Typical correlations and entanglement in random tensor network states
Mostly based on joint work with David Pérez-García, available at arXiv:1906.11682
Cécilia Lancien
Institut Fourier Grenoble & CNRS
Journée théorie CPTGA – September 27 2021
Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 1
Outline
1 Background and motivations
2 Random MPS and PEPS: construction and statement of the main questions & results
3 Typical correlation length in a random TNS through typical spectral gap of its transfer operator
4 Typical correlation length in a random TNS through spectral gap of its parent Hamiltonian
5 Open questions and perspectives
Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 2
The curse of dimensionality when dealing with many-body quantum systems
A quantum system composed ofN d-dimensional subsystems has dimensiondN.
−→
Exponential growth of the dimension with the number of subsystems.However,‘physically relevant’ statesof many-body quantum systems are often well approximated by so-calledtensor network states (TNS), which form a small subset of the global state space.
Tensor network state construction on
(
Cd)
⊗N: Take a graphGwithNvertices andLedges.Put at each vertexva tensor
|χ
vi ∈
Cd⊗ (
Cq)
⊗δ(v)to get a tensor|˜ χ
Gi ∈ (
Cd)
⊗N⊗ (
Cq)
⊗2L. Contract together the indices of|˜ χ
Gi
associated to a same edge to get a tensor|χ
Gi ∈ (
Cd)
⊗N.degree ofv
pure state on(Cd)⊗N(up to normalization)
−→
Ifδ(
v) 6
rfor allv, then such state is described by at mostNqrdparameters (linear inN).• •
• •
• • Gwith 6 vertices and 7 edges
• •
• •
• •
|˜ χ
Gi ∈ (
Cd)
⊗6⊗ (
Cq)
⊗14• •
• •
• •
|χ
Gi ∈ (
Cd)
⊗6d-dimensional indices:physicalindices.q-dimensional indices:bondindices. In this talk:The underlying graphGis a regular lattice in dimension 1 or 2.
−→ |χ
Gi
is amatrix product state (MPS)or aprojected entangled pair state (PEPS).Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 3
The curse of dimensionality when dealing with many-body quantum systems
A quantum system composed ofN d-dimensional subsystems has dimensiondN.
−→
Exponential growth of the dimension with the number of subsystems.However,‘physically relevant’ statesof many-body quantum systems are often well approximated by so-calledtensor network states (TNS), which form a small subset of the global state space.
Tensor network state construction on
(
Cd)
⊗N: Take a graphGwithNvertices andLedges.Put at each vertexva tensor
|χ
vi ∈
Cd⊗ (
Cq)
⊗δ(v)to get a tensor|˜ χ
Gi ∈ (
Cd)
⊗N⊗ (
Cq)
⊗2L. Contract together the indices of|˜ χ
Gi
associated to a same edge to get a tensor|χ
Gi ∈ (
Cd)
⊗N.degree ofv
pure state on(Cd)⊗N(up to normalization)
−→
Ifδ(
v) 6
rfor allv, then such state is described by at mostNqrdparameters (linear inN).• •
• •
• • Gwith 6 vertices and 7 edges
• •
• •
• •
|˜ χ
Gi ∈ (
Cd)
⊗6⊗ (
Cq)
⊗14• •
• •
• •
|χ
Gi ∈ (
Cd)
⊗6 d-dimensional indices:physicalindices.q-dimensional indices:bondindices.In this talk:The underlying graphGis a regular lattice in dimension 1 or 2.
−→ |χ
Gi
is amatrix product state (MPS)or aprojected entangled pair state (PEPS).Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 3
The curse of dimensionality when dealing with many-body quantum systems
A quantum system composed ofN d-dimensional subsystems has dimensiondN.
−→
Exponential growth of the dimension with the number of subsystems.However,‘physically relevant’ statesof many-body quantum systems are often well approximated by so-calledtensor network states (TNS), which form a small subset of the global state space.
