Tri-unitary quantum circuits

Cheryne Jonay, Vedika Khemani, and Matteo Ippoliti

Department of Physics, Stanford University, Stanford, CA 94305
(Dated: October 26, 2021)
We introduce a novel class of quantum circuits that are unitary along three distinct “arrows of
time”. These dynamics share some of the analytical tractability of “dual-unitary” circuits, while
exhibiting distinctive and richer phenomenology. We find that two-point correlations in these dy-
namics are strictly confined to three directions in (1 + 1)-dimensional spacetime – the two light cone
edges, δx = ±vδt, and the static worldline δx = 0. Along these directions, correlation functions
are obtained exactly in terms of quantum channels built from the individual gates that make up
the circuit. We prove that, for a class of initial states, entanglement grows at the maximum al-
lowed speed up to an entropy density of at least one half of the thermal value, at which point it
becomes model-dependent. Finally, we extend our circuit construction to 2 + 1 dimensions, where
arXiv:2106.07686v2 [quant-ph] 23 Oct 2021

two-point correlation functions are confined to the one-dimensional edges of a tetrahedral light cone
– a subdimensional propagation of information reminiscent of “fractonic” physics.

I. INTRODUCTION of these models, which only allows correlations between

points at “light-like” separations18,19 , as can be seen from
The dynamics of quantum many-body systems play Fig. 1(a).
a central role in many areas of physics, from non- In this work, we introduce a new class of minimal cir-
equilibrium statistical mechanics to applied quantum in- cuit models which we call “tri-unitary”. They are a family
formation science. Spatiotemporal correlations of lo- of lattice models in one and two spatial dimensions for
cal operators are among the most useful physical char- which the dynamics are unitary along three space-time
acterizations of such systems: they diagnose the ap- directions, as sketched (for the one-dimensional case) in
proach to equilibrium1 (or lack thereof2 ), encode trans- Fig. 1(b). These models allow us to extend and gen-
port coefficients3 , and are more readily measurable in ex- eralize results obtained on “dual-unitary” models. The
periments than global properties such as entanglement. phenomenology we uncover is richer in several ways: in
However, they are notoriously hard to compute for gen- one dimension, correlations are pinned to three, not two,
eral interacting many-body systems. possible directions in spacetime; while two of these di-
Until recently, spatiotemporal correlations could be rections are the edges of a light cone, i.e. “maximum-
calculated exactly only in the presence of integrability3,4 , velocity” worldlines just like in the dual-unitary case, the
where the evolution is constrained by extensively many third one is the static worldline, δx = 0. This means that
conservation laws. On the opposite end of the spectrum, it is possible for information to remain “stuck” in place in
much analytical progress on entanglement, thermaliza- tri-unitary circuits. In two dimensions, the phenomenol-
tion, quantum information scrambling, quantum chaos, ogy is even more distinctive: correlations are pinned to
and the emergence of hydrodynamics has been achieved three rays in (2 + 1)-dimensional spacetime, namely they
via random circuit models5–16 . The usefulness of ran- propagate at maximum velocity along three special di-
dom unitary circuits stems from the fact that they retain rections in two-dimensional space, while vanishing every-
only the absolutely fundamental features of quantum dy- where else.
namics: unitarity and spatial locality. Additional struc- Tri-unitary circuits have the property of remaining uni-
ture can then be reintroduced in controlled ways, for in- tary under six-fold rotations of (1+1)-dimensional space-
stance through symmetries and conservation laws10,11,17 , time. Rather than swapping space and time (“spacetime
or temporal periodicity and the associated set of (Flo- duality”)34–38 , these rotations mix the two nontrivially.
quet) eigenvalues and eigenvectors. However, random cir- The ensuing dynamics is thus even more “agnostic” about
cuit models are (by design) better suited to the study of the roles of space and time. This points to interest-
“locally averaged” quantities, such as entanglement, than ing connections with recent ideas about the emergence
of spatiotemporally-resolved ones such as correlators. of spacetime from tensor networks39–41 , and with holo-
It is only recently that a class of local, interacting cir- graphic quantum error correcting codes – tensor network
cuit models was proposed where correlation functions can architectures that have been recently employed as toy
be calculated analytically18,19 . The key feature of those models of the AdS/CFT correspondence42 . In both tri-
models is that the dynamics are unitary in the space unitary circuits and holographic codes, one has a highly
direction, as well as the time direction, as sketched in isotropic tensor network that is agnostic about the role
Fig. 1(a). Such “dual-unitary” models have been studied of space and time. The key distinction between the two
extensively in recent years8,18–33 , leading to a plethora is that, while holographic codes employ “perfect tensors”
of exact results on key aspects of quantum many-body that always hide information in maximally non-local de-
dynamics. In particular, the study of correlations is grees of freedom, the elementary gates in our tri-unitary
drastically simplified due to the peculiar causal structure circuits allow for information to remain localized and ac-
 2

(a) dual-unitary t x̃
o1 o2 o1 o2

(a) Uio11,i,o22 = = = Ũoo12,i,i12

i1 i2 i1 i2
x t̃

(b) tri-unitary (b) U U † = = = = =1

(c) Ũ Ũ † = = = = =1

FIG. 1. Causality structure of dual- and tri-unitary circuits.

FIG. 2. Dual unitarity is a spacetime duality given by the
(a) Dual-unitary circuits have strict causal light cones un-
index reshuffling depicted in a). In b) (c)) we show how the
der two orthogonal arrows of time, t (left) and t̃ (right). As
contraction of U (Ũ ) with its Hermitian conjugate simplifies
a result, correlations vanish everywhere except on the light
to an identity in the folded picture.
rays δx = ±δt. (b) Tri-unitary circuits have strict causal
light cones under three arrows of time, t (left), t̃ (center) and
t̆ (right) at 2π/3 angles with each other. As a result, cor-
relations vanish everywhere except on the intersection of all II. REVIEW OF DUAL-UNITARY DYNAMICS
causally-allowed regions: the light rays δx = ±vδt and the
static worldline δx = 0.
We begin this review section by introducing dual-
unitary gates, which serve as the elementary building
blocks of dual-unitary circuits. These are two-qubit uni-
cessible, but only along certain directions. tary gates with the special property of being unitary
More concretely, the advent of digital quantum simula- when read “sideways”, i.e. as a space evolution. More
tors is making it possible to engineer time evolutions that precisely, we can associate a “spacetime dual” gate Ũ to
realize target tensor networks by specific sequences of each gate U via a reshuffling of indices, Ũio11io22 = Uii12oo12 ,
unitary gates and, potentially, projective measurements. see Fig. 2(a). Then, dual-unitary gates must satisfy
This flexibility allows one to explore a wide range of U U † = U † U = 1, as well as Ũ Ũ † = Ũ † Ũ = 1. Those
unitary43–45 and non-unitary37,46–50 dynamics, including identities are depicted pictorially in Fig.2(b-c). As in
ones obtained by rotating the “arrow of time” 18,19,34–36,38 . Ref. [18], we use the “folded” picture, where the unitary U
Our work thus adds a new direction to this program, by is overlaid with its adjoint U † ; we also denote contraction
broadening the possibilities for causal structure in tensor with an identity matrix by an open circle ◦ terminating
networks that are realizable dynamically. We note that a folded tensor’s leg.
tri-unitary circuits have the maximal number of unitary Arranging dual-unitary gates in a brickwork pat-
“arrows of time” that can be embedded in flat (1 + 1)- tern in (1 + 1)-dimensional spacetime yields a dual-
dimensional spacetime, and thus represent in a sense the unitary circuit. These circuits realize examples of
“most isotropic” architecture for tensor networks that are “maximally chaotic” dynamics and allow considerable
realizable dynamically in one spatial dimension. In addi- analytical tractability: exact results have been ob-
tion, (2 + 1)-dimensional tri-unitary circuits extend this tained for their spectral form factor8,29,30 , state21,25
program to three-dimensional tensor networks, a much and operator22,23,31 entanglement growth, two-point
less explored area. correlation functions18,19,24,26 and out-of-time-ordered
The paper is structured as follows. In Sec. II we re- correlators27 . A remarkable physical property of these
view the definition and properties of dual-unitary cir- circuits, underlying several of the above results, stems
cuits, which will serve as a starting point for the following from their causal structure. Conventionally, unitarity
discussion. We introduce tri-unitary circuits (in one spa- and the strict geometric locality of the circuit structure
tial dimension) in Sec. III, and analytically derive their forbid correlations between points displaced by a space-
correlation functions in Sec. IV. We then prove an ex- time vector (x, t) if |x| > |t|, i.e. if the two points lie out-
act result about entanglement growth in these circuits in side each other’s (past or future) light cone. By the same
Sec. V, and extend our construction to two spatial di- token however, in dual-unitary circuits one can view the
mensions in Sec. VI. We conclude by summarizing our evolution “sideways” and rule out correlations if |t| > |x|
results and pointing to directions for future research in (in a sufficiently large system). As a result, correlations
Sec. VII. are only possible on the light cone’s boundary, |x| = |t|.
 3

So any correlations must propagate exactly at the speed (a) (b)

of light.
To set up the discussion in the following sections,
⋯ ⋯⋯ ⋯
it is helpful to review the behavior of two-point func-
tions in dual-unitary circuits in greater detail. Infinite- t t
temperature two-point functions obey18,19
1
C ab (x, t) = Tr(a0 (t)bx (0))
2L x x
1 X

= L δx,µt Tr Mµt (a0 )bx , (1)
2 µ=±1 FIG. 3. Tri-unitary circuit layout. (a) Tri-untiary gates
(hexagons) are arranged at the sites of a triangular lattice
where the M± are “transfer matrices” (quantum chan- in spacetime, with manifest six-fold rotational symmetry. (b)
nels) which move an operator along either the left or Same circuit drawn with rectangular gates and vertical qubit
right fronts of the light cone, and are obtained directly worldlines.
from the gates U that make up the circuits51 . (Here one
unit of t is defined as half of a brickwork layer.) Cor- “arrows of time”, they remain unitary under four-fold ro-
relations are then obtained analytically by diagonalizing tations of spacetime. It is thus interesting to general-
the single-qubit quantum channels M± . Since these are ize this idea and expand the domain of nontrivial many-
trace-preserving and unital (i.e. M± (1) = 1), one only body quantum dynamics amenable to analytical treat-
has to consider the traceless subspace spanned by X, Y ment via rotations in spacetime. To this end, we intro-
and Z Pauli matrices; denoting the three eigenvalues in duce a new class of “tri-unitary” circuits, which admit
this subspace by {λ±,i : i = 1, 2, 3}, for traceless opera- three unitary arrows of time and remain unitary under
tors a, b, we have six-fold rotations of (1+1)-dimensional spacetime. In this
X X Section we first present the architecture of such circuits
C ab (x, t) = δx,µt cab t
µ,i λµ,i , (2) in (1 + 1) dimensions, then the tri-unitarity condition on
µ=±1 i=1,2,3
individual gates, and finally a realization of a family of
tri-unitary circuits as time-dependent local Hamiltonians
where the coefficients cab are overlaps between the oper- of the kicked Ising type.
ators a, b and the eigenmodes of M± . This places strong
constraints on the time dependence of correlators. In par-
ticular, depending on the eigenvalues λµ,i , one has the fol- A. Tri-unitary circuits
lowing possible types of behavior18,28 : (i) non-interacting
(all eigenvalues are 1, correlations are constant on the
Much like dual-unitary circuits are built from dual-
light cone); (ii) interacting non-ergodic (some but not all
unitary gates arranged on a square lattice in spacetime,
eigenvalues are 1, some correlations are constant while
tri-unitary circuits require their own special set of gates
others decay); (iii) ergodic non-mixing (none of the eigen-
arranged on a triangular lattice in spacetime. This is
values are 1, but there is at least one unimodular eigen-
shown in Fig. 3(a), where the hexagons represent three-
value, some correlations oscillate indefinitely); and (iv)
qubit gates (constraints on these gates are discussed in
ergodic and mixing (all |λµ,i | < 1, all correlations decay
the following). We note that this construction does
exponentially).
not extend beyond three arrows of time, at least in flat
We conclude this review by recalling the general
Minkowski spacetime: there are no Bravais lattices with
parametrization of dual-unitary gates on two qubits,
coordination higher than 6, and thus no natural ways of
Ud.u. [φ] ≡ u1 u2 · SWAP1,2 · CP1,2 (φ) · v1 v2 , (3) building “n-unitary circuits” for n > 3.
Additionally, Fig. 3(b) shows the same circuit archi-
where u1,2 , v1,2 ∈ SU (2) are arbitrary single-qubit gates tecture with more familiar rectangular gates and vertical
(tensor product symbols are omitted) and CP(φ) ≡ qubit worldlines, clarifying the layout for practical imple-
φ mentations. Notice there is a spatial unit cell of 4 qubits,
e−i 4 (Z1 −1)(Z2 −1) is a controlled-phase gate (despite the
with an important difference between even qubits (which
slightly different notation, the parametrization in Eq. (3)
take part in all the interactions) and odd ones (which
is equivalent to the one in Ref. [18]).
skip every other layer). Formally, a brick-work layer of
the resulting unitary circuit is given by Uo Ue , with
III. TRI-UNITARITY O
Ue = (Ux,x+1,x+2 ⊗ 1x+3 ), Uo = T2 Ue (T† )2 , (4)
x∈4Z
As we have seen, the usefulness of dual-unitary cir-
cuits for studying quantum dynamics largely stems from where U is a three-qubit gate and T is the one-site trans-
their unusual causal structure: having two equally valid lation operator. Notice that in each layer, one out of
 4 4

