Introduction
Two-dimensional optical lattices split a quantum gas into an array of parallel one-dimensional tubes, the standard setting for one-dimensional Bose gases [1 , 2 ] , from the quantum Newton’s cradle [3 ] to quantitative tests of generalized hydrodynamics [4 ] and the observation of hydrodynamization [5 ] . With magnetic lanthanides [6 , 7 ] the same geometry acquires a long-range dipolar interaction [8 , 9 ] , which breaks integrability by a tunable amount [10 , 11 ] and stabilizes strongly correlated states against collapse [12 ] . The dipolar interaction also acts between tubes. Single-particle tunneling falls exponentially with lattice depth; the dipolar coupling falls only as the inverse cube of the separation. There is therefore a regime in which the tubes exchange energy and momentum but not particles, as observed in dysprosium bilayers through sympathetic cooling and coupled collective oscillations [13 ] . Each of the N t N_t N t tubes then keeps its own particle number and its own condensate phase, a U ( 1 ) N t U(1)^{N_t} U ( 1 ) N t symmetry instead of the single U ( 1 ) U(1) U ( 1 ) of a coupled superfluid.
With tunneling, an array of tubes undergoes a dimensional crossover, or deconfinement transition, into an anisotropic three-dimensional condensate [14 , 15 ] . Without it, the U ( 1 ) N t U(1)^{N_t} U ( 1 ) N t structure is that of the sliding phases of coupled XY models and Luttinger liquids [16 , 17 , 18 , 19 ] , in which interchain couplings leave each chain’s phase free and modify its exponents. For dipolar bosons in a planar array, Kollath, Meyer and Giamarchi obtained the phase diagram of this sliding Luttinger liquid, with stripe and checkerboard density-wave order at strong intra-tube correlations and a collapse instability at strong dipolar coupling, and proposed the transfer of center-of-mass momentum between tubes as an experimental probe [20 ] . Momentum transfer without particle exchange is Coulomb drag in condensed matter [21 ] , including between one-dimensional wires [22 ] ; dipolar drag has been proposed for pairs of pancake-shaped clouds [23 ] and computed for bilayer Fermi gases [24 ] . Summed over an extended array, the same process is the transverse transport of longitudinal momentum, the quantity that a shear viscosity summarizes in a fluid, studied in quantum gases down to dilute Bose condensates [25 ] . Prior work on the dipolar tube array concerns its equilibrium phases, the inter-tube contribution to integrability breaking and thermalization [10 ] , and, in the modeling of experiments, the self-consistent mean-field potential that the array exerts on each tube, the tubes otherwise remaining independent [26 ] . How longitudinal momentum spreads across many tubes in time has, to our knowledge, not been addressed.
Tube occupancies are fixed at loading and vary from tube to tube, through the envelope of the parent cloud and through fluctuations about it [4 , 10 , 26 ] . The collective modes of the array couple through the local densities, so these variations are a quenched random potential for the transverse modes even though the Hamiltonian is uniform. Anderson localization [27 ] of cold atoms has used externally imposed random or quasiperiodic potentials [28 , 29 , 30 ] , and Bogoliubov quasiparticles in such potentials have been studied along the tube axis [31 , 32 , 33 ] ; for bosonic excitations, stability further requires every mode energy to be positive [34 ] . Here the disorder is intrinsic, it acts across the tubes, and it enters as a random density-dependent rescaling of the coupling matrix. The transverse problem is therefore closer to the harmonic chain with random masses [35 , 36 ] than to a particle in a random potential.
We study how longitudinal momentum spreads across an array of tubes coupled by density interactions, and how this dynamics changes when the tube populations are unequal. Within zero-temperature Bogoliubov theory, we examine both the momentum correlator, termed the amplitude, and its squared magnitude, the intensity. In a uniform array the intensity spreads ballistically. Random density imbalance localizes the transverse collective modes and arrests this spreading; at fixed longitudinal momentum, the spatial decay of the time-averaged intensity measures the localization length. Equal-time correlations, by contrast, are only weakly affected by disorder and primarily reflect the interaction range. Section II introduces the model and these observables. Sections III and IV treat the uniform and disordered arrays, respectively. Section V examines the instability at strong inter-tube repulsion, and Sec. VI discusses experimental probes and extensions. Appendix A estimates the experimental time and length scales and the conditions under which tunneling can be neglected (with other Appendices devoted to fleshing out details of calculations).
Models and Bogoliubov theory
Model Hamiltonians
We consider a linear array of N t N_t N t equidistant Bose tubes with spacing a a a Figure 1 a, described by bosonic field operators Ψ i ( x ) \Psi_i(x) Ψ i ( x ) annihilating a particle of mass M M M at position x x x along tube i i i , with densities n ^ i = Ψ i † Ψ i \hat{n}_i=\Psi_i^\dagger\Psi_i n ^ i = Ψ i † Ψ i . The Hamiltonian is
Figure 1. (a) Array of N t N_t N t tubes with spacing a a a , intra-tube contact interaction U c U_c U c and inter-tube density coupling V i j ( x − x ′ ) V_{ij}(x-x') V ij ( x − x ′ ) . (b) Dispersion of the uniform array E q , k / U c E_{q,k}/U_c E q , k / U c against k ξ k\xi k ξ for the dipolar kernel at V = 0.3 V=0.3 V = 0.3 : solid, V = 0 V=0 V = 0 ; shaded, the range over transverse q q q ; dashed, the phonon limit E = n 0 U c k E=\sqrt{n_0U_c}\,k E = n 0 U c k and the free-particle limit ϵ k + n 0 U c \epsilon_k+n_0U_c ϵ k + n 0 U c . Inset: g q ( k ) g_q(k) g q ( k ) across the transverse zone at k ξ ≈ 0.6 k\xi\approx0.6 k ξ ≈ 0.6 (dotted line), against its V = 0 V=0 V = 0 value (horizontal line).
H = ∑ i H i + 1 2 ∑ i ≠ j H i j , (1) \mathcal{H}=\sum_i\mathcal{H}_i+\frac{1}{2}\sum_{i\ne j}\mathcal{H}_{ij}, \tag*{(1)} H = i ∑ H i + 2 1 i = j ∑ H ij , (1)
where H i \mathcal{H}_i H i is the intratube Hamiltonian and H i j \mathcal{H}_{ij} H ij the interaction between distinct tubes,
H i = ∫ d x [ ℏ 2 2 M ∂ x Ψ i † ∂ x Ψ i + U c 2 n ^ i 2 ( x ) ] , (2) \mathcal{H}_i=\int dx\left[\frac{\hbar^2}{2M}\partial_x\Psi_i^\dagger\partial_x\Psi_i+\frac{U_c}{2}\hat{n}_i^2(x)\right], \tag*{(2)} H i = ∫ d x [ 2 M ℏ 2 ∂ x Ψ i † ∂ x Ψ i + 2 U c n ^ i 2 ( x ) ] , (2)
H i j = ∫ d x d x ′ V i j ( x − x ′ ) n ^ i ( x ) n ^ j ( x ′ ) , (3) \mathcal{H}_{ij}=\int dx\,dx'\,V_{ij}(x-x')\hat{n}_i(x)\hat{n}_j(x'), \tag*{(3)} H ij = ∫ d x d x ′ V ij ( x − x ′ ) n ^ i ( x ) n ^ j ( x ′ ) , (3)
with U c > 0 U_c>0 U c > 0 the contact interaction. We consider two inter-tube potentials, nearest-neighbor (n.n.) and dipolar,
V i j ( x ) = { V δ ( x ) δ ∣ i − j ∣ , 1 , V d [ x 2 + a 2 ( i − j ) 2 ] − 3 / 2 . (4) V_{ij}(x)=
\begin{cases}
V\delta(x)\delta_{\lvert i-j\rvert,1},\\
V_d\left[x^2+a^2(i-j)^2\right]^{-3/2}.
\end{cases} \tag*{(4)} V ij ( x ) = { V δ ( x ) δ ∣ i − j ∣ , 1 , V d [ x 2 + a 2 ( i − j ) 2 ] − 3/2 . (4)
The dipoles are polarized normal to the plane of the array, so that the dipolar coupling is repulsive and the same for every pair of tubes. The two kernels are normalized to the same nearest-neighbor bond at k → 0 k\to0 k → 0 ,
V = V 1 ( 0 ) = 2 V d / a 2 V=V_1(0)=2V_d/a^2 V = V 1 ( 0 ) = 2 V d / a 2 (see Eq. (14) ), and V V V denotes this bond for both below.
Single-particle tunneling between tubes, t ⊥ t_\perp t ⊥ , is set to zero,
t ⊥ = 0 ; (5) t_\perp=0; \tag*{(5)} t ⊥ = 0 ; (5)
the time and momentum scales below which this is justified in a deep lattice are estimated in Appendix A. We use ℏ = M = 1 \hbar=M=1 ℏ = M = 1 , measure energies in units of n 0 U c n_0U_c n 0 U c with n 0 n_0 n 0 the mean density, lengths along the tubes in units of the healing length ξ = 1 / n 0 U c \xi=1/\sqrt{n_0U_c} ξ = 1/ n 0 U c , and set a = 1 a=1 a = 1 (for the realization of Appendix A, a ≃ 1.2 ξ a\simeq1.2\,\xi a ≃ 1.2 ξ ). Longitudinal momentum is written k k k and transverse quasimomentum q q q throughout.
Symmetry sectors and effective disorder
Both models conserve the particle number N ^ i = ∫ d x n ^ i ( x ) \hat{N}_i=\int dx\,\hat{n}_i(x) N ^ i = ∫ d x n ^ i ( x ) of each tube; equivalently H \mathcal{H} H is invariant under Ψ i → e i α i Ψ i \Psi_i\to e^{i\alpha_i}\Psi_i Ψ i → e i α i Ψ i for every i i i separately. This U ( 1 ) N t U(1)^{N_t} U ( 1 ) N t symmetry is the one familiar from sliding phases [16 , 17 , 18 , 20 ] . The Bogoliubov ansatz of Sec. II C breaks it by picking a pattern of α i \alpha_i α i ; nothing we compute depends on the α i \alpha_i α i . Translation symmetry along the tubes conserves the total longitudinal momentum of the array,
P ^ = ∑ i P ^ i = ∑ i ∫ d k k a i , k † a i , k , (6) \hat{P}=\sum_i\hat{P}_i=\sum_i\int dk\,k\,a_{i,k}^\dagger a_{i,k}, \tag*{(6)} P ^ = i ∑ P ^ i = i ∑ ∫ d k k a i , k † a i , k , (6)
where P ^ i \hat{P}_i P ^ i is the momentum of tube i i i and a i , k a_{i,k} a i , k its Fourier modes. Unlike N ^ i \hat{N}_i N ^ i , the individual P ^ i \hat{P}_i P ^ i are not conserved: momentum hops between tubes through the inter-tube interaction, and M i j ( t ) = ⟨ P ^ i ( t ) P ^ j ( 0 ) ⟩ c M_{ij}(t)=\langle\hat{P}_i(t)\hat{P}_j(0)\rangle_c M ij ( t ) = ⟨ P ^ i ( t ) P ^ j ( 0 ) ⟩ c , Eq. (24) below, is the object that measures this dynamics.
We distinguish (i) the uniform, or clean, array with n i = n 0 n_i=n_0 n i = n 0 , and (ii) disordered arrays with independent tube densities,
⟨ n i ⟩ = n 0 , Var ( n i ) = σ 2 n 0 2 , (7) \langle n_i\rangle=n_0,\qquad\operatorname{Var}(n_i)=\sigma^2n_0^2, \tag*{(7)} ⟨ n i ⟩ = n 0 , Var ( n i ) = σ 2 n 0 2 , (7)
drawn from the uniform distribution
n i ∈ [ n 0 − w , n 0 + w ] , w = 3 σ n 0 . (8) n_i \in[n_0-w,n_0+w], \qquad w=\sqrt{3}\,\sigma n_0. \tag*{(8)} n i ∈ [ n 0 − w , n 0 + w ] , w = 3 σ n 0 . (8)
Positivity of every density requires
σ < 1 3 ≃ 0.577. (9) \sigma< \frac{1}{\sqrt{3}} \simeq0.577. \tag*{(9)} σ < 3 1 ≃ 0.577. (9)
This represents loading fluctuations in an optical lattice. The localization length of the disordered array is written ξ l o c \xi_{\mathrm{loc}} ξ loc and its inverse γ ≡ 1 / ξ l o c \gamma\equiv1/\xi_{\mathrm{loc}} γ ≡ 1/ ξ loc .
Bogoliubov theory
We expand Ψ i ( x ) = n i e i α i + δ Ψ i ( x ) \Psi_i(x)=\sqrt{n_i}e^{i\alpha_i}+\delta\Psi_i(x) Ψ i ( x ) = n i e i α i + δ Ψ i ( x ) , retain terms quadratic in the fluctuations, and write a i , k a_{i,k} a i , k for the Fourier modes of δ Ψ i \delta\Psi_i δ Ψ i along the tube. The resulting Hamiltonian is
H = ∫ k > 0 d k ψ k † H ( k ) ψ k , (10) \mathcal{H}=\int_{k>0}dk\,\psi_k^\dagger H(k)\psi_k, \tag*{(10)} H = ∫ k > 0 d k ψ k † H ( k ) ψ k , (10)
where
ψ k = ( a 1 , k , … , a N t , k , a 1 , − k † , … , a N t , − k † ) T (11) \psi_k=(a_{1,k},\ldots,a_{N_t,k},a^\dagger_{1,-k},\ldots,a^\dagger_{N_t,-k})^T \tag*{(11)} ψ k = ( a 1 , k , … , a N t , k , a 1 , − k † , … , a N t , − k † ) T (11)
is the Nambu spinor and
H ( k ) = ( T ^ + G ^ G ^ G ^ T ^ + G ^ ) , (12) H(k)=
\begin{pmatrix}
\hat{T}+\hat{G} & \hat{G}\\
\hat{G} & \hat{T}+\hat{G}
\end{pmatrix}, \tag*{(12)} H ( k ) = ( T ^ + G ^ G ^ G ^ T ^ + G ^ ) , (12)
with T i j = ϵ k δ i j T_{ij}=\epsilon_k\delta_{ij} T ij = ϵ k δ ij , ϵ k = k 2 / 2 \epsilon_k=k^2/2 ϵ k = k 2 /2 , and the interaction matrix
G i j ( k ) = U c n i δ i j + ( 1 − δ i j ) V i j ( k ) n i n j , (13) G_{ij}(k)=U_c n_i\delta_{ij}+(1-\delta_{ij})V_{ij}(k)\sqrt{n_i n_j}, \tag*{(13)} G ij ( k ) = U c n i δ ij + ( 1 − δ ij ) V ij ( k ) n i n j , (13)
where V i j ( k ) = ∫ d x e i k x V i j ( x ) V_{ij}(k)=\int dx\,e^{ikx}V_{ij}(x) V ij ( k ) = ∫ d x e ik x V ij ( x ) and the chemical potential of each tube, μ i = U c n i + ∑ j ≠ i V i j ( 0 ) n j \mu_i=U_c n_i+\sum_{j\ne i}V_{ij}(0)n_j μ i = U c n i + ∑ j = i V ij ( 0 ) n j , has been subtracted. Density imbalance enters both the diagonal and the bonds of G ^ \hat{G} G ^ .
For dipolar interactions the partial Fourier transform is
V i j ( k ) = 2 V d ∣ k ∣ a ∣ i − j ∣ K 1 ( ∣ k ∣ a ∣ i − j ∣ ) , (14) V_{ij}(k)=\frac{2V_d|k|}{a|i-j|}K_1(|k|a|i-j|), \tag*{(14)} V ij ( k ) = a ∣ i − j ∣ 2 V d ∣ k ∣ K 1 ( ∣ k ∣ a ∣ i − j ∣ ) , (14)
with K 1 K_1 K 1 the modified Bessel function of the second kind; it reduces to 2 V d / a 2 ∣ i − j ∣ 2 2V_d/a^2|i-j|^2 2 V d / a 2 ∣ i − j ∣ 2 as k → 0 k\to0 k → 0 and decays as e − ∣ k ∣ a ∣ i − j ∣ e^{-|k|a|i-j|} e − ∣ k ∣ a ∣ i − j ∣ at large separation, so the transverse range of G ^ ( k ) \hat{G}(k) G ^ ( k ) is 1 / k 1/k 1/ k . The n.n. kernel has V i j ( k ) = V δ ∣ i − j ∣ , 1 V_{ij}(k)=V\delta_{|i-j|,1} V ij ( k ) = V δ ∣ i − j ∣ , 1 with no k k k dependence.
At each k k k we diagonalize G ^ ( k ) \hat{G}(k) G ^ ( k ) with an orthogonal matrix U g ( k ) U_g(k) U g ( k ) ,
U g T ( k ) G ^ ( k ) U g ( k ) = diag [ g 1 ( k ) , … , g N t ( k ) ] , (15) U_g^T(k)\hat{G}(k)U_g(k)=\operatorname{diag}[g_1(k),\ldots,g_{N_t}(k)], \tag*{(15)} U g T ( k ) G ^ ( k ) U g ( k ) = diag [ g 1 ( k ) , … , g N t ( k )] , (15)
so that g m ( k ) g_m(k) g m ( k ) is the m m m -th eigenvalue of G ^ ( k ) \hat{G}(k) G ^ ( k ) ; we omit the argument k k k where no confusion arises. In the uniform array the eigenvectors are transverse plane waves, Eq. (32) ; with disorder they are obtained numerically, once per realization and k k k . Since T ^ = ϵ k 1 \hat{T}=\epsilon_k\mathbf{1} T ^ = ϵ k 1 , the same rotation in particle and hole space block diagonalizes H ( k ) H(k) H ( k ) ,
H ~ ( k ) = U g T H ( k ) U g = ⨁ m ( ϵ k + g m g m g m ϵ k + g m ) , (16) \widetilde{H}(k)=U_g^T H(k)U_g=\bigoplus_m
\begin{pmatrix}
\epsilon_k+g_m & g_m\\
g_m & \epsilon_k+g_m
\end{pmatrix}, \tag*{(16)} H ( k ) = U g T H ( k ) U g = m ⨁ ( ϵ k + g m g m g m ϵ k + g m ) , (16)
where U g = diag ( U g , U g ) U_g=\operatorname{diag}(U_g,U_g) U g = diag ( U g , U g ) , in terms of the generalized-tube operators
b m , k = ∑ i ( U g ) i m a i , k . (17) b_{m,k}=\sum_i(U_g)_{im}a_{i,k}. \tag*{(17)} b m , k = i ∑ ( U g ) im a i , k . (17)
All mixing between physical tubes is contained in U g U_g U g . Each block is diagonalized by a single-mode Bogoliubov transformation,
β m , k = u m , k b m , k − v m , k b m , − k † \beta_{m,k}=u_{m,k}b_{m,k}-v_{m,k}b^\dagger_{m,-k} β m , k = u m , k b m , k − v m , k b m , − k †
u m , k = cosh θ m , k , v m , k = sinh θ m , k , (18) u_{m,k}=\cosh\theta_{m,k},\qquad v_{m,k}=\sinh\theta_{m,k}, \tag*{(18)} u m , k = cosh θ m , k , v m , k = sinh θ m , k , (18)
cosh ( 2 θ m , k ) = ϵ k + g m E m , k , E m , k = ϵ k ( ϵ k + 2 g m ) , (19) \cosh(2\theta_{m,k})=\frac{\epsilon_k+g_m}{E_{m,k}},\qquad E_{m,k}=\sqrt{\epsilon_k(\epsilon_k+2g_m)}, \tag*{(19)} cosh ( 2 θ m , k ) = E m , k ϵ k + g m , E m , k = ϵ k ( ϵ k + 2 g m ) , (19)
H = ∫ k > 0 d k ∑ m E m , k ( β m , k † β m , k + β m , − k † β m , − k ) + c o n s t . (20) \mathcal{H}=\int_{k>0}dk\sum_m E_{m,k}\left(\beta^\dagger_{m,k}\beta_{m,k}+\beta^\dagger_{m,-k}\beta_{m,-k}\right)+\mathrm{const.} \tag*{(20)} H = ∫ k > 0 d k m ∑ E m , k ( β m , k † β m , k + β m , − k † β m , − k ) + const. (20)
At small k k k , E m , k ≃ ∣ k ∣ g m E_{m,k}\simeq|k|\sqrt{g_m} E m , k ≃ ∣ k ∣ g m : each transverse channel has its own longitudinal sound velocity. Because U g U_g U g is orthogonal, the transverse eigenvectors are normalized and the localization of the Bogoliubov modes can be read directly from the eigenvectors of G ^ \hat{G} G ^ ; the non-unitary part of the symplectic transformation is confined to the squeeze within each block.
