Stochastic Integral Representation for the Dynamics of Disordered Systems

Ivana Kureciˇ c´1 and Tobias J. Osborne2 1Max Planck Institute of Quantum Optics, Hans-Kopfermann-Str. 1, 85748 Garching, Germany 2Institut fur¨ Theoretische Physik, Leibniz Universitat¨ Hannover, Appelstr. 2, 30167 Hannover, Germany

The dynamics of interacting quantum systems in the presence of disorder is studied and an exact rep- resentation for disorder-averaged quantities via Itoˆ stochastic calculus is obtained. The stochastic integral representation affords many advantages, including amenability to analytic approximation, applicability to interacting systems, and compatibility with existing tensor network methods. The integral may be ex- panded to produce a series of approximations, the first of which already includes all diffusive corrections and, further, is manifestly completely positive. The addition of fluctuations leads to a convergent series of systematic corrections. As examples, expressions for the density of states, spectral form factor, and out-of-time-order correlators for the Anderson model are obtained.

Since Anderson’s disovery of the localisation phe- averaged propagator of an arbitrary quantum system. The nomenon [1], quantum systems in the presence of disor- derivation of this representation is reminiscent, in parts, of der have attracted considerable interest. The localisation field-theoretic approaches to disordered systems and also phenomenon is a result of quantum coherent effects, thus separately to recent calculations [22, 23] of Prosen and presenting many challenges and exhibiting rich physics. coworkers. However, there are crucial differences. For ex- During the past 60 years there has been excellent progress ample, there is no mapping to the nonlinear sigma model. in the study of disordered systems in the single-particle Also, the representation is provably exact and not an ap- regime, culminating in a solid understanding of these phys- proximation. We exploit the integral to produce a series ical phenomena (see, e.g., the recent review [2]). These ad- of approximations, the first of which already includes all vances are the result of several powerful techniques which diffusive corrections and, further, is manifestly the result have been developed to quantitatively study disordered sys- of a completely positive evolution, thus conserving proba- tems, including numerical methods [3], the renormalisa- bilities. The addition of fluctuations leads to a convergent tion group [4], and field-theoretic methods [5] such as the series of systematic corrections. Using this expansion we replica trick [6, 7], the supersymmetric method [8, 9], and compute the density of states and out-of-time-order corre- the Keldysh formalism [10–12]. lations for the Anderson model. With the recent discovery of the intriguing phenomenon Disordered quantum systems, Brownian motions, and of many body localisation (MBL) [13] (see also the recent stochastic calculus.—Here we introduce the systems un- reviews [14–17]), renewed interest in disordered systems der consideration and give a brief notational summary of has emerged. As a collective many particle effect, MBL stochastic calculus. We consider quantum systems with a is proportionally more challenging to study than its single- Hamiltonian of the form m particle counterpart. The most progress in this field has X come from perturbation theory and direct numerical sim- H(x) = H0 + xjDj, (1) ulation via Monte Carlo sampling. In general, these tech- j=1 niques are limited either to weak disorder or finite system where H0 is a fixed Hamiltonian (typically a kinetic energy sizes (of the order of 20 spins), and few samples. Ex- term), D represent disordered terms (e.g., a local mag- ploiting the most powerful field-theoretic approaches to j netic field or potential energy terms), and xj are random study the strongly interacting quantum spin systems ex- variables drawn from the Gaussian distribution [24] with hibiting MBL appears to be a difficult task (for some recent the probability density function

arXiv:1809.04341v1 [quant-ph] 12 Sep 2018 progress, see [18]). 2 − x Some intriguing new directions — emerging from holo- e 2γ2 gγ (x) ≡ √ , graphic arguments in high energy physics [19] — are also 2πγ being explored in the study of the complex dynamics of quantum systems such as the Anderson model [20]. Here a where γ is its standard deviation. This class of models is central role is played by out-of-time-order correlation func- sufficiently general to describe a diverse variety of models tions (OTOCs) [21] as signatures of quantum chaos. The from single impurity models, the Anderson model in arbi- behaviour of OTOCs for complex disordered systems is a trary dimensions, as well as MBL systems. We illustrate central goal in the study of new phases of disordered mat- the results in terms of the one-dimensional Anderson tight- ter. binding model with periodic boundary conditions, where

Motivated by the twin challenges of MBL and calculat- N X ing OTOCs for disordered systems, in this Letter we intro- H0 ≡ 2I − |j + 1ihj| + |jihj + 1|, (2) duce a stochastic integral representation for the disorder- j=1 2

