n i x iy i 1 κ n (x) 2 (y) i x 2 iy i x i,j (1 κx iy j ) 2 i Λ 1 (x i ) Λ 1 (y i ) dx idy i Theorem (T Widom): The unit circle κ = 1 is a natural boundary for S.

64 Applying this to K N = H N (Λ)H N (Λ 1 ) followed by a deformation of contours gives S = n=1 S n with 1 κ 2n 1 1 i S n = (n!) 2 π 2n x ( ( )) iy i κ n 1 2 i x det iy i 1 κx i y j Λ 1 (x i ) dx i dy i Λ 1 (y i ) where i i κ := k 2 and Λ 1 (x) = (1 x)(1 κx) Alternate representation of S n (Cauchy determinant identity): 1 κ n(n+1) (n!) 2 π 2n i x iy i 1 κ n (x) 2 (y) i x 2 iy i x i,j (1 κx iy j ) 2 i Λ 1 (x i ) Λ 1 (y i ) dx idy i Theorem (T Widom): The unit circle κ = 1 is a natural boundary for S. Proof proceeds by four lemmas.

65 Let ɛ 1 be a nth root of unity and we wish to consider behavior of S as κ ɛ radially. Look at d l S n /dκ l main contribution will come from i x iy i (1 κ n (x) 2 (y) i x 2 iy i ) l+1 i,j (1 κx iy j ) 2 i Λ 1 (x i ) Λ 1 (y i ) dx idy i

66 Let ɛ 1 be a nth root of unity and we wish to consider behavior of S as κ ɛ radially. Look at d l S n /dκ l main contribution will come from i x iy i (1 κ n (x) 2 (y) i x 2 iy i ) l+1 i,j (1 κx iy j ) 2 i Λ 1 (x i ) Λ 1 (y i ) dx idy i Lemma 1: The integral is bounded when l < 2n 2 1 and it is of order log(1 κ ) 1 when l = 2n 2 1. Main idea: Various approximations bound by the integral 2nδ which will give first part of lemma. 0 r 2n2 l 2 dr

67 Let ɛ 1 be a nth root of unity and we wish to consider behavior of S as κ ɛ radially. Look at d l S n /dκ l main contribution will come from i x iy i (1 κ n (x) 2 (y) i x 2 iy i ) l+1 i,j (1 κx iy j ) 2 i Λ 1 (x i ) Λ 1 (y i ) dx idy i Lemma 1: The integral is bounded when l < 2n 2 1 and it is of order log(1 κ ) 1 when l = 2n 2 1. Main idea: Various approximations bound by the integral 2nδ which will give first part of lemma. Lemma 2: 0 r 2n2 l 2 dr ( d dκ )2n2 1 S n log(1 κ) 1 In differentiating the other terms are O(1).

68 Lemma 3: If ɛ m 1, then All integrals are bounded as κ ɛ. ( ) d 2n 2 1 S m = O(1). dκ

69 Lemma 3: If ɛ m 1, then All integrals are bounded as κ ɛ. Lemma 4: (Main lemma) ( ) d 2n 2 1 S m = O(1). dκ m>n ( ) d 2n 2 1 S m = O(1) dκ For κ sufficiently close to ɛ, all integrals we get by differentiating the integrals for S m are at most A m m m. (A can depend upon n but not on m.) The extra (m!) 2 appearing in gives a bounded sum.

70 For T < T c : β 1 χ = M,N Z = M 2 0 What about χ? [ σ00 σ MN M 2 0] = M 2 0 n=1 χ (2n) M,N Z [det(i K M,N ) 1] In the last equality we used the Fredholm expansion and product formulas for the determinants appearing in the integrands to find

71 For T < T c : β 1 χ = M,N Z = M 2 0 What about χ? [ σ00 σ MN M 2 0] = M 2 0 n=1 χ (2n) M,N Z [det(i K M,N ) 1] In the last equality we used the Fredholm expansion and product formulas for the determinants appearing in the integrands to find χ (n) (s) = 1 1 (1 + j n! (2π i) 2n x 1 j )(1 + j y 1 j ) C r C r (1 j x j)(1 j y j) (x j x k )(y j y k ) dx j dy j (x j x k 1)(y j y k 1) D(x j, y j ; s) j<k j D(x, y; s) = s + s 1 (x + x 1 )/2 (y + y 1 )/2, s := 1/ k

72 Nickel Singularities Standard estimates show that χ (s) is holomorphic for s > 1 ( k < 1). Definition: A Nickel singularity of order n is a point s 0 on the unit circle such that the real part of s 0 is the average of the real parts of two nth roots of unity. Note that D(x, y; s) vanishes when R(s) = R(x) + R(y) 2 Theorem (T Widom) When n is even χ (n) extends to a C function on the unit circle except at the Nickel singularities of order n.

73 Some ideas in the proof A partition of unity allows one to localize.

74 Some ideas in the proof A partition of unity allows one to localize. Each potential singular factor in the integrand of χ (n) is represented as an exponential integral over R +.

75 Some ideas in the proof A partition of unity allows one to localize. Each potential singular factor in the integrand of χ (n) is represented as an exponential integral over R +. The gradient of the exponent in the resulting integrand is approximately a linear combination with positive coefficients of certain vectors one from each factor.

76 Some ideas in the proof A partition of unity allows one to localize. Each potential singular factor in the integrand of χ (n) is represented as an exponential integral over R +. The gradient of the exponent in the resulting integrand is approximately a linear combination with positive coefficients of certain vectors one from each factor. Unless s is a Nickel singularity the convex hull of these vectors does not contain 0 this allows a lower bound for the length of the gradient.

77 Some ideas in the proof A partition of unity allows one to localize. Each potential singular factor in the integrand of χ (n) is represented as an exponential integral over R +. The gradient of the exponent in the resulting integrand is approximately a linear combination with positive coefficients of certain vectors one from each factor. Unless s is a Nickel singularity the convex hull of these vectors does not contain 0 this allows a lower bound for the length of the gradient. Application of the divergence theorem gives the bound O(1) same bound after differentiating with with respect to s any number of times.

78 Some ideas in the proof A partition of unity allows one to localize. Each potential singular factor in the integrand of χ (n) is represented as an exponential integral over R +. The gradient of the exponent in the resulting integrand is approximately a linear combination with positive coefficients of certain vectors one from each factor. Unless s is a Nickel singularity the convex hull of these vectors does not contain 0 this allows a lower bound for the length of the gradient. Application of the divergence theorem gives the bound O(1) same bound after differentiating with with respect to s any number of times. Follows that χ (n) extends to a C function excluding the Nickel singularities.

79 Remarks: What we don t have is a Lemma 4 that says the sum of the other terms don t cancel the Nickel singularities.

80 Remarks: What we don t have is a Lemma 4 that says the sum of the other terms don t cancel the Nickel singularities. It appears that χ (n) satisfies a linear differential equation with only regular singular points. (This follows from work of Kashiwara.) Using this result we get that for n even χ (n) extends analytically across the unit circle except at the Nickel singularities.

81 Remarks: What we don t have is a Lemma 4 that says the sum of the other terms don t cancel the Nickel singularities. It appears that χ (n) satisfies a linear differential equation with only regular singular points. (This follows from work of Kashiwara.) Using this result we get that for n even χ (n) extends analytically across the unit circle except at the Nickel singularities. Thank you for your attention!