Tensor networks for complex quantum systems

Román Orús1, 2
1
Donostia International Physics Center, Paseo Manuel de Lardizabal 4, E-20018 San Sebastián, Spain
2
Ikerbasque Foundation for Science, Maria Diaz de Haro 3, E-48013 Bilbao, Spain
Tensor network states and methods have erupted in recent years. Originally developed in the
context of condensed matter physics and based on renormalization group ideas, tensor networks lived
a revival thanks to quantum information theory and the understanding of entanglement in quantum
many-body systems. Moreover, it has been not-so-long realized that tensor network states play a key
role in other scientific disciplines, such as quantum gravity and artificial intelligence. In this context,
here we provide an overview of basic concepts and key developments in the field. In particular, we
briefly discuss the most important tensor network structures and algorithms, together with a sketch
arXiv:1812.04011v2 [cond-mat.str-el] 22 Jun 2019

on advances related to global and gauge symmetries, fermions, topological order, classification of
phases, entanglement Hamiltonians, AdS/CFT, artificial intelligence, the 2d Hubbard model, 2d
quantum antiferromagnets, conformal field theory, quantum chemistry, disordered systems, and
many-body localization.

I. INTRODUCTION 1d Hamiltonians. White understood that the relevant

degrees of freedom to be kept are, in fact, entanglement
Tensor network (TN) states [1] are representations of degrees of freedom of the wave-function. Later on, it was
complex quantum states in terms of tensor building proven that the state produced by DMRG is a “finitely-
blocks that can sometimes be regarded as fundamental. correlated state”, i.e., essentially what we call today Ma-
This is somehow similar to building up a wavefunction trix Product State (MPS) [5], or equivalently, what some
with LEGO R
pieces. Tensors are the bricks building up mathematicians call Tensor Train [6]. Moreover, it was
the quantum state, and quantum entanglement plays the also understood that DMRG was a variational optimiza-
role of the glue amongst the different pieces. The in- tion algorithm over MPS [7]. The success of DMRG was
dividual tensors codify the key properties of the overall explosive, and the method quickly became the tool of
wavefunction, intuitively similar to how DNA codifies the reference for 1d quantum lattice systems at low energies.
key properties of a human being. This is a naive insight A second wave of results came around the 2000’s.
but also very powerful: not only has it helped us to re- Physicists working on quantum information theory
produce the theoretical properties of complex quantum started to wonder about entanglement in low-energy
systems, but also to develop new numerical simulation eigenstates of Hamiltonians, and its behaviour in quan-
algorithms that go well beyond the limits of what could tum phase transitions. Developments such as the en-
be simulated a few years ago. It helps us in so many ways tanglement entropy of a block for 1d quantum critical
that the field of tensor networks has become very active systems [8] helped to understand that entanglement had
and interdisciplinary. Our purpose with this review is a structure. Moreover, it was also discovered that MPS
to provide an entry-point text on key aspects and recent are particularly well-suited for gapped 1d quantum lat-
developments, pinpointing important references. tice systems with local interactions [9], precisely due to
their entanglement structure. Subsequent developments
generalized this to other scenarios: higher dimensional
systems with Projected Entangled Pair States (PEPS)
A. Some history [12], critical systems with the Multiscale Entanglement
Renormalization Ansatz (MERA) [13], and more.
One does not wake up one day and think all of a sud- From then on, developments on TN methods contin-
den about the TN structure of complex quantum states. ued mostly at the crossover of condensed matter physics
Instead, TNs have evolved progressively over the years. and quantum information. But in addition to this, it
Kramers and Wannier proposed an approximation to the came as a pleasant surprise that TNs were also relevant
2-dimensional (2d) classical Ising model [2] that is con- in other scientific areas. Examples are quantum gravity,
sidered as a precursor of TN variational methods. Later where the MERA has been proposed to be linked to ge-
on, Baxter [3] made use of similar ideas when dealing ometry of space, e.g.. via the Anti-de Sitter / Conformal
with classical partition functions. But it was around the Field Theory correspondence (AdS/CFT) [14], and arti-
90’s when people investigated more seriously how such ficial intelligence, where it has been proven that neural
structures were intimately linked to wave-functions of networks have a TN structure [15]. And more applica-
quantum lattice systems. In particular, a key develop- tions are being found every day, sometimes in the most
ment was that of Density Matrix Renormalization Group unexpected places. Essentially, anywhere that there is
(DMRG) [4] by White. DMRG was constructed as a tech- a structure of correlations, there is also a TN, meaning
nique to keep the relevant degrees of freedom in a renor- that there is room to apply our knowledge on quantum
malization procedure targeting low-energy eigenstates of many-body entanglement.
 2

Ψ = ∑ Ψi i i1 i 2 ! i N (SVD), which is strongly tied to the Schmidt decomposi-

1 2 ...i N
i 's tion of quantum states. This is specified in the Tech-
nical Box. The Schmidt decomposition states that it
is possible to write a bipartite quantum state in terms
of orthonormal correlated basis for the two parties (the
Schmidt basis), together with χ real and positive coeffi-
cients (the Schmidt coefficients). Parameter χ is called
the Schmidt rank. At the level of the coefficients of the
FIG. 1: The coefficient of the quantum state of a many-body wave-function, if orthonormal basis for the A/B vector
system with N sites (say, N spins) can be understood as a ten- spaces are considered (see the box), then the Schmidt de-
sor with exponentially many coefficients in the system’s size. composition is equivalent to the Singular Value Decom-
The structure of this tensor is accounted for by a tensor net- position of the matrix containing the coefficients, and
work, which is a network of tensors interconnected by ancil- Schmidt coefficients correspond to the singular values.
lary indices that take into account the structure and amount As we shall see, the SVD is one of the key tools in TN
of entanglement in the quantum many-body state. This is algorithms. The interested reader is referred to, e.g.,
represented here diagrammatically, where shapes correspond
Ref. [1] for more details.
to tensors, lines to indices, and lines connecting shapes to con-
tracted (summed) common indices. The tensor network on Another important concept is that of canonical form.
the right hand side is intended to be “generic”. More useful Mathematically, the canonical form of a TN is that in
examples of tensor networks are discussed in Fig. 2. which all the bond indices simultaneously correspond to
orthonormal Hilbert spaces. For a TN without loops this
can be achieved by playing with sequential SVDs, see,
B. Some maths e.g., Refs. [1, 40–42] for more details. For TNs with loops,
however, this is formally not possible. On top of provid-
ing a useful representation of the TN, the canonical form
TNs are representations of quantum many-body states
helps a lot in simplifying tensor contractions as well as
based on their local entanglement structure. They arise
in providing truncation schemes in TN algorithms [40].
naturally whenever one has a tensor product structure
(more generally, a tensor category or a braided fusion
category). Take for instance a quantum many-body sys-
II. MAIN TENSOR NETWORK STRUCTURES
tem of N spins-1/2. Any wave-function of the system
can be described, in a computationally inefficient way, in
terms of O(2N ) coefficients once an individual basis for Let us now introduce some of the most important TN
spins is chosen. As such, these coefficients can be un- structures. These are shown diagrammatically in Fig. 2,
derstood as a tensor with N indices, where each index and a comparison of their properties is provided in Table
takes two possible values (say, spin “up” and “down”). I. In what follows we sketch briefly their motivation and
For our purposes, a tensor will be, simply, a multidimen- main properties.
sional array of complex numbers. A plausible motivation
for TNs is that one can replace this tensor by a network (i) Matrix Product States (MPS) [5]: these are 1d ar-
of interconnected tensors with less coefficients, see Fig. 1 rays of tensors, as in Fig. 2(a). Generally speaking, they
for a generic case and Fig. 2 for useful examples. This correspond to low-energy eigenstates of gapped 1d local
construction defines a TN, and it depends on O(poly(N )) Hamiltonians [9]. They also satisfy a 1d area law for
parameters only, assuming that the rank of the intercon- the entanglement entropy of a block (see Table I). Ex-
necting indices is upper-bounded by a constant χ (some- pectation values of local observables can be efficiently
times also called D, as we shall see later [179]). This computed, and have a finite correlation length. Hence,
constant is called “bond dimension”. Similarly, inter- MPS cannot formally represent the entanglement struc-
connecting indices in the network are also called “bond ture of a quantum critical system. However, they are
indices”. In practice they provide the structure of the extremely efficient to manipulate, and over the years
many-body entanglement in the quantum state, and pa- people have used a rich variety of techniques based on
rameter χ turns out to be a quantitative measure of the finite-size scaling and finite-entanglement scaling [10] to
entanglement present in the quantum state. As an ex- use them in order to extract properties of 1d quantum
ample, χ = 1 corresponds to a separable product state critical systems, see Ref. [11] for a recent example.
(i.e., mean field theory), whereas any χ > 1 provides non-
trivial entanglement properties. In addition, it is possible (ii) Projected Entangled Pair States (PEPS) [12]:
to see that TN states satisfy the so-called area-law for the these are 2d arrays of tensors, as in Fig. 2(b) for the par-
entanglement entropy [16], and characterize the relevant ticular case of a square lattice. They are known to cap-
corner of the Hilbert space for a quantum many-body ture the correct correlation structure of low-energy eigen-
systems at low energies. states of 2d local Hamiltonians satisfying the 2d area law
A key mathematical ingredient in numerical algorithms as well as of 2d thermal states [9, 17]. Unlike MPS, PEPS
dealing with TNs is the Singular Value Decomposition can handle critical correlation functions [18], but cannot
 3

be contracted both efficiently and exactly [19]. This is

the reason why people developed approximate methods Schmidt & Singular Value Decompositions
to manipulate them, some of which are sketched later
in this review. They can also handle topological order,
both chiral [20] and non-chiral [21]. Still, it is not fully The Schmidt Decomposition (SD) states that a
understood if they can handle chiral topological order bipartite quantum state |ΨiAB ∈ HA ⊗ HB can
with gapped bulk excitations. always be written as
χ
X
(iii) Tree Tensor Networks (TTN) [22]: these are tree- |ΨiAB = λα |αiA |αiB , (1)
like structures as in Fig. 2(c). By construction, TTNs α=1
have a finite correlation length and an entropy that, on
average, satisfies a 1d area law. They are therefore well with {|αiA }, {|αiB } orthonormal basis (the
suited for gapped 1d systems, though they have also been Schmidt basis) respectively for Hilbert spaces HA
used quite extensively to deal with 1d critical systems and HB , λα > 0 (the Schmidt coefficients), and
[23], as well as 2d systems [24]. By construction, they χ ≤ min(dA , dB ) (the Schmidt rank ) with dA/B
codify a Wilsonian renormalization structure, if the ten- the dimension of HA/B . Proving the SD is easy
sors codify a coarse-graining of the Hilbert space and are using the Singular Value Decomposition P (SVD):
therefore isometries. write the state as |ΨiAB = ij Ψij |iiA |jiB in
some arbitrary basis for A and B. The SVD of
(iv) Multiscale Entanglement Renormalization Ansatz the matrix Ψ of coefficients reads Ψ = U ΛV † ,
(MERA) [13]: these are structures such as those in with U † U = Iχ , V † V = Iχ (i.e., they are isome-
Fig. 2(d), where we show the case of a 1d MERA. They tries), and Λ a χ × χ diagonal matrix with real
are essentially like a TTN of isometries, but including positive entriesΨ(the
= ∑singular
Ψi i ...i i1 ivalues).
2
! iN In terms of
12 N

the so-called “disentanglers”, which are unitary opera- matrix components,i ' s this reads
tors that account for entanglement amongst neighbour- χ χ
ing sites. Thus, MERAs are made from unitaries and
X X
Uiα Λαβ V † Uiα λα V †