Tensor network state construction on
(
Cd)
⊗N: Take a graphGwithNvertices andLedges.Put at each vertexva tensor
|χ
vi ∈
Cd⊗ (
Cq)
⊗δ(v)to get a tensor|˜ χ
Gi ∈ (
Cd)
⊗N⊗ (
Cq)
⊗2L. Contract together the indices of|˜ χ
Gi
associated to a same edge to get a tensor|χ
Gi ∈ (
Cd)
⊗N.degree ofv
pure state on(Cd)⊗N(up to normalization)
−→
Ifδ(
v) 6
rfor allv, then such state is described by at mostNqrdparameters (linear inN).• •
• •
• • Gwith 6 vertices and 7 edges
• •
• •
• •
|˜ χ
Gi ∈ (
Cd)
⊗6⊗ (
Cq)
⊗14• •
• •
• •
|χ
Gi ∈ (
Cd)
⊗6d-dimensional indices:physicalindices.q-dimensional indices:bondindices.
In this talk:The underlying graphGis a regular lattice in dimension 1 or 2.
−→ |χ
Gi
is amatrix product state (MPS)or aprojected entangled pair state (PEPS).Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 3
Motivations behind TNS and related questions
A TNS representation is thought to be amathematically tractabledescription of the state of a system composed ofmany subsystems having a certain geometryand subject to interactions respecting this geometry.
Example:Being a TNS is (conjectured to be) approximately equivalent to being a ground state of a gapped local Hamiltonian(Hastings, Landau/Vazirani/Vidick...)
composed of terms which act non-trivially only on nearby sites spectral gap lower bounded by a constant independent ofN
−→
In condensed-matter physics, TNS are used as Ansatz in computations: optimization over a manageable number of parameters, even for largeN.General question:Among the variousqualitativeintuitions about TNS, which ones are at least true generically? And canquantitativestatements be made about the generic case?
Idea:Sample a TNS at random and study the characteristics that it exhibits with high probability. In particular, what is itstypical amount of correlations and entanglement?
Regime we are interested in: larged
,
q, ideally anyN.Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 4
Motivations behind TNS and related questions
A TNS representation is thought to be amathematically tractabledescription of the state of a system composed ofmany subsystems having a certain geometryand subject to interactions respecting this geometry.
Example:Being a TNS is (conjectured to be) approximately equivalent to being a ground state of a gapped local Hamiltonian(Hastings, Landau/Vazirani/Vidick...)
composed of terms which act non-trivially only on nearby sites spectral gap lower bounded by a constant independent ofN
−→
In condensed-matter physics, TNS are used as Ansatz in computations: optimization over a manageable number of parameters, even for largeN.General question:Among the variousqualitativeintuitions about TNS, which ones are at least true generically? And canquantitativestatements be made about the generic case?
Idea:Sample a TNS at random and study the characteristics that it exhibits with high probability. In particular, what is itstypical amount of correlations and entanglement?
Regime we are interested in: larged
,
q, ideally anyN.Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 4
Motivations behind TNS and related questions
A TNS representation is thought to be amathematically tractabledescription of the state of a system composed ofmany subsystems having a certain geometryand subject to interactions respecting this geometry.
Example:Being a TNS is (conjectured to be) approximately equivalent to being a ground state of a gapped local Hamiltonian(Hastings, Landau/Vazirani/Vidick...)
composed of terms which act non-trivially only on nearby sites spectral gap lower bounded by a constant independent ofN
−→
In condensed-matter physics, TNS are used as Ansatz in computations: optimization over a manageable number of parameters, even for largeN.General question:Among the variousqualitativeintuitions about TNS, which ones are at least true generically? And canquantitativestatements be made about the generic case?
Idea:Sample a TNS at random and study the characteristics that it exhibits with high probability. In particular, what is itstypical amount of correlations and entanglement?
Regime we are interested in: larged
,
q, ideally anyN.Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 4
Motivations behind TNS and related questions
A TNS representation is thought to be amathematically tractabledescription of the state of a system composed ofmany subsystems having a certain geometryand subject to interactions respecting this geometry.
Example:Being a TNS is (conjectured to be) approximately equivalent to being a ground state of a gapped local Hamiltonian(Hastings, Landau/Vazirani/Vidick...)
composed of terms which act non-trivially only on nearby sites spectral gap lower bounded by a constant independent ofN
−→
In condensed-matter physics, TNS are used as Ansatz in computations: optimization over a manageable number of parameters, even for largeN.General question:Among the variousqualitativeintuitions about TNS, which ones are at least true generically? And canquantitativestatements be made about the generic case?
Idea:Sample a TNS at random and study the characteristics that it exhibits with high probability.
In particular, what is itstypical amount of correlations and entanglement?
