Typical amount of entanglement in multipartite quantum systems
Quantifying multipartite entanglement through tensor norms
Cécilia Lancien
Institut de Mathématiques de Toulouse & CNRS
Sakura Meeting – July 1 2021
Outline
1 Tensor norms and entanglement
2 Quantifying the typical amount of entanglement/correlations in multipartite pure states
Tensor norms in Banach spaces
LetA1, . . . ,AMbe Banach spaces. Givenx∈A1⊗ · · · ⊗AM, itsprojective tensor normis kxkA1⊗π···⊗πAM:= inf
( r
∑
k=1
|αk|:aki ∈Ai,kakik61,x=
r
∑
k=1
αka1k⊗ · · · ⊗akM,r∈N )
,
and itsinjective tensor normis
kxkA1⊗ε···⊗εAM:= sup{hb1⊗ · · · ⊗bM|xi:bi∈A∗i,kbik61}.
These norms aredualto one another: for allx∈A1⊗ · · · ⊗AM, kxkA1⊗π···⊗πAM= sup
hy|xi:kykA∗
1⊗ε···⊗εA∗M61 , kxkA1⊗ε···⊗εAM= sup
hy|xi:kykA∗
1⊗π···⊗πA∗M61 .
The projective and injective norms are examples oftensor norms: for alla1∈A1, . . . ,aM∈AM, ka1⊗ · · · ⊗aMkA1⊗π···⊗πAM=ka1⊗ · · · ⊗aMkA1⊗ε···⊗εAM=ka1k · · · kaMk. And they areextremalamong such norms: for any other tensor normk · konA1⊗ · · · ⊗AM,
k · kA1⊗ε···⊗εAM6k · k6k · kA1⊗π···⊗πAM.
Tensor norms in Banach spaces
LetA1, . . . ,AMbe Banach spaces. Givenx∈A1⊗ · · · ⊗AM, itsprojective tensor normis kxkA1⊗π···⊗πAM:= inf
( r
∑
k=1
|αk|:aki ∈Ai,kakik61,x=
r
∑
k=1
αka1k⊗ · · · ⊗akM,r∈N )
,
and itsinjective tensor normis
kxkA1⊗ε···⊗εAM:= sup{hb1⊗ · · · ⊗bM|xi:bi∈A∗i,kbik61}.
These norms aredualto one another: for allx∈A1⊗ · · · ⊗AM, kxkA1⊗π···⊗πAM= sup
hy|xi:kykA∗
1⊗ε···⊗εA∗M61 , kxkA1⊗ε···⊗εAM= sup
hy|xi:kykA∗
1⊗π···⊗πA∗M61 .
The projective and injective norms are examples oftensor norms: for alla1∈A1, . . . ,aM∈AM, ka1⊗ · · · ⊗aMkA1⊗π···⊗πAM=ka1⊗ · · · ⊗aMkA1⊗ε···⊗εAM=ka1k · · · kaMk. And they areextremalamong such norms: for any other tensor normk · konA1⊗ · · · ⊗AM,
k · kA1⊗ε···⊗εAM6k · k6k · kA1⊗π···⊗πAM.
Tensor norms in Banach spaces
LetA1, . . . ,AMbe Banach spaces. Givenx∈A1⊗ · · · ⊗AM, itsprojective tensor normis kxkA1⊗π···⊗πAM:= inf
( r
∑
k=1
|αk|:aki ∈Ai,kakik61,x=
r
∑
k=1
αka1k⊗ · · · ⊗akM,r∈N )
,
and itsinjective tensor normis
kxkA1⊗ε···⊗εAM:= sup{hb1⊗ · · · ⊗bM|xi:bi∈A∗i,kbik61}.
These norms aredualto one another: for allx∈A1⊗ · · · ⊗AM, kxkA1⊗π···⊗πAM= sup
hy|xi:kykA∗
1⊗ε···⊗εA∗M61 , kxkA1⊗ε···⊗εAM= sup
hy|xi:kykA∗
1⊗π···⊗πA∗M61 .
The projective and injective norms are examples oftensor norms: for alla1∈A1, . . . ,aM∈AM, ka1⊗ · · · ⊗aMkA1⊗π···⊗πAM=ka1⊗ · · · ⊗aMkA1⊗ε···⊗εAM=ka1k · · · kaMk.
And they areextremalamong such norms: for any other tensor normk · konA1⊗ · · · ⊗AM, k · kA1⊗ε···⊗εAM6k · k6k · kA1⊗π···⊗πAM.