5 in fact, one can doa even

to construct more:
“perfect it is known
tensor”, that thetensor
i.e. a 6-qubit 5- such
4 6 qubit perfect quantum error-correcting code53,54 can be
that any bipartition of its legs into 3 inputs and 3 out-
(a)(a) U = = used to construct a “perfect tensor”, i.e. a 6-qubit ten- 42
{⇡/3, 2⇡/3} puts that
sor such (not anynecessarily
bipartitioncontiguous)
of its legsyields
into a3 unitary
inputs gate .
and There are(not10 such partitions; by contrast, tri-unitarity
1 3
3 outputs necessarily contiguous) yields a uni-
tary only
gate constraints
. There arethree bipartitions.
partitions; Perfect tensors play
2
42
10 such by contrast,
40,42,55
an important
tri-unitarity role in toy
only constraints threemodels of holography
bipartitions. Perfect ,
(b) (b) U U † = = = =1 where
tensors (alsosuchknown highly isotropic maximally
as “absolutely tensors are arranged on a
entangled
lattice
states” 55–57
)inplay
curved space; this
an important results
role in toyinmodelsa quantumof error
40,42,58
holography
correcting , where
code thatsuchencodeshighlybulkisotropic
spacetimetensors degrees of
are arranged
freedom on intoa the
lattice in curved
edge, space; this
in a manner results of the
reminiscent
(c) (c) Ũ Ũ † = = = =1 in a AdS/CFT
quantum error correcting codeWhile that encodes
correspondence. perfect bulktensors can
spacetime degrees viewed
be naturally of freedom into the edge,
as quantum in a man- codes,
error-correcting
ner reminiscent of the AdS/CFT
generic tri-unitary gates can correspondence.
be viewed While as “defective”
perfect tensors can be naturally viewed as quantum error-
quantum codes: the encoded qubit is vulnerable to er-
correcting codes, generic tri-unitary gates can be viewed
(d) (d) Ŭ Ŭ † = = = rors on aquantum specific codes:
physical thequbit.
encodedThus, qubit crucially, some
=1
as “defective” is vul-
of the encoded information can be
nerable to errors on a specific physical qubit. Thus, cru- accessed locally, but
only along a special direction. As
cially, some of the encoded information can be accessed we shall see in Sec. IV,
FIG.
FIG. 4.4. (a) A three-qubit gate U , normally
(a) A three-qubit gate U , normally drawn as a drawn as a this
locally, geometrically
but only along aconstrained
special leakage
direction. of
As information
we shall has
rectangle (left), can
rectangle (left), canbe berepresented
represented as as a hexagon
a hexagon in order
in order to tosee in a striking
Sec. IV, effect on the behavior
this geometrically of two-point
constrained leakagecorrelations
of
facilitate spacetimerotations
facilitate spacetime rotations(right).
(right).LegsLegs
areare labeled
labeled fromfrominformation
in generic has tri-unitary
a striking effectcircuits.
on the behavior of two-
⇤
11 to 6. (b)
to 6. UnitarityofofUU
(b) Unitarity . .The
Theblue
blue hexagon
hexagon represents
represents U ∗ ; U ;point correlations
While a complete in generic characterization
tri-unitary circuits.of all tri-unitary
∗ ⇤
the pink hexagon represents U U in the
the pink hexagon represents U ⊗ U in the “folded picture”,
⌦ “folded picture”, While a complete characterization
gates is a goal for future work, here of we
all will
tri-unitary
focus on a par-
where
where each leg isis implicitly
each leg implicitlydoubled.
doubled.Open Open circles
circles representgates is a goal for future work, here we will focus on
represent
contraction with
ticularly simple family of tri-unitary gates, which are sim-
contraction with 11 operators.
operators.(c,(c,d) d)
Unitarity
Unitarityof Ũof and Ŭ : Ŭ :a particularly simple family of tri-unitary gates59 , which
Ũ and
picking different triplets of (contiguous) input legs still yieldsare simple to construct and sufficient to realize all the
picking different triplets of (contiguous) input legs still yields ple to construct and sufficient to realize all
the
relevant
a unitary operator.
a unitary operator. physics. These gates are constructed out of controlled-
relevant physics. These gates i are constructed out of
phase gates CPij (CP
controlled-phase gates
) ⌘(φ) e ≡4 (Zei−i 1)(Z
φ j 1)
between pairs of
4 (Zi −1)(Zj −1) be-
ij
tweenqubits
pairs of followed by a SWAP
qubits followed by a gate
SWAPbetween qubits 1 and
gate between
every 4 qubits is not acted upon. 3, in analogy to the dual-unitary gate parametrization in
every 4 qubits is not acted upon. qubits 1 and 3, in analogy to the dual-unitary gate
The architecture in Fig. 3(a) is manifestly invariant
The architecture in Fig. 3(a) is manifestly invariant Eq. (3):
parametrization in Eq. (3):
under six-fold rotations of spacetime, however the circuit
under
resultingsix-fold
from rotations
such rotation of spacetime,
is, in general, however the circuit
not unitary.
resulting
For this tofrombe the such case, rotation
we must is,constrain
in general, the not unitary.
choice of Ut.u.
Ut.u. [φ] =[ u1] u= 2 uu 3 1·uSWAP SWAP
2 u3 · 3,1 · CP3,1 · CP
3,1 (φ v1(v3 3 ) · v1 v3
3 ) ·3,1
For this to gates
three-qubit be theU .case, we must constrain the choice of · CP2,3· (φ CP · v(2 ·2CP
2 )2,3 ) · 1,2 · CP
v2 (φ 1 ) ·1,2
w1(w12)w·3 w , 1w 2 w3 , (6)
(6)
three-qubit gates U .
where ui , viu
where , iw, ivare arbitrary single-qubit gates acting on
i , wi are arbitrary single-qubit gates acting on
B. Tri-unitary gates qubitqubit
i (tensor product
i (tensor symbols
product are omitted).
symbols This fam-
are omitted). This fam-
B. Tri-unitary gates ily of gates is sketched in Fig. 5(a). Their
ily of gates is sketched in Fig. 5(a). Their tri-unitarity tri-unitarity
We denote the input legs as (1, 2, 3) and the output legs is made apparent by drawing them in a symmetric way,
is made apparent by drawing them in a symmetric way,
by (4, 5, 6), as shown in Fig. 4(a). The requirement of as in Fig. 5(d): a 2π/3 rotation results in a permuta-
We denote the input legs as (1, 2, 3) and the output legstion as of thein single-qubit
Fig. 5(d): gates a 2⇡/3 androtation
the φ angles, results in a permuta-
preserving
tri-unitarity is that these tensors be unitary under three tion of the single-qubit gates and the angles, preserving
by (4, 5, 6), as shown in Fig. 4(a).
distinct choices for the arrow of time, i.e., whether we The requirement ofthe family of gates as a whole; rotations by π/3 on the
tri-unitarity is that these tensors be
map qubits (1, 2, 3) 7→ (4, 5, 6) (the “conventional” arrow unitary under threeother hand produce the transpose of a gate in the family, on the
the family of gates as a whole; rotations by ⇡/3
distinct choices for(4,the
of time, Fig.4(b)), 1, 2)arrow
7→ (5, of 6, 3)time, i.e., whether
(clockwise π/3 ro- wewhich other hand
is still produce
unitary. Thus the Ũ and Ŭ of
U ,transpose areaallgate in the family,
unitary,
map
tation,qubits (1, 2, or
Fig.4(c)), 3) (5,7! 4,(4,1)5,7→ (the
6) (6, 3, 2)“conventional”
(clockwise 2π/3 arrow which
making U is still
tri-unitary. unitary. Thus
Intuitively, one
U , Ũ and
obtains Fig.
Ŭ are all unitary,
5(d)
of time, Fig.4(b)),
rotation, Fig.4(d)). (4, This selects
1, 2) 7! (5,three (clockwise
6, 3)distinct ⇡/3 ro-
contrac- by making
“sliding” the
U tri-unitary.
controlled-phase Intuitively,
gates one
past obtains
the SWAP Fig. 5(d)
tation,
tions ofFig.4(c)),
U and U or ∗
and (5,sets
4, 1)them equal
7! (6, 3, 2)the(clockwise
three-qubit2⇡/3 in Fig.
by 5(a),
“sliding”followingthe the steps
controlled-phase sketched in
gates Fig 5(b,c).
past the SWAP
identity operator.
rotation, Fig.4(d)). More specifically,
This selects threewe candistinct
define trans-
contrac- More inrigorously,
Fig. 5(a), we note
following that SWAP
the steps
1,3 CPsketched
1,3 (φ 3 ) equals
in Fig 5(b,c).
formed
tions of gates
U and Ũ and
U ⇤ Ŭ and viasets them equal the three-qubitthe dual-unitary More rigorously, gate Ud.u. we [φ 3 ] of that
note Eq. (3) SWAPacting on qubits
1,3 CP1,3 ( 3 ) equals
1 and 3; thus the vertical (yellow)
the dual-unitary gate Ud.u. [ 3 ] of Eq. (3) acting CP gate in Fig. 5(d),on qubits
identity operator. More specifically, we can define trans-which
Uaa14aa25aa36 = Ũaa45aa16aa23 = Ŭaa56aa43aa12 .
formed gates Ũ and Ŭ via (5) 1 and 3; thus the vertical (yellow) CP gate intheFig. 5(d),
at first glance appears ill-defined (coupling
same qubit at different times), is in fact equivalent via
which at first glance appears ill-defined (coupling the
(notice the counterclockwise rotation of
a6 aindices). Then, spacetime duality to a unitary gate acting (simultane-
U a4 a5 a6
a gate U is tri-unitary = Ũ a5 a
a1 a2 a3 if itasatisfies
6 a3
4 a1 a2
= Ŭ
allathree
3 a2
.
5 a4 a1 identities
52 (5) same
ously) on two qubit at different
distinct qubits. times), is in fact equivalent via
U U = Ũ Ũ = Ŭ Ŭ = 1, shown in Fig. 4(b-d).
† † † spacetime duality
This family of gates is specified to a unitary gate acting(nine
by 31 parameters (simultane-
(notice the counterclockwise rotation
First of all, it is natural to ask whether the conditions of indices). Then, ously) on two distinct qubits.
SU (2) gates, with three angles each, the three φ inter-
aingate
Eq. (5)U iscantri-unitary
be satisfied if at
it all.
satisfies all three
The answer identities52 action This
is positive; phases, family and of a gates
globalisphase).
specifiedWhile by 31not parameters
a full (nine
U U † = Ũ Ũ † = Ŭ Ŭ † = 1, shown in Fig. 4(b-d). SU (2) gates, with three angles each, the three inter-
First of all, it is natural to ask whether the conditions action phases, and a global phase). While not a full
in Eq. (5) can be satisfied at all. The answer is positive; in parametrization of tri-unitary gates (e.g., it does not con-
fact, one can do even more: it is known that the 5-qubit tain the perfect tensor56 ), this is nonetheless a very large
53,54
perfect quantum error-correcting code can be used space that offers a rich set of possibilities for dynamics.
 55

(a) (b) (c) (d) b J b J b J b J b

A
J00 J10 J20 J30 J40 J50 J60 J70
B
= = = b10 b30 b50 b70

FIG.
FIG. 6.6. Schematic
Schematicofofthe
thekicked
kickedIsing
Isingmodel
modelrealization
realizationofof
tri-unitary
tri-unitary dynamics.
dynamics. Circles
Circlesrepresent
representqubits, qubits,partitioned
partitioned
into
intotwo
twosublattices
sublatticesAAand
andB; B;the
theJ Jand andJ 0Jbonds
0
bondsdenote
denoteZZZZ
0
Ising
Isinginteractions;
interactions;b bandandb b0are
areon-site
on-sitetransverse
transversefields.
fields.For
For
0 0
FIG. 5.
FIG. 5. (a)
(a) Diagrammatic
Diagrammatic representation
representation of of the
the gate
gate UUt.u. [ ]
t.u. [φ]
|J|
|J| ==|b|
|b|==⇡/4 π/4(and
(andarbitrary
arbitraryb b, 0 ,J J)0 )the
theFloquet
Floquetoperator
operator
in Eq.
in Eq. (6),
(6), consisting
consisting of of two-qubit
two-qubit CP CP gates
gates (dots
(dots joined
joined Eq.
Eq.(7)
(7)isistri-unitary.
tri-unitary.
by horizontal
by horizontal lines),
lines), single-qubit
single-qubit rotations
rotations (squares),
(squares), and and aa
gate (exchange
SWAP gate
SWAP (exchange of of left
left and
and right
right qubit
qubit worldlines).
worldlines). (b,c)(b,c)
This diagram
This diagram can can be
be recast
recast inin aa manifestly
manifestly tri-unitary
tri-unitaryform formby by explore
angles, nontrivial tri-unitary
coupling the gates Ut.u.
two sublattices. [ ] of
This arbitrary
allows one to
“sliding” one-
“sliding” one- and
and two-qubit
two-qubit gatesgates past
past the SWAP. The
the SWAP. The lastlast angles, where the three arrows of time are fully on the
explore nontrivial tri-unitary gates Ut.u. [φ] of arbitrary
step (sliding
step (sliding one
one end
end of
of the
the yellow
yellow CPCP gate
gate past
past the
the SWAP)
SWAP) same footing.
φ angles, where the three arrows of time are fully on the
relies on dual-unitarity of SWAP · CP. (d) The resulting dia-
relies on dual-unitarity of SWAP · CP. (d) The resulting dia- same footing.
gram is
gram is manifestly
manifestly invariant
invariant under
under rotations
rotations of of spacetime
spacetime by by
(up to
2⇡/3 (up to aa permutation
permutation of of parameters),
parameters), thus
thus thethe gate
gate isis
2π/3 IV. CORRELATIONS IN TRI-UNITARY
tri-unitary.
tri-unitary. CIRCUITS
IV. CORRELATIONS IN TRI-UNITARY
CIRCUITS
tain the perfect tensor 56
), this gates
is nonetheless a very For concreteness we consider spatiotemporally uniform
parametrization of tri-unitary (e.g., it does notlarge
con- (i.e. clean, Floquet) tri-unitary circuits, built out of a
space that offers a rich60set of possibilities for dynamics.
tain the perfect tensor ), this is nonetheless a very large Fortri-unitary
single concreteness gate weUconsider
arrangedspatiotemporally uniform
in spacetime according
space that offers a rich set of possibilities for dynamics. (i.e.
to Fig.clean,
3(a). We Floquet)
consider tri-unitary
two-pointcircuits, builtbetween
correlations out of a
C. Kicked Ising model realization
single
two tri-unitary
traceless, gate Uoperators
single-site arranged ainand spacetime according
b, displaced by
xtosites
Fig.and
3(a). We consider
t time steps (again two-point
one unit correlations between
of t consists of a
C. Kicked Ising model realization two brickwork
half traceless, layer),
single-site operators
at infinite a and b, displaced by
temperature:
While the tri-unitary circuits constructed above would x sites and t time steps (again one unit of t consists of a
most naturally be realized on gate-based digital quantum half brickwork 1infinite temperature:
While the tri-unitary circuits constructed above would
simulators, they may also be realized as time-dependent C ablayer),
(x, t) =at L Tr(a0 (t)bx (0)) . (8)
most naturally be realized on gate-based digital quantum 2
Hamiltonians. As an example, we consider a spin chain 1
simulators, they may also be realized as time-dependent ab
. forces C ab(8)
partitioned in two sublattices, A and B, evolving un- The usual lightCcone (x,for
t) =time evolution
Tr(a0 (t)bxwith(0)) U
Hamiltonians. As an example, we consider a spin chain 2L
der a generalized kicked-Ising dynamics as follows. The to vanish if |x| > v|t|, with v = 2 (each three-qubit gate
partitioned in two sublattices, A and B, evolving un-
spin chain’s Hamiltonian alternates between
P dynamics three forms:
P follows. can
Themove usualinformation
light cone for by two
timesites). Normally
evolution with U(for U uni-
forces C ab
der a generalized kicked-Ising as The
a transverse field, HX = b n2A Xn + n2B b0n Xn ; a tary but not
to vanish if tri-unitary),
|x| > v|t|, with correlations
v = 2 (each can three-qubit
exist anywhere gate
spin chain’s Hamiltonian alternates between three forms:
next-nearestP neighbor Ising coupling
P in the PA sublattice, inside
can move the light cone; however,
information the two
by two sites). additional
Normally (fordirec-
U uni-
aHtransverse
AA
= J field,
Z H
Z X = ; b the
and Xn +interaction
n∈Asame
0
n∈B bn X ; a
HnZAA tions of unitarity ( Ũ and Ŭ ) pose further restrictions. As
next-nearest
Z n2A n
neighbor
n+2
Ising coupling in the A sublattice, tary but not tri-unitary), correlations can exist anywhere
with
AA an additional nearest-neighbor P
P Ising coupling (cou- sketched in Fig. 1, each direction of unitarity rules out
H = J n∈A ZAn Z ; and the =same J interaction HZAA inside the light cone; however, the two additional direc-
pling
Z sublattices and
n+2B), HZAB n n Zn Zn+1 . The
0
correlations in two ⇡/3 wedges of spacetime that are in
with an additional nearest-neighbor Ising coupling (cou- tions of unitarity (Ũ and Ŭ ) pose further restrictions. As
setup is summarized schematically in
PFig. 6. (In addi- the “present” relative to the arrow of time. The inter-
pling sublattices A andarbitrary = n Jn0 Znfields
B), HZABlongitudinal Zn+1h. The sketched in Fig. 1, each direction of unitarity rules out
tion one could include n Zn section of these three constraints rules out almost all of
setup correlations in two π/3 wedges of spacetime that are in
in the HZ Hamiltonians.) The Floquet unitary(In
is summarized schematically in Fig. 6. addi-
is given spacetime, leaving only the lines x = ±vt, x = 0.
tion the “present” relative to the arrow of time. The inter-
by one could include arbitrary longitudinal fields hn Zn More formally, using the tri-unitarity identities U U † =
in the HZ Hamiltonians.) The Floquet unitary is given section of these three constraints rules out almost all of
Ũ Ũ † = Ŭ Ŭ † = 1, in their diagrammatic form, Fig. 4(b-
by AA AA AB spacetime, leaving only the lines x = ±vt, x = 0.
U = e iHX e iHZ e iHX e i(HZ +HZ ) .
F (7) d), we can explicitly reduce the tensor contraction ex-
More formally,
pressing C ab†(x, t) using
and show the tri-unitarity
that it vanishesidentities U† =
away Ufrom
†
in their diagrammatic
an exampleform,with 0Fig.
< x4(b-
AA AA AB
At J =UFb =
−iHX −iHZ
= e⇡/4 (and
−iHX −i(HZ +HZ )
e for earbitrary, e .
possibly position- (7) Ũ Ũ special
these = 1, Fig.
= Ŭ Ŭ lines: 7 shows <
dependent J , b , h), the above evolution is tri-unitary.
0 0 d), we can explicitly reduce the tensor
vt, in whichabunitarity of U and Ŭ causes the correlator contraction ex-
At
ThisJ is=most π/4 (and
b = easily seen forby arbitrary,
setting J 0 = possibly
0 first:position-
in that pressing C (x, t) and show that
to vanish. Symmetrically, for vt < x < 0, one would it vanishes away from
dependent
case, sublattice A realizes a kicked-Ising is
J 0
, b 0
, h), the above evolution tri-unitary.
chain at the theseunitarity
need special lines:
of U Fig.
and Ũ7 shows
to reachan example
the samewith 0<x<
result.
This is most point,
dual-unitary easily as seen by setting
in Ref. [8], while = 0 first: B
J 0 sublattice in con-
that vt,Forin xwhich
= 0 orunitarity
x = ±vt,ofweU can andexactly
Ŭ causes the correlator
express the cor-
case,
sists ofsublattice
decoupledAsites. realizes
Thisarealizes
kicked-Ising chain atgate
the tri-unitary the to vanish.
relations Symmetrically,
in terms of iteratedfor −vt < x of
application 0, one
< one would
of three
dual-unitary
Ut.u. [ ], Eq. point,
(6), with as in 1Ref.
= [8], while sublattice B con-
2 = 0 and 3 = ⇡ (i.e. a
need unitarity of U and Ũ to reach
quantum channels, M± (for correlations along x = ±vt) the same result.
sists of decoupled
dual-unitary gate U sites. This
d.u. [⇡] realizes
on sites the3,tri-unitary
1 and while site gate
2 is andFor M0x = (for0 or x = ±vt, we
correlations can xexactly
along = 0),express
as shownthe cor-
in
untouched).
U t.u. [φ], Eq. (6), with and
Turning on 1the J2 couplings preserves
φ = φ 0 = 0 φ 3 = π (i.e.thisa relations in terms of iterated application
Fig. 8(a). This phenomenology is similar to that of one of three
of
dual-unitary
structure, butgate d.u. [π] on sites
alsoUintroduces 1 and 3,1,2
nontrivial while site 2 is
interaction quantum channels,
dual-unitary circuits,Mwhich± (foralsocorrelations along
exhibit left- andx= ±vt)
right-
untouched).
angles, coupling Turningthe twoon sublattices.
the J 0 couplings preserves
This allows onethis
to and M0correlations
moving (for correlationsalong xalong = ±vt; as shown in
x =the0),“non-moving”
structure, but also introduces nontrivial φ1,2 interaction Fig. 8(a). This phenomenology is similar to that of
 6 66 6
6 b666 6
b6
b b b
bxx bx bxx bx bx x bx x
x
bx
bx bx bx bx
b bbbxbxbx b b
bbxxxbxx = x x bxbbxx x = bxbbxx x
= = = =
= =