We will need the overlap of a physical tube with two transverse modes,
M i , k m n = [ U g T p i U g ] m n = ( U g ) i m ( U g ) i n , \mathcal{M}^{mn}_{i,k}=[U_g^T p_i U_g]_{mn}=(U_g)_{im}(U_g)_{in}, M i , k mn = [ U g T p i U g ] mn = ( U g ) im ( U g ) in ,
( p i ) a b = δ a i δ b i , (21) (p_i)_{ab}=\delta_{ai}\delta_{bi}, \tag*{(21)} ( p i ) ab = δ ai δ bi , (21)
with p i p_i p i the projector onto tube i i i ; it is symmetric in m , n m,n m , n and satisfies ∑ i M i , k m m = ∑ m M i , k m m = 1 \sum_i\mathcal{M}^{mm}_{i,k}=\sum_m\mathcal{M}^{mm}_{i,k}=1 ∑ i M i , k mm = ∑ m M i , k mm = 1 .
The spectrum (19) is real when every g m ( k ) > 0 g_m(k)>0 g m ( k ) > 0 . Writing D = diag ( n i ) D=\operatorname{diag}(n_i) D = diag ( n i ) and
M i j ( k ) = U c δ i j + ( 1 − δ i j ) V i j ( k ) , (22) M_{ij}(k)=U_c\delta_{ij}+(1-\delta_{ij})V_{ij}(k), \tag*{(22)} M ij ( k ) = U c δ ij + ( 1 − δ ij ) V ij ( k ) , (22)
Eq. (13) is a congruence of the density-independent matrix M M M ,
G ^ ( k ) = D 1 / 2 M ( k ) D 1 / 2 , (23) \hat{G}(k)=D^{1/2}M(k)D^{1/2}, \tag*{(23)} G ^ ( k ) = D 1/2 M ( k ) D 1/2 , (23)
so by Sylvester’s law of inertia G ^ \hat{G} G ^ and M M M have the same numbers of positive, negative and zero eigenvalues for any positive densities. The stability boundary is therefore independent of the density imbalance throughout the range (9); the thresholds, and what happens beyond them, are given in Sec. V.
Momentum correlations
We start by defining the average connected correlator
Π i j ( t ) = ⟨ P ^ i ( t ) P ^ j ( 0 ) ⟩ c , (24) \Pi_{ij}(t)=\langle\hat{P}_i(t)\hat{P}_j(0)\rangle_c, \tag*{(24)} Π ij ( t ) = ⟨ P ^ i ( t ) P ^ j ( 0 ) ⟩ c , (24)
evaluated in the Bogoliubov ground state. In the generalized-tube basis (17) the momentum of tube i i i is
P ^ i = ∫ d k k ∑ m , n M i , k m n ( b m , k † b n , k − b m , − k † b n , − k ) , (25) \hat{P}_i=\int dk\, k\sum_{m,n}\mathcal{M}^{mn}_{i,k}\left(b^\dagger_{m,k}b_{n,k}-b^\dagger_{m,-k}b_{n,-k}\right), \tag*{(25)} P ^ i = ∫ d k k m , n ∑ M i , k mn ( b m , k † b n , k − b m , − k † b n , − k ) , (25)
and with
b m , k ( t ) = u m , k β m , k e i E m , k t + v m , k β m , − k † e − i E m , k t (26) b_{m,k}(t)=u_{m,k}\beta_{m,k}e^{iE_{m,k}t}+v_{m,k}\beta^\dagger_{m,-k}e^{-iE_{m,k}t} \tag*{(26)} b m , k ( t ) = u m , k β m , k e i E m , k t + v m , k β m , − k † e − i E m , k t (26)
(the opposite sign convention for the time evolution replaces Π i j \Pi_{ij} Π ij by its complex conjugate) the zero-temperature contractions give (Appendix B)
Π i j ( t ) = ∫ d k k 2 ∑ m , n e i E m n , k t O i j , k m n sinh 2 ϑ m n , k , (27) \Pi_{ij}(t)=\int dk\, k^2\sum_{m,n}e^{i\mathcal{E}_{mn,k}t}\mathcal{O}^{mn}_{ij,k}\sinh^2\vartheta_{mn,k}, \tag*{(27)} Π ij ( t ) = ∫ d k k 2 m , n ∑ e i E mn , k t O ij , k mn sinh 2 ϑ mn , k , (27)
where
E m n , k = E m , k + E n , k , ϑ m n , k = θ m , k − θ n , k , (28) \mathcal{E}_{mn,k}=E_{m,k}+E_{n,k},\qquad\vartheta_{mn,k}=\theta_{m,k}-\theta_{n,k}, \tag*{(28)} E mn , k = E m , k + E n , k , ϑ mn , k = θ m , k − θ n , k , (28)
and
O i j , k m n = M i , k m n M j , k n m . (29) \mathcal{O}^{mn}_{ij,k}=\mathcal{M}^{mn}_{i,k}\mathcal{M}^{nm}_{j,k}. \tag*{(29)} O ij , k mn = M i , k mn M j , k nm . (29)
The spatial structure is carried by the eigenvectors of G ^ \hat{G} G ^ through O \mathcal{O} O , the bosonic coherence by sinh 2 ϑ \sinh^2\vartheta sinh 2 ϑ . Since P ^ i \hat{P}_i P ^ i is bilinear, it creates a pair of quasiparticles from the vacuum: only sums of mode energies appear, and the spectrum of Π \Pi Π at fixed k k k has a threshold at 2 min m E m , k 2\min_m E_{m,k} 2 min m E m , k . Each of the two mode amplitudes in O \mathcal{O} O connects tube i i i to tube j j j , which is the origin of the factor of two in Eq. (61) .
Whereas we typically expect correlation functions to be useful for characterizing important physics, Π i j \Pi_{ij} Π ij suffers from two handicaps: (i) the BEC ground state is an eigenstate of the total momentum with zero fluctuation, so ∑ j Π i j ( t ) = 0 \sum_j\Pi_{ij}(t)=0 ∑ j Π ij ( t ) = 0 at all t t t [Eq. (44) ]; (ii) disordered wave functions are random and so are the overlaps in Eq. (29) , hence suppressing Π i j \Pi_{ij} Π ij on averaging. We can circumvent these difficulties following the prescription from the theory of disorder-averaged Green functions, where the single-particle Green function is similarly compromised. To this end we define ∣ Φ i ⟩ = P ^ i ∣ 0 ⟩ |\Phi_i\rangle=\hat{P}_i|0\rangle ∣ Φ i ⟩ = P ^ i ∣0 ⟩ and corresponding amplitude and intensity, respectively,
Π i j ( t ) = ⟨ Φ i ∣ e − i ( H ^ − E 0 ) t ∣ Φ j ⟩ , (30) \Pi_{ij}(t)=\langle\Phi_i|e^{-i(\hat{H}-E_0)t}|\Phi_j\rangle, \tag*{(30)} Π ij ( t ) = ⟨ Φ i ∣ e − i ( H ^ − E 0 ) t ∣ Φ j ⟩ , (30)
∣ Π i j ( t ) ∣ 2 = ∣ ⟨ Φ i ∣ e − i H ^ t ∣ Φ j ⟩ ∣ 2 . (31) |\Pi_{ij}(t)|^2=\left|\langle\Phi_i|e^{-i\hat{H}t}|\Phi_j\rangle\right|^2. \tag*{(31)} ∣ Π ij ( t ) ∣ 2 = ⟨ Φ i ∣ e − i H ^ t ∣ Φ j ⟩ 2 . (31)
Late time asymptotics of ∣ Π i j ( t ) ∣ 2 |\Pi_{ij}(t)|^2 ∣ Π ij ( t ) ∣ 2 will be used to characterize ballistic transport next and localization in Sec. IV.
For n i = n 0 n_i=n_0 n i = n 0 the interaction matrix is translationally invariant in the tube index, so its eigenvectors are Bloch waves and the channel index of Sec. IIC acquires the meaning of a transverse momentum rather than a bare ordering label. We accordingly write q q q in place of m m m throughout this section. In the thermodynamic limit the eigenvalues are
g q ( k ) = n 0 [ U c + 2 ∑ r = 1 ∞ V r ( k ) cos ( q r ) ] , (32) g_q(k)=n_0\left[U_c+2\sum_{r=1}^{\infty}V_r(k)\cos(qr)\right], \tag*{(32)} g q ( k ) = n 0 [ U c + 2 r = 1 ∑ ∞ V r ( k ) cos ( q r ) ] , (32)
where q q q is the transverse quasimomentum and V r ( k ) V_r(k) V r ( k ) the interaction matrix element at separation r r r . Nearest-neighbor coupling has V r ( k ) = V δ r , 1 V_r(k)=V\delta_{r,1} V r ( k ) = V δ r , 1 , while the dipolar kernel with a = 1 a=1 a = 1 has
V r ( k ) = 2 V d ∣ k ∣ r K 1 ( ∣ k ∣ r ) . (33) V_r(k)=\frac{2V_d|k|}{r}K_1(|k|r). \tag*{(33)} V r ( k ) = r 2 V d ∣ k ∣ K 1 ( ∣ k ∣ r ) . (33)
The two kernels differ in how far this can be taken. For nearest neighbors the sum terminates, g q = n 0 ( U c + 2 V cos q ) g_q=n_0(U_c+2V\cos q) g q = n 0 ( U c + 2 V cos q ) , with no residual k k k dependence. The dipolar sum has no closed form at general k k k ; it converges rapidly and we evaluate it numerically, except as k → 0 k\to0 k → 0 , where it reduces to the piecewise-quadratic form of Eq. (106) and fixes the threshold of Sec. V. The dispersion of the uniform array is E q , k = ϵ k [ ϵ k + 2 g q ( k ) ] E_{q,k}=\sqrt{\epsilon_k[\epsilon_k+2g_q(k)]} E q , k = ϵ k [ ϵ k + 2 g q ( k )] .
In a complex Bloch basis the mode-pair overlap is
O i j , k q , q ′ = 1 N t 2 e i ( q ′ − q ) ( i − j ) , (34) \mathcal{O}^{q,q'}_{ij,k}=\frac{1}{N_t^2}e^{i(q'-q)(i-j)}, \tag*{(34)} O ij , k q , q ′ = N t 2 1 e i ( q ′ − q ) ( i − j ) , (34)
so that the transverse Fourier transform of the correlator reads (q ± = q ± Q 2 q_{\pm}=q\pm\frac{Q}{2} q ± = q ± 2 Q )
Π Q ( t ) = ∫ d q d k k 2 e i E q − , q + , k t sinh 2 ϑ q − , q + , k . (35) \Pi_Q(t)=\int dq\,dk\,k^2e^{i\mathcal{E}_{q_-,q_+,k}t}\sinh^2\vartheta_{q_-,q_+,k}. \tag*{(35)} Π Q ( t ) = ∫ d q d k k 2 e i E q − , q + , k t sinh 2 ϑ q − , q + , k . (35)
Amplitude
Kinematics
At fixed k k k the amplitude is a sum over pairs of transverse modes, and its frequency content is the weighted joint density of states of the pair continuum,
Π Q ( ω ∣ k ) = ∑ q S Q ( q ) δ ( ω − Ω Q ( q ) ) , (36) \Pi_Q(\omega|k)=\sum_q S_Q(q)\delta\left(\omega-\Omega_Q(q)\right), \tag*{(36)} Π Q ( ω ∣ k ) = q ∑ S Q ( q ) δ ( ω − Ω Q ( q ) ) , (36)
with Ω Q ( q ) = E q − , k + E q + , k \Omega_Q(q)=E_{q_-,k}+E_{q_+,k} Ω Q ( q ) = E q − , k + E q + , k , S Q ( q ) = sinh 2 ( θ q − , k − θ q + , k ) S_Q(q)=\sinh^2(\theta_{q_-,k}-\theta_{q_+,k}) S Q ( q ) = sinh 2 ( θ q − , k − θ q + , k ) and Π Q ( k , t ) = ∫ d ω e i ω t Π Q ( ω ∣ k ) \Pi_Q(k,t)=\int d\omega\,e^{i\omega t}\Pi_Q(\omega|k) Π Q ( k , t ) = ∫ d ω e iω t Π Q ( ω ∣ k ) , the k 2 k^2 k 2 and normalization factors being those of Eq. (35) . In the variables q ± q\pm q ± both Ω Q \Omega_Q Ω Q and S Q S_Q S Q are even in q q q , so q = 0 q=0 q = 0 and q = π q=\pi q = π are critical points of Ω Q \Omega_Q Ω Q on every Q Q Q and all other critical points come in pairs ± q c \pm q_c ± q c . The support of Eq. (36) is the range of $\Omega_Q(q), {xref}Fig. 2`: the doubled band
[ 2 E min , 2 E max ] [2E_{\min},2E_{\max}] [ 2 E m i n , 2 E m a x ] at Q = 0 Q = 0 Q = 0 , where S Q ∝ Q 2 S_Q \propto Q^2 S Q ∝ Q 2 , narrowing toward Q = π Q = \pi Q = π . Each critical point is a van Hove singularity, Π Q ( ω ∣ k ) ∝ S Q ( q c ) ∣ Ω Q ′ ( q c ) ∣ − 1 / 2 ∣ ω − Ω Q ( q c ) ∣ − 1 / 2 \Pi_Q(\omega|k) \propto S_Q(q_c)|\Omega'_Q(q_c)|^{-1/2}|\omega-\Omega_Q(q_c)|^{-1/2} Π Q ( ω ∣ k ) ∝ S Q ( q c ) ∣ Ω Q ′ ( q c ) ∣ − 1/2 ∣ ω − Ω Q ( q c ) ∣ − 1/2 , one-sided at the edges and two-sided inside the band.
For the nearest-neighbor kernel in the phonon limit, E q , k → k g q E_{q,k} \to k\sqrt{g_q} E q , k → k g q with g q = 1 + 2 V cos q g_q = 1 + 2V\cos q g q = 1 + 2 V cos q , the structure is explicit to second order in V V V :
Ω Q ( q ) k = 2 + 2 V cos Q 2 cos q − V 2 2 [ 1 + cos Q cos 2 q ] . (37) \frac{\Omega_Q(q)}{k} = 2 + 2V\cos\frac{Q}{2}\cos q - \frac{V^2}{2}\left[1+\cos Q\cos2q\right]. \tag*{(37)} k Ω Q ( q ) = 2 + 2 V cos 2 Q cos q − 2 V 2 [ 1 + cos Q cos 2 q ] . (37)
The point q = 0 q = 0 q = 0 is the maximum for all Q Q Q , the upper edge 2 + 2 V cos Q 2 − V 2 cos 2 Q 2 2 + 2V\cos\frac{Q}{2} - V^2\cos^2\frac{Q}{2} 2 + 2 V cos 2 Q − V 2 cos 2 2 Q . The point q = π q = \pi q = π is the minimum for Q < Q c Q < Q_c Q < Q c , with Q c = π − 2 V Q_c = \pi- 2V Q c = π − 2 V to this order; beyond Q c Q_c Q c a pair of minima splits off at q = π ± δ q = \pi\pm\delta q = π ± δ , cos δ = cos Q 2 / ( V ∣ cos Q ∣ ) \cos\delta= \cos\frac{Q}{2}/(V|\cos Q|) cos δ = cos 2 Q / ( V ∣ cos Q ∣ ) , and q = π q = \pi q = π becomes a local maximum inside the band (dashed line in Figure 2 ), where the singularity is two-sided; at Q c Q_c Q c itself the curvature at q = π q = \pi q = π vanishes and the onset is ∣ ω − ω c ∣ − 3 / 4 |\omega-\omega_c|^{-3/4} ∣ ω − ω c ∣ − 3/4 . The band narrows from width 4 V 4V 4 V at Q = 0 Q = 0 Q = 0 to V 2 V^2 V 2 at Q = π Q = \pi Q = π , where δ = π / 2 \delta= \pi/2 δ = π /2 and the support is [ 2 − V 2 , 2 ] [2-V^2,2] [ 2 − V 2 , 2 ] : the first-order terms cancel for every q q q because g q + π = 2 − g q g_{q+\pi}=2-g_q g q + π = 2 − g q for a bipartite kernel. The dipolar kernel has the same critical-point structure, with Q c Q_c Q c further from π \pi π (Q c = 0.87 Q_c = 0.87 Q c = 0.87 at k = 0.1 k = 0.1 k = 0.1 ) and a band of width O ( V ) O(V) O ( V ) at Q = π Q = \pi Q = π , since the even-separation bonds do not cancel; its interior line crosses the band from the upper edge at Q c Q_c Q c to the lower edge at Q = π Q = \pi Q = π . Two consequences run through the rest of the section: there is no weight below 2 min q E q , k 2\min_q E_{q,k} 2 min q E q , k at any Q Q Q , and the staggered pair Q = π Q = \pi Q = π is nearly dispersionless, so it dephases last.
Figure 2. Support of the pair spectrum Π Q ( ω ∣ k ) \Pi_Q(\omega|k) Π Q ( ω ∣ k ) , Eq. (36) , at k = 0.1 k = 0.1 k = 0.1 (grey), for (a) the dipolar kernel at V = 0.3 V = 0.3 V = 0.3 and (b) the n.n. kernel at the matched bond V = 0.181 V = 0.181 V = 0.181 . Solid lines are the band edges, critical points of Ω Q ( q ) \Omega_Q(q) Ω Q ( q ) at q = 0 q = 0 q = 0 and q = π q = \pi q = π or at the pair ± q c \pm q_c ± q c ; the dashed line is the interior critical point that exists for Q > Q c Q > Q_c Q > Q c (dotted). Every line is an inverse-square-root singularity; the weight S Q S_Q S Q , not shown, vanishes as Q 2 Q^2 Q 2 at small Q Q Q and is largest near Q = π Q = \pi Q = π . For (b), Eq. (37) gives Q c = π − 2 V = 2.78 Q_c = \pi- 2V = 2.78 Q c = π − 2 V = 2.78 and the support [ 2 − V 2 , 2 ] [2 - V^2, 2] [ 2 − V 2 , 2 ] at Q = π Q = \pi Q = π , against 2.77 2.77 2.77 and [ 1.968 , 2.003 ] [1.968, 2.003] [ 1.968 , 2.003 ] from the full dispersion.
Equal time, spatial decay profiles
The squeeze factor in Eq. (35) separates into functions of the two transverse momenta,
sinh 2 ( θ q − θ q ′ ) = s q c q ′ + c q s q ′ − 2 h q h q ′ . (38) \sinh^2(\theta_q-\theta_{q'})=s_qc_{q'}+c_qs_{q'}-2h_qh_{q'}. \tag*{(38)} sinh 2 ( θ q − θ q ′ ) = s q c q ′ + c q s q ′ − 2 h q h q ′ . (38)
where
s q = ϵ k + g q 2 E q , k − 1 2 , c q = ϵ k + g q 2 E q , k + 1 2 . (39) s_q=\frac{\epsilon_k+g_q}{2E_{q,k}}-\frac{1}{2},\qquad c_q=\frac{\epsilon_k+g_q}{2E_{q,k}}+\frac{1}{2}. \tag*{(39)} s q = 2 E q , k ϵ k + g q − 2 1 , c q = 2 E q , k ϵ k + g q + 2 1 . (39)
and
h q = g q 2 E q , k . (40) h_q=\frac{g_q}{2E_{q,k}}. \tag*{(40)} h q = 2 E q , k g q . (40)
Equation (38) is the hyperbolic angle-addition identity, and it factorizes because the two mode labels enter ϑ q q ′ \vartheta_{qq'} ϑ q q ′ only through their own squeeze angles. This is what allows the double transverse-mode sum to collapse. For x q ∈ { s q , c q , h q } x_q\in\{s_q,c_q,h_q\} x q ∈ { s q , c q , h q } define the Fourier coefficients
X r ( k , t ) = ∫ d q 2 π e i E q , k t − i q r x q , (41) X_r(k,t)=\int\frac{dq}{2\pi}e^{iE_{q,k}t-iqr}x_q, \tag*{(41)} X r ( k , t ) = ∫ 2 π d q e i E q , k t − i q r x q , (41)
with r = i − j r=i-j r = i − j the transverse separation. All three integrands are even in q q q , so X − r = X r X_{-r}=X_r X − r = X r , and
Π r ( t ) = 2 ∫ d k k 2 [ S r ( k , t ) C r ( k , t ) − H r 2 ( k , t ) ] . (42) \Pi_r(t)=2\int dk\,k^2\left[S_r(k,t)C_r(k,t)-H_r^2(k,t)\right]. \tag*{(42)} Π r ( t ) = 2 ∫ d k k 2 [ S r ( k , t ) C r ( k , t ) − H r 2 ( k , t ) ] . (42)
This expression is exact within the quadratic theory at all r r r and t t t .
The coherence factors satisfy the pointwise identity
c q s q − h q 2 = ( ϵ k + g q ) 2 − g q 2 4 E q , k 2 − 1 4 = 0 , (43) c_qs_q-h_q^2=\frac{(\epsilon_k+g_q)^2-g_q^2}{4E_{q,k}^2}-\frac{1}{4}=0, \tag*{(43)} c q s q − h q 2 = 4 E q , k 2 ( ϵ k + g q ) 2 − g q 2 − 4 1 = 0 , (43)
which holds by the definition of E q , k E_{q,k} E q , k . Applying Parseval’s theorem to Eq. (42) then gives
∑ j Π i j ( t ) = 0 , (44) \sum_j\Pi_{ij}(t)=0, \tag*{(44)} j ∑ Π ij ( t ) = 0 , (44)
the correlator form of the conservation law in Eq. (6) , and a dynamical statement rather than an equal-time one. Since Π i i ( t ) ≠ 0 \Pi_{ii}(t)\ne0 Π ii ( t ) = 0 , Eq. (44) forces a zero crossing in ∣ i − j ∣ |i-j| ∣ i − j ∣ : the profile cannot be a simple positive exponential at any time, which matters for the fitting procedure of Sec. IV B.