Ψij = βj
= αj
,
isometries, and have a number of remarkable properties. α,β=1 α=1
For instance, they can handle the entanglement entropy
of critical 1d systems [25]. Moreover, they are efficiently or in diagrammatic form
contractable. They have an extra holographic dimen-
sion that encodes a renormalization scale, related to the ⇤
= V† ,
<latexit sha1_base64="Nz6aF8F5pGEVg4CvNbJ5BmTvEVI=">AAAB7nicbVBNS8NAFHypX7V+VT16WSyCp5KooMeiFw8eKlhbaEPZbDbt0s0m7L4IJfRHePGgiFd/jzf/jds2B20dWBhm5rHvTZBKYdB1v53Syura+kZ5s7K1vbO7V90/eDRJphlvsUQmuhNQw6VQvIUCJe+kmtM4kLwdjG6mfvuJayMS9YDjlPsxHSgRCUbRSu3enY2GtF+tuXV3BrJMvILUoECzX/3qhQnLYq6QSWpM13NT9HOqUTDJJ5VeZnhK2YgOeNdSRWNu/Hy27oScWCUkUaLtU0hm6u+JnMbGjOPAJmOKQ7PoTcX/vG6G0ZWfC5VmyBWbfxRlkmBCpreTUGjOUI4toUwLuythQ6opQ9tQxZbgLZ68TB7P6t553b2/qDWuizrKcATHcAoeXEIDbqEJLWAwgmd4hTcndV6cd+djHi05xcwh/IHz+QMMT49f</latexit>

so-called Entanglement Renormalization [26]. Finally,

<latexit sha1_base64="2wsinhV7OEj9020G2B+xBypL2+k=">AAAB6HicbVBNS8NAEJ34WetX1aOXxSJ4KokKehGKXjy2YD+gDWWznbRrN5uwuxFK6C/w4kERr/4kb/4bt20O2vpg4PHeDDPzgkRwbVz321lZXVvf2CxsFbd3dvf2SweHTR2nimGDxSJW7YBqFFxiw3AjsJ0opFEgsBWM7qZ+6wmV5rF8MOME/YgOJA85o8ZK9ZteqexW3BnIMvFyUoYctV7pq9uPWRqhNExQrTuemxg/o8pwJnBS7KYaE8pGdIAdSyWNUPvZ7NAJObVKn4SxsiUNmam/JzIaaT2OAtsZUTPUi95U/M/rpCa89jMuk9SgZPNFYSqIicn0a9LnCpkRY0soU9zeStiQKsqMzaZoQ/AWX14mzfOKd1Fx65fl6m0eRwGO4QTOwIMrqMI91KABDBCe4RXenEfnxXl3PuatK04+cwR/4Hz+AI13jMM=</latexit>

U <latexit sha1_base64="tkdqA29D3HUtINdrbBNUFgGzmq0=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lU0GPRi8cWTFtoQ9lsJ+3azSbsboRS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8MBVcG9f9dgpr6xubW8Xt0s7u3v5B+fCoqZNMMfRZIhLVDqlGwSX6hhuB7VQhjUOBrXB0N/NbT6g0T+SDGacYxHQgecQZNVZq+L1yxa26c5BV4uWkAjnqvfJXt5+wLEZpmKBadzw3NcGEKsOZwGmpm2lMKRvRAXYslTRGHUzmh07JmVX6JEqULWnIXP09MaGx1uM4tJ0xNUO97M3E/7xOZqKbYMJlmhmUbLEoygQxCZl9TfpcITNibAllittbCRtSRZmx2ZRsCN7yy6ukeVH1Lqtu46pSu83jKMIJnMI5eHANNbiHOvjAAOEZXuHNeXRenHfnY9FacPKZY/gD5/MHsdeM2w==</latexit> <latexit sha1_base64="M6XjO2YjplBJZ3CsJjpnNJSS4kQ=">AAAB8HicbVBNS8NAEN3Ur1q/qh69LBbBU0lU0GPRi8cK9kPaWDabSbp0dxN2N0IJ/RVePCji1Z/jzX/jts1BWx8MPN6bYWZekHKmjet+O6WV1bX1jfJmZWt7Z3evun/Q1kmmKLRowhPVDYgGziS0DDMcuqkCIgIOnWB0M/U7T6A0S+S9GafgCxJLFjFKjJUe2o/9kMQxqEG15tbdGfAy8QpSQwWag+pXP0xoJkAayonWPc9NjZ8TZRjlMKn0Mw0poSMSQ89SSQRoP58dPMEnVglxlChb0uCZ+nsiJ0LrsQhspyBmqBe9qfif18tMdOXnTKaZAUnni6KMY5Pg6fc4ZAqo4WNLCFXM3orpkChCjc2oYkPwFl9eJu2zunded+8uao3rIo4yOkLH6BR56BI10C1qohaiSKBn9IreHOW8OO/Ox7y15BQzh+gPnM8fqueQUA==</latexit>

MERA is believed to be linked to the AdS/CFT cor-

respondence in quantum gravity [14]. where we used Λαβ = λα δαβ and the fact that
there are at most χ non-zero singular values λα .
(v) Branching MERA (bMERA) [27]: this is shown in Inserting the above equation into |ΨiAB and re-
Fig. 2(e). It is similar to the MERA, with the additional arranging the sums produces the SD in Eq.(1).
fact that at every renormalization scale, the MERA de- The conclusion from this small exercise is that
couples into several copies. In this way one can have an the SD for a quantum state is the same as the
arbitrary scaling of the entanglement entropy of a block, SVD for its coefficients, provided that one has
thus allowing to reproduce the entanglement structure orthonormal basis for the Hilbert spaces HA/B .
of systems that violate the area-law. Physically speak- This has important consequences in approxima-
ing, this structure accounts for the decoupling of degrees tion schemes of TN methods: truncating in sin-
of freedom at different renormalization scales (e.g., spin- gular values is related to truncating in Schmidt
charge separation in electronic solid-state systems). coefficients, either exactly or approximately (as
in the case of non-orthonormal basis for HA/B ).
(vi) Operators and mixed states: operators can be con- And this is linked to truncating in the entangle-
veniently represented by TNs, e.g., by Matrix Product ment of the quantum state, since the entangle-
Operators (MPO) in 1d [28] and Projected Entangled ment entropy S between A and B satisfies
Pair Operators (PEPO) in 2d [29], see Fig. 2(f,g). Mixed
χ
states can also be represented by such structures, and X
S=− λ2α log λ2α ≤ log χ.
also by purification-like structures such as Matrix Prod-
α=1
uct Density Operators (MPDO) in 1d [30], see Fig. 2(h).
The advantage of MPDOs is that they are positive by
construction. However, the amount of correlations that
they can carry for a fixed bond dimension is typically
lower than that of a generic MPO with similar bond di- (vii) Continuous TNs: finally, we would like to men-
mension [31]. tion that the structures above allow for a continuum
 4

MPS 2d PEPS TTN 1d MERA 1d bMERA (b)

S(L) O(1) O(L) O(1) O(log L) O(L) (a)
hOi exact approx. exact exact exact
ξ <∞ ≤∞ <∞ ≤∞ ≤∞
Tensors any any any unit./isom. unit./isom.
Can. form obc, ∞ no yes – – (d)
(c)
TABLE I: Comparison of several properties for some of the
TNs discussed in the main text: entanglement entropy S(L)
of a block of length L (i.e., with L sites in 1d and L×L in 2d),
calculation of a local expectation value hOi, correlation length
ξ, constraints on tensors, and exact canonical form. MERA
and bMERA are built from unitary and isometric tensors, and
we therefore do not consider a canonical form for them. For
MPS, “obc” stands for “open boundary conditions”.
(e)
... ... ... ...
limit, therefore becoming an ansatz for low-energy func- ... ... ... ... ...
tionals of quantum field theories, as well as to opera-
tors on those functionals. For instance, one has continu- ... ... ... ...
ous MPS (cMPS) [32], continuous MERA (cMERA) [33]
and continuous PEPS (cPEPS) [34]. While cMPS and
cMERA have been used for a variety of applications al-
... ...
ready, cPEPS are still to be very much explored.

(g) (h)
III. MAIN ALGORITHMS (f)

Let us now sketch very briefly the main ideas behind

some of the most important numerical algorithms using
TNs. Our purpose here is to explain the basic idea be- FIG. 2: Several tensor networks in diagrammatic notation.
hind key families of numerical TN methods, leaving out (a) Matrix Product State (MPS). (b) Projected Entangled
technical details about implementation. Readers inter- Pair State (PEPS) for a square lattice. (c) Tree Tensor Net-
ested in more information are addressed to the specific work (TTN). (d) Multiscale Entanglement Renormalization
papers explaining the details of each technique. Let us Ansatz (MERA). (e) Branching MERA. (f) Matrix Prod-
also stress that numerical TN methods are also difficult uct Operator (MPO). (g) Projected Entangled Pair Opera-
to classify according to a unique criteria. Here we do it as tor (PEPO). (h) Matrix Product Density Operator (MPDO),
follows: we first introduce methods to obtain 1d states, where tensors in the upper row are the hermitian conjugates
then methods to contract 2d TNs, then methods to ob- (with respect to up/down indices) of the tensors in the lower
row, so that the whole operator is hermitian and positive.
tain 2d states, and finally mixed methods where TNs are
combined with other techniques.
tensor parameters, which can be solved via standard lin-
ear algebra. The method also makes use of other tricks,
A. Methods to obtain 1d states such as the canonical form for open boundary conditions
in order to improve stability and performance [35]. It
can also be adapted to translationally-invariant systems
(i) Density Matrix Renormalization Group (DMRG) in the thermodynamic limit, the so-called infinite-DMRG
[4]: by far the most famous TN method. DMRG was (iDMRG) [36]. There are also extensions of the method
originally proposed as a renormalization procedure over to deal with periodic boundary conditions [37] as well as
the ground-state wave-function of a 1d quantum lattice low-energy excitations [38, 39]. For open boundary con-
system. In the modern formulation, it is a variational ditions and infinite-size systems the computational cost
optimization algorithm over the family of MPS [35]. In is O(χ3 ), with χ the MPS bond dimension. For peri-
particular, one sweeps back and forth over the different odic boundary conditions, if no further approximations
tensors of an MPS, with tensor coefficients treated as are introduced the computational cost is O(χ5 ).
variational parameters in order to minimize the expecta-
tion value of a given Hamiltonian. At every step, the op- (ii) Time-Evolving Block Decimation (TEBD) [40]:
timization can be written as a quadratic problem for the this method, introduced by Vidal in 2004, allows to com-
 5

pute the time evolution of an MPS. If the evolution is in to the variational optimization of the expectation value
real-time, then it allows to compute dynamics as long as of a Hamiltonian, the so-called Variational Uniform MPS
the state remains slightly entangled. If it is in imaginary- (VUMPS) algorithm [47]. The asymptotic computational
time, then it allows to approximate ground states and is a cost of this family of methods for 1d systems is similar
plausible alternative to DMRG. TEBD has also been ap- to the one of other MPS approaches, i.e., O(χ3 ) for an
plied successfully to infinite-size systems (infinite-TEBD) infinite-size MPS of bond dimension χ.
[41]. The asymptotic scaling of the computational cost
is similar to the one of DMRG, i.e., O(χ3 ) for an MPS (vi) Other: people have used other strategies to de-
of bond dimension χ with open boundary conditions. velop new algorithms focusing on the properties of MPS.
TEBD is based on the canonical form of an MPS [42] Examples are alternative methods to simulate time evo-
and the truncation of the MPS bond dimension via SVD, lution in 1d [48], and to simulate open quantum dynam-
which provides an optimal local approximation between ics, including non-Markovian processes [49] and effective
MPS formally justified only for TNs without loops. This small reservoirs [50].
truncation scheme is similar to the one in the so-called
“2-site DMRG”, where tensors for two sites (and not just
one) are optimized by a variational update following an B. Methods to contract 2d tensor networks
SVD.
Here we consider methods to contract a TN without
(iii) Tree Tensor Networks (TTN): in addition to open indices, focusing mostly on methods to compute ef-
MPS, TTNs are also a useful tool to study 1d sys- fective environments from 2d PEPS. This is a key step in
tems. Concerning this, the procedures used in DMRG the simulation of 2d quantum lattice systems. The same
and TEBD have been extended also to TTNs to study methods can also be considered for 2d classical partition
1d gapped and critical systems [22, 23]. This is possible functions, and with some modifications they can also be
given the absence of loops in the TTN. The computa- used for 3d classical and quantum systems. In fact,
tional cost depends on the specifics of the tree, but typ- many of these methods were originally developed in the
ically it can be boiled down to O(χ4 ), with χ the bond context of 2d classical statistical mechanics, see for in-
dimension of the TTN [22]. stance the 1982 book by Baxter in Ref. [3].