Reminder: separability vs entanglement in multipartite quantum systems
H1, . . . ,HMcomplex Hilbert spaces. In this talk: of finite, but usually large, dimension.
−→Hi≡Cdi withdi1, for 16i6M.
Definition [Separability and entanglement]
A stateρonH1⊗ · · · ⊗HMis calledseparableif it is a convex combination of product states, i.e.
positive semidefinite operator with trace 1 onH1⊗ · · · ⊗HM
ρ=
r
∑
k=1
λkρk1⊗ · · · ⊗ρkM,with (
λk>0,16k6r,∑rk=1λk=1 ρki state onHi,16k6r,16i6M . Otherwise it is calledentangled.
[Note: Ifρis a pure state, i.e.ρ=|ψihψ|for some unit vectorψ∈H1⊗ · · · ⊗HM, thenρis separable iffψ=ψ1⊗ · · · ⊗ψMfor some unit vectorsψi∈Hi, 16i6M.]
Fact:If a multipartite quantum system is in a separable state, there is no intrinsically quantum correlation between its subsystems. It thus does not provide any advantage over a classical system in information processing tasks.
−→Characterizing and quantifying the entanglement of multipartite quantum states is an important issue in practice.
Reminder: separability vs entanglement in multipartite quantum systems
H1, . . . ,HMcomplex Hilbert spaces. In this talk: of finite, but usually large, dimension.
−→Hi≡Cdi withdi1, for 16i6M.
Definition [Separability and entanglement]
A stateρonH1⊗ · · · ⊗HMis calledseparableif it is a convex combination of product states, i.e.
positive semidefinite operator with trace 1 onH1⊗ · · · ⊗HM
ρ=
r
∑
k=1
λkρk1⊗ · · · ⊗ρkM,with (
λk>0,16k6r,∑rk=1λk=1 ρki state onHi,16k6r,16i6M . Otherwise it is calledentangled.
[Note: Ifρis a pure state, i.e.ρ=|ψihψ|for some unit vectorψ∈H1⊗ · · · ⊗HM, thenρis separable iffψ=ψ1⊗ · · · ⊗ψMfor some unit vectorsψi∈Hi, 16i6M.]
Fact:If a multipartite quantum system is in a separable state, there is no intrinsically quantum correlation between its subsystems. It thus does not provide any advantage over a classical system in information processing tasks.
−→Characterizing and quantifying the entanglement of multipartite quantum states is an important issue in practice.
Characterizing entanglement through tensor norms
•Pure state entanglement:
Banach spaces
Cdi,k · k`di 2
, 16i6M.
Notation:∀x∈Cd,kxk`d
2:= ∑dk=1|xk|21/2
vector 2-norm A pure stateψ∈Cd1⊗ · · · ⊗CdMis s.t.kψk`d1···dM
2
=1.
And it is separable iffkψk`d1
2⊗ε···⊗ε`dM2 =kψk`d1
2⊗π···⊗π`dM2 =1, where by definition:
kψk`d1
2⊗ε···⊗ε`dM2 := supn
hϕ1⊗ · · · ⊗ϕM|ψi:ϕi∈Cdi,kϕik`di 2
=1o , kψk`d1
2⊗π···⊗π`dM2 := inf r
∑
k=1
|αk|:χki ∈Cdi,kχkik`di 2
=1,ψ= ∑r
k=1
αkχk1⊗ · · · ⊗χkM
.
−→Ifkψkε1 orkψkπ1, thenψis ‘very’ entangled.
•Mixed state entanglement: Banach spaces
M
di(C),k · kSdi1
, 16i6M.
Notation:∀X∈
M
d(C),kXkSd1:= Tr|X|
matrix 1-norm A mixed stateρ∈
M
d1(C)⊗ · · · ⊗M
dM(C)is s.t.ρ>0 andkρkS1d1···dM
=1. And it is separable iffkρkSd1
1⊗π···⊗πSdM1 =1(Rudolph, Pérez-García), where by definition: kρkS1d1⊗π···⊗πSdM1
:= inf r
∑
k=1
|αk|:τki ∈
M
di(C),kτkikS1di
=1,ρ= ∑r
k=1
αkτk1⊗ · · · ⊗τkM
.
−→Ifkρkπ1, thenρis ‘very’ entangled.
