PHYSICAL REVIEW B 90, 125154 (2014)

Unconstrained tree tensor network: An adaptive gauge picture for enhanced performance
M. Gerster,1 P. Silvi,1 M. Rizzi,2 R. Fazio,3,4 T. Calarco,1 and S. Montangero1
1
Institut für Quanteninformationsverarbeitung, Universität Ulm, D-89069 Ulm, Germany
2
Institut für Physik, Johannes-Gutenberg-Universitaet, D-55128 Mainz
3
NEST, Scuola Normale Superiore & Istituto Nanoscienze CNR, I-56126 Pisa, Italy
4
Centre for Quantum Technologies, National University of Singapore, 3 Science Drive 2, 117543 Singapore
(Received 13 June 2014; revised manuscript received 12 September 2014; published 29 September 2014)

We introduce a variational algorithm to simulate quantum many-body states based on a tree tensor network
ansatz which releases the isometry constraint usually imposed by the real-space renormalization coarse graining.
This additional numerical freedom, combined with the loop-free topology of the tree network, allows one to
maximally exploit the internal gauge invariance of tensor networks, ultimately leading to a computationally
flexible and efficient algorithm able to treat open and periodic boundary conditions on the same footing. We
benchmark the novel approach against the 1D Ising model in transverse field with periodic boundary conditions
and discuss the strategy to cope with the broken translational invariance generated by the network structure. We
then perform investigations on a state-of-the-art problem, namely, the bilinear-biquadratic model in the transition
between dimer and ferromagnetic phases. Our results clearly display an exponentially diverging correlation
length and thus support the most recent guesses on the peculiarity of the transition.

DOI: 10.1103/PhysRevB.90.125154 PACS number(s): 05.30.−d, 02.70.−c, 64.60.De, 05.50.+q

I. INTRODUCTION of critical states [27], both in those cases where area laws
are logarithmically violated [e.g., in one-dimensional (1D)
The approach of simulating quantum many-body states
critical systems] and in those where area laws are satisfied (e.g.,
with a tailored microscopical variational ansatz has been
bosonic critical systems in two or higher dimensions) [16].
refreshed in the last decade thanks to the introduction of
TTN states show a smooth computational scaling with the
tensor network states. Despite being originally related [1–5]
tensor network bond dimension m for the involved algebraic
to density matrix renormalization group [6–8] schemes, these
operations [e.g., for binary TTNs it is never higher than
variational approaches have been engineered to encompass
O(m4 )]. This allows one to push numerical precision and de-
a wide variety of physical situations [9,10], thus widening
scription capabilities by sensitively increasing m, making the
the capabilities of the traditional numerical renormalization
TTN ansatz a potentially competitive method for simulating
group (RG) approach. Generally, tensor networks encode
quantum many-body states. On the other hand, TTNs suffer
in a compact, numerically efficient way, many-body wave-
more of a kind of entanglement clusterization, which is much
function amplitudes over a real-space local basis expansion:
more alleviated in other approaches, such as MERA, thanks to
the main reason for this real-space choice is the fact that,
the presence of disentangling operations in their structure.
since typical Hamiltonians are characterized by two-body
It is important to stress that the traditional scheme for
interactions which decay sufficiently fast with the pairwise
simulating quantum lattice models with TTN states [28]
distance, physically meaningful states (e.g., ground states, relies on a particular selection of the internal tensor network
lowest excited states, thermal states) obey precise scaling gauge symmetry. In accordance with the RG flow picture
laws on entanglement entropy under a real-space bipartition by Wilson, the tensors are fixed to be isometric operators
[11–14]. Such entanglement scaling can be precisely encoded in the real-space renormalization direction. Although this
in the tensor network paradigm [15,16] and led to the design gauge selection has historical motivation and, moreover, it
of various planar tensor network geometries, such as matrix guarantees some useful mathematical properties in the thermo-
product states (MPSs) [1], projected entangled pair states dynamical limit (namely, the complete positivity of the causal
(PEPS) [17], and complete graph states [10]. maps [21,22,24,29]), it confers no advantage in the simulation
A physically sensible class of tensor network architectures of finite-size systems, where actually it is more a hindrance.
is the hierarchical (or holographic [18]) tensor networks, which Instead, if no “a priori” rigid selection of the isometric gauge
have the key feature of combining the usual local quantum from bottom to top is made, one can always adjust the tensor
space numerical renormalization together with a real-space network gauge to gain a computational enhancement in the
coarse graining, much like in the original RG picture by algorithm. This type of manipulation is particularly useful for
Wilson [19]. Tree tensor networks (TTNs) [20,21], multiscale tensor networks without closed loops in their topology, like
entanglement renormalization ansatz (MERA) states [22–24], TTN, for which gauge flexibility translates into a simplification
and the recently introduced branching MERA [25] are the of the variational algorithm into a simple eigenvalue problem
most prominent examples of hierarchical tensor networks. (as it happens, for instance, for open boundary MPSs compared
The fact that their network structure naturally embeds a scale to periodic ones).
invariance makes them the ideal choice for representing critical In this manuscript we describe in detail an algorithm to
(gapless) quantum phases of matter, which are characterized find the ground state of quantum lattice Hamiltonians, based
by conformal invariance [26]. Moreover, hierarchical tensor on unconstrained (i.e., gauge-adaptive) tree tensor networks.
networks can indeed satisfy the scaling rules of entanglement We discuss thoroughly the computational cost scaling with

1098-0121/2014/90(12)/125154(10) 125154-1 ©2014 American Physical Society

M. GERSTER et al. PHYSICAL REVIEW B 90, 125154 (2014)

numerical parameters, first of all the tensor network bond (a)

dimension m. We test the algorithm on one-dimensional Λ[0,1]
quantum models in both open (OBCs) and periodic boundary m m
L layers
conditions (PBCs) using the quantum Ising model as a
benchmark, and then investigating the bilinear-biquadratic
spin-1 model in proximity of the interface between the
Λ[1,1] Λ[1,2]
ferromagnetic and dimer phases. In the latter, a peculiar
exponential scaling of correlation lengths has been conjectured
to explain the traditional toughness of the numerical problem.
The nice agreement of our data with the most recent state-of- Λ[2,1] Λ[2,2] Λ[2,3] Λ[2,4]
the-art calculations corroborates the validity of the variational
strategy presented here. We also address the problem of 1 2 3 4 5 6 7 8
restoration of translational invariance, which is broken by the
N sites
TTN architecture; we inquire for which physical quantities
it is meaningful to average incoherently over translations as
opposed to local evaluation over a highly entangled cluster of (b) 1 8
sites. Ultimately, such investigation reveals quite a different
behavior between local observations and correlations.
2 7