(iv) 1d MERA: the MERA [13] can also be used as a (i) Boundary MPS: this was the first method used to
variational ansatz to approximate ground-state proper- approximate the environment of a site on a 2d PEPS,
ties. For the case of 1d systems, this has been done both both for finite [12] and infinite [51] systems. The idea is
for gapped [13] and critical [25] systems. The variational to contract the full 2d lattice starting from a boundary
optimization is however tricker than for, e.g., DMRG, (either an actual boundary for finite systems, or placing
because of the constraints on the tensors (they must be it manually at infinity for systems in the thermodynamic
unitaries and isometries). Such optimization can be done limit), with boundary tensors forming an MPS. The algo-
by using a number of techniques, as explained in detail rithm proceeds by contracting 1d MPOs made of lattice
in Ref. [43]. Overall, the computational cost depends on tensors into the MPS, mimicking a non-unitary time evo-
the actual structure of the MERA. For instance, for the lution that can be treated, e.g., with the techniques from
so-called binary MERA the cost is O(χ9 ), whereas for Ref. [42]. For a PEPS of bond dimension D and a bound-
the so-called ternary MERA the cost is O(χ8 ), and for ary MPS of bond dimension χ, the computational cost
the so-called modified-binary MERA it is O(χ7 ) [25]. of this scheme for a 2d square lattice is O(χ3 D6 + χ2 D8 )
for a horizontal/vertical boundary, and O(χ3 D4 + χ2 D6 )
(v) Tangent Space Methods [44]: a whole new family for a diagonal boundary [1, 51].
of methods is based on the idea that MPS can be under-
stood as a manifold, in the sense of differential geome- (ii) Corner Transfer Matrices (CTM): this is a pop-
try. With this in mind, it is possible to obtain a variety ular approach nowadays because it is relatively easy to
of highly-efficient and accurate methods using the con- implement and produces good-quality results. Focusing
cept of tangent space of the MPS manifold. A remark- on a 2d square lattice of tensors, the idea is to find renor-
able example is the Time-Dependent Variational Prin- malized (coarse-grained) approximations to the tensors
ciple (TDVP) algorithm [45] for MPS. This algorithm amounting to the contraction of all the tensors on the cor-
uses concepts of differential geometry to compute the ners. Such tensors are the CTMs, which are well-known
time evolution of an MPS either in real or imaginary- objects in the context of exactly-solvable statistical me-
time, without the need of a Trotter decomposition (as in chanical models. There are several schemes for dealing
TEBD) and preserving naturally all the spatial symme- and finding such CTMs [3, 52–54]. The computational
tries of the physical system. The idea also allows for the cost depends on the specifics of the implementation. To
computation of low-energy excitations and energy bands have an idea, the method from Ref. [53] has a cost of
by means of an ansatz with well-defined momentum [46]. O(χ3 D6 ) for a PEPS of bond dimension D and a CTM
The formalism of the tangent space can also be applied of bond dimension χ.
 6

(iii) Tensor coarse-graining: in this approach the main of the cylinder or the width of the stripe. This approach
idea is to coarse-grain the network, by finding new renor- has been very successful in determining properties of the
malized tensors amounting for a “zoom-out” and which 2d t−J model [65] as well as the spin liquid nature of the
take into account the main features of the TN at long dis- ground state of the Kagome Heisenberg Antiferromagnet
tances. This is similar to a Kadanoff blocking for clas- (KHAF) [66]. An interesting evolution of 2d DMRG is
sical statistical mechanical models. Many schemes fall to combine position and momentum basis for both di-
into this category, each one with its own advantages and rections, showing better performance in some situations
drawbacks: Tensor Renormalization Group (TRG) [55], [67]. Still, the computational cost of 2d DMRG is even-
Second Renormalization Group (SRG) [56], Higher Or- tually doomed for large 2d systems due to an exponential
der TRG (HOTRG), Higher Order SRG (HOSRG) [57], entanglement wall in the transverse direction that cannot
Tensor Entanglement-Filtering Renormalization (TEFR) be handled by an MPS with finite bond dimension.
[58], Tensor Network Renormalization (TNR) [59], loop-
TNR [60] and TNR+ [61]. The specifics of the implemen- (ii) 2d TTNs: TTNs have also been used to study 2d
tation for each one of these cases is different. However, systems, since their structure can be easily adapted for
the idea is that one typically contracts tensors, defining such geometries [24]. Still, and as in the 2d DMRG case,
a new lattice in terms of some new tensors that amount the underlying TN structure has inherently 1d built-
for the contraction. The new tensors are then renormal- in correlations because of the absence of loops in the
ized by truncating their bond indices with some isome- network, which again implies exponential entanglement
tries. The prescription for how to find such isometries is walls and therefore only relatively good accuracy depend-
what defines the different schemes. Some schemes (such ing on the system and regime. Still, they can be pretty
as TNR) also remove local entanglement before blocking useful since their associated algorithms are quite efficient
the tensors. The computational cost of each scheme can (as in 1d). For instance, they have been used to study
also be very different depending on the implementation. confinement / deconfinement transitions of 2d Z2 lattice
As an example, for a classical partition function with a gauge theories [68].
TN structure of bond dimension χ on a square lattice,
the SRG method has a cost of O(χ10 ) [62]. (iii) PEPS: numerical algorithms based on PEPS are
well-suited to tackle 2d systems. A PEPS has inherently-
(iv) Nested tensor network [63]: this approach is not a built truly 2d correlations and, as such, is a natural
contraction scheme in itself (as the previous cases), but ansatz to study a wide variety of 2d systems. People have
rather the idea of projecting the tensors of a 3d TN on a considered the ansatz for finite-size systems, the so-called
2d plane, so that the resulting 2d TN can be contracted finite-PEPS [12, 69], but also for infinite systems, the so-
by any of the three strategies (i) – (iii) above. An exam- called infinite-PEPS (iPEPS) [51]. There are different
ple is the TN for the norm of a 2d PEPS with bond di- ways of optimizing the tensors of a PEPS to obtain ap-
mension D, where bra and ket tensors are shifted with re- proximations of ground states. For instance, variational
spect to each other, thus producing a new 2d TN similar updates (both for finite [69] and infinite [70] PEPS), as
to a 2d partition function with bond dimension D. Such well as imaginary-time evolution via simple [71], full [51]
a 2d TN can thus be contracted more efficiently than the and fast-full [72] updates. Simple updates are very ef-
one obtained with a double-layer approach, which has ficient but not necessarily accurate, while full and fast
bond dimension D2 . full updates take into account the effect of the environ-
ment when optimizing a PEPS tensor and are therefore
slower but more accurate. Such environments are usu-
C. Methods to obtain 2d states ally computed via the renormalization methods to con-
tract 2d TNs explained in the previous section, i.e., using
Let us now focus on techniques to obtain TN states boundary MPS, CTMs, and tensor coarse-graining. Fur-
for 2d quantum lattice systems. As we shall see, in some thermore, recently it was also shown how to compute
cases the techniques explained before for 1d systems and excited states with PEPS for 2d systems [73], and how
2d TN contractions will turn out to be fundamental. to do accurate extrapolations in the bond dimension [74].
The computational cost of PEPS algorithms heavily de-
(i) 2d DMRG: people have considered using DMRG pends on the type of algorithm chosen. For instance, for
also to study 2d systems [64]. Even if, by construction, a square lattice, simple-update algorithms with a mean-
DMRG produces an MPS and is therefore a priory better- field environment [75] have a cost of O(D5 ), whereas full
suited to deal with gapped 1d local Hamiltonians, the and fast full updates have a cost of O(χ3 D6 ) (and with a
technique has also seen a lot of success in 2d because prefactor that could be large), with χ and D respectively
of its efficiency. The idea behind 2d DMRG is to have a the environment and PEPS bond dimensions.
stripe, or wrap the 2d system around a cylinder, and then
use MPS as an ansatz for the 2d lattice following a snake (iv) 2d MERA: The MERA has also been used as a
pattern. The true 2d properties of the system are recov- variational ansatz to approximate ground states of 2d
ered by doing careful finite-size scaling with the thickness quantum lattice systems, see for instance Refs. [76–78].
 7

The computational cost of the approach strongly depends (v) More synergies: while the most relevant “com-
on the type of lattice as well as the specific choice of bined” methods are sketched above, there have been
unitaries and disentanglers. For instance, the approach in further developments in other directions which are also
Ref. [76] for an infinite square lattice has a computational worth mentioning. For instance, the so-called Entangle-
cost of O(χ16 ), with χ the MERA bond dimension. ment Continuous Unitary Transformations (eCUT) [86]
showed how to mix the idea of continuous unitary trans-
formations and Wegner’s flow [87] with TNs, by truncat-
D. Combined methods ing the flow equation in its operator-entanglement con-
tent. Moreover, TNs have also been used in combination
There have also been developments where TNs have with perturbation theory, e.g., in Ref. [88] it was shown
been combined with other existing methods, or where how to construct an exact 2d PEPS up to a given order
TNs have been useful to understand other existing tech- in perturbation theory. Another fruitful combination has
niques. Some of these developments are briefly sketched been that of TNs and the randomized SVD [89], which is
in this section. proven to improve the efficiency of numerical algorithms
such as TBED, DMRG and TRG [90]. TN states have
(i) Monte Carlo TNs: Monte Carlo methods have been also been used to develop generalized Lanczos methods
used together with TN techniques in several ways. For [91]. Finally, TN states were also useful to understand
instance, people have implemented Monte Carlo sam- the mathematical structure of exactly-solvable systems,
pling to do variational optimizations over TNs, as well as e.g., the algebraic Bethe ansatz [92], the fermionic Fourier
approximate calculations of effective environments [79]. transform [93], the XY spin chain [94], and Kitaev’s hon-
In the context of string-bond and plaquette-entangled eycomb model [95].
states, Monte Carlo has been used also as a sampling
technique for optimization and expectation value cal-
culation [80]. The combination with Monte Carlo al- IV. THE ROLE OF SYMMETRIES
lows in principle to reach higher bond dimension in the
calculations, at the cost of the length of the sampling. In this section we briefly overview the effect of symme-
For instance, the cost of these methods when combined tries in TNs. We sketch why having a symmetry implies
with MPS for periodic boundary conditions is typically a tensor factorization, and its important consequences.
O(N χ3 ), with χ the MPS bond dimension and N the We also comment some basic aspects of fermionic ten-
length of the (finite) 1d system, and with a prefactor sor networks as a special case, as well as gauge symme-
that depends on the number of samples. tries, topological order, and the classification of quantum
phases of matter.
(ii) TNs for Density Functional Theory (DFT): DFT
is one of the most popular numerical approaches to
perform ab-initio calculations of real materials and A. Global symmetries
molecules. In this context, TNs (and specially MPS)
were used to produce systematic approximations to the The implementation of global symmetries in TN al-
exchange-correlation potential of electronic systems [81]. gorithms has been considered (at least) since the early
years of DMRG, and it is well-known that it can lead
(iii) TNs for Dynamical Mean-Field Theory (DMFT): to important computational advantages. People have ex-
TNs have also found important applications in DMFT. ploited this fact specially in DMRG [96], the 1d MERA
In particular, in a series of works [82] it has been shown [97] and 2d PEPS [98], both for abelian symmetries such
that MPS techniques can be used as a high-accuracy and as U (1) particle-conservation, but also for non-abelian
low-cost impurity solver, including applications to non- symmetries such as SU (2) rotation-invariance. For fur-
equilibrium systems. ther details we refer the reader to, e.g., Refs.[97, 99, 100],
which provide in-depth overviews. In addition to numer-
(iv) TNs and wavelets: recently, a number of connec- ical advantages, an important theoretical consequence is
tions have also been stablished between wavelet transfor- that the so-called spin networks appear naturally from
mations and TNs. More specifically, in Ref. [83] it was TN states with symmetries. Spin networks are used, e.g.,
shown how Daubechies wavelets could be used to build an in loop quantum gravity to describe quantum states of
analytic approximation to the ground state of the 1d crit- space at a certain point in time [101]. In this sense, TNs
ical Ising model which, in turn, correspond to instances with symmetries turn out to offer not only efficient nu-
of the 1d MERA. Additionally, Ref. [84] showed how the merical methods for complex quantum systems, but also
structure of wavelet transformations adapts to that of a a very intriguing connection between quantum entangle-
quantum circuit, and Ref. [85] showed how the ground ment and gravity.
state of some fermionic systems could be understood via When dealing with local symmetries in TN states,
entanglement renormalization also using the language of Schur’s lemma [102] ultimately implies that symmetric
wavelets. tensors can be decomposed in two pieces: one completely
 8