==
= ==
== ==
aa0 aa00 a0 a0a0 a0 a0a0 a0 (a)(a)
(a)
0 a0
a0 a0 a0 a0 (a)
bbxx bbx bbxxbbx bxbbx bx
aaa000aa00 x aa0a0a0
a 0 0
x
a0aa00a0 x a0aa00a0 (a)(a)
(a)
(a)
bbxxxbbxx bbbxbxbx
xx b b
bxbxx x
=
== =
=
= =
=
bxbxb bx bx bbx bx
x x
== =
==
= ==
= =
bxbbxxbx bxbbxxbx
aa00a aa00aa0
a0 a0aa0a0 b)b)
0 0
a0aa0a0 a0aa0a0 b)b)
(b)
0
aaa0a
FIG.
FIG.
FIG. 0a00 7.
0
7. Correlation function
Correlationfunction ha0ha
functionha (0)b
0(0)b
aa0a0a0for 0 < x < vt
(t)i
a
xx(t)i0 0for 0 < x < vt
for 0 < x < vt
0 0
(b)
FIG. 7. Correlation function ha00(0)bxx(t)i for 0 < x < vt
(0)b (t)i a0aa00a0 b)= vt.
(away
(away from special rays x x ±vt). Hexagons
Hexagons are
are “folded”
“folded” FIG.
FIG. 8.8.8.8.(a)
(a)(a) Correlation function
a0a00a0 haha
a0a0a0 function forforfor xb)
xb)b)
(a)Correlation
==
= 0, 0,
±vt). 0 (0)b x (t)i,
(away
FIG. (away 7.7.
from⇤ special rays
Correlation function =0,ha0, ±vt).
±vt).
(0)b
Hexagons
Hexagons
(t)i for 0 <
are
are x
“folded”
“folded” FIG.
FIG. Correlation
Correlation function
function 0 (0)b
ha ha x (t)i,
0 (0)b
0 (0)b x (t)i,for
x (t)i, x=x=vt. =vt.vt.
gates
FIG.
FIG.
gates
FIG.
FIG.gates
gates 7.
U ⌦U
U ⌦U , ⇤open circles
Correlation
Correlation
, open circles
are
function
function
are
function
function are
are
contractions
contractions
ha
ha ha
ha (0)b
(0)b
00(0)b
(0)b
contractions
contractions
00 0 xx (t)i
xx(t)i
x(t)ito
to
(t)i 1,
to
to for
for
for
1,
for 1,1,full
full
0000<<circles
circles
<
full
full <x xxx<
circles
circles << < <vt
are
arevt
vtvtvt
are
are Unlike
Unlike
UnlikeFig.
Unlike
Fig.
Fig. 7,7,the
Fig.7,7,thethe correlator
correlator
thecorrelator remains
remains
correlatorremains connected
connected
remainsconnected connectedand andandandis isgiven
is isgiven given
given
(away
the from special
operators. rays xx= = 0, ±vt).
equality Hexagons
follows from are “folded”
unitarity of by iteration ofCorrelation
aquantum
quantum channel, .xAnalogous
(away
(away
the
(away
(away
the
gates
a,
the a,
b
UU
from
from special
special
b operators.
⌦U ⇤
∗ ,⇤ ,open
Therays
rays first
x =
first
The first
circles are
equality
=
= 0,
0, 0,
0, ±vt).
±vt).
±vt).
±vt).
equality follows
equality
contractions
Hexagons
Hexagons
followsHexagons
Hexagons
follows
to
from
1,fromfrom
full
are
unitarity
are
are
are unitarity
unitarity
circles
“folded”
“folded”
“folded”
“folded” of
are ofof by FIG.
FIG.
FIG.
FIG.by
by 8.8.8.(a)
iteration
8.
iteration
iteration (a)(a)
(a) ofCorrelation
ofCorrelation
aof
Correlation
aaquantum
quantum function
channel,
function
functionchannel,
channel, ha
M
haha
0 M
0ha
(0)b
0
M0 .+(0)b
(0)b
+(0)b
0
M +x
x.Analogous
x
x (t)i,
(t)i,
.+(t)i, for
(t)i, for
for
x Analogous
Analogous for x=xresults
x xresults
= =vt.
=results
vt.
results vt.vt.
;U
gates;Uthe
gates second ⇤
from
, open open unitarity
circles
unitarity
circles are of
of either
arecontractions
contractions
either
contractions or
ortoto ;Ŭ;Ŭ ;the
;full
the
full last
circles
last
circles from
from are
are would
Unlike
would be
be obtained
Fig. 7,7,
obtained the for
forcorrelator vtand
remains
and connected0(in terms
and of
isgiven
ofgiven
gates
gates are
are contractions toto full
full circles
circles areare Unlike
Unlike Fig.
Fig. 7, the
the correlator remains =0=00(in
connected
connected terms
and
and isof
isgiven
is given
U
U U
;a,the ⌦U
⊗U⌦U
second from unitarity
unitarity of
of either
either ŨŨŨ
Ũ or
or Ŭ
Ŭ1,
1,1, 1, the
the last
last from
from Unlike
would Fig. be 7,obtained
the correlator
xxx=x=
for remains
vtvt xconnected
and xx
=x = and
(in terms M M
the b operators. The first equality follows from unitarity of would be obtained for == vt and (in terms of MM
unitarity
the
unitarity
the
the
the unitarity
unitarity ofofŬ .Ŭ . Thus
operators.
a,a,b boperators. Thethe
The
Thus first
first
first
first
the correlator
equality
thecorrelator
equality
equality
equality
correlator
correlator becomes
follows
becomes
follows
follows
follows becomes
becomes from
from
from
from disconnected
unitarity
disconnected
unitarity
unitarity
unitarity
disconnected
disconnected of ofof
of and
and
by by
by
byand iteration
iteration
iteration
iteration
M M respectively).
ofof
0 respectively).
ofof aquantum
quantum
aaquantum
respectively). (b)
(b) The
The
channel,
channel,
(b) The quantum
channel,
quantum
MM M
quantumM .++.+ .channels
Analogous
.channels
Analogous
Analogous
Analogous
channels MM ,, M
results
,results
results
resultsM ,0 M
and 0M
M 0 0 respectively). (b) The quantum +
+ + channels M M M 0
;; ;the
the second from unitarity of either or ;;; ;the last from
UU Ũ Ŭ 0 0
(and
(and
U
U U ;(and
(and vanishes
the secondfor
second
vanishes from traceless
from
for unitarity
unitarity
unitarity
unitarity
traceless a,a,
a, of
of ofb).
of
b).
b). InIn
In
either
either
either
either In the
the Ũ Ũ
Ũ
Ũ
the reflected
reflected
or
oror
orŬŬ
Ŭ Ŭ
reflected
the reflected ;
the (x
the
the
the (x !
7
(x last
last
last
last !
7
(x 7! x) from from
from
fromx)
x) and
and would
would
would
would M
andM
and M be
be :
be:
be the
the
obtained
M++: : the diagrams
diagrams
obtained
obtained
obtained
thediagrams for for
for x
diagramsshowxshow
show
=x= = Tr
−vt
vtTr
vtand
bM
vt
showTrTrbM bM 2 2
and
and
and x(a)
xx(a)
= x
2 ==0= 0
0with
with (in
(in
0 (in
(interms
µ µ= terms
terms
terms
= of
withµ µ== (left),of
of
M(left),
(left),
ofM
M M
(left),
µ (a) with
+ 2
unitarity of ŬŬ..diagram,
Thus the a, b).
correlator becomes disconnected !
7 x) + µbMµ
µ (a)
−
version
version
unitarity
unitarity
unitarity
unitarity
version
version ofofthis
ofofthis
Ŭ . diagram,
Thusthe
Thus the
the
the the
the last
last
correlator
correlator
correlator
last
last step
step
correlator step
step would
would
becomes
becomes
becomes
becomes
would
would instead
instead disconnected
disconnected
disconnected
instead
instead require
require
disconnected require
require (center),
(center),
and
and
0and
0and00 M MMM
(center),
(center),
0 (right).
++(right).
respectively).
respectively).
0respectively).
00respectively).
+ + (right).(b)
(right). (b)(b) The
The
The quantum
quantum
quantum
quantum channels
channels
channels