Characterizing entanglement through tensor norms
•Pure state entanglement:
Banach spaces
Cdi,k · k`di 2
, 16i6M.
Notation:∀x∈Cd,kxk`d
2:= ∑dk=1|xk|21/2
vector 2-norm A pure stateψ∈Cd1⊗ · · · ⊗CdMis s.t.kψk`d1···dM
2
=1.
And it is separable iffkψk`d1
2⊗ε···⊗ε`dM2 =kψk`d1
2⊗π···⊗π`dM2 =1, where by definition:
kψk`d1
2⊗ε···⊗ε`dM2 := supn
hϕ1⊗ · · · ⊗ϕM|ψi:ϕi∈Cdi,kϕik`di 2
=1o , kψk`d1
2⊗π···⊗π`dM2 := inf r
∑
k=1
|αk|:χki ∈Cdi,kχkik`di 2
=1,ψ= ∑r
k=1
αkχk1⊗ · · · ⊗χkM
.
−→Ifkψkε1 orkψkπ1, thenψis ‘very’ entangled.
•Mixed state entanglement:
Banach spaces
M
di(C),k · kSdi1
, 16i6M.
Notation:∀X∈
M
d(C),kXkSd1:= Tr|X|
matrix 1-norm A mixed stateρ∈
M
d1(C)⊗ · · · ⊗M
dM(C)is s.t.ρ>0 andkρkSd1···dM1
=1.
And it is separable iffkρkSd1
1⊗π···⊗πSdM1 =1(Rudolph, Pérez-García), where by definition:
kρkS1d1⊗π···⊗πSdM1
:= inf r
∑
k=1
|αk|:τki ∈
M
di(C),kτkikS1di
=1,ρ= ∑r
k=1
αkτk1⊗ · · · ⊗τkM
.
−→Ifkρkπ1, thenρis ‘very’ entangled.
Particular case of bipartite pure states
Let us look at the previous definitions in the case of pure states whenM=2 andd1=d2=:d.
We can identify|ψi=
d
∑
k,l=1
ψkl|kli ∈Cd⊗Cd withMψ=
d
∑
k,l=1
ψkl|kihl| ∈
M
d(C). Then clearly,kψk2=kMψk2.And theSchmidt decompositionofψcorresponds to thesingular value decompositionofMψ:
|ψi=
r
∑
k=1
p
λk|ekfki ←→Mψ=
r
∑
k=1
p
λk|ekihfk|,
withr6dtheSchmidt rankofψ,∑rk=1λk=1,{ek}rk=1,{fk}rk=1orthonormal sets inCd. Sokψkε= max16k6r√
λk=kMψk∞andkψkπ=∑rk=1
√
λk=kMψk1.
−→Checking bipartite pure state separability is easy. Quantitatively, for allψ∈Cd⊗Cds.t.kψk2=1,√1
d 6kψkε61 and 16kψkπ6√ d. But no such simple characterization forM>2 (no equivalent of the Schmidt decomposition).
Particular case of bipartite pure states
Let us look at the previous definitions in the case of pure states whenM=2 andd1=d2=:d.
We can identify|ψi=
d
∑
k,l=1
ψkl|kli ∈Cd⊗Cd withMψ=
d
∑
k,l=1
ψkl|kihl| ∈
M
d(C). Then clearly,kψk2=kMψk2.And theSchmidt decompositionofψcorresponds to thesingular value decompositionofMψ:
|ψi=
r
∑
k=1
p
λk|ekfki ←→Mψ=
r
∑
k=1
p
λk|ekihfk|,
withr6dtheSchmidt rankofψ,∑rk=1λk=1,{ek}rk=1,{fk}rk=1orthonormal sets inCd. Sokψkε= max16k6r√
λk=kMψk∞andkψkπ=∑rk=1
√
λk=kMψk1.
−→Checking bipartite pure state separability is easy. Quantitatively, for allψ∈Cd⊗Cds.t.kψk2=1,√1
d 6kψkε61 and 16kψkπ6√ d. But no such simple characterization forM>2 (no equivalent of the Schmidt decomposition).
Particular case of bipartite pure states
Let us look at the previous definitions in the case of pure states whenM=2 andd1=d2=:d.