determined by the symmetry which acts on the subspaces (a)

of irreducible representations, and another that contains k (c, γ c ) (c, oc )
the actual degrees of freedom of the tensor, and which
acts on a degeneracy subspace. As an example, for a O = P Q
symmetric tensor Oijk with three indices (i, j, k), the
Wigner-Eckart theorem implies that it can be decom-
posed as Oijk = (P abc )αa βb γc (Qabc )ma nb oc , where index i i j
(a, α a ) (b, βb ) (a, ma ) (b, nb )
decomposes as i ≡ (a, αa , ma ) (and similarly for the rest), €
P abc is a degeneracy tensor that contains all the degrees (b)
of freedom of Oijk , and the structural tensor Qabc is com-
pletely fixed by the symmetry group. This is represented
in the TN of Fig. 3(a). In practice, this means that the
symmetry constraint heavily reduces the degrees of free-
=∑
dom of the tensor. As a consequence, the whole TN also
factorizes in two pieces: a TN of degeneracy tensors, and (c)
a TN of structural tensors. The second of the pieces is a
€
i1 i2 i1 i2
spin network, see Fig. 3(b). Thus, a symmetric TN for a X X i2 i1 j1 j 2 = δi1 j1 δi2 j 2 S(P(i1),P(i2 ))
#−1 if P(i1) = P(i2 ) = −1
quantum state |Ψi of N sites is a superposition of expo- S(P(i1 ),P(i2 )) = $
nentially many spin networks with N open indices, with i2 i1 j1 j2 %+1 otherwise
€ € €
coefficients determined by TNs of degeneracy tensors. € €
There are many reasons why it is a good idea to im- FIG. 3: (a)€ A 3-index symmetric
€ tensor O decomposes into
plement symmetries in TN algorithms whenever this is € a €degeneracy
€ tensor€ P and an intertwiner of the symmetry
possible. For instance, they allow to simulate systems group Q (which, e.g., for SU (2) is a Clebsch-Gordan coeffi-
with specific quantum numbers. But most importantly, cient). (b) A symmetric TN decomposes as the sum of spin
symmetries also allow for a compact description and ma- networks, with coefficients given by TNs of degeneracy ten-
nipulation of the TN, thus leading to more efficient algo- sors, and the sum being done over the degrees of freedom of
internal indices. (c) In fermionic TNs, crossings between lines
rithms that can reach larger bond dimension.
in diagrammatic notation are replaced by fermionic SWAP
operators X. These operators take into account that the an-
ticonmutation relation of fermionic operators in second quan-
B. Fermionic systems tization.

TN methods can also be used to simulate fermionic

systems in any dimension, and directly in second quan-
tization language. Fermionic statistics can be imple- ferent due to the local (gauge) nature of the symmetry
mented in TN algorithms in different but equivalent [108, 109]. Many works have implemented the formalism
ways [78, 103–107]. The graphical implementation from of gauge-invariant TNs, focusing mostly on 1d systems
Ref. [103] is perhaps the simplest one to describe. This and sometimes in 2d. In particular, Z2 lattice gauge the-
is based on two “fermionization” rules: (i) use parity- ories (LGT) in (1+1)d have been considered with DMRG
symmetric tensors, and (ii) replace crossings in the pla- [110]. For the Schwinger model, i.e., quantum electro-
nar representation of the TN by fermionic SWAP gates. dynamics (QED) in (1+1) dimensions, DMRG (without
The first rule is justified because fermionic parity, i.e., MPS formulation) was considered in several works [111],
whether the total number of fermions is even or odd, is whereas MPS simulations have been done to compute
a good Z2 symmetry for fermionic systems. The sec- the chiral condensate [112] as well as thermal properties
ond rule is also justified because fermionic SWAP gates, [113], the mass spectrum [114], the Schwinger effect [115],
defined as in Fig. 3(c), take correctly into account the the effect of truncation in the gauge variable [116], the
anticommutation of fermionic operators in second quan- case of several fermionic flavours [117], and the thermo-
tization. In the end, fermoinic TN algorithms can be dynamic limit of DMRG and its possible extrapolation
programmed in the same way as their bosonic counter- to (2+1) dimensions [118]. The consequences of gauge
parts, but taking into account these two simple rules, symmetry in the MPS of the Schwinger model was first
which only imply a subleading increase in the computa- elaborated in Ref. [109]. A gauge-invariant MPS ansatz
tional cost of the algorithm. was also used to compute the confining potential [119]
as well as the scattering of two quasiparticles [120]. TN
simulations have also been implemented recently for non-
C. Gauge symmetries abelian lattice gauge theories in (1+1)d [121]. For higher-
dimensional systems, gauge-invariant TN ansatzs have
Gauge symmetries can also be implemented naturally also been proposed analytically [108, 122, 123], which
in the framework of TNs, leading to a similar TN de- can be used as variational wave-functions to study lat-
composition as for global symmetries, but slightly dif- tice gauge theories in (2+1)d.
 9

D. Topological order and classification of phases A B

TN states are the natural language for topologically-
ordered systems, which can in turn be understood using
gauge symmetries. There are several developments along
this direction, both from the perspectives of analytics and
numerics. From the analytics point of view, it has been
proven that eigenstates of string-net models [124] admit
an exact tensor network description [21, 125], where ten-
sors have specific gauge symmetries. This means that
all non-chiral topologically-ordered 2d phases on a lat- FIG. 4: 2d PEPS on an cylinder. To compute the entangle-
tice admit a TN description. Moreover, it has also been ment Hamiltonian HE , one makes a bipartition of the system
shown that certain PEPS for fermionic systems as well as (A vs B), and computes the reduced density matrix ρ of one
for spin systems can handle chiral topological order, al- of the subsystems using standard contraction techniques for
beit the corresponding parent Hamiltonians (which can PEPS (e.g., boundary-MPS with periodic boundary condi-
tions). The eigenvalues of ρ define then the eigenvalues of
be obtained directly from the PEPS tensors) have ei-
HE , which can be grouped according to quantum numbers
ther gapless bulk excitations or long-range interactions such as the vertical momentum (since there is translation in-
[20]. It is therefore a theoretical challenge to under- variance if we rotate the cylinder around its axis).
stand the correct TN framework to describe chiral topo-
logical states with gapped bulk and short-range parent
Hamiltonians. From the numerical perspective, TN al-
gorithms have been used to compute phase diagrams
of topological systems under perturbations, e.g., using
TTNs [68] as well as 2d PEPS [126] and DMRG on cylin-
ders [127]. In addition to these developments, TNs have Such reduced density matrix can be written as ρ ∝ e−HE ,
also proven to be an extremely useful tool in computing with HE the entanglement Hamiltonian. The claim is
symmetry-protected topological order in 1d [129], repre- that HE describes the fundamental degrees of freedom
senting fractional quantum Hall states [128], simulating of the projection of the quantum state on a boundary.
anyonic systems [130], and describing theoretically topo- TNs turn out to be the natural arena to investigate such
logical quantum computation [131] as well as symmetry- entanglement Hamiltonians for a variety of systems, and
enriched topological order [132]. A separate set of re- very specially for 2d PEPS [141]. The main idea in such
sults, but related to the developments on topological or- derivations is to wrap the 2d PEPS around a cylinder,
der, has concerned the theoretical classification of quan- and then study the entanglement spectra (i.e., the eigen-
tum phases of matter [133], for which TNs have also been values of HE ) of half of the cylinder versus the other half,
useful. For instance, MPS and PEPS were used to clas- see Fig. 4. Such eigenvalues can be grouped in terms of
sify phases of 1d and 2d quantum spin systems [134–136]. their (vertical) momentum quantum number, and encode
The classification of fermionic topological phases has also very useful information about the PEPS, such as possible
been studied using fermionic MPOs and fermionic PEPS gapless edge states and chiral topological order [20].
[137–139].

From the studies so far, there is an interesting cor-

V. HOLOGRAPHY
respondence. Consider PEPS that are ground states of
2d Hamiltonians with local interactions. It looks like, if
Several notions related to holography also play a key the 2d system is gapped and not topologically ordered,
role in TN states. Here we sketch briefly how TNs provide HE is usually a 1d Hamiltonian with short-range inter-
a natural bulk-boundary correspondence via the so-called actions. However, for 2d critical systems, then HE is
entanglement Hamiltonians. Moreover, we will also com- a 1d Hamiltonian with long-range interactions. Also, if
ment (very descriptively) on the connection to the ge- the 2d system is gapped and topologically ordered, then
ometry of space-time in quantum gravity and the idea HE is essentially a projector. All this is very interest-
that space-time may emerge from quantum many-body ing, because it justifies the numerical observation that
entanglement. environment calculations in, e.g., infinite 2d PEPS con-
verge quickly with very few iterations of a boundary MPS
[9]. It seems, therefore, that 2d PEPS for ground states
A. Entanglement Hamiltonians of gapped 2d systems with no topological order can be
contracted efficiently with good accuracy, even if this is
Li and Haldane [140] pointed out that the eigenvalues not the case for a generic PEPS [19] or even on average
of the reduced density matrix of a bipartition codify im- [142]. Some steps towards turning this observation into
portant information about the boundaries of the system. a mathematical theorem have already been taken [143].
 10