channels MM MM ,−,M,M ,M
00M000 0
(and
unitarity
unitarity
(and vanishes
ofof for
Ũ .Ũfor traceless a, b). In the reflected
reflected(x (x7! x)
(and
(and
(and vanishes
vanishes .fortraceless
tracelessa, b).In InIn
Inthe the
the reflected
thereflected
reflected (x(x and
and ::the
the
:the
the diagrams
diagrams show with
with (left),
(left),
2 22 2
unitarity
unitarity traceless a,a,
a,b).b).
b). (x 7! !
7→7!7!−x) x)x)
x)x) and
and M MM M
++ :+
+ diagrams
diagrams show
show TrTrTr
bMbM
bMbM µ(a)
µ (a)
µ (a)
µ (a)with with µµµ=µ===−(left),(left),
version
version
version
version
version of
ofof this
this
this diagram,
diagram,
diagram, the
the
the
the last
last
last
last step
step
step
step would
would
would
would instead
instead
insteadinstead require
require
require require (center), (right).
the last step would instead require (center),
000(center),
0(center), ++++ (right).
(right).
(right).
unitarity
unitarityof
unitarity
unitarity ofofŨŨ. .
unitarity Ũ . ily
ily yields
yields 11⌦3 ⌦3on the three remaining legs; thus all M
⌦3on
terms
ily
ily yields
yields 11⌦3 on onthe thethethree threeremaining
three remaininglegs;
remaining legs;
legs;⌦4 thus
thusthusallallallM MµM µ
tri-unitarygate Uas
µ
terms
termsofofthe thetri-unitary gategate
gate UUU asasas (whose
(whose definition
definition involves
involves contraction
contraction with
with 1 11⌦4 ⌦4on⌦4on fourfour
µ
(whose
(whosedefinition definition involves
involves contraction
contraction withwith 1 on on fourfour
ilyily
ilyily yields
yields
yields
yields
non-contiguous
wenon-contiguous
survey 1in⌦3 ⌦3
⌦3
11more⌦3 on on on
legs) the
legs) thethe
detail three
three
reduce three
reducein Appendix remaining
remaining
remaining
remaining
to to exact
exact A. legs;
legs; legs;
legs;
erasure
erasure thus
thus thus
thus all
all
channels, all
all
channels, MM µM
M
auto-correlator
termsofofthe x = 0,1 in contrast,
gate is unique to the tri- non-contiguous
non-contiguous
along the rays while legs)
legs) otherreduce
reduce decay); toto exact
exact
by erasure
erasure
introducing channels,
channels,
generic
µµ µ
terms
dual-unitary
terms
terms the circuits,
Mtri-unitary == which
tri-unitary 1Tr gategate
gate also
UUU ††as
as
as exhibit
as left- ]] ] and right- (whose contraction with on four
⌦4
(a) [U
[U
†U†(1
(1 ⌦
⌦⌦ 111 ⌦⌦⌦ a)U
a)U (9)
(9)
(9) (whose
M M (whose
(whose
Finally
(a)
(a) = definition
= definition
definition
we
Tr(a)1/2,note
Tr(a)1/2, involves
that involves
involves
andand ifallall
U contraction
is contraction
contraction
two-point
two-point
a perfect with with
with
correlations
correlations
tensor, 1 ⌦4
11 ⌦4
1
then on
⌦4
on four
on
vanish
vanish
con- four
four
unitary setting. M (a) Tr2,3
2,3 [U
[U (1
(1 ⌦ 1 ⌦ a)U
a)U ] (9) Mµ µµ (a)
µ (a)=
single-qubit =Tr(a)1/2,
gates one and
and all
realizes alltwo-point
two-point correlations
correlations vanish
vanish
moving correlations along 4 x = ±vt; the “non-moving” M Tr(a)1/2, toall ⇤other behaviors, which
2,3
4 2,3
non-contiguous
non-contiguous
non-contiguous
non-contiguous
everywhere.
everywhere.
tracting any three legs)
legs) legs)
legs reduce
reduce in U to⌦
reduce toUexact
exactexact
exact with erasure
erasure 1⌦3 channels,
erasure
erasure channels,
channels,
channels,
necessar-
The
auto-correlator quantum channels 11 1
in M
contrast, (µ = 0, ±) are given in weeverywhere.
everywhere.
survey in more detail in Appendix A.
ais unique ]]]]]]]to the(10) tri-
†
MM0 (a) x(a)=
(a) ==0, 1Tr
= Tr Tr 2,3 [U
[U
[U
µ
[U
[U
† †
††(1
(1
†
††(1
(1(1 ⌦⌦
⌦
⌦
⌦ ⌦ 111
1a ⌦
⌦ ⌦ ⌦ ⌦ a)Ua)U
a)U
1)U a)U (9)
(9)
(9)
(10) (9) M
ily
MM M (a)(a)
yields
(a)
(a) == == 1Tr(a)1/2,
Tr(a)1/2, on the
Tr(a)1/2,
⌦3 andand
and
threeallall all two-point
two-point
two-point
two-point
remaining correlations
correlations
correlations
correlations
legs; thus all vanish
vanish vanish
vanish
M
terms ofsetting.
unitary theM tri-unitary
0 0 (a)
M =444 gate Tr2,3
2,3
1,3
1,32,3[U
1,3[U
U [U(1 as (1
(1 ⌦⌦ a⌦
a⌦ ⌦ 1)U 1)U
1)U ] (10)
(10)
µµ µ µ µ
0
4
1,3
Finally
everywhere.
everywhere.
everywhere.
everywhere. we note
(whose definition involves contraction that if U is a perfect with tensor, 1 ⌦4 then con-
on four
The quantum channels 11111Tr †(µ
M
M MM
M M
M+0(a)
0
+ 0 (a)
(a)
(a)
(a)
0 == =
(a) =4444Tr
=
= 1Tr
Tr Tr
Tr
Tr
M
1,3
1,3
1,3
1,21,3[U
[U
[U
[U [U
[U
µ
[U
††††(1
†(a
†(1⌦⌦
††(1
(1
(a ⌦=
⌦
⌦ ⌦
⌦ 1a aaa
11 1
⌦
⌦
⌦
0,
⌦⌦
⌦ ⌦⌦
±)
1)U1)U
1)U
1)U
1)U1)U
1)U ]]]are
]]]]]]
given
(11)
in
(10)
(10)
(10)
(10)
(11)
(11)
(9)
tractingV.
non-contiguous V.any
V. V.
three
legs) legs
ENTANGLEMENT
ENTANGLEMENT
ENTANGLEMENT
⊗3 ENTANGLEMENT GROWTH
reduce in U to ⊗exact U ∗GROWTHwith
GROWTH
erasure
GROWTH 1⊗3channels,necessar-
terms of the M tri-unitary
+
+ (a) = 4 44 gate
Tr1,2
2,3 1,2
1,2 U[U (1as(a(a ⌦ 1 ⌦a)U 1)U (11)
M ilyµ (a) yields 1 on the
= Tr(a)1/2, andthree all two-point remaining legs; thusvanish
correlations all Mµ
1111 V.V. ENTANGLEMENT
ENTANGLEMENT GROWTH
GROWTH
where MisMM the (a)
(a) = =
partial 11 Tr Trtrace [U [U ††††
††over
(a (a
(a⌦⌦⌦ 11sites
sites
1 1⌦⌦⌦ ⌦1)U i1)U and ]]]] j.j. These (11)
(11)
(11)
(11) (whose
Another
everywhere.
Another definition
V.V.question ENTANGLEMENTinvolves
ENTANGLEMENTthat is contraction
made GROWTH
analyticallywith
GROWTH 1 ⊗4
on
tractable four
Anotherquestion thatthatisisismade analytically tractable
where where TrTr ij M is M the
is + (a)
(a)
the = =
partial
partial Tr Tr trace
1,2 [U
trace [U over
(aover ⌦ sites i1)U
1)U iand j.j.These Another question that made analytically
i ]and These
Tr + 1,2
where ij
Tr M
ij is
+
0
+
the
(a)
(a) =
= partial
4 4 44 Tr
Tr
1,21,2
trace
[U
2,3 (up(1
1,3 [U (1 ⌦
over
⊗aa1SWAP, ⌦
sites
⊗1)U a)U ]and (10)
These
(9) non-contiguous question legs) reduce tomade exact analytically
erasure tractable
channels,
maps
maps mapsare are quantum
ij
arequantum
−
quantumchannels channels
4 (up to
toto they corre- byby tri-unitarity
tri-unitarity is the
isisisthe growth
growth of of entanglement.
ofofentanglement. We We showshow
maps
where
where are quantum
isthe
isis thethe
channels
aapartial
partial
4
channels
11 trace trace
(up
(up over toasites aaSWAP,
sites
SWAP,
SWAP, and
they
they
they corre-
corre-
corre-
These
byby tri-unitarity
tri-unitarity
Another
MAnother question
question
thethe
andthat
that
growth
growth
all
aisais two-point
made
made
entanglement.
entanglement.
correlations
analytically
analytically We
vanish
tractable
tractable show
††over and These here (a)
that, = Tr(a)1/2,
starting from class ofof
spond
where
where
spond spond
spond
to
Tr
to
Tr Tr
Trto
to
coupling
ij
ij ijis
coupling
ij coupling
coupling
M the
(a) =
qubit
partial
partial
a
aqubit qubit
qubit
Tr
to
trace
trace
to to
[Uto
two
two over
two
(atwo
ancillas
over
⌦ sites
ancillas
ancillas
ancillas
1 sites
⌦
iiin
1)U
iin
iand
in
] in
an
andan j.
an infinite-
j.
an
j. These
These
infinite-
j.infinite-
(11)
infinite- here µ
here
here
Another
Another
that,
that,
that,
question
question
starting
starting
starting from
fromthat
that
from ais is
aclassmade
made
class
class ofsimple,
of
analytically
analytically
simple,
simple,
simple,
short-range- tractable
tractable
short-range-
short-range-
maps
maps are
are quantum
quantum channels
channels (up(up toto
⊗ a SWAP,⊗ SWAP, 1)U they
they (10)
corre-
corre- byeverywhere.
by
entangled tri-unitarity
tri-unitarity
V.initial is is thethe
states, growth
growththe of of entanglement.
entanglement. WeWe show
show
thehalf-chain entanglement en-
temperature
maps
maps arequantum
are M +
quantum (a)
state, = evolving
channels
channels Tr 1,2
44 1,3unitarily, [U (up (1
(uptotoaaand a SWAP, SWAP, ]
discarding theycorre-
they the
corre- by bytri-unitarity
entangled tri-unitarity initial is
ENTANGLEMENTisstates,
thethegrowth
states, growth
the ofofentanglement.
half-chain entanglement.
GROWTH entanglement WeWeshow show
en-
temperature
temperature state,
0
state, evolving
evolving unitarily,
unitarily, and and discarding
discarding the
the entangled
entangled initial
initial states, half-chain entanglement
temperature
spond
spond
ancillas).
spond
spond toto
to tocoupling
coupling
coupling state,
Furthermore,
coupling
evolving
aaaaqubitqubit
qubitqubitthey to
toto
are
to
unitarily,
two
two twounital,
two ancillas
ancillas
ancillas
ancillas
andin
i.e. in
M in
in
discarding
an
anan
an(1) infinite-
infinite-
infinite-
infinite-
= 1.the here
here
tropyherethat,
here
tropy that,
inthat,
that,
in the
the starting
starting
thermodynamic
starting
starting
thermodynamic from
from from
from aaathe aclass
class half-chain
limit
class
class ofof
limit ofof simple,
simple,
growssimple,
simple,
grows
entanglement
short-range-
atshort-range-
at the the maxi-
short-range-
short-range- maxi-
en-
ancillas).
whereancillas).
ancillas). Tr Furthermore,
Furthermore,
is the
Furthermore, partial 1 theythey
trace
they are are †over
are unital,
unital,
sites
unital, i.e.i.e.
i.e.
i M
and M ±,0
M±,0 j. (1)(1) =
These
(1) = =1.1.
1. tropy
tropy inin the
the thermodynamic
thermodynamic limit
limit grows
grows at the maxi-
temperature
temperature
As a
temperature
temperature result, ij M state,
state,
the (a)
state,
state, = evolving
evolving
tracelessevolving
evolving Tr [U unitarily,
unitarily,
subspaceunitarily,
unitarily, (a ⊗ (spanned
1 and
⊗
and and 1)U discarding
discarding
anddiscarding ] by
discarding
±,0
±,0 the the
X,the
(11)the
the mumentangled
entangled
entangled
entangled velocity initial
initial allowed
initial
initial states,
states, by
states,
states, thethe
the
the the half-chain
half-chain
circuit’s
half-chain
half-chain entanglement
entanglement
geometry, entanglement
entanglement S(t) en-
= en-
ten-
en-
As
maps
AsAsa aaresult,
areresult,
result, quantumthe
+ thetraceless
the traceless
channels
traceless 4 subspace
1,2 subspace
(up
subspace to a (spanned (spanned
SWAP,
(spanned by
they by
by thethe
corre-
the X,X,
X, mum
mum
Another
mum velocity
velocity question
velocity allowed
allowed
allowed that byby byisthe
the made
the circuit’s
circuit’s geometry,
geometry,
analytically
circuit’s geometry, S(t)
tractableS(t) = =t t
ancillas).
ancillas).
,Z Pauli Furthermore,
Furthermore,
matrices) does they
they not areare mix unital,
unital, with i.e. i.e.
the (intropy
tropy bits).in in the
the For thermodynamic
thermodynamic
a finite subsystem limit
limit of grows
grows qubits,at at thethe
we maxi-
maxi-
find
1subspace,
Y
ancillas).Z
ancillas). Furthermore,
Furthermore, they
they areare unital,
unital, i.e.i.e. M M 1 (1)(1) ==== 1.1. (intropy
tropy
(in bits).in
bits). in thethe
For
For thermodynamic
thermodynamic
a a finite
finite subsystem
subsystem limit
limit of
of `grows
grows qubits,at at the
the we maxi-
maxi-
find
Y ,Y
spond ZPauli
,,aall toPauli matrices)
couplingmatrices) a qubit does does not
to not two mix mix with
ancillas with the the
inM M ±,0
1
an subspace,
±,0 (1)
subspace,
(1)
infinite- 1. 1. by (intri-unitarity
bits). For is the a finitegrowth ofcircuit’s
subsystem entanglement. of`` We show
t)`t)qubits, we tfind
±,0
As Y
where
As Z
a Pauli
result,
result, is matrices)
the
the partial
traceless
traceless does trace not
subspace
subspace over mix sites with
(spanned
(spanned the
and
±,0
by1by subspace,
theThese
the mummum velocity
velocity (each allowed
allowed boundary by by the the circuit’s
contributing geometry,
geometry, but only
S(t) up=
and
AsAs
and and aall
temperature
Tr
all information
aresult,
result,ij
information the
information thetraceless
state,
about
traceless
about
evolvingabout correlations
subspace
subspace
correlations
correlations
unitarily, (spanned and
is
(spanned is is
i
obtained
obtained
obtained
discarding byj.
bythe bythe
by bydi-
X,
the
X,
di-
X,di-
X, S(t)
mummum
S(t)
hereS(t) = = ==
that,
2t2t
velocity
velocity
V.
2t2t (each
(each
starting allowed
allowed
ENTANGLEMENT
boundary
boundary
from by by the the
asubsystem
class circuit’s
circuit’s
contributing
contributing
of of simple, geometry,
geometry,
GROWTH but
S(t)
short-range- only
S(t)
S(t) = = =tt t
up
YYYand
maps , ,,Z
agonalizing
ZZ all
are
Pauli
Pauli
ZPauli information
quantummatrices)
matrices)
the about
channels
doesdoes
33 3matrices not correlations
not(up mix to with
mix
↵ a⌘ with SWAP, isthe
the obtained
11they by di-
corre-
subspace,
11µsubspace, (in
to (inS(t)
a bits).
bits).
time t ForFor
= (eacha
`/4, a boundary
finite
finiteat subsystem
which contributing
point of
the ` ` qubits,
qubits,
entropy t) butwe we only
density find
find up
,agonalizing
agonalizing
Yancillas).
agonalizing
Pauli the the33 ⇥
matrices)
matrices)
Furthermore,
the 33 ⇥
⇥ does does
matrices
matrices
they
matrices
notnot
are
MMmix mix
Mµ↵ ↵ with
µ↵ ⌘⌘
unital, with Tr(
i.e.
Tr( the
Tr( M↵M
the M M (subspace,
( (( ))/2,
subspace,
(1) = ))/2,
))/2,
1. (in(in
toto
entangled
to
bits).
bits).
aaatimetime
time
For
For
tt t===`/4,
initial a a
`/4,
states,
`/4,
finite
finite atatthe
at which
which subsystem
subsystem
which point
point
half-chain point
of of
the
the `
the
` qubits,
qubits,
entropy
entanglemententropy
wewe
density
en- find
find
density
spond
and
and
where
and all to
all coupling
information
information
µinformation
all = ±,±,
information 0 0and ⇥
a ↵, 3
qubit
about
about
about = to
x, two
correlations
correlations
y,
correlations z. M µ ancillas ⌘ is is Tr( in↵↵
obtained
obtained an
±,0
↵ Mµ µ
infinite-
µ by by ))/2,
di-di- isis
S(t)
S(t) S/` = = 2t2t
1/2; (each
(each afterwards,
(each
(each boundary
boundary
boundary
boundary contributing
contributing
entanglementcontributing
contributing growth
t) t) butbut
t)but but only
only
becomes up
onlyup
only up up
and
where
As where aallresult,
µ µ
= = ±, the
0 andand about
traceless
↵, ↵, = correlations
= x,
subspace
x, y, y, z. z. µ
(spanned isisobtainedobtained by the byby di-
di- is
S(t)
S(t)
tropy S/`
S/` in==2t
=
isaAnother
2t
the1/2;
1/2; afterwards,
afterwards,
thermodynamic entanglement
entanglement
limit grows growth
growth
t)
at the becomes
maxi-
where
agonalizing
temperature
agonalizing
We can µ =
observe±,
thethe 0
state,
3 3and
⇥all
⇥ evolving
3 four
3 ↵, matrices
matrices of= the x,
unitarily,
M y, z.
↵ ↵
behaviors
M ⌘ ⌘ and
Tr( Tr( discarding
listed M M µin
( ( Sec.
X,
the
II
))/2,
))/2, to to aS/`time
time
model-dependent.
to to a a time
time
= tquestion
t t1/2;
= = `/4, afterwards,
`/4, that
at
atat at which
which is
which
which
made
entanglement
point
point point
point
analytically
thethe
thethe entropygrowth
entropy
entropy
entropy
tractable
becomes
density
density
density
density
agonalizing
agonalizing
,We We can can
Pauli the
observe the
observe 3
matrices) 3 ⇥ ⇥ all
allall
3 four
3 doesmatrices
matrices
four of
ofnot the the M
mix
↵ ↵
behaviors
behaviors
M µ µ ⌘⌘Tr(
µwith the Tr( listed
↵↵M
listed M µinµ Sec. IIII
in
µ( (
↵ subspace, Sec.
))/2,
))/2, mum model-dependent.
model-dependent.
velocity t = =
allowed`/4,`/4,
theby the circuit’s geometry, S(t) = t
Ywhere
ancillas).
where
(from
Z We µcan observe
Furthermore,
non-interacting
µ= == = ±, ±,0000andand ↵, ↵, tofourthey of
ergodic
= = x,x,arethe
y, y, z. behaviors
µ
unital,
and
z. mixing) i.e. M
↵1
listed within in Sec.
(1) the
= 1.II isby
isis
is In
S/` tri-unitarity
model-dependent.
S/` this
== == Section,
1/2;
1/2; is we
afterwards,
afterwards,
afterwards,
afterwards,
growth
will allow of the
entanglement
entanglement
entanglement
entanglement
entanglement.
local growth
growth Hilbert
growth
growth
We
becomes
becomes
spaceshow
becomes
becomes
where
where
(from
and (from all non-interacting
non-interacting
µ µ ±,
information ±, and
and ↵,
about↵,to to= ergodic
ergodic
= x, x, y, y, z. and
z. and mixing)
mixing) ±,0
within
within the
the (in In In
S/`
S/`
bits). this
this Section,
Section,
1/2;
1/2;
For a finitewe
we will
will
subsystem allow
allow the
the
of local
localqubits, Hilbert we space
find
As (from
We We
a can
can non-interacting
result, observe
observethe allall
traceless four four toofcorrelations
ergodic
of the the
subspace and
behaviors
behaviors (spanned ismixing)obtained
listed
listed within
in
by in bySec.
Sec.
the di-
IIthe
II here
dimension In that,
model-dependent.
model-dependent. this starting
totoSection,
be from
we willa class
allow of the `
simple,local short-range-
Hilbert space
family
family We Wecan
family
agonalizing
of
can
of gates
observe
observe
ofgatesgates
the
in
in in Eq.
allall
Eq. four
Eq. (6)
four
(6)
matrices (6) and
ofof
andthe
andthe Fig.
behaviors
Fig.Fig. 5.
behaviors
↵ 5.5. The The The simplest
listed
listed ininSec.
simplest
simplest ex-
Sec. X,IIII
ex-
ex- model-dependent.
model-dependent.
dimension
dimension
S(t) toto
(each bebe q qq 2,2,
boundary 2,and and
2,and
measure
measure
measure
contributing
entropy
entropy
entropy but
ininunits
inunits
only up
ofof
family
(from
(from
ample
Y , Z Pauli
is of gates
non-interacting
non-interacting
given
3
matrices)by
⇥
in 3Eq.
setting to
does (6)
toergodic and
ergodic
not M
⌘ mix µFig.
and
and
0
⌘
with
and mixing) 5.Tr(
mixing) The
the
all ↵ M (
simplest
within
within
µ
subspace,
single-qubit
1 ))/2,
the ex-
the ln(q)In In=
entangled
dimension this
this
(i.e.
2tSection,
initial
Section,
bits forbe states,
we
qqwe = will
will2, the and
etc). half-chain
allow
allow measure
the
As theinlocal
local
the
t)Hilbert
entanglement
entropy Hilbert
case of units
space
space
dual- en-of
(from
(from
ample
whereample
ample µisnon-interacting
non-interacting
isgiven
=
is
given
±,
given0 andbyby
by setting toto=
↵,setting
setting
ergodic
ergodic
x, i
iy,i⌘ ⌘andand
⌘0⌦05.
z.Fig. 0and and
and
mixing)
mixing) all withinthe
within
allsingle-qubit
single-qubit the to aInIn
ln(q)
ln(q)
tropy time thisthe
this
(i.e.
(i.e.
in(i.e.tto Section,
Section,
=bits
bits `/4,forforwe
thermodynamic
bits for qqwe
at = will
will
which
= 2,2, etc). allow
allow
etc).
point
limit
etc).
the
As
As the
the
As in
growslocal
local
in
in
the Hilbert
the
entropy
the at Hilbert
case the
case ofof
density space
space
dual-
maxi- dual-
family
family
and
gates
family
familygates
of
alltoof of
to 1,
of1,
gates
gates
information giving
gates
gates
inin Eq.
ininall
giving U
Eq.
Eq.
about
U=
Eq. (6)
(6)
(6)
(6)SWAP and
and
andandiFig.
correlations
SWAP Fig.
Fig. 5.
5.15. The
is:TheThe
The
:allallsimplest
obtained
all M single-qubit
simplest
µ act
simplest
simplest byact
ex-
ex-
di-
as
ex-ex-
as
dimension
dimension
ln(q)
untiary
dimension
dimension
untiary
untiary
circuits,to be be itqitis
totoafterwards,
circuits,
circuits, be be q q
itqqitis
q and
=
ishelpful
2,2, and
and
helpful
helpful
2, 2,25
2,
and
measure
measure
to consider
measure
measure
to
to consider
consider
entropy
entropy a suitable
entropy
entropy acase
in in
suitablein
units
units
inS(t)classof
units
units
of
class oftof
gates We
gates isto to
can 1,giving
observe U four = = SWAP
of the 1,3
behaviors
⌦ ⌦ 1 21
: all
listed M in
M act
Sec. as
II ismum S/`
untiary = 1/2;
velocity circuits, allowed is=by entanglement
the
helpful circuit’s
to consider growth
geometry, a becomes
suitable class
thegiving SWAP :all all act as (i.e. bits for etc). As in the case of dual-
1,3 2 µ
ample
ample
agonalizing
identity isgivengiven
1,
channels, 3byby
×and 3setting
setting
Umatrices
thus = all 1,3
i icorrelations
⌘M 00and
αβ
⌘µ1,3 ≡⌦ and 2 12all
Tr(σ along Msingle-qubit
single-qubit
Mµ
the
(σµthe spe-
β ))/2, of ln(q)
ln(q) (i.e.
“solvable” bits
initial for states
q= 2, 2, .
25etc).
While As thein the
task of of
classifying dual- =
ample
ample identity
identity
(from is given
channels,
ischannels,
given
non-interacting bybyand andsetting
setting thus
to thus all
ergodic all ⌘correlations
i icorrelations
⌘ 00and
and and
mixing) all all
αalong
along single-qubit
single-qubit
µ
within the spe-
spe-
the ofof
(in “solvable”
ln(q)
model-dependent.
ln(q) bits). (i.e.
“solvable”
(i.e. bits
bits initial
initial
Forinitial for for
aititthe states
states
q
finiteq = = 2,252, .
25
subsystem . etc).
etc).While
While AsAs the
thein
ofthein task
the
task the
qubits, ofcase
case of
we of
classifying dual-
dual-
find
identity
gates
gates to to= channels,giving
giving and thus x,isall
SWAP
SWAP correlations :2as: all allalong µ µact the
actas spe-
as ofsuch
untiary
untiary
all “solvable”circuits,
circuits,
states states
isishelpful
helpful .toWhile
to consider
consider is` aistask
asuitable of classifying
suitable class
class
where
cial rays
cialrays
gates
gates µto
rays
to are
1,1,±,giving
are andU
constant.
0constant.
giving
U=
Uα, βThis
= =
=This SWAP
SWAP isis
y,intuitive,
z.1,3⌦⌦11
intuitive,
1,3
⌦121 : as
:2as the
allall
MM
the circuit
µcircuit
µact act in
inin
asas all
untiary
untiary such statesinin
circuits,
circuits, in the
itthe tri-unitary
isiswill
itboundary tri-unitary
helpful
helpful totothe case
case
consider
consider left
isat) aleft for forfuture
suitable
suitable future class
class
cial
family are areconstant.
ofchannels,
gates in andEq. This
(6) and intuitive,
Fig. 5. 2The the thecircuit
simplest ex- ofall such states thetri-unitary case
1,
ofIn this Section, we .allow local Hilbert space
1, U= 1,3⌦ M M all such (each
states in 25 25 contributing
tri-unitary case isofofbut
left only
for up
future
thiscial
identity
identity
We rays
can
limit channels,
observe
reduces constant.
all
to andafour thus thus
sequenceThis
of allall
the is intuitive,
correlations
1,3
correlations
behaviors
of SWAPs asalong
along
listed
acting the
in circuit
the
on spe-
spe-
Sec.
odd IIin S(t)
work, “solvable”
“solvable”
=
a 2t
particularly initial
initial states
states
simple 25 .While
family
25 While the
of the
such task
task states classifying
classifying
is given
this
identity
identity limit reduces
channels,
channels, and to
and athus sequence
thus allall of
correlations
correlations SWAPs acting
along
along the on
the odd
spe-
spe- ofwork,
of
work, “solvable”a
“solvable” particularly
a astates
particularly initial
initial simple
states
states
simple . family
. While
While
family of
thethesuch
ofentropy
such task
task states
of
states of is
is given
classifying
classifying
ample
thisthislimit
cial
cial
(from
qubits rays is are
limit
rays given
reduces
arereduces
non-interacting
(giving
bytosetting
constant.
constant.
left- toa to
and asequence
Thissequence
This
ergodic
of0ofSWAPs
⌘intuitive,
isiisintuitive,
right-movers), and
and
SWAPs mixing)
while asall
asacting
the single-qubit
acting
the
even
ononodd
circuit
circuit
within qubits odd
in
thein dimension
byto
all asuch
allwork,
such
Bell timepairs
to bein
particularly
states
t intercalated
= qinthe
`/4, the 2,
at and
which
simple
tri-unitary
tri-unitary
by
measurepoint
family
disentangled case
case the
of isentropy
issuch left
left
sites:
in
states
for units
for isgiven
density
future
future of
given
cial
cial
gates
qubitsqubits rays
raysto (giving
(givingare
1,aregiving left-and
constant.
constant.
left- Uaand This right-movers),
This
SWAP isisintuitive,
right-movers), intuitive,
⌦SWAPs 1 while : while asthe
asall the even
even circuit
circuit
act qubits
qubits inin
as byby
allall Bell
such
such
Bell(i.e. pairs
pairsstates
states
bits intercalated
ininqthe
intercalated
for the tri-unitary
byby disentangled
tri-unitary
etc). disentangled
Asofof case
case
in the isgrowthsites:
iscaseleftfor
left
sites: fordual-
of future
future
are qubits
this
this
family of(giving
limit
limit
inert reduces
reduces
(“non-movers”).
gates left-
in to to
Eq. anda= right-movers),
sequence
sequence
(6) Letting
and ofofSWAPs
1,3
Fig. 5.6=26=The 0whilemakesacting
acting M even
simplestµ on
the on qubits
oddodd
dy-
ex-
ln(q)
is
work, by
work, S/` Bell
aa= pairs
particularly
particularly
1/2; afterwards,
intercalated
O
=
simple
simple 2, entanglement
by disentangled
family
family such
such states
states sites: isbecomes
isgiven
given
this
this areinert
identity
are inert
limit
limit (“non-movers”).
reduces
reduces
channels,
(“non-movers”). to
andtoaathus sequence
sequence Letting
all
Letting of ofSWAPs
correlations iSWAPs
6
= 0 0 makes
acting
acting
along
makes the on
thethe
on odd
spe-
dy-dy-
odd work,
work,
untiary
model-dependent.a a particularly
particularly
circuits, it is
OO simple
simple
helpful tofamily
family consider of of such
such a states
states
suitable isis
class given
given
qubits
are inert
qubits
namics
ample (giving
(giving
isinteracting
given left-
(“non-movers”).
left-
by but andnon-ergodic
and
setting right-movers),
φLetting
right-movers), ≡ 0 i(ZZ and
i
i while6=while 0 makes
correlators
all evenqubits
even
single-qubit qubits
theare dy- by byBell Bellpairs pairs i intercalated
| |0intercalated
= O |B|B +
i+xby
by disentangled
disentangled
|0i|0ix 1 |0i sites:
sites: , , (12)(12)
cial
namicsnamics
qubits
qubits rays interacting
(giving
(givingare
interacting left-
left-
constant. but but
and
and non-ergodic
right-movers),
right-movers),
This
non-ergodicisi intuitive, (ZZ (ZZ whilewhile
as correlators
the even
even
correlators circuit qubits
qubits are
in
are ofby by Bellpairs
Bell
“solvable” pairs
| initial
i iintercalated
0intercalated
== states |B 25+
. by
i+ by
iWhile disentangled
x disentangled
2,x
2,x |0ithe x task |0i
1|0i
x+1
of
x+1 sites:
sites:
classifying
, (12)
are
arenamics
along
gates inert
inertthe
tothe (“non-movers”).
(“non-movers”).
interacting
rays
1, while
giving but
Uother = SWAP Letting
Letting
non-ergodic
decay); by⊗i iintroducing
6
= (ZZ
6
= 0
2 :00makes
0
1introducing makes
makes
allcorrelators the
Mµgenericthe
act dy-dy-
are
as In this Section, | 0 i x24Z
0 =
O O |B allow
we tri-unitary
will
x ix 2,xthe
2,x |0ixlocal
x 1
1 |0iHilbert
x+1
x+1 , space (12)
are
are
this
along along inert
inert
limit
the rays rays
(“non-movers”).
reduces while
(“non-movers”).
while tobut a other
other sequence decay);
Letting
Letting
decay); of by
1,3 by
SWAPs iintroducing
6= 6= makes
acting the
on generic
theodd
generic dy-
dy- all such states in the
x24Z
O O ++ case is left for future (12)
namics
namics
along
single-qubit
identity the interacting
interacting
rays
channels, gates while
one
and but other
realizes
thus non-ergodic
non-ergodicdecay);all
all correlations other i(ZZ
by (ZZ introducing
behaviors,correlators
correlators
along are
are
generic
which | | i i
= = x24Z |B |B i i |0i|0i |0i
|0i , , (12)
eventhe spe- dimension |to|p01ibe qq ≥ and measure entropy ,in units of
x24Z
1 2,|B
0 0 P x
+ x 2,x2,x x x 1 1 x+1 x+1
single-qubit
namics
qubits
namics
single-qubit interacting
(giving
interacting gates
gatesleft- one one
but
and
but realizes
realizes non-ergodic
right-movers),
non-ergodic allall other (ZZ (ZZwhile behaviors,
correlators
correlators which
qubits areare work, |Ba(i.e.
+particularly 01i=P = simple |B +
xix 2,x
afamily
iBell |0i
of
|0i xsuch 1 |0i
x 1 |0i states ,isqudits,
given (12)
(12)
allother A.behaviors, which
along the rays while other decay); by introducing generic with P pair
2,x state on x+1
x+1 two
along
we
cialsingle-qubit
survey
raysthe rays
arein while
gates
more
constant. other
one
detail This indecay);
realizesAppendix
is by
intuitive, other introducing as behaviors,
the generic
circuit which
in with
ln(q) |B i+⌘ i bits
⌘ 1 for x24Z qqq 1|jji
x24Z
P =1
2,
|jji etc).
a Bell As
pair in the
state oncase two of dual-
qudits,
are
along
we we
along survey
inert
survey the
the rays
in in
rays more
while
(“non-movers”).
while
more detail
other
detail other in in
decay); Appendix
decay);
Letting
Appendix by by 6
= introducing
A. 0A.
introducingmakes the generic
dy-
generic with
by with Bell |B
|B +
pairsi ⌘
i+ ⌘ pqp
intercalated
q1
j=0
j=0
x24Z
x24Z q |jji
1
|jji a Bell
by disentangled
ato Bell pair state
pair state on
sites: two
on state qudits,
two qudits,
single-qubit
single-qubit
we Finally survey wewe in gates
gatesmore
note oneone
that detail ifrealizes
a realizes in aall
is isAppendix all other
other
i behaviors,
behaviors,
A. which
which qpPj=0
this
namics
single-qubit
Finally
limit
Finally
single-qubit reduces
interacting
we gates
notenote
gates to
one
that that
but
one sequence
realizesifU
non-ergodic
ifrealizes
UinU is a⌦ allaperfect
ofperfect
all SWAPs
⇤other
other
perfect (ZZ
tensor, acting
tensor,
correlators
behaviors,
behaviors,
tensor, ⌦3 then
then
thenon con-
odd
con-
are
which
which
con-
and
untiary
and|0i
and
with
with |0i
|0i
|B |B ++
could
circuits,
could
could
i i⌘ ⌘
be
p1p1be
Pbe P
replaced
qit
P
is1j=0
1helpful
q qreplaced
replaced |jji
|jji a
byby
a by
Bell
Bell
any any
any
single-qudit
consider
pair
pair single-qudit
single-qudit
state
state
a suitable
on on twotwo state
state
qudits,
qudits,
(a(a
class
(a(a
wewe
tracting
qubits survey
survey
Finally any
(giving in
inwe more
morenote
three
left- detail
detail
that
legs
and in
if
in Appendix
Appendix
right-movers),
UU is a U perfect⇤with A. A.
while tensor,
1 even then
⌦3necessar- qubits con- of and
short-range-entangled
“solvable”
with|B
with |0i + could q1 be
initial O j=0q replaced
state
states
j=0q 1 1 25 would
. by
While
aawouldBell
Bell any
cause
the
pair
pair single-qudit
a
task
state
state minorof
ontwo state
change
classifying
twochangequdits,
along
wewe tracting
tracting the
survey
survey any
rays
in in three
while
more
more
three other
detail
detail
legs legs inin
decay);
in Appendix
U
Appendix ⌦ by
U⇤ introducing with
A. A. ⌦3 necessar-
1 generic
+
short-range-entangled pq1p
|B i| i⌘⌘i =q q j=0 state
|jji
|jji would cause a onminor qudits,
change
ififin U U ⇤with necessar- short-range-entangled state cause a minor
+
Finally
Finally
tracting anywe wenote
any note
threethatthat legs UU UisU
inis a⌦ aperfect
perfect tensor,
with 1 1⌦3
tensor, then
then con-
con-
necessar- andand could
could
short-range-entangled
|0i
|0i be be replaced |B
replaced
j=0 i by
state wouldby any|0i
any |0i
single-qudit
single-qudit
cause ,
a minor state
state (12) (a(a
change
are inert
single-qubit
Finally
Finally (“non-movers”).
we we gates
note
note one that
that realizes
ififin Letting
UU
Uin isUisa⌦ all a⌦
⌦ φ
other
perfect
perfect i 6
= 0 makes
behaviors,
tensor,
tensor, ⌦3 then
the
thenwhich dy-
con-
con- all
andand such
|0i states
|0icould
0
couldbe inbe the
replaced
replaced tri-unitary
x 2,x
by byany any
x
case
1
is left
x+1
single-qudit
single-qudit forstatefuture
state (a(a
tracting
tracting
namics any
any
interacting three
three legs
legs U U⇤ ⇤with with 11 ⌦3 necessar-
necessar- short-range-entangled
short-range-entangled x24Z statewould
state would cause
cause aaminor minorchange change
tracting
tracting anyany three threebut legslegsnon-ergodic
ininUU⌦⌦UU⇤ ⇤(ZZ with withcorrelators
11 ⌦3 ⌦3
necessar-
necessar- are work, a particularly
short-range-entangledstate
short-range-entangled simple statewould family
wouldcause of such
causeaaminor states is
minorchange given
change
 7