Regime we are interested in: larged
,
q, ideally anyN.Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 4
Outline
1 Background and motivations
2 Random MPS and PEPS: construction and statement of the main questions & results
3 Typical correlation length in a random TNS through typical spectral gap of its transfer operator
4 Typical correlation length in a random TNS through spectral gap of its parent Hamiltonian
5 Open questions and perspectives
Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 5
A simple model of random translation-invariant MPS
Nparticles on a circle• • • • • N
Pick a tensor
|χi ∈
Cd⊗ (
Cq)
⊗2whose entries are independent complex Gaussians with mean 0 and variance 1/
dq.Repeat it on all sites and contract neighboringq-dimensional indices.
−→
Obtained tensor|χ
Ni ∈ (
Cd)
⊗N:random translation-invariant MPS with periodic boundary conditions(typically almost normalized).d q q
|χi = ∑
dx=1 q
∑
i,j=1
gxij
|
xiji
1-site tensorN
|χ
Ni =
d
∑
x1,...,xN=1
q∑
i1,...,iN=1
gx1iNi1
· · ·
gxNiN−1iN|
x1· · ·
xNi
Associatedtransfer operator:T
:
Cq⊗
Cq→
Cq⊗
Cq, obtained by contracting thed-dimensional indices of|χi
and|¯ χi
.T
= ∑
dx=1
q∑
i,j,k,l=1
gxij
¯
gxkl|
ikih
jl|
=
1d∑
dx=1
Gx
⊗
G¯
xTheGx’s are independentq
×
qmatrices whose entries are independent complex Gaussians with mean 0 and variance 1/
q.Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 6
A simple model of random translation-invariant MPS
Nparticles on a circle• • • • • N
Pick a tensor
|χi ∈
Cd⊗ (
Cq)
⊗2whose entries are independent complex Gaussians with mean 0 and variance 1/
dq.Repeat it on all sites and contract neighboringq-dimensional indices.
−→
Obtained tensor|χ
Ni ∈ (
Cd)
⊗N:random translation-invariant MPS with periodic boundary conditions(typically almost normalized).d q q
|χi = ∑
dx=1 q
∑
i,j=1
gxij
|
xiji
1-site tensorN
|χ
Ni =
d
∑
x1,...,xN=1
q∑
i1,...,iN=1
gx1iNi1
· · ·
gxNiN−1iN|
x1· · ·
xNi
Associatedtransfer operator:T
:
Cq⊗
Cq→
Cq⊗
Cq, obtained by contracting thed-dimensional indices of|χi
and|¯ χi
.T
=
d
∑
x=1
q∑
i,j,k,l=1
gxij
¯
gxkl|
ikih
jl|
=
1dd
∑
x=1
Gx
⊗
G¯
xTheGx’s are independentq
×
qmatrices whose entries are independent complex Gaussians with mean 0 and variance 1/
q.Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 6
A simple model of random translation-invariant PEPS
MNparticles on a torus• • • • •
• • • • •
• • • • • N
M
Pick a tensor
|χi ∈
Cd⊗ (
Cq)
⊗4whose entries are independent complex Gaussians with mean 0 and variance 1/
dq2.Repeat it on all sites and contract neighboringq-dimensional indices.
−→
Obtained tensor|χ
MNi ∈ (
Cd)
⊗MN:random translation-invariant PEPS with periodic boundary conditions(typically almost normalized).q d q
q q
|χi = ∑
dx=1 q
∑
i,j,i0,j0=1
gxiji0j0
|
xiji0j0i
1-site tensor M
|χ
Mi ∈ (
Cd⊗
Cq⊗
Cq)
⊗M 1-column tensorAssociatedtransfer operator:TM
: (
Cq⊗
Cq)
⊗M→ (
Cq⊗
Cq)
⊗M, obtained by contracting the d-dimensional indices of|χ
Mi
and|¯ χ
Mi
.TM
=
dM1qM d∑
x1,...,xM=1 q
∑
i1,j1,...,iM,jM=1
Gx1iMi1
⊗
G¯
x1jMj1
⊗ · · · ⊗
GxMiM−1iM⊗
G¯
xMjM−1jM
TheGxij’s are independentq
×
qmatrices whose entries are independent complex Gaussians with mean 0 and variance 1/
q.Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 7
A simple model of random translation-invariant PEPS
MNparticles on a torus• • • • •
• • • • •
• • • • • N
M
Pick a tensor
|χi ∈
Cd⊗ (
Cq)
⊗4whose entries are independent complex Gaussians with mean 0 and variance 1/
dq2.Repeat it on all sites and contract neighboringq-dimensional indices.