B. Emergent geometry

In the framework of holography, there is a very sugges-

tive connection between entanglement, TNs, and quan-
tum gravity: it looks like the MERA is a lattice real-
ization of a space with some geometry, where curva-
ture is somehow linked to entanglement. The observation
implies that space-time geometry may emerge from the
underlying structure of entanglement in complex quan-
tum states. An instance that has been studied in some
detail is the possible relation between MERA and the
AdS/CFT or gauge / gravity duality [144]. This connec-
tion between TNs and quantum gravity was originally
noticed by Swingle [14], and later on investigated by sev-
eral authors [145–147, 151]. More specifically, for a scale-
invariant MERA, the tensors in the bulk can be under- FIG. 5: The entanglement entropy of a block of length L for
stood as a discretized AdS geometry, whereas the indices a 1d MERA is upper bounded as SL ≤ log χ × ∂ΩL , with
∂ΩL the boundary of region ΩL in the TN (i.e., the number
at the boundary correspond to the local Hilbert spaces
of links crossed by the blue line) and χ the bond dimension.
obtained after a discretization of a CFT, see Fig. 5. The One quickly realizes that ∂ΩL = O(log L), and therefore one
connection can be made more formal by taking the con- has that SL = O(log L) for the 1d MERA. This calculation
tinuum MERA [33] and evaluating the metric of the re- matches the behavior from CFTs in (1+1)d, and corresponds,
sulting smooth space in the bulk, with the curvature of precisely, to the lattice version of the Ryu-Takatanagi pre-
the geometry being linked to the density of disentanglers scription to compute the entanglement entropy in AdS/CFT
[146, 147]. As of today, the connection is very intriguing [150]. Entanglement is thus the area in holographic space of
and has motivated a lot of research, specially from the the minimal surface separating the two regions. This is one
string theory community. In particular, there have also of the key observations that motivates the analogy between
been claims that MERA does not actually correspond to MERA and AdS/CFT.
and AdS geometry, but rather to a de-Sitter (dS) geome-
try [148]. In a recent work, however, Milstead and Vidal
showed that MERA is in fact neither AdS nor dS, but neural networks in new ways. One must however be
rather a lightcone geometry [149]. In any case, and even careful, since in general neural networks are characterized
if this connection is certainly suggestive and remarkable, by nonlinear functions, whereas TNs are linear and there-
the role played by TNs in the quantization of gravity is fore obey the superposition principle. In Ref. [152] it was
still unclear. also shown that there is an equivalence between restricted
Boltzmann machines (a simple type of neural network)
and TN states. In Ref. [153] Boltzmann machines were
VI. ARTIFICIAL INTELLIGENCE also shown to be connected to some classes of TN states
in arbitrary dimensions. In addition to this, in Refs.[154–
In this section we will comment on the recent obser- 157] it has been shown how MPS and TTNs can be used
vation that neural networks (such as those used in deep for supervised and unsupervised learning tasks of classi-
learning) are in fact particular cases of TNs, as well as fying images. Finally, in Ref. [158] it has been discussed
on the use of MPS to improve some methods of artificial how quantum circuits based on MPS and TTNs could be
intelligence. Additionally, we will also sketch the result used to implement machine-learning tasks in near-term
that syntactic relations in language have a TN structure quantum devices, and in Ref. [159] it has been explored
that is inherited by probabilistic language models. how probabilistic graphical models motivate the concept
of “generalized TN”, where information from a tensor
can be copied and reused in other parts of the network,
A. Machine learning thus allowing for new types of variational wave-functions.

Several promising connections between TNs and ma-

chine learning have been put forward recently. In B. Language models
Ref. [15] it was shown that deep learning architectures
can actually be understood using the language of quan- From the perspective of computational linguistics, it
tum entanglement. To name a couple of examples, con- has also been discovered recently that probabilistic lan-
volutional networks correspond to specific cases of TTNs, guage models used for speech and text recognition have
and recurrent neural networks correspond to MPS. More actually a TN structure. This is a consequence of the
generically, the whole machinery of quantum information fact that Chomsky’s MERGE operation can be under-
and entanglement theory can be applied to understand stood as a physical coarse-graining of information [160].
 11

Such probabilistic models usually have the form of a C. Conformal field theory
TTN or even an MPS, i.e., loop-free TNs. In turn, this
also matches the empirical observation that convolutional We discussed before the role of the MERA TN in the
neural networks are quite good at language processing. AdS/CFT correspondence. However, we would like to
In connection with the results commented in the previ- stress a number of recent results where TNs directly tar-
ous section, it is clear that this is indeed so because such get properties of CFTs, not necessarily with the holo-
neural networks are TTNs, which encode the RG struc- graphic duality in mind. In this respect, and always in
ture of language found in Ref. [160], and are therefore the context of (1+1)d CFTs, Ref. [165] showed how to
naturally well-suited for this task. do the quotient of a MERA representation of the vacuum
taking it to a thermal state. Ref. [166] showed how to
describe and coarse-grain the partition function of the 2d
critical Ising model in the presence of topological confor-
VII. FURTHER TOPICS mal defects. Moreover, Ref. [167] studied how space-time
symmetries are reflected in the cMERA of a free boson
There are many other interesting results that have to CFT, Ref. [168] explored the TN description of conformal
do with TN states and methods, and which were not dis- transformations, and Ref. [169] proposed the interpreta-
cussed so far. For obvious reasons we cannot summarize tion of such TNs as path integrals on curved spacetime.
all of them here. Nevertheless, here we sketch a few of In addition to this, and in connection to topological or-
them which we believe are worth pointing out. der, Ref. [170] investigated the mapping of topological
quantum field theories to CFTs using 2d TNs. This fam-
ily of results establishes a valuable dictionary between
TNs and CFTs.
A. 2d Hubbard model

D. Quantum chemistry
The Hubbard model tries to capture the dynamics of
electrons hopping on a lattice. In 2d, it is believed to be
related to high-temperature superconductivity, but un- TN numerical methods have also found important ap-
fortunately its phase diagram has not been conclusively plications in quantum chemistry. For instance, DMRG
determined yet, even in the single-band approximation. has been used in this context already since some time
In this context, it is worth mentioning that the best vari- ago [171]. The reordering of the fermionic orbitals has
ational ground state energies computed so far for this also been considered with MPS simulations [172]. For a
model in the strongly correlated regime have been with discussion on recent TN approaches to quantum chem-
TNs, specifically with the iPEPS algorithm for fermions istry, see also Ref. [173].
[74]. Other related simpler fermionic models, such as the
t − J model, have also been simulated successfully us-
ing a variety of TN techniques, including 2d DMRG and E. Disorder and many-body localization
iPEPS [65, 161].
Several TN methods have been proposed to deal with
disordered systems [174]. Moreover, many-body local-
ized (MBL) phases have also been studied using tailored
B. 2d quantum antiferromagnetism numerical TN methods. For instance, in Ref. [175] a
spectral tensor network was explored to represent all the
The antiferromagetic Heisenberg model on the Kagome spectrum of energy eigenstates. In Ref. [176] a varia-
lattice is the archetypical example of a frustrated mag- tional method over unitary MPOs was proposed to diag-
net. Its ground state has remained elusive until recently, onalize MBL Hamiltonians. An alternative TN encoding
where 2d DMRG simulations unveiled that it is a quan- of all eigenstates of MBL systems in 1d was proposed in
tum spin liquid [66]. Other TN methods have also at- Ref. [177]. Finally, TNs were used to prove the robust-
tacked this problem, including 2d MERA with a specific ness of MBL phases with SPT order [178].
disentangling structure [77], Projected Entangled Sim-
plex States (PESS) [162], and iPEPS on coarse-grained
lattices [163]. Some of these simulations are compatible VIII. OUTLOOK
with a gapless quantum spin liquid ground state, but so
far could not produce better energies than those obtained This paper is an overview of developments around TN
with 2d DMRG. Other models of quantum antiferromag- states and methods along different directions. We try
netism have also been studied via TNs. For example, to collect valuable information, including basic notions
iPEPS have been used extensively in the study of SU (N ) and bits of the overall current perspective. It is of course
magnets [164]. impossible to cover all the developments, but hopefully
 12

we were able to summarize some of the key ones. Im- • Gauge symmetry: an operation that leaves invari-
portantly, notice that applications go well beyond quan- ant the system, and which has the freedom to act
tum science. It is thus important to keep in mind that, differently at every point in the system.
in such cross-over applications (e.g., to artificial intelli-
gence), typical properties of quantum mechanics are lost, • Topological order: a type of order in quantum mat-
such as unitarity. In such contexts one must therefore ter entirely due to global entanglement properties,
keep in mind that the TNs codify not a entanglement and which does not exist classically. Other charac-
structure of a complex system, but rather the structure terizations: excitations are anyonic, the topological
of its generalized correlations, shall these be quantum or entanglement entropy is non-zero, ground states are
not. topologically degenerate, reduced density matrices
Our vision is that TNs will continue finding numeri- are locally equivalent...
cal and theoretical applications, both along established
research directions, but also along new ones yet to be dis- • Chiral system: a system that breaks time-reversal
covered. The lesson we learnt over the years is that wher- symmetry.
ever one finds correlations, there may be a TN behind. • Renormalization: the process of removing degrees
And sometimes this leads to unexpected connections and of freedom that are not relevant to describe a com-
pleasant surprises, pushing forward the boundaries of sci- plex system at different scales of some physical vari-
ence. able (energy, length...).

• AdS space: a geometric space with negative curva-

Acknowledgements.- I acknowledge financial sup-
ture.
port from Ikerbasque, DIPC and DFG, as well as dis-
cussions over the years with many people on several of • CFT: a quantum field theory with conformal sym-
the topics presented here. I also acknowledge M. Rizzi metry, which includes scale invariance. Low-energy
and P. Schmoll for a critical reading, and Iulia Georgescu field theories of quantum critical systems are usu-
for suggesting me to write this paper. ally CFTs.

Glossary.- • Neural network: a mathematical structure resem-

bling the brain, made of a concatenation on non-
• Area-law: property by which the entanglement en- linear functions (neurons) according to layers, and
tropy of a region scales proportionally to the size which can be optimized (trained) to recognize pat-
of the boundary of the region. terns, images, etc.
• Parent Hamiltonian: Hamiltonian that has a given
• Boltzmann machine: a specific type of neural net-
PEPS or MPS as unique ground state.
work where the target is to reproduce some Gibbs
• Gap: energy difference between the lowest-energy thermal probabilities.
eigenstate of a Hamiltonian and the first excited
state. • MERGE: linguistic operation introduced by Noam
Chomsky, which picks up two entities (e.g, noun
• Correlation length: nonmathematically, this is the and adjective) and produces a new one from the
length scale at which correlations are sizeable in a two (e.g., noun phrase).
many-body system.
• Many-body localization: property of interacting
• Global symmetry: an operation that leaves invari- quantum many-body systems with disorder leading
ant the system, and which acts equally on the whole to a phase of matter which does not self-thermalize.
system.

[1] R. Orús, Ann. Phys.-New York 349 117158 (2014); F. (1941).

