Entanglement and the density matrix renormalisation group

in the generalised Landau paradigm

Laurens Lootens,1, 2, ∗ Clement Delcamp,3 and Frank Verstraete1, 2
1
Department of Applied Mathematics and Theoretical Physics, University of Cambridge,
Wilberforce Road, Cambridge, CB3 0WA, United Kingdom
2
Department of Physics and Astronomy, Ghent University, Krijgslaan 281, 9000 Gent, Belgium
3
Institut des Hautes Études Scientifiques, Bures-sur-Yvette, France
We leverage the interplay between gapped phases and dualities of symmetric one-dimensional quan-
tum lattice models to demonstrate that every phase is efficiently characterised by the maximal
breaking of the dual (genereralised) symmetry whose structure encodes the quasiparticle excita-
tions. This result has strong implications for the complexity of simulating many-body systems
using variational tensor network methods. For every phase in the phase diagram, the dual repre-
arXiv:2408.06334v1 [quant-ph] 12 Aug 2024

sentation of the ground state that breaks all symmetries minimises both the entanglement entropy
and the required number of variational parameters. We demonstrate the applicability of this idea
by developing a generalised density matrix renormalisation group algorithm that works on (dual)
constrained Hilbert spaces, and quantify the computational gains obtained over traditional DMRG
methods in a perturbed Heisenberg model. Our work testifies to the usefulness of generalised
non-invertible symmetries and their formal category theoretic description for the nuts and bolts
simulation of strongly correlated systems.

Sec. I | Introduction follows that ground states of local gapped Hamiltoni-

ans exhibit an area law for the entanglement entropy
The concept of symmetry breaking forms one of the cor- [15]. Such ground states can efficiently be represented in
nerstones of many-body physics. This was already recog- terms of variational matrix product states (MPS) [16],
nized by Pierre Curie in 1894 [1, 2]: “Phenomenon may and the corresponding density matrix renormalisation
exist in a medium having the same characteristic sym- group (DMRG) has effectively broken down the expo-
metry or the symmetry of a subgroup of its characteristic nential complexity wall for simulating ground states of
symmetry. What is necessary, is that certain elements interacting one-dimensional quantum lattice models [17–
of symmetry do not exist. Dissymmetry is what creates 19]. The less entanglement in the ground state, the better
the phenomenon.” Landau [3] formalised this idea, and the algorithm works.
demonstrated that different gapped phases are charac- This entanglement-based approach has also unveiled
terised by local order parameters which transform non- ways of distinguishing different topological phases of mat-
trivially under the global symmetry. In 1971, Kadanoff ter via the entanglement degrees of freedom of the MPS:
and Ceva [4] turned the tables around and realised that besides the subgroup describing the remaining symmetry,
non-local disorder parameters can be used to characterise phases are also labelled by the possible ways in which the
symmetric phases. Their non-local order parameter was entanglement degrees of freedom transform projectively
obtained by a Kramers-Wannier duality transformation under this subgroup [20, 21]. Furthermore, it facilitates
of the local order parameter in the Ising model [5]. Non- the definition of symmetry operators that do not act in an
local order parameters became much more prominent af- on-site manner, but rather encode a correlated action on
ter the discovery of phases of matter that could only neighbouring sites, often on a constrained Hilbert space
be characterized through such non-local order parame- without tensor product structure. Mathematically, such
ters [6, 7]. These topological phases can be constructed generalised symmetries [22–32] are described by a fusion
through dualities relating theories with global symme- category [33], and they can be represented explicitly as
tries to dual (gauge) theories [8–11], and their discovery matrix product operators (MPO) [34–38]. The phases of
challenged the standard Landau paradigm of symmetry such systems are classified by a choice of module cate-
breaking [12]. gory compatible with those symmetries [27, 39–43], and
The modern approach to characterising phases of mat- the generalised Landau paradigm [29, 44–47] entails that
ter uses the language of entanglement and quantum cir- the inclusion of these generalised symmetries yields a
cuits. Two Hamiltonians with a given symmetry are in complete classification of gapped phases. Interestingly,
the same phase if and only if there exists a symmetry- the same classification labels all possible dualities, or all
preserving constant-depth quantum circuit that trans- possible ways in which a (generalised) symmetry can be
forms their ground states into each other [13, 14]. For gauged [48–50]. The explicit intertwiners representing
the case of one-dimensional quantum lattice models, it those dualities on the lattice are again of the MPO form
[38], and they map local order parameters to non-local
ones. By incorporating all possible boundary conditions,
one can show that these dualities are unitary and hence
∗ ll708@cam.ac.uk preserve the full spectrum of the Hamiltonian [51, 52].
 2

In this paper, we connect the concepts of symmetry lustrate our algorithm with a Heisenberg-like model in
breaking, dualities and entanglement, and come to the various phases, before presenting the general mathemat-
remarkable insight that for every Hamiltonian, one can ical framework.
determine an optimal dual Hamiltonian whose ground In retrospect, there is one scenario where this has al-
states exhibit the minimal amount of entanglement. ready been used to great success. Indeed, when applying
More precisely, consider a Hamiltonian with a (gener- our method to an ordinary symmetric phase, it is essen-
alised) symmetry and with a ground state in a certain tially equivalent to using symmetric tensors [53–55] in
phase. By a (twisted) gauging of the remaining sym- which the variational degrees of freedom effectively build
metries, the corresponding dual Hamiltonian will spon- up the (non-invertible) symmetry-broken ground states
taneously break its dual symmetry completely, and the of our dual constrained Hilbert space. By defining Hamil-
corresponding ground states exhibit the minimal amount tonians directly in the dual space [38, 51], our method
of entanglement when compared to all other dual models. yields a much simpler way of implementing symmetry-
As all redundancy due to the symmetries is eliminated, preserving tensor networks. This strategy was also used
those dual ground states have the minimal possible num- in a visionary paper by Sierra and Nishino from 1997 [56],
ber of variational parameters, and this leads to a large in which the Heisenberg model was studied by means of
reduction in complexity for their simulation. The ground its dual interaction round a face Hamiltonian. It took the
states in the original theory can be obtained by multiply- development of a general theory of dualities for the most
ing the optimal ones with duality intertwiners in the form general kind of symmetry in terms of matrix product op-
of MPO’s [38, 51], which reintroduces the multiplicities in erators to turn this is into a practical algorithm for sim-
the entanglement spectra by enlarging the bond dimen- ulating ground states of Hamiltonians with (generalised)
sion. The fact that generalised symmetries emerge in a symmetries that is optimal even beyond the trivially sym-
very natural way, even in cases where the original Hamil- metric phase.
tonian only exhibits global on-site symmetries, makes a
strong case for the computational usefulness of the gen-
eralised Landau paradigm. Sec. II | Illustrative example
The mathematical underpinnings of this result can be
summarised in a commutative diagram: Consider an open chain of length L, and assign to ev-
ery site i spin-1 degrees of freedom with spin operators
Symmetry Kinematics Hamiltonian (Six , Siy , Siz ). These degrees of freedom are governed by
the nearest neighbour Hamiltonian [57]
Phase Entanglement
L−1
ÿ