−→
Obtained tensor|χ
MNi ∈ (
Cd)
⊗MN:random translation-invariant PEPS with periodic boundary conditions(typically almost normalized).q d q
q q
|χi = ∑
dx=1 q
∑
i,j,i0,j0=1
gxiji0j0
|
xiji0j0i
1-site tensor M
|χ
Mi ∈ (
Cd⊗
Cq⊗
Cq)
⊗M 1-column tensorAssociatedtransfer operator:TM
: (
Cq⊗
Cq)
⊗M→ (
Cq⊗
Cq)
⊗M, obtained by contracting the d-dimensional indices of|χ
Mi
and|¯ χ
Mi
.TM
=
dM1qM d∑
x1,...,xM=1 q
∑
i1,j1,...,iM,jM=1
Gx1iMi1
⊗
G¯
x1jMj1
⊗ · · · ⊗
GxMiM−1iM⊗
G¯
xMjM−1jM
TheGxij’s are independentq
×
qmatrices whose entries are independent complex Gaussians with mean 0 and variance 1/
q.Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 7
Correlations in a TNS
MPS:M=
1. PEPS:M>
1.A
,
Bobservables on(
Cd)
⊗M, i.e. on 1 site for an MPS and on 1 column ofMsites for a PEPS.Goal:Quantify the correlations between the outcomes of AandB, when performed on ‘distant’ sites or columns.
A A
:
Cd→
CdM
A
A
: (
Cd)
⊗M→ (
Cd)
⊗MCompute the value on the TNS
|χ
MNi
of the observableA⊗ I
⊗k⊗
B⊗ I
⊗N−k−2, i.e. vχ(
A,
B,
k) := hχ
MN|
A⊗ I
⊗k⊗
B⊗ I
⊗N−k−2|χ
MNi.
Compare it to the product of the values on
|χ
MNi
ofA⊗ I
⊗N−1andI
⊗k+1⊗
B⊗ I
⊗N−k−2, i.e. vχ(
A)
vχ(
B) := hχ
MN|
A⊗ I
⊗N−1|χ
MNihχ
MN|
B⊗ I
⊗N−1|χ
MNi.
by translation-invariance of|χMNi Correlationsin the TNS
|χ
MNi
:γ
χ(
A,
B,
k) :=
vχ(
A,
B,
k) −
vχ(
A)
vχ(
B)
. Question:Do we haveγ
χ(
A,
B,
k) −→
kN→∞0? And if so, at which speed?
A B
N
k
A B
'
?Nk1
×
Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 8
Correlations in a TNS
MPS:M=
1. PEPS:M>
1.A
,
Bobservables on(
Cd)
⊗M, i.e. on 1 site for an MPS and on 1 column ofMsites for a PEPS.Goal:Quantify the correlations between the outcomes of AandB, when performed on ‘distant’ sites or columns.
A A
:
Cd→
CdM
A
A
: (
Cd)
⊗M→ (
Cd)
⊗MCompute the value on the TNS
|χ
MNi
of the observableA⊗ I
⊗k⊗
B⊗ I
⊗N−k−2, i.e.vχ
(
A,
B,
k) := hχ
MN|
A⊗ I
⊗k⊗
B⊗ I
⊗N−k−2|χ
MNi.
Compare it to the product of the values on
|χ
MNi
ofA⊗ I
⊗N−1andI
⊗k+1⊗
B⊗ I
⊗N−k−2, i.e.vχ
(
A)
vχ(
B) := hχ
MN|
A⊗ I
⊗N−1|χ
MNihχ
MN|
B⊗ I
⊗N−1|χ
MNi.
by translation-invariance of|χMNi Correlationsin the TNS
|χ
MNi
:γ
χ(
A,
B,
k) :=
vχ(
A,
B,
k) −
vχ(
A)
vχ(
B)
. Question:Do we haveγ
χ(
A,
B,
k) −→
kN→∞0? And if so, at which speed?
A B
N
k
A B
'
?Nk1
×
Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 8
Main result: random TNS typically exhibit fast exponential decay of correlations
Intuition:For any TNS
|χ
MNi ∈ (
Cd)
⊗MN, the correlations between two 1-site or 1-column observables decay exponentially with the distance separating the two sites or columns, i.e. there existC(χ),τ(χ) >
0 such that, for anykNand any observablesA,
Bon(
Cd)
⊗M,γ
χ(
A,
B,
k) 6
C(χ)
e−τ(χ)kk
Ak
∞k
Bk
∞.