by Bell pairs intercalated by disentangled sites: “solvable” initial state in Eq. (12). Then, unitarity of
O Ũ and Ŭ allows further gate elisions starting from the
|ψ0 i = |B + ix−2,x |0ix−1 |0ix+1 , (12) corners and iterating all the way to the cut, leaving only
x∈4Z a one-dimensional column of gates, Fig. 9(c). Keeping
Pq−1 track of the initial state normalization, we find Tr ρnA =
with |B + i ≡ √1q j=0 |jji a Bell pair state on two qudits, [(e|χ)q −n ]t+1 , where (e|χ) denotes the contraction of the
and |0i could be replaced by any single-qudit state (a two permutations62 :
short-range-entangled state would cause a minor change q
X q
X
to the proof). Notice that the first unitary layer to act (e|χ) = hi1 , . . . in |i2 , . . . in , i1 i = 1=q.
on |ψ0 i is Ue in Eq. (4), with three-qubit gates acting i1 ,...in =1 i=1
on triplets (x, x + 1, x + 2), x ∈ 4Z, while the initial
entanglement is between qubits (x − 2, x), (x + 2, x + 4), We conclude Tr ρA (t)n = q (1−n)(t+1) , i.e. Sn = t + 1 (in
etc. units of log q). For the other kind of entanglement cut
We consider a semi-infinte contiguous subsystem A = (adjacent to a qubit that has not been acted on by a gate
{x < xcut } in an infinite chain x ∈ Z. The initial state in the last layer), the same derivation yields Sn (t) = t.
in Eq. (12) makes the computation of Tr ρnA analytically (Notice that at t = 0 this cut dependence correctly re-
tractable, for all integer n ≥ 2, thus giving all the Renyi duces to whether or not one of the Bell pairs in the initial
entanglement entropies Sn = 1/(1 − n) log Tr(ρn ), n ≥ 2. state straddles the cut.) Thus at a fixed cut in space, en-
We will find that Sn (t) is independent of n and equals tanglement alternately grows by 2 (if a gate acts across
either t or t + 1, depending on the entanglement cut xcut the cut) and 0 (otherwise), for an average entanglement
and the parity of t. The fact that Sn (t) is independent of velocity of exactly 1.
n for all integer n ≥ 2 implies that all the Renyi entropies, For a finite subsystem with two edges, the derivation
as well as the von Neummann entropy, in fact coincide. proceeds unchanged for each entanglement cut, giving
This allows us to use the word ‘entropy’ in an unqualified Sn (t) = 2t + c (c = 0, 1 or 2 depending on the cut lo-
sense in this setting. cations and the parity of t as explained above), as long
The derivation is illustrated diagrammatically in as t < `/4 (` being the number of qubits of the subsys-
Fig. 9, for the case of an entanglement cut xcut between tem). After this point, the backward lightcones emanat-
two qubits that have both been acted on by the last layer ing from each cut intersect, and it is not possible to elide
of unitary gates (half of possible cuts are of this kind; gates further based on tri-unitarity alone. Thus Sn (t)
the other half are reduced to this kind by eliding the last is guaranteed to grow at the maximal speed only up to
layer, which does not affect entanglement, and setting t ' `/4, when it satisfies Sn (t) ' `/2; after that time
t 7→ t − 1 in the following derivation). ρnA consists of the behavior may change. In fact it is easy to find an ex-
n replicas of the doubled circuit U ⊗ U ∗ : one “ket” and treme example in which entanglement growth abruptly
one “bra” per replica. The tensor legs at the final time stops at t = `/4: if the odd sublattice is entirely decou-
step t must be contracted appropriately: each “ket” leg pled from the even one, the tri-unitary circuit breaks up
is paired to a “bra” leg according to a permutation π in into a dual-unitary circuit and a set of inert, disentan-
the symmetric group on n elements9,61 . In Ā (which is gled qubits; then, at time t = `/4, the even sublattice
traced out to produce ρA ), each “ket” leg is contracted saturates to maximal entanglement (`/2), while the odd
to the “bra” leg from the same replica, i.e. the pairing sublattice remains disentangled.
is given by the trivial (or identity) permutation, denoted The above results hold for all integer n ≥ 2. This
by e. In A instead, each “ket” leg is contracted with the however implies that they also hold for non-integer n, as
“bra” leg in the following replica (in order to implement well as for n → 1 (von Neumann entropy). One can
the product ρnA ), giving a cyclic permutation χ. see this as follows. Considering a finite subsystem A
As a consequence of tri-unitarity, whenever a “stack” for simplicity, let ρA be the reduced density matrix and
of gates (U ⊗ U ∗ )⊗n has three identical permutations {λα : α = 1, . . . q |A| } be its eigenvalues. Our results
g ⊗3 (g = e or χ) on three adjacent legs, the gates can state that
X
be elided, and the permutations g ⊗3 moved over to the Tr ρnA = λnα = q (1−n)S , (13)
three output legs. By using unitarity of U alone, one α
can elide the circuit everywhere outside the backward
light cone of the entanglement cut, turning the tensor for integer n ≥ 2, where S is an n-independent integer
network of Fig. 9(a) into that of Fig. 9(b). We note that value, as derived earlier. The above can be rewritten as
contraction between a permutation e or χ and one of X
the single-qudit initial states on the odd qudits is simply (q S λα )n = q S . (14)
⊗n α
Tr |0i h0| = 1; similarly all Bell pairs contained entirely
outside the backward light cone give 1. Crucially, Bell For the sum to stay finite as n → ∞, we see that
pairs that straddle the light cone carry a permutation q S λα ≤ 1 must hold for all α; moreover, to match the
(e or χ) into an input leg for the gates at the bottom right-hand side when n → ∞, exactly q S of the entangle-
corners of Fig. 9(b): this is the reason for the choice of ment eigenvalues (note that S is an integer) must satisfy
 8