and Dj ≡ |jihj|. We study disorder-averaged dynamical where the product is taken from right to left and δ(f(x)) = quantities, such as the disorder-averaged propagator: δ(x2 −x1) ··· δ(xn −xn−1). The second step is to use the itH identity S(t) = Ex[e ], (3) 1 Z ∞ δ(x) ≡ eikx dk, (9) whose Fourier transform yields the density of states (DOS), 2π the density operator: −∞ and to carry out the integral over the xj variables. This −itH itH ρ(t) = Ex[e |ψ0ihψ0|e ], (4) leaves an integral over k1, k2, . . . , kn−1: 1 Z nγ2 Pn 2 spectral form factors: − 2 j=1(kj−1−kj ) S(t) = 1 e Fe(k) dk, n− 2 (−β−it)H 2k (2π) γ Ex[| tr(e )| ], (5) (10) and OTOCs: where we’ve introduced two additional k variables, k0 = kn = 0, as well as the function Ex[hA(0)B(t)C(0)D(t)i]. (6) n Y These quantities are intimately interrelated (see, e.g., [25]), Fe(k) ≡ Fej(kj−1 − kj), (11) and can all be calculated in terms of S(t) by taking tensor j=1 copies of the Hamiltonian [26]. where Z ∞  √ 2 − √ 1 x −i n γy it We make use of stochastic calculus, but it is not at 2nγ j 2 (A+xj B) Fej(y) = e e n dxj. (12) all necessary to be familiar with this formalism, as all −∞ derivations can be understood with little more than a pass- The third step is to expand the exponents of the operator ex- ing familiarity with Gaussian integrals and a tolerance ponentials in F s, collect terms to O(1/n2), and then carry for lengthy derivations using discretisations [27]. At this e out the integrals over x : stage it is sufficient to comment that a Brownian motion or j Z n Wiener process Wt is characterised by the following four Y F (kj−1 − kj) dµ, (13) properties: (1) W0 = 0; (2) Wt is (almost surely) continu- j=1 ous; (3) Wt has independent increments; and (4) Wt − Ws is distributed according to the normal distribution with a where mean of zero and variance t − s for 0 ≤ s ≤ t. We will it t2 F (y) ≡ + (A + iγ2nyB) − (nγ2 − n2γ4y2)B2 also encounter stochastic differential equations (SDE) of I n 2n2 the form dy = f ds + g dW . At this stage, it is sufficient (14) to regard these as equations that, upon discretisation, define and s new random variables y in terms of W . For further details 2 n−1 2 n(nγ ) − nγ kT Mk on stochastic calculus see, e.g., [28, 29]. dµ ≡ e 2 dk1 ··· dkn−1, (15) (2π)n−1 A stochastic integral representation.— In this section we summarise the salient features of the derivation of our in- with tegral representation, with an emphasis on the physical  2 −1 0 0 ··· 0 foundations of the argument. As explained previously, for −1 2 −1 0 ··· 0 Hamiltonians of the form (1), it is sufficient to restrict our  0 −1 2 −1 0 M =   = 2I − Pn−1. (16) attention to the study of the quantity  . . .   . .. .  itH(x) S(t) ≡ Ex[e ]. (7) 0 ··· 0 −1 2 For simplicity, we focus on Hamiltonians of the form In physical terms, µ is the (discretisation of the) equilib- rium probability distribution for a free particle with the H(x) = A+xB, where x is a single Gaussian-distributed 2 random variable, and A and B are N × N matrices. (The Hamiltonian H = −∇ , which is diffusively moving on extension to Hamiltonians of the form from Eq. (1) is the interval [0, 1] with vanishing Dirichlet boundary condi- entirely straightforward and requires no additional tech- tions. niques.) We derive our representation in four steps. The By employing the approximation itH it 2 t2 2 2 first step is to break the propagator e into n small pieces n (A+inγ (kj−1−kj )B)− 2n γ B n F (y) ≈ e , (17)  it H  e n , introduce n independent Gaussian-distributed valid to O(1/n2), we have already arrived at an approxi- variables xj, j = 1, 2, . . . , n, and enforce equality via mation of great utility in tensor-network simulations (this delta functions: will be the subject of a forthcoming paper): n n Z kxk2 Z 2 1 − Y it (A+x B) Y it A−tγ2(k −k )B− t γ2B2 √ S(t) = √ e 2nγ2 δ(f(x)) e n j dx, S(t) ≈ e n j−1 j 2n dµ+O(1/ n). 2πγ j=1 j=1 (8) (18) 3

The final step is to recognise the integral measure in the The stochastic Dyson series: diffusions and limit of n → ∞ as the path measure for the Brownian fluctuations.—In this section we describe a Dyson se- bridge [28], which is a continuous-time stochastic process ries procedure to develop the integral in a power series with the same conditional probability distribution as the in the disorder parameter γ. This series has several Wiener process, but subject to the condition that at s = 0 extremely desirable features with the following physical and s = 1 it is pinned to 0, i.e., B1 = 0. The Brownian interpretations. The first term in the series already ex- bridge is defined by plicitly incorporates disorder corrections and describes a completely positive evolution with diffusive behaviour. zs = Ws − sWs=1. (19) The subsequent terms incorporate quantum fluctuation corrections around the diffusions. This solution exhibits Note that the increments of the Brownian bridge are not excellent large t behaviour (in contrast to a direct Dyson independent. This allows us to take the continuum limit: series approximation of the propagator followed by a Z disorder average). R 1 K ds+γt R 1 B dz S(t) = T e 0 0 dµ, (20) We focus, again for simplicity, on the simplified case H = A + xB, and make an expansion of the inte- where T is the time-ordering operation, gral Eq. (20) for S(t) in powers of the stochastic term 1 γt R B dz in the exponent. We do this by first defining 2 2 0 γ t sK −sK K = itA − B2, (21) B(s) ≡ e Be and writing 2 Z −K γt R 1 B(s) dz e S(t) = T e 0 dµ(z). (27) and the increment dz obeys the stochastic differential equa- tion: Expanding the exponential with a standard Dyson series z leads to dz = − ds + dW. (22) 1 − s  Z 1 −K e S(t) = E I + γt B(s) dz+ It is important to note that Eq. (20) is an equality — this 0 formula is not an approximation. In this way we have ob- γ2t2 Z 1 Z 1  tained a representation of the operator S via the operator T [B(s1)B(s2)] dz1dz2 + ··· . (28) 2 SDE: 0 0 The next step is to employ the covariance of the Brownian dS = itAS ds + γtBS dz. (23) bridge, E[z(s)z(t)] = min{s(1 − t), t(1 − s)} = Cst, from which we derive E[dz(s)dz(t)] = (δ(s−t)−1)dsdt, This representation may be subjected to a variety of solu- so that, to O(γ2), we have tion and approximation techniques, from direct sampling, Z 1 Z s1 moment expansions, and the Dyson series. These will all −K 2 2 e S(t) = I − γ t ds1 ds2 B(s2)B(s1)+ be the subject of future studies. 0 0 By following the derivation described above we can im- γ2t2 Z 1 mediately write down the stochastic integral representation B(s)2 ds. (29) for Hamiltonians H of the form Eq. (1): 2 0 (We can compute higher-order terms using the classi- Z R 1 Pm R 1 K ds+γt Dj dzj cal Wick’s theorem, e.g., [z(s1)z(s2)z(s3)z(s4)] = S(t) = T e 0 j=1 0 dµ(z), (24) E Cs1s2 Cs3s4 + Cs1s3 Cs2s4 + Cs1s4 Cs2s3 . This will be the subject of a future paper.) where The expansion Eq. (29) admits a very pleasing physical 2 2 m γ t 2 2 2 itA− 2 B γ t X 2 interpretation: the O(1) term S1(t) = e already K = itH0 − Dj , (25) explicitly incorporates the effects of disorder in the form 2 2 2 j=1 − γ t B2 of diffusive corrections e 2 . When we apply this with technique to the disorder-averaged density operator ρ(t), it leads to the expression zj dz = − ds + dW , (26) L j 1 − s j ρ(t) ≈ e [ρ(0)], (30) where and Wj are m independent Brownian motions. γ2t2 An alternative derivation of our representation may be L(X) ≡ it[A, X] − {B,X} + γ2t2BXB (31) found by employing the Lie-Trotter formula eA+B ≈ 2 eAeB + O(k[A, B]k2) and an operator Hubbard- is a generator of Lindblad form, meaning that the evolution Stratonovich transformation. eL is completely positive, hence physical. 4