Correlation lengthin the TNS|χ
MNi
:ξ(χ) :=
1/τ(χ)
.Our main result (informal)
For random translation-invariant MPS and PEPS with periodic boundary conditions this intuition is generically true, with a short correlation length.
going to 0 asd,qgrow with probability going to 1 (exponentially) asd,qgrow
Two ways of proving this:(suited to different dimensional regimes)
Show that thetransfer operatorassociated to the random TNS generically has alarge (upper) spectral gap.
Show that theparent Hamiltonianof the random TNS generically has alarge (lower) spectral gap.
going to 1 asd,qgrow
going to 1 asd,qgrow
Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 9
Main result: random TNS typically exhibit fast exponential decay of correlations
Intuition:For any TNS
|χ
MNi ∈ (
Cd)
⊗MN, the correlations between two 1-site or 1-column observables decay exponentially with the distance separating the two sites or columns, i.e. there existC(χ),τ(χ) >
0 such that, for anykNand any observablesA,
Bon(
Cd)
⊗M,γ
χ(
A,
B,
k) 6
C(χ)
e−τ(χ)kk
Ak
∞k
Bk
∞.
Correlation lengthin the TNS|χ
MNi
:ξ(χ) :=
1/τ(χ)
.Our main result (informal)
For random translation-invariant MPS and PEPS with periodic boundary conditions this intuition is generically true, with a short correlation length.
going to 0 asd,qgrow with probability going to 1 (exponentially) asd,qgrow
Two ways of proving this:(suited to different dimensional regimes)
Show that thetransfer operatorassociated to the random TNS generically has alarge (upper) spectral gap.
Show that theparent Hamiltonianof the random TNS generically has alarge (lower) spectral gap.
going to 1 asd,qgrow
going to 1 asd,qgrow
Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 9
Main result: random TNS typically exhibit fast exponential decay of correlations
Intuition:For any TNS
|χ
MNi ∈ (
Cd)
⊗MN, the correlations between two 1-site or 1-column observables decay exponentially with the distance separating the two sites or columns, i.e. there existC(χ),τ(χ) >
0 such that, for anykNand any observablesA,
Bon(
Cd)
⊗M,γ
χ(
A,
B,
k) 6
C(χ)
e−τ(χ)kk
Ak
∞k
Bk
∞.
Correlation lengthin the TNS|χ
MNi
:ξ(χ) :=
1/τ(χ)
.Our main result (informal)
For random translation-invariant MPS and PEPS with periodic boundary conditions this intuition is generically true, with a short correlation length.
going to 0 asd,qgrow with probability going to 1 (exponentially) asd,qgrow
Two ways of proving this:(suited to different dimensional regimes)
Show that thetransfer operatorassociated to the random TNS generically has alarge (upper) spectral gap.
Show that theparent Hamiltonianof the random TNS generically has alarge (lower) spectral gap.
going to 1 asd,qgrow
going to 1 asd,qgrow
Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 9
Outline
1 Background and motivations
2 Random MPS and PEPS: construction and statement of the main questions & results
3 Typical correlation length in a random TNS through typical spectral gap of its transfer operator
4 Typical correlation length in a random TNS through spectral gap of its parent Hamiltonian
5 Open questions and perspectives
Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 10
Correlation length in a TNS and spectrum of its transfer operator
|χ
MNi ∈ (
Cd)
⊗MNan MPS (M=
1) or a PEPS (M>
1).Tits associated transfer operator on
(
Cq⊗
Cq)
⊗M. By reading diagrams representingvχ(
A,
B,
k)
,vχ(
A)
,vχ(
B)
‘horizontally’ instead of ‘vertically’, its correlation function can be re-written as:
γ
χ(
A,
B,
k) =
Tr ˜
ATkBT˜
N−k−2− Tr ˜
ATN−1Tr ˜
BTN−1.
A A
:
Cd→
CdA
A
˜ :
Cq⊗
Cq→
Cq⊗
CqImportant consequence:
Denote by
λ
1(
T)
andλ
2(
T)
the largest and second largest eigenvalues ofT (with multiplicities). There existsC(
T) >
0 such that, settingε(
T) = |λ
2(
T)|/|λ
1(
T)|
, we haveγ
χ(
A,
B,
k) 6
C(
T)ε(
T)
kk
Ak
∞k
Bk
∞.