(a) Ā A comes model-dependent.

This phenomenology is illustrated in Fig. 10, which
shows the results of exact numerical simulations of en-
tanglement growth under Floquet tri-unitary circuits, for
… … a subsystem A of ` = 13 qubits in a chain of length
t =5

L = 27, starting from the solvable initial state Eq. (12).

The circuit setup is sketched in Fig. 10(a); we choose
open boundary conditions, so that the maximum entan-
⋯ q −n q −n q −n q −n ⋯ glement velocity is 1 rather than 2. We find that all the
(b) circuits, as predicted, exhibit unit entanglement velocity
up to S = 6 bits (the integer part of `/2); after that point,
the curves for different circuits visibly split up, Fig. 10(b).
= At φ = 0 we have a non-interacting SWAP circuit, where
all entanglement is due to the ballistic motion of the six
Bell pairs present in the initial state; with an infinite
q −n q −n bath Ā, the entropy of A would plateau at 6 bits forever,
q −n q −n q −n q −n however finite L induces periodic oscillations between 0
and 6 bits (depending on how many Bell pairs straddle
(c) the entanglement cut at any given time). Adding weak
= q −n(t+1) × = q −n(t+1) × = q (1−n)(t+1) interactions φ > 0, the oscillating behavior dictated by
the motion of Bell pairs gradually morphs into a steady
ballistic trend. Finally, at strong interactions φ ∼ π/2
we observe fast growth all the way up to the Page value,
though the entanglement velocity becomes sub-maximal
immediately after t = 6. The same behavior is seen in
FIG. 9. Entanglement growth in tri-unitary circuits: dia-
grammatic derivation for the nth Renyi entropy of a state a Floquet circuit made of perfect tensors (with random
evolved for time t = 5. (a) Tensor network diagram for single-qubit gates ensuring the circuit is non-Clifford).
Tr(ρA (t)n ). Hexagons represent n replicas of the “folded” gate Finally, we note that while the non-interacting be-
U ⊗ U ∗ ; the initial state (bottom) is an in Eq. (12), with tri- haviour (φ = 0 in Fig.10) may be captured in dual-
angles denoting single-qudit |0i states and arcs denoting the unitary circuits with suitable initial states, the interact-
Bell pairs |B + i up to the normalization q −n (one factor of 1/q ing behaviour (φ > 0 in Fig.10) is unique to tri-unitary
for each replica of |B + ihB + |) indicated explicitly; circles at circuits. This is because in dual-unitary circuits the en-
the top represent contraction of the n “ket” and n “bra” legs tropy must grow linearly until saturation, whereas in tri-
according to the identity permutation e (in Ā) or the cyclic unitary circuits the entropy must only grow linearly until
permutation χ (in A). (b) Using only unitarity of U , all gates
half the saturation value. Along with the results on cor-
outside the backward light cone of the entanglement cut are
elided. Tri-unitarity and the choice of a “solvable” initial state relation functions in Sec. IV, this is an example of how
allow further progress: gates can be elided from the bottom tri-unitariy circuits give rise to a richer variety of be-
right (left) corner by using unitarity of Ũ (Ŭ ). (c) Gates at haviors than dual-unitary ones, while at the same time
the entanglement cut are elided last, resulting in t + 1 direct retaining some of the analytical tractability.
contractions between pertmutations (e|χ) = q. The net re-
sult is Tr ρA (t)n = q (1−n)(t+1) , thus the nth Renyi entropy (in
units of ln(q)) is Sn = t + 1 for all integers n ≥ 2. VI. HIGHER DIMENSION

Having used tri-unitarity to derive results on cor-

q S λα = 1, while all others must vanish. This constraints relations and entanglement in (1 + 1) dimensions, it
the reduced density matrix to the form ρA = P/q S , where is interesting to ask about possible applications to
P is a projector of rank q S : thus the entanglement spec- higher dimensions. The most straightforward exten-
trum is flat and all the entropies (Renyi or von Neumann) sion of dual-unitary circuits consists of “gluing” (1 + 1)-
coincide. Based on this fact we simply refer to ‘entropy’ dimensional dual-unitary circuits in an additional spatial
in the following. dimension33 ; this however results in highly anisotropic
In generic tri-unitary circuits, one expects entangle- (2 + 1)-dimensional models, with a “spacetime duality”
ment to saturate to a thermal (infinite-temperature) transformation that only acts on two of the three dimen-
volume-law given by the Page value63 SPage = ` − sions.
22`−L−1 / ln(2) (in bits) for a subsystem of ` qubits in A different approach, which reproduces the phe-
a one-dimensional chain of length L ≥ 2`, up to cor- nomenology of (1 + 1)-dimensional dual-unitary circuits
rections exponentially small in `. Thus entanglement more faithfully, is to use “multi-unitary gates” arranged
should continue to increase even after the maximum- on the vertices of a hypercubic lattice in any dimen-
velocity growth regime stops; however, the behavior be- sion18 . The tri-unitary gates introduced here can be
 9

(a) A Ā (a) t (b) t̃ t

t̃ t̆

(b)
t̆

FIG. 11. (a) Tri-unitary tensor represented as a hexagon in

(1 + 1)-dimensional spacetime; each leg is associated to a ver-
tex. Each contiguous bipartition of legs (associated to one
of the three arrows of time t, t̃, t̆) defines a unitary gate.
(b) The same tensor viewed as a cube in (2 + 1)-dimensional
spacetime; each leg is associated to a face and labelled by the
direction perpendicular to it. Each arrow of time crosses the
cube through a pair of opposite vertices. The correspondence
between legs in the two panels is indicated by the labels in
(a).

FIG. 10. Numerical results for entanglement growth in Flo- in Fig. 11(b).
quet tri-unitary circuits. (a) Circuit layout: a subsystem A
of 13 qubits in a chain of length L = 27, with open bound- How do these constraints affect infinite-temperature
ary conditions and initial state as in Eq. (12). (b) Numeri- correlation functions? Generalizing the idea in Fig. 1
cal results for the second Renyi entropy of subsystem A, SA to the present context, we see that each arrow of time
(measured in bits), from exact time evolution. The gates are restricts two-point correlations to two octants, or tetra-
Ut.u. [φ] from Eq. (6) with φi ≡ φ, variable φ, and Haar- hedral light cones: e.g., for two operators separated by
random single-qubit gates (data averaged over 10 realiza- a vector (x, y, z), unitarity along t requires x, y, z ≥ 0
tions). For φ = 0, the model is non-interacting and the 6 (future light cone) or x, y, z ≤ 0 (past light cone) for
Bell pairs in the initial state move ballistically, bouncing off connected correlators not to vanish. The other six oc-
the walls and causing SA to periodically oscillate between tants correspond to “spacelike separations” relative to
0 and 6 bits. As interactions are turned on (φ > 0), the
t and thus cannot have correlations. Similarly, unitar-
entropy growth up to t = 6 remains exactly unchanged, as
expected; past t = 6, the oscillations gradually give way
ity along t̃ restricts correlations to two distinct octants,
to ballistic growth. We also show data for a circuit made −x, y, z ≤ 0 or −x, y, z ≥ 0; and analogously for t̆. Thus
of perfect tensors (dressed with random single qubit gates). unitarity about all three arrows of time limits correla-
The dashed line denotes the Page value for this bipartition, tions to lines: namely the x, y, and z rays (which are
1
SPage = |A| − 4 ln 2
. the intersection of all three tetrahedral light cones).
More rigorously, this result can be derived by analyz-
ing the tensor network contraction that corresponds to
directly employed in one such construction, as we dis- the correlation function. In Fig. 12 we derive in detail
cuss in the following. Specifically, we present one con- the fact that correlators vanish at all points strictly in-
struction of (2 + 1)-dimensional quantum circuits built side the light cone, x, y, z > 0. This only requires uni-
out of 3-qubit tri-unitary gates (as defined in Sec. III B), tarity of U and of one between Ũ and Ŭ , which is a
and exactly derive their two-point correlation functions. less restrictive condition than tri-unitarity. This less-
Like in the (1 + 1)-dimensional case, correlations are con- restrictive condition allows correlations on some surfaces
fined to three special rays; however, in this case the three at the boundary of the light cone, e.g. the quadrant
rays are not co-planar. Having three valid arrows of time x = 0, y, z > 0 (where the derivation of Fig. 12 would
thus offers the possibility of genuine, intrinsically (2 + 1)- not carry through without invoking unitarity of Ŭ ). On
dimensional tri-unitary circuits. the other hand, full tri-unitarity rules out all correlations
We consider a cubic lattice, Z3 , with a tri-unitary gate except for the three lines x, y, z; there, the same tensor
U at each vertex. Each leg of U connects to one of network analysis shows that correlations are given by it-
its 6 nearest neighbors. It is convenient to name the eration of the same quantum channels M±,0 found in the
legs of U as x, y, z and x̄, ȳ, z̄, based on which neighbor (1 + 1)-dimensional case, Eq. (9), (10), (11). The chan-
they connect to, see Fig. 11; tri-unitarity of U means nels do not describe left/right/non-“movers”, as they did
that the maps (x̄, ȳ, z̄) 7→ (x, y, z), (x, ȳ, z̄) 7→ (x̄, y, z), in the (1+1)-dimensional case of Sec. IV, but rather three
and (x, y, z̄) 7→ (x̄, ȳ, z) are unitary. These define three equivalent directions in space: thre projections of x, y, z
equally valid arrows of time: t ∝ (1, 1, 1), t̃ ∝ (−1, 1, 1), onto the spatial plane t⊥ ≡ {x + y + z = 0}, as sketched
and t̆ ∝ (−1, −1, 1). Each of these crosses the cube in Fig. 13(a). These three directions in the plane, wich
through a pair of diagonally opposite vertices64 , as shown we call ξ, η and ζ, form 2π/3 relative angles and are
 10

b(x, y, z) (a) (b)

(a) (b) z ζ
z t∝x+y+z t
ζ t⊥
y
η
x ξ y
a(0,0,0) x ξ η
(c) (d)

FIG. 13. Correlations in (2 + 1)-dimensional tri-unitary cir-

cuits. (a) Correlations vanish everywhere except for the rays
= = x, y and z. The arrow of time t ∝ x + y + z defines spatial
planes t⊥ ; the directions ξ, η and ζ on the plane are defined
by projecting x, y and z. (b) Snapshot of a spatial plane.
Qubits live on the sites of a kagome lattice (circles), and two-
point correlations propagate ballistically at maximum velocity
along the high-symmetry directions ξ, η and ζ on the lattice.
FIG. 12. Two-point correlation functions vanish inside the
light cone for (2 + 1)-dimensional tri-unitary circuits. (a)
The physical arrow of time is t ∝ (1, 1, 1). (b) Ten-
sor network contraction for the infinite-temperature correla- as a local quantum circuit on two-dimensional qubit ar-
tion ha(0, 0, 0)b(x, y, z)i. Cubes represent “folded” tri-unitary rays. This is slightly subtle: the space-like slices of the
gates U ⊗ U ∗ ; legs contracted to non-trivial operators (a and (2 + 1)-dimensional tensor network described above, i.e.
b) are shown, while all other exposed faces are implicitly con- the planes t⊥ orthogonal to t, intersect the qubit world-
tracted to the identity. Unitarity of U (along t) simplifies the lines at locations that vary with time over the course of
tensor network to a hyperrectangle – the intersection of a’s one period (meant here as a period of the circuit architec-
forward light cone and b’s backward light cone. Unitarity of
ture; the gates need not be time-periodic). Thus it looks
Ũ causes the elision of gates that have an exposed (1, −1, −1)
vertex (starting from the bottom right arrow). (c) This pro-
like the qubit array itself should change its geometric
cess can be iterated until no more vertices of that kind remain structure during the dynamics. However, in Appendix B
exposed. Then, unitarity of Ũ † elides all gates with vertex we show that it is possible to sidestep this issue and im-
(−1, 1, 1) exposed (starting from the top left arrow). (d) It- plement the desired circuit with local gates on a static
eration of this step results in disconnected tensors, and thus 2D array of qubits by introducing two ancillas for ev-
a vanishing connected correlator. ery system qubit. The qubits are arranged on a kagome
lattice, with system qubits occupying one of three sub-
lattices and ancillas occupying the other two (Fig. 13(b)
high-symmetry lines in the underlying lattice structure – shows only the system qubits). All ancillas are initialized
a kagome lattice (see Appendix B), shown in Fig. 13(b). in a fiduciary state, say |0i, and are returned to this ini-
Correlations travel along ξ, η and ζ at the same (maxi- tial state after each period; meanwhile, the system qubits
mal) velocity, and vanish everywhere else. undergo the desired time evolution68 .
This phenomenology generalizes that of (1 + 1)-
dimensional dual- and tri-unitary circuits in an interest-
ing way. Correlations are not pinned to manifolds of VII. CONCLUSION AND OUTLOOK
co-dimension 1 (the surface of a suitable light cone), as
one may have guessed, but rather to manifolds of di- The dynamics of many-body quantum systems out of
mension 1 (lines). This is particularly striking in non- equilibrium is a notoriously hard problem for both the-
ergodic models, where this result implies that some op- ory and computation. Models that afford a degree of
erators (eigenmodes with λµ,i = 1) move ballistically analytical control or solvability are extremely useful in
along one of three special directions in two-dimensional elucidating phenomena and principles that may apply to
space, Fig. 13; These directions are picked by the un- more general, less tractable scenarios; yet such solvable
derlying lattice structure and circuit architecture. This models are rare.
“subdimensional” propagation of information is reminis- In this work, we have introduced a new, large family
cent of fractonic excitations pinned to subdimensional of quantum many-body evolutions, dubbed tri-unitary
manifolds65–67 , albeit in a dynamical, driven setting, in circuits, in which crucial properties including correla-
which there is no meaningful notion of energy and the tions and entanglement are analytically tractable. This
only quantity moving along subdimensional manifolds is tractability stems from their peculiar causal structure,
information. which features three distinct “arrows of time” under
Finally we remark on implementing this construction which the dynamics are unitary. This builds on previous
 11