1 1

0.8 0.8

0.6 0.6 0.4 0.4 0.2

0.2 0

-0.2 0 0 123 4 5 0 1 2 3 4 5

1 itH 1 itH 2 FIG. 1. The quantity X(t) ≡ N E[tr(e )] for the Anderson FIG. 2. The spectral form factor N 2 E[| tr(e )| ] for the An- model on 30 sites (the x axis is time in units where ~ = 1). (It derson model on 30 sites (the x axis is time in units where turns out that the results for 30 sites are already indistinguishable ~ = 1). Again, the results for 30 sites are indistinguishable from the N = ∞ limit.) Shown black is the result of numerical from the N = ∞ limit. Shown black is the result of numer- sampling with 100 samples. Shown blue is X(t) = X0(t) + ical sampling with 100 samples. Shown blue is the zeroth or- X2(t), the second order result calculated via the stochastic Dyson der diffusion-corrected term from the stochastic Dyson series. series Eq. (34). Shown for comparison, in grey, is the zeroth order result from or- dinary time-dependent perturbation theory. Note that the zeroth- order stochastic Dyson series result already incorporates the de- Because the diffusive solution generically supplies an phasing decay resulting from the disorder average. exponential suppression in t, we see that all the sub- sequence fluctuation corrections are exponentially sup- 1 PN 2πi kl √ e N |ki, which have the corresponding eigen- pressed. Further, one can argue that the resulting series N k=1 in O(γ) is actually convergent, in contrast to some field- 2π

γ2t2 values: ωl = i 2 − 2 cos N l t − 2 . Using these theoretic approaches. observations we have (see the supplementary material), in Application 1: the density of states for the Anderson the limit N → ∞: model.—Here we detail the calculations for the Anderson 2 2 2 2it− γ t sin(2t) model. We focus on the quantity (ii) = γ te 2 . (36) 4 1 itH(x) Similarly, to second order in γ, we find: X(t) ≡ E[tr(e )], (32) N 2 2  2 2  2  2it− γ t γ t γ t X(t) = e 2 1 + J (2t) − sin(2t) . whose Fourier transform directly yields the density of 2 0 4 states. Following the previous section, we develop a Dyson (37) series for this representation in powers of γ: Taking the Fourier transform of this solution gives us the