−→
Correlations between two 1-site or 1-column observables decay exponentially with the distance separating the two sites or columns, at a rate|log ε(
T)|
, i.e.ξ(χ) =
1/|log ε(
T)|
. Question:For a random transfer operatorTwhat are the typical values of|λ
1(
T)|
and|λ
2(
T)|
?Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 11
Correlation length in a TNS and spectrum of its transfer operator
|χ
MNi ∈ (
Cd)
⊗MNan MPS (M=
1) or a PEPS (M>
1).Tits associated transfer operator on
(
Cq⊗
Cq)
⊗M. By reading diagrams representingvχ(
A,
B,
k)
,vχ(
A)
,vχ(
B)
‘horizontally’ instead of ‘vertically’, its correlation function can be re-written as:
γ
χ(
A,
B,
k) =
Tr ˜
ATkBT˜
N−k−2− Tr ˜
ATN−1Tr ˜
BTN−1.
A A
:
Cd→
CdA
A
˜ :
Cq⊗
Cq→
Cq⊗
CqImportant consequence:
Denote by
λ
1(
T)
andλ
2(
T)
the largest and second largest eigenvalues ofT (with multiplicities).There existsC
(
T) >
0 such that, settingε(
T) = |λ
2(
T)|/|λ
1(
T)|
, we haveγ
χ(
A,
B,
k) 6
C(
T)ε(
T)
kk
Ak
∞k
Bk
∞.
−→
Correlations between two 1-site or 1-column observables decay exponentially with the distance separating the two sites or columns, at a rate|log ε(
T)|
, i.e.ξ(χ) =
1/|log ε(
T)|
.Question:For a random transfer operatorTwhat are the typical values of
|λ
1(
T)|
and|λ
2(
T)|
?Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 11
Correlation length in a TNS and spectrum of its transfer operator
|χ
MNi ∈ (
Cd)
⊗MNan MPS (M=
1) or a PEPS (M>
1).Tits associated transfer operator on
(
Cq⊗
Cq)
⊗M. By reading diagrams representingvχ(
A,
B,
k)
,vχ(
A)
,vχ(
B)
‘horizontally’ instead of ‘vertically’, its correlation function can be re-written as:
γ
χ(
A,
B,
k) =
Tr ˜
ATkBT˜
N−k−2− Tr ˜
ATN−1Tr ˜
BTN−1.
A A
:
Cd→
CdA
A
˜ :
Cq⊗
Cq→
Cq⊗
CqImportant consequence:
Denote by
λ
1(
T)
andλ
2(
T)
the largest and second largest eigenvalues ofT (with multiplicities).There existsC
(
T) >
0 such that, settingε(
T) = |λ
2(
T)|/|λ
1(
T)|
, we haveγ
χ(
A,
B,
k) 6
C(
T)ε(
T)
kk
Ak
∞k
Bk
∞.
−→
Correlations between two 1-site or 1-column observables decay exponentially with the distance separating the two sites or columns, at a rate|log ε(
T)|
, i.e.ξ(χ) =
1/|log ε(
T)|
. Question:For a random transfer operatorTwhat are the typical values of|λ
1(
T)|
and|λ
2(
T)|
?Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 11
Typical spectral gap of the transfer operator of a random MPS
Theorem [Typical spectral gap of the random MPS transfer operator T ]
There exist constantsc,
C>
0 such that, for anyd,
q∈
N,P
|λ
1(
T)| >
1− √
Cd and
|λ
2(
T)| 6 √
C d>
1−
e−cq.
Corollary [Typical correlation length in the random MPS |χ
Ni ]
There exist constantsc,
C0>
0 such that, for anyd,
q,
N∈
N,P
ξ(χ) 6
C0
log
d>
1−
e−cq.
Remark:Similar results obtained on slightly different models, withT composed of independent Haar unitaries(Hastings, Pisier)or blocks of a Haar unitary(González-Guillén/Junge/Nechita). Motivation: Construction ofquantum expanders. This Gaussian model provides a new example. Proof idea:Approximate first eigenvector forT: maximally entangled state
|ψi ∈
Cq⊗
Cq. Observing that|λ
1(
T)| > |hψ|
T|ψi|
and|λ
2(
T)| 6 k
T(I − |ψihψ|)k
∞, we show that:1 E
|hψ|
T|ψi| =
1 andEk
T(I − |ψihψ|)k
∞6
C/ √
d(Gaussian moment computations).