We can identify|ψi=
d
∑
k,l=1
ψkl|kli ∈Cd⊗Cd withMψ=
d
∑
k,l=1
ψkl|kihl| ∈
M
d(C). Then clearly,kψk2=kMψk2.And theSchmidt decompositionofψcorresponds to thesingular value decompositionofMψ:
|ψi=
r
∑
k=1
p
λk|ekfki ←→Mψ=
r
∑
k=1
p
λk|ekihfk|,
withr6dtheSchmidt rankofψ,∑rk=1λk=1,{ek}rk=1,{fk}rk=1orthonormal sets inCd. Sokψkε= max16k6r√
λk=kMψk∞andkψkπ=∑rk=1
√
λk=kMψk1.
−→Checking bipartite pure state separability is easy.
Quantitatively, for allψ∈Cd⊗Cds.t.kψk2=1,√1
d 6kψkε61 and 16kψkπ6√ d.
But no such simple characterization forM>2 (no equivalent of the Schmidt decomposition).
Geometric measure of entanglement
Definition [Geometric measure of entanglement(Shimony, Wei/Goldbart, Zhu/Chen/Hayashi)] Letψ∈H1⊗ · · · ⊗HMbe a pure state. Itsgeometric measure of entanglement (GME)is
E(ψ) :=−log sup
|hϕ1⊗ · · · ⊗ϕM|ψi|2:ϕi∈Hi,kϕik2=1 . By definition,E(ψ) =−2logkψkε. SoE(ψ) =0 iffψis separable.
−→Eis afaithful entanglement measurefor multipartite pure states.
Remark:The definition of the GME can be extended to mixed states onH1⊗ · · · ⊗HM(but it is not an entanglement measure anymore), as
E(ρ) :=−log sup{hϕ1⊗ · · · ⊗ϕM|ρ|ϕ1⊗ · · · ⊗ϕMi:ϕi∈Hi,kϕik2=1}
=−log sup{Tr(ρσ) :σseparable state onH1⊗ · · · ⊗HM}. Fact:For any unit vectorψ∈(Cd)⊗M,kψkε> √d1M−1, i.e.E(ψ)6(M−1) logd. This can be checked recursively, starting from the bipartite case:
For any unit vectorψ∈Cd1⊗Cd2,kψkε> √1d, whered:= min(d1,d2).
maximal possible Schmidt rank ofψ Question:Are multipartite pure states generically ‘very’ or ‘little’ entangled?
−→What is the typical value of the GME for a unit vectorψ∈(Cd)⊗Msampled at random?
Geometric measure of entanglement
Definition [Geometric measure of entanglement(Shimony, Wei/Goldbart, Zhu/Chen/Hayashi)] Letψ∈H1⊗ · · · ⊗HMbe a pure state. Itsgeometric measure of entanglement (GME)is
E(ψ) :=−log sup
|hϕ1⊗ · · · ⊗ϕM|ψi|2:ϕi∈Hi,kϕik2=1 . By definition,E(ψ) =−2logkψkε. SoE(ψ) =0 iffψis separable.
−→Eis afaithful entanglement measurefor multipartite pure states.
Remark:The definition of the GME can be extended to mixed states onH1⊗ · · · ⊗HM(but it is not an entanglement measure anymore), as
E(ρ) :=−log sup{hϕ1⊗ · · · ⊗ϕM|ρ|ϕ1⊗ · · · ⊗ϕMi:ϕi∈Hi,kϕik2=1}
=−log sup{Tr(ρσ) :σseparable state onH1⊗ · · · ⊗HM}.
Fact:For any unit vectorψ∈(Cd)⊗M,kψkε> √d1M−1, i.e.E(ψ)6(M−1) logd. This can be checked recursively, starting from the bipartite case:
For any unit vectorψ∈Cd1⊗Cd2,kψkε> √1d, whered:= min(d1,d2).
maximal possible Schmidt rank ofψ Question:Are multipartite pure states generically ‘very’ or ‘little’ entangled?
−→What is the typical value of the GME for a unit vectorψ∈(Cd)⊗Msampled at random?
Geometric measure of entanglement
Definition [Geometric measure of entanglement(Shimony, Wei/Goldbart, Zhu/Chen/Hayashi)] Letψ∈H1⊗ · · · ⊗HMbe a pure state. Itsgeometric measure of entanglement (GME)is
E(ψ) :=−log sup
|hϕ1⊗ · · · ⊗ϕM|ψi|2:ϕi∈Hi,kϕik2=1 . By definition,E(ψ) =−2logkψkε. SoE(ψ) =0 iffψis separable.
−→Eis afaithful entanglement measurefor multipartite pure states.