Verstraete, J. I. Cirac, and V. Murg, Adv. Phys. 57,143 [3] R. J. Baxter, Physica A 106, pp18-27 (1981); R. J.
(2008); J. I. Cirac and F. Verstraete, J. Phys. A: Math. Baxter, Exactly Solved Models in Statistical Mechanics,
Theor. 42, 504004 (2009); R. Augusiak, F. M. Cucchi- Academic Press, London, (1982); R. J. Baxter, J. Math.
etti, and M. Lewenstein, in Modern Theories of Many- Phys. 9, 650 (1968); R.J. Baxter, J. Stat. Phys. 19 461
Particle Systems in Condensed Matter Physics, Lect. (1978).
Not. Phys. 843, 245-294 (2012); J. Eisert, Modeling [4] S. R. White, Phys. Rev. Lett.69, 28632866 (1992); S.
and Simulation 3, 520 (2013); N. Schuch, QIP, Lec- R. White, Phys. Rev. B 48, 10345 (1992).
ture Notes of the 44th IFF Spring School (2013); J. [5] M. Fannes, B. Nachtergaele, R. F. Werner, Commun.
Biamonte, V. Bergholm, arXiv:1708.00006. Math. Phys. 144, 443-490 (1992); A. Klümper, A.
[2] H. A. Kramers, G. H. Wannier, Phys. Rev. 60, 263 Schadschneider, J. Zittartz, J. Phys. A 24, L955 (1991);
 13

A. Klümper, A. Schadschneider, J. Zittartz, Europhys. (2004); V. Murg, J.I. Cirac, B. Pirvu, F. Verstraete,
Lett. 24, 293 (1993). New J. Phys. 12, 025012 (2010); E. Mascarenhas, H.
[6] W. Hackbusch, Tensor spaces and numerical tensor cal- Flayac, V. Savona, Phys. Rev. A 92, 022116 (2015); J.
culus, Springer (2012). Cui, J. I. Cirac, M. C. Bañnuls, Phys. Rev. Lett. 114,
[7] S. Östlund, S. Rommer, Phys. Rev. Lett. 75, 3537 220601 (2015); B.-B Chen, L. Chen, Z. Chen, W. Li, A.
(1995). Weichselbaum, Phys. Rev. X 8, 031082 (2018); A. A.
[8] C. Holzhey, F. Larsen, F. Wilczek, Nucl.Phys. B 424 Gangat, T. I, Y.-J. Kao, Phys. Rev. Lett. 119, 010501
443-467 (1994); G. Vidal, J. I. Latorre, E. Rico, A. Ki- (2017).
taev, Phys. Rev. Lett. 90 227902 (2003); J. I. Latorre, [29] P. Czarnik, J. Dziarmaga, Phys. Rev. B 92, 035152
E. Rico, G. Vidal, QIC 4 48-92 (2004); P. Calabrese, J. (2015); P. Czarnik, M. M. Rams, J. Dziarmaga, Phys.
Cardy, JSTAT 0406:P06002 (2004). Rev. B 94, 235142 (2016); A. Kshetrimayum, H.
[9] M. B. Hastings, Phys. Rev. B 73, 085115 (2006); M. B. Weimer, R. Orús, Nat. Comm. 8, 1291 (2017); A.
Hastings, JSTAT, P08024 (2007). Kshetrimayum, M. Rizzi, J. Eisert, R. Orús, Phys. Rev.
[10] L. Tagliacozzo, R. R. de Oliveira, S. Iblisdir, J. I. La- Lett. 122, 070502 (2019).
torre, Phys. Rev. B 78, 024410 (2008); F. Pollmann, [30] F. Verstraete, J. J. Garcia-Ripoll, J. I. Cirac, Phys. Rev.
S. Mukerjee, A. Turner, J. E. Moore, Phys. Rev. Lett. Lett. 93, 207204 (2004); A. H. Werner, D. Jaschke, P.
102, 255701 (2009). Silvi, M. Kliesch, T. Calarco, J. Eisert, S. Montangero,
[11] Y. Zou, A. Milsted, G. Vidal, Phys. Rev. Lett. 121, Phys. Rev. Lett. 116, 237201 (2016).
230402, 2018. [31] G. De las Cuevas, N. Schuch, D. Pérez-Garcı́a, J. I.
[12] F. Verstraete, J. I. Cirac, cond-mat/0407066. Cirac, New J. Phys. 15, 123021 (2013).
[13] G. Vidal, Phys. Rev. Lett. 99, 220405 (2007). [32] F. Verstraete, I. Cirac, Phys. Rev. Lett. 104, 190405
[14] B. Swingle, Phys. Rev. D 86, 065007 (2012). (2010); J. Haegeman, J. I. Cirac, T. J. Osborne, H. Ver-
[15] Y. Levine, O. Sharir, N. Cohen, A. Shashua, schelde, F. Verstraete, Phys. Rev. Lett. 105, 251601
arXiv:1803.09780. (2010); J. Haegeman, J. I. Cirac, T. J. Osborne, F.
[16] M.M. Wolf, F. Verstraete, M.B. Hastings, J.I. Cirac, Verstraete, Phys. Rev. B 88, 085118 (2013); D. Jen-
Phys. Rev. Lett. 100, 070502 (2008); A. Riera, J. I. nings, J. Haegeman, T. J. Osborne, F. Verstraete,
Latorre, Phys. Rev. A 74, 052326 (2006). arXiv:1212.3833.
[17] D. Perez-Garcı́a, F. Verstraete, J. I. Cirac, M. M. Wolf, [33] J. Haegeman, T. J. Osborne, H. Verschelde, F. Ver-
Quant. Inf. Comp. 8, 0650-0663 (2008). straete, Phys. Rev. Lett. 110, 100402 (2013).
[18] F. Verstraete, M. M. Wolf, D. Pérez-Garcı́a, J. I. Cirac, [34] C. Brockt, J. Haegeman, D. Jennings, T. J. Osborne,
Phys. Rev. Lett. 96, 220601 (2006). F. Verstraete, arXiv:1210.5401
[19] N. Schuch, M. M. Wolf, F. Verstraete, J. I. Cirac, Phys. [35] U. Schollwöck, Rev. Mod. Phys. 77, 259 (2005); U.
Rev. Lett. 98, 140506 (2007). Schollwöck, Annals of Physics 326, 96 (2011).
[20] J. Dubail, N. Read, arXiv:1307.7726; T.B. Wahl, H.-H. [36] I. P. McCulloch, arXiv:0804.2509; G. M. Crosswhite, A.
Tu, N. Schuch, J.I. Cirac, Phys. Rev. Lett. 111, 236805 C. Doherty, G. Vidal, Phys. Rev. B 78, 035116 (2008).
(2013); T. B. Wahl, S. T. Haßler, H.-H. Tu, J. I. Cirac, [37] D. Porras, F. Verstraete, J. I. Cirac, Phys. Rev. Lett.
N. Schuch, arXiv:1405.0447; M. Mambrini, R. Orús, D. 93, 227205 (2004); P. Pippan, S. R. White, H. G. Ev-
Poilblanc, Phys. Rev. B 94, 205124 (2016). ertz, Phys. Rev. B 81, 081103(R) (2010); B. Pirvu, F.
[21] O. Buerschaper, M. Aguado, G. Vidal, Phys. Rev. B 79, Verstraete, G. Vidal, Phys. Rev. B 83, 125104 (2011).
085119 (2009); Z.-C. Gu, M. Levin, B. Swingle, X.-G. [38] D. Porras, F. Verstraete, J. I. Cirac, Phys. Rev. B 73,
Wen, Phys. Rev. B 79, 085118 (2009); O. Buerschaper, 014410 (2006).
J. M. Mombelli, M. Christandl, M. Aguado, J. Math. [39] N. Chepiga, F. Mila, Phys. Rev. B 96, 054425 (2017).
Phys. 54, 012201 (2013). [40] G. Vidal, Phys. Rev. Lett. 91, 147902 (2003); G. Vidal,
[22] Y. Shi, L. Duan, G. Vidal, Phys. Rev. A 74, 022320 Phys. Rev. Lett. 939, 040502 (2004).
(2006). [41] G. Vidal, Phys. Rev. Lett. 98, 070201 (2007).
[23] P. Silvi, V.Giovannetti, S.Montangero, M.Rizzi, J. I. [42] R. Orús, G. Vidal, Phys. Rev. B 78, 155117 (2008).
Cirac, R. Fazio, Phys. Rev. A 81, 062335 (2010). [43] G. Evenbly, G. Vidal, Phys. Rev. B 79, 144108 (2009).
[24] L. Tagliacozzo, G. Evenbly, G. Vidal, Phys. Rev. B 80, [44] J. Haegeman, T. J. Osborne, F. Verstraete, Phys. Rev.
235127 (2009); V. Murg, Ö. Legeza, R. M. Noack, F. B 88, 075133 (2013); L. Vanderstraeten, J. Haegeman,
Verstraete, Phys. Rev. B 82, 205105 (2010). F. Verstraete, arXiv:1810.07006.
[25] G. Evenbly, G. Vidal, chapter 4 in the book Strongly [45] J. Haegeman, J. I. Cirac, T. J. Osborne, I. Pizorn, H.
Correlated Systems. Numerical Methods, edited by A. Verschelde, F. Verstraete, Phys. Rev. Lett. 107, 070601
Avella and F. Mancini, Springer Series in Solid-State (2011).
Sciences, Vol. 176, 2013, arXiv:1109.5334. [46] J. Haegeman, B. Pirvu, D. J. Weir, J. I. Cirac, T. J.
[26] G. Vidal, Phys. Rev. Lett. 99, 220405 (2007); For an Osborne, H. Verschelde, F. Verstraete, Phys. Rev. B
introduction, see e.g.: G. Vidal, chapter of the book 85, 100408(R) (2012);
Understanding Quantum Phase Transitions, edited by [47] V. Zauner-Stauber, L. Vanderstraeten, M.T. Fishman,
Lincoln D. Carr, Taylor & Francis, Boca Raton, 2010, F. Verstraete, J. Haegeman, Phys. Rev. B 97, 045145
arXiv:0912.1651v2. (2018).
[27] G. Evenbly, G. Vidal, Phys. Rev. Lett. 112, 240502 [48] M. C. Bañuls, M. B. Hastings, F. verstraete, J. I. Cirac,
(2014); G. Evenbly, G. Vidal, Phys. Rev. B 89, 235113 Phys. Rev. Lett. 102, 240603 (2009); A. J. Daley, C.
(2014). Kollath, U. Schollwöck, G. Vidal, J. Stat. Mech.: Theor.
[28] M. Zwolak, G. Vidal, Phys. Rev. Lett. 93, 207205 Exp. (2004) P04005.
[49] F. A. Pollock, C. Rodriguez-Rosario, T. Frauenheim, M.
 14