At fixed longitudinal momentum the decay is governed by the branch point of the integrand of Eq. (41) lying nearest the real q q q axis: one at ∣ Im q ∣ = ∣ κ ∣ |\operatorname{Im}q|=|\kappa| ∣ Im q ∣ = ∣ κ ∣ makes each Fourier coefficient decay at κ \kappa κ , and Π r \Pi_r Π r , being bilinear in them, at 2 κ 2\kappa 2 κ . The two kernels place it differently. For nearest-neighbor coupling g ( q ) g(q) g ( q ) is entire, so the branch point of E q , k E_{q,k} E q , k controls κ E ( k ) = Im z k \kappa_E(k)=\operatorname{Im}z_k κ E ( k ) = Im z k with z k = ( k 2 / 2 + 2 n 0 U ˉ c ) / ( 4 n 0 V ) z_k=(k^2/2+2n_0\bar U_c)/(4n_0V) z k = ( k 2 /2 + 2 n 0 U ˉ c ) / ( 4 n 0 V ) . For the dipolar kernel the Bessel sum converges only inside ∣ Im q ∣ < ∣ κ ∣ |\operatorname{Im}q|<|\kappa| ∣ Im q ∣ < ∣ κ ∣ , and for the parameters used here that strip edge is nearer, so the rate is ∣ k ∣ |k| ∣ k ∣ itself. Hence
Π r N N ( k , 0 ) ∼ C N N ( k ) r − 2 e − 2 κ E ( k ) r , (45) \Pi_r^{\mathrm{NN}}(k,0) \sim C_{\mathrm{NN}}(k) r^{-2}e^{-2\kappa_E(k)r}, \tag*{(45)} Π r NN ( k , 0 ) ∼ C NN ( k ) r − 2 e − 2 κ E ( k ) r , (45)
Π r d i p ( k , 0 ) ∼ C d i p ( k ) r − 3 e − 2 ∣ k ∣ r , (46) \Pi_r^{\mathrm{dip}}(k,0) \sim C_{\mathrm{dip}}(k) r^{-3}e^{-2|k|r}, \tag*{(46)} Π r dip ( k , 0 ) ∼ C dip ( k ) r − 3 e − 2∣ k ∣ r , (46)
with C ( k ) C(k) C ( k ) smooth. The two exponents differ because the correlator factorizes into transforms carrying reciprocal powers of E q , k E_{q,k} E q , k , whose singular behavior coincides for the dipolar kernel and cancels for the nearest-neighbor one; the derivation is given in Appendix D. Both forms are confirmed against Eq. (42) in Figure 3 , prefactors included. The dipolar correlator therefore carries an interaction-induced exponential envelope even though its transverse eigenmodes are Bloch waves.
Figure 3. Spatial structure of the momentum correlator at V = 0.3 V = 0.3 V = 0.3 . Panel (a) shows the fixed-k k k profile of the uniform array at k = 1 k = 1 k = 1 for both kernels, from Eq. (42) . Dotted lines give the predicted asymptotic forms of Eqs. (45) and (46) , anchored to the data at r = 8 r = 8 r = 8 : r − 2 e − 2 κ E r r^{-2}e^{-2\kappa_E r} r − 2 e − 2 κ E r with 2 κ E = 2.73 2\kappa_E = 2.73 2 κ E = 2.73 for nearest neighbors, and r − 3 e − 2 ∣ k ∣ r r^{-3}e^{-2|k|r} r − 3 e − 2∣ k ∣ r for dipolar. Both the decay rates and the algebraic prefactors are reproduced. Panel (b) shows the same profiles after integration over longitudinal momentum, on log-log axes, against the r − 7 r^{-7} r − 7 law of Eq. (49) and the r − 7 / 2 e − 2 r arccosh ( U c / 2 V ) r^{-7/2}e^{-2r\operatorname{arccosh}(U_c/2V)} r − 7/2 e − 2 r arccosh ( U c /2 V ) form of Eq. (48) . Panel (c) shows the normalized long-time average of Eq. (63) for disordered dipolar arrays at N t = 1000 N_t = 1000 N t = 1000 and k = 1 k = 1 k = 1 , whose slope is the localization rate 4 γ 4\gamma 4 γ of Eq. (62) . Dotted lines give that slope with γ \gamma γ from the transfer matrix of Appendix I3, nothing being fitted to these curves; the measured decay approaches it from above as N t / ξ l o c N_t/\xi_{\mathrm{loc}} N t / ξ loc grows, the shortfall at weak disorder being the finite-ring bias quantified in Figure 8 . Curves are truncated at the roundoff floor, and the rates quoted in Sec. IV B are fitted over d / ξ l o c ∈ [ 3 , 8 ] d/\xi_{\mathrm{loc}} \in[3,8] d / ξ loc ∈ [ 3 , 8 ] .
The longitudinal-momentum integral treats the two kernels differently, and the difference is qualitative. For nearest-neighbor coupling κ E \kappa_E κ E is minimized at k = 0 k=0 k = 0 with quadratic curvature, κ E ( k ) = κ 0 + a k 2 + O ( k 4 ) \kappa_E(k)=\kappa_0+ak^2+O(k^4) κ E ( k ) = κ 0 + a k 2 + O ( k 4 ) , where
a = 1 4 n 0 V ( U c / 2 V ) 2 − 1 (47) a=\frac{1}{4n_0V\sqrt{(U_c/2V)^2-1}} \tag*{(47)} a = 4 n 0 V ( U c /2 V ) 2 − 1 1 (47)
sets the width in k k k that contributes at large r r r . The amplitude vanishes quadratically as well, since the explicit k 2 k^2 k 2 in Eq. (27) survives the cancellation of the 1 / k 1/k 1/ k divergences between S r C r S_r C_r S r C r and H 2 H^2 H 2 . Laplace integration then contributes a further r − 3 / 2 r^{-3/2} r − 3/2 , and
Π r ( 0 ) ∼ C r 7 / 2 exp [ − 2 r arccosh ( U c 2 V ) ] . (48) \Pi_r(0)\sim\frac{C}{r^{7/2}}\exp\left[-2r\operatorname{arccosh}\left(\frac{U_c}{2V}\right)\right]. \tag*{(48)} Π r ( 0 ) ∼ r 7/2 C exp [ − 2 r arccosh ( 2 V U c ) ] . (48)
For dipolar interactions the fixed-k k k rate in Eq. (46) instead vanishes linearly as k → 0 k\to0 k → 0 , so no exponential survives the integration and
Π r ( 0 ) ∼ C r 7 . (49) \Pi_r(0)\sim\frac{C}{r^7}. \tag*{(49)} Π r ( 0 ) ∼ r 7 C . (49)
The dipolar profile of the uniform array is thus a pure power law. Both regimes are shown in Figure 3 . This power-law background does not enter the localization analysis of Sec. IV, which works at fixed k k k and with the long-time average, whose profile in the uniform array is flat (Sec. IV B).
Late time asymptotics, Q → 0
Expanding Eq. (35) to order Q 2 Q^2 Q 2 ,
Π Q ( t ) = Q 2 A Λ ( t ) + O ( Q 4 ) , \Pi_Q(t)=Q^2 A_\Lambda(t)+O(Q^4), Π Q ( t ) = Q 2 A Λ ( t ) + O ( Q 4 ) ,
A Λ ( t ) = 1 4 ∫ 0 Λ d k k 2 ∫ 0 2 π d q ( g q ′ ) 2 ( ϵ k + 2 g q ) 2 e 2 i E q , k t , (50) A_\Lambda(t)=\frac{1}{4}\int_0^\Lambda dk\,k^2\int_0^{2\pi}dq\,\frac{(g'_q)^2}{(\epsilon_k+2g_q)^2}e^{2iE_{q,k}t}, \tag*{(50)} A Λ ( t ) = 4 1 ∫ 0 Λ d k k 2 ∫ 0 2 π d q ( ϵ k + 2 g q ) 2 ( g q ′ ) 2 e 2 i E q , k t , (50)
where g q ′ = ∂ q g q g'_q=\partial_q g_q g q ′ = ∂ q g q and Λ \Lambda Λ is a longitudinal ultraviolet cutoff. We define η = A Λ ( 0 ) \eta=A_\Lambda(0) η = A Λ ( 0 ) for later use, see also Appendix I for details of setting the cutoff.
The long-time behavior follows from the k → 0 k\to0 k → 0 endpoint of Eq. (50) and holds for either kernel. There E q , k → k g q ( 0 ) E_{q,k}\to k\sqrt{g_q(0)} E q , k → k g q ( 0 ) , so the phase is linear in k k k with no stationary point, while the amplitude vanishes as k 2 k^2 k 2 ; the endpoint expansion of Appendix E then gives
A ( t ) ∼ c 3 t 3 , t → ∞ , (51) A(t)\sim\frac{c_3}{t^3},\qquad t\to\infty, \tag*{(51)} A ( t ) ∼ t 3 c 3 , t → ∞ , (51)
with c 3 c_3 c 3 fixed by the transverse interaction through ∫ d q ( g q ′ ) 2 / g q 7 / 2 \int dq\,(g'_q)^2/g_q^{7/2} ∫ d q ( g q ′ ) 2 / g q 7/2 . The t − 3 t^{-3} t − 3 decay makes ∫ 0 ∞ d t A ( t ) \int_0^\infty dt\,A(t) ∫ 0 ∞ d t A ( t ) finite; it is a sum rule over the pair spectrum, and the next paragraph explains why it is no transport coefficient. The approach to the power law is shown in Figure 4 .
Figure 4. Temporal decay of the small-Q Q Q response of the uniform array at V = 0.3 V = 0.3 V = 0.3 for both kernels. The left panel shows ∣ Im A ( t ) ∣ |\operatorname{Im} A(t)| ∣ Im A ( t ) ∣ against the t − 3 t^{-3} t − 3 law of Eq. (51) , with c 3 c_3 c 3 evaluated from Eq. (102) . The nearest-neighbor curve uses the closed form of Eq. (52) ; the dipolar curve is Eq. (50) by direct quadrature, that reduction not applying when the band disperses with k k k . The right panel shows the local exponent α ( t ) = − d ln A / d ln t \alpha(t) = -d\ln A/d\ln t α ( t ) = − d ln A / d ln t , which overshoots 3 before settling onto it once the subleading t − 5 t^{-5} t − 5 correction has died off. The cutoff used for it is discussed in Appendix I5.
Equations (50) and (51) should be read against the hydrodynamics of a conserved density. There the transverse Fourier component of the density relaxes as χ Q e − D Q 2 t \chi_Qe^{-DQ^2t} χ Q e − D Q 2 t : the static susceptibility χ Q \chi_Q χ Q tends to a constant as Q → 0 Q\to0 Q → 0 , the time dependence enters only through Q 2 t Q^2t Q 2 t , and the rate vanishes with Q Q Q because the total is conserved. The total longitudinal momentum ∑ i P ^ i \sum_i\hat P_i ∑ i P ^ i is conserved here as well, but the Bogoliubov ground state is an eigenstate of it with zero fluctuation, so Π Q ( t ) \Pi_Q(t) Π Q ( t ) vanishes identically, and the small-Q Q Q correlator is Q 2 A ( t ) Q^2A(t) Q 2 A ( t ) with the Q Q Q and t t t dependence factorized. What decays in A ( t ) A(t) A ( t ) is the phase coherence of the two-quasiparticle continuum of Eq. (35) : P ^ i \hat P_i P ^ i creates a pair with energy E q , k + E q + Q , k E_{q,k}+E_{q+Q,k} E q , k + E q + Q , k , the spectrum of Π \Pi Π has a threshold at 2 min q E q , k 2\min_q E_{q,k} 2 min q E q , k , and vanishes as ω 2 \omega^2 ω 2 after the k k k integration, and the t − 3 t^{-3} t − 3 law is the k → 0 k\to0 k → 0 endpoint of that continuum. Nothing diffusive appears at any Q Q Q because the quadratic theory has no quasiparticle collisions. A momentum disturbance on one tube crosses the array as a wave packet with the transverse group velocity ∂ q E q , k \partial_qE_{q,k} ∂ q E q , k , and the width of the momentum profile grows ballistically until the ring closes (Sec. IV D). A viscosity, the ω → 0 \omega\to0 ω → 0 limit of a stress correlator, would require the intraband channel E m − E n E_m-E_n E m − E n , which is empty at T = 0 T=0 T = 0 and opens with thermal quasiparticles, together with the collisions that make it dissipative. We therefore speak of transverse momentum transport throughout, and of η \eta η as the equal-time small-Q Q Q coefficient: it measures how far the momentum profile extends at t = 0 t=0 t = 0 , which is set by the interaction range, and Sec. IV C finds it accordingly insensitive to localization.
For nearest-neighbor coupling the same result is available in closed form. Since g q g_q g q is then independent of k k k , the change of variable k = 2 g q x k=2\sqrt{g_q}x k = 2 g q x reduces Eq. (50) to
A Λ ( t ) = 1 2 ∫ 0 2 π d q ( g q ′ ) 2 g q J ( 4 g q t , Λ 2 g q ) , (52) A_\Lambda(t)=\frac{1}{2}\int_0^{2\pi}dq\,\frac{(g'_q)^2}{\sqrt{g_q}}J\left(4g_qt,\frac{\Lambda}{2\sqrt{g_q}}\right), \tag*{(52)} A Λ ( t ) = 2 1 ∫ 0 2 π d q g q ( g q ′ ) 2 J ( 4 g q t , 2 g q Λ ) , (52)
with the master integral
J ( s , X ) = ∫ 0 X d x x 2 ( 1 + x 2 ) 2 e i s x 1 + x 2 , (53) J(s,X)=\int_0^Xdx\,\frac{x^2}{(1+x^2)^2}e^{isx\sqrt{1+x^2}}, \tag*{(53)} J ( s , X ) = ∫ 0 X d x ( 1 + x 2 ) 2 x 2 e i s x 1 + x 2 , (53)
which is ultraviolet finite at t = 0 t=0 t = 0 and reproduces Eq. (51) with the same coefficient. Appendix F collects the transverse integrals. At the instability threshold the logarithmic divergence of the nearest-neighbor one is cancelled by a vanishing prefactor, so the singular behavior of η \eta η there originates in the longitudinal integral instead. The long-time coefficient behaves differently, since ∫ d q ( g q ′ ) 2 / g q 7 / 2 \int dq\,(g_{q}')^2/g_q^{7/2} ∫ d q ( g q ′ ) 2 / g q 7/2 carries a higher power of g q g_q g q and does diverge as g π → 0 g_\pi\to0 g π → 0 .
Intensity
The effective mean-square distance the excitations have moved via ( d i j = min ( ∣ i − j ∣ , N t − ∣ i − j ∣ ) ) (d_{ij} = \min(|i-j|,N_t-|i-j|)) ( d ij = min ( ∣ i − j ∣ , N t − ∣ i − j ∣ ))
R 2 ( k , t ) ≡ ∑ j ≠ i d i j 2 ∣ Π i j ( k , t ) ∣ 2 ∑ j ≠ i ∣ Π i j ( k , t ) ∣ 2 . (54) R^2(k,t) \equiv\frac{\sum_{j\ne i}d_{ij}^2|\Pi_{ij}(k,t)|^2}{\sum_{j\ne i}|\Pi_{ij}(k,t)|^2}. \tag*{(54)} R 2 ( k , t ) ≡ ∑ j = i ∣ Π ij ( k , t ) ∣ 2 ∑ j = i d ij 2 ∣ Π ij ( k , t ) ∣ 2 . (54)
It is relatively straightforward to see that this quantity is not dominated by small Q Q Q regime where Π Q \Pi_Q Π Q becomes suppressed
∑ r ∣ Π r ( k , t ) ∣ 2 = 1 N t ∑ Q ∣ Π Q ( k , t ) ∣ 2 , \sum_r|\Pi_r(k,t)|^2 = \frac{1}{N_t}\sum_Q|\Pi_Q(k,t)|^2, r ∑ ∣ Π r ( k , t ) ∣ 2 = N t 1 Q ∑ ∣ Π Q ( k , t ) ∣ 2 ,
∑ r r 2 ∣ Π r ( k , t ) ∣ 2 = 1 N t ∑ Q ∣ ∂ Q Π Q ( k , t ) ∣ 2 . (55) \sum_r r^2|\Pi_r(k,t)|^2 = \frac{1}{N_t}\sum_Q|\partial_Q\Pi_Q(k,t)|^2. \tag*{(55)} r ∑ r 2 ∣ Π r ( k , t ) ∣ 2 = N t 1 Q ∑ ∣ ∂ Q Π Q ( k , t ) ∣ 2 . (55)
The denominator is the probability that the pair created on tube i i i is found, at time t t t , as a pair on a single tube. Written in real space, Π r ( k , t ) = N t − 2 ∑ q , q ′ S ( q , q ′ ) e i ( q − q ′ ) r + i ( E q , k + E q ′ , k ) t \Pi_r(k,t) = N_t^{-2}\sum_{q,q'}S(q,q')e^{i(q-q')r+i(E_{q,k}+E_{q',k})t} Π r ( k , t ) = N t − 2 ∑ q , q ′ S ( q , q ′ ) e i ( q − q ′ ) r + i ( E q , k + E q ′ , k ) t with S = sinh 2 ( θ q − θ q ′ ) S = \sinh^2(\theta_q-\theta_{q'}) S = sinh 2 ( θ q − θ q ′ ) , and for 1 ≪ t ≪ N t / v 1 \ll t \ll N_t/v 1 ≪ t ≪ N t / v the double sum is dominated by the stationary points v q = − v q ′ v_q=-v_{q'} v q = − v q ′ , r / t r/t r / t . v q = ∂ q E q , k v_q=\partial_qE_{q,k} v q = ∂ q E q , k : the two quasiparticles must have moved to the same tube, so only pairs with equal transverse velocity contribute. These are the critical points of Ω Q ( q ) \Omega_Q(q) Ω Q ( q ) , ∂ q Ω Q = 0 \partial_q\Omega_Q=0 ∂ q Ω Q = 0 , i.e. the van Hove lines of Figure 2 , which for the pairs that carry the signal lie at Q Q Q between Q c Q_c Q c and π \pi π . Each such pair gives ∣ Π r ∣ 2 ∝ t − 2 |\Pi_r|^2\propto t^{-2} ∣ Π r ∣ 2 ∝ t − 2 , uniformly in r r r inside the cone ∣ r ∣ < v max t |r|<v_{\max}t ∣ r ∣ < v m a x t , so the denominator decays as 1 / t 1/t 1/ t while the numerator, with r 2 ∝ t 2 r^2\propto t^2 r 2 ∝ t 2 , grows as t t t . The spread is ballistic, R 2 ( k , t ) = v b 2 t 2 R^2(k,t)=v_b^2t^2 R 2 ( k , t ) = v b 2 t 2 with
v b 2 = ∫ d q d q ′ v q 2 S ( q , q ′ ) 2 δ ( v q + v q ′ ) ∫ d q d q ′ S ( q , q ′ ) 2 δ ( v q + v q ′ ) . (56) v_b^2 = \frac{\int dq\,dq'\,v_q^2S(q,q')^2\delta(v_q+v_{q'})}{\int dq\,dq'\,S(q,q')^2\delta(v_q+v_{q'})}. \tag*{(56)} v b 2 = ∫ d q d q ′ S ( q , q ′ ) 2 δ ( v q + v q ′ ) ∫ d q d q ′ v q 2 S ( q , q ′ ) 2 δ ( v q + v q ′ ) . (56)
the mean-square group velocity of the co-moving pairs, weighted by their creation amplitude. At k = 1 k = 1 k = 1 Eq. (56) gives v b 2 = 5.55 × 10 − 3 v_b^2 = 5.55\times10^{-3} v b 2 = 5.55 × 1 0 − 3 for the dipolar kernel at V = 0.3 V = 0.3 V = 0.3 and 5.38 × 10 − 3 5.38\times10^{-3} 5.38 × 1 0 − 3 for nearest neighbors at the matched bond, against v max = 0.162 v_{\max}=0.162 v m a x = 0.162 and 0.164 0.164 0.164 ; the profile is roughly uniform inside the cone with a sharp front at v max t v_{\max}t v m a x t , and Figure 5 shows R 2 ( k , t ) R^2(k,t) R 2 ( k , t ) on the ballistic line over two decades in t t t before the ring closes at t ≃ N t / 2 v max t\simeq N_t/2v_{\max} t ≃ N t /2 v m a x and R 2 R^2 R 2 settles on the flat-ring value N t 2 / 12 N_t^2/12 N t 2 /12 . At t ≲ 10 t\lesssim10 t ≲ 10 the profile is still the equal-time one, R 2 ≃ 1 R^2\simeq1 R 2 ≃ 1 , and the first departure is quadratic in t t t with a smaller coefficient fixed by the energy moments ∑ q ( E q , k + E q + Q , k ) n S \sum_q(E_{q,k}+E_{q+Q,k})^nS ∑ q ( E q , k + E q + Q , k ) n S , n ≤ 2 n\leq2 n ≤ 2 , which we have checked against the numerics but do not need. Nothing in Eq. (56) is dissipative: v b v_b v b is set by the Bogoliubov dispersion alone, and the only decay in the uniform array, that of the denominator, is the separation of the two quasiparticles of a pair. With disorder the same R 2 ( k , t ) R^2(k,t) R 2 ( k , t ) follows the ballistic line until t ∼ ξ l o c / v b t\sim\xi_{\mathrm{loc}}/v_b t ∼ ξ loc / v b and then saturates (Figure 5 , Sec. IV D).