results on dual-unitary circuits, whose unitarity under First, while we have provided a large (31-parameter) fam-
two distinct arrows of time has enabled the derivation of ily of tri-unitary gates on qubits, it would be interest-
a plethora of exact results. ing to obtain a full parametrizations of all tri-unitary
Tri-unitary circuits generalize and extend the construc- gates (whether on qubits or higher-dimensional qudits).
tion of dual-unitary circuits in several important ways. Second, regarding the phenomenology of these dynam-
The different circuit architecture, featuring three-qubit ics, it would be interesting to obtain more general re-
gates arranged at the vertices of a triangular lattice in sults on the approach to thermalization, the growth of
spacetime, results in a different symmetry – rather than entanglement from generic (non-“solvable”) initial states,
exchanging space and time, it mixes the two nontriv- and other diagnostics of quantum chaos such as the spec-
ially. This has sharp consequences in the phenomenol- tral form factor and out-of-time-ordered correlators. Re-
ogy of these systems: correlations are allowed to propa- garding circuit architectures, we have argued that tri-
gate along three special directions in spacetime, namely unitary circuits saturate the number of possible “arrows
the light rays δx = ±vδt as well as the static worldline of time” in flat (1 + 1)-dimensional spacetime, due to the
δx = 0. Information can thus move at the “speed of absence of regular lattices with higher symmetry; how-
light” or not move at all – the latter a qualitatively dif- ever, more exotic generalizations may be possible, e.g.
ferent possibility absent in dual-unitary circuits. While via quasicrystalline tilings or on curved spaces – the latter
tri-unitary circuits are expected to be strongly chaotic potentially connecting to ideas in quantum gravity40,42 ,
in general, it is intriguing to speculate that this feature as well as recent explorations of quantum simulation in
(the presence of strictly non-moving operators) might in- curved spaces70,71 . Finally, we note that the triangular
spire constructions of tractable circuit models of localiza- circuit structure is not invariant under “spacetime du-
tion12 . A richer phenomenology also arises in the growth ality” – a π/2 rotation maps the circuit to a sequence
of quantum entanglement. In tri-unitary circuits (start- of non-local matrix-product operators (with finite bond
ing from a class of “solvable” initial states), entropy grows dimension represented by a spacelike qubit worldline).
ballistically at the maximal velocity, but only up to an This poses a challenge in deriving results about the spec-
entropy density of half the maximum, at which point the tral form factor of these circuits, as it makes the trans-
behavior may change based on the model (in contrast fer matrix employed in Ref. [8] non-unitary and non-
with dual-unitary circuits where the maximum-velocity local. On the other hand, this may present opportuni-
growth persists up to maximum density). ties for the study of non-unitary dynamics via spacetime
duality34–36 , by allowing access to potentially distinctive
Another novel aspect of tri-unitarity is the possibil- types of non-unitary evolutions involving matrix-product
ity of genuine higher-dimensional extensions. Higher- operators rather than local circuits.
dimensional constructions of dual-unitary circuits33 are
highly anisotropic: having two co-planar arrows of
time, they are effectively stacks of coupled (1 + 1)- ACKNOWLEDGMENTS
dimensional dual-unitary layers69 . In other words, the
action of spacetime duality by necessity exchanges time We acknowledge useful discussions with Tibor
and one spatial direction, leaving out any others36 . Rakovszky, Yuri Lensky, and Bruno Bertini. This work
On the contrary, we have constructed tri-unitary cir- was supported with funding from the Defense Advanced
cuits in (2 + 1)-dimensional spacetime that treat all di- Research Projects Agency (DARPA) via the DRINQS
mensions on the same footing, due to the presence of program (M.I.), the Sloan Foundation through a Sloan
three non-coplanar arrows of time. Interestingly, cor- Research Fellowship (V.K.) and by the US Department
relations in these circuits are pinned to three special of Energy, Office of Science, Basic Energy Sciences, un-
light-rays – correlations propagate at maximal velocity der Early Career Award No. DE-SC0021111 (V.K. and
along three high-symmetry directions of the underlying C.J.). The views, opinions and/or findings expressed are
lattice, at 2π/3 angles with each other. The propaga- those of the authors and should not be interpreted as rep-
tion of information along sub-dimensional manifolds is resenting the official views or policies of the Department
in itself a novel feature of these circuits, reminiscent of of Defense or the U.S. Government. M.I. was funded
quasiparticles with sub-dimensional mobility in fractonic in part by the Gordon and Betty Moore Foundation’s
systems65–67 , though in a completely different (driven, EPiQS Initiative through Grant GBMF8686. Numerical
non-equilibrium) context. simulations were performed on Stanford Research Com-
Our work opens several directions for future research. puting Center’s Sherlock cluster.

1
Luca D’Alessio, Yariv Kafri, Anatoli Polkovnikov, and https://doi.org/10.1080/00018732.2016.1198134.
2
Marcos Rigol, “From quantum chaos and eigenstate Rahul Nandkishore and David A. Huse, “Many-body local-
thermalization to statistical mechanics and thermo- ization and thermalization in quantum statistical mechan-
dynamics,” Advances in Physics 65, 239–362 (2016), ics,” Annual Review of Condensed Matter Physics 6, 15–
 12

38 (2015), https://doi.org/10.1146/annurev-conmatphys- channels,” Physical Review B 100 (2019), 10.1103/phys-

031214-014726. revb.100.064309.
3 20
B. Bertini, F. Heidrich-Meisner, C. Karrasch, T. Prosen, M Akila, D Waltner, B Gutkin, and T Guhr, “Particle-
R. Steinigeweg, and M. Žnidarič, “Finite-temperature time duality in the kicked ising spin chain,” Journal
transport in one-dimensional quantum lattice models,” of Physics A: Mathematical and Theoretical 49, 375101
Rev. Mod. Phys. 93, 025003 (2021). (2016).
4 21
Vincenzo Alba, Bruno Bertini, Maurizio Fagotti, Lorenzo Bruno Bertini, Pavel Kos, and Toma ž Prosen, “Entangle-
Piroli, and Paola Ruggiero, “Generalized-Hydrodynamic ment spreading in a minimal model of maximal many-body
approach to Inhomogeneous Quenches: Correlations, quantum chaos,” Phys. Rev. X 9, 021033 (2019).
22
Entanglement and Quantum Effects,” arXiv e-prints Bruno Bertini, Pavel Kos, and Tomaz Prosen, “Operator
, arXiv:2104.00656 (2021), arXiv:2104.00656 [cond- entanglement in local quantum circuits i: Chaotic dual-
mat.stat-mech]. unitary circuits,” SciPost Physics 8 (2020), 10.21468/sci-
5
Adam Nahum, Jonathan Ruhman, Sagar Vijay, and postphys.8.4.067.
23
Jeongwan Haah, “Quantum entanglement growth under Bruno Bertini, Pavel Kos, and Tomaz Prosen, “Opera-
random unitary dynamics,” Physical Review X 7 (2017), tor entanglement in local quantum circuits ii: Solitons in
10.1103/physrevx.7.031016. chains of qubits,” SciPost Physics 8 (2020), 10.21468/sci-
6
C. W. von Keyserlingk, Tibor Rakovszky, Frank Pollmann, postphys.8.4.068.
24
and S. L. Sondhi, “Operator hydrodynamics, otocs, and en- Pavel Kos, Bruno Bertini, and Toma ž Prosen, “Corre-
tanglement growth in systems without conservation laws,” lations in perturbed dual-unitary circuits: Efficient path-
Phys. Rev. X 8, 021013 (2018). integral formula,” Phys. Rev. X 11, 011022 (2021).
7 25
Vedika Khemani, David A. Huse, and Adam Nahum, Lorenzo Piroli, Bruno Bertini, J. Ignacio Cirac, and
“Velocity-dependent lyapunov exponents in many-body Tomaž Prosen, “Exact dynamics in dual-unitary quantum
quantum, semiclassical, and classical chaos,” Physical Re- circuits,” Physical Review B 101 (2020), 10.1103/phys-
view B 98 (2018), 10.1103/physrevb.98.144304. revb.101.094304.
8 26
Bruno Bertini, Pavel Kos, and Toma ž Prosen, “Exact Boris Gutkin, Petr Braun, Maram Akila, Daniel Walt-
spectral form factor in a minimal model of many-body ner, and Thomas Guhr, “Local correlations in dual-unitary
quantum chaos,” Phys. Rev. Lett. 121, 264101 (2018). kicked chains,” arXiv e-prints , arXiv:2001.01298 (2020),
9
Adam Nahum, Sagar Vijay, and Jeongwan Haah, “Oper- arXiv:2001.01298 [cond-mat.stat-mech].
27
ator spreading in random unitary circuits,” Phys. Rev. X Pieter W. Claeys and Austen Lamacraft, “Maximum veloc-
8, 021014 (2018). ity quantum circuits,” Physical Review Research 2 (2020),
10
Vedika Khemani, Ashvin Vishwanath, and David A. 10.1103/physrevresearch.2.033032.
28
Huse, “Operator spreading and the emergence of dissipa- Pieter W. Claeys and Austen Lamacraft, “Ergodic and
tive hydrodynamics under unitary evolution with conser- nonergodic dual-unitary quantum circuits with arbitrary
vation laws,” Physical Review X 8 (2018), 10.1103/phys- local hilbert space dimension,” Physical Review Letters
revx.8.031057. 126 (2021), 10.1103/physrevlett.126.100603.
11 29
Tibor Rakovszky, Frank Pollmann, and C. W. von Keyser- Ana Flack, Bruno Bertini, and Toma ž Prosen, “Statis-
lingk, “Diffusive hydrodynamics of out-of-time-ordered cor- tics of the spectral form factor in the self-dual kicked ising
relators with charge conservation,” Phys. Rev. X 8, 031058 model,” Phys. Rev. Research 2, 043403 (2020).
30
(2018). Bruno Bertini, Pavel Kos, and Tomaž Prosen, “Random
12
Anushya Chandran and C. R. Laumann, “Semiclassical matrix spectral form factor of dual-unitary quantum cir-
limit for the many-body localization transition,” Physical cuits,” Communications in Mathematical Physics (2021),
Review B 92 (2015), 10.1103/physrevb.92.024301. 10.1007/s00220-021-04139-2.
13 31
Winton Brown and Omar Fawzi, “Scrambling speed of ran- Isaac Reid and Bruno Bertini, “Entanglement barriers in
dom quantum circuits,” arXiv e-prints , arXiv:1210.6644 dual-unitary circuits,” Phys. Rev. B 104, 014301 (2021).
32
(2012), arXiv:1210.6644 [quant-ph]. Felix Fritzsch and Toma ž Prosen, “Eigenstate thermal-
14
Patrick Hayden and John Preskill, “Black holes as mirrors: ization in dual-unitary quantum circuits: Asymptotics of
quantum information in random subsystems,” Journal of spectral functions,” Phys. Rev. E 103, 062133 (2021).
33
High Energy Physics 2007, 120–120 (2007). Ryotaro Suzuki, Kosuke Mitarai, and Keisuke Fujii, “Com-
15
Amos Chan, Andrea De Luca, and J. T. Chalker, “Solu- putational power of one- and two-dimensional dual-unitary
tion of a minimal model for many-body quantum chaos,” quantum circuits,” (2021), arXiv:2103.09211 [quant-ph].
34
Physical Review X 8 (2018), 10.1103/physrevx.8.041019. Matteo Ippoliti and Vedika Khemani, “Postselection-free
16
S. J. Garratt and J. T. Chalker, “Many-body delocaliza- entanglement dynamics via spacetime duality,” Phys. Rev.
tion as symmetry breaking,” Phys. Rev. Lett. 127, 026802 Lett. 126, 060501 (2021).
35
(2021). Matteo Ippoliti, Tibor Rakovszky, and Vedika Khe-
17
Aaron J. Friedman, Amos Chan, Andrea De Luca, and mani, “Fractal, logarithmic and volume-law entangled non-
J. T. Chalker, “Spectral statistics and many-body quantum thermal steady states via spacetime duality,” arXiv e-prints
chaos with conserved charge,” Physical Review Letters 123 , arXiv:2103.06873 (2021), arXiv:2103.06873 [quant-ph].
36
(2019), 10.1103/physrevlett.123.210603. Tsung-Cheng Lu and Tarun Grover, “Entanglement transi-
18
Bruno Bertini, Pavel Kos, and Tomaž Prosen, “Exact tions via space-time rotation of quantum circuits,” arXiv e-
correlation functions for dual-unitary lattice models in prints , arXiv:2103.06356 (2021), arXiv:2103.06356 [quant-
1+1 dimensions,” Physical Review Letters 123 (2019), ph].
37
10.1103/physrevlett.123.210601. Michael Foss-Feig, David Hayes, Joan M. Dreiling, Caro-
19
Sarang Gopalakrishnan and Austen Lamacraft, “Unitary line Figgatt, John P. Gaebler, Steven A. Moses, Juan M.
circuits of finite depth and infinite width from quantum Pino, and Andrew C. Potter, “Holographic quantum algo-
 13

rithms for simulating correlated spin systems,” Phys. Rev. the context of holographic quantum error correcting codes,
Research 3, 033002 (2021). in Ref. [73] as “perfect tangles” for modular tensor cate-
38
Eli Chertkov, Justin Bohnet, David Francois, John Gae- gories, and in in Ref. [74] as “planar maximally entangled
bler, Dan Gresh, Aaron Hankin, Kenny Lee, Ra’anan states”.
53
Tobey, David Hayes, Brian Neyenhuis, Russell Stutz, Charles H. Bennett, David P. DiVincenzo, John A. Smolin,
Andrew C. Potter, and Michael Foss-Feig, “Holo- and William K. Wootters, “Mixed-state entanglement and
graphic dynamics simulations with a trapped ion quan- quantum error correction,” Phys. Rev. A 54, 3824–3851
tum computer,” arXiv e-prints , arXiv:2105.09324 (2021), (1996).
54
arXiv:2105.09324 [quant-ph]. Raymond Laflamme, Cesar Miquel, Juan Pablo Paz, and
39
Brian Swingle, “Spacetime from entanglement,” An- Wojciech Hubert Zurek, “Perfect quantum error correcting
nual Review of Condensed Matter Physics 9, 345– code,” Phys. Rev. Lett. 77, 198–201 (1996).
55
358 (2018), https://doi.org/10.1146/annurev-conmatphys- Wolfram Helwig, Wei Cui, José Ignacio Latorre, Arnau Ri-
033117-054219. era, and Hoi-Kwong Lo, “Absolute maximal entanglement
40
Patrick Hayden, Sepehr Nezami, Xiao-Liang Qi, Nathaniel and quantum secret sharing,” Phys. Rev. A 86, 052335
Thomas, Michael Walter, and Zhao Yang, “Holographic (2012).
56
duality from random tensor networks,” Journal of High Dardo Goyeneche, Daniel Alsina, José I. Latorre, Arnau
Energy Physics 2016, 9 (2016). Riera, and Karol Życzkowski, “Absolutely maximally en-
41
Jordan Cotler, Xizhi Han, Xiao-Liang Qi, and Zhao tangled states, combinatorial designs, and multiunitary
Yang, “Quantum causal influence,” Journal of High Energy matrices,” Phys. Rev. A 92, 032316 (2015).
57
Physics 2019 (2019), 10.1007/jhep07(2019)042. Tomasz Linowski, Grzegorz Rajchel-Mieldzioć, and Karol
42
Fernando Pastawski, Beni Yoshida, Daniel Harlow, and Życzkowski, “Entangling power of multipartite unitary
John Preskill, “Holographic quantum error-correcting gates,” Journal of Physics A: Mathematical and Theoreti-
codes: toy models for the bulk/boundary correspon- cal 53, 125303 (2020).
58
dence,” Journal of High Energy Physics 2015 (2015), Jose I. Latorre and German Sierra, “Holographic codes,”
10.1007/jhep06(2015)149. arXiv e-prints , arXiv:1502.06618 (2015), arXiv:1502.06618
43
Frank Arute, Kunal Arya, Ryan Babbush, Dave Bacon, [quant-ph].
59
et al., “Quantum supremacy using a programmable super- A similar construction of “planar maximally entangled
conducting processor,” Nature 574, 505–510 (2019). states” in Ref. [74] yields a subset of this family.
44 60
Xiao Mi, Pedram Roushan, Chris Quintana, Salvatore The gates defined in Eq. (6) have at most two bits of
Mandra, et al., “Information Scrambling in Computa- entropy between non-contiguous bipartitions of the type
tionally Complex Quantum Circuits,” arXiv e-prints , (1, 2, 6) 7→ (3, 4, 5). As such, they cannot express the per-
arXiv:2101.08870 (2021), arXiv:2101.08870 [quant-ph]. fect tensor, whose entropy across any bipartition of the 6
45
Matteo Ippoliti, Kostyantyn Kechedzhi, Roderich Moess- legs is maximal (3 bits).
61
ner, S. L. Sondhi, and Vedika Khemani, “Many-body Romain Vasseur, Andrew C. Potter, Yi-Zhuang You, and
physics in the NISQ era: quantum programming a discrete Andreas W. W. Ludwig, “Entanglement transitions from
time crystal,” arXiv e-prints , arXiv:2007.11602 (2020), holographic random tensor networks,” Phys. Rev. B 100,
arXiv:2007.11602 [cond-mat.dis-nn]. 134203 (2019).
46
Brian Skinner, Jonathan Ruhman, and Adam Nahum, 62 −1
In general one has (σ|τ ) = q |σ τ | , where |g| denotes the
“Measurement-induced phase transitions in the dynamics number of cycles in the permutation g. In our case, g = χ
of entanglement,” Phys. Rev. X 9, 031009 (2019). contains a single cycle.
47
Yaodong Li, Xiao Chen, and Matthew P. A. Fisher, 63
Don N. Page, “Average entropy of a subsystem,” Physical
“Quantum zeno effect and the many-body entanglement Review Letters 71, 1291–1294 (1993).
transition,” Phys. Rev. B 98, 205136 (2018). 64
Unitarity of the mapping (x̄, y, z̄) 7→ (x, ȳ, z), correspond-
48
Michael J. Gullans and David A. Huse, “Dynamical pu- ing to the last pair of vertices of the cube, is not assumed.
rification phase transition induced by quantum measure- 65
Rahul M. Nandkishore and Michael Hermele, “Fractons,”
ments,” Phys. Rev. X 10, 041020 (2020). Annual Review of Condensed Matter Physics 10, 295–
49
Michael Foss-Feig, Stephen Ragole, Andrew Potter, Joan 313 (2019), https://doi.org/10.1146/annurev-conmatphys-
Dreiling, Caroline Figgatt, John Gaebler, Alex Hall, 031218-013604.
Steven Moses, Juan Pino, Ben Spaun, Brian Neyenhuis, 66
Sagar Vijay, Jeongwan Haah, and Liang Fu, “A new kind
and David Hayes, “Entanglement from tensor networks on of topological quantum order: A dimensional hierarchy
a trapped-ion QCCD quantum computer,” arXiv e-prints of quasiparticles built from stationary excitations,” Phys.
, arXiv:2104.11235 (2021), arXiv:2104.11235 [quant-ph]. Rev. B 92, 235136 (2015).
50
Crystal Noel, Pradeep Niroula, Andrew Risinger, Laird 67
Michael Pretko, “Subdimensional particle structure of
Egan, Debopriyo Biswas, Marko Cetina, Alexey V. Gor- higher rank u(1) spin liquids,” Phys. Rev. B 95, 115139
shkov, Michael Gullans, David A. Huse, and Christopher (2017).
Monroe, “Observation of measurement-induced quantum 68
Quantum circuits enhanced with ancillas in this way are
phases in a trapped-ion quantum computer,” arXiv e-prints equivalent to the class of quantum cellular automata75–78 ,
, arXiv:2106.05881 (2021), arXiv:2106.05881 [quant-ph]. i.e. unitary transformations that preserve the locality of
51
Namely, we have M+ (a) ≡ Tr1 (U † a1 U ) and M− (a) ≡ operators but are not necessarily realizable via finite-depth
Tr2 (U † a2 U ) where U acts on qubits 1, 2 and a1 ≡ a ⊗ 1, local circuits.
a2 ≡ 1 ⊗ a. 69
52
This definition of dual-unitary circuits, based on the
We note that tensors obeying these same constraints (for number of unitary arrows of time, excludes the higher-
generic numbers of legs, thus also including dual-unitarity) dimensional generalizations mentioned in Ref. [18].
were introduced in Ref. [72] as “block-perfect tensors” in
 14