Quasiparticles H= hi,0 + J1 hi,1 + J2 hi,2 , (1)
i=1

The arrows denote equivalences between the fusion cate- with coupling constants J1 , J2 ∈ R and local operators
gories that organise the symmetry, the symmetric Hamil-
tonians and the quasiparticle excitations. More precisely, hi,0 := Six Si+1
x
+ Siy Si+1
y
+ Siz Si+1
z
,
the category that describes the algebraic properties of the
Hamiltonian and the category that describes the symme- hi,1 := (Six Si+1
x
)2 + (Siy Si+1
y
)2 + (Siz Si+1
z
)2 ,
try satisfy an equivalence determined by the kinematical hi,2 := {Six , Siy }Si+1
z y
+ {Siz , Six }Si+1 + {Siy , Siz }Si+1
x
.
degrees of freedom on which both the Hamiltonian and
the symmetry act. Similarly, the phase of the Hamil- While the term hi,0 defines the spin-1 antiferromagnetic
tonian determines an equivalence between the category Heisenberg model, which is SO(3) symmetric, the terms
that describes the symmetry and the one that describes hi,1 and hi,2 are perturbations breaking the symmetry
the fusion properties of quasiparticle excitations. Com- down to the finite subgroup A4 ⊂ SO(3) of orientation-
bining these two provides an equivalence between the preserving symmetries of the tetrahedron. Specifically,
Hamiltonian and quasiparticle categories, which governs A4 is isomorphic to the semidirect product Z3 ⋉D2 , where
the properties of the entanglement degrees of freedom in Z3 is generated by any cyclic permutation ofx (S x , S y , Szz ),
the optimal tensor network representation of the ground whereas D2 ∼ = Z2 × Z2 is generated by eiπS and eiπS .
state. We explain this diagram in more detail in figure 4. In the presence of a symmetry A4 , we distinguish seven
The only price we pay by considering a dual model is possible gapped phases. These are labelled by a sub-
that it may not be defined on a tensor product Hilbert group H ⊆ A4 characterising the symmetry of the ground
space due to kinematical constraints introduced by the states, together with a class [ψ] labeling the projective
gauging procedure. We deal with this aspect by devel- representations that classify possible symmetry protected
oping a variant of the DMRG algorithm that directly topological (SPT) orders [20]. Physically, a non-trivial
incorporates these constraints, and demonstrate that all [ψ] signals the presence of edge modes that transform
the building blocks for state-of-the-art implementations projectively under the action of the symmetry H. Up
of tensor networks algorithms are still in place. We il- to isomorphisms, A4 counts five subgroups, namely A4 ,
 3

D2 , Z3 , Z2 as well as the trivial one. Out of these sub- structure factors, and not the specific representation cho-
groups, only A4 and D2 have non-trivial second cohomol- sen. Interpreting eq. (3) as a tensor network equation,
ogy groups, which are both isomorphic to Z2 . one can ask the following question: Is there another set
In order to proceed with our analysis, it is crucial to of tensors, generically depicted as
rewrite the local operators entering the definition of the m
Hamiltonian (1) in such a way that the symmetry A4 is 1 m2
V1 V1
manifest. Invoking a finite group version of the Wigner- V2 V2
i ÿ i
Eckart theorem, we know that any operator transforming ≡ |V1 , m1 ⟩|V2 , m2 ⟩⟨V3 , m3 | ,
V3 m1 ,m2 V3
trivially under A4 must be expressible as a linear combi- m3
nation of Clebsch–Gordan coefficients. The group A4 pos- m3
sesses three one-dimensional irreducible representations (4)
{0, 1, 1∗ } and a single three-dimensional one 3 satisfying also indexed by triplets of representations {V1 , V2 , V3 ⊂
3b3 ∼ = 0 ‘ 1 ‘ 1∗ ‘ 2 · 3. Given three irreducible represen- V1 b V2 }, satisfying eq. (3)? Keeping {hn }n the same,
tations V1 , V2 and V3 such that V3 ⊂ V1 bV2 , we interpret replacing the intertwining maps in eq. (2) by these new
the intertwining map V3 → V1 b V2 as the tensor tensors would result in an isomorphic algebra of local
operators, and thus a dual model with Hamiltonian [38]
v1
v2
V1 V1
V2 V2
i ÿ i 3 3
≡ |V1 , v1 ⟩ |V2 , v2 ⟩ ⟨V3 , v3 | , ÿÿ
v1 ,v2 hdual
i,n = hn (V, i, j) i V j . (5)
V3 V3
v3 V i,j
v3 3 3

where the sums are over basis vectors, and i enumerates When dealing with a symmetry A4 , one can find collec-
the different ways V1 b V2 decomposes into V3 . In this tions of tensors of the form (4) satisfying eq. (3) for any
equation, the diagram on the r.h.s. depicts the Clebsch– H ⊆ A4 and [ψ] ∈ H 2 (H, U(1)) [58]. This is because
Gordan coefficients valued in C. In this notation, we can dualities correspond to (twisted) gauging maps of the
show that the local operators hi,n , n ∈ {0, 1, 2}, are all of symmetry A4 , and there are as many ways to gauge a
the form symmetry as there are ways to spontaneously break it.
Typically, the dual Hamiltonian acts on a distinct mi-
3 3 croscopic Hilbert space, which is not necessarily a tensor
product space as suggested by the graphical notation.
ÿÿ
hi,n ≡ hn (V, i, j) i V j , (2)
V i,j
This is consistent with the gauging interpretation, lead-
3 3 ing to theories with gauge degrees of freedom that satisfy
local Gauß constraints. Given a pair (H, [ψ]), degrees of
where hn (V, i, j) ∈ C. Notice that in this formulation, freedom of the resulting dual model are associated with
the state space of a given spin-1 degree of freedom is irreducible ψ-projective representations [38], and as such
spanned by |3, v⟩, with v = 1, . . . , 3. we label the model by Repψ (H). Instead of Clebsch–
Gordan coefficients, the tensors are found to evaluate to
Racah W -coefficients involving linear representations of
II.A. Dualities G and ψ-projective representations of H. Crucially, the
initial Hamiltonian can be transmuted into any of its du-
Importantly, we have the following equality of intertwin- als via an MPO [51]:
ing maps V4 → V1 b V2 b V3
3 3 3 3
i V j
V2 V2
i k V 3 3 = 3 3 , (6)
V1 V5 V1 3
V3 V6 j
j l i V
ÿÿ
V6 ,kl
V4
= FVV41 V2 V3 V5 ,ij V4
, 3 3 3 3
V6 k,l
(3) which is true for any V , i and j.

V6 ,kl
where the ‘F -symbols’ FVV41 V2 V3 V5 ,ij ∈ C are provided
by the Racah W-coefficients of A4 . One can use this
identity to show that the structure factors of the algebra II.B. Numerical simulations
generated by the local symmetric operators {hi,n }i,n only
depends on the F -symbols and coefficients {hn }n ; the We consider points (J1 , J2 ) = {(1, 1), (−2, −5), (−5, 1)}
eigenvalues of the full Hamiltonian only depend on those in the phase diagram of the model (1) corresponding to
 4

the A4 SPT phase, the A4 symmetric phase and the D2 A4 SPT phase
symmetric phase, respectively. For each point, we simu- Original Rep(Z3 )
late all the seven dual Hamiltonians labelled by Repψ (H),

log(λ)
−1
which result from the ψ-twisted gauging of H ⊆ G, the −2
first of which (coined ”Original”) corresponds to the stan-
−3
dard DMRG algorithm without symmetries and hence
no constraints. The ground states of these dual mod- Rep(D2 ) Repψ (D2 )

log(λ)
−1
els are found using the generalised DMRG algorithm
−2
(c.f. app. A), which performs a variational optimisation
within the subspace of the constrained Hilbert space that −3
is spanned by MPSs of the form Rep(A4 ) Repψ (A4 )

log(λ)
−1

−2
3 3 3
−3
Ai−1 Ai Ai+1 . (7)
Rep(A4 ), rescaled Rep(Z2 )

log(λ)
−1
Due to the constraints, the MPS matrices carry gauge −2
degrees of freedoms associated with ψ-projective repre-
−3
sentations of H, which are shared by neighbouring sites.
Repψ (D2 )