X(t) = X0(t) + X1(t) + X2(t) + ··· , (33) density of states (this is calculated in the Supplementary material). We see that the diffusion correction gives a sim- where ple convolution of the DOS for the tight-binding model 2 2 1 K 2it− γ t with a Gaussian of width γ. The second-order corrections X0(t) = tr(e ) −→ e 2 J0(2t), (34) N N→∞ incorporate the effects of level repulsion. We depict the quantity X(t), calculated to second order, in Fig. 1 instead J0(t) is the Bessel function of the first kind, K ≡ itT − 2 2 of the DOS as it is easier to see the difference between the γ t X (t) = 0 X (t) 2 I, 1 , and 2 is a complicated expression. numerical solution and the approximation in the temporal We can express X2(t) as a sum of two integrals, (i) + (ii), domain. γ2t2 where (i) = 2 X0(t) and Application 2: spectral form factors and out-of-time- order correlation functions.—Here we approximate the N 1 X Z 1 Z s1 k = 1 spectral form factor Eq. (5) and the OTOC Eq. (6) (ii) = γ2t2 ds ds hj|e(1−s1+s2)K |ji× N 1 2 for the Anderson model. This is expedited by noting that j=1 0 0 the spectral form factor may be calculated from S(2)(t), hj|e(s1−s2)K |ji. (35) the average propagator for two copies of the Hamiltonian, H = H ⊗ − ⊗ H . The OTOC is equal to To go further we must diagonalise K. This is achieved (2) 1 I2 I1 2
upon introducing the eigenvectors of K, |Wli ≡ tr A ⊗ B ⊗ C ⊗ DS(4)(t)SWAP1234 , (38) 5 where S(4)(t) is the average propagator for four copies of International School of Physics “Enrico Fermi”, Vol. 151 the Hamiltonian, H(4) = H1 ⊗I234 −I1 ⊗H2 ⊗I34 +I12 ⊗ (IOS Press, The Netherlands, 2003). [7] I. V. Lerner, B. L. Althsuler, V. I. Fal’ko, and T. Gia- H3 ⊗I4 −I123 ⊗H4, and SWAP1234 is the unitary permuta- tion that cycles the four copies. We can now directly apply marchi, eds., Strongly Correlated Fermions and Bosons in Low-Dimensional Disordered Systems, Nato Science Series the integral representation to both S(2)(t) and S(4)(t) and derive the diffusion approximations: II: (Springer, The Netherlands, 2002). [8] K. Efetov, Supersymmetry in disorder and chaos (Cam- 2 2 it(H ) − γ t PN (D )2 bridge University Press, Cambridge, 1997). S (t) = e 0 (2k) 2 j=1 j (2k) . (39) (2k) [9] M. R. Zirnbauer, J. Phys. A 29, 7113 (1996). Choosing the initial local perturbation to be A = C = [10] A. Kamenev and A. Levchenko, Adv. Phys. 58, 197 (2009). |1ih1| and the observation site to be B = D = |`ih`| [11] A. Kamenev and A. Andreev, Phys. Rev. B 60, 2218 (1999). yields [12] A. Kamenev, Field Theory of Non-Equilibrium Systems, 1st ed. (Cambridge University Press, New York, 2011). h0`0`|S(4)(t)|`0`0i (40) [13] D. M. Basko, I. L. Aleiner, and B. L. Altshuler, Ann. Phys. 321, 1126 (2006). for the OTOC. We have depicted the spectral form factor in [14] F. Alet and N. Laflorencie, (2018), arXiv:1711.03145. Fig 2. The OTOC is numerically more intensive and will [15] R. Nandkishore and D. A. Huse, Ann. Rev. Conden. Ma. P. be the subject of a forthcoming study. 6, 15 (2015). Conclusions and further directions.—In this Letter we [16] E. Altman and R. Vosk, Ann. Rev. Conden. Ma. P. 6, 383 have introduced an exact representation for the disorder- (2015). averaged propagator of a quantum system in terms of a [17] D. A. Abanin and Z. Papic,´ Ann. Phys. 529, 1700169 stochastic integral over a time-ordered operator expression (2017). which involves a Lindblad generator and a temporally ran- [18] Y. Liao, A. Levchenko, and M. S. Foster, Ann. Phys. 386, dom external field, with the path measure given by the 97 (2017), arXiv:1706.07066. Brownian bridge. This expression was then expanded in [19] A. Kitaev, “Alexei Kitaev: 2015 Break- through Prize Fundamental Physics Symposium,” a stochastic Dyson series to yield a power series in the dis- Https://www.youtube.com/watch?v=OQ9qN8j7EZI. order parameter whose O(1) term explicitly includes dif- [20] B. Swingle and D. Chowdhury, Phys. Rev. B 95 (2017), fusive disorder effects. Fluctuations around the diffusive arXiv:1608.03280. solution arise from the higher-order terms. The representa- [21] A. I. Larkin and Y. N. Ovchinnikov, Sov. Phys. JETP 28, tion was exploited to calculate the density of states, spectral 1200 (1969). form factor, and OTOC for the Anderson model. Much fur- [22] P. Kos, M. Ljubotina, and T. Prosen, Phys. Rev. X 8, 021062 ther work remains to be done, including investigating the (2018) . higher-order corrections, applying tensor-network numeri- [23] B. Bertini, P. Kos, and T. Prosen, (2018), cal methods, studying moments, and applications to MBL. arXiv:1805.00931. This will be the subject of several future papers. [24] The restriction to a Gaussian-distributed disorder is not fun- Acknowledgments.—This work was supported by the damental and the techniques developed here are easily gen- eralised to any divisible disorder distribution. DFG through SFB 1227 (DQ-mat), the RTG 1991, and the [25] J. Cotler, N. Hunter-Jones, J. Liu, and B. Yoshida, J. High Max Planck Society through the IMPRS-QST. Energy Phys. 2017, 48 (2017). [26] For example, we can calculate ρ(t) from S2(t) = it(H⊗ − ⊗HT ) Ex[e I I ]. [27] The elementary derivation via Gaussian integrals of the main representation is carried out in the supplementary ma- [1] P. W. Anderson, Phys. Rev. 109, 1492 (1958). terial. [2] A. Lagendijk, B. van Tiggelen, and D. S. Wiersma, Phys. [28] B. Øksendal, Stochastic Differential Equations, sixth ed., Today 62, 24 (2009). Universitext (Springer-Verlag, Berlin, 2003). [3] O. Schenk, M. Bollhofer,¨ and R. Romer,¨ SIAM Rev. 50, 91 [29] C. W. Gardiner, Handbook of Stochastic Methods: For (2008). Physics, Chemistry and the Natural Sciences, 2nd ed., [4] E. Abrahams, P. W. Anderson, D. C. Licciardello, and T. V. Springer Series in Synergetics No. 13 (Springer-Verlag, Ramakrishnan, Phys. Rev. Lett. 42, 673 (1979). Berlin, 1997). [5] F. Wegner, Z. Phys. B Con. Mat. 35, 207 (1979). [6] B. Altshuler, V. Tognetti, and A. Tagliacozzo, eds., Quan- tum Phenomena in Mesoscopic Systems, Proceedings of the 6

