Characterizing expansion, classically and quantumly
Perspective on “Quasirandom quantum channels”
T. Bannink, J. Briët, F. Labib, H. Maassen
Cécilia Lancien
Institut de Mathématiques de Toulouse & CNRS
QIT Journal Club – 18 Novembre 2020
Motivations
A graph isexpandingif it is simultaneouslysparse(i.e. ‘economical’) andhighly connected (i.e. modeling a network which is robust against perturbations, converges fast to equilibrium etc).
−→Important structural feature in many contexts.
Question:Given a regular graph, even if it is sparse, how ‘close’ is it to the complete graph from an ‘operational’ point of view?
Obvious properties of the complete graph:
1 fast mixing of a random walk supported on it,
2 uniformity of edge density between subsets of vertices.
−→Is measuring closeness to the complete graph in the sense of property (1) or (2) equivalent? Known:
Random regular graphs exhibit both (1) and (2) with high probability.
For any graph, (1) implies (2)(Hoory/Linial/Widgerson). The opposite implication holds only for dense graphs(Chung/Graham/Wilson)or vertex-transitive sparse graphs(Conlon/Zhao).
−→For those, the two notions of ‘quasi-randomness’ defined by (1) and (2) are equivalent. What about a quantum analogue of the problem?
1 Walk on a regular graph→Evolution under iterative application of a unital quantum channel.
2 Graph with uniform edge density→Quantum channel with uniform output directions.
−→Are these two ‘random looking’ properties of unital quantum channels in fact equivalent?
Motivations
A graph isexpandingif it is simultaneouslysparse(i.e. ‘economical’) andhighly connected (i.e. modeling a network which is robust against perturbations, converges fast to equilibrium etc).
−→Important structural feature in many contexts.
Question:Given a regular graph, even if it is sparse, how ‘close’ is it to the complete graph from an ‘operational’ point of view?
Obvious properties of the complete graph:
1 fast mixing of a random walk supported on it,
2 uniformity of edge density between subsets of vertices.
−→Is measuring closeness to the complete graph in the sense of property (1) or (2) equivalent?
Known:
Random regular graphs exhibit both (1) and (2) with high probability.
For any graph, (1) implies (2)(Hoory/Linial/Widgerson). The opposite implication holds only for dense graphs(Chung/Graham/Wilson)or vertex-transitive sparse graphs(Conlon/Zhao).
−→For those, the two notions of ‘quasi-randomness’ defined by (1) and (2) are equivalent. What about a quantum analogue of the problem?
1 Walk on a regular graph→Evolution under iterative application of a unital quantum channel.
2 Graph with uniform edge density→Quantum channel with uniform output directions.
−→Are these two ‘random looking’ properties of unital quantum channels in fact equivalent?
Motivations
A graph isexpandingif it is simultaneouslysparse(i.e. ‘economical’) andhighly connected (i.e. modeling a network which is robust against perturbations, converges fast to equilibrium etc).
−→Important structural feature in many contexts.
Question:Given a regular graph, even if it is sparse, how ‘close’ is it to the complete graph from an ‘operational’ point of view?
Obvious properties of the complete graph:
1 fast mixing of a random walk supported on it,
2 uniformity of edge density between subsets of vertices.
−→Is measuring closeness to the complete graph in the sense of property (1) or (2) equivalent?
Known:
Random regular graphs exhibit both (1) and (2) with high probability.
For any graph, (1) implies (2)(Hoory/Linial/Widgerson). The opposite implication holds only for dense graphs(Chung/Graham/Wilson)or vertex-transitive sparse graphs(Conlon/Zhao).
−→For those, the two notions of ‘quasi-randomness’ defined by (1) and (2) are equivalent.
What about a quantum analogue of the problem?
1 Walk on a regular graph→Evolution under iterative application of a unital quantum channel.
2 Graph with uniform edge density→Quantum channel with uniform output directions.
−→Are these two ‘random looking’ properties of unital quantum channels in fact equivalent?