2 These two quantities have a small probability of deviating from their average (local version of Gaussian concentration inequality).
by minimax principle for singular values, doing as ifTwere Hermitian
Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 12
Typical spectral gap of the transfer operator of a random MPS
Theorem [Typical spectral gap of the random MPS transfer operator T ]
There exist constantsc,
C>
0 such that, for anyd,
q∈
N,P
|λ
1(
T)| >
1− √
Cd and
|λ
2(
T)| 6 √
C d>
1−
e−cq.
Corollary [Typical correlation length in the random MPS |χ
Ni ]
There exist constantsc,
C0>
0 such that, for anyd,
q,
N∈
N,P
ξ(χ) 6
C0
log
d>
1−
e−cq.
Remark:Similar results obtained on slightly different models, withT composed of independent Haar unitaries(Hastings, Pisier)or blocks of a Haar unitary(González-Guillén/Junge/Nechita). Motivation: Construction ofquantum expanders. This Gaussian model provides a new example. Proof idea:Approximate first eigenvector forT: maximally entangled state
|ψi ∈
Cq⊗
Cq. Observing that|λ
1(
T)| > |hψ|
T|ψi|
and|λ
2(
T)| 6 k
T(I − |ψihψ|)k
∞, we show that:1 E
|hψ|
T|ψi| =
1 andEk
T(I − |ψihψ|)k
∞6
C/ √
d(Gaussian moment computations).
2 These two quantities have a small probability of deviating from their average (local version of Gaussian concentration inequality).
by minimax principle for singular values, doing as ifTwere Hermitian
Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 12
Typical spectral gap of the transfer operator of a random MPS
Theorem [Typical spectral gap of the random MPS transfer operator T ]
There exist constantsc,
C>
0 such that, for anyd,
q∈
N,P
|λ
1(
T)| >
1− √
Cd and
|λ
2(
T)| 6 √
C d>
1−
e−cq.
Corollary [Typical correlation length in the random MPS |χ
Ni ]
There exist constantsc,
C0>
0 such that, for anyd,
q,
N∈
N,P
ξ(χ) 6
C0
log
d>
1−
e−cq.
Remark:Similar results obtained on slightly different models, withT composed of independent Haar unitaries(Hastings, Pisier)or blocks of a Haar unitary(González-Guillén/Junge/Nechita). Motivation: Construction ofquantum expanders. This Gaussian model provides a new example.
Proof idea:Approximate first eigenvector forT: maximally entangled state
|ψi ∈
Cq⊗
Cq. Observing that|λ
1(
T)| > |hψ|
T|ψi|
and|λ
2(
T)| 6 k
T(I − |ψihψ|)k
∞, we show that:1 E
|hψ|
T|ψi| =
1 andEk
T(I − |ψihψ|)k
∞6
C/ √
d(Gaussian moment computations).
2 These two quantities have a small probability of deviating from their average (local version of Gaussian concentration inequality).
by minimax principle for singular values, doing as ifTwere Hermitian
Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 12
Typical spectral gap of the transfer operator of a random MPS
Theorem [Typical spectral gap of the random MPS transfer operator T ]
There exist constantsc,
C>
0 such that, for anyd,
q∈
N,P
|λ
1(
T)| >
1− √
Cd and
|λ
2(
T)| 6 √
C d>
1−
e−cq.
Corollary [Typical correlation length in the random MPS |χ
Ni ]
There exist constantsc,
C0>
0 such that, for anyd,
q,
N∈
N,P
ξ(χ) 6
C0
log
d>
1−
e−cq.
Remark:Similar results obtained on slightly different models, withT composed of independent Haar unitaries(Hastings, Pisier)or blocks of a Haar unitary(González-Guillén/Junge/Nechita). Motivation: Construction ofquantum expanders. This Gaussian model provides a new example.
Proof idea:Approximate first eigenvector forT: maximally entangled state
|ψi ∈
Cq⊗
Cq. Observing that|λ
1(
T)| > |hψ|
T|ψi|
and|λ
2(
T)| 6 k
T(I − |ψihψ|)k
∞, we show that:1 E
|hψ|
T|ψi| =
1 andEk
T(I − |ψihψ|)k
∞6
C/ √
d(Gaussian moment computations).