SUPPLEMENTARY MATERIAL

DISORDER AVERAGE OF THE PROPAGATOR FOR AN ARBITRARY IMPURITY MODEL

In this section we exemplify the general method employed to obtain disorder averages of dynamical processes for disordered systems. The prototype system we consider has a Hamiltonian of the type H(x) = A + xB, (41)

where A and B are D×D matrices and x ∈ R. The parameter x is chosen randomly according to the Gaussian distribution

2 − x e 2γ2 gγ (x) ≡ √ , (42) 2πγ with variance γ2. We want to be able to calculate disorder averages of all kinds of things. To begin, we’ll focus on a simple nontrivial case, namely the disorder-averaged propagator: Z it(A+xB) Sγ (t) ≡ gγ (x)e dx. (43)

This is, generally, hard to do. One approach to take is to sample x many times and form the empirical average. This works well enough for this example, but scales badly as we reach bigger systems. Also, the large-time limit is noisy. We are going to need to complete the square numerous times. To this end, we record the following formula: √ β 2 β2 − αx2 + βx = − αx − √ + . (44) 2 α 4α

Before we describe our approach we will evaluate Sγ (t) in the case where [A, B] = 0, as this is rather instructive; we find 2 2 itA − γ t B2 Sγ (t) = e e 2 . (45) The first step is to break up the evolution into small pieces: Z n  it (A+xB) Sγ (t) ≡ gγ (x) e n dx. (46)

The next step is to introduce n independent variables x1, x2, . . . , xn and enforce equality via delta functions: n ! x2 x2 1 Z 1 n it − 2 − 2 Y (A+xj B) Sγ (t) = √ e 2nγ ··· e 2nγ δ(x2 − x1) ··· δ(xn − xn−1) e n dx1 ··· dxn. (47) 2πγ j=1 We eliminate the delta functions by employing the formula 1 Z ∞ δ(x) ≡ eikx dk, (48) 2π −∞ and find: n ! Z kxk2 1 1 − ik (x −x ) ik (x −x ) Y it (A+x B) S (t) = √ e 2nγ2 e 1 2 1 ··· e n−1 n n−1 e n j dx ··· dx dk ··· dk . (49) γ (2π)n−1 1 n 1 n−1 2πγ j=1

Introducing two auxiliary variables, k0 = kn = 0, we find: n ! Z kxk2 1 1 − ix (k −k ) ix (k −k ) Y it (A+x B) S (t) = √ e 2nγ2 e 1 0 1 ··· e n n−1 n e n j dx ··· dx dk ··· dk . (50) γ (2π)n−1 1 n 1 n−1 2πγ j=1 Consider the expression

2 xj − ix (k −k ) e 2nγ2 e j j−1 j . (51) 7

In the exponent we use: 1 α = , and β = i(k − k ). (52) 2nγ2 j−1 j Complete the square to reexpress it as:

x2  √ 2 j 1 n nγ2 − ix (k −k ) − √ xj −i γ(kj−1−kj ) − (k −k )2 e 2nγ2 e j j−1 j = e 2nγ 2 e 2 j−1 j . (53)

Substitute this value into Sγ (t):

n ! 1 1 Z nγ2 Pn 2 − (kj−1−kj ) Y Sγ (t) = √ e 2 j=1 Fej(kj−1 − kj) dk1 ··· dkn−1, (54) (2π)n−1 2πγ j=1 where Z ∞  √ 2 − √ 1 x −i n γ(k −k ) it 2nγ j 2 j−1 j (A+xj B) Fej(kj−1 − kj) = e e n dxj. (55) −∞

We turn to the calculation of the Fe operators. We expand the values in powers of 1/n and retain only the terms to O(n−1):

∞ 2 Z 1 2 it − 2 (xj −inγ (kj−1−kj )) (A+xj B) Fej(kj−1 − kj) = e 2nγ e n dxj. (56) −∞ Then we introduce

κj ≡ n(kj−1 − kj), (57)

dW (this will later become − ds ), so that

∞ 2 Z 1 2 it − 2 (xj −iγ κj ) (A+xj B) Fej(kj−1 − kj) = e 2nγ e n dxj. (58) −∞ The next step is to expand the exponential in its powers: ∞ 2 Z 1 2 2  it t  − 2 (xj −iγ κj ) 2 F (k − k ) ≈ e 2nγ + (A + x B) − (A + x B) + ··· dx . ej j−1 j I j 2 j j (59) −∞ n 2n The O(1) term is ∞ ∞ Z 1 2 2 Z 1 2 √ − 2 (xj −iγ κj ) − 2 xj e 2nγ dxj = e 2nγ dxj = 2πnγ. (60) −∞ −∞ The first-order contribution is calculated using ∞ ∞ Z 1 2 2 Z 1 2 √ − 2 (xj −iγ κj ) 2 − 2 xj 2 xje 2nγ dxj = (xj + iγ κj)e 2nγ dxj = iγ κj 2πnγ. (61) −∞ −∞ Thus, we find: ∞ it Z 1 2 2 it√ − 2 (xj −iγ κj ) 2
e 2nγ (A + xjB) dxj = 2πnγ A + iγ κjB . (62) n −∞ n The second-order contribution may be calculated using