II. STATE ARCHITECTURE AND ALGORITHM

We consider a one-dimensional lattice with N = 2L sites,
where each site has a local Hilbert dimension of d. Our
3 6
ansatz to approximate a many-body quantum state | =
 N
 χ1 ...χN
{χ} i=1 |χi  on such a lattice, where the strings
χ = (χ1 . . . χN ) label the configurations of the N sites in
some local “canonical” basis (with χi = 1 . . . d), is displayed 4 5
in Fig. 1. It is a binary TTN [20,21,28], a hierarchical
FIG. 1. (a) Structure of the binary TTN with L layers and N sites.
structure consisting of L layers of tensors [l,n] with three
The maximal bond dimension is m. (b) The same TTN displayed in
indices each, where l = 0 . . . L − 1 indicates the layer and
a PBC configuration, showing its natural capability of treating OBC
n = 1 . . . 2l denotes the horizontal position of the tensor. and PBC on the same footing.
The sketch follows the usual convention of drawing tensor
indices as “legs” or “links”; joining two tensor legs has the
usual meaning of a contraction, i.e., a summation over the In this manuscript, we are going to apply this variational
corresponding indices of the tensor elements product. All ansatz to approximate the ground state of nearest-neighbor
tensors in Fig. 1 have three legs, except for the top tensor [0,1] interacting spin-Hamiltonians on a lattice. The application
that is two-legged, which can be viewed as the contraction to bosons is straightforward, while extension to fermions is
of a three-legged tensor with a vector encompassing a wave carried out via standard Jordan-Wigner transformation [31].
function on a renormalized (degenerate) manifold [30]. The For convenience, we write the local fields and the spin-spin
physical sites of the chain are represented by the dots attached interactions separately, resulting in a Hamiltonian of the
to the bottom of the lowest layer of tensors. Each tensor following form:
effectively merges two sites into a single “virtual” site,   
allowing one to interpret the tensors as coarse-graining linear H = H[n] + λα H← H→
α[n] α[n+1]
, (1)
n n α
maps in a RG flow. Labeling the sites (either physical or
virtual) with [l,n] (l = 1 . . . L: layer coordinate; n = 1 . . . 2l : where H[n] (local term), H← [n]
(left interaction term), and H→
[n]
horizontal coordinate), the tensor [l,n] maps the two sites (right interaction term) are on-site operators acting on site n.
[l + 1,2n − 1] and [l + 1,2n] to the site [l,n]. Consequently, The index α accounts for the fact that the interaction term can
the full Hilbert spaces of the sites in layer l have dimension consist of several tensor-product contributions, weighted by
L−l
M(l) = d 2 . Such dimension is exponentially growing in their couplings λα . A strong point of the TTN ansatz is that
the number 2L−l of physical sites blocked together in layer l, it adapts comparatively well to both the OBC (n = 1 . . . N −
and therefore a numerically efficient representation of such 1) setting and the PBC (n = 1 . . . N and N + 1 ≡ 1) setting.
degrees of freedom requires some kind of space truncation. The In the two cases, the computational costs are equal, and the
most easily controlled truncation method is fixing a maximally numerical precisions compatible.
allowed value m (an upper bound to the so-called “bond A key requirement for any tensor network representation is
L−l
dimension”), resulting in M(l) = min(d 2 ,m). Moreover, the its efficient contractibility, which is instrumental to gaining
total number of tensors in the binary tree network is N − 1, access to physically sensible information on the quantum
and since we are not introducing additional constraints in the many-body state, such as its energy or expectation values
variational picture, the total number of parameters in the TTN of observables. Indeed, a prominent advantage of TTN
representation ultimately scales like O(N m3 ). architectures is that they are algebraically contractible, thus

125154-2
UNCONSTRAINED TREE TENSOR NETWORK: AN . . . PHYSICAL REVIEW B 90, 125154 (2014)

R
Q 1
5
Λ = Q , with = 11 =
QR dec.
Q∗
1
4
5
FIG. 2. Isometrization of a three-legged tensor . The arrow
indicates the leg with respect to which the tensor is isometric. In 2
the picture, Q∗ is the element-wise complex complex conjugate of 1 3
Q, and it is also drawn as vertically reflected.
3
2
providing exact expectation values efficiently, without the need 5
4
of stochastic sampling of their variational data [32,33]. One can
easily identify two properties of TTN that make this possible:
the first is the loopless structure of the tree network and the 3
5
second is the exploitation of a flexible, adaptive isometric
gauge selection. The latter is based on a straightforward
generalization of the QR decomposition [34] applied to FIG. 3. (Color online) TTN isometrization with respect to the
three-legged tensors , which produces a directed-isometric tensor highlighted in red, i.e., all isometrization arrows point towards
tensor Q and a matrix R, such that this tensor. To obtain this gauge, QR-decompose all tensors in order
of decreasing distance (indicated by red numbers).
α1 α2 α3 = Qα1 α2 β1 Rβ1 α3 , Q∗α1 α2 β1 Qα1 α2 β2 = δβ1 β2 , (2)
where Q is isometric with respect to the third leg. (In the
graphical notation of the TTN, we draw an outgoing arrow H←α[l,n]
, and H→
α[l,n]
, resulting from performing the isometric
from Q on that leg.) In Fig. 2 we report Eq. (2) expressed mapping operation sketched in Fig. 4. By identifying the
in graphical notation. We say that the TTN is isometrized effective Hamiltonian terms on the physical sites layer as the
with respect to tensor [l,n] if all the other tensors in the original Hamiltonian, i.e., H[L,n] ≡ H[n] , H↔ α[L,n]
≡ H↔α[n]
,
tree are isometrized in the direction of this tensor according the mapping is a well-defined operation for every link. On the
to the network structure, i.e., all the arrows are pointing other hand, we do not have to explicitly define effective identity
towards [l,n] . Note that due to the loopless topology of the operators N [l,n] , since the identity operator 1 is invariant under
TTN, such an isometric gauge is always unique (apart from the mapping by construction. The mapping is always to be
unitary gauge transformations) and efficiently attainable. Just carried in the direction of the adaptive gauge isometrization:
perform the QR-decomposition equation (2) for all the tensors, starting from contracting the Hamiltonian pieces, which are
starting from the most distant ones (in the network metric, farther from the node [l,n] in the network metric, and then
i.e., the graph distance of two nodes in the tree network)
and absorbing the gauge matrices R into tensors yet to be
isometrized, until reaching [l,n] . This concept is illustrated
in Fig. 3. Also, note that in the adaptive isometric gauge Λ[l,n] Λ[l,n] Λ[l,n]
the calculation of the many-body state norm | simply
collapses to ∗ [l,n] [l,n]
α1 α2 α3 α1 α2 α3 , regardless of the node position H[l,n] = H[l+1,2n−1] 11 + 11 H[l+1,2n] + [l+1,2n−1]
H←
[l+1,2n]
H→
in the network, since all the other partial contractions cancel
to identities due to the isometry condition.
Such a flexible gauge selection provides a crucial advantage Λ∗[l,n] Λ∗[l,n] Λ∗[l,n]
along the search for the best TTN representation of the
ground state of the Hamiltonian H in Eq. (1). Expressed in
a variational sense, the task consists in searching the set of
Λ[l,n] Λ[l,n]
tensors {[l,n] } such that E = |H |/| is minimal.
For practical values of N and m, the complete space of
variational parameters of all tensors combined is too large to [l,n]
H← = 11 [l+1,2n]
H←
,
[l,n]
H→ = [l+1,2n−1]
H→ 11
allow a successful application of a direct search optimization;
instead, the approach pursued here relies on an iterative
Λ∗[l,n] Λ∗[l,n]
strategy, optimizing one tensor at a time while assuming all
other tensors to be fixed. A sensible reason to choose this
strategy is that the optimization problem for a single tensor FIG. 4. Mapping operation induced by the tensor [l,n] for the
[l,n] actually reduces to a simple eigenvalue problem for an case of upward-directed isometrization arrow. The operation maps the
[l,n]
effective Hamiltonian Heff acting on the degrees of freedom effective Hamiltonian terms at sites [l + 1,2n − 1] and [l + 1,2n] to
[l,n]
of  alone, in an analogous fashion to the density matrix the new terms at site [l,n]. For the two other possible isometrization
RG with single center site [35]. To see this, we define iteratively directions similar mapping operations hold. The index α has been
for each virtual site [l,n] the effective Hamiltonian terms H[l,n] , dropped for simplicity.