log(λmin )
These are represented by the blue lines in eq. (7). For Original Rep(Z3 ) Rep(D2 )
−4
the three phases, we plot the entanglement spectra of the Rep(Z2 ) Rep(A4 ) Repψ (A4 )
−5
dual ground states coloured by the different objects la-
belling the gauge degrees of freedom, as well as the mem- −6
100 KB 200 KB 300 KB 400 KB 500 KB
ory requirements to reach a specific minimal Schmidt
value λmin serving as an error measure. Our findings
are displayed on fig. 1, 2 and 3 and analysed below: Figure 1. Entanglement spectra of the dual models in the
middle of the ground state on 60 sites in the A4 SPT phase of
• A4 SPT, fig. 1: In this phase, entanglement degrees the initial model (J1 = 1, J2 = 1). The colour of a Schmidt
of freedom of the unique ground state transform as pro- value indicates the object that labels the corresponding gauge
jective representations of A4 . The three irreducible pro- degree of freedom. Bottom: The memory required to store
a ground state MPS tensor in the bulk at double precision
jective representations of A4 are two-dimensional, which
for a given truncation error λmin . The ground state of the
explains the two-fold degeneracy for every Schmidt value. Repψ (A4 ) model minimizes the entanglement and the number
Comparing the entanglement spectra of the various dual of variational parameters for a fixed truncation error.
models, we observe that the entanglement is minimised
in the Repψ (D2 ) and Repψ (A4 ) models. Additionally, the
ground state of the Repψ (A4 ) model requires the least happen to be labelled by irreducible representations of
amount of memory, making this model the most efficient D2 . A more subtle case is the Rep(A4 ) model, which
one to simulate. Crucially, this dual model possesses has a non-invertible Rep(A4 ) symmetry that is only par-
a non-invertible Rep(A4 ) symmetry, whereby symmetry tially broken in the ground state subspace. Although
operators are labelled by representations of A4 , which the remaining symmetry constrains the entanglement
√
happens to be completely broken in the ground state spectrum, the degeneracy is hidden as a fixed 3 ratio
subspace. An important subtlety with a non-invertible (the ”quantum dimension” of the irrep) of consecutive
symmetry breaking phase is that the different ground Schmidt values labelled by the 3d and a 1d irreducible
states, being related by non-trivial MPOs, can have dif- representation. We make this manifest by rescaling the
ferent entanglement properties. In this particular exam- spectrum appropriately.
ple, the ground state labelled by the 3d irreducible repre-
sentation requires more entanglement and therefore more • A4 symmetric, fig. 2: In this phase, entanglement de-
variational parameters than the other ground states. In grees of freedom of the unique ground state transform as
practice, it is possible to avoid this ground state by bi- linear representations of A4 , the three-dimensional irre-
asing the initial MPS towards the other less entangled ducible representation 3 explaining occurrences of three-
ground states. fold degeneracy in the entanglement spectrum. Compar-
All the other dual models admit ground states that ing the entanglement spectra of the various dual mod-
preserve some symmetry, resulting in these cases in de- els, we observe that the entanglement is minimised in
generacy in the entanglement spectrum as well excessive the Rep(A4 ) model. Additionally, it is the model whose
memory requirements. This is ostensible in the Rep(D2 ) ground state requires the least amount of memory, mak-
model, where every Schmidt value has a four-fold degen- ing this model the most efficient one to simulate. As for
eracy due to its ground state preserving a dual A4 sym- the previous phase, this optimal dual model has a non-
metry, the D2 subgroup of which permuting the gauge invertible symmetry Rep(A4 ), which also happens to be
degrees of freedom that label the Schmidt values, which completely broken in the ground state subspace.
 5

A4 symmetric phase D2 symmetric phase

Original Rep(Z3 ) Original Rep(Z3 )
log(λ)

log(λ)
−1 −1

−2 −2

−3 −3
Rep(D2 ) Repψ (D2 ) Rep(D2 ) Repψ (D2 )
log(λ)

log(λ)
−1 −1

−2 −2

−3 −3
ψ
Rep(A4 ) Rep (A4 ) Rep(A4 ) Repψ (A4 )
log(λ)

log(λ)
−1 −1

−2 −2

−3 −3
Rep(A4 ), rescaled Rep(Z2 ) Rep(A4 ), rescaled Rep(Z2 )
log(λ)

log(λ)
−1 −1

−2 −2

−3 −3
ψ
Repψ (D2 )
log(λmin )

log(λmin )
Original Rep(Z3 ) Rep(D2 ) Rep (D2 ) Original Rep(Z3 ) Rep(D2 )
−4 −4
ψ
Rep(Z2 ) Rep(A4 ) Rep (A4 ) Rep(Z2 ) Rep(A4 ) Repψ (A4 )
−5 −5

−6 −6
100 KB 200 KB 300 KB 400 KB 500 KB 20 KB 40 KB 60 KB 80 KB

Figure 2. Top: Entanglement spectra of the dual models in Figure 3. Top: Entanglement spectra of the dual models in
the middle of the ground state on 60 sites in the A4 symmetric the middle of the ground state on 60 sites in the D2 sym-
phase of the original model (J1 = −2, J2 = −5). The colour metric phase of the original model (J1 = −5, J2 = 1). The
of a Schmidt value indicates the object that labels the cor- colour of a Schmidt value indicates the object that labels the
responding gauge degree of freedom. Bottom: The memory corresponding gauge degree of freedom. Bottom: The mem-
required to store a ground state MPS tensor in the bulk at ory required to store a ground state MPS tensor in the bulk
double precision for a given truncation error λmin . The ground at double precision for a given truncation error λmin . The
state of the Rep(A4 ) model minimises the entanglement and ground state of the Rep(D2 ) model minimizes the entangle-
number of variational parameters for a fixed truncation error. ment and the number of variational parameters for a fixed
truncation error.

All the other dual models admit ground states sharing

the same entanglement spectrum as the initial model, dif-
fering only in the labelling of the Schmidt values and the The remaining dual models admit ground states show-
improvement in the memory requirements. These dual ing more entanglement, as a consequence of Schmidt val-
models possess either invertible or non-invertible sym- ues in the initial ground state becoming degenerate. This
metries, and the ground states break various amounts is most easily understood in the Rep(Z3 ) model that pos-
thereof, but never is the whole symmetry broken. sesses a non-invertible Rep(A4 ), which is fully preserved
by the unique ground state. The one-dimensional irre-
• D2 symmetric, fig. 3: In this phase, entanglement de- ducible representations act by permuting gauge degrees
grees of freedom of the ground states transform as linear of freedom labelled by irreducible representations of Z3 ,
representations of D2 . Since irreducible representations enforcing an additional three-fold degeneracy. A similar
of D2 are all one-dimensional, no additional degeneracy in explanation holds for the degenerate Schmidt values in
the entanglement spectrum is enforced, the visible two- the Rep(A4 ) and Repψ (A4 ) models.
fold degeneracies must originate from a hidden symmetry
that might involve time reversal combined with a physi-
cal on-site action. Comparing the entanglement spectra What is the main lesson of this analysis? Regardless of
of the various dual models, we observe that the entan- the phase, the optimal dual model to simulate is always
glement is minimised in the initial model, as well as the the single one whose dual symmetry is completely broken
Repψ (D2 ), Rep(D2 ) and Rep(Z2 ) models. However, the in the ground state subspace. This is the model obtained
Rep(D2 ) model stands out as requiring the least amount by performing the (possibly ψ-twisted) gauging of the
of memory, making this model the most efficient one to symmetry H that is preserved within the ground state
simulate. This optimal dual model has an A4 symmetry, subspace of the initial model. Incidentally, this method
which happens to be completely broken in the ground may also be employed to identify a gapped phase by com-
state subspace. paring entanglement spectra of dual models.
 6