Remark:The definition of the GME can be extended to mixed states onH1⊗ · · · ⊗HM(but it is not an entanglement measure anymore), as
E(ρ) :=−log sup{hϕ1⊗ · · · ⊗ϕM|ρ|ϕ1⊗ · · · ⊗ϕMi:ϕi∈Hi,kϕik2=1}
=−log sup{Tr(ρσ) :σseparable state onH1⊗ · · · ⊗HM}. Fact:For any unit vectorψ∈(Cd)⊗M,kψkε> √d1M−1, i.e.E(ψ)6(M−1) logd.
This can be checked recursively, starting from the bipartite case:
For any unit vectorψ∈Cd1⊗Cd2,kψkε> √1d, whered:= min(d1,d2).
maximal possible Schmidt rank ofψ
Question:Are multipartite pure states generically ‘very’ or ‘little’ entangled?
−→What is the typical value of the GME for a unit vectorψ∈(Cd)⊗Msampled at random?
Geometric measure of entanglement
Definition [Geometric measure of entanglement(Shimony, Wei/Goldbart, Zhu/Chen/Hayashi)] Letψ∈H1⊗ · · · ⊗HMbe a pure state. Itsgeometric measure of entanglement (GME)is
E(ψ) :=−log sup
|hϕ1⊗ · · · ⊗ϕM|ψi|2:ϕi∈Hi,kϕik2=1 . By definition,E(ψ) =−2logkψkε. SoE(ψ) =0 iffψis separable.
−→Eis afaithful entanglement measurefor multipartite pure states.
Remark:The definition of the GME can be extended to mixed states onH1⊗ · · · ⊗HM(but it is not an entanglement measure anymore), as
E(ρ) :=−log sup{hϕ1⊗ · · · ⊗ϕM|ρ|ϕ1⊗ · · · ⊗ϕMi:ϕi∈Hi,kϕik2=1}
=−log sup{Tr(ρσ) :σseparable state onH1⊗ · · · ⊗HM}. Fact:For any unit vectorψ∈(Cd)⊗M,kψkε> √d1M−1, i.e.E(ψ)6(M−1) logd.
This can be checked recursively, starting from the bipartite case:
For any unit vectorψ∈Cd1⊗Cd2,kψkε> √1d, whered:= min(d1,d2).
maximal possible Schmidt rank ofψ Question:Are multipartite pure states generically ‘very’ or ‘little’ entangled?
−→What is the typical value of the GME for a unit vectorψ∈(Cd)⊗Msampled at random?
Outline
1 Tensor norms and entanglement
2 Quantifying the typical amount of entanglement/correlations in multipartite pure states
GME of uniformly distributed multipartite pure states
Theorem [Typicalεnorm of a random unit vector(Aubrun/Szarek)]
There exist constantsc,C,c0>0 s.t., forψ∈(Cd)⊗Ma uniformly distributed unit vector,
P c
rMlogM
dM−1 6kψkε6C
rMlogM dM−1
!
>1−e−c0dMlogM.
Consequence:Forψ∈(Cd)⊗Ma uniformly distributed unit vector, whendorMis large, E(ψ) = (M−1) logd−log(MlogM) +O(1)with high probability.
−→A random multipartite pure state is typically close to maximally entangled.
Proof idea:Observe thatψ∼g/kgk2, whereg∈(Cd)⊗Mhas independent complex Gaussian entries with mean 0 and variance 1.
•By the standard Gaussian concentration inequality:P
kgk2≷√
dM(1±ε)
6e−dMε2.
•Set
V
:=ϕ1⊗ · · · ⊗ϕM :ϕi∈Cd,kϕik2=1 , so thatEkgkε=Esupϕ∈V|hϕ|gi|. To estimate the latter quantity, use results about suprema of Gaussian processes. Upper bound: ‘small’ covering subset ofV
. Lower bound: ‘large’ separated subset ofV
.Conclusion:Ekgkεis of order√
dMlogM.
Then show thatg7→ kgkεalso concentrates around its average.
Question:Are ‘interesting’ multipartite pure states really captured by the uniform distribution?
GME of uniformly distributed multipartite pure states
Theorem [Typicalεnorm of a random unit vector(Aubrun/Szarek)]
There exist constantsc,C,c0>0 s.t., forψ∈(Cd)⊗Ma uniformly distributed unit vector,
P c
rMlogM
dM−1 6kψkε6C
rMlogM dM−1
!