125154-3
M. GERSTER et al. PHYSICAL REVIEW B 90, 125154 (2014)

more demanding and unstable than a SEP when addressed

H[l,n] numerically [36].
After these considerations, the ground-state search algo-
[l,n]
rithm is summarized as follows:
Heff |Λ[l,n] = Λ[l,n] + Λ[l,n] + Λ[l,n] (i) Initialize all tensor entries:
(a) by picking a random state in the TTN manifold;
H[l+1,2n−1] H[l+1,2n]
(b) by selecting a particularly symmetric or meaningful
state (e.g., a ferromagnetic product state); and
(c) by performing some iterations of an exponentially
[l,n]
H→
[l,n]
H←
growing DMRG-like procedure.
In the following, we focus on strategy (a) to prove that the
algorithm is robust, as its convergence is ultimately insensitive
+ Λ[l,n] + Λ[l,n] + Λ[l,n] to the initialization.
(ii) Select a tensor [l,n] in the network. Isometrize the
[l+1,2n−1] [l+1,2n] [l+1,2n] [l+1,2n−1]
TTN in the direction of [l,n] and perform the mapping
H← H→ H← H→
operations according to the directed network. Optimize [l,n]
by solving Eq. (4).
FIG. 5. (Color online) Definition of the action of the effective (iii) From [l,n] move to the next tensor [l ,n ] , adjusting
Hamiltonian Heff[l,n]
on the degrees of freedom of the tensor [l,n] . the isometrization and updating the effective Hamiltonian
(Again, the index α has been dropped for simplicity.) terms. Note that only tensors located on the path connecting
[l,n] and [l ,n ] are affected by this move (see Fig. 6). Having
determined the new effective Hamiltonian, optimize [l ,n ] via
proceeding closer and closer, until the effective Hamiltonian Eq. (4) again.
acts only on [l,n] itself. According to this picture, the (iv) Repeat (iii), targeting each tensor in the tree by
calculation of the energy expectation value for the TTN can be following some “sweeping” pattern (e.g., the one indicated
[l,n] [l,n]
written as E = [l,n] |Heff |[l,n] , where the action of Heff in Fig. 6); stop when convergence in the ground-state energy
on |  is meant as indicated in Fig. 5. The minimization
[l,n]
is reached, according to some precision threshold.
problem for [l,n] , subject to the constraint of normalization, The sweeping action is the key point which pushes the TTN
is then easily solved by the method of Lagrange multipliers. representation beyond the simple numerical real-space RG
We then write the following Lagrangian, flow. Optimizing the same tensor multiple times while tuning
[l,n]
L(|[l,n] ,[l,n] |,λ) = [l,n] |Heff |[l,n]  the environment (i.e., the rest of the network surrounding
it) makes the algorithm a complete variational approach,
−λ ([l,n] |1|[l,n]  − 1) , (3) guaranteed to converge in the TTN manifold. Summing up,
we have reduced the energy minimization problem to a
and since tensors different from [l,n] are assumed to be fixed, sequence of QR decompositions, linear mapping operations,
the Euler-Lagrange equation simply reads and SEPs, all of which can be carried out in a numerically
[l,n]
Heff |[l,n]  = λ 1|[l,n]  , [l,n] |1|[l,n]  = 1 , (4) stable fashion with a computational cost of O(m4 ). Let us
stress that this scaling behavior is independent of the boundary
[l,n]
which is a standard eigenvalue problem (SEP) for Heff . By conditions chosen, in contrast to MPS ansätze; the only change
linearity, it is readily seen that the normalized eigenvector accompanying a switch from PBCs to OBCs consists in
[l,n] omitting the terms that mediate the interaction between the
of Heff corresponding to the lowest eigenvalue is the best
choice for [l,n] to minimize E. We purposely highlighted the physical sites 1 and N when calculating the action of the
identity operators in Eq. (4) in order to stress the benefit that is effective Hamiltonian (Fig. 5). These terms occur only for
obtained from the isometric gauge selection. In fact, choosing the outermost tensors in a layer (i.e., n = 1 or n = 2l ) and
a different gauge would require us to substitute the 1 for a thus produce a subleading change in the overall computational
(nontrivial) effective identity operator N [l,n] , turning Eq. (4) cost. Moreover, we remark that the O(m4 ) contraction cost
into a generalized eigenvalue problem, which is significantly relies only on the loop-free network structure and the fact that

(a) (b) (c)

*

Λ[l ,n ] Λ[l,n] Λ[l ,n ]

Λ[l,n]

FIG. 6. (Color online) Generic optimization move: (a) after optimizing [l,n] , the tensor [l ,n ] is targeted for optimization. (b) Only tensors
and effective Hamiltonian terms located on the path connecting [l,n] and [l ,n ] (colored in blue) need to be updated in order to enable the
optimization of [l ,n ] . (c) Targeting each tensor in the tree multiple times results in a sweeping pattern. After completing a sweep, resume at
the top (as indicated by the encircled marks).