Paternostro, K. Modi, Phys. Rev. A 97, 012127 (2018); [75] S. S. Jahromi, R. Orús, arXiv:1808.00680.
A. Strathearn, P. Kirton, D. Kilda, J. Keeling, B. W. [76] G. Evenbly, G. Vidal, Phys. Rev. Lett. 102, 180406
Lovett, Nat. Comm. 9, 3322 (2018). (2009).
[50] I. A. Luchnikov, S. V. Vintskevich, H. Ouerdane, S. N. [77] G. Evenbly, G. Vidal, Phys. Rev. Lett. 104, 187203
Filippov, arXiv:1812.00043. (2010).
[51] J. Jordan, R. Orús, G. Vidal, F. Verstraete, J. I. Cirac, [78] P. Corboz, G. Evenbly, F. Verstraete, G. Vidal, Phys.
Phys. Rev. Lett. 101, 250602 (2008); J. Jordan, R. Rev. A 81, 010303(R) (2010); P. Corboz, G. Vidal,
Orús, G. Vidal, Phys. Rev. B 79, 174515 (2009). Phys. Rev. B 80, 165129 (2009).
[52] T. Nishino, K. Okunishi, J. Phys. Soc. Jpn. 65 pp. 891- [79] A. W. Sandvik, G. Vidal, Phys. Rev. Lett. 99, 220602
894 (1996); T. Nishino, K. Okunishi, J. Phys. Soc. Jp. (2007); L. Wang, I. Pizorn, F. Verstraete, Phys. Rev. B
66, 3040 (1997). 83, 134421 (2011).
[53] R. Orús, G. Vidal, Phys. Rev. B 80, 094403 (2009); R. [80] N. Schuch, M. M. Wolf, F. Verstraete, and J. I. Cirac,
Orús, Phys. Rev. B 85, 205117 (2012). Phys. Rev. Lett. 100, 040501 (2008); F. Mezzacapo, N.
[54] L. Vanderstraeten, M. Mariën, F. Verstraete, J. Haege- Schuch, M. Boninsegni, J. I. Cirac New J. Phys. 11,
man, Phys. Rev. B 92, 201111 (2015). 083026 (2009).
[55] M. Levin, C. P. Nave, Phys. Rev. Lett. 99, 120601 [81] M. Lubasch, J. I. Fuks, H. Appel, A. Rubio, J. I. Cirac,
(2007). M. C. Bañuls, New J. Phys. 18, 083039 (2016).
[56] Z. Y. Xie, H. C. Jiang, Q. N. Chen, Z. Y. Weng, T. [82] F. A. Wolf, I. P. McCulloch, O. Parcollet, U. Schollwöck,
Xiang, Phys. Rev. Lett. 103, 160601 (2009); H. H. Zhao, Phys. Rev. B 90, 115124 (2014); F. A. Wolf, I. P. Mc-
Z. Y. Xie, Q. N. Chen, Z. C. Wei, J. W. Cai, T. Xiang, Culloch, U. Schollwöck, Phys. Rev. B 90, 235131 (2014);
Phys. Rev. B 81, 174411 (2010). F. A. Wolf, A. Go, I. P. McCulloch, A. J. Millis, U.
[57] Z. Y. Xie, J. Chen, M. P. Qin, J. W. Zhu, L. P. Yang, Schollwöck, Phys. Rev. X 5, 041032 (2015).
T. Xiang, Phys. Rev. B 86, 045139 (2012). [83] G. Evenbly, S. R. White, Phys. Rev. Lett. 116, 140403
[58] Z.-C. Gu, X.-G. Wen, Phys. Rev. B 80, 155131 (2009). (2016).
[59] G. Evenbly, G. Vidal, Phys. Rev. Lett. 115, 180405 [84] G. Evenbly, S. R. White, Phys. Rev. A 97, 052314
(2015); G. Evenbly, Phys. Rev. B 95, 045117 (2017); Q. (2018).
Hu, A. Franco-Rubio, G. Vidal, arXiv:1809.05176. [85] J. Haegeman, B. Swingle, M. Walter, J. Cotler, G. Even-
[60] S. Yang, Z.-C. Gu, X.-G. Wen, Phys. Rev. Lett. 118, bly, V. B. Scholz, Phys. Rev. X 8, 011003 (2018).
110504 (2017). [86] S. Sahin, K. P. Schmidt, R. Orús, Eur. Lett. 117, 20002
[61] M. Bal, M. Mariën, J. Haegeman, F. Verstraete, Phys. (2017).
Rev. Lett. 118, 250602 (2017). [87] F. Wegner, Ann. Phys. (Leipzig) 3, 77 (1994); S. D.
[62] H.-H. Zhao, Z.-Y. Xie, T. Xiang, M. Imada, Phys. Rev. Glazek, K. G. Wilson, Phys. Rev. D 49, 4214 (1994);
B 93, 125115 (2016). S. D. Glazek, K. G. Wilson, Phys. Rev. D 48, 5863
[63] Z. Y. Xie, H. J. Liao, R. Z. Huang, H. D. Xie, J. Chen, (1993); S. Kehrein, The Flow Equation Approach to
Z. Y. Liu, T. Xiang, Phys. Rev. B 96, 045128 (2017). Many-Particle Systems (Springer Science & Business
[64] E. M. Stoudenmire, S. R. White, Ann. Rev. of Cond. Media, 2006).
Matt. Phys. 3: 111-128 (2012). [88] L. Vanderstraeten, M. Mariën, J. Haegeman, N. Schuch,
[65] S. R. White, D. J. Scalapino, Phys. Rev. B 55, R14701 J. Vidal, F. Verstraete, Phys. Rev. Lett. 119, 070401
(1997); S. R. White, D. J. Scalapino Phys. Rev. Lett. (2017).
80, 1272 (1998). [89] N. Halko, P. G. Martinsson, J. A. Tropp, SIAM Review
[66] H. C. Jiang, Z. Y. Weng, D. N. Sheng, Phys. Rev. Lett. 53, 217 (2011).
101, 117203 (2008); S. Yan, D. A. Huse, S. White, Sci- [90] D. Tamascelli, R. Rosenbach, M. B. Plenio, Phys. Rev.
ence 332, 1173 (2011); S. Depenbrock, I. P. McCulloch, E 91, 063306 (2015); L. Kohn, F. Tschirsich, M. Keck,
U. Schollwöck, Phys. Rev. Lett. 109, 067201 (2012). M. B. Plenio, D. Tamascelli, S. Montangero, Phys. Rev.
[67] G. Ehlers, S. R. White, R. M. Noack, Phys. Rev. B 95, E 97, 013301 (2018); S. Morita, R. Igarashi, H.-H. Zhao,
125125 (2017). N. Kawashima, Phys. Rev. E 97, 033310 (2018).
[68] L. Tagliacozzo, G. Vidal, Phys. Rev. B 83, 115127 [91] R.-Z. Huang, H.-J. Liao, Z.-Y. Liu, H.-D. Xie, Z.-Y. Xie,
(2011). H.-H. Zhao, J. Chen, T. Xiang, arXiv:1611.09574.
[69] V. Murg, F. Verstraete, J. I. Cirac, Phys. Rev. A 75, [92] V. Murg, V. E. Korepin, F. Verstraete, Phys. Rev. B
033605 (2007). 86, 045125 (2012).
[70] L. Vanderstraeten, J. Haegeman, P. Corboz, F. Ver- [93] A. J. Ferris, Phys. Rev. Lett. 113, 010401 (2014).
straete, Phys. Rev. B 94, 155123 (2016); P. Corboz, [94] M. M. Rams, V. Zauner, M. Bal, J. Haegeman, F. Ver-
Phys. Rev. B 94, 035133 (2016). straete, Phys. Rev. B 92, 235150 (2015).
[71] H. C. Jiang, Z. Y. Weng, T. Xiang, Phys. Rev. Lett. [95] P. Schmoll, R. Orús, Phys. Rev. B 95, 045112 (2017).
101, 090603 (2008). [96] I. McCulloch, M. Gulacsi, Eur. Phys. Lett. 57, 852
[72] M. Lubasch, I. Cirac, M.C. Bañuls, Phys. Rev. B 90, (2002); I. P. McCulloch, J. Stat. Mech. (2007) P10014;
064425 (2014); H. N. Phien, J. A. Bengua, H. D. Tuan, S. Singh, H.-Q. Zhou and G. Vidal, New J. Phys. 033029
P. Corboz, R. Orús, Phys. Rev. B 92, 035142 (2015). (2010).
[73] L. Vanderstraeten, J. Haegeman, F. Verstraete, [97] S. Singh, R. N. C. Pfeifer and G. Vidal, Phys. Rev. B
arXiv:1809.06747. 83, 115125 (2011); Phys. Rev. B 86, 195114 (2012).
[74] P. Corboz, Phys. Rev. B 93, 045116 (2016); P. Corboz, [98] B. Bauer, P. Corboz, R. Orús and M. Troyer, Phys. Rev.
P. Czarnik, G. Kapteijns, L. Tagliacozzo, Phys. Rev. X B 83, 125106 (2011); T. Liu, W. Li, A. Weichselbaum, J.
8, 031031 (2018); M. Rader, A. Läuchli, Phys. Rev. X von Delft and G. Su, Phys. Rev. B 91, 060403 (2015);
8, 031030 (2018). W. Li, A. Weichselbaum, J. von Delft and H.-H. Tu,
 15

Phys. Rev. B 91, 224414 (2015); S.-J. Ran, W. Li, S.- K. Van Acoleyen, arXiv:1702.08838.
S. Gong, A. Weichselbaum, J. von Delft and G. Su, [117] M. C. Bañuls, K. Cichy, J. I. Cirac, S. Kühn, Phys. Rev.
Phys. Rev. B 97, 075146 (2018); D. Poilblanc and M. Lett. 118, 071601 (2017); M. C. Bañuls, K. Cichy, J. I.
Mambrini, Phys. Rev. B 96, 014414 (2017); J.-Y. Chen, Cirac, K. Jansen, S. Kühn, H. Saito, arXiv:1611.01458;
L. Vanderstraeten, S. Capponi and D. Poilblanc, Phys. M. C. Bañuls, K. Cichy, J. I. Cirac, K. Jansen, S. Kühn,
Rev. B 98, 184409 (2018); C. Hubig, SciPost Phys. 5, H. Saito, EPJ Web of Conferences 137, 04001 (2017).
047 (2018). [118] K. Zapp, R. Orús, Phys. Rev. D 95, 114508 (2017).
[99] A. Weichselbaum, Ann. Phys. 327, 2972-3047 (2012); [119] B. Buyens, J. Haegeman, H. Verschelde, F. Verstraete,
P. Silvi, F. Tschirsich, M. Gerster, J. Jünemann, D. K. Van Acoleyen, Phys. Rev. X 6, 041040 (2016).
Jaschke, M. Rizzi and S. Montangero, arXiv:1710.03733. [120] T. Pilcher, M. Dalmonte, E. Rico, P. Zoller, S. Mon-
[100] P. Schmoll, S. Singh, M. Rizzi, R. Orús, tangero, Phys. Rev. X 6, 011023 (2016).
arXiv:1809.08180. [121] S. Kühn, E. Zohar, J. I. Cirac, M. C. Bañuls, JHEP 07
[101] R. Penrose, Angular momentum: an approach to combi- (2015) 130; P. Silvi, E. Rico, M. Dalmonte, F. Tschir-
natorial space-time, 1971. (http://math.ucr.edu/home/ sich, S. Montangero, Quantum 1, 9 (2017).
baez/penrose/); C. Rovelli, L. Smolin, Physical Review [122] E. Zohar, M. Burrello, T. B. Wahl, J. I. Cirac, Annals
D 52, 5743 (1995); C. Rovelli, Living Reviews in Rela- of Physics (2015), pp. 385-439.
tivity (http://www.livingreviews.org/lrr-2008-5). [123] E. Zohar, M. Burrello, T. B. Wahl, J. I. Cirac, Annals
[102] J. F. Cornwell, Group Theory in Physics, (Academic of Physics 374 84-137 (2016).
Press, 1997). [124] M. A. Levin X.-G. Wen, Phys. Rev. B 71, 045110
[103] P. Corboz, R. Orús, B. Bauer, G. Vidal, Phys. Rev. B (2005).
81, 165104 (2010). [125] M. Aguado and G. Vidal, Phys. Rev. Lett. 100, 070404
[104] C. V. Kraus, N. Schuch, F. Verstraete, J. I. Cirac, Phys. (2008); R. Koenig, B. W. Reichardt, G. Vidal, Phys.
Rev. A 81, 052338 (2010). Rev. B 79, 195123 (2009).
[105] I. Pizorn, F. Verstraete, Phys. Rev. B 81, 245110 (2010); [126] S. Dusuel, M. Kamfor, R. Orús, K. P. Schmidt, J. Vidal,
Q.-Q. Shi, S.-Hao Li, J.-Hui Zhao, H.- Qiang Zhou, Phys. Rev. Lett. 106, 107203 (2011); M. D. Schulz, S.
arXiv:0907.5520. Dusuel, R. Orús, J. Vidal, K. P. Schmidt, New J. Phys.
[106] C. Pineda, T. Barthel, J. Eisert, Phys. Rev. A 81, 14, 025005 (2012).
050303(R) (2010); T. Barthel, C. Pineda, J. Eisert, [127] M. P. Zaletel, R. S. K. Mong, F. Pollmann, Phys. Rev.
Phys. Rev. A 80, 042333 (2009). Lett. 110, 236801 (2013);
[107] R. Orús, Eur. Phys. J. B 87, 280 (2014). [128] B. Estienne, Z. Papic, N. Regnault, B. A. Bernevig,
[108] L. Tagliacozzo, A. Celi, M. Lewenstein, Phys. Rev. X Phys. Rev. B 87, 161112(R) (2013); B. Estienne, N.
4, 041024 (2014); P. Silvi, E. Rico, T. Calarco, S. Mon- Regnault, B. A. Bernevig, arXiv:131.2936;
tangero, New J. Phys. 16 103015 (2014); E. Rico, T. [129] F. Pollmann, A. Turner, E. Berg, M. Oshikawa, Phys.
Pichler, M. Dalmonte, P. Zoller, S. Montangero, Phys. Rev. B 81 064439 (2010); F. Pollmann, A. Turner, Phys.
Rev. Lett. 112, 201601 (2014). Rev. B 86, 125441 (2012); H.-W. Chang, Y.-D. Hsieh,
[109] B. Buyens, J. Haegeman, K. Van Acoleyen, H. Ver- Y.-J. Kao, arXiv:1305.2663.
schelde, F. Verstraete, Phys. Rev. Lett. 113, 091601 [130] R. N. C. Pfeifer, P. Corboz, O. Buerschaper, M. Aguado,
(2014); B. Buyens, K. Van Acoleyen, J. Haegeman, F. M. Troyer, G. Vidal, Phys. Rev. B 82, 115126 (2010); S.
Verstraete, PoS(LATTICE2014)308. Singh, R. N. C. Pfeifer, G. Vidal, G. K. Brennen, Phys.
[110] T. Sugihara, J. High Energy Phys. 7 (2005) 022. Rev. B 89, 075112 (2014).
[111] T. M. R. Byrnes, P. Sriganesh, R. J. Bursill, [131] N. Bultinck, M. Mariën, D. J. Williamson, M. B.
C. J. Hamer, Nucl. Phys. B, Proc. Suppl. 109, Sahinoglu, J. Haegeman, F. Verstraete, Ann. of Phys.
202 (2002); T. M. Byrnes, P. Sriganesh, R. J. 378, 183-233 (2017).
Bursill, C. J. Hamer, Phys. Rev. D 66, 013002 [132] D. J. Williamson, N. Bultinck, F. Verstraete,
(2002); T.M.R. Byrnes, Ph.D. thesis, The Univer- arXiv:1711.07982.
sity of New South Wales, 2003, http://books.google.co [133] X.-G. Wen, Rev. Mod. Phys. 89, 41004 (2017).
.jp/books?id=FwaiXgJWM2MC. [134] X. Chen, Z.-C. Gu, X.-G. Wen, Phys. Rev. B 83, 035107
[112] M. C. Bañuls, K. Cichy, J. I. Cirac, K. Jansen, H. Saito, (2011).
PoS(LATTICE 2013)332. [135] N. Schuch, D. Pérez-Garcı́a, J. I. Cirac, Phys. Rev. B
[113] H. Saito, M. C. Bañuls, K. Cichy, J. I. Cirac, K. Jansen, 84, 165139 (2011).
arXiv:1412.0596; M. C. Bañuls, K. Cichy, J. I. Cirac, K. [136] O. Buerschaper, Ann. Phys. 351, 447-476 (2014).
Jansen, H. Saito, Phys. Rev. D 92, 034519 (2015); H. [137] D. J. Williamson, N. Bultinck, J. Haegeman, F. Ver-
Saito, M. C. Bañuls, K. Cichy, J. I. Cirac, K. Jansen, straete, arXiv:1609.02897.
proceeding of the 33rd International Symposium on Lat- [138] N. Bultinck, D. J. Williamson, J. Haegeman, F. Ver-
tice Field Theory, 14-18 July 2015, Kobe in Japan; M. straete, J. Phys. A: Math. Theor. 51 025202 (2017).
C. Bañuls, K. Cichy, K. Jansen, H. Saito, Phys. Rev. [139] N. Bultinck, D. J. Williamson, J. Haegeman, F. Ver-
D 93, 094512 (2016); B. Buyens, F. Verstraete, K. Van straete, Phys. Rev. B 95, 075108 (2017).
Acoleyen, Phys. Rev. D 94, 085018 (2016). [140] H. Li, F. D. M. Haldane, Phys. Rev. Lett. 101, 010504
[114] M. C. Bañuls, K. Cichy, K. Jansen, J. I. Cirac, JHEP (2008).
11 (2013) 158; B. Buyens, J. Haegeman, F. Verstraete, [141] J. I. Cirac, D. Poilblanc, N. Schuch, F. Verstraete, Phys.
K. Van Acoleyen, PoS(LATTICE 2015)280. Rev. B 83, 245134 (2011).
[115] B. Buyens, J. Haegeman, F. Hebenstreit, F. Verstraete, [142] J. Haferkamp, D. Hangleiter, J. Eisert, M. Gluza,
K. Van Acoleyen, arXiv:1612.00739. arXiv:1810.00738.
[116] B. Buyens, S. Montangero, J. Haegeman, F. Verstraete, [143] M. J. Kastoryano, A. Lucia, D. Pérez-Garcı́a,
 16