>1−e−c0dMlogM.
Consequence:Forψ∈(Cd)⊗Ma uniformly distributed unit vector, whendorMis large, E(ψ) = (M−1) logd−log(MlogM) +O(1)with high probability.
−→A random multipartite pure state is typically close to maximally entangled.
Proof idea:Observe thatψ∼g/kgk2, whereg∈(Cd)⊗Mhas independent complex Gaussian entries with mean 0 and variance 1.
•By the standard Gaussian concentration inequality:P
kgk2≷√
dM(1±ε)
6e−dMε2.
•Set
V
:=ϕ1⊗ · · · ⊗ϕM :ϕi∈Cd,kϕik2=1 , so thatEkgkε=Esupϕ∈V|hϕ|gi|. To estimate the latter quantity, use results about suprema of Gaussian processes.Upper bound: ‘small’ covering subset of
V
. Lower bound: ‘large’ separated subset ofV
.Conclusion:Ekgkεis of order√
dMlogM.
Then show thatg7→ kgkεalso concentrates around its average.
Question:Are ‘interesting’ multipartite pure states really captured by the uniform distribution?
GME of uniformly distributed multipartite pure states
Theorem [Typicalεnorm of a random unit vector(Aubrun/Szarek)]
There exist constantsc,C,c0>0 s.t., forψ∈(Cd)⊗Ma uniformly distributed unit vector,
P c
rMlogM
dM−1 6kψkε6C
rMlogM dM−1
!
>1−e−c0dMlogM.
Consequence:Forψ∈(Cd)⊗Ma uniformly distributed unit vector, whendorMis large, E(ψ) = (M−1) logd−log(MlogM) +O(1)with high probability.
−→A random multipartite pure state is typically close to maximally entangled.
Proof idea:Observe thatψ∼g/kgk2, whereg∈(Cd)⊗Mhas independent complex Gaussian entries with mean 0 and variance 1.
•By the standard Gaussian concentration inequality:P
kgk2≷√
dM(1±ε)
6e−dMε2.
•Set
V
:=ϕ1⊗ · · · ⊗ϕM :ϕi∈Cd,kϕik2=1 , so thatEkgkε=Esupϕ∈V|hϕ|gi|. To estimate the latter quantity, use results about suprema of Gaussian processes.Upper bound: ‘small’ covering subset of
V
. Lower bound: ‘large’ separated subset ofV
.Conclusion:Ekgkεis of order√
dMlogM.
Then show thatg7→ kgkεalso concentrates around its average.
‘Physical’ states of many-body quantum systems and tensor network states
Curse of dimensionalityin many-body quantum systems: A system composed ofM d-dimensional subsystems has dimensiondM, which is exponential inM.
However, ‘physically relevant’ states of many-body quantum systems, such asground states of gapped local Hamiltonians, are (conjectured to be) well approximated by so-calledtensor network states (TNS), which form a small subset of the global state space(Hastings, Landau/Vazirani/Vidick).
Tensor network state on(Cd)⊗M:Take a graphGwithMvertices andLedges.
Put at each vertexva tensorχv∈Cd⊗(Cq)⊗δ(v)to get a tensorˆχG∈(Cd)⊗M⊗(Cq)⊗2L. Contract together the indices ofˆχGassociated to a same edge to get a tensorχG∈(Cd)⊗M.
−→Ifδ(v)6δfor allv, thenχGis described by at mostMqδdparameters, which is linear inM.
• •
• •
• • Gwith 6 vertices and 7 edges
• •
• •
• •
ˆχG∈(Cd)⊗6⊗(Cq)⊗14
• •
• •
• • χG∈(Cd)⊗6 d-dimensional indices:physicalindices.q-dimensional indices:bondindices.
If the underlying graphGis 1-dimensional (line or circle),χGis amatrix product state (MPS).
‘Physical’ states of many-body quantum systems and tensor network states
Curse of dimensionalityin many-body quantum systems: A system composed ofM d-dimensional subsystems has dimensiondM, which is exponential inM.
However, ‘physically relevant’ states of many-body quantum systems, such asground states of gapped local Hamiltonians, are (conjectured to be) well approximated by so-calledtensor network states (TNS), which form a small subset of the global state space(Hastings, Landau/Vazirani/Vidick). Tensor network state on(Cd)⊗M:Take a graphGwithMvertices andLedges.