125154-4
UNCONSTRAINED TREE TENSOR NETWORK: AN . . . PHYSICAL REVIEW B 90, 125154 (2014)

r
l
Ψ|O[n] |Ψ = O[n] = O[n] r
sc

FIG. 8. (Color online) Location of the central site sc in a 1D

binary TTN.

FIG. 7. (Color online) Expectation value of a local observable identify the location sc of these sites by alternately following
O [n] . Due to the isometry condition, the calculation always reduces the left (l) or right (r) branch down the tree, starting from the
to a contraction involving only three tensors. top (see Fig. 8). Algebraically, this can be written as

L/2
tensors have three legs. This means that the binary TTN ansatz sc = 1 + 2L−2l = (N + 2)/3 , (5)
can naturally be extended to other lattices and dimensionalities l=1
(e.g., two dimensions, Cayley trees [9]) without increasing the
numerical effort for the contractions. where x is the floor function of x. In order to clarify the
We conclude this section by describing how to extract situation for the 1D-TTN, we use a benchmark Ising model
expectation values of the TTN state |. This is particularly to compute local observables and two-point correlators at all
convenient for local observables O [n] , having support on a lattice sites and compare them with their respective system-
single lattice site n. By isometrizing the TTN with respect to wide averages. The results of this analysis are presented at the
[L−1, n/2 ] (i.e., the tensor directly attached to site n; x is end of the next section.
the ceiling function of x), the expectation value |O [n] | IV. BENCHMARKING OF THE ALGORITHM
collapses to a contraction of only three tensors, as indicated in
Fig. 7. In the case of two-point correlators O [n] O [n+r]  (two To test the algorithm outlined in Sec. II, we consider the
local observables separated by a distance r), a similar proce- spin- 12 Ising model in a transverse field with PBCs, defined by
dure can be adopted; the only difference is that now nontrivial the Hamiltonian
contractions arise for all tensors on the network path connect- 
PBC
ing the two sites n and n + r (which is unique thanks to the H =− σnx σn+1
x
+ h σnz , (6)
loop-free network structure). Given that the maximum number n=1
of tensors on this path scales logarithmically in the lattice size
N, two-point correlators can be computed very efficiently. where σnα (α = x,y,z) is a Pauli matrix acting on spin n and
the parameter h is the external magnetic field. We focus on
the critical point of the model at |h| = 1, where, as previously
III. TRANSLATIONAL INVARIANCE IN
stated and argued, e.g., in Refs. [21,24,38], the entanglement
THE TTN ARCHITECTURE
scaling of TTN should prove more useful. This analytically
Before benchmarking the presented algorithm, let us solvable model [39,40] constitutes a commonly employed
analyze the issue of the translational invariance of the lattice, excellent benchmark for the quantities of our interest, namely,
which is broken by the design of the TTN ansatz. Indeed, it the ground-state energy E, the central charge c, and the
is clear from Fig. 1 that some sites are better connected with spin-correlation functions C α (r) = σnα σn+r
α
 − σnα σn+r
α
,
their immediate environment (i.e., the neighboring sites) than including their respective critical exponents ηα .
others [37]. For instance, a very poorly connected environment First of all we report the convergence behavior of the
occurs for the sites at N/2 and N/2 + 1, which are nearest algorithm, shown on the left-hand side of Fig. 9. Convergence
neighbors, and yet they are renormalized together (in the to the ground state as a function of the number of optimization
RG-flow picture) only at the last step, i.e., at the very top sweeps is fast, with roughly ten sweeps being sufficient to
of the tree network. The translational symmetry breaking reliably achieve the global minimum in the TTN state space.
induced by the TTN design makes one wonder what is the Only for very big system sizes starting from N = 1024
best strategy to obtain the most accurate observation results sites did we observe the emergence of local minima of the
when we are simulating a translationally invariant model. A optimization problem (visible, e.g., in the curve for m = 40
legitimate question is whether it is beneficial or detrimental in the upper-left plot of Fig. 9), but even in those cases the
to measure at one site (or region) in particular rather than algorithm converges to the global minimum after performing
translationally averaging the measurements over the lattice. a few extra sweeps. As can be seen from the upper-right plot in
An analysis in this direction has been recently developed for a Fig. 9, the error of the ground-state energy per site tends to be
two-dimensional TTN design and reported in Ref. [37]. In that size-independent with increasing N , a feature the TTN shares
scenario, a significantly more accurate description of local with other hierarchical tensor network ansätze [29] when
quantities has been obtained by focusing on “central sites,” simulating a (1+1)D gapless system. An intuitive argument
which are defined by the criterion that the Hilbert space of their to motivate this is that, in order to describe the entanglement
immediate environment is largest. In our 1D setting we can of a localized region with given precision, one has to pick

125154-5
M. GERSTER et al. PHYSICAL REVIEW B 90, 125154 (2014)

0 log2 (N )
-2 N = 1024: 4 5 6 7 8 9 10
-4 N=
N = 32: 16 256
0 -4 -6

log 10 (ΔE/N)
32 512
-2 -6 -8 64 1024
0 4 8 12 m= -10 128 fit
log10 (ΔE/N)

-4
-12 -4
10
-6 20 -6 N →∞

log10 (ΔE/N )
-14
-8 30 -8
-16
40 -10
-10 50
-12
60
-12 -14 mach. prec.
70
-16
0 2 4 6 8 10
10 100
m
number of sweeps

FIG. 9. (Color online) Ising model: Error of the ground-state energy per site E/N as a function of the number of optimization sweeps
for two different system sizes N = 32 and N = 1024 (left). Error as a function of the system size (right upper). The same quantity as a
function of the bond dimension m. The data for N = 32 is fitted by am−b exp(−c m), with a = 12.8, b = 5.5, c = 0.11. Extrapolation to the
thermodynamic limit suggests a polynomial decay of the form E ∝ m−3.3 (right-lower).