Figure 5. Spatial variance R 2 ( k , t ) R^2(k,t) R 2 ( k , t ) of the intensity profile, Eq. (54) , at k = 1 k=1 k = 1 , N t = 1000 N_t=1000 N t = 1000 : uniform dipolar (V = 0.3 V=0.3 V = 0.3 ) and n.n. (V = 0.181 V=0.181 V = 0.181 ) arrays, the ballistic law v b 2 t 2 v_b^2t^2 v b 2 t 2 of Eq. (56) (dotted), the flat-ring value N t 2 / 12 N_t^2/12 N t 2 /12 (dash-dotted), and disordered dipolar arrays at three σ \sigma σ (4 realizations, 20 source tubes each), which leave the ballistic line at t ∼ ξ l o c / v b t\sim\xi_{\mathrm{loc}}/v_b t ∼ ξ loc / v b and saturate at the values of Figure 10 .
Disordered arrays
Density imbalance turns G ^ ( k ) \hat{G}(k) G ^ ( k ) into a one-dimensional disordered matrix in the tube index, with the diagonal U c n i U_c n_i U c n i and the bonds V i j ( k ) n i n j V_{ij}(k)\sqrt{n_i n_j} V ij ( k ) n i n j both random. Each k k k is a separate localization problem, and for the dipolar kernel the range of G ^ ( k ) \hat{G}(k) G ^ ( k ) is 1 / k 1/k 1/ k , so the eigenvectors cannot decay faster than the bonds: the localization rate is bounded, γ ( k ) ≤ ∣ k ∣ \gamma(k)\leq|k| γ ( k ) ≤ ∣ k ∣ , and at k → 0 k\to0 k → 0 , where the bonds fall as 1 / r 2 1/r^2 1/ r 2 , localization is at most algebraic. The analysis below is at fixed k k k , mostly k = 1 k=1 k = 1 , where γ \gamma γ is well inside this bound except at the strongest disorder (Figure 8 ); the k k k integration is performed only where the uniform-array asymptotics require it. We write Π i j ( k , t ) \Pi_{ij}(k,t) Π ij ( k , t ) and Π Q ( k , t ) \Pi_Q(k,t) Π Q ( k , t ) for the correlator at fixed k k k , and Π i j ( t ) \Pi_{ij}(t) Π ij ( t ) , Π Q ( t ) \Pi_Q(t) Π Q ( t ) for its k k k integral; in this section an unadorned Π i j \Pi_{ij} Π ij inside a long-time average is the fixed-k k k object.
Figure 8. Localization rate from three diagnostics at k = 1 k=1 k = 1 , N t = 1000 N_t=1000 N t = 1000 , 100 realizations; dipolar at V = 0.3 V=0.3 V = 0.3 (left), n.n. at the matched bond 0.181 (right). Top: transfer-matrix γ \gamma γ (black), one quarter of the decay rate of Eq. (63) (green), 2 I t y p 2I_{\mathrm{typ}} 2 I typ (red); dashed, Eq. (58) ; dotted, the clean IPR floor; dash-dotted (left), the kernel bound γ ≤ ∣ k ∣ \gamma\le|k| γ ≤ ∣ k ∣ . Bottom: the same two diagnostics divided by γ \gamma γ .
Localization of the collective modes
We characterize the transverse extent of mode m m m by its inverse participation ratio (IPR),
I m = ∑ i = 1 N t ( M i , k m m ) 2 = ∑ i = 1 N t ( U g ) i m 4 , (57) \mathcal{I}_m=\sum_{i=1}^{N_t}\left(\mathcal{M}_{i,k}^{mm}\right)^2=\sum_{i=1}^{N_t}(U^g)_{im}^4, \tag*{(57)} I m = i = 1 ∑ N t ( M i , k mm ) 2 = i = 1 ∑ N t ( U g ) im 4 , (57)
a property of the eigenvectors of G ^ \hat{G} G ^ alone. In the uniform array real standing waves give I m = 3 / 2 N t \mathcal{I}_m=3/2N_t I m = 3/2 N t (a complex plane wave would give 1 / N t 1/N_t 1/ N t ); Figure 6 shows I \mathcal{I} I moving off this floor to an N t N_t N t -independent value as σ \sigma σ increases, for both kernels and several k k k .
Figure 6. Typical IPR, Eq. (57) , of band-center modes against N t N_t N t at V = 0.3 V = 0.3 V = 0.3 , for the n.n. kernel and the dipolar kernel at three k k k ; color gives σ \sigma σ , 100 realizations per point. Dashed: the clean value 3 / 2 N t 3/2N_t 3/2 N t .
The asymptotic localization rate γ = 1 / ξ l o c \gamma=1/\xi_{\mathrm{loc}} γ = 1/ ξ loc is obtained from the transfer matrix of G ^ ( k ) \hat{G}(k) G ^ ( k ) (Appendix I3). It depends on where the mode sits in the band, and γ \gamma γ denotes its minimum over the band: the most extended modes dominate the long-distance tail of the correlator, so that is the rate the correlator measures. At weak disorder the rate at the band center g = n 0 U c g=n_0U_c g = n 0 U c is given by the Born approximation,
γ B o r n = σ 2 n 0 2 U c 2 8 t e f f 2 , t e f f = ∣ ∑ r r t r sin ( q c r ) ∣ , (58) \gamma_{\mathrm{Born}}=\frac{\sigma^2 n_0^2 U_c^2}{8t_{\mathrm{eff}}^2},\qquad t_{\mathrm{eff}}=\left|\sum_r r t_r\sin(q_c r)\right|, \tag*{(58)} γ Born = 8 t eff 2 σ 2 n 0 2 U c 2 , t eff = r ∑ r t r sin ( q c r ) , (58)
where t r = n 0 V r ( k ) t_r=n_0V_r(k) t r = n 0 V r ( k ) is the uniform-array bond at separation r r r and q c q_c q c the band-center quasimomentum, g c = n 0 U c g_c=n_0U_c g c = n 0 U c , so that 2 t e f f = ∣ ∂ q g ∣ q c 2t_{\mathrm{eff}}=|\,\partial_q g\,|_{q_c} 2 t eff = ∣ ∂ q g ∣ q c is the band-center group velocity and Eq. (58) is Eq. (19) evaluated there (Appendix I3); for nearest neighbors q c = π / 2 q_c=\pi/2 q c = π /2 and t e f f = t 1 t_{\mathrm{eff}}=t_1 t eff = t 1 . Since t e f f ∝ V t_{\mathrm{eff}}\propto V t eff ∝ V , γ \gamma γ depends on σ \sigma σ and V V V only through σ n 0 U c / t e f f \sigma n_0U_c/t_{\mathrm{eff}} σ n 0 U c / t eff in this regime; Figure 7 shows the disorder sweep, the coupling scan, and both kernels collapsing onto Eq. (58) . Beyond the Born regime the kernels part: the dipolar rate falls below the line, the n.n. rate rises above it, and in both cases σ \sigma σ and V V V no longer enter through one combination, because the effective random potential g ( 1 / n i − 1 / n 0 ) g(1/n_i-1/n_0) g ( 1/ n i − 1/ n 0 ) is not Gaussian and carries no factor of V V V at second order (Appendix I3). Here and in all localization figures the n.n. coupling is matched to the dipolar first-neighbor bond at the k k k used, V n n = V k K 1 ( k ) = 0.181 V_{\mathrm{nn}}=VkK_1(k)=0.181 V nn = V k K 1 ( k ) = 0.181 at k = 1 k=1 k = 1 , so that the two arrays are compared at equal bond and comparable N t / ξ l o c N_t/\xi_{\mathrm{loc}} N t / ξ loc .
Figure 7. Localization rate at the band center against σ n 0 U c / t e f f \sigma n_0 U_c/t_{\mathrm{eff}} σ n 0 U c / t eff , k = 1 k = 1 k = 1 . Filled circles: disorder sweep at V = 0.3 V = 0.3 V = 0.3 ; open symbols: coupling scan at V = 0.15 V = 0.15 V = 0.15 and 0.1 0.1 0.1 . Dashed: Eq. (58) .
The IPR follows the disorder dependence of γ \gamma γ , reproducing its power law and magnitude, but 2 I t y p / γ 2\mathcal{I}_{\mathrm{typ}}/\gamma 2 I typ / γ drifts across the disorder range (Figure 8 ): the fourth moment in Eq. (57) weights the log-normal amplitude fluctuations riding on the envelope (Appendix I4), and I t y p \mathcal{I}_{\mathrm{typ}} I typ averages over the band whereas γ \gamma γ is its minimum. The drift is systematic, so the IPR detects localization but does not measure γ \gamma γ ; we quote the transfer-matrix rate throughout.
Long-time momentum correlator
For one disorder realization the fixed-k k k correlator is
Π i j ( k , t ) = ∑ m , n c i j m n e i ε m n , k t , (59) \Pi_{ij}(k,t) = \sum_{m,n} c_{ij}^{mn} e^{i\varepsilon_{mn,k}t}, \tag*{(59)} Π ij ( k , t ) = m , n ∑ c ij mn e i ε mn , k t , (59)
with c i j m n c_{ij}^{mn} c ij mn the overlap and coherence factors of Eq. (27) , and for a nondegenerate spectrum its infinite-time average is
⟨ ∣ Π i j ∣ 2 ⟩ t = 2 ∑ m , n ∣ c i j m n ∣ 2 (60) \left\langle\left|\Pi_{ij}\right|^2 \right\rangle_t = 2\sum_{m,n} \left|c_{ij}^{mn}\right|^2 \tag*{(60)} ⟨ ∣ Π ij ∣ 2 ⟩ t = 2 m , n ∑ c ij mn 2 (60)
(the factor two from c i j m n = c j i n m c_{ij}^{mn}=c_{ji}^{nm} c ij mn = c j i nm ; c i i m m = 0 c_{ii}^{mm}=0 c ii mm = 0 ), evaluated in closed form without a time grid (Appendix § 0.1 ). A nonlinear functional of the time record is needed: at any fixed t t t the correlator is a smooth matrix function of G ^ \hat{G} G ^ , Eq. (81) , whose decay in ∣ i − j ∣ |i-j| ∣ i − j ∣ is set by the range of the kernel and the spectrum of G ^ \hat{G} G ^ , not by the localization of its eigenvectors. The time average kills all cross terms and leaves a sum of squared mode amplitudes, which does decay at the localization rate. A localized eigenvector has envelope e − γ r e^{-\gamma r} e − γ r ; the pair overlap (29) contains two amplitudes connecting i i i to j j j , so
∣ Π i j ∣ ∼ e − 2 γ ∣ i − j ∣ (61) \left|\Pi_{ij}\right| \sim e^{-2\gamma|i-j|} \tag*{(61)} ∣ Π ij ∣ ∼ e − 2 γ ∣ i − j ∣ (61)
and
⟨ ∣ Π i j ∣ 2 ⟩ t ∼ e − 4 γ ∣ i − j ∣ . (62) \left\langle\left|\Pi_{ij}\right|^2 \right\rangle_t \sim e^{-4\gamma|i-j|}. \tag*{(62)} ⟨ ∣ Π ij ∣ 2 ⟩ t ∼ e − 4 γ ∣ i − j ∣ . (62)
We average ln ⟨ ∣ Π i j ∣ 2 ⟩ t \ln\langle|\Pi_{ij}|^2\rangle_t ln ⟨ ∣ Π ij ∣ 2 ⟩ t over realizations and over tube pairs at fixed separation (the annealed average is dominated by rare slowly decaying realizations and does not converge with the fit window), and plot
ln [ ⟨ ∣ Π i j ∣ 2 ⟩ t ⟨ ∣ Π i j ∣ 2 ⟩ t 0 ] , (63) \ln\left[\frac{\langle|\Pi_{ij}|^2\rangle_t}{\langle|\Pi_{ij}|^2\rangle_{t_0}}\right], \tag*{(63)} ln [ ⟨ ∣ Π ij ∣ 2 ⟩ t 0 ⟨ ∣ Π ij ∣ 2 ⟩ t ] , (63)
where the uniform-array denominator is independent of ∣ i − j ∣ |i-j| ∣ i − j ∣ beyond a few sites (the dephased profile of the uniform array is flat on the ring, Sec. IV D) and only fixes the overall scale. The rate is fitted over 3 ≤ d / ξ l o c ≤ 8 3 \le d/\xi_{\mathrm{loc}} \le8 3 ≤ d / ξ loc ≤ 8 (Appendix § 0.1 ).
Figure 8 shows the correlator rate approaching 4 γ 4\gamma 4 γ as N t / ξ l o c N_t/\xi_{\mathrm{loc}} N t / ξ loc grows, reaching it at the strongest disorder, where γ \gamma γ also comes closest to the kernel bound ∣ k ∣ |k| ∣ k ∣ (dash-dotted line in Figure 8 ). At weak disorder the fit window is capped by the ring and the minimum-image profile flattens, biasing the fitted slope low; the correlator is an unbiased measurement of γ \gamma γ for N t ≫ ξ l o c N_t \gg\xi_{\mathrm{loc}} N t ≫ ξ loc .
Equal-time response
On the disordered ring Π Q = ∑ i j e i Q ( i − j ) Π i j \Pi_Q = \sum_{ij} e^{iQ(i-j)}\Pi_{ij} Π Q = ∑ ij e i Q ( i − j ) Π ij with Q = 2 π n / N t Q = 2\pi n/N_t Q = 2 π n / N t remains well defined, although Q Q Q no longer labels the modes, and
Π Q ( k , 0 ) = k 2 ∑ m , n sinh 2 ϑ m n , k ∣ F m n ( Q ) ∣ 2 , \Pi_Q(k,0) = k^2 \sum_{m,n} \sinh^2\vartheta_{mn,k}\left|F^{mn}(Q)\right|^2, Π Q ( k , 0 ) = k 2 m , n ∑ sinh 2 ϑ mn , k ∣ F mn ( Q ) ∣ 2 ,
F m n ( Q ) = ∑ i e i Q i ( U g ) i m ( U g ) i n . (64) F^{mn}(Q) = \sum_i e^{iQi}(U_g)_{im}(U_g)_{in}. \tag*{(64)} F mn ( Q ) = i ∑ e i Q i ( U g ) im ( U g ) in . (64)
Expanding e i Q ( i − j ) e^{iQ(i-j)} e i Q ( i − j ) , the constant term vanishes by the sum rule (44) and the linear one by symmetry, so
∫ 0 Λ d k Π Q ( k , 0 ) = Q 2 η + O ( Q 4 ) , \int_0^\Lambda dk\,\Pi_Q(k,0)=Q^2\eta+O(Q^4), ∫ 0 Λ d k Π Q ( k , 0 ) = Q 2 η + O ( Q 4 ) ,
η = − 1 2 ∫ 0 Λ d k ∑ i , j d i j 2 Π i j ( k , 0 ) , (65) \eta=-\frac{1}{2}\int_0^\Lambda dk\sum_{i,j}d_{ij}^2\Pi_{ij}(k,0), \tag*{(65)} η = − 2 1 ∫ 0 Λ d k i , j ∑ d ij 2 Π ij ( k , 0 ) , (65)
with d i j = min ( ∣ i − j ∣ , N t − ∣ i − j ∣ ) d_{ij}=\min(|i-j|,N_t-|i-j|) d ij = min ( ∣ i − j ∣ , N t − ∣ i − j ∣ ) the minimum-image separation; in the uniform array this returns A Λ ( 0 ) A_\Lambda(0) A Λ ( 0 ) of Eq. (50) up to the factor N t / 2 π N_t/2\pi N t /2 π (Appendix H). Denoting by η ( σ ) \eta(\sigma) η ( σ ) the average of Eq. (65) over realizations at disorder strength σ \sigma σ , at small disorder
η ( σ ) η ( 0 ) = 1 − c σ 2 + O ( σ 4 ) , (66) \frac{\eta(\sigma)}{\eta(0)}=1-c\sigma^2+O(\sigma^4), \tag*{(66)} η ( 0 ) η ( σ ) = 1 − c σ 2 + O ( σ 4 ) , (66)
where the computed suppression is independent of N t N_t N t to within a few percent (Figure 14 ) and of V V V to within 1% over V = 0.3 V=0.3 V = 0.3 –0.05 0.05 0.05 .
Figure 14. Suppression 1 − η ( σ ) / η ( 0 ) 1-\eta(\sigma)/\eta(0) 1 − η ( σ ) / η ( 0 ) against N t N_t N t , normalized to its value at N t = 125 N_t=125 N t = 125 , for four σ \sigma σ ; dipolar solid, n.n. dashed, V = 0.3 V=0.3 V = 0.3 . Antithetic sampling as described in the text.
This weak, size- and coupling-independent suppression follows from perturbation theory in the inter-tube coupling. Taking the diagonal of G ^ \hat{G} G ^ , g i = n i U c g_i=n_iU_c g i = n i U c , as the unperturbed problem, each mode sits on a single tube, and the off-diagonal element G ^ i j \hat{G}_{ij} G ^ ij mixes tubes i i i and j j j with amplitude G ^ i j / ( g i − g j ) \hat{G}_{ij}/(g_i-g_j) G ^ ij / ( g i − g j ) . The overlap (29) needs one such amplitude in each of its two modes, so the leading term of Π i j ( k , 0 ) \Pi_{ij}(k,0) Π ij ( k , 0 ) is second order in G ^ i j \hat{G}_{ij} G ^ ij and involves only the two tubes i i i and j j j ,
Π i j ( k , 0 ) ≃ − 2 k 2 G ^ i j 2 sinh 2 ( θ i , k − θ j , k ) ( g i − g j ) 2 , (67) \Pi_{ij}(k,0)\simeq-2k^2\hat{G}_{ij}^2\frac{\sinh^2(\theta_{i,k}-\theta_{j,k})}{(g_i-g_j)^2}, \tag*{(67)} Π ij ( k , 0 ) ≃ − 2 k 2 G ^ ij 2 ( g i − g j ) 2 sinh 2 ( θ i , k − θ j , k ) , (67)
with g i = n i U c g_i=n_iU_c g i = n i U c , G ^ i j = n i n j V i j ( k ) \hat{G}_{ij}=\sqrt{n_i n_j}V_{ij}(k) G ^ ij = n i n j V ij ( k ) , and θ i , k \theta_{i,k} θ i , k the squeeze angle of Eq. (19) at g = g i g=g_i g = g i ; the ratio stays finite as g i → g j g_i\to g_j g i → g j , where it agrees to the same order with the exact two-tube result. Through Eq. (65) , η \eta η is then a sum over pairs, weighted by d i j 2 V i j ( k ) 2 d_{ij}^2V_{ij}(k)^2 d ij 2 V ij ( k ) 2 , of a function of n i n_i n i and n j n_j n j alone, and its disorder average gives Eq. (66) with c c c independent of V V V and of N t N_t N t by construction. Evaluated with the distribution (8), at the cutoffs of Appendix I 5, it gives c = 0.40 c=0.40 c = 0.40 (dipolar) and 0.27 0.27 0.27 (n.n.), and with no free parameter it follows the computed η ( σ ) / η ( 0 ) \eta(\sigma)/\eta(0) η ( σ ) / η ( 0 ) over the whole disorder range, Figure 9 , including the O ( σ 4 ) O(\sigma^4) O ( σ 4 ) correction that steepens the dipolar curve. The equal-time response is set by the interaction range because it is the momentum profile before anything has propagated; the localization length does not enter.
Figure 9. Disorder dependence of the equal-time coefficient η \eta η , Eq. (65) , at N t = 1000 N_t=1000 N t = 1000 and V = 0.3 V=0.3 V = 0.3 . Points: full calculation, 100 realizations. Lines: the two-site average of Eq. (67) , with no free parameter. Cutoffs as in Appendix I 5.