arXiv:1709.07691. Phys. Rev. B 93, 201113 (2016).

[144] J. M. Maldecena, Adv. Theor. Math. Phys. 2, 231 [165] B. Czech, G. Evenbly, L. Lamprou, S. McCandlish, X.-
(1998); S. S. Gubser, I. R. Klebanov, A. M. Polyakov, L. Qi, J. Sully, G. Vidal, Phys. Rev. B 94, 085101
Phys. Lett. B 428, 105 (1998); E. Witten, Adv. Theor. (2016).
Math. Phys. 2, 253 (1998). [166] M. Hauru, G. Evenbly, W. W. Ho, D. Gaiotto, G. Vidal,
[145] G. Evenbly, G. Vidal, J Stat Phys 145:891-918 (2011). Phys. Rev. B 94, 115125 (2016).
[146] B. Swingle, arXiv:1209.3304. [167] Q. Hu, G. Vidal, Phys. Rev. Lett. 119, 010603 (2017).
[147] M. Nozaki, S. Ryu, T. Takayanagi, JHEP 10 (2012) [168] A. Milsted, G. Vidal, arXiv:1805.12524.
193; A. Mollabashi, M. Nozaki, S. Ryu, T. Takayanagi, [169] A. Milsted, G. Vidal, arXiv:1807.02501.
JHEP 03 (2014) 098. [170] M. Bal, D. J. Williamson, R. Vanhove, N. Bultinck, J.
[148] C. Bény, New J. Phys. 15 (2013) 023020; B. Czech, L. Haegeman, F. Verstraete, arXiv:1801.05959.
Lamprou, S. McCandlish, J. Sully, arXiv:1512.01548; R. [171] G. K.-L. Chan, J. J. Dorando, D. Ghosh, J.
S. Kunkolienkar, K. Banerjee, International Journal of Hachmann, E. Neuscamman, H. Wang and T. Yanai,
Modern Physics D 26, 1750143 (2017); N. Bao, C.-Jun arXiv:0711.1398.
Cao, S. M. Carroll, A. Chatwin-Davies, Phys. Rev. D [172] C. Krumnow, L. Veis, Ö. Legeza and J. Eisert, Phys.
96, 123536 (2017). Rev. Lett. 117, 210402 (2016).
[149] A. Milsted, G. Vidal, arXiv:1812.00529. [173] S. Szalay, M. Pfeffer, V. Murg, G. Barcza, F. Verstraete,
[150] S. Ryu, T. Takayanagi, Phys. Rev. Lett. 96, 181602 R. Schneider, Ö. Legeza, Int. J. Quant. Chem. 115, 1342
(2006) (2015).
[151] H. Matsueda, arXiv:1208.5103; H. Matsueda, [174] B. Paredes, F. Verstraete, J. I. Cirac, Phys. Rev. Lett.
arXiv:1208.5305; J. Molina-Vilaplana, P. Sodano, JHEP 95, 140501 (2005); A. M. Goldsborough, R. A. Römer,
10 011 (2011); J. Molina-Vilaplana, arXiv:1109.5592; H. Phys. Rev. B 89, 214203, 2014; A. M. Goldsborough,
Matsueda, M. Ishihara, Y. Hashizume, arXiv:1208.0206; R. A. Römer, EPL 111, 26004-6 (2015); A. M. Golds-
H. Matsueda, arXiv:1208.5106; J. Molina-Vilaplana, borough, G. Evenbly, Phys. Rev. B 96, 155136 (2017).
JHEP 05 024 (2013); J. Molina-Vilaplana, J. Prior, [175] A. Chandran, J. Carrasquilla, I. H. Kim, D. A. Abanin,
Gen. Relativ. Gravit 46:1823 (2014); N. Bao, G. G. Vidal, Phys. Rev. B 92, 024201 (2015).
Penington, J. Sorce, A. Wall, arXiv:1812.01171; P. [176] F. Pollmann, V. Khemani, J. I. Cirac, S. L. Sondhi,
Hayden, S. Nezami, X.-L. Qi, N. Thomas, M. Walter, Phys. Rev. B 94, 041116 (2016).
Z. Yang, JHEP 11 009 (2016); X.-L. Qi, Z. Yang, [177] T. B. Wahl, A. Pal, S. H. Simon, Phys. Rev. X 7, 021018
arXiv:1801.05289; R. Vasseur, A. C. Potter, Y.-Z. You, (2017).
A. W. W. Ludwig, arXiv:1807.07082; F. Pastawski, B. [178] T. B. Wahl, Phys. Rev. B 98, 054204 (2018).
Yoshida, D. Harlow, J. Preskill, JHEP 06 149 (2015); [179] Depending on the situation one may use χ, or D, or
T. Hartman, J. Maldacena, JHEP 05 014 (2013). both. For instance, we use χ for the bond dimension of
[152] J. Chen, S. Cheng, H. Xie, L. Wang, T. Xiang, Phys. MPS and MERA. However, when dealing with 2d tensor
Rev. B 97, 085104 (2018). networks such as a PEPS, one uses D for the PEPS
[153] I. Glasser, N. Pancotti, M. August, I. D. Rodriguez, I. bond dimension, and reserves χ for the bond dimension
I. Cirac, Phys. Rev. X 8, 011006 (2018). of the environment tensors of 2d numerical algorithms,
[154] E. M. Stoudenmire, D. J. Schwab, Adv. Neur. Inf. Proc. see Sec. III B.
Syst. 29, 4799 (2016).
[155] E. M. Stoudenmire, Quant. Sci. and Tech. 3, 034003
(2018).
[156] D. Liu, S.-J. Ran, P. Wittek, C. Peng, R. Blázquez
Garcı́a, G. Su, M. Lewenstein, arXiv:1710.04833.
[157] Z.-Y. Han, J. Wang, H. Fan, L. Wang, P. Zhang, Phys.
Rev. X 8, 031012 (2018).
[158] W. Huggins, P. Patel, K. B. Whaley, E. M. Stoudenmire,
arXiv:1803.11537.
[159] I. Glasser, N. Pancotti, J. I. Cirac, arXiv:1806.05964.
[160] A. J. Gállego, R. Orús, arXiv:1708.01525.
[161] P. Corboz, S. R. White, G. Vidal, M. Troyer, Phys. Rev.
B 84, 041108 (2011); P. Corboz, T. M. Rice, M. Troyer,
Phys. Rev. Lett. 113, 046402 (2014).
[162] Z. Y. Xie, J. Chen, J. F. Yu, X. Kong, B. Normand, T.
Xiang, Phys. Rev. X 4, 011025 (2014).
[163] T. Picot, M. Ziegler, R. Orús, D. Poilblanc, Phys. Rev.
B 93, 060407 (2016).
[164] P. Corboz, A.M. Laeuchli, K. Penc, M. Troyer, F. Mila,
Phys. Rev. Lett 107, 215301 (2011); B. Bauer, P. Cor-
boz, A. M. Laeuchli, L. Messio, K. Penc, M. Troyer, F.
Mila, Phys. Rev. B 85, 125116 (2012); P. Corboz, K.
Penc, F. Mila, A. M. Laeuchli, Phys. Rev. B 86, 041106
(2012); P. Corboz, M. Lajko, K. Penc, F. Mila, A. M.
Laeuchli, Phys. Rev. B 87, 195113 (2013); P. Nataf, M.
Lajko, P. Corboz, A. M. Laeuchli, K. Penc, F. Mila,