70
Alicia J. Kollár, Mattias Fitzpatrick, and Andrew A. maximally entangled states,” Phys. Rev. A 102, 012427
Houck, “Hyperbolic lattices in circuit quantum electrody- (2020).
75
namics,” Nature 571, 45–50 (2019). B. Schumacher and R. F. Werner, “Reversible quantum cel-
71
Igor Boettcher, Przemyslaw Bienias, Ron Belyansky, Ali- lular automata,” arXiv e-prints , quant-ph/0405174 (2004),
cia J. Kollár, and Alexey V. Gorshkov, “Quantum sim- arXiv:quant-ph/0405174 [quant-ph].
76
ulation of hyperbolic space with circuit quantum electro- P. Arrighi, “An overview of quantum cellular automata,”
dynamics: From graphs to geometry,” Phys. Rev. A 102, Natural Computing 18, 885–899 (2019).
77
032208 (2020). Terry Farrelly, “A review of Quantum Cellular Automata,”
72
Robert J. Harris, Nathan A. McMahon, Gavin K. Bren- Quantum 4, 368 (2020).
78
nen, and Thomas M. Stace, “Calderbank-shor-steane holo- Lorenzo Piroli and J. Ignacio Cirac, “Quantum cellular au-
graphic quantum error-correcting codes,” Phys. Rev. A 98, tomata, tensor networks, and area laws,” Phys. Rev. Lett.
052301 (2018). 125, 190402 (2020).
73 79
Johannes Berger and Tobias J. Osborne, “Perfect tangles,” S. Aravinda, Suhail Ahmad Rather, and Arul Lakshmi-
arXiv e-prints , arXiv:1804.03199 (2018), arXiv:1804.03199 narayan, “From dual-unitary to quantum Bernoulli cir-
[quant-ph]. cuits: Role of the entangling power in constructing a quan-
74
Mehregan Doroudiani and Vahid Karimipour, “Planar tum ergodic hierarchy,” arXiv e-prints , arXiv:2101.04580
(2021), arXiv:2101.04580 [quant-ph].

Appendix A: Examples of correlations

In this appendix we show that a two-parameter family of gates within the parametrization of Eq. (6) can realize all
the correlation behaviors of the “ergodic hierarchy” reviewed in Sec. II. We define the gate
P P
Zi −ig
U (φ, g) = SWAP3,1 CP3,1 (φ) CP2,3 (φ) CP1,2 (φ)eiφ/2 i e i Xi

φ P
= SWAP3,1 e−i 2 (Z1 Z2 +Z2 Z3 +Z3 Z1 ) e−ig i Xi
(A1)
which is Ut.u. from Eq. (6) with φi ≡ φ, ui ≡ vi ≡ 1, and wi ≡ e−igXi . We can compute the transfer matrices
 
1 0 0 0
  2 
0 cos φ2 0 0 
  2 
Mµ =  0 φ

 (A2)
 0 cos(2g) cos 2 sin(2g) 
  2 
φ
0 0 − sin(2g) cos 2 cos(2g)

which are independent of µ = ±, 0 due to the symmetry of the gate. Varying parameters φ and g, we can realize the
entire “ergodic hierarchy” of correlation functions:
(i) non-interacting: setting g = φ = 0 yields U(i) = U (0, 0) = SWAP3,1 . The transfer matrices are Mµ = I (the
identity channel), thus the eigenvalues
λ(i)
µ = {1, 1, 1, 1} (A3)
2
(ii) interacting, non-ergodic: setting φ 6= 0 with g = 0 gives (up to a global phase) U(ii) = SWAP1,3 e−i(φ/8)Ztot
where Ztot = Z1 + Z2 + Z3 ; thus two-point functions of Zi remain constant along the rays while those of Xi , Yi
decay:
2 2
λ(ii)
µ = {1, cos(φ/2) , cos(φ/2) , 1} (A4)

(iii) ergodic, non-mixing: setting g = π/2 adds a π-pulse about the x axis to the previous drive, U(iii) =
2
SWAP1,3 e−i(φ/8)Ztot X1 X2 X3 , causing Z correlations to oscillate:
(iii) 2 2
λ0,± = {1, − cos(φ/2) , cos(φ/2) , −1} (A5)
Thus time-averaged Z correlators decay, but instantaneous ones do not.
(iv) ergodic, mixing: obtained for generic values of g, φ. We find the eigenvalues
2 1 1 1 1
λ(iv)
µ = {1, cos(φ/2) , cos(2g)(3 + cos(φ)) − f (φ, g), cos(2g)(3 + cos(φ)) + f (φ, g)} (A6)
p 4 8 4 8
f (φ, g) = −13 − 20 cos(φ) + 2 cos(4g)(3 + cos(φ))2 + cos(2φ)) (A7)
 15
14

= =

i z
=e 2
ig x
=e 2

FIG. 14. Left: case (i), non-interactingFIG.

gate.14.
Right:
Left: cases (ii)-(iv),
case (i), right:equal
case CP gates with variable single-qubit rotations.
(ii)-(iv)

whichwith
are independent of µ
eigenoperators 0 due
±,h(φ,
{1,=X, g)Yto±the
Z},symmetry
with of the gate. Varying parameters and g, we can realize the
entire “ergodic hierarchy” of correlation functions:
2
(i) non-interacting: setting h(φ, g) ==e2i(3g+φ)
0 yieldsf U
(φ, g)=+U4e 2i(3g+φ)
cos(2g) .sin(φ/2) . (A8)
g= (i) (0, 0) = SWAP 3,1 The transfer matrices are Mµ = I (the
identity channel), thus the eigenvalues
Thus all correlations decay exponentially, without time averaging.
(i)
Finally one may add an extreme case, dubbed “Bernoulliµ = {1, 1, 1, 1}79 , These are represented by the perfect tensor42
circuits” (A3)
, in
which all transfer matrices are erasure channels, Mµ (O) = Tr(O)1/2, and thus all correlators decay to 0 immediately.
This
(ii) property follows
interacting, from quantum
non-ergodic: error correction,
setting 6= 0 with gnamely from
= 0 gives (upthetofact that no
a global information
phase) about1,3the
U(ii) = SWAP e i(encoded
2
/8)Ztot
qubit where
shouldZbe accessible from any single one of the physical qubits. (The Mµ along
tot = Z1 + Z2 + Z3 ; thus two-point functions of Zi remain constant
channels
the correspond to encoding
rays while those of Xi , Yai
logicaldecay:
qubit into 5 physical qubits, discarding (tracing out) 4 of them, and retaining the last one as output).
(ii) 2 2
µ = {1, cos( /2) , cos( /2) , 1} (A4)
Appendix B: Details on two-dimensional circuit
(iii) ergodic, non-mixing: setting g = ⇡/2 adds a ⇡-pulse about the x axis to the previous drive, U(iii) =
HereSWAP
we discuss i( the implementation
/8)Z 2
tot X X X , causing of theZ3-dimensional
correlations to tri-unitary
oscillate: tensor network of Sec. VI as a quantum circuit
1,3 e 1 2 3
in two spatial dimensions.
To review, our construction involves a(iii) cubic lattice (x, /2) 2 ∈ Z3 with
y, z) , cos( /2) , tri-unitary
2 gates at each vertex; the qubit(A5)
0,± = {1, cos( 1}
worldlines travel in three directions, x, y and z (given by x ∈ R, y, z ∈ Z and permutations thereof), intersecting
at gates;
Thuseach gate thus has
time-averaged 3 input qubits
Z correlators decay,andbut 3instantaneous
output qubits. onesThere
do not.are three equally valid axes of time, we
take t = x + y + z as the physical or “laboratory” time for concreteness. Expressing the above tensor network as
a (iv) ergodic, mixing:
two-dimensional quantum obtained
circuitforis generic valuesstraightforward
not entirely of g,
√ . We find the eigenvalues
because the lattice
√ configuration of qubits in the
space-like plane t⊥ (spanned e.g. by ξ = (2, −1, −1)/ 6 and η = (−1, −1, 2)/ 6) changes during the dynamics:
naïvely, the qubits (iv) would have to 2 be 1 physically moved on the 1 plane1during the evolution. 1However, this can be
µ = {1, cos( /2) , cos(2g)(3 + cos( )) f ( , g), cos(2g)(3 + cos( )) + f ( , g)} (A6)
avoided by introducingp ancilla qubits,4 as we explain in the following. 8 4 8
Let us for simplicity
f ( , g) = assume 13 all gates )U+ in
20 cos( the circuit
2 cos(4g)(3 are equal;
+ cos( ))2 + the
cos(2circuit
)) then is time-periodic, with a Floquet (A7)
period of t = 3: (x, y, z) 7→ (x + 1, y + 1, z + 1) is the shortest translation along the temporal direction t that leaves
with eigenoperators
the lattice invariant. Three {1,layers
X, h(of, g)Y ± Z},
unitary withtake place within each Floquet period – at times t = 3n, 3n + 1,
gates
and 3n + 2. In between two layers of unitary gates, say at t = 3n + k + 1/2, the location 2
of qubits on the spatial plane
is given by intersecting the qubit g) = e2i(3g+
h( ,worldlines )
withf (the + 4e2i(3g+
, g)planes )
x + y +cos(2g)
z = k sin( /2)which
+ 1/2, . yields a kagome lattice(A8)for
any k ∈ Z3 . However, the three kagome lattices are distinct: namely the blue sublattice in Fig. 15 for k = 0, the red
one forThus
k = all correlations
1, and the greendecay exponentially,
one for k = 2. The without
union oftime theseaveraging.
three lattices is itself a kagome lattice.
In order to implement the (2 + 1)-dimensional tri-unitary
Finally one may add an extreme case, dubbed Bernoulling circuits69 , These dynamics as a local circuit in 2-dimensional
are represented by the perfectspace,
tensorwe42
,
place a physical qubit on every site of the kagome lattice obtained above; however, the
in which all transfer matrices are erasure channels, Mµ (O) = Tr(O)1/2, and thus all correlators decay to 0 immediately. state of interest |ψi is stored
only
This on the blue
property sublattice,
follows while the error
from quantum rest ofcorrection,
the qubitsnamely(red and green
from thesublattices)
fact that noareinformation
ancillas initialized
about thein encoded
a trivial
⊗2N
product state, say . We then perform the tri-unitary gates of Eq. (6)
qubit should be accessible from any single one of the physical qubits. (The Mµ channels correspond to encoding
|0i by acting on triplets of blue sites witha
controlled-phase gates (thick lines in Fig. 15), plus any single-qubit rotations;
logical qubit into 5 physical qubits, discarding (tracing out) 4 of them, and retaining the last one as output). then we swap each qubit with the
diametrically opposite vertex of the hexagonal plaquette where the gate has acted (arrows in Fig. 15). As a result,
⊗N
the blue sites are now occupied by trivial states |0i , while the state of interest is written on the red sites. This
process implements the t = 0 layer Appendix
of unitary B: gates.
DetailsThe on ttwo-dimensional
= 1 and 2 layers circuitare implemented analogously, as shown
in the other panels of Fig. 15, with the state of interest moving to the green sites and finally back to the blue sites,
completing a Floquet
Here we discuss the cycle. Because after
implementation of thea 3-dimensional
period the state of the ancillas
tri-unitary tensor is unchanged,
network of Sec.this evolution
VI as belongs
a quantum to
circuit
the
in two class of
QCaspatial quantum circuits augmented by ancillas. This class is equivalent to quantum cellular automata75–78 .
dimensions.
 16

| ψ⟩ | 0⟩ | 0⟩ | 0⟩ | ψ⟩ | 0⟩ | 0⟩ | 0⟩ | ψ⟩

FIG. 15. (2 + 1)-dimensional tri-unitary circuit implementation in two spatial dimensions. Qubits are arranged on a kagome
lattice and partitioned into three sublattices (red, green and blue). The state of interest occupies only one sublattice at any
given time; the other two store ancillas in a fixed product state, e.g. |0i⊗2N . A floquet cycle consists of three layers of unitary
gates. Before the first layer (left), the state |ψi lives in the blue sublattice. Tri-unitary gates (thick lines) couple triplets of blue
qubits as shown; then, SWAP gates (thin arrows) act on pairs of blue and red qubits, moving the state to the red sublattice.
The second (center) and third (right) layers of unitary gates are applied similarly. The state of interest changes sublattice at
each layer, and returns to the blue sublattice after three layers (a Floquet period, if the gates are time-independent).

We note that in this case there is no fundamental obstruction to realizing the evolution as a low-depth local circuit,
merely a technical inconvenience (the three-qubit gates would have to act on triplets of qubits separated by several
lattice spacings, as opposed to the present case where all interactions take place around a plaquette).
If all interactions are turned off, the system evolves by a sequence of SWAP gates along three special directions; it
is clear then that opereators propagate ballistically, at fixed velocity, along one of three high-symmetry directions in
the kagome lattice, at 2π/3 angles with each other. These are the projections of x, y and z on the t⊥ plane. More
surprisingly, this phenomenology is robust to the addition of arbitrary tri-unitary interactions, except correlators on
the special rays generically decay exponentially in time rather than being 1 – their behavior is again dictated by the
quantum channels M0,± , as in the (1 + 1)-dimensional case.