a fixed value m and not one that scales with the length L. improvements lies in the fact that the adaptive TTN avoids the
In a critical scenario, this can happen only if a given m numerically unstable generalized eigenvalue problems [36]
captures the critical correlation scaling, i.e., the logarithmic and requires fewer isometrization operations.
entanglement area law violation. In the lower-right graph we In order to assess how the adaptive TTN deals with
plot the dependency of the error on the bond dimension m. In criticality and, consequently, strongly correlated systems, we
double-logarithmic representation the curves clearly exhibit a report the spin-correlation functions and the von Neumann
negative curvature, meaning that the decay is subpolynomial entropy in Fig. 11. Since we are at criticality, we expect
for any finite N. In the relevant range of m the behavior can be
described well by a combined polynomial-exponential decay
of the form E(m) = am−b exp(−c m), where a, b, c are fit 100
parameters dependent on the system size (see Fig. 9). In the
10−1 0
thermodynamic limit N → ∞ the decay seems to be governed slope:
log10 |C α |

by a polynomial behavior of the form E(m) ∝ m−3.3 . -2

0.2500
10−2 Cx 2.2507
In Fig. 10 we compare the performances of the adaptive
|C α |

-4 Cy 2.0008
TTN and a constrained TTN. In the latter, each tensor [l,n] is Cz
constrained to be an isometry in the direction of the RG flow, 10−3
0 0.3 0.6 0.9 1.2 1.5
namely, from bottom to top [20,21]. The comparison clearly −4 log10 N
π
π
sin N r
10
shows that the adaptive TTN algorithm exhibits a significantly
improved convergence behavior. Additionally, the adaptive
20 40 60 80 100 120
TTN provides a computational speed-up; in our benchmarks
the time required to perform one optimization sweep decreased r
by about a factor of 3. As mentioned above, the reason for these 1.8
1.6
0 1.8
constrained TTN, adaptive TTN, 1.4 1.5 slope: 0.502
-2 N = 256: N = 256: m= 1.2
S

1.2
S
log10 (ΔE/N )

20 0.9
30 1 0.6
-4
40 0.8 0 0.4 0.8 1.2 1.6 2
1
50 3
log2 N
π
π
sin N
-6 0.6
60 20 40 60 80 100 120
-8 70
80
0 2 4 6 8 10 0 2 4 6 8 10 FIG. 11. (Color online) (Top) Spin-correlation functions and fit-
number of sweeps ted critical exponents for a PBC quantum Ising chain of N = 128
sites, bond dimension m = 100. The exact exponents are ηx =
FIG. 10. (Color online) Error of the ground-state energy per 0.25, ηy = 2.25, ηz = 2. In order to account for the TTN-induced
site E/N as a function of the number of optimization sweeps. translational symmetry breaking, the correlations of distance r are
Comparison between a constrained and an adaptive TTN. The obtained by averaging over lattice locations within branches wide
generalized eigenvalue problems present in the constrained TTN enough to span r (see text). (Bottom) Von Neumann entropy of a
algorithm cause strong instabilities in the energy minimization, N = 256 (m = 100) chain, for partitions of length and with fitted
hindering a reliable convergence behavior. central charge. The exact central charge is c = 0.5.

125154-6
UNCONSTRAINED TREE TENSOR NETWORK: AN . . . PHYSICAL REVIEW B 90, 125154 (2014)

y
m = 050 m = 100 |ΔCn,n+r |
m = 030 m = 070 -4 64 10−5
m = 050 m = 100

log10 |ΔC y |
-4 -6
-8 10−6
-5 -10 32
10−7
-6 -12
log10 |ΔEn |

-7 -14 0 10−8

r
log10 |ΔC y |
-8
-6 10−9
-9 -32
-8 10−10
-10
sc
-11 -10 10−11
16 32 48 64 80 96 112 128 32 64 96 128 32 64 96 128
n n n

FIG. 12. (Color online) Error of the local energy En as a function of the lattice site n for different bond dimensions m, in the critical
quantum Ising model with PBC defined in Eq. (6). The solid lines are the errors of the system-wide averages. The vertical dashed line indicates
y
the location of the center site sc (left). Error of the spin two-point correlator Cn,n+r for bond dimension m = 100 (right); the two insets show
the data at fixed distance r together with system-wide averages for r = 8 (top) and r = −32 (bottom).

α
the correlation functions to decay according to a power law correlators Cn,n+r for every distance r and pair position n, we
C α (r) ∝ [crd(r)]−ηα , where crd(r) ≡ N/π sin(π r/N) is the can easily address the of question whether or not it is profitable

chord function giving the effective distance of two sites at to calculate the average N α
n=1 Cn,n+r /N for a given pairwise
distance r arranged on a ring of N sites [26,41,42]. The y
distance r. Figure 12 shows the error Cn,n+r of the spin
critical exponents are analytically known [42] to be ηx = 1/4, correlations in y direction as a function of n and r. Immediately
ηy = 9/4, and ηz = 2. The via fitting numerically obtained noticeable is a triangular pattern, indicating that with increas-
exponents in the upper panel of Fig. 11 show an agreement ing r the accuracy abruptly degrades each time a major branch
of the order of O(10−4 ), presenting state-of-the-art accuracy in the tree is trespassed. The zoomed panels show that for this
to the best of our knowledge [29]. The von Neumann entropy reason it is advantageous (especially for short distances) to
S( ) ≡ −Tr[ρ ln ρ ] of a partition of sites for PBC critical measure the correlator within a branch wide enough to span
(1+1)D systems is known to behave according to S( ) = the given distance, i.e., within the dark-blue areas of Fig. 12.
c/3 log2 [crd( )] + const, where c is the central charge of the Averaging, in this case, is mainly detrimental and results in
underlying conformal field theory [43], and for the Ising model significantly bigger errors, up to 2 orders of magnitude.
we have c = 1/2. Again, we have very good agreement of the
exact value with the numerically determined central charge
c = 0.502 shown in the lower panel of Fig. 11, proving that V. BILINEAR-BIQUADRATIC SPIN-1 CHAIN
the TTN ansatz is well suited for reproducing a system at After having benchmarked the adaptive TTN algorithm
criticality. In particular, we showed that even with a relatively on the spin-1/2 quantum Ising model, in this section we
small bond dimension m, it is perfectly meaningful to simulate turn to a numerically very challenging transition in SU (2)
critical systems even with hundreds of sites, and indeed achieve invariant spin-1 chains. We consider the bilinear-biquadratic
a high precision of the results. 
Hamiltonian (H ≡ N−1 n=1 hn,n+1 ):
We conclude this section by addressing the question raised
in Sec. III, i.e., whether or not it is beneficial to average over 
N−1
expectation values. For this purpose we compute the local H = cos θ (Sn · Sn+1 ) + sin θ (Sn · Sn+1 )2 , (7)
energy En = −σnx σn+1 x
 + σnz  for the critical Ising model n=1
Eq. (6) and plot the absolute value of its error En as a function
of the site location (see Fig. 12). Also shown is the error of whose phase diagram [53,54] is provided in Fig. 13 as a
 function of the parameterizing angle θ . While most phase
the system-wide average N n=1 En /N. It is clear from Fig. 12
that the averaged quantity is about 1 order of magnitude more boundaries have been set on firm grounds relatively early [44–
accurate than the vast majority of the individual measurements 48], the region between the translationally broken dimer and
[including the location of the central sites sc as defined in the rotationally broken ferromagnet has been the subject of a
Eq. (5)]. This is due to the fact that the En are scattered long and controversial debate [44,49,50,54–60]. An intermedi-
above and below the exact value, resulting in an average ate nematic phase, restoring the translational invariance while
that is more accurate. Although isolated sites more accurate mildly breaking the rotational one by a quadrupolar moment,
than the average value exist, these do not occur at specific has been conjectured [49,50] to be in the range −3π/4 ≡ θF 
locations independent of m, which is why this behavior has θ  θC  −0.67π . The competing order parameters would
to be attributed to numerical fluctuations rather than to some then be the dimerization D [48],
systematic reason. Therefore, Fig. 12 strongly suggests that
Dn = |hn−1,n − hn,n+1 | , (8)
averaging over all sites is the best way to extract local observ-
ables from a 1D-TTN. We proceed to analyze the averaging which can be nonzero in 1D finite chains, and a quadrupole
issue for correlation functions. Since it is efficient to calculate moment Q (with e a unit vector pointing to the solid