II.C. Symmetry-preserving tensor networks Sec. III | General framework

We close our study of this example with an observation The results presented above hold much more broadly.
that sheds light on the computational gains that we just Consider a generalised symmetry in a one-dimensional
observed. In the symmetric phase, the Hamiltonian of quantum lattice model, i.e., a symmetry whose opera-
the dual model Rep(A4 )—resulting from gauging the A4 tors are not necessarily invertible [24–29, 31, 32, 49]. The
symmetry of the initial model—is transmuted into the symmetry operators take the form of (typically non-local)
initial one via an MPO that takes a particularly simple MPOs [35–38, 51]. Mathematically, a finite generalised
form. Specifically, the building blocks of this MPO eval- symmetry can be modelled by a so-called fusion cate-
uate to Clebsch–Gordan coefficients. By definition, the gory [59], extending the group theoretical formalism of
ground state of the initial model is obtained by acting ordinary symmetries. Similar to ordinary symmetries,
with this MPO upon the ground state of the dual model generalised symmetries can be spontaneously broken—
Rep(A4 ). Graphically, it yields an MPS of the form as we already witnessed in our examples—and may be
gauged provided that there is no ’t Hooft anomaly. Fur-
thermore, gapped phases with a generalised symmetry
3 3 3
are still in one-to-one correspondence with the different
ways to gauge a (sub)symmetry thereof [27, 40], and thus
dualities. This is the essence of the generalised Landau
3 3 3
paradigm [26, 27, 40, 42, 60, 61].
Ai−1 Ai Ai+1 , (8) Consider a one-dimensional quantum lattice model
with a generalised symmetry. Suppose the symmetry
i.e., a linear combination of A4 -invariant tensors. But, is completely broken in the ground state subspace. By
decomposing an MPS in this way is precisely the start- gauging the symmetry, which amounts to identifying the
ing point of symmetry preserving tensor network algo- corresponding symmetry operators as well as the cor-
rithms (c.f. app. A) [53–55]. Concisely, in this form, responding symmetry broken states, we obtain a dual
the symmetry of the state can be exploited to yield a model with a dual symmetry that is fully preserved by the
computational gain, by targeting for instance a specific unique ground state. Crucially, gauging this dual sym-
charge sector. By directly computing the ground state metry recovers the initial model. More generally, there
of the dual Rep(A4 ) model, our method provides a re- is always a way to gauge the (sub)symmetry that is pre-
markably simple way to implement symmetry preserving served in the ground state subspace of a model so as to
tensor network algorithms, which is applicable for any yield a dual model whose dual generalised symmetry is
symmetry group, including generalised symmetries. The completely broken (c.f. app. C). Practically, this dual
projection onto a particular charge sector is enforced lo- model is obtained by extending the approach followed
cally via the Gauß constraints together with the choice of in our series of examples: We write the Hamiltonian in
boundary conditions. Importantly however, our previous terms of tensors that make the generalised symmetry of
analysis revealed that the model Rep(A4 ) is the optimal the model manifest; these tensors satisfy equations gener-
one to simulate if and only if the initial model is in the alised eq. (3); there is a different set of solutions to these
A4 symmetric phase. By accessing the other dual mod- equations, which correspond to the relevant gauging of
els, our approach overcomes this limitation and yields the the preserved subsymmetry, that yield the dual model
optimal DMRG simulation in all possible phases. (c.f. app. C).
In practice, symmetric tensor networks have found We claim that the optimal way of simulating the phase
most of their utility in models with continuous Lie group of a given model amounts to simulating the dual phase
symmetries such as SU(2). While we restricted ourselves of this dual model where the dual symmetry is com-
to a finite group, our results readily generalise to these pletely broken, after which we recover the original ground
cases, implying for instance that a ground state preserv- state by acting with the MPO that transmutes the corre-
ing an SU(2) symmetry is most efficiently parameterised sponding Hamiltonians into each other. Broadly speak-
in terms of a dual non-invertible Rep(SU(2)) symmetry ing, the reasoning is that any symmetry translates into
breaking ground state. Irreducible representations with constraints amongst the variational parameters so they
high spins are given weight zero, so we only have to deal are not all independent. The associated redundancy un-
with a finite number of labels, just as in standard sym- equivocally translates into a suboptimal use of computa-
metric DMRG codes. This scenario turns out to be equiv- tional resources, as we observed in the example above.
alent to simulation a dual interaction-round-a-face model In the optimal dual phase, at least one of the ground
[56]. The different symmetry-broken ground states can states has the property that the action of any dual sym-
be related by acting with MPO symmetry operators. As metry operator on it yields an orthogonal ground state.
a consequence, these ground states do not have the same For this maximal symmetry breaking state, all order pa-
entanglement, showing the importance of properly ini- rameters are strictly local. This follows from the fact that
tialising the DMRG algorithm to favour the ground states the action of the MPOs representing the dual symmetries
with the least amount of entanglement. map such a ground state into a different one, and hence
 7

the expectation value of any non-local string order op- Acknowledgments: We would like to thank Jacob
erator vanishes exponentially in the number of sites on Bridgeman, Lukas Devos, José Garre-Rubio, Jutho
which it acts [62]. Conversely, all quasiparticle excita- Haegeman and Sukhwinder Singh for interesting dis-
tions on top of this maximal symmetry breaking ground cussions and useful comments. This work has re-
state correspond to domain wall excitations. In the case ceived funding from EOS (grant No. 40007526),
of an infinite one dimensional lattice model, these excita- IBOF (grant No. IBOF23/064), BOF-GOA (grant
tions are created by the action of the symmetry MPOs on No. BOF23/GOA/021). LL is supported by an
one half of the chain [63]. As in the usual ansatz for topo- FWO postdoctoral fellowship (grant No. 12AUN24N)
logical excitations in MPS, additional variational degrees and an EPSRC postdoctoral fellowship (grant No.
of freedom characterising the precise nature of the exci- EP/Y020456/1).
tations emerge at the endpoint of the MPO [64, 65]. The
domain wall excitations of the dual symmetry breaking References
model are mapped to the quasiparticle excitations of the
original model, which can be of a very different nature [1] P. Curie, Sur la symétrie dans les phénomènes
(spinon, holon, etc.). Due to the unitarity of the duality physiques, symétrie d’un champ électrique et d’un champ
transformation however, they retain the same algebraic magnétique, J. Phys. Theor. Appl. 3, 393 (1894).
[2] E. Castellani and J. Ismael, Which Curie’s Principle?,
properties, and in particular their fusion properties are
Philosophy of Science 83, 1002 (2016).
described by the same categorical data (c.f. figure 4). [3] L. D. Landau, On the theory of phase transitions, Zh.
Eksp. Teor. Fiz. 7, 19 (1937).
[4] L. P. Kadanoff and H. Ceva, Determination of an opeator
Sec. IV | Outlook algebra for the two-dimensional Ising model, Phys. Rev.
B 3, 3918 (1971).
[5] H. A. Kramers and G. H. Wannier, Statistics of the two-
In appendix (A) we develop an algorithm that performs dimensional ferromagnet. Part 1., Phys. Rev. 60, 252
the optimal (two-site) DMRG simulation of any gapped (1941).
phase with an arbitrary generalised symmetry. In sym- [6] F. J. Wegner, Duality in Generalized Ising Models and
metric phases, it offers a simpler and more versatile im- Phase Transitions Without Local Order Parameters, J.
plementation of symmetry-preserving tensor network al- Math. Phys. 12, 2259 (1971).
gorithms, as it bypasses many aspects of traditional im- [7] M. den Nijs and K. Rommelse, Preroughening transitions
plementations such as fusion trees. Our formalism will be in crystal surfaces and valence-bond phases in quantum
particularly interesting for studying second order phase spin chains, Phys. Rev. B 40, 4709 (1989).
transitions, defects in critical theories, and boundary ef- [8] X. G. Wen, Topological Order in Rigid States, Int. J.
Mod. Phys. B 4, 239 (1990).
fects, as dual models exhibit different finite-entanglement
[9] A. Lopez and E. Fradkin, Fractional quantum hall effect
scaling regimes [66]. Currently, our algorithm only takes and chern-simons gauge theories, Phys. Rev. B 44, 5246
advantage of internal symmetries, since it exploits an ap- (1991).
proach to dualities that has been tailored to this type of [10] T. Kennedy and H. Tasaki, Hidden Z2 ×Z2 symmetry
symmetry. It will be very interesting to generalise this breaking in Haldane-gap antiferromagnets, Phys. Rev. B
approach to different types of symmetry, such as spa- 45, 304 (1992).
tial symmetries and time reversal, which lead to further [11] B. I. Halperin, P. A. Lee, and N. Read, Theory of the
refinenements of phase diagrams. Note also that essen- half-filled Landau level, Phys. Rev. B 47, 7312 (1993).
tially the same ideas can be be used to construct more [12] X.-G. Wen, Quantum Field Theory of Many-Body Sys-
efficient algorithms for simulating quantum spin systems tems: From the Origin of Sound to an Origin of Light
using neural network states [67]. and Electrons (Oxford University Press, 2007).
[13] M. B. Hastings, Locality in quantum systems (2010),
One of the main merits of our method is that it is arXiv:1008.5137 [math-ph].
systematically extensible to higher dimensional models, [14] S. Bravyi, M. B. Hastings, and F. Verstraete, Lieb-
and to tensor network network algorithms in terms of Robinson Bounds and the Generation of Correlations and
projected entangled pair states [68]. In fact, taking ad- Topological Quantum Order, Phys. Rev. Lett. 97, 050401
vantage of symmetries in higher dimensions is expected (2006), arXiv:quant-ph/0603121.
to be even more beneficial than in (1+1)d. Although for- [15] M. B. Hastings, An area law for one-dimensional quan-
mally more challenging, many aspects of dualities in two- tum systems, J. Stat. Mech. 0708, P08024 (2007),
dimensional quantum lattice models have been worked arXiv:0705.2024 [quant-ph].
out [69]. Specifically, the relevant tensor network oper- [16] F. Verstraete and J. I. Cirac, Matrix product states rep-
ators are already known for a large class of generalised resent ground states faithfully, Phys. Rev. B 73, 094423
(2006).
symmetries [70–72]. For instance, a (2+1)d model with
[17] S. R. White, Density matrix formulation for quantum
an ordinary global symmetry in the symmetric phase can renormalization groups, Phys. Rev. Lett. 69, 2863 (1992).
be understood as a dual 1-form symmetry breaking phase [18] U. Schollwoeck, The density-matrix renormalization
obtained by gauging the global symmetry. Our approach group in the age of matrix product states, Annals Phys.
reveals that (2+1)d symmetric tensor network methods 326, 96 (2011), arXiv:1008.3477 [cond-mat.str-el].
[54, 55, 73] effectively implement this duality. [19] J. I. Cirac, D. Pérez-Garcı́a, N. Schuch, and F. Ver-
 8