2 These two quantities have a small probability of deviating from their average (local version of Gaussian concentration inequality).
by minimax principle for singular values, doing as ifTwere Hermitian
Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 12
Typical spectral gap of the transfer operator of a random PEPS
Assumption
(?)
:The dimensionsdandqgrow polynomially with the number of sitesM.More precisely:d
,
qsatisfyd'
Mαandq'
Mβwithα >
8 and(α +
1)/
3< β < (α −
2)/
2.Theorem [Typical spectral gap of the random PEPS transfer operator T
M]
There exist constantsc,
C>
0 such that, for anyM∈
Nandd,
q∈
Nsatisfying(?)
,P
|λ
1(
TM)| >
1−
CMα/2−1−β and
|λ
2(
TM)| 6
C Mα/2−1−β>
1−
e−cM3β−α.
Corollary [Typical correlation length in the random PEPS |χ
MNi ]
There exist constantsc
,
C0>
0 such that, for anyN>
M∈
Nandd,
q∈
Nsatisfying(?)
, Pξ(χ) 6
C0
(α/
2−
1− β) log
M>
1−
e−cM3β−α.
Main issue:Results are valid only in the regime whered
,
qgrow polynomially withM.−→
Enforce this scaling byblocking: Redefine 1 site as being a sublattice oflog
Msites. Proof idea:Some kind of recursion procedure that uses the MPS results as building blocks. Approximate first eigenvector forTM:|ψi
⊗M∈ (
Cq⊗
Cq)
⊗M.Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 13
Typical spectral gap of the transfer operator of a random PEPS
Assumption
(?)
:The dimensionsdandqgrow polynomially with the number of sitesM.More precisely:d
,
qsatisfyd'
Mαandq'
Mβwithα >
8 and(α +
1)/
3< β < (α −
2)/
2.Theorem [Typical spectral gap of the random PEPS transfer operator T
M]
There exist constantsc,
C>
0 such that, for anyM∈
Nandd,
q∈
Nsatisfying(?)
,P
|λ
1(
TM)| >
1−
CMα/2−1−β and
|λ
2(
TM)| 6
C Mα/2−1−β>
1−
e−cM3β−α.
Corollary [Typical correlation length in the random PEPS |χ
MNi ]
There exist constantsc
,
C0>
0 such that, for anyN>
M∈
Nandd,
q∈
Nsatisfying(?)
, Pξ(χ) 6
C0
(α/
2−
1− β) log
M>
1−
e−cM3β−α.
Main issue:Results are valid only in the regime whered
,
qgrow polynomially withM.−→
Enforce this scaling byblocking: Redefine 1 site as being a sublattice oflog
Msites. Proof idea:Some kind of recursion procedure that uses the MPS results as building blocks. Approximate first eigenvector forTM:|ψi
⊗M∈ (
Cq⊗
Cq)
⊗M.Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 13
Typical spectral gap of the transfer operator of a random PEPS
Assumption
(?)
:The dimensionsdandqgrow polynomially with the number of sitesM.More precisely:d
,
qsatisfyd'
Mαandq'
Mβwithα >
8 and(α +
1)/
3< β < (α −
2)/
2.Theorem [Typical spectral gap of the random PEPS transfer operator T
M]
There exist constantsc,
C>
0 such that, for anyM∈
Nandd,
q∈
Nsatisfying(?)
,P
|λ
1(
TM)| >
1−
CMα/2−1−β and
|λ
2(
TM)| 6
C Mα/2−1−β>
1−
e−cM3β−α.
Corollary [Typical correlation length in the random PEPS |χ
MNi ]
There exist constantsc
,
C0>
0 such that, for anyN>
M∈
Nandd,
q∈
Nsatisfying(?)
, Pξ(χ) 6
C0
(α/
2−
1− β) log
M>
1−
e−cM3β−α.
Main issue:Results are valid only in the regime whered
,
qgrow polynomially withM.−→
Enforce this scaling byblocking: Redefine 1 site as being a sublattice oflog
Msites. Proof idea:Some kind of recursion procedure that uses the MPS results as building blocks. Approximate first eigenvector forTM:|ψi
⊗M∈ (
Cq⊗
Cq)
⊗M.Cécilia Lancien Typical correlations and entanglement in random tensor network states Journée théorie CPTGA – September 27 2021 13