125154-7
M. GERSTER et al. PHYSICAL REVIEW B 90, 125154 (2014)

-2.64 -2.778
-2.78 N = 128

r
-2.66

me
fit
-2.782

Tri
Fer -2.68 -2.784
rom
ag θ
-2.7
Haldane 75 90 105 120

E
-2.72 m

Dimerized
-2.74
E(m → ∞)
? -2.76
θF exact: E∞ = −2.7968...
-2.78
fit: E∞ + a · (1/N), E∞ = −2.7967
θ = − π/ 2 -2.8

0
1/256

1/128

1/64

1/32

1/16
FIG. 13. (Color online) Phase diagram of the bilinear-
biquadratic model of Eq. (7), with four firmly established phases: 1/N
ferromagnetic, trimerized [44], Haldane antiferromagnet [45,46],
FIG. 14. (Color online) Extrapolation to the thermodynamic
and dimerized [47,48]. According to the most recent studies and
limit of the ground-state energy at the purely biquadratic point
the data obtained here with adaptive TTN, no intermediate nematic
θ = −π/2. The inset shows the extrapolation in the bond dimension
phase (originally hypothesized location indicated by red hatched
for N = 128 using a fit similar to the one indicated in Fig. 9.
region [49,50]) is emerging around θF (black dot). The white dot
identifies the purely biquadratic point [47,51,52] whose Bethe ansatz
solution in Eq. (11) we use to benchmark the numerical precision of
our algorithm. and then considering the thermodynamic limit N → ∞ by
a linear fit in 1/N to eliminate edge effects, as shown in
the main plot of Fig. 14. Thus the estimated energy ETTN =
angle ), −2.796 70(1) has a precision of 10−4 .
Qn, = (Sn · e )2 − 2/3 , (9)
of which only quasi-long-range order in correlations can be
measured in 1D finite chains [see Eq. (12)]. Increasingly 1.8 0
powerful numerical techniques [44,54–60] and improved 1.6 -2
ln(D)

theoretical analyses [61–63] have shrunk the nematic window 1.4

-4
more and more by accumulating evidence of an exponen- 1.2
-6
tially vanishing D accompanied by an exponentially growing 1 9 12 15 18 21 24
D

quadrupole correlation length ξQ . The latter accounts for the 0.8 π/Δθ
N=
extreme difficulty in ruling out a quasi-long-range nematic or- 0.6 16
der based on only numerical data of finite-size chains. A recent 0.4 32
description of the Berry phases of quantum fluctuations [61] 64
0.2
128
predicted the following scalings with θ = θ − θF : 0
-0.75 -0.7 -0.65 -0.6 -0.55 -0.5
D ≈ exp (−π 2 /(8 θ )), (10a) θ/π

ξQ ≈ exp(π 2/ θ ) . (10b)
4
The data we obtain from our adaptive TTN procedure actually 3.8
nicely agree with this prediction, as we argue below, and 3.6
3.4
thus corroborate the validity of the presented method also in
ln(ξQ )

3.2
nontrivial cases. 3
First, we assess the numerical precision of the adaptive 2.8 N = 32
TTN algorithm by analyzing its performances in the purely 2.6 lin. fit N = 64
biquadratic point [47,51,52] at θ = −π/2 (empty circle in 2.4 lin. fit N = 128
Fig. 13). The ground-state energy in the thermodynamic limit 2.2
3 3.5 4 4.5 5
can be analytically obtained by Bethe ansatz (after a mapping
to a spin-1/2 XXZ chain) as [47,51,64] π/Δθ
 ∞

√ 1  1 FIG. 15. (Color online) (Top) Dimerization parameter D for var-
E/N = −1 − 2 5 + √ ious system sizes. Each data point is the result of an extrapolation
4 k=1 1 + [(3 + 5)/2]2k m → ∞ in the bond dimension. The dashed lines shown in the inset
= −2.796 863 43... (11) are linear fits. (Bottom) Quadrupolar correlation length ξQ , obtained
from fitting the correlation function Eq. (12). Each C Q (r) has been
and will serve as a reference for our finite-bond and finite-size extrapolated to m → ∞ (for every r) before the correlation length
extrapolation procedure. This consists in taking first the limit was extracted. Reported error bars are fit errors. The large error bars
m → ∞ for every chain length N (according to the empirical for N = 32 are caused by the fact that the chain is too small to contain
formula indicated in Fig. 9), as illustrated in the inset of Fig. 14, meaningful information on the long-range property ξQ .