straete, Matrix product states and projected entangled thogonality, Commun. Math. Phys. 402, 2691 (2023),
pair states: Concepts, symmetries, theorems, Rev. Mod. arXiv:2211.01947 [math.QA].
Phys. 93, 045003 (2021). [38] L. Lootens, C. Delcamp, G. Ortiz, and F. Verstraete, Du-
[20] X. Chen, Z.-C. Gu, and X.-G. Wen, Complete classifica- alities in one-dimensional quantum lattice models: Sym-
tion of one-dimensional gapped quantum phases in inter- metric hamiltonians and matrix product operator inter-
acting spin systems, Phys. Rev. B 84, 235128 (2011). twiners, PRX Quantum 4, 020357 (2023).
[21] N. Schuch, D. Pérez-Garcı́a, and I. Cirac, Classifying [39] A. Kitaev and L. Kong, Models for Gapped Bound-
quantum phases using matrix product states and pro- aries and Domain Walls, Commun. Math. Phys. 313, 351
jected entangled pair states, Phys. Rev. B 84, 165139 (2012), arXiv:1104.5047 [cond-mat.str-el].
(2011). [40] Z. Komargodski, K. Ohmori, K. Roumpedakis, and
[22] A. Feiguin, S. Trebst, A. W. W. Ludwig, M. Troyer, S. Seifnashri, Symmetries and strings of adjoint QCD2 ,
A. Kitaev, Z. Wang, and M. H. Freedman, Interacting JHEP 03, 103, arXiv:2008.07567 [hep-th].
anyons in topological quantum liquids: The golden chain, [41] R. Thorngren and Y. Wang, Fusion category symmetry.
Phys. Rev. Lett. 98, 160409 (2007). Part II. Categoriosities at c = 1 and beyond, JHEP 07,
[23] N. Bultinck, M. Mariën, D. J. Williamson, M. B. 051, arXiv:2106.12577 [hep-th].
Şahinoğlu, J. Haegeman, and F. Verstraete, Anyons and [42] K. Inamura, On lattice models of gapped phases
matrix product operator algebras, Annals Phys. 378, 183 with fusion category symmetries, JHEP 03, 036,
(2017), arXiv:1511.08090 [cond-mat.str-el]. arXiv:2110.12882 [cond-mat.str-el].
[24] D. Aasen, R. S. K. Mong, and P. Fendley, Topological de- [43] J. Garre-Rubio, L. Lootens, and A. Molnár, Classifying
fects on the lattice: I. the ising model, Journal of Physics phases protected by matrix product operator symmetries
A: Mathematical and Theoretical 49, 354001 (2016). using matrix product states, Quantum 7, 927 (2023).
[25] M. Buican and A. Gromov, Anyonic Chains, Topological [44] D. Gaiotto and J. Kulp, Orbifold groupoids, JHEP 02,
Defects, and Conformal Field Theory, Commun. Math. 132, arXiv:2008.05960 [hep-th].
Phys. 356, 1017 (2017), arXiv:1701.02800 [hep-th]. [45] F. Apruzzi, F. Bonetti, I. n. Garcı́a Etxebarria, S. S.
[26] C.-M. Chang, Y.-H. Lin, S.-H. Shao, Y. Wang, and Hosseini, and S. Schafer-Nameki, Symmetry TFTs from
X. Yin, Topological Defect Lines and Renormaliza- String Theory, Commun. Math. Phys. 402, 895 (2023),
tion Group Flows in Two Dimensions, JHEP 01, 026, arXiv:2112.02092 [hep-th].
arXiv:1802.04445 [hep-th]. [46] L. Bhardwaj, L. E. Bottini, D. Pajer, and S. Schafer-
[27] R. Thorngren and Y. Wang, Fusion category symmetry. Nameki, Categorical Landau Paradigm for Gapped
Part I. Anomaly in-flow and gapped phases, JHEP 04, Phases (2023), arXiv:2310.03786 [cond-mat.str-el].
132, arXiv:1912.02817 [hep-th]. [47] S.-J. Huang and M. Cheng, Topological hologra-
[28] D. Aasen, P. Fendley, and R. S. K. Mong, Topological De- phy, quantum criticality, and boundary states (2023),
fects on the Lattice: Dualities and Degeneracies (2020), arXiv:2310.16878 [cond-mat.str-el].
arXiv:2008.08598 [cond-mat.stat-mech]. [48] J. Frohlich, J. Fuchs, I. Runkel, and C. Schweigert, Du-
[29] L. Kong, T. Lan, X.-G. Wen, Z.-H. Zhang, and H. Zheng, ality and defects in rational conformal field theory, Nucl.
Algebraic higher symmetry and categorical symmetry: A Phys. B 763, 354 (2007), arXiv:hep-th/0607247.
holographic and entanglement view of symmetry, Phys. [49] L. Bhardwaj and Y. Tachikawa, On finite symmetries
Rev. Res. 2, 043086 (2020). and their gauging in two dimensions, JHEP 03, 189,
[30] D. S. Freed, G. W. Moore, and C. Teleman, Topo- arXiv:1704.02330 [hep-th].
logical symmetry in quantum field theory (2022), [50] Y. Tachikawa, On gauging finite subgroups, SciPost
arXiv:2209.07471 [hep-th]. Phys. 8, 015 (2020).
[31] S. Schafer-Nameki, ICTP lectures on (non-)invertible [51] L. Lootens, C. Delcamp, and F. Verstraete, Dualities
generalized symmetries, Phys. Rept. 1063, 1 (2024), in one-dimensional quantum lattice models: Topological
arXiv:2305.18296 [hep-th]. sectors, PRX Quantum 5, 010338 (2024).
[32] S.-H. Shao, What’s Done Cannot Be Undone: TASI [52] L. Lootens, C. Delcamp, D. Williamson, and F. Ver-
Lectures on Non-Invertible Symmetries (2023), straete, Low-depth unitary quantum circuits for duali-
arXiv:2308.00747 [hep-th]. ties in one-dimensional quantum lattice models (2023),
[33] P. Etingof, S. Gelaki, D. Nikshych, and V. Ostrik, Tensor arXiv:2311.01439 [quant-ph].
categories, Vol. 205 (American Mathematical Soc., 2016). [53] I. P. McCulloch and M. Gulácsi, The non-abelian density
[34] M. B. Şahinoğlu, D. Williamson, N. Bultinck, M. Mariën, matrix renormalization group algorithm, Europhysics
J. Haegeman, N. Schuch, and F. Verstraete, Character- Letters 57, 852 (2002).
izing Topological Order with Matrix Product Operators, [54] S. Singh, R. N. C. Pfeifer, and G. Vidal, Tensor net-
Annales Henri Poincare 22, 563 (2021), arXiv:1409.2150 work decompositions in the presence of a global symme-
[quant-ph]. try, Phys. Rev. A 82, 050301 (2010).
[35] R. Vanhove, M. Bal, D. J. Williamson, N. Bultinck, [55] A. Weichselbaum, Non-abelian symmetries in tensor net-
J. Haegeman, and F. Verstraete, Mapping topological works: A quantum symmetry space approach, Annals of
to conformal field theories through strange correlators, Physics 327, 2972 (2012).
Phys. Rev. Lett. 121, 177203 (2018). [56] G. Sierra and T. Nishino, The density matrix renormal-
[36] L. Lootens, J. Fuchs, J. Haegeman, C. Schweigert, and ization group method applied to interaction round a face
F. Verstraete, Matrix product operator symmetries and hamiltonians, Nuclear Physics B 495, 505 (1997).
intertwiners in string-nets with domain walls, SciPost [57] A. Prakash, C. G. West, and T.-C. Wei, Detection of
Phys. 10, 053 (2021). gapped phases of a one-dimensional spin chain with on-
[37] J. C. Bridgeman, L. Lootens, and F. Verstraete, In- site and spatial symmetries, Phys. Rev. B 94, 045136
vertible Bimodule Categories and Generalized Schur Or- (2016).
 9