Z ∞ 2 Z ∞ Z ∞ − 1 x −iγ2κ − 1 x2 − 1 x2 √ 2 2nγ2 ( j j ) 2 2 2nγ2 j 2 2nγ2 j 4 2 xj e dxj = (xj + iγ κj) e dxj = xj e dxj − γ κj 2πnγ −∞ −∞ −∞ √ (63) 2 4 2 = (nγ − γ κj ) 2πnγ. Then we calculate 2 ∞ 2 ∞ t Z 1 2 2 t Z 1 2 2 − 2 (xj −iγ κj ) 2 − 2 (xj −iγ κj ) 2 2 2 − e 2nγ (A + x B) dx = − e 2nγ (A + 2x {A, B} + x B ) dx 2 j j 2 j j j (64) 2n −∞ 2n −∞ 8

2 2 and find that only the xj B term is of the order of t/n: 2 ∞ 2 t Z 1 2 2 t √ − 2 (xj −iγ κj ) 2 2 2 4 2 − e 2nγ (A + x B) dx = − B (nγ − γ κ ) 2πnγ. 2 j j 2 j (65) 2n −∞ 2n Putting this all together, we find: √  2  it 2 t 2 2 4 2 2 Fej(kj−1 − kj) ≈ 2πnγ + (A + iγ n(kj−1 − kj)B) − (nγ − n γ (kj−1 − kj) )B . (66) I n 2n2 √ It is convenient to take out the overall factor 2πnγ and define

Fej(kj−1 − kj) Fj(kj−1 − kj) = √ . (67) 2πnγ

Now that we have the formula for the F and Fe operators, we can put together the normalisations and we find: n ! Z 2 1 1 2 n − nγ Pn (k −k )2 Y S (t) = √ (2πnγ ) 2 e 2 j=1 j−1 j F (k − k ) dk ··· dk . (68) γ (2π)n−1 j j−1 j 1 n−1 2πγ j=1 Let’s work out the normalisation required to make the integral over the ks a probability measure. To this end we write the exponent in matrix form:

2 2 − nγ Pn (k −k )2 − nγ kT Mk e 2 j=1 j−1 j = e 2 , (69)

where k = (k1, . . . , kn−1) and the (n − 1) × (n − 1) matrix M is given by  2 −1 0 0 ··· 0 −1 2 −1 0 ··· 0  0 −1 2 −1 0 M =   = 2I − Pn−1. (70)  . . .   . .. .  0 ··· 0 −1 2

Pn−1 is the adjacency matrix of the path graph. The matrix M can be diagonalised, which gives the eigenvalues πj  λ = 2 − 2 cos , j = 1, 2, . . . , n − 1. (71) j n The corresponding eigenvector is given by 1 πj  1 2πj  1 (n − 1)πj  v = sin , sin ,..., sin . (72) j 2 n 2 n 2 n p (These eigenvectors need to be normalised; the normalisation of vj is (n)/8.) We’ve also gathered, from the Laplace expansion (twice), that det(M) = n. We need the gaussian integral formula

Z r n − 1 xT Ax (2π) e 2 dx ··· dx = . (73) 1 n det A In our case, we have s Z 2 n−1 − nγ kT Mk (2π) e 2 dk ··· dk = , (74) 1 n−1 n(nγ2)n−1 which means that s 2 n−1 2 n(nγ ) − nγ kT Mk dµ ≡ e 2 dk ··· dk (75) (2π)n−1 1 n−1 is a probability measure. 9

Putting this together, we find: s n−1 Z n ! Z n ! 1 1 2 n (2π) Y Y S (t) = √ (2πnγ ) 2 F (k − k ) dµ = F (k − k ) dµ. (76) γ (2π)n−1 n(nγ2)n−1 j j−1 j j j−1 j 2πγ j=1 j=1 The next step is to identify the measure dµ. Our contention is that it is the (discretised) Brownian bridge measure. i t A itA Before we do this, we make a basic consistency check: suppose B = 0; then we have that F ≈ e n , so Sγ (t) = e , as it should. The other case we can directly calculate is for A = 0. Using the exponential form,

it 2 t2 2 2 n (A+inγ (kj−1−kj )B)− 2n γ B Fj(kj−1 − kj) ≈ e , (77) we find, in the case of A = 0,

2 t2 2 2 −tγ (kj−1−kj )B− γ B Fj(kj−1 − kj) ≈ e 2n . (78)

Substituting this into Sγ (t) we find: n ! n Z Z  2 t2 2 2  Y Y −tγ (kj−1−kj )B− γ B Sγ (t) = Fj(kj−1 − kj) dµ = e 2n dµ. (79) j=1 j=1 The sum in the exponential collapses to zero and we are left with

2 2 − γ t B2 Sγ (t) = e 2 , (80) as required. The next stage of our argument is to realise Sγ (t) as the expectation value of the operator n n it 2 t2 2 2 Y Y n (A+inγ (kj−1−kj )B)− 2n γ B Fj(kj−1 − kj) ≈ e ≡ U(k) (81) j=1 j=1 over random paths k ≡ (k0 = 0, k1, . . . , kn−1, kn = 0), sampled according to the path measure dµ. We will eventually identify these paths as coming from the Brownian bridge. This requires several steps; we begin by first discussing the discretisation of Brownian motion and then move onto the bridge. Recall that a Brownian motion or Wiener process Wt is characterised by the following four properties:

1. W0 = 0;

2. Wt is (almost surely) continuous;

3. Wt has independent increments; and

4. Wt − Ws is distributed according to the normal distribution with zero mean and variance t − s for 0 ≤ s ≤ t.

This characterisation provides us with a recipe to approximate Wt: first discretise the interval [0, t] into n subintervals: t t 2t (n−1)t [0, t] = [0, n ) ∪ [ n , n ) ∪ · · · ∪ [ n , t]. (82) Define