125154-8
UNCONSTRAINED TREE TENSOR NETWORK: AN . . . PHYSICAL REVIEW B 90, 125154 (2014)

We then proceed to evaluate, for various values of θ ∈ the transition point. Moreover, we investigated how well the
[θF , − π/2], the dimerization D of Eq. (8). In order to avoid TTN ansatz can restore the broken translational invariance in
issues related to resonant superpositions of different dimerized 1D, and verified that while for local quantities averaging over
configurations, we resort to open boundary conditions. We translations is a winning strategy, for correlation functions
avoid end effects of the open chain by measuring the evaluating over highly entangled clusters is a more suitable
dimerization only at sites that are sufficiently far away from approach.
the edges. The data presented in the upper plot of Fig. 15 The redirectionable isometric gauge presented here also
nicely agree with the prediction of Eq. (10a), displaying a allows for a convenient, natural treatment of preserved global
clear exponential behavior in 1/ θ . symmetries, both Abelian and non-Abelian, which result in
Finally, we proceed to compute the correlation of a block-diagonal structure of the tensor entries with clear,
quadrupole moments at a long distance r averaged over the unique selection rules and structure constants [65–67]. (In
solid angle: this framework, a simultaneous double-tensor optimization
 scheme would help in circumventing the related intermediate
d
C (r) =
Q
Qn, Qn+r,  − Qn, Qn+r,  charge-targeting issues [35].) Similarly, lattice gauge sym-
4π metries, which in contrast to global symmetries act on local
2  α 2 α 2 2 2 compact supports of the 1D chain of sites, can be rigorously
= Sn Sn+r − Snα α
Sn+r
15 α and efficiently cast in tensor network language [68–71],
particularly in a TTN ansatz.
1  αβ αβ αβ An additional research direction is offered by the recently
+ T T − Tnαβ Tn+r , (12)
15 α<β n n+r introduced time-dependent variational principle for tensor
network states [72], whose purpose is to efficiently describe
αβ β
with the anticommutator Tn ≡ {Snα ,Sn }. The correlation out-of-equilibrium dynamics. Indeed, it can be elegantly and
length ξQ is then extracted by fitting the m → ∞ extrapolations successfully tailored to TTN, free of tensor gauge complica-
with C Q (r) = a r −η exp(−r/ξQ ), where a is a prefactor and tions, thanks to the loop-free geometry of the tree network.
η is an exponent. The results are shown in the lower panel We stress again that the flexibility of the gauge-adaptive
of Fig. 15 and provide √quite clear evidence of an exponential TTN ansatz presented here, and especially the fact that its
growth of ξQ with 1/ θ . In summary, these results support scaling is independent from the specific (open or periodic)
Eq. (10) and therefore indicate the nonexistence of the nematic boundary conditions, makes it an extremely promising tool to
phase. attack and solve open complex many-body problems spanning
from condensed matter to quantum information. For spin
VI. CONCLUSIONS and bosonic systems, the presented approach can be trivially
extended to include lattice dimensions higher than 1 with
In this manuscript we introduced, motivated, and discussed the same polynomial scaling of the computational cost. For
an algorithm for simulating ground states of quantum many- fermionic systems this extension is less trivial, since it requires
body Hamiltonians on a lattice, based on a gauge-adaptive tree additional tensors that keep track of the parity in order to
tensor network ansatz. We stressed how the manipulation at respect the fermionic anticommutation relations. Nevertheless,
runtime of the tensor network gauge, combined with the loop- as was shown in Refs. [73–76], this task is still realizable
free topology of the tree network, plays an instrumental role in without increasing the scaling of the computational cost.
enhancing the performance of the algorithm, first by reducing
a generalized eigenvalue problem into a simple eigenvalue
problem. We characterized the computational scalings with
ACKNOWLEDGMENTS
bond dimension m and argued how the algorithm speed-
up allows for pushing to higher numerical precision. We We thankfully acknowledge J.I. Cirac for inspiring this
tested the algorithm on 1D quantum models, with various work, feedback, and discussions. The authors acknowledge
boundary conditions. First we benchmarked it against the financial support from the EU through SIQS, the German
exactly solvable Ising model in a transverse field, extracting Research Foundation (DFG) via the SFB/TRR21, the Italian
very precise critical exponents with PBCs. Then we explored MIUR via PRIN Project No. 2010LLKJBX, and computa-
the bilinear-biquadratic spin-1 model in a region known to tional resources provided by the bwUniCluster and bwGRiD
be numerically challenging, being able to confirm an intricate projects in Ulm and by the MOGON Cluster of the ZDV Data
exponential explosion of (nematic) correlation lengths towards Center in Mainz.

[1] F. Verstraete, D. Porras, and J. I. Cirac, Phys. Rev. Lett. 93, [5] S. Rommer and S. Östlund, Phys. Rev. B 55, 2164 (1997).
227205 (2004). [6] U. Schollwöck, Ann. Phys. 326, 96 (2011).
[2] G. Vidal, Phys. Rev. Lett. 91, 147902 (2003). [7] G. De Chiara, M. Rizzi, D. Rossini, and S. Montangero, J.
[3] D. Perez-Garcia, F. Verstraete, M. M. Wolf, and J. I. Cirac, Comput. Theor. Nanosci. 5, 1277 (2008).
Quantum Info. Comput. 7, 401 (2007). [8] S. R. White, Phys. Rev. Lett. 69, 2863 (1992).
[4] S. Östlund and S. Rommer, Phys. Rev. Lett. 75, 3537 [9] V. Murg, F. Verstraete, O. Legeza, and R. M. Noack, Phys. Rev.
(1995). B 82, 205105 (2010).

125154-9
M. GERSTER et al. PHYSICAL REVIEW B 90, 125154 (2014)