[58] V. Ostrik, Module categories over the Drinfeld double of a algorithms for uniform matrix product states, Phys. Rev.
finite group, International Mathematics Research Notices B 97, 045145 (2018).
2003, 1507 (2003). [77] J. Haegeman and F. Verstraete, Diagonalizing transfer
[59] P. Etingof, D. Nikshych, and V. Ostrik, On fusion cate- matrices and matrix product operators: A medley of ex-
gories (2002), arXiv:math/0203060. act and computational methods, Annual Review of Con-
[60] T.-C. Huang, Y.-H. Lin, and S. Seifnashri, Construction densed Matter Physics 8, 355 (2017).
of two-dimensional topological field theories with non- [78] M. Müger, From subfactors to categories and topology i:
invertible symmetries, JHEP 12, 028, arXiv:2110.02958 Frobenius algebras in and morita equivalence of tensor
[hep-th]. categories, Journal of Pure and Applied Algebra 180, 81
[61] L. Bhardwaj, L. E. Bottini, D. Pajer, and S. Schäfer- (2003).
Nameki, Gapped Phases with Non-Invertible Symme- [79] A. Molnar, A. R. de Alarcón, J. Garre-Rubio, N. Schuch,
tries: (1+1)d (2023), arXiv:2310.03784 [hep-th]. J. I. Cirac, and D. Pérez-Garcı́a, Matrix product opera-
[62] J. Haegeman, V. Zauner, N. Schuch, and F. Verstraete, tor algebras I: representations of weak Hopf algebras and
Shadows of anyons and the entanglement structure projected entangled pair states (2022), arXiv:2204.05940
of topological phases, Nature communications 6, 8284 [quant-ph].
(2015). [80] D. Barter, J. C. Bridgeman, and R. Wolf, Computing
[63] J. Haegeman, B. Pirvu, D. J. Weir, J. I. Cirac, T. J. Os- associators of endomorphism fusion categories, SciPost
borne, H. Verschelde, and F. Verstraete, Variational ma- Phys. 13, 029 (2022).
trix product ansatz for dispersion relations, Phys. Rev.
B 85, 100408 (2012).
[64] V. Zauner-Stauber, L. Vanderstraeten, J. Haegeman, Sec. A | Generalised DMRG
I. P. McCulloch, and F. Verstraete, Topological nature of
spinons and holons: Elementary excitations from matrix
product states with conserved symmetries, Phys. Rev. B In order to compute the ground states of the various mod-
97, 235155 (2018). els appearing in our examples, we implemented a version
[65] M. Mariën, J. Haegeman, P. Fendley, and F. Ver- of the two-site DMRG algorithm for finite chains with
straete, Condensation-driven phase transitions in per- open boundary conditions [74]. The DMRG algorithm
turbed string nets, Phys. Rev. B 96, 155127 (2017). is a variational algorithm within the subspace of MPSs,
[66] R.-Z. Huang, L. Zhang, A. M. Läuchli, J. Haegeman, which we recall is a class of wavefunctions that implement
F. Verstraete, and L. Vanderstraeten, Emergent confor- the area law of gapped phases at the microscopic level,
mal boundaries from finite-entanglement scaling in ma- thereby specifically targeting the physical corner of the
trix product states, Phys. Rev. Lett. 132, 086503 (2024).
total Hilbert space. Briefly, the DMRG algorithm pro-
[67] T. Vieijra, C. Casert, J. Nys, W. De Neve, J. Haegeman,
J. Ryckebusch, and F. Verstraete, Restricted boltzmann
ceeds as follows: The wavefunction being a multilinear
machines for quantum states with non-abelian or anyonic function of the variables in all local tensors, the global
symmetries, Phys. Rev. Lett. 124, 097201 (2020). optimisation problem can iteratively be solved using an
[68] F. Verstraete, V. Murg, and J. I. Cirac, Matrix prod- alternating least squares approach [75]. The two-site ver-
uct states, projected entangled pair states, and varia- sion proceeds by blocking two sites together, before solv-
tional renormalization group methods for quantum spin ing the combined least-squares problem, and finally us-
systems, Adv. Phys. 57, 143 (2008). ing a singular value decomposition in order to split the
[69] K. Inamura and K. Ohmori, Fusion surface models: 2+1d two-site tensor into two one-site tensors. This two-site
lattice models from fusion 2-categories, SciPost Phys. 16, version is typically preferred as it allows for an easier re-
143 (2024). distribution of Schmidt coefficients in the different tensor
[70] J. Haegeman, K. Van Acoleyen, N. Schuch, J. I. Cirac,
blocks.
and F. Verstraete, Gauging quantum states: From global
to local symmetries in many-body systems, Phys. Rev. X Typically, MPSs are taken to span a subspace of a
5, 011024 (2015). tensor product space. But, an important feature of the
[71] C. Delcamp, Tensor network approach to electromagnetic models we consider is that they are typically not defined
duality in (3+1)d topological gauge models, JHEP 08, on a tensor product space. Rather, states need to satisfy
149, arXiv:2112.08324 [cond-mat.str-el]. some local kinematical constraints in the form of gener-
[72] C. Delcamp and A. Tiwari, Higher categorical symme- alised Gauß laws. Thus, we require MPSs that explicitly
tries and gauging in two-dimensional spin systems (2023), enforce these kinematical constraints. For our illustra-
arXiv:2301.01259 [hep-th]. tive example, this is accomplished by considering tensors
[73] B. Bauer, P. Corboz, R. Orús, and M. Troyer, Implement- of the form
ing global abelian symmetries in projected entangled-pair
state algorithms, Phys. Rev. B 83, 125106 (2011). (V1 3V2 ,i)
[74] S. R. White, Density-matrix algorithms for quantum
renormalization groups, Phys. Rev. B 48, 10345 (1993). 3 3
ÿ
[75] F. Verstraete, D. Porras, and J. I. Cirac, Density matrix A ≡ (V1 ,d1 ) (V2 ,d2 ) (A1)
A
renormalization group and periodic boundary conditions: V1 ,V2
A quantum information perspective, Phys. Rev. Lett. 93, d1 ,d2 ,i
227205 (2004). · |V1 , d1 ⟩ |V1 3V2 , i⟩ ⟨V2 , d2 | ,
[76] V. Zauner-Stauber, L. Vanderstraeten, M. T. Fishman,
F. Verstraete, and J. Haegeman, Variational optimization where V1 , V2 are summed over ψ-projective irreducible
 10