∆wj ≡ W jt − W (j−1)t . (83) n n

According to the fourth property, we know that ∆wj is distributed according the normal distribution with the probability distribution function

r 2 n − n x p (x) ≡ e t 2 . (84) t,n 2πt We also have that n X Wt = ∆wj. (85) j=1 The probability distribution function for our discretisation is hence s n (∆w )2 n − n Pn j e t j=1 2 . (86) (2πt)n 10

Pj Let’s change variables from ∆wj to wj ≡ k=1 ∆wk. The Jacobian for this change of variables has the matrix elements: ( ∂wj 1, k ≤ j, [J]jk ≡ = (87) ∂∆wk 0, otherwise, so that J is the lower-triangular matrix with 1s in all the nonzero entries: 1 0 0 ··· 0 1 1 0 ··· 0 J ≡  . . . .  . (88)  . . .. .  1 1 1 ··· 1

The inverse matrix J−1 is given by  1 0 0 ··· 0 −1 1 0 ··· 0 −1   D ≡ J ≡  0 −1 1 ··· 0 . (89)  . . . .   . . .. .  0 0 · · · −1 1

Thus we can express the probability density function for wj as s n T T n − n w D Dw e t 2 . (90) (2πt)n

According to Donsker’s theorem, the continuum limit (N → ∞) of this construction tends (in distribution) to Wt. Note that the matrix DT D is a triangular matrix of the form  2 −1 0 ··· 0  −1 2 −1 ··· 0   0 −1 2 ··· 0  T   D D =  . . . .  . (91)  . . .. .   . . .   0 0 ··· 2 −1 0 0 · · · −1 1

The path measure for kj is, however, slightly different in a crucial way. To understand this measure, we introduce the variables j b ≡ w − w , j = 0, 1, . . . , n. (92) j j n n

These variables are the discretisation of the Brownian bridge Bt, which is a continuous-time stochastic process with the same conditional probability distribution as the Wiener process, but subject to the condition that at t = 1 it is pinned to 0, i.e., B1 = 0. The Brownian bridge is defined by

Bt = Wt − tWt=1. (93) Note that the increments of the Brownian bridge are not independent. These random variables have the property that b0 = bn = 0. In terms of the variables ∆wj we have:

j n X j X b = ∆w − ∆w , j = 0, 1, . . . , n. (94) j k n k k=1 k=1

The Jacobian relating b with ∆wj has matrix elements

( j −1 ∂bj 1 − n , k ≤ j, [Γ ]jk ≡ = j (95) ∂∆wk − n , otherwise. 11

This matrix has the form  1 1 1 1  1 − n − n − n · · · − n 1 − 2 1 − 2 − 2 · · · − 2   n n n n  1 − 3 1 − 3 1 − 3 · · · − 3  −1  n n n n  Γ =   . (96)  . . .. .   . . . .   1 1 1 1   n n n · · · −1 + n  0 0 0 ··· 0 The matrix Γ (for which Γ−1 is the partial left inverse) is then given by  1 0 0 0 ··· 0 −1 1 0 0 ··· 0    0 −1 1 0 ··· 0 Γ ≡  . . . . .  . (97)  ......   . . . .   0 ··· 0 −1 1 0 0 ··· 0 0 −1 0 Since the (n − 1) × (n − 1) matrix M is determined by the (n − 1) × (n − 1) submatrix of ΓT Γ, we have that its inverse M−1 is given by the corresponding (n − 1) × (n − 1) submatrix of Γ−1(Γ−1)T . Calculating the matrix elements, we have: −1  k
j
[M ]j,k = min j 1 − n , k 1 − n . (98) Now that we have a formula for the inverse of M, we are able to calculate moments via the generating function: s 2 n−1 Z 2 n(nγ ) − nγ kT Mk `T k 1 `T M−1` e 2 e dk ··· dk = e 2nγ2 . (99) (2π)n−1 1 n−1 Thus we obtain, e.g., for j ≤ j0: 0 ∂ ∂ 1 T −1 1 1 j  j  2 ` M ` −1 hk k 0 i = e 2nγ = [M ] 0 = 1 − . j j 2 j,j 2 (100) ∂`j ∂`k `=0 nγ γ n n

The continuum limit

Taking the continuum limit will yield a coupled set of stochastic differential equations (SDE). Our starting point is the expression

s n ! 2 n−1 Z 2 n(nγ ) − nγ kT Mk Y S (t) = e 2 F (k − k ) dk ··· dk . (101) γ (2π)n−1 j j−1 j 1 n−1 j=1

We first scale out γ by defining lj = γkj; we obtain:

s n Z n  ! n − n lT Ml Y 1 S (t) = e 2 F (l − l ) dl ··· dl . (102) γ (2π)n−1 j γ j−1 j 1 n−1 j=1

Thus, by substituting for Fj it follows:

s n Z n  2  n − n lT Ml Y it t 2 2 2 2 2 S (t) = e 2 + A + tγ(∆l )B − (nγ − n γ (∆l ) )B dl ··· dl , (103) γ (2π)n−1 I n j 2n2 j 1 n−1 j=1 where

∆lj = lj − lj−1. (104)

As we explained earlier, lj may be identified with a discretisation of the standard Brownian bridge. 12