[10] K. H. Marti, B. Bauer, M. Reiher, M. Troyer, and F. Verstraete, [43] P. Calabrese and J. Cardy, J. Stat. Mech.: Theory Exp. (2004)
New J. Phys. 12, 103008 (2010). P06002.
[11] G. Vidal, J. I. Latorre, E. Rico, and A. Kitaev, Phys. Rev. Lett. [44] A. Läuchli, G. Schmid, and S. Trebst, Phys. Rev. B 74, 144426
90, 227902 (2003). (2006).
[12] J. I. Latorre, E. Rico, and G. Vidal, Quantum Info. Comput. 4, [45] F. D. M. Haldane, Phys. Lett. A 93, 464 (1983).
48 (2004). [46] F. D. M. Haldane, Phys. Rev. Lett. 50, 1153 (1983).
[13] F. Verstraete and J. I. Cirac, Phys. Rev. B 73, 094423 (2006). [47] M. N. Barber and M. T. Batchelor, Phys. Rev. B 40, 4621 (1989).
[14] J. Eisert, M. Cramer, and M. Plenio, Rev. Mod. Phys. 82, 277 [48] Y. Xian, Phys. Lett. A 183, 437 (1993).
(2010). [49] A. V. Chubukov, J. Phys.: Condens. Matter 2, 1593 (1990).
[15] T. Barthel, M. Kliesch, and J. Eisert, Phys. Rev. Lett. 105, [50] A. V. Chubukov, Phys. Rev. B 43, 3337 (1991).
010502 (2010). [51] J. B. Parkinson, J. Phys. C 20, L1029 (1987); ,21, 3793 (1988).
[16] G. Evenbly and G. Vidal, Phys. Rev. B 89, 235113 (2014). [52] A. Klümper, EPL 9, 815 (1989).
[17] F. Verstraete, M. M. Wolf, D. Perez-Garcia, and J. I. Cirac, Phys. [53] N. Papanicolaou, Nucl. Phys. B 305, 367 (1988).
Rev. Lett. 96, 220601 (2006). [54] G. Fáth and J. Sólyom, Phys. Rev. B 51, 3620 (1995).
[18] J. Molina, JHEP 05 (2013) 024. [55] N. Kawashima, Prog. Theor. Phys. Suppl. 145, 138 (2002).
[19] K. G. Wilson, Rev. Mod. Phys. 47, 773 (1975). [56] B. A. Ivanov and A. K. Kolezhuk, Phys. Rev. B 68, 052401
[20] Y.-Y. Shi, L.-M. Duan, and G. Vidal, Phys. Rev. A 74, 022320 (2003).
(2006). [57] K. Buchta, G. Fáth, Ö. Legeza, and J. Sólyom, Phys. Rev. B 72,
[21] P. Silvi, V. Giovannetti, S. Montangero, M. Rizzi, J. I. Cirac, 054433 (2005).
and R. Fazio, Phys. Rev. A 81, 062335 (2010). [58] M. Rizzi, D. Rossini, G. De Chiara, S. Montangero, and
[22] G. Vidal, Phys. Rev. Lett. 99, 220405 (2007). R. Fazio, Phys. Rev. Lett. 95, 240404 (2005).
[23] M. Rizzi, S. Montangero, and G. Vidal, Phys. Rev. A 77, 052328 [59] D. Porras, F. Verstraete, and J. I. Cirac, Phys. Rev. B 73, 014410
(2008). (2006).
[24] V. Giovannetti, S. Montangero, and R. Fazio, Phys. Rev. Lett. [60] R. Orús, T.-C. Wei, and H.-H. Tu, Phys. Rev. B 84, 064409
101, 180503 (2008). (2011).
[25] G. Evenbly and G. Vidal, J. Stat. Phys. 145, 891 (2011). [61] S. Hu, A. M. Turner, K. Penc, and F. Pollmann, Phys. Rev. Lett.
[26] G. Mussardo, Statistical Field Theory (Oxford University Press, 113, 027202 (2014).
Oxford, UK, 2009). [62] R. Moessner, S. L. Sondhi, and E. Fradkin, Phys. Rev. B 65,
[27] J. Cardy, J. Phys. A 42, 504005 (2009). 024504 (2001).
[28] L. Tagliacozzo, G. Evenbly, and G. Vidal, Phys. Rev. B 80, [63] T. Grover and T. Senthil, Phys. Rev. Lett. 98, 247202 (2007).
235127 (2009). [64] E. S. Sørensen and A. P. Young, Phys. Rev. B 42, 754 (1990).
[29] G. Evenbly and G. Vidal, Phys. Rev. B 79, 144108 (2009). [65] S. Singh, R. N. C. Pfeifer, and G. Vidal, Phys. Rev. A 82, 050301
[30] M. Aguado and G. Vidal, Phys. Rev. Lett. 100, 070404 (2010).
(2008). [66] A. Weichselbaum, Ann. Phys. 327, 2972 (2012).
[31] C. D. Batista and G. Ortiz, Phys. Rev. Lett. 86, 1082 (2001). [67] P. Silvi, arXiv:1205.4198 [quant-ph].
[32] L. Wang, I. Pizorn, and F. Verstraete, Phys. Rev. B 83, 134421 [68] E. Rico, T. Pichler, M. Dalmonte, P. Zoller, and S. Montangero,
(2011). Phys. Rev. Lett. 112, 201601 (2014).
[33] A. J. Ferris and G. Vidal, Phys. Rev. B 85, 165146 (2012). [69] B. Buyens, J. Haegeman, K. Van Acoleyen, H. Verschelde, and
[34] R. A. Horn and C. R. Johnson, Matrix Analysis (Cambridge F. Verstraete, Phys. Rev. Lett. 113, 091601 (2014).
University Press, Cambridge, UK, 2012). [70] P. Silvi, E. Rico, T. Calarco, and S. Montangero,
[35] S. R. White, Phys. Rev. B 72, 180403 (2005). arXiv:1404.7439 [quant-ph].
[36] Z. Bai, J. Demmel, J. Dongarra, A. Ruhe, and H. van der Vorst, [71] L. Tagliacozzo, A. Celi, and M. Lewenstein, arXiv:1405.4811
Templates for the Solution of Algebraic Eigenvalue Problems: [cond-mat.str-el].
A Practical Guide, Vol. 11 (Society for Industrial and Applied [72] J. Haegeman, J. I. Cirac, T. J. Osborne, I. Pizorn, H. Verschelde,
Mathematics, Philadelphia, PA, 2000). and F. Verstraete, Phys. Rev. Lett. 107, 070601 (2011).
[37] A. J. Ferris, Phys. Rev. B 87, 125139 (2013). [73] T. Barthel, C. Pineda, and J. Eisert, Phys. Rev. A 80, 042333
[38] G. Vidal, Phys. Rev. Lett. 101, 110501 (2008). (2009).
[39] E. Lieb, T. Schultz, and D. Mattis, Ann. Phys. (Amsterdam, [74] P. Corboz, G. Evenbly, F. Verstraete, and G. Vidal, Phys. Rev.
Neth.) 16, 407 (1961). A 81, 010303 (2010).
[40] P. Pfeuty, Ann. Phys. (Amsterdam, Neth.) 57, 79 (1970). [75] P. Corboz, R. Orús, B. Bauer, and G. Vidal, Phys. Rev. B 81,
[41] M. A. Cazalilla, J. Phys. B 37, S1 (2004). 165104 (2010).
[42] S. Sachdev, Quantum Phase Transitions (Cambridge University [76] C. V. Kraus, N. Schuch, F. Verstraete, and J. I. Cirac, Phys. Rev.
Press, Cambridge, UK, 2011). A 81, 052338 (2010).

125154-10