Put at each vertexva tensorχv∈Cd⊗(Cq)⊗δ(v)to get a tensorˆχG∈(Cd)⊗M⊗(Cq)⊗2L. Contract together the indices ofˆχGassociated to a same edge to get a tensorχG∈(Cd)⊗M.
−→Ifδ(v)6δfor allv, thenχGis described by at mostMqδdparameters, which is linear inM.
• •
• •
• • Gwith 6 vertices and 7 edges
• •
• •
• •
ˆχG∈(Cd)⊗6⊗(Cq)⊗14
• •
• •
• • χG∈(Cd)⊗6 d-dimensional indices:physicalindices.q-dimensional indices:bondindices.
If the underlying graphGis 1-dimensional (line or circle),χGis amatrix product state (MPS).
‘Physical’ states of many-body quantum systems and tensor network states
Curse of dimensionalityin many-body quantum systems: A system composed ofM d-dimensional subsystems has dimensiondM, which is exponential inM.
However, ‘physically relevant’ states of many-body quantum systems, such asground states of gapped local Hamiltonians, are (conjectured to be) well approximated by so-calledtensor network states (TNS), which form a small subset of the global state space(Hastings, Landau/Vazirani/Vidick). Tensor network state on(Cd)⊗M:Take a graphGwithMvertices andLedges.
Put at each vertexva tensorχv∈Cd⊗(Cq)⊗δ(v)to get a tensorˆχG∈(Cd)⊗M⊗(Cq)⊗2L. Contract together the indices ofˆχGassociated to a same edge to get a tensorχG∈(Cd)⊗M.
−→Ifδ(v)6δfor allv, thenχGis described by at mostMqδdparameters, which is linear inM.
• •
• •
• • Gwith 6 vertices and 7 edges
• •
• •
• •
ˆχG∈(Cd)⊗6⊗(Cq)⊗14
• •
• •
• • χG∈(Cd)⊗6 d-dimensional indices:physicalindices.q-dimensional indices:bondindices.
If the underlying graphGis 1-dimensional (line or circle),χGis amatrix product state (MPS).
A simple model of random translation-invariant MPS
Mparticles on a circle
• • • • • M
Pick a tensorχ∈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 tensorχM∈(Cd)⊗M:random translation-invariant MPS with periodic boundary conditions(typically almost normalized).
d q q
|χi= ∑d
i=1 q
∑
a,a0=1
giaa0|iaa0i
M
|χMi= ∑d
i1,...,iM=1
q
∑
a1,...,aM=1
gi1aMa1· · ·giMaM−1aM
|i1· · ·iMi
Associatedtransfer operator: T:Cq⊗Cq→Cq⊗Cq, obtained by contracting the
d-dimensional indices ofχandχ¯. T= ∑d
i=1
q
∑
a,a0,b,b0=1
giaa0¯gibb0|abiha0b0|
=: ∑d
i=1
Gi⊗G¯i
Remark:The parameterqquantifies the amount of bipartite entanglement: Across any bipartite cut preserving the ordering of subsystems,χMhas Schmidt rank at mostq2dM/2.
area vs volume law Now what about genuinely multipartite entanglement?
−→Ifq=1,χM=χ⊗Mis separable. But what can we say forq1?
A simple model of random translation-invariant MPS
Mparticles on a circle
• • • • • M
Pick a tensorχ∈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 tensorχM∈(Cd)⊗M:random translation-invariant MPS with periodic boundary conditions(typically almost normalized).
d q q
|χi= ∑d
i=1 q
∑
a,a0=1
giaa0|iaa0i
M
|χMi= ∑d
i1,...,iM=1
q
∑
a1,...,aM=1
gi1aMa1· · ·giMaM−1aM
|i1· · ·iMi
Associatedtransfer operator:
T:Cq⊗Cq→Cq⊗Cq, obtained by contracting the
d-dimensional indices ofχandχ¯. T= ∑d
i=1
q
∑
a,a0,b,b0=1
giaa0¯gibb0|abiha0b0|
=: ∑d
i=1
Gi⊗G¯i
Remark:The parameterqquantifies the amount of bipartite entanglement: Across any bipartite cut preserving the ordering of subsystems,χMhas Schmidt rank at mostq2dM/2.
area vs volume law Now what about genuinely multipartite entanglement?
−→Ifq=1,χM=χ⊗Mis separable. But what can we say forq1?