representations of some subgroup H ⊆ A4 , i over basis Performing a singular value decomposition yields a fac-
vectors in the space of intertwining maps V1 b 3 → V2 , torisation of the form B V2 = M V2 ΣV2 (N V2 )† , where M
while d1 , d2 label the remaining variational degrees of and N are unitary matrices, while ΣV2 is a diagonal ma-
freedom. Notice that both entanglement degrees of free- trix. The entries of Σ are all positive and are referred
dom labelled by (V, d) and physical degrees of freedom to as the singular values of B V2 . Truncating the singular
labelled by (V1 3V2 , i) carry gauge degrees of freedom rep- values to the desired precision λmin yields the low-rank
resented by blue lines, which are shared by neighbouring approximation
physical degrees of freedom, as suggested by our graphi-
cal notation. Typically, the dimension of the space of in- λÿ
min
(V ,d i ) ∗
tertwining maps V1 b 3 → V2 depends on (V1 , V2 ), which [B V2 ](V13 ,d13 i12 ) ≈ [M V2 ]k(V1 ,d1 i1 ) [ΣV2 ]kk [N V2 ]k(V3 ,d3 i2 ) .
is incompatible with a tensor product Hilbert space. By k=1
construction, the action of the Hamiltonian (5) leaves the
constrained Hilbert space invariant, and thus explicitly Repeating this operation for all ψ-projective representa-
preserves the structure of such an MPS. For the case of an tions V2 of A, one finally defines MPS tensors Ai and
on-site symmetry described by a finite group G, the ten- Ai+1
sors that build up the dual Hamiltonians are be obtained (V1 3V2 ,i1 ) (V2 ,k)
as Racah W-coefficients of linear representations of G [Ai ](V1 ,d1 ) := [M V2 ]k(V1 ,d1 i1 )
with (projective) representations of a subgroup H, which (V 3V3 ,i2 ) (V3 ,d3 ) ∗ (A6)
and [Ai+12 ](V2 ,k) := [ΣV2 ]kk [N V2 ]k(V3 ,d3 i2 ) ,
can readily be computed by restricting G-representations
to the subgroup.
Our algorithm proceeds like the standard two-site respectively , so that (A4) is approximated by
DMRG, but all the basic operations are tailored to
preserve the kinematical constraints, which amounts to (V1 3V2 ,i1 ) (V2 3V3 ,i2 )

maintaining the structure displayed in eq. (A1). First 3 3

of all, by using block-diagonal basis transformations on
the entanglement space, any MPS of the form (A1) can (V1 ,d1 ) Ai Ai+1 (V3 ,d3 ) . (A7)
be brought into the left canonical form defined by the
condition: We can then repeat the same steps for the sites i + 1 and
i + 2, and keep on sweeping from left to right and then
Āi from right to left.
!
= . (A2) In our simulations, we initialise the bulk of the MPS
3
with random matrices of a given dimension per block V2 ,
Ai while on the boundary we restrict ourselves to a single
one-dimensional block. This corresponds to a choice of
In left canonical form, the Schmidt values λ of the re- boundary condition for the MPO transmuting the Hamil-
duced density matrix obtained by tracing out all the sites tonian of the model we are simulating into the initial one.
to the right of the site i are then given by the spectrum In the very specific cases where our algorithm boils to
of ρi defined as the standard symmetry-preserving DMRG, this amounts
to the customary fixing of the total charge sector of
Āi Āi+1 Āi+2 the state. Finally, note that our approach is not spe-
ρi ≡ . (A3) cific to the two-site DMRG. In particular, uniform MPS
3 3 3
algorithms for infinite chains (including VUMPS and
Ai Ai+1 Ai+2 ‘pulling-through’ algorithms) can also be implemented
[76, 77].
As reviewed above, a crucial step of the two-site DMRG
algorithm amounts to decomposing the two-site MPS
tensor solving the combined least-squares problem into Sec. B | Symmetries in tensor networks
single-site MPS tensors A. Specifically, consider a tensor
whose entries are of the form
Consider a Hamiltonian with an ordinary symmetry en-
(V1 3V2 ,i1 ) (V2 3V3 ,i2 ) coded into a (finite) group G. Suppose the Hamiltonian
is in the G symmetric phase. We claim that the optimal
3 3
way of simulating this phase via the DMRG algorithm is
(V1 ,d1 ) B (V3 ,d3 ) . (A4) to simulate the dual model obtained by gauging G, be-
fore acting with the MPO transmuting the Hamiltonian
Keeping the gauge degree of freedom V2 fixed, one con- of the dual model into the initial one. As commented in
siders the matrix B V2 with entries the main text, the building blocks of the MPO for this
duality evaluate to Clebsch–Gordan coefficients. More
(V ,d i ) (V ,d )
[B V2 ](V31 ,d31 i21 ) := [B (V1 3V2 ,i1 )(V2 3V3 ,i2 ) ](V31 ,d31 ) . (A5) precisely, in the case of the model (2), where the local
 11

Hilbert space is spanned by |3, v⟩, v = 1, . . . , 3, we have recoupling theory for symmetric tensors based on fusion
the following identification: trees.
As commented in the main text, this decomposition
v v is tailored to the symmetric phase, for which entangle-
3 3
ment degrees of freedom of the unique ground state trans-
form as linear representations of A4 . In contrast, this is
v1 V1 i V2 v2 ≡ (V1 ,v1 ) (V2 ,v2 ) , (B1) clearly not suited to the A4 SPT phase, for which en-
3 tanglement degrees of freedom transform as projective
i
representations of A4 . Indeed, it is well known that using
standard symmetric tensor networks in an SPT phase is
whereby the MPO tensor acts on the space of intertwin- more costly because it forces the edge modes to trans-
ing maps V1 b 3 → V2 . Furthermore, we commented in form according to a linear representation, which requires
the main text that in this very specific case, our pro- additional long-range entanglement.
cedure amounts to directly simulating the initial phase
using symmetry-preserving tensor networks. Generally,
Sec. C | Mathematical formalism
symmetry-preserving tensor network algorithms exploit
a specific expression for the tensors that explicitly en-
forces the symmetry. Concretely, consider an MPS in We sketch here the mathematical formalism underpin-
the Hilbert space of the model (5). Generically, entan- ning the results presented in the main text. We en-
glement degrees of freedom of the MPS tensor live in courage the reader to consult ref. [33] for detailed def-
a vector space U . Let us suppose that the MPS ten- initions. Consider any local one-dimensional quantum
sors are invariant under the action of A4 . As already lattice model with a generalised symmetry encoded into
exploited in the main text, the Wigner–Eckart theorem a fusion category C. In the presence of such a generalised
stipulates that the tensors are expressible as linear combi- symmetry, gapped phases are characterised by a choice
nation of Clebsch–Gordan coefficients. More concretely, of (indecomposable) C-module category, whose objects la-
since the vector space U is equipped with an A4 ac- bel the degenerate ground states of the phase [27, 40].
tion, it can be decomposed into irreducible representa- The same module categories also characterise the differ-
tions, i.e. U ∼ ent ways to gauge (sub)symmetries of the model. After
À
= V ⟨U, V ⟩V , where the direct sum is
over irreducible representations of A4 and ⟨V, U ⟩ ∈ Z≥0 . performing the gauging operation associated with a C-
It follows that we can decompose u = 1, . . . , dim U as module category R, the dual symmetry of the resulting
∗
(V, v, d) ≡ (V, v) b (V, d), where v = 1, . . . , dim V and model is encoded into the so-called Morita dual CR of
∗
d = 1, . . . , ⟨U, V ⟩. Using this notation, the MPS tensors C with respect to R. The fusion category CR is defined
decompose as follows: to be the category FunC (R, R) of C-module endofunctors
of R [33, 78], the fusion structure being provided by the
v composition of C-module functors. Crucially, R is also a
∗ ∗ ∗
CR module category and we have (CR )R ≃ C. In words,
v 3 ∗
there is always a way to gauge a subsymmetry of CR so
3
v1 V1 i V2 v2 as to recover the initial model.
ÿ Let us examine this gauging operation in practice. For
u1 u2 ≡ (V1 ,d1 ) i (V2 ,d2 ) ,
conciseness, we focus on nearest-neighbour Hamiltoni-
i
(B2) ans, but longer range interactions can be accommodated
revealing in particular the sparse block structure of the just as easily. As for the example studied in the main
tensors. As per our graphical calculus, the matrices on text, it is crucial to write the Hamiltonian in such a
the r.h.s. labelled by i are in fact of the form (A1). way that the generalised symmetry C is manifest. Under
From a symmetric tensor network viewpoint, this de- some mild mathematical assumptions about the symme-
composition is used to target a specific charge sector of try MPOs [36, 79], it follows from a generalised Wigner–
the Hilbert space, thereby reducing computational costs, Eckart theorem [37] that any local symmetric operator
while explicitly enforcing the symmetry [53–55]. Let us is expressible in terms of generalised Clebsch-Gordan co-
now assume that the MPS is the unique ground state of efficients. Specifically, given a Hamiltonian of the form
řL−1 ř
the A4 symmetric phase. In this case, eq. (B2) precisely H = i=1 n hi,n , the local operators can always be put
recovers eq. (8) under the identification (B1) such that in the form
the entanglement degrees of freedom of the MPS ground
Y1 Y2
state of the dual model are labelled by pairs (V, d); our al- ÿÿ
gorithm then produces results that agree with symmetry- hi,n ≡ hn ({Y }, i, j) i Y5 j , (C1)
preserving DMRG. However, when comparing our algo- {Y } i,j
Y3 Y4
rithm to current state-of-the-art implementations, our
approach turns out to be practically much simpler since it in terms of tensors evaluating to the generalised Clebsch-
does not require the conventional implementation of the Gordan coefficients. The graphical notation mimics that
 12