Transverse momentum spreading
The saturated spread of the momentum profile is the long-time limit of the spatial variance (54), with numerator and denominator averaged separately,
R ∞ 2 ( σ , N t ) = ∑ j ≠ i d i j 2 ⟨ ∣ Π i j ∣ 2 ⟩ t ∑ j ≠ i ⟨ ∣ Π i j ∣ 2 ⟩ t , (68) R_\infty^2(\sigma,N_t)=\frac{\sum_{j\ne i}d_{ij}^2\langle| \Pi_{ij}|^2\rangle_t}{\sum_{j\ne i}\langle|\Pi_{ij}|^2\rangle_t}, \tag*{(68)} R ∞ 2 ( σ , N t ) = ∑ j = i ⟨ ∣ Π ij ∣ 2 ⟩ t ∑ j = i d ij 2 ⟨ ∣ Π ij ∣ 2 ⟩ t , (68)
with the same long-time average as Sec. IV B. In the uniform array the dephased profile is uniform on the ring,
R ∞ 2 ≃ N t 2 12 , (69) R_\infty^2\simeq\frac{N_t^2}{12}, \tag*{(69)} R ∞ 2 ≃ 12 N t 2 , (69)
for either kernel; at any finite σ \sigma σ it saturates at an N t N_t N t -independent value that decreases with γ \gamma γ , Figure 10 , reached from the ballistic law of Sec. III B at t ∼ ξ l o c / v b t\sim\xi_{\mathrm{loc}}/v_b t ∼ ξ loc / v b (Figure 5 ). The spread is ballistic on the ring in the uniform array and arrested at ξ l o c \xi_{\mathrm{loc}} ξ loc when disordered. In both cases it is propagation in the pair channel of Sec. III A 3; a dc momentum conductivity would need the intraband channel E m − E n E_m-E_n E m − E n , which opens at finite temperature with the same overlaps M i , k m n \mathcal{M}_{i,k}^{mn} M i , k mn , and we do not pursue it here.
Figure 10. Saturated transverse spread R ∞ 2 R_{\infty}^{2} R ∞ 2 , Eq. (68) , against N t N_t N t at k = 1 k=1 k = 1 ; color gives σ \sigma σ . Solid, filled: dipolar at V = 0.3 V=0.3 V = 0.3 ; dashed, open: n.n. at the matched bond 0.181 0.181 0.181 . Dotted: N t 2 / 12 N_t^2/12 N t 2 /12 .
Instability of the uniform condensate
By Sec. II C the stability of the uniform condensate is decided by the interaction matrix of the uniform array. Its minimum eigenvalue is at q = π q=\pi q = π for both kernels,
g π = n 0 ( U c − 2 V ) (70) g_{\pi}=n_0(U_c-2V) \tag*{(70)} g π = n 0 ( U c − 2 V ) (70)
for n.n. coupling and, at k → 0 k\to0 k → 0 ,
g π = n 0 ( U c − π 2 6 V ) (71) g_{\pi}=n_0\left(U_c-\frac{\pi^2}{6}V\right) \tag*{(71)} g π = n 0 ( U c − 6 π 2 V ) (71)
for dipolar coupling, from the alternating sum 2 ∑ r ( − 1 ) r / r 2 = − π 2 / 6 2\sum_r(-1)^r/r^2=-\pi^2/6 2 ∑ r ( − 1 ) r / r 2 = − π 2 /6 . The critical couplings are
V c n n = U c 2 , V c d i p = 6 U c π 2 ≃ 0.60 U c , (72) V_c^{\mathrm{nn}}=\frac{U_c}{2},\qquad V_c^{\mathrm{dip}}=\frac{6U_c}{\pi^2}\simeq0.60U_c, \tag*{(72)} V c nn = 2 U c , V c dip = π 2 6 U c ≃ 0.60 U c , (72)
the dipolar one being the collapse threshold of the Gaussian sliding-liquid theory of Ref. [20 ] . The long-range kernel tolerates the larger coupling because tubes at even separation sit on the same sublattice of the q = π q=\pi q = π pattern and stabilize it, which is why the alternating sum, − 1.64 -1.64 − 1.64 , is smaller in magnitude than the − 2 -2 − 2 of a single bond. The dipolar g π ( k ) g_{\pi}(k) g π ( k ) rises with k k k (the threshold is
0.703 U c 0.703U_c 0.703 U c at k = 0.5 k=0.5 k = 0.5 ), so the first unstable mode is at k → 0 k\to0 k → 0 , q = π q=\pi q = π . Density imbalance lowers the smallest eigenvalue, by an order of magnitude across the accessible range (Figure 11 ), without ever changing its sign, as Eq. (23) requires; the threshold is the same at every σ \sigma σ to four digits.
Figure 11. Disorder-averaged minimum eigenvalue of G ^ \hat{G} G ^ at V = 0.3 V=0.3 V = 0.3 , k = 1 k=1 k = 1 , N t = 1000 N_t=1000 N t = 1000 , 100 realizations.
Unstable modes and the state beyond the threshold
For g π < 0 g_{\pi}<0 g π < 0 the modes with
ε k < − 2 g π ( k ) , i.e. k 2 < − 4 g π , (73) \varepsilon_k<-2g_{\pi}(k),\qquad\text{i.e.}\quad k^2<-4g_{\pi}, \tag*{(73)} ε k < − 2 g π ( k ) , i.e. k 2 < − 4 g π , (73)
have imaginary E π , k E_{\pi,k} E π , k and grow at the rate Γ k = ε k ( − ε k − 2 g π ) \Gamma_k=\sqrt{\varepsilon_k(-\varepsilon_k-2g_{\pi})} Γ k = ε k ( − ε k − 2 g π ) , which vanishes at k → 0 k\to0 k → 0 and is largest at ε k = − g π \varepsilon_k=-g_{\pi} ε k = − g π , Γ max = ∣ g π ∣ \Gamma_{\max}=|g_{\pi}| Γ m a x = ∣ g π ∣ (for the dipolar kernel, with g π ( k ) g_{\pi}(k) g π ( k ) rising with k k k , the maximum sits at somewhat smaller k k k ). The unstable pattern is therefore two-dimensional in the ( x , i ) (x,i) ( x , i ) plane: a longitudinal density wave of wavevector k ∗ ≃ − 2 g π k_*\simeq\sqrt{-2g_{\pi}} k ∗ ≃ − 2 g π in each tube, antiphase between neighboring tubes. A purely transverse modulation, k = 0 k=0 k = 0 with alternate tubes fuller and emptier, is not available, since each N ^ i \hat{N}_i N ^ i is conserved; the transverse q = π q=\pi q = π pattern can only be realized by redistributing density along the tubes. The quadratic theory says nothing about the final state. The natural endpoint, with every k < − 4 g π k<\sqrt{-4g_{\pi}} k < − 4 g π unstable, is coarsening to phase separation along each tube into dense and dilute segments, staggered between neighbors, as in the miscible–immiscible transition of a two-component condensate to which the problem reduces at N t = 2 N_t=2 N t = 2 [37 , 38 , 39 ] (threshold V = U c V=U_c V = U c there, U c / 2 U_c/2 U c /2 in the extended array because both neighbors soften the q = π q=\pi q = π mode). Whether a periodic density wave at k ∗ k_* k ∗ can instead be stabilized requires the terms dropped from Eq. (10) ; Ref. [20 ] describes the state past the collapse as bunching within the tubes. The modulation energy of such a state is a Gross–Pitaevskii problem in the extended model, not a refinement of this one.
A modeling caveat: U c U_c U c here is purely intra-tube and V V V purely inter-tube, so V / U c V/U_c V / U c is an anisotropy ratio. A physical dipolar gas also has a dipolar contribution to the intra-tube coupling [20 ] , which shifts U c U_c U c and hence the threshold in terms of the bare dipolar strength; the “dipolar stronger than contact” regime is in any case not reached, since the array orders at V d ≃ 0.30 U c V_d \simeq0.30 U_c V d ≃ 0.30 U c .
Outlook
We have developed a novel approach to characterize transverse momentum transport using both standard average correlator (termed “amplitude”) and its variance (“intensity”). Both can be obtained in an experimental setup where momentum is imparted locally to one tube, with other tubes measured at later times, shot by shot, so that the same set of shots can be used to obtain both the amplitude and the intensity. Since each loading is a new realization of the density imbalance, the shot average performs the disorder average, and the two quantities measure the two sides of the paper: the amplitude obeys the sum rule and is suppressed by disorder, the intensity carries the ballistic spread v b 2 t 2 v_b^2t^2 v b 2 t 2 of the uniform array and its saturation at ξ l o c \xi_{\mathrm{loc}} ξ loc in the disordered one. On the theory side, the most immediate extension is to include heretofore neglected effects of finite temperature and interchain hopping. We expect a more conventional “normal” viscosity to emerge from incoherent hopping of excitations.
Acknowledgments
Claude (Anthropic; Opus 5.5 and Fable 5.1) was used throughout this work: to debug and extend the authors’ numerical routines, to write most of the analysis and plotting code, to assist with derivations, including the closed form of the uniform-array correlator and the stationary-phase analysis of Sec. III B, to check formulas, and to draft and restructure parts of the text. Figures are plots of computed data produced by this code. The authors directed and verified all AI-assisted output; the numerical and analytical checks are described in Appendix I.
Time and length scales
Numbers are for 162 D y {}^{162}\mathrm{Dy} 162 Dy in the lattice of Ref. [10 ] : λ = 741 n m \lambda= 741\ \mathrm{nm} λ = 741 nm (E R = h 2 / 2 M λ 2 = h × 2.24 k H z E_R = h^2/2M\lambda^2 = h \times2.24\ \mathrm{kHz} E R = h 2 /2 M λ 2 = h × 2.24 kHz ), lattice depth s ≡ V 0 / E R = 18 s \equiv V_0/E_R = 18 s ≡ V 0 / E R = 18 unless stated, transverse confinement ω ⊥ = 2 s E R / ℏ = 2 π × 19 k H z \omega_\perp= 2\sqrt{s}E_R/\hbar= 2\pi\times19\ \mathrm{kHz} ω ⊥ = 2 s E R /ℏ = 2 π × 19 kHz , a s = 100 a 0 a_s = 100\,a_0 a s = 100 a 0 , n 0 = 3 μ m − 1 n_0 = 3\ \mu\mathrm{m}^{-1} n 0 = 3 μ m − 1 , μ = 9.9 μ B \mu= 9.9\,\mu_B μ = 9.9 μ B , dipoles polarized normal to the plane of the array. Order-unity factors are tracked only where stated.
Healing length ξ ≡ ℏ / M n 0 U c ≃ 0.32 μ m \xi\equiv\hbar/\sqrt{M n_0 U_c} \simeq0.32\,\mu\mathrm{m} ξ ≡ ℏ/ M n 0 U c ≃ 0.32 μ m ; sound velocity v = n 0 U c / M ≃ 1.2 m m / s v = \sqrt{n_0 U_c/M} \simeq1.2\,\mathrm{mm/s} v = n 0 U c / M ≃ 1.2 mm/s ; ξ / v = ℏ / n 0 U c \xi/v = \hbar/n_0 U_c ξ / v = ℏ/ n 0 U c .
Intra-tube interaction energy g U ≡ n 0 U c = 2 ℏ ω ⊥ a s n 0 = h × 600 H z g_U \equiv n_0 U_c = 2\hbar\omega_\perp a_s n_0 = h \times600\ \mathrm{Hz} g U ≡ n 0 U c = 2ℏ ω ⊥ a s n 0 = h × 600 Hz (Lieb–Liniger γ ≃ 1.1 \gamma\simeq1.1 γ ≃ 1.1 ). The short time scale is t s h o r t ≡ ℏ / g U = 0.26 m s t_{\mathrm{short}} \equiv\hbar/g_U = 0.26\ \mathrm{ms} t short ≡ ℏ/ g U = 0.26 ms : the phonon at k = 1 / ξ k = 1/\xi k = 1/ ξ has E k = ϵ k ( ϵ k + 2 g U ) = 1.12 g U E_k = \sqrt{\epsilon_k(\epsilon_k + 2g_U)} = 1.12\,g_U E k = ϵ k ( ϵ k + 2 g U ) = 1.12 g U with ϵ k = g U / 2 \epsilon_k = g_U/2 ϵ k = g U /2 , so t s h o r t t_{\mathrm{short}} t short is 1 / 2 π 1/2\pi 1/2 π of its period, and the crossover from linear to quadratic dispersion sits at k ξ ≃ 2 k\xi\simeq2 k ξ ≃ 2 .
Tube spacing a = λ / 2 = 371 n m = 1.15 ξ a = \lambda/2 = 371\ \mathrm{nm} = 1.15\,\xi a = λ /2 = 371 nm = 1.15 ξ . This is not a coincidence of one experiment: for a single retro-reflected lattice a / ξ = π 2 s a s n 0 a/\xi= \pi\sqrt{2\sqrt{s}a_s n_0} a / ξ = π 2 s a s n 0 , independent of λ \lambda λ .
Nearest-tube bond g V ≡ V n 0 = 2 V d n 0 / a 2 = h × 56 H z g_V \equiv Vn_0 = 2V_d n_0/a^2 = h \times56\ \mathrm{Hz} g V ≡ V n 0 = 2 V d n 0 / a 2 = h × 56 Hz , so g V / g U = 0.093 g_V/g_U = 0.093 g V / g U = 0.093 ; t V ≡ ℏ / g V = 2.9 m s t_V \equiv\hbar/g_V = 2.9\ \mathrm{ms} t V ≡ ℏ/ g V = 2.9 ms , τ V ≡ t V / t s h o r t = 10.8 \tau_V \equiv t_V/t_{\mathrm{short}} = 10.8 τ V ≡ t V / t short = 10.8 . At k = 1 / ξ k = 1/\xi k = 1/ ξ the bond is reduced by κ ( k a ) K 1 ( k a ) = 0.54 \kappa(ka)K_1(ka) = 0.54 κ ( k a ) K 1 ( k a ) = 0.54 , to h × 30 H z h \times30\ \mathrm{Hz} h × 30 Hz . The ratio V / U c = 4 M V d / ( s ℏ 2 a s ) V/U_c = 4MV_d/(\sqrt{s}\hbar^2a_s) V / U c = 4 M V d / ( s ℏ 2 a s ) is likewise independent of λ \lambda λ ; it is set by s s s and a s a_s a s . The transverse dispersion of the collective modes is generated by g V g_V g V through g q ( k ) = g U + 2 n 0 ∑ r V r ( k ) cos q r g_q(k) = g_U + 2n_0\sum_r V_r(k)\cos qr g q ( k ) = g U + 2 n 0 ∑ r V r ( k ) cos q r .
Transverse dispersion. As k → 0 k \to0 k → 0 , E q , k = ℏ k g q / M E_{q,k} = \hbar k\sqrt{g_q/M} E q , k = ℏ k g q / M with g 0 = g U + ( π 2 / 3 ) g V g_0 = g_U + (\pi^2/3)g_V g 0 = g U + ( π 2 /3 ) g V and g π = g U − ( π 2 / 6 ) g V = 0.85 g U g_{\pi} = g_U - (\pi^2/6)g_V = 0.85\,g_U g π = g U − ( π 2 /6 ) g V = 0.85 g U , so Δ ω ( k ) ≡ E 0 , k − E π , k = 0.22 k v \Delta\omega(k) \equiv E_{0,k} - E_{\pi,k} = 0.22kv Δ ω ( k ) ≡ E 0 , k − E π , k = 0.22 k v . At k = 1 / ξ k = 1/\xi k = 1/ ξ , with the Bessel-reduced kernel and the full Bogoliubov form, Δ ω = 0.09 g U / ℏ = 2 π × 54 H z \Delta\omega= 0.09\,g_U/\hbar= 2\pi\times54\ \mathrm{Hz} Δ ω = 0.09 g U /ℏ = 2 π × 54 Hz , giving t Δ ≡ 1 / Δ ω = 2.9 m s ≃ τ V t_\Delta\equiv1/\Delta\omega= 2.9\ \mathrm{ms} \simeq\tau_V t Δ ≡ 1/Δ ω = 2.9 ms ≃ τ V : the transverse hop time per tube is ℏ / g V ( k ) \hbar/g_V(k) ℏ/ g V ( k ) .
Single-boson tunneling J J J between tubes, the t ⊥ t_\perp t ⊥ of Eq. (5), from the deep-lattice (Mathieu) asymptotics J / E R ≃ ( 4 / π ) s 3 / 4 e − 2 s J/E_R \simeq(4/\sqrt{\pi})s^{3/4}e^{-2\sqrt{s}} J / E R ≃ ( 4/ π ) s 3/4 e − 2 s : J / h = 9.0 H z J/h = 9.0\ \mathrm{Hz} J / h = 9.0 Hz at s = 18 s = 18 s = 18 (the value quoted in Ref. [10 ] ) and 0.020 H z 0.020\ \mathrm{Hz} 0.020 Hz at s = 60 s = 60 s = 60 .
Josephson frequency of the Bogoliubov mean-field state. With hopping, ϵ ~ ( q ) \tilde{\epsilon}(q) ϵ ~ ( q ) acquires ϵ ∥ ( q ) = 2 J ( 1 − cos q ) \epsilon_\parallel(q) = 2J(1-\cos q) ϵ ∥ ( q ) = 2 J ( 1 − cos q ) and the k = 0 k=0 k = 0 , q = π q=\pi q = π mode has E 2 = ϵ ⊥ ( ϵ ⊥ + 2 g π ) ≃ 8 J g π E^2 = \epsilon_\perp(\epsilon_\perp+ 2g_\pi) \simeq8Jg_\pi E 2 = ϵ ⊥ ( ϵ ⊥ + 2 g π ) ≃ 8 J g π , so ω J = 8 J g π / ℏ \omega_J = \sqrt{8Jg_\pi/\hbar} ω J = 8 J g π /ℏ , i.e. ω J / 2 π = 8 ( J / h ) ( g π / h ) ≃ 2.6 g U / h \omega_J/2\pi= \sqrt{8(J/h)(g_\pi/h)} \simeq2.6\sqrt{g_U/h} ω J /2 π = 8 ( J / h ) ( g π / h ) ≃ 2.6 g U / h (factor 16 in place of 8 for a square array). Values: 190 Hz at s = 18 s = 18 s = 18 , 9 Hz at s = 60 s = 60 s = 60 . The exponential suppression of J J J enters only through its square root.
Infrared cutoff on k k k . Setting Δ ω ( Λ I R ) = ω J \Delta\omega(\Lambda_{\mathrm{IR}}) = \omega_J Δ ω ( Λ IR ) = ω J with the small-k k k form of item 5 gives Λ I R ξ = ℏ ω J / ( 0.22 g U ) \Lambda_{\mathrm{IR}}\xi= \hbar\omega_J/(0.22g_U) Λ IR ξ = ℏ ω J / ( 0.22 g U ) : 1.4 at s = 18 s = 18 s = 18 and 0.07 at s = 60 s = 60 s = 60 . (Comparing instead the hopping correction to E q , k E_{q,k} E q , k , which grows as ϵ ∥ q / k v \epsilon_\parallel q/k v ϵ ∥ q / k v toward small k k k , with Δ ω \Delta\omega Δ ω gives the same crossover within a factor 1.5.) At s = 18 s = 18 s = 18 the entire phonon regime, including k ξ ≤ 1 k\xi\le1 k ξ ≤ 1 , is transversely locked.
Self-consistency of t ⊥ = 0 t_\perp= 0 t ⊥ = 0 requires Λ I R ξ ≪ 1 \Lambda_{\mathrm{IR}}\xi\ll1 Λ IR ξ ≪ 1 . Demanding Λ I R ξ = 0.1 \Lambda_{\mathrm{IR}}\xi= 0.1 Λ IR ξ = 0.1 gives $\omega_J/2\pi = 13\ \mathrm{Hz},
J / h = 0.04 J/h = 0.04 J / h = 0.04 Hz, J / E R = 1.9 × 10 − 5 J/E_R = 1.9 \times10^{-5} J / E R = 1.9 × 1 0 − 5 , i.e. s ≃ 54 s \simeq54 s ≃ 54 . Since Λ I R ξ ∝ J / g U \Lambda_{\mathrm{IR}}\xi\propto\sqrt{J/g_U} Λ IR ξ ∝ J / g U , lowering g U g_U g U to reach the Bogoliubov regime raises the required depth (γ ≃ 0.3 \gamma\simeq0.3 γ ≃ 0.3 at this density needs s ≃ 58 s \simeq58 s ≃ 58 ).
Disorder scale. Density imbalance enters G ^ \hat{G} G ^ as σ g U \sigma g_U σ g U on the diagonal and σ g V ( k ) \sigma g_V(k) σ g V ( k ) on the bonds. The fixed-k k k localization results additionally require 4 J ≪ σ g V ( k ) 4J \ll\sigma g_V(k) 4 J ≪ σ g V ( k ) : at s = 18 s = 18 s = 18 this is 36 Hz against 30 σ 30\sigma 30 σ Hz and fails for σ ≲ 1 \sigma\lesssim1 σ ≲ 1 ; at s = 60 s = 60 s = 60 it holds comfortably.