A simple model of random translation-invariant MPS
Mparticles on a circle
• • • • • M
Pick a tensorχ∈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 tensorχM∈(Cd)⊗M:random translation-invariant MPS with periodic boundary conditions(typically almost normalized).
d q q
|χi= ∑d
i=1 q
∑
a,a0=1
giaa0|iaa0i
M
|χMi= ∑d
i1,...,iM=1
q
∑
a1,...,aM=1
gi1aMa1· · ·giMaM−1aM
|i1· · ·iMi
Associatedtransfer operator:
T:Cq⊗Cq→Cq⊗Cq, obtained by contracting the
d-dimensional indices ofχandχ¯. T= ∑d
i=1
q
∑
a,a0,b,b0=1
giaa0¯gibb0|abiha0b0|
=: ∑d
i=1
Gi⊗G¯i
Remark:The parameterqquantifies the amount of bipartite entanglement: Across any bipartite cut preserving the ordering of subsystems,χMhas Schmidt rank at mostq2dM/2.
area vs volume law Now what about genuinely multipartite entanglement?
−→Ifq=1, = ⊗Mis separable. But what can we say forq1?
Correlations in an MPS
LetA,Bbe 1-site observables, i.e. observables onCd. Goal:Quantify the correlations between the outcomes of AandB, when performed on ‘distant’ sites.
A A:Cd→Cd
B B:Cd→Cd
Compute the value on the MPSχMof the observableA1⊗Ik⊗B1⊗IM−k−2, i.e. vχ(A,B,k) :=hχM|A1⊗Ik⊗B1⊗IM−k−2|χMi.
Compare it to the product of the values onχMofA1⊗IM−1andIk+1⊗B1⊗IM−k−2, i.e. vχ(A)vχ(B) :=hχM|A1⊗IM−1|χMihχM|Ik+1⊗B1⊗IM−k−2|χMi. Correlationsin the MPSχM:γχ(A,B,k) :=
vχ(A,B,k)−vχ(A)vχ(B) . Question:Do we haveγχ(A,B,k) −→
kM→∞0? And if so, at which speed?
A B
M
k
A B
'?
Mk1 ×
Correlations in an MPS
LetA,Bbe 1-site observables, i.e. observables onCd. Goal:Quantify the correlations between the outcomes of AandB, when performed on ‘distant’ sites.
A A:Cd→Cd
B
B:Cd→Cd Compute the value on the MPSχMof the observableA1⊗Ik⊗B1⊗IM−k−2, i.e.
vχ(A,B,k) :=hχM|A1⊗Ik⊗B1⊗IM−k−2|χMi.
Compare it to the product of the values onχMofA1⊗IM−1andIk+1⊗B1⊗IM−k−2, i.e.
vχ(A)vχ(B) :=hχM|A1⊗IM−1|χMihχM|Ik+1⊗B1⊗IM−k−2|χMi.
Correlationsin the MPSχM:γχ(A,B,k) :=
vχ(A,B,k)−vχ(A)vχ(B) . Question:Do we haveγχ(A,B,k) −→
kM→∞0? And if so, at which speed?
A B
M
k
A B
'?
Mk1 ×
Exponential decay of correlations in random translation-invariant MPS
Clearly, separability implies no correlation between 1-site observables:
IfχM=χ⊗M, thenγχ(A,B,k) =0 for anyk6Mand any observablesA,BonCd.
Intuition:In an MPSχM, the correlations between 1-site observables decay exponentially with the distance separating the sites, i.e. there existC(χ),τ(χ)>0 s.t., for anykMand any observablesA,BonCd,
γχ(A,B,k)6C(χ)e−τ(χ)kkAk∞kBk∞. Correlation lengthin the MPSχM:ξ(χ) :=1/τ(χ).
Theorem [Typical correlation length of a random MPS(Lancien/Pérez-García)] There exist constantsC,c0>0 s.t., forχM∈(Cd)⊗Ma random translation-invariant MPS,
P
ξ(χ)6 C logd
>1−e−c0q.
Proof idea:Letλ1(T),λ2(T)be the two largest eigenvalues of the transfer operatorT and set ε(T) :=|λ2(T)|/|λ1(T)|. Then,γχ(A,B,k)6C(T)ε(T)kkAk∞kBk∞. Soξ(χ) =1/|logε(T)|. We can then prove thatP
|λ1(T)|>1−√C
d and|λ2(T)|6 √Cd
>1−e−c0q. spectral analysis for a non-normal random matrix with tensor product structure