∗ ∗ ∗
of eq. (5) and encodes in particular the fact that the C R CR CP Q CR
Hilbert space of a model with a generalised symmetry is
generically not a tensor product space. More concretely, P Q → ∗
CP Q
there is a (possibly not unique) choice of C-module cat-
egory R such that local operators can be expressed as ∗
CP ∗
CP
∗
(C1) where {Y } label objects in CR , and the generalised
Clesbch-Gordan coefficients are given by the so-called Figure 4. The MPO representation of the symmetry C is de-
∗
module associator of R, as a CR -module category (c.f. termined by a choice of module category R, which in turn
[36, 38, 51] for details). fixes a fusion category CR∗
that governs the algebra of sym-
∗
It follows from the defining axioms of the CR -module metric Hamiltonians. The phase of this Hamiltonian is given
category R that the tensors in eq. (C1) satisfy an ana- by a module category P over C, which determines the fusion
logue to eq. (3): category CP∗ describing the quasiparticle excitations. Combin-
ing these two module categories we obtain Q = FunC (R, P),
which describes the entanglement degrees of freedom of the
Y2 Y2 optimal tensor network description of the ground state. In-
i k Y deed, by dualising and replacing R by Q, we end up in the
Y1 Y5 Y1 3 ∗
j
Y3 Y6 maximal symmetry breaking phase CP of the dual symmetry
l ∗
ÿÿ
Y6 ,kl CP .
Y4
= FYY41 Y2 Y3 Y5 ,ij Y4
,
Y6 k,l
(C2) category is given by the symmetry fusion category itself.

Y6 ,kl ′
where the ‘F -symbols’ FYY41 Y2 Y3 Y5 ,ij ∈ C enters the defi- Given P ≃ FunCR ∗ (R, Q), this is achieved by choosing R

nition of the fusion category CR ∗

. Together with the com- to be FunC (R, P) ≃ Q. The symmetry of this optimal
∗
plex coefficients {hn }n , these F -symbols govern the al- dual model is then given by CP , which also encodes the
gebra generated by local symmetric operators {hi,n }i,n . quasiparticle excitations.
Within this framework, performing a gauging operation For many fusion categories of interest, all possible
simply amounts to picking a different CR ∗
-module cate- module categories — solutions to eq. (C2) —have been
′
gory R . This means replacing the tensors in eq. (C1) obtained. Importantly, the data of the module functors
by a new set of tensors that now evaluate to generalised needed to describe the MPO intertwiners relating the dif-
Clebsch–Gordan coefficients given by the module associ- ferent dual models can be obtained as a representation
ator of R′ . Crucially, this new set of tensors still sat- theoretic problem [37, 80], that numerically can be re-
isfy eq. (C2), so the algebra of local symmetric opera- duced to a linear algebra problem. This MPO intertwiner
tors generated by (C1) is isomorphic to the initial one. allows us to relate the ground state of the optimal dual
The dual symmetry of the resulting model is then en- model defined with respect to the module category Q to
coded into the fusion category (CR ∗ ∗
)R′ . Similarly to the the ground state of the initial model defined with respect
examples discussed in the main text, Hamiltonians asso- to the module category R:
∗
ciated with different choices of CR -module categories can
R R R R
be transmuted into each other via an MPO. Mathemat-
ically, such an operator is described by an object in the P
′ ∗
∗ (R, R ) of C -module functors from R to
category FunCR R
Q Q Q Q
R′ , in such a way that the building blocks of the MPO
are provided by the module structure of such a functor Ai−1 Ai Ai+1 , (C3)
[51].
Let us now suppose that the initial model, which is where the individual tensors of which are determined
∗
defined with respect to the CR -module category R, is by the data of the category of module functors P =
in the phase associated with the C-module category P. FunC (R, Q) [51]. In this way, we obtain the optimal
Our goal is to find a dual model whose dual symme- tensor network description of the original ground state.
try is completely broken in the ground state subspace. The various fusion categories and the Morita equivalences
To achieve this, we must understand how to relate the between them are summarised in the commutative dia-
phase of the dual model to the phase of the initial model, grams of figure 4.
which is not immediate given that the symmetry struc- To conclude, we revisit our example in light of this gen-
∗ ∗
tures differ. It follows from (CR )R ≃ C that every C- eral formalism. When dealing with an ordinary (invert-
module category P is of the form FunCR ∗ (R, Q) for some ible) symmetry A4 , the corresponding fusion category C
∗
CR -module category Q, the module structure being pro- is the category VecA4 of A4 -graded vector spaces. The
∗
vided by the composition of CR -module functors. Con- different ways to gauge subsymmetries of A4 are pro-
∗
versely, the C-module category P uniquely specifies a CR - vided by VecA4 -module categories, which are known to
module category, namely FunC (R, P). The optimal dual be classified by pairs (A, [ψ]) as defined in the main text
model whose dual symmetry is completely broken in the [58]. In particular, choosing A = G and ψ = 1 amounts
ground state subspace is obtained whenever the module to the (untwisted) gauging of A, and the corresponding
 13

VecA4 -module category is equivalent to the category Vec light on some of the suboptimal simulations: For in-
of complex vector spaces. The Morita dual (VecA4 )∗Vec of stance, consider the A4 -symmetric phase and the dual
VecA4 with respect to Vec can be checked to be equiva- model labelled by R′ = Repψ (D2 ); the symmetry is
lent to the category Rep(A4 ) of representations of A4 , in found to be (Rep(A4 ))∗Repψ (D2 ) ≃ VecA4 , while the dual
agreement with our results. When writing the Hamilto- phase is that associated with the VecA4 -module category
nian as in eq. (1), we are choosing the Rep(A4 )-module FunRep(A4 ) (Repψ (D2 ), Rep(A4 )) ≃ Vec; together with the
category R = Vec such that the module associator boils
fact that Repψ (D2 ) is equivalent to Vec as category, this
down to the ordinary Clebsch–Gordan coefficients of A4 .
explains why the numerical results were the same for this
We obtain the various dual models by choosing differ-
dual model as for the initial one. In a similar vein, in the
ent Rep(A4 )-module categories. Specifically, the dual
A4 SPT phase, the dual model obtained by choosing R′ =
model resulting for the ψ-twisted gauging of H is ob-
Rep(D2 ) has a (Rep(A4 ))∗Rep(D2 ) ≃ VecA4 symmetry; the
tained by choosing the module category R′ = Repψ (H)
of ψ-projective representations of H, and the dual sym- dual phase is associated with the VecA4 -module category
metry is encoded into (Rep(A4 ))∗Repψ (H) . FunRep(A4 ) (Rep(D2 ), Repψ (A4 )), which is equivalent to Vec
as a category, meaning that the whole symmetry is pre-
We now suppose that the initial model is in
served, in agreement with our numerical results. Finally,
the phase associated with the VecA4 -module category
let us examine the D2 symmetric phase and the dual
FunRep(A4 ) (Vec, Repϕ (K)). The A4 SPT, A4 symmet-
model obtained by choosing the VecA4 -module category
ric, and D2 symmetric phases considered in the main
R′ = Rep(Z3 ); the dual symmetry (Rep(A4 ))∗Rep(Z3 ) ≃
text are obtained by choosing Repϕ (K) to be equal
Rep(A4 ) is also fully preserved since the module category
to Repψ (A4 ), Rep(A4 ), and Rep(D2 ), respectively. The
over it is FunRep(A4 ) (Rep(Z3 ), Rep(D2 )), which happens
optimal model is always found to be that given by
to be equivalent to Vec, in agreement with our numerical
R′ = Repϕ (K) = Q, which amounts to performing a
results.
ϕ-twisted gauging of K, as predicted. Let us also shed