Temperature. At T > 0 T > 0 T > 0 the inter-tube phase locking of items 7–9 survives only below the temperature at which the Josephson coupling orders the array. In the random-phase approximation of [14 , 15 ] ,
( 2 π k B T c ℏ v n 0 ) 2 − 1 / ( 2 K ) = F ( K ) z J ℏ v n 0 , \left(\frac{2\pi k_B T_c}{\hbar v n_0}\right)^{2-1/(2K)} = F(K)\frac{zJ}{\hbar v n_0}, ( ℏ v n 0 2 π k B T c ) 2 − 1/ ( 2 K ) = F ( K ) ℏ v n 0 z J ,
with F ( K ) = π 2 A B ( K ) sin ( π / 4 K ) B 2 ( 1 / 8 K , 1 − 1 / 4 K ) F(K) = \pi^2 A_B(K)\sin(\pi/4K) B^2(1/8K,1 - 1/4K) F ( K ) = π 2 A B ( K ) sin ( π /4 K ) B 2 ( 1/8 K , 1 − 1/4 K ) , K = π ℏ n 0 / M v ≃ 3.0 K = \pi\hbar n_0/Mv \simeq3.0 K = π ℏ n 0 / M v ≃ 3.0 the Luttinger parameter, A B ≃ ( K / π ) 1 / ( 2 K ) A_B \simeq(K/\pi)^{1/(2K)} A B ≃ ( K / π ) 1/ ( 2 K ) , B B B the beta function and z z z the number of neighboring tubes; for K ≫ 1 K \gg1 K ≫ 1 it reduces to k B T c ≃ 2 π z J ℏ 2 n 0 2 / M k_B T_c \simeq2\pi\sqrt{zJ\hbar^2n_0^2/M} k B T c ≃ 2 π z J ℏ 2 n 0 2 / M . Above T c T_c T c the array is incoherently coupled, and the remaining condition is that k B T k_B T k B T stay below the collective-mode energies probed, h × 600 h \times600 h × 600 Hz = 30 nK at k ξ = 1 k\xi= 1 k ξ = 1 . For a chain of tubes (z = 2 z = 2 z = 2 ), k B T c = h × 750 k_B T_c = h \times750 k B T c = h × 750 Hz = 36 nK at s = 18 s = 18 s = 18 , above that scale, so at this depth there is no window in which the zero-temperature results with t ⊥ = 0 t_\perp= 0 t ⊥ = 0 apply. At s = 60 s = 60 s = 60 , k B T c = h × 27 k_B T_c = h \times27 k B T c = h × 27 Hz = 1.3 nK, and the window is 1.3 nK ≲ T ≲ \lesssim T \lesssim ≲ T ≲ 30 nK. Since T c T_c T c falls as J \sqrt{J} J , it drops below 30 nK only for s ≳ 20 s \gtrsim20 s ≳ 20 ; at s ≃ 40 s \simeq40 s ≃ 40 , k B T c ≃ 5 k_B T_c \simeq5 k B T c ≃ 5 nK, and the range is 5 nK ≲ T ≲ \lesssim T \lesssim ≲ T ≲ 30 nK.
Derivation of the momentum correlator
We begin from Eq. (25) and use the time evolution in Eq. (26) . At zero temperature the contractions needed for the four-operator expectation value include
⟨ β m , − k β n , − k † β m ′ , − k ′ β n ′ , − k ′ † ⟩ = δ m n δ m ′ n ′ , (74) \langle\beta_{m,-k}\beta^\dagger_{n,-k}\beta_{m',-k'}\beta^\dagger_{n',-k'} \rangle= \delta_{mn}\delta_{m'n'}, \tag*{(74)} ⟨ β m , − k β n , − k † β m ′ , − k ′ β n ′ , − k ′ † ⟩ = δ mn δ m ′ n ′ , (74)
⟨ β m , − k β n , k β m ′ , k ′ † β n ′ , − k ′ † ⟩ = δ m n ′ δ n m ′ δ k k ′ . (75) \langle\beta_{m,-k}\beta_{n,k}\beta^\dagger_{m',k'}\beta^\dagger_{n',-k'} \rangle= \delta_{mn'}\delta_{nm'}\delta_{kk'}. \tag*{(75)} ⟨ β m , − k β n , k β m ′ , k ′ † β n ′ , − k ′ † ⟩ = δ m n ′ δ n m ′ δ k k ′ . (75)
The first contraction produces a time-independent factorized term and contributes to the disconnected part of the momentum correlator, removed by the connected definition in Eq. (24) . The second gives the dynamical contribution
Π i j ( t ) = ∫ d k k 2 ∑ m , n M i , k m n M j , k m n e i E m n , k t ( u m 2 v n 2 + v m 2 u n 2 − 2 u m v m u n v n ) , (76) \Pi_{ij}(t) = \int dk\, k^2 \sum_{m,n} \mathcal{M}^{mn}_{i,k}\mathcal{M}^{mn}_{j,k}e^{iE_{mn,k}t}\left(u_m^2v_n^2+v_m^2u_n^2-2u_mv_mu_nv_n\right), \tag*{(76)} Π ij ( t ) = ∫ d k k 2 m , n ∑ M i , k mn M j , k mn e i E mn , k t ( u m 2 v n 2 + v m 2 u n 2 − 2 u m v m u n v n ) , (76)
where all coherence factors are at the same k k k , which we suppress, and the terms arise from pair creation and annihilation in the two transverse modes. Using u m = cosh θ m u_m = \cosh\theta_m u m = cosh θ m and v m = sinh θ m v_m = \sinh\theta_m v m = sinh θ m ,
u m 2 v n 2 + v m 2 u n 2 − 2 u m v m u n v n = ( sinh θ m cosh θ n − cosh θ m sinh θ n ) 2 = sinh 2 ( θ m − θ n ) , (77) u_m^2v_n^2+v_m^2u_n^2-2u_mv_mu_nv_n
= (\sinh\theta_m\cosh\theta_n-\cosh\theta_m\sinh\theta_n)^2
= \sinh^2(\theta_m-\theta_n), \tag*{(77)} u m 2 v n 2 + v m 2 u n 2 − 2 u m v m u n v n = ( sinh θ m cosh θ n − cosh θ m sinh θ n ) 2 = sinh 2 ( θ m − θ n ) , (77)
which reduces Eq. (76) to Eq. (27) .
Properties of M M M
The definition in Eq. (21) gives
M i , k m n = M i , k n m , (78) \mathcal{M}^{mn}_{i,k} = \mathcal{M}^{nm}_{i,k}, \tag*{(78)} M i , k mn = M i , k nm , (78)
so the matrix is symmetric in the two transverse-mode indices. Orthogonality of U g U_g U g further gives
∑ m M i , k m m = 1 , ∑ i M i , k m m = 1. (79) \sum_m \mathcal{M}^{mm}_{i,k} = 1,\qquad\sum_i \mathcal{M}^{mm}_{i,k} = 1. \tag*{(79)} m ∑ M i , k mm = 1 , i ∑ M i , k mm = 1. (79)
The first identity resolves the weight of a physical tube over transverse modes; the second normalizes the tube weight of each mode and is what makes Eq. (57) an inverse participation ratio.
We recast Eq. (42) before taking asymptotics. With Σ q = ( ϵ k + g q ) / 2 E q , k \Sigma_q = (\epsilon_k+g_q)/2E_{q,k} Σ q = ( ϵ k + g q ) /2 E q , k , Eqs. (39) and (40) give C r − S r = δ r , 0 C_r-S_r=\delta_{r,0} C r − S r = δ r , 0 , so the bracket of Eq. (42) becomes a difference of squares, Σ r 2 − H r 2 = ( Σ r − H r ) ( Σ r + H r ) \Sigma_r^2-H_r^2=(\Sigma_r-H_r)(\Sigma_r+H_r) Σ r 2 − H r 2 = ( Σ r − H r ) ( Σ r + H r ) , whose factors are the transforms of
P q = Σ q − h q = ϵ k 2 E q , k , M q = Σ q + h q = E q , k 2 ϵ k , (80) P_q = \Sigma_q-h_q = \frac{\epsilon_k}{2E_{q,k}},\qquad M_q = \Sigma_q+h_q = \frac{E_{q,k}}{2\epsilon_k}, \tag*{(80)} P q = Σ q − h q = 2 E q , k ϵ k , M q = Σ q + h q = 2 ϵ k E q , k , (80)
using E q , k 2 = ϵ k ( ϵ k + 2 g q ) E_{q,k}^2=\epsilon_k(\epsilon_k+2g_q) E q , k 2 = ϵ k ( ϵ k + 2 g q ) in the second. Treating r = 0 r=0 r = 0 separately,
Π r ( k , 0 ) = 2 k 2 [ P r M r − 1 4 δ r , 0 ] , (81) \Pi_r(k,0)=2k^2\left[P_rM_r-\frac{1}{4}\delta_{r,0}\right], \tag*{(81)} Π r ( k , 0 ) = 2 k 2 [ P r M r − 4 1 δ r , 0 ] , (81)
exactly, for either kernel. Since P q M q = 1 / 4 P_qM_q=1/4 P q M q = 1/4 pointwise, Parseval gives the sum rule of Eq. (44) . The two factors carry reciprocal powers of E q , k E_{q,k} E q , k , and the asymptotics of Eqs. (45) and (46) reduce to their behavior near the singularity nearest the real axis.
Contour integral
Both P r P_r P r and M r M_r M r are Fourier coefficients of the form of Eq. (41) at t = 0 t=0 t = 0 ,
X r = ∫ − π π d q 2 π e − i r q x q . (82) X_r=\int_{-\pi}^{\pi}\frac{dq}{2\pi}e^{-irq}x_q. \tag*{(82)} X r = ∫ − π π 2 π d q e − i r q x q . (82)
with x q x_q x q standing for P q P_q P q or M q M_q M q . Both are 2 π 2\pi 2 π -periodic, g q g_q g q being a cosine series in the integer tube separation, and both are analytic on the real axis, since stability gives g q > 0 g_q > 0 g q > 0 and hence E q , k > 0 E_{q,k} > 0 E q , k > 0 there for k ≠ 0 k \ne0 k = 0 . Writing q 0 = ς 0 − i κ q_0 = \varsigma_0 - i\kappa q 0 = ς 0 − iκ for the branch point lying nearest the real axis, with κ > 0 \kappa> 0 κ > 0 its depth below it, x q x_q x q is then analytic in the strip ∣ Im q ∣ < κ |\operatorname{Im} q| < \kappa ∣ Im q ∣ < κ . The value of κ \kappa κ is kernel dependent and is determined below.
Take the rectangle with corners ± π \pm\pi ± π and ± π − i λ \pm\pi-i\lambda ± π − iλ , for any λ < κ \lambda< \kappa λ < κ . It encloses no branch point, and its vertical sides cancel by periodicity, so Cauchy’s theorem equates the top and bottom edges,
X r = ∫ − π − i λ π − i λ d q 2 π e − i q r x q . (83) X_r = \int_{-\pi-i\lambda}^{\pi-i\lambda} \frac{dq}{2\pi} e^{-iqr}x_q. \tag*{(83)} X r = ∫ − π − iλ π − iλ 2 π d q e − i q r x q . (83)
Substituting q = q ′ − i λ q=q'-i\lambda q = q ′ − iλ with q ′ q' q ′ real gives e − i q r = e − i q ′ r e − λ r e^{-iqr}=e^{-iq'r}e^{-\lambda r} e − i q r = e − i q ′ r e − λ r , and the constant factor comes out of the integral:
X r = e − λ r ∫ − π π d q ′ 2 π e − i q ′ r x q ′ − i λ . (84) X_r=e^{-\lambda r}\int_{-\pi}^{\pi}\frac{dq'}{2\pi}e^{-iq'r}x_{q'-i\lambda}. \tag*{(84)} X r = e − λ r ∫ − π π 2 π d q ′ e − i q ′ r x q ′ − iλ . (84)
The remaining integrand is bounded, so ∣ X r ∣ ≲ e − λ r |X_r|\lesssim e^{-\lambda r} ∣ X r ∣ ≲ e − λ r for every λ < κ \lambda<\kappa λ < κ : the decay rate is the distance to the nearest branch point. At the instability of Sec. V the strip closes, κ → 0 \kappa\to0 κ → 0 , and the exponential decay is lost. Since x q x_q x q is real on the real axis, its singularities are mirror images in that axis and occur in conjugate pairs. For r > 0 r>0 r > 0 the factor e − λ r e^{-\lambda r} e − λ r of (84) decays only for λ > 0 \lambda>0 λ > 0 , so the contour can be moved downward alone and the relevant branch point is the one below the real axis, with Im q 0 = − κ \operatorname{Im}q_0=-\kappa Im q 0 = − κ . The case r < 0 r<0 r < 0 follows from the evenness of Π r \Pi_r Π r .
The power of r r r follows from the type of the branch point, which for both kernels below is of square-root type. Let x q ≃ A ( q − q 0 ) α x_q\simeq A(q-q_0)^\alpha x q ≃ A ( q − q 0 ) α near q 0 q_0 q 0 , with α > − 1 \alpha>-1 α > − 1 , and deform the real-axis contour of (82) into a keyhole about the cut descending from q 0 q_0 q 0 . The arc at its tip contributes O ( ϵ α + 1 ) O(\epsilon^{\alpha+1}) O ( ϵ α + 1 ) and drops out, while the two sides of the cut do not cancel: parametrizing q = q 0 − i s q=q_0-is q = q 0 − i s with s > 0 s>0 s > 0 , their difference is proportional to sin ( π α ) s α \sin(\pi\alpha)s^\alpha sin ( π α ) s α , and e − i q r = e − i ς 0 r e − κ r e − s r e^{-iqr}=e^{-i\varsigma_0r}e^{-\kappa r}e^{-sr} e − i q r = e − i ς 0 r e − κ r e − sr . The remaining r r r dependence is therefore
∫ 0 ∞ d s s α e − s r = Γ ( α + 1 ) r − α − 1 , (85) \int_0^\infty ds\,s^\alpha e^{-sr}=\Gamma(\alpha+1)r^{-\alpha-1}, \tag*{(85)} ∫ 0 ∞ d s s α e − sr = Γ ( α + 1 ) r − α − 1 , (85)
the factor e − s r e^{-sr} e − sr confining the integral to s ≲ 1 / r s\lesssim1/r s ≲ 1/ r , so that a region of width 1 / r 1/r 1/ r with integrand of size r − α r^{-\alpha} r − α produces the power. Collecting the factors,
∣ X r ∣ ∼ r − α − 1 e − κ r (86) |X_r|\sim r^{-\alpha-1}e^{-\kappa r} \tag*{(86)} ∣ X r ∣ ∼ r − α − 1 e − κ r (86)
up to a constant, with corrections down by powers of 1 / r 1/r 1/ r ; the accompanying phase e − i ς 0 r e^{-i\varsigma_0r} e − i ς 0 r plays no role below. Applied to (81) , the exponential rates of P r P_r P r and M r M_r M r add and their algebraic exponents add, so that Π r ∝ r − ( α P + α M ) − 2 e − 2 κ r \Pi_r\propto r^{-(\alpha_P+\alpha_M)-2}e^{-2\kappa r} Π r ∝ r − ( α P + α M ) − 2 e − 2 κ r .
Short range interactions
Here g q = n 0 ( U c + 2 V cos q ) g_q=n_0(U_c+2V\cos q) g q = n 0 ( U c + 2 V cos q ) is entire, so the only branch points are those of E q , k E_{q,k} E q , k , at the zeros of ϵ k + 2 g q \epsilon_k+2g_q ϵ k + 2 g q . These satisfy
cos q = − z k , z k = k 2 / 2 + 2 n 0 U c 4 n 0 V , (87) \cos q=-z_k,\qquad z_k=\frac{k^2/2+2n_0U_c}{4n_0V}, \tag*{(87)} cos q = − z k , z k = 4 n 0 V k 2 /2 + 2 n 0 U c , (87)
and stability requires g π > 0 g_\pi>0 g π > 0 , that is U c > 2 V U_c>2V U c > 2 V , so z k ≥ U c / 2 V > 1 z_k\ge U_c/2V>1 z k ≥ U c /2 V > 1 and no real solution exists. Setting q = π + i y q=\pi+iy q = π + i y and using cos ( π + i y ) = − cosh y \cos(\pi+iy)=-\cosh y cos ( π + i y ) = − cosh y gives cosh y = z k \cosh y=z_k cosh y = z k , so that
q 0 = π ± i κ E , κ E = arccosh z k , (88) q_0=\pi\pm i\kappa_E,\qquad\kappa_E=\operatorname{arccosh}z_k, \tag*{(88)} q 0 = π ± i κ E , κ E = arccosh z k , (88)
of which the lower controls r > 0 r>0 r > 0 , with ς 0 = π \varsigma_0=\pi ς 0 = π . The zero is simple, since 2 g q 0 ′ = − 4 n 0 V sin q 0 = − 4 i n 0 V sinh κ E ≠ 0 2g'_{q_0}=-4n_0V\sin q_0=-4in_0V\sinh\kappa_E\ne0 2 g q 0 ′ = − 4 n 0 V sin q 0 = − 4 i n 0 V sinh κ E = 0 , so E q , k ∝ ( q − q 0 ) 1 / 2 E_{q,k}\propto(q-q_0)^{1/2} E q , k ∝ ( q − q 0 ) 1/2 and (80) gives
α P = − 1 2 , α M = + 1 2 . (89) \alpha_P=-\frac{1}{2},\qquad\alpha_M=+\frac{1}{2}. \tag*{(89)} α P = − 2 1 , α M = + 2 1 . (89)
Hence P r ∼ r − 1 / 2 e − κ E r P_r\sim r^{-1/2}e^{-\kappa_Er} P r ∼ r − 1/2 e − κ E r and M r ∼ r − 3 / 2 e − κ E r M_r\sim r^{-3/2}e^{-\kappa_Er} M r ∼ r − 3/2 e − κ E r , and
Π r ( k , 0 ) ∼ r − 2 e − 2 κ E r (90) \Pi_r(k,0)\sim r^{-2}e^{-2\kappa_Er} \tag*{(90)} Π r ( k , 0 ) ∼ r − 2 e − 2 κ E r (90)
as stated in (45) . The same rate follows from a special-function representation: s q s_q s q , c q c_q c q , and h q h_q h q are all of the form ( A + B cos q ) ± 1 / 2 (A+B\cos q)^{\pm1/2} ( A + B cos q ) ± 1/2 up to constants, whose Fourier coefficients are Legendre functions of the second kind,
S r , C r , H r ∝ Q r − 1 / 2 ( z k ) , (91) S_r,\ C_r,\ H_r\propto Q_{r-1/2}(z_k), \tag*{(91)} S r , C r , H r ∝ Q r − 1/2 ( z k ) , (91)
and writing z k = cosh η z_k=\cosh\eta z k = cosh η the large-order behavior
Q ν ( cosh η ) ≃ π 2 ν sinh η e − ( ν + 1 / 2 ) η (92) Q_\nu(\cosh\eta)\simeq\sqrt{\frac{\pi}{2\nu\sinh\eta}}e^{-(\nu+1/2)\eta} \tag*{(92)} Q ν ( cosh η ) ≃ 2 ν sinh η π e − ( ν + 1/2 ) η (92)
reproduces r − 1 / 2 e − κ E r r^{-1/2}e^{-\kappa_Er} r − 1/2 e − κ E r per coefficient.