We define k Y  it t2  X ≡ + A + tγ(∆l )B − (nγ2 − n2γ2(∆l )2)B2 . (105) k I n j 2n2 j j=1 Using this expression we have:  it t2  X = + A + tγ(∆l )B − (nγ2 − n2γ2(∆l )2)B2 X . (106) k+1 I n k 2n2 k k

The difference between Xk+1 − Xk is thus it t2  ∆X ≡ X − X = A + tγ(∆l )B − (nγ2 − n2γ2(∆l )2)B2 X . (107) k k+1 k n k 2n2 k k We define t dz = ∆l (108) s=k/n k n and

Xs=k/n ≡ ∆Xk, (109) and write the increment ∆Xk as dXs in the limit of N → ∞. By putting this together we obtain the system of stochastic differential equations:

dXs = itAXs ds + tγBXs dzs, z (110) dz = − s ds + dW . s 1 − s s This is a consequence of the fact that, in the distribution, 1 1 (∆l )2 = − , (111) k n n2 and the second term is negligible in the limit.

CALCULATIONS FOR THE ANDERSON MODEL

Here we detail the calculations for the Anderson model. We focus on the quantity 1 X(t) ≡ [tr(eitH(x))]. (112) nE By employing the stochastic integral representation we have :

Z 2 2 1  R 1 itT − γ t ds+γt Pn R 1 D dz  X(t) = tr T e 0 2 I j=1 0 j j dµ(z). (113) n By following the previous section, we develop a Dyson series for this representation in powers of γ:

X(t) = X0(t) + X1(t) + X2(t) + ··· , (114) where 1 X (t) = tr(eK ), (115) 0 n with γ2t2 K ≡ itT − , (116) 2 I

X1(t) = 0, (117) 13

and X2(t) is a complicated expression. We can express X2(t) = (i) + (ii) in terms of two integrals: n n γ2t2 1 X Z 1   1 X Z 1 Z s1   X (t) = tr e(1−s)K D esK ds − γ2t2 ds ds tr e(1−s1)K D e(s1−s2)K D es2K . 2 2 n j n 1 2 j j j=1 0 j=1 0 0 (118) The first is simple: n γ2t2 1 X Z 1   γ2t2 tr e(1−s)K D esK ds = X (t). (119) 2 n j 2 0 j=1 0 We can slightly simplify the second integral: n 1 X Z 1 Z s1 (ii) = γ2t2 ds ds hj|e(1−s1+s2)K |jihj|e(s1−s2)K |ji. (120) n 1 2 j=1 0 0 To go further, we must diagonalise K. This is achieved upon introducing the eigenvectors of K, n 1 X 2πi kl |W i ≡ √ e n |ki, (121) l n k=1 which have the corresponding eigenvalues:  2π  γ2t2 ω = i 2 − 2 cos l t − . (122) l n 2 Note that

1 2πi kl hj|W i = √ e n , (123) l n so that 1 hj|W ihW |ji = . (124) l l n Using these observations we have: n Z 1 Z s1 2 2 1 X (1−s +s )ω (s −s )ω = γ t ds ds e 1 2 k1 e 1 2 k2 . (ii) 2 1 2 (125) n 0 0 k1,k2=1 Expanding the exponents: n 2 2 Z 1 Z s1 2 2 2it− γ t 1 X 2it(1−s +s ) cos 2π k 2it(s −s ) cos 2π k = γ t e 2 ds ds e 1 2 ( n 1)e 1 2 ( n 2). (ii) 2 1 2 (126) n 0 0 k1,k2=1 We now take the limit of N → ∞ and define 2π 2π z ≡ k , and z ≡ k , (127) 1 n 1 2 n 2 with n 2π 1 X 1 Z 2π dz2 = dz1 ≈ ; ≈ dz1. (128) n n 2π 0 k1=1 In this way we obtain

2 2 1 s1  2π 2π  γ t Z Z 1 Z 1 Z 2 2 2it− 2it(1−s1+s2) cos(z1) 2it(s1−s2) cos(z2) (ii) = γ t e 2 ds1 ds2 e dz1 × e dz2 . (129) 0 0 2π 0 2π 0

We recognise these integrals as representations for the Bessel function of the first kind J0(x):

2 2 Z 1 Z s1 2 2 2it− γ t (ii) = γ t e 2 ds1 ds2 J0(2t(1 − s1 + s2))J0(2t(s1 − s2)). (130) 0 0 14

Using the series representation, X (−1)l J (x) = x2l, (131) 0 22l(l!)2 l=0 and integrating explicitly, one can show that the double integral evaluates to: Z 1 Z s1 sin(2t) ds1 ds2 J0(2t(1 − s1 + s2))J0(2t(s1 − s2)) = , (132) 0 0 4t so we have

2 2 2 2it− γ t sin(2t) (ii) = γ te 2 . (133) 4

Similarly, for X0(t) we find: n 2 2 2 2 1 2it− γ t X 2it cos 2π k 2it− γ t X (t) = e 2 e ( n ) = e 2 J (2t). (134) 0 n 0 k=1 Thus we have, to second order in γ:

 2 2  2 2 2 2 2 γ t 2it− γ t γ t 2it− γ t X(t) = 1 + e 2 J (2t) − e 2 sin(2t). (135) 2 0 4 Taking the Fourier transform of X(t) gives us the density of states:

2 ! k 1   2 2 √ 2 2 2   2 √ 2 2 2 i γ d 2 − π γ k rect( 2 − 2π ) i d γ 2 − π γ k 2 X(k) = 1 + 2πγ e 2 ? − 2πγ e 2 ?(δ(k− )−δ(k)). b 2 q π 2π 2 dk 2 1 2 2π dk 8i 1 − π (k − π ) (136)