Dipolar interactions
No such representation is available, and the branch point now lies in g q g_q g q itself. Using K 1 ( z ) ≃ π / 2 z e − z K_1(z)\simeq\sqrt{\pi/2z}e^{-z} K 1 ( z ) ≃ π /2 z e − z in (33) , the interaction matrix element at large separation is
V r ( k ) ≃ 2 V d π ∣ k ∣ 2 r − 3 / 2 e − ∣ k ∣ r . (93) V_r(k)\simeq2V_d\sqrt{\frac{\pi|k|}{2}}\,r^{-3/2}e^{-|k|r}. \tag*{(93)} V r ( k ) ≃ 2 V d 2 π ∣ k ∣ r − 3/2 e − ∣ k ∣ r . (93)
In ∑ r V r ( k ) cos ( q r ) \sum_r V_r(k)\cos(qr) ∑ r V r ( k ) cos ( q r ) , written as 1 2 ∑ r V r ( e i q r + e − i q r ) \frac{1}{2}\sum_rV_r(e^{iqr}+e^{-iqr}) 2 1 ∑ r V r ( e i q r + e − i q r ) , it is the first half that can diverge in the lower half plane. With u ~ = e i q e − ∣ k ∣ \tilde u=e^{iq}e^{-|k|} u ~ = e i q e − ∣ k ∣ the divergent part is
∑ r ≥ 1 r − 3 / 2 u ~ r = Li 3 / 2 ( u ~ ) , (94) \sum_{r\ge1}r^{-3/2}\tilde u^r=\operatorname{Li}_{3/2}(\tilde u), \tag*{(94)} r ≥ 1 ∑ r − 3/2 u ~ r = Li 3/2 ( u ~ ) , (94)
which converges for ∣ u ~ ∣ < 1 |\tilde u|<1 ∣ u ~ ∣ < 1 , that is for Im q > − ∣ k ∣ \operatorname{Im}q>-|k| Im q > − ∣ k ∣ , and is singular only at u ~ = 1 \tilde u=1 u ~ = 1 . That point requires Re q = 0 \operatorname{Re}q=0 Re q = 0 in addition to Im q = − ∣ k ∣ \operatorname{Im}q=-|k| Im q = − ∣ k ∣ , so the branch points are
q 0 = ± i ∣ k ∣ , ς 0 = 0. (95) q_0=\pm i|k|,\qquad\varsigma_0=0. \tag*{(95)} q 0 = ± i ∣ k ∣ , ς 0 = 0. (95)
above and below the zone center rather than the zone corner. The expansion of the polylogarithm about u ~ = 1 \tilde{u}=1 u ~ = 1 is
Li 3 / 2 ( u ~ ) = ζ ( 3 2 ) − 2 π ( − ln u ~ ) 1 / 2 + O ( ln u ~ ) , (96) \operatorname{Li}_{3/2}(\tilde{u})=\zeta\left(\frac{3}{2}\right)-2\sqrt{\pi}(-\ln\tilde{u})^{1/2}+O(\ln\tilde{u}), \tag*{(96)} Li 3/2 ( u ~ ) = ζ ( 2 3 ) − 2 π ( − ln u ~ ) 1/2 + O ( ln u ~ ) , (96)
and − ln u ~ ≃ − i ( q − q 0 ) -\ln\tilde{u}\simeq-i(q-q_0) − ln u ~ ≃ − i ( q − q 0 ) near the branch point, so that
g q = g q 0 + A ( q − q 0 ) 1 / 2 + … (97) g_q=g_{q_0}+A(q-q_0)^{1/2}+\ldots \tag*{(97)} g q = g q 0 + A ( q − q 0 ) 1/2 + … (97)
with g q 0 g_{q_0} g q 0 finite because ζ ( 3 / 2 ) \zeta(3/2) ζ ( 3/2 ) converges. This finiteness is what separates the two kernels. In the stable regime ϵ k + 2 g q 0 ≠ 0 \epsilon_k+2g_{q_0}\ne0 ϵ k + 2 g q 0 = 0 , so E q 0 , k ≡ E 0 ≠ 0 E_{q_0,k}\equiv E_0\ne0 E q 0 , k ≡ E 0 = 0 and
E q , k = E 0 [ 1 + 2 A ( q − q 0 ) 1 / 2 ϵ k + 2 g q 0 ] 1 / 2 ≃ E 0 + E 0 A ϵ k + 2 g q 0 ( q − q 0 ) 1 / 2 , (98) \begin{aligned}
E_{q,k}=E_0\left[1+\frac{2A(q-q_0)^{1/2}}{\epsilon_k+2g_{q_0}}\right]^{1/2} \\
&\simeq E_0+\frac{E_0A}{\epsilon_k+2g_{q_0}}(q-q_0)^{1/2}, \tag*{(98)}
\end{aligned} E q , k = E 0 [ 1 + ϵ k + 2 g q 0 2 A ( q − q 0 ) 1/2 ] 1/2 ≃ E 0 + ϵ k + 2 g q 0 E 0 A ( q − q 0 ) 1/2 , (98)
so E q , k E_{q,k} E q , k and 1 / E q , k 1/E_{q,k} 1/ E q , k acquire singular parts of the same type and
α P = α M = + 1 2 . (99) \alpha_P=\alpha_M=+\frac{1}{2}. \tag*{(99)} α P = α M = + 2 1 . (99)
Both transforms therefore behave as r − 3 / 2 e − ∣ k ∣ r r^{-3/2}e^{-|k|r} r − 3/2 e − ∣ k ∣ r , and
Π r ( k , 0 ) ∝ r − 2 e − 2 ∣ k ∣ r (100) \Pi_r(k,0)\propto r^{-2}e^{-2|k|r} \tag*{(100)} Π r ( k , 0 ) ∝ r − 2 e − 2∣ k ∣ r (100)
as stated in Eq. (46) . The contrast with Eq. (90) is entirely in the exponents: P q P_q P q and M q M_q M q carry opposite powers of E q , k E_{q,k} E q , k , so their singular exponents cancel when E q , k E_{q,k} E q , k vanishes at q 0 q_0 q 0 and add when it does not.
Two limits are worth recording. The branch points of E q , k E_{q,k} E q , k exist for the dipolar kernel as well, at the complex q q q solving ϵ k + 2 g q = 0 \epsilon_k+2g_q=0 ϵ k + 2 g q = 0 ; for the couplings used here they lie further from the real axis than Eq. (95) , and they would take over on approach to the instability of Sec. V, where g q g_q g q approaches zero and κ E \kappa_E κ E of Eq. (88) descends toward the axis. And as k → 0 k\to0 k → 0 the branch points of Eq. (95) themselves reach the real axis at q = 0 q=0 q = 0 , where g q g_q g q acquires the ∣ q ∣ |q| ∣ q ∣ non-analyticity of Eq. (106) ; the exponential rate vanishes with k k k , which is why no exponential survives the longitudinal integration and Eq. (49) is a pure power law.
Long-time asymptotics
The t − 3 t^{-3} t − 3 law follows from the k → 0 k\to0 k → 0 endpoint of Eq. (50) for either kernel, and we give that argument first. As k → 0 k\to0 k → 0 we have ϵ k → 0 \epsilon_k\to0 ϵ k → 0 , g q ( k ) → g q ( 0 ) g_q(k)\to g_q(0) g q ( k ) → g q ( 0 ) , and E q , k → c q k E_{q,k}\to c_qk E q , k → c q k with c q = g q ( 0 ) c_q=\sqrt{g_q(0)} c q = g q ( 0 ) , so the phase is linear in k k k and has no stationary point while the amplitude vanishes as k 2 k^2 k 2 . With the retarded prescription t → t − i 0 + t\to t-i0^+ t → t − i 0 + ,
∫ 0 ∞ d k k 2 e 2 i c q k t = − 2 i ( 2 c q t ) 3 , (101) \int_0^\infty dk\,k^2 e^{2ic_qkt}=\frac{-2i}{(2c_qt)^3}, \tag*{(101)} ∫ 0 ∞ d k k 2 e 2 i c q k t = ( 2 c q t ) 3 − 2 i , (101)
and inserting this into Eq. (50) gives
Im A ( t ) ∼ t → ∞ − 1 64 t 3 ∫ 0 2 π d q ( g q ′ ) 2 g q 7 / 2 , (102) \operatorname{Im} A(t)\underset{t\to\infty}{\sim}-\frac{1}{64t^3}\int_0^{2\pi}dq\,\frac{(g'_q)^2}{g_q^{7/2}}, \tag*{(102)} Im A ( t ) t → ∞ ∼ − 64 t 3 1 ∫ 0 2 π d q g q 7/2 ( g q ′ ) 2 , (102)
which is Eq. (51) . The derivation assumes nothing about the kernel beyond g q ( 0 ) > 0 g_q(0)>0 g q ( 0 ) > 0 and the convergence of the q q q integral, and both hold in either case. For nearest-neighbor coupling g q = n 0 ( U c + 2 V cos q ) g_q=n_0(U_c+2V\cos q) g q = n 0 ( U c + 2 V cos q ) is smooth and bounded away from zero throughout the stable regime. For the dipolar kernel g q ( 0 ) g_q(0) g q ( 0 ) is the piecewise-quadratic form of Eq. (106) , likewise positive there, while g q ′ g'_q g q ′ acquires a jump at q = 0 q=0 q = 0 from the ∣ q ∣ |q| ∣ q ∣ term; it remains bounded, so the kink changes the coefficient and not the power. The integral diverges as g π → 0 g_\pi\to0 g π → 0 , which is the threshold singularity discussed in Sec. III A 3.
For nearest-neighbor coupling the reduction of Eq. (52) makes the same limit explicit. The phase x 1 + x 2 x\sqrt{1+x^2} x 1 + x 2 of Eq. (52) likewise has no stationary point, and the same endpoint expansion gives J ( s , ∞ ) = − 2 i / s 3 + O ( s − 5 ) J(s,\infty)=-2i/s^3+O(s^{-5}) J ( s , ∞ ) = − 2 i / s 3 + O ( s − 5 ) , which returns Eq. (102) on substitution. This is the argument above in the rescaled variable rather than a second derivative, and the independent checks are numerical. Direct evaluation of Eq. (52) approaches the coefficient 64 from below as the subleading t − 5 t^{-5} t − 5 correction dies off, reaching 63.99 at t = 80 t=80 t = 80 , and quadrature of Eq. (50) for the dipolar kernel reproduces the same power. The local exponent in Figure 4 is shown only to t = 50 t=50 t = 50 : the coefficient is evaluated pointwise as A ( t ) t 3 A(t)t^3 A ( t ) t 3 and stays accurate well beyond that, whereas the numerical derivative of ln A \ln A ln A amplifies roundoff once ∣ Im A ∣ |\operatorname{Im} A| ∣ Im A ∣ falls to 10 − 8 10^{-8} 1 0 − 8 .
Small-Q Q Q interaction integrals
For nearest-neighbor interactions we define
α = n 0 U c , β = 2 n 0 V , k e l l = 2 β α + β , (103) \alpha=n_0U_c,\qquad\beta=2n_0V,\qquad k_{\mathrm{ell}}=\sqrt{\frac{2\beta}{\alpha+\beta}}, \tag*{(103)} α = n 0 U c , β = 2 n 0 V , k ell = α + β 2 β , (103)
in terms of which the transverse integral at t = 0 t=0 t = 0 is
∫ 0 2 π d q ( g q ′ ) 2 g q = 8 3 α + β [ α E ( k e l l ) − ( α − β ) K ( k e l l ) ] , (104) \int_0^{2\pi}dq\,\frac{(g'_q)^2}{\sqrt{g_q}}=\frac{8}{3}\sqrt{\alpha+\beta}\left[\alpha\mathop{\mathrm{E}}(k_{\mathrm{ell}})-(\alpha-\beta)\mathop{\mathrm{K}}(k_{\mathrm{ell}})\right], \tag*{(104)} ∫ 0 2 π d q g q ( g q ′ ) 2 = 3 8 α + β [ α E ( k ell ) − ( α − β ) K ( k ell ) ] , (104)
where K \mathrm{K} K and E \mathrm{E} E are complete elliptic integrals with parameter m = k e l l 2 m=k_{\mathrm{ell}}^2 m = k ell 2 . The logarithmic divergence of K as $V\to U_c/2 is cancelled by the factor α − β \alpha-\beta α − β , so the transverse integral stays finite at threshold and the singular behavior of η \eta η there comes from the longitudinal integral instead.
For dipolar interactions g q ( k ) g_q(k) g q ( k ) depends on both momenta. In the k → 0 k\to0 k → 0 limit,
∑ r = 1 ∞ cos ( q r ) r 2 = π 2 6 − π ∣ q ∣ 2 + q 2 4 , q ∈ [ − π , π ] , (105) \sum_{r=1}^{\infty}\frac{\cos(qr)}{r^2}=\frac{\pi^2}{6}-\frac{\pi|q|}{2}+\frac{q^2}{4},\qquad q\in[-\pi,\pi], \tag*{(105)} r = 1 ∑ ∞ r 2 cos ( q r ) = 6 π 2 − 2 π ∣ q ∣ + 4 q 2 , q ∈ [ − π , π ] , (105)
so that
g q d i p ( 0 ) = n 0 [ U c + 2 V ( π 2 6 − π ∣ q ∣ 2 + q 2 4 ) ] , (106) g_q^{\mathrm{dip}}(0)=n_0\left[U_c+2V\left(\frac{\pi^2}{6}-\frac{\pi|q|}{2}+\frac{q^2}{4}\right)\right], \tag*{(106)} g q dip ( 0 ) = n 0 [ U c + 2 V ( 6 π 2 − 2 π ∣ q ∣ + 4 q 2 ) ] , (106)
which is piecewise quadratic in q q q and, evaluated at q = π q=\pi q = π , gives Eq. (71) and the dipolar threshold of Eq. (72) . At finite k k k the dipolar g q ( k ) g_q(k) g q ( k ) has no closed form, and we evaluate the q q q integral in Eq. (50) numerically.
Momentum correlator in the tube basis
The time-evolved operators in the original tube basis are
a i , k ( t ) = ∑ m [ u i m , k β m , k e i E m , k t + v i m , k β m , − k † e − i E m , k t ] , (107) a_{i,k}(t)=\sum_m\left[u_{im,k}\beta_{m,k}e^{iE_{m,k}t}+v_{im,k}\beta^\dagger_{m,-k}e^{-iE_{m,k}t}\right], \tag*{(107)} a i , k ( t ) = m ∑ [ u im , k β m , k e i E m , k t + v im , k β m , − k † e − i E m , k t ] , (107)
with
u i m , k = ( U g ) i m u m , k , v i m , k = ( U g ) i m v m , k . (108) u_{im,k}=(U_g)_{im}u_{m,k},\qquad v_{im,k}=(U_g)_{im}v_{m,k}. \tag*{(108)} u im , k = ( U g ) im u m , k , v im , k = ( U g ) im v m , k . (108)
Using the contractions of Appendix B,
Π i j ( t ) = ∫ d k k 2 ∑ m , n e i ε m n , k t Q i j m n , (109) \Pi_{ij}(t)=\int dk\,k^2\sum_{m,n}e^{i\varepsilon_{mn,k}t}Q_{ij}^{mn}, \tag*{(109)} Π ij ( t ) = ∫ d k k 2 m , n ∑ e i ε mn , k t Q ij mn , (109)
where
Q i j m n = u i m u j m v i n v j n + v i m v j m u i n u j n − u i m v j m u j n v i n − u j m v i m u i n v j n . (110) Q_{ij}^{mn}=u_{im}u_{jm}v_{in}v_{jn}+v_{im}v_{jm}u_{in}u_{jn}-u_{im}v_{jm}u_{jn}v_{in}-u_{jm}v_{im}u_{in}v_{jn}. \tag*{(110)} Q ij mn = u im u j m v in v j n + v im v j m u in u j n − u im v j m u j n v in − u j m v im u in v j n . (110)
Substituting Eq. (108) reproduces the factorized overlap of Eq. (29) and therefore Eq. (27) .
Small-Q Q Q coefficient in the Bloch basis
Equation (65) defines η \eta η from the eigenbasis of G ^ \hat{G} G ^ , which is the form we evaluate with disorder. Here we check that it returns A Λ ( 0 ) A_\Lambda(0) A Λ ( 0 ) of Eq. (50) when the array is uniform. For n i = n 0 n_i=n_0 n i = n 0 the eigenvectors are Bloch waves, ( U g ) i m = e i q m i / N t (U_g)_{im}=e^{iq_m i}/\sqrt{N_t} ( U g ) im = e i q m i / N t , so the overlap of Eq. (64) is
F m n ( Q ) = 1 N t ∑ i e i Q i e i ( q n − q m ) i = δ Q , q m − q n , (111) F^{mn}(Q)=\frac{1}{N_t}\sum_i e^{iQ_i}e^{i(q_n-q_m)i}=\delta_{Q,q_m-q_n}, \tag*{(111)} F mn ( Q ) = N t 1 i ∑ e i Q i e i ( q n − q m ) i = δ Q , q m − q n , (111)
and the double mode sum collapses to a single one,
Π Q ( k , 0 ) = k 2 ∑ q sinh 2 ( θ q + Q , k − θ q , k ) , (112) \Pi_Q(k,0)=k^2\sum_q\sinh^2(\theta_{q+Q,k}-\theta_{q,k}), \tag*{(112)} Π Q ( k , 0 ) = k 2 q ∑ sinh 2 ( θ q + Q , k − θ q , k ) , (112)
which is Eq. (35) at t = 0 t=0 t = 0 . Expanding at small Q Q Q gives sinh 2 ( θ q + Q − θ q ) ≃ Q 2 ( θ q ′ ) 2 \sinh^2(\theta_{q+Q}-\theta_q)\simeq Q^2(\theta'_q)^2 sinh 2 ( θ q + Q − θ q ) ≃ Q 2 ( θ q ′ ) 2 . To obtain θ q ′ \theta'_q θ q ′ , differentiate cosh 2 θ q , k = ( ε k + g q ) / E q , k \cosh2\theta_{q,k}=(\varepsilon_k+g_q)/E_{q,k} cosh 2 θ q , k = ( ε k + g q ) / E q , k from Eq. (19) . Since E q , k 2 = ε k ( ε k + 2 g q ) E_{q,k}^2=\varepsilon_k(\varepsilon_k+2g_q) E q , k 2 = ε k ( ε k + 2 g q ) gives E q , k ′ = ε k g q ′ / E q , k E'_{q,k}=\varepsilon_k g'_q/E_{q,k} E q , k ′ = ε k g q ′ / E q , k ,
d d q cosh 2 θ q , k = g q ′ E q , k − ( ε k + g q ) ε k g q ′ E q , k 3 = ε k g q g q ′ E q , k 3 , (113) \frac{d}{dq}\cosh2\theta_{q,k}=\frac{g'_q}{E_{q,k}}-\frac{(\varepsilon_k+g_q)\varepsilon_k g'_q}{E_{q,k}^3}=\frac{\varepsilon_k g_q g'_q}{E_{q,k}^3}, \tag*{(113)} d q d cosh 2 θ q , k = E q , k g q ′ − E q , k 3 ( ε k + g q ) ε k g q ′ = E q , k 3 ε k g q g q ′ , (113)
while sinh 2 θ q , k = cosh 2 2 θ q , k − 1 = g q / E q , k \sinh2\theta_{q,k}=\sqrt{\cosh^2 2\theta_{q,k}-1}=g_q/E_{q,k} sinh 2 θ q , k = cosh 2 2 θ q , k − 1 = g q / E q , k . Equating this to 2 θ q ′ sinh 2 θ q , k 2\theta'_q\sinh2\theta_{q,k} 2 θ q ′ sinh 2 θ q , k gives
θ q ′ = ε k g q ′ 2 E q , k 2 = g q ′ 2 ( ε k + 2 g q ) . (114) \theta'_q=\frac{\varepsilon_k g'_q}{2E_{q,k}^2}=\frac{g'_q}{2(\varepsilon_k+2g_q)}. \tag*{(114)} θ q ′ = 2 E q , k 2 ε k g q ′ = 2 ( ε k + 2 g q ) g q ′ . (114)
Substituting into Eq. (112) and taking ∑ q → N t ∫ d q / 2 π \sum_q\to N_t\int dq/2\pi ∑ q → N t ∫ d q /2 π ,
η = N t 8 π ∫ 0 1 d k k 2 ∫ 0 2 π d q ( g q ′ ) 2 ( ε k + 2 g q ) 2 = N t 2 π Re A Λ ( 0 ) , (115) \eta=\frac{N_t}{8\pi}\int_0^1 dk\,k^2\int_0^{2\pi}dq\,\frac{(g'_q)^2}{(\varepsilon_k+2g_q)^2}=\frac{N_t}{2\pi}\operatorname{Re}A_\Lambda(0), \tag*{(115)} η = 8 π N t ∫ 0 1 d k k 2 ∫ 0 2 π d q ( ε k + 2 g q ) 2 ( g q ′ ) 2 = 2 π N t Re A Λ ( 0 ) , (115)
the integrand being that of Eq. (50) at t = 0 t=0 t = 0 . The factor N t / 2 π N_t/2\pi N t /2 π is the density of states converting the mode sum to the continuum q q q integral, and is the normalization absorbed after Eq. (35) ; it is independent of σ \sigma σ and of the kernel, so it cancels in the ratio η ( σ ) / η ( 0 ) \eta(\sigma)/\eta(0) η ( σ ) / η ( 0 ) that every result in Sec. IV C uses.
Numerical methods
Unless stated otherwise, the finite-array calculations use periodic boundary conditions, N t = 1000 N_t=1000 N t = 1000 tubes, U c = n 0 = 1 U_c=n_0=1 U c = n 0 = 1 , and 100 disorder realizations. All localization results are obtained at fixed longitudinal momentum, that is for the integrand of Eq. (27) rather than the k k k -integrated correlator. We use the minimum-image distance
d i j = min ( ∣ i − j ∣ , N − ∣ i − j ∣ ) , (116) d_{ij}=\min(|i-j|,N-|i-j|), \tag*{(116)} d ij = min ( ∣ i − j ∣ , N − ∣ i − j ∣ ) , (116)
which preserves translation invariance in the uniform finite ring. Each analytical reduction was checked against direct evaluation, and the implementation against its known limits: the uniform-array correlator, the sum rule, the congruence of Eq. (23) , the standing-wave IPR and the three-term recursion for nearest neighbors are reproduced, the threshold located by bisection is constant across σ \sigma σ as Sec. II C requires, and no realization was dynamically unstable.
Momentum correlator
In the uniform array the transforms of Eq. (41) are evaluated by FFT, so that Eq. (42) returns the whole profile Π r \Pi_r Π r at a cost of three transforms per k k k , against O ( N t 4 ) O(N_t^4) O ( N t 4 ) for the direct double mode sum of Eq. (27) over all tube pairs. This is how the uniform-array curves of Figure 3 are obtained. With disorder the same factorization is used in matrix rather than transform form. For each realization we construct G ^ ( k ) \hat{G}(k) G ^ ( k ) and diagonalize it once. Defining
s m = sinh 2 θ m , c m = cosh 2 θ m , h m = sinh θ m cosh θ m , (117) s_m=\sinh^2\theta_m,\qquad c_m=\cosh^2\theta_m,\qquad h_m=\sinh\theta_m\cosh\theta_m, \tag*{(117)} s m = sinh 2 θ m , c m = cosh 2 θ m , h m = sinh θ m cosh θ m , (117)
the fixed-k k k correlator follows from Eq. (38) as
Π i j ( k , t ) = 2 k 2 [ A i j ( t ) B i j ( t ) − C i j 2 ( t ) ] , (118) \Pi_{ij}(k,t)=2k^2\left[A_{ij}(t)B_{ij}(t)-C_{ij}^2(t)\right], \tag*{(118)} Π ij ( k , t ) = 2 k 2 [ A ij ( t ) B ij ( t ) − C ij 2 ( t ) ] , (118)
where
A i j ( t ) = ∑ m ( U g ) i m s m e i E m , k t ( U g ) j m , (119) A_{ij}(t)=\sum_m (U_g)_{im}s_m e^{iE_{m,k}t}(U_g)_{jm}, \tag*{(119)} A ij ( t ) = m ∑ ( U g ) im s m e i E m , k t ( U g ) j m , (119)
B i j ( t ) = ∑ m ( U g ) i m c m e i E m , k t ( U g ) j m , (120) B_{ij}(t)=\sum_m (U_g)_{im}c_m e^{iE_{m,k}t}(U_g)_{jm}, \tag*{(120)} B ij ( t ) = m ∑ ( U g ) im c m e i E m , k t ( U g ) j m , (120)
C i j ( t ) = ∑ m ( U g ) i m h m e i E m , k t ( U g ) j m . (121) C_{ij}(t)=\sum_m (U_g)_{im}h_m e^{iE_{m,k}t}(U_g)_{jm}. \tag*{(121)} C ij ( t ) = m ∑ ( U g ) im h m e i E m , k t ( U g ) j m . (121)
The factorized form replaces the explicit double mode sum by three matrix products. For the infinite-time squared correlator we evaluate Eq. (60) without a finite time grid: the identity sinh 2 ( θ m − θ n ) = s m c n + c m s n − 2 h m h n \sinh^2(\theta_m-\theta_n)=s_mc_n+c_ms_n-2h_mh_n sinh 2 ( θ m − θ n ) = s m c n + c m s n − 2 h m h n expands into six terms, each a product of a function of m m m with a function of n n n , so the double mode sum collapses into matrix products built from ∣ U g ∣ 2 \lvert U_g\rvert^2 ∣ U g ∣ 2 and the result is the T → ∞ T\to\infty T → ∞ limit exactly.
Averaging and fit windows
We average spatial profiles geometrically over disorder realizations and over tube pairs at fixed separation, since the arithmetic mean is dominated by rare large-amplitude samples and does not converge as the window is pushed outward. We fit over
3 ≤ d / ξ l o c ≤ 8 , (122) 3\le d/\xi_{\mathrm{loc}}\le8, \tag*{(122)} 3 ≤ d / ξ loc ≤ 8 , (122)
with ξ l o c = 1 / γ \xi_{\mathrm{loc}}=1/\gamma ξ loc = 1/ γ . A window fixed in absolute distance cannot serve, since ξ l o c \xi_{\mathrm{loc}} ξ loc varies by nearly two orders of magnitude across the disorder grid: at weak disorder a window of a few tens of sites lies within a single localization length and is not asymptotic, while at strong disorder the squared correlator, decaying at 4 γ 4\gamma 4 γ , has already reached the roundoff floor there. The window is clipped by the largest available separation and by the precision floor, which makes the weakest-disorder points the most sensitive to finite size. That bias is set by N t / ξ l o c N_t/\xi_{\mathrm{loc}} N t / ξ loc and not by the kernel. For nearest neighbors at V = 0.3 V=0.3 V = 0.3 , where ξ l o c \xi_{\mathrm{loc}} ξ loc is about three times longer than at the matched coupling of Sec. IV A, the ratio at σ = 0.25 \sigma=0.25 σ = 0.25 is 0.78 against 0.90, and at the two weakest disorders the window cannot reach 3 ξ l o c 3\xi_{\mathrm{loc}} 3 ξ loc at all; at σ = 0.50 \sigma=0.50 σ = 0.50 , where N t / ξ l o c N_t/\xi_{\mathrm{loc}} N t / ξ loc is large at both couplings, the two agree within a few percent.
Lyapunov exponent
At fixed k k k we represent the eigenvalue problem for G ^ ( k ) \hat{G}(k) G ^ ( k ) by a banded transfer matrix. With the off-diagonal coupling truncated at range b b b , the state vector is
x i = ( ψ i + b − 1 , … , ψ i − b ) T , (123) x_i=(\psi_{i+b-1},\ldots,\psi_{i-b})^{\mathrm{T}}, \tag*{(123)} x i = ( ψ i + b − 1 , … , ψ i − b ) T , (123)
so the transfer matrix has dimension 2 b 2b 2 b and Lyapunov exponents in positive and negative pairs. The localization rate is the smallest positive exponent, which controls the slowest asymptotic decay. We propagate an orthonormal frame with periodic QR reorthogonalization and accumulate the logarithms of the diagonal entries of R R R . Production calculations use b = 5 b=5 b = 5 , chains of 1.5 × 10 5 1.5\times10^5 1.5 × 1 0 5 sites, independent of the ring size N t N_t N t used elsewhere, and four independent chains, whose spread gives the error bars in Figure 8 . At sufficiently small k k k the dipolar coupling becomes effectively long ranged and a finite-band transfer matrix is no longer a controlled approximation.
The Lyapunov exponent depends on spectral position as well as on disorder strength, Figure 12 , and the dependence follows from the congruence of Eq. (23) . Writing G ^ ψ = g ψ \hat{G}\psi=g\psi G ^ ψ = g ψ as M ϕ = g D − 1 ϕ M\phi=gD^{-1}\phi M ϕ = g D − 1 ϕ with ϕ = D 1 / 2 ψ \phi=D^{1/2}\psi ϕ = D 1/2 ψ leaves the disorder-free band of M M M intact and turns the density fluctuations into a random potential − ( g / n 0 ) u i -(g/n_0)u_i − ( g / n 0 ) u i to first order, whose strength grows with g g g ; since D 1 / 2 D^{1/2} D 1/2 is bounded, ψ \psi ψ and ϕ \phi ϕ share their Lyapunov exponent. The Born approximation for this potential gives
Figure 12. Transfer-matrix Lyapunov exponent across the band at k = 1 k = 1 k = 1 , scaled by σ 2 \sigma^2 σ 2 , for the dipolar kernel at V = 0.3 V = 0.3 V = 0.3 (left) and for nearest neighbors at V = 0.181 V = 0.181 V = 0.181 , the matched first-neighbor bond of Sec. IV A (right). Dots mark the minimum at each σ \sigma σ , which is the rate the correlator measures, and the dotted line marks the band center g = n 0 U c g = n_0U_c g = n 0 U c , where the rate of Fig. 7 is taken. The dashed line is the random-mass Born result of Eq. (19), with no free parameter, and the weakest disorder follows it across the band. The Born rate is the ratio of the random potential, which grows with g g g , to the squared group velocity. For nearest neighbors the velocity is symmetric about the band center, so the growing potential shifts the minimum below it. For the dipolar kernel at k = 1 k = 1 k = 1 the band is asymmetric, and the asymmetry offsets this shift, placing the minimum at the band center throughout the Born regime; it leaves the center only once that regime ends. Each curve is one chain of 1.2 × 10 5 1.2 \times10^5 1.2 × 1 0 5 sites, with the same disorder sequence at every g g g .
γ ( g q ) ≃ σ 2 g q 2 2 ( ∂ q g q ) 2 , (124) \gamma(g_q)\simeq\frac{\sigma^2g_q^2}{2\left(\partial_qg_q\right)^2}, \tag*{(124)} γ ( g q ) ≃ 2 ( ∂ q g q ) 2 σ 2 g q 2 , (124)
which reduces to Eq. (58) at g = n 0 U c g=n_0U_c g = n 0 U c and follows the transfer matrix across the band at weak disorder. The potential grows with g g g while the group velocity vanishes at both band edges, and the minimum of Eq. (124) sits where the two balance: at g ∗ = n 0 ( U c − 4 V 2 / U c ) g^*=n_0(U_c-4V^2/U_c) g ∗ = n 0 ( U c − 4 V 2 / U c ) for nearest neighbors, below the band center, and at g ≃ n 0 U c g\simeq n_0U_c g ≃ n 0 U c for the dipolar band at k = 1 k=1 k = 1 .
Two rates therefore enter the paper, and each comparison uses the one it tests. Figure 7 tests Eq. (58) , a statement at fixed energy, and uses the rate at the band center at g = n 0 U c g=n_0U_c g = n 0 U c for both kernels; for nearest neighbors this is q = π / 2 q=\pi/2 q = π /2 , where the Kappus–Wegner anomaly [40 ] lowers the rate below Eq. (58) by 9% as σ → 0 \sigma\to0 σ → 0 . At finite σ \sigma σ it competes with the nonlinearity of the random potential, whose exact form g ( 1 / n i − 1 / n 0 ) g(1/n_i-1/n_0) g ( 1/ n i − 1/ n 0 ) has a variance exceeding g 2 σ 2 / n 0 2 g^2\sigma^2/n_0^2 g 2 σ 2 / n 0 2 by a term of relative order σ 2 \sigma^2 σ 2 , so the band-center rate departs from Eq. (58) by a few percent in either direction until the Born regime itself ends. Figure 3 (c) and Figure 8 compare with the long-time correlator, whose tail is a sum over all modes and therefore decays at the smallest rate in the band. There we scan the band at each σ \sigma σ with chains of 2 × 10 4 2\times10^4 2 × 1 0 4 sites, locate the minimum positive exponent, and evaluate it with the production chains. The two rates agree for the dipolar kernel at weak disorder and differ for nearest neighbors, where the band-center rate is the larger; this is why the nearest-neighbor minimum lies below the band-center Born line of Eq. (58) .
Eigenvector envelopes
Figure 13 compares two averages of the eigenvector amplitude at distance d d d from the mode maximum, the typical envelope ⟨ ln ∣ ψ ∣ ⟩ \langle\ln|\psi|\rangle ⟨ ln ∣ ψ ∣ ⟩ and the annealed envelope ln ⟨ ∣ ψ ∣ ⟩ \ln\langle|\psi|\rangle ln ⟨ ∣ ψ ∣ ⟩ . The amplitude at fixed d d d is log-normally distributed, so the annealed average is dominated by rare large amplitudes and lies above the typical one by an amount that grows with d d d ; only the typical envelope decays at γ \gamma γ . The IPR, a fourth moment of ∣ ψ ∣ |\psi| ∣ ψ ∣ , weights the same upper tail, which is one source of the drift of 2 ℓ t y p / γ 2\mathit{\ell}_{\mathrm{typ}}/\gamma 2 ℓ typ / γ in Fig. 8.
Figure 13. Typical (blue) and annealed (red) eigenvector envelopes against the distance from the mode maximum, σ = 0.18 \sigma= 0.18 σ = 0.18 , k = 1 k = 1 k = 1 , N t = 1000 N_t = 1000 N t = 1000 , 100 realizations, middle third of the band; black, half the annealed envelope. Dipolar at V = 0.3 V = 0.3 V = 0.3 (left) and n.n. at the matched bond 0.181 (right). Dotted: numerical (roundoff) floor.
Evaluation of A ( t ) A(t) A ( t ) and η \eta η
The function A ( t ) A(t) A ( t ) enters only Figure 4 and is a uniform-array quantity. For nearest neighbors we use the closed form of (52) , which integrates to Λ → ∞ \Lambda\to\infty Λ → ∞ ; for the dipolar kernel we evaluate (50) by Gauss–Legendre quadrature in q q q and k k k . The t − 3 t^{-3} t − 3 law places a specific demand on the cutoff. It is an asymptotic statement about the k → 0 k \to0 k → 0 endpoint, and a finite cutoff leaves an endpoint term oscillating as e 2 i E ( Λ ) t e^{2iE(\Lambda)t} e 2 i E ( Λ ) t that decays more slowly than t − 3 t^{-3} t − 3 , so the integrand must be negligible at Λ \Lambda Λ rather than merely truncated there. The dipolar integrand falls as e − 2 k e^{-2k} e − 2 k and is negligible at Λ = 10 \Lambda= 10 Λ = 10 , beyond which α ( t ) \alpha(t) α ( t ) no longer moves.
The coefficient η \eta η is evaluated with disorder, realization by realization, in the eigenbasis of G ^ ( k ) \hat{G}(k) G ^ ( k ) : we compute Π Q ( k , 0 ) \Pi_Q(k,0) Π Q ( k , 0 ) from (64) at the smallest ring wavevector Q = 2 π / N t Q = 2\pi/N_t Q = 2 π / N t , divide by Q 2 Q^2 Q 2 , and integrate over k k k by Gauss–Legendre quadrature, normalized so that the uniform array returns A Λ ( 0 ) A_\Lambda(0) A Λ ( 0 ) of (50) . For nearest neighbors G ^ \hat{G} G ^ is independent of k k k and one diagonalization per realization serves every quadrature node; for the dipolar kernel G ^ ( k ) \hat{G}(k) G ^ ( k ) is diagonalized at each node.
The cutoff for η \eta η is set by the large-k k k behavior of its integrand, which is fixed by the kernel rather than by the disorder. For ε k ≫ g m \varepsilon_k \gg g_m ε k ≫ g m the squeeze angles are small, θ m , k ≃ g m / k 2 \theta_{m,k} \simeq g_m/k^2 θ m , k ≃ g m / k 2 , so sinh 2 ϑ m n , k ≃ ( g m − g n ) 2 / k 4 \sinh^2\vartheta_{mn,k} \simeq(g_m-g_n)^2/k^4 sinh 2 ϑ mn , k ≃ ( g m − g n ) 2 / k 4 , and the mode sum of (64) collapses to a commutator with the transverse phase,
∑ m , n ( g m − g n ) 2 ∣ F m n ( Q ) ∣ 2 = ∑ i , j ∣ G ^ i j 2 ∣ ∣ e i Q i − e i Q j ∣ 2 ≈ Q 2 ∑ i ≠ j d i j 2 V i j ( k ) 2 n i n j , (125) \sum_{m,n}(g_m-g_n)^2\left|F^{mn}(Q)\right|^2
=
\sum_{i,j}\left|\hat{G}_{ij}^{2}\right|\left|e^{iQ_i}-e^{iQ_j}\right|^2
\approx Q^2\sum_{i\ne j}d_{ij}^2V_{ij}(k)^2n_in_j, \tag*{(125)} m , n ∑ ( g m − g n ) 2 ∣ F mn ( Q ) ∣ 2 = i , j ∑ G ^ ij 2 e i Q i − e i Q j 2 ≈ Q 2 i = j ∑ d ij 2 V ij ( k ) 2 n i n j , (125)
for any density profile. The large-k k k integrand of η \eta η therefore falls as k − 2 k^{-2} k − 2 with a k k k dependence set by the kernel alone, and since distinct tubes are independent, ⟨ n i n j ⟩ = n 0 2 \langle n_i n_j\rangle= n_0^2 ⟨ n i n j ⟩ = n 0 2 for i ≠ j i\ne j i = j and its disorder average equals the uniform-array value, which in the form of (50) is 4 ( g q ′ ) 2 / k 2 4(g'_q)^2/k^2 4 ( g q ′ ) 2 / k 2 . For nearest neighbors g q = n 0 ( U c + 2 V cos q ) g_q = n_0(U_c+2V\cos q) g q = n 0 ( U c + 2 V cos q ) carries no k k k , and
∫ Λ ∞ d k 4 ( g q ′ ) 2 k 2 = 4 ( g q ′ ) 2 Λ , (126) \int_\Lambda^\infty dk\,\frac{4(g'_q)^2}{k^2}
=
\frac{4(g'_q)^2}{\Lambda}, \tag*{(126)} ∫ Λ ∞ d k k 2 4 ( g q ′ ) 2 = Λ 4 ( g q ′ ) 2 , (126)
an algebraic tail. For the dipolar kernel g q ′ = − 2 n 0 V k ∑ r K 1 ( k r ) sin q r g'_q=-2n_0V_k\sum_r K_1(kr)\sin qr g q ′ = − 2 n 0 V k ∑ r K 1 ( k r ) sin q r from Eq. (14) , and K 1 ( z ) → π / 2 z e − z K_1(z)\to\sqrt{\pi/2z}e^{-z} K 1 ( z ) → π /2 z e − z leaves the r = 1 r=1 r = 1 term dominant, so g q ′ ∼ e − k g'_q\sim e^{-k} g q ′ ∼ e − k and
∫ Λ ∞ d k 4 ( g q ′ ) 2 k 2 ∼ e − 2 Λ 2 Λ 2 . (127) \int_{\Lambda}^{\infty} dk\,\frac{4(g'_q)^2}{k^2}\sim\frac{e^{-2\Lambda}}{2\Lambda^2}. \tag*{(127)} ∫ Λ ∞ d k k 2 4 ( g q ′ ) 2 ∼ 2 Λ 2 e − 2Λ . (127)
The dipolar truncation is therefore negligible at modest Λ \Lambda Λ , and Λ = 6 \Lambda=6 Λ = 6 suffices, while the nearest-neighbor one is reduced only in proportion to 1 / Λ 1/\Lambda 1/Λ . For nearest neighbors the exact Λ → ∞ \Lambda\to\infty Λ → ∞ value is known in the uniform array, since ∫ 0 ∞ d k k 2 ( ϵ k + 2 g q ) − 2 = π / 2 g q \int_0^\infty dk\,k^2(\epsilon_k+2g_q)^{-2}=\pi/2\sqrt{g_q} ∫ 0 ∞ d k k 2 ( ϵ k + 2 g q ) − 2 = π /2 g q gives η = π 8 ∫ d q ( g q ′ ) 2 / g q \eta=\frac{\pi}{8}\int dq\,(g'_q)^2/\sqrt{g_q} η = 8 π ∫ d q ( g q ′ ) 2 / g q , the elliptic integral of Eq. (104) , and the computed η \eta η at σ = 0 \sigma=0 σ = 0 approaches it with the 1 / Λ 1/\Lambda 1/Λ deficit that Eq. (126) predicts. That value cannot set the cutoff, being a uniform-array result with no counterpart at finite σ \sigma σ . The cutoff is instead set on the quantity we report: Λ \Lambda Λ is raised until the suppression 1 − η ( σ ) / η ( 0 ) 1-\eta(\sigma)/\eta(0) 1 − η ( σ ) / η ( 0 ) stops moving, which happens well before η \eta η itself has converged, since by Eq. (125) the truncated tail is common to numerator and denominator on average, to leading order in g / ϵ k g/\epsilon_k g / ϵ k . We take Λ = 800 \Lambda=800 Λ = 800 for nearest neighbors and split the quadrature at k = 10 k=10 k = 10 to keep the small-k k k region resolved. The same fact sets how the two kernels weight the disorder: the nearest-neighbor k − 2 k^{-2} k − 2 tail places much of η \eta η at large k k k , where its average is independent of σ \sigma σ , so the suppression is diluted and c c c is smaller, while the dipolar integrand never reaches that regime.
At that cutoff the nearest-neighbor suppression is a small difference of two large numbers, and the scatter of single realizations is dominated by a first-order term. Since η \eta η is translation invariant on the uniform ring, η ( σ ) − η ( 0 ) = c 1 ∑ i u i + O ( u 2 ) \eta(\sigma)-\eta(0)=c_1\sum_i u_i+O(u^2) η ( σ ) − η ( 0 ) = c 1 ∑ i u i + O ( u 2 ) , which averages to zero but is of order σ N t \sigma\sqrt{N_t} σ N t in each realization, against a signal of order σ 2 \sigma^2 σ 2 ; in Eq. (125) it is carried by n i n j − n 0 2 ≃ n 0 2 ( u i + u j ) n_i n_j-n_0^2\simeq n_0^2(u_i+u_j) n i n j − n 0 2 ≃ n 0 2 ( u i + u j ) . We remove it by antithetic sampling, pairing each profile n i n_i n i with its mirror 2 n 0 − n i 2n_0-n_i 2 n 0 − n i , equally likely under the distribution of Eq. (8) and positive, which cancels all odd orders in u i u_i u i exactly; for nearest neighbors at σ = 0.25 \sigma=0.25 σ = 0.25 this reduces the statistical error sixfold at fixed cost. The size scan of Figure 14 uses these pairs, with the realization count raised at small N t N_t N t , to 400 pairs at N t = 125 N_t=125 N t = 125 , where diagonalization is cheap. The suppression is then independent of N t N_t N t to within a few percent for both kernels between N t = 125 N_t=125 N t = 125 and N t = 1000 N_t=1000 N t = 1000 . For the dipolar kernel at N t = 1000 N_t=1000 N t = 1000 it is also independent of the coupling to within 1% across V = 0.3 V=0.3 V = 0.3 , 0.1 and 0.05, at σ = 0.25 \sigma=0.25 σ = 0.25 and 0.5, while ξ l o c \xi_{\mathrm{loc}} ξ loc changes by more than an order of magnitude.
References