Motivations
A graph isexpandingif it is simultaneouslysparse(i.e. ‘economical’) andhighly connected (i.e. modeling a network which is robust against perturbations, converges fast to equilibrium etc).
−→Important structural feature in many contexts.
Question:Given a regular graph, even if it is sparse, how ‘close’ is it to the complete graph from an ‘operational’ point of view?
Obvious properties of the complete graph:
1 fast mixing of a random walk supported on it,
2 uniformity of edge density between subsets of vertices.
−→Is measuring closeness to the complete graph in the sense of property (1) or (2) equivalent?
Known:
Random regular graphs exhibit both (1) and (2) with high probability.
For any graph, (1) implies (2)(Hoory/Linial/Widgerson). The opposite implication holds only for dense graphs(Chung/Graham/Wilson)or vertex-transitive sparse graphs(Conlon/Zhao).
−→For those, the two notions of ‘quasi-randomness’ defined by (1) and (2) are equivalent.
What about a quantum analogue of the problem?
1 Walk on a regular graph→Evolution under iterative application of a unital quantum channel.
2 Graph with uniform edge density→Quantum channel with uniform output directions.
−→Are these two ‘random looking’ properties of unital quantum channels in fact equivalent?
Spectral expansion and uniformity for regular graphs
G= (V,E)ad-regular graph, with vertex setVand edge setE(dedges at each vertex).
Aits normalized adjacency matrix, i.e. the|V| × |V|matrix s.t.Auv=e(u,v)
| {z }
number of edges between verticesuandv
/dfor eachu,v∈V.
•Property (1) is related tospectral expansion:
λ1(A),λ2(A), . . .eigenvalues ofA, with|λ1(A)|>|λ2(A)|>· · ·. SinceGis regular,λ1(A) =1 (all-one vector as associated eigenvector), and the spectral expansion parameter ofGis
λ(G) :=|λ2(A)|; the smallestλ(G),the most spectrally expandingG.
•Property (2) is calleduniformity, and is an edge expansion property: The uniformity parameterε(G)ofGis the smallestεs.t.
∀S,T⊂V,
∑
u∈S,v∈T
Auv−|S||T|
|V|
6ε|V|; the smallestε(G),the most uniformG.
−→Two ways of defining and quantifying expansion of a regular graphG: spectral perspective (throughλ(G)) and geometrical perspective (throughε(G)).
Known:If a graphGis spectrally expanding (i.e. ifλ(G)is small), then it is necessarily geometrically expanding (i.e.ε(G)is small as well). The converse (i.e.ε(G)small impliesλ(G) small) is known to hold only for some classes of graphs.
Spectral expansion and uniformity for regular graphs
G= (V,E)ad-regular graph, with vertex setVand edge setE(dedges at each vertex).
Aits normalized adjacency matrix, i.e. the|V| × |V|matrix s.t.Auv=e(u,v)
| {z }
number of edges between verticesuandv
/dfor eachu,v∈V.
•Property (1) is related tospectral expansion:
λ1(A),λ2(A), . . .eigenvalues ofA, with|λ1(A)|>|λ2(A)|>· · ·. SinceGis regular,λ1(A) =1 (all-one vector as associated eigenvector), and the spectral expansion parameter ofGis
λ(G) :=|λ2(A)|; the smallestλ(G),the most spectrally expandingG.
•Property (2) is calleduniformity, and is an edge expansion property: The uniformity parameterε(G)ofGis the smallestεs.t.
∀S,T⊂V,
∑
u∈S,v∈T
Auv−|S||T|
|V|
6ε|V|; the smallestε(G),the most uniformG.
−→Two ways of defining and quantifying expansion of a regular graphG: spectral perspective (throughλ(G)) and geometrical perspective (throughε(G)).
Known:If a graphGis spectrally expanding (i.e. ifλ(G)is small), then it is necessarily geometrically expanding (i.e.ε(G)is small as well). The converse (i.e.ε(G)small impliesλ(G) small) is known to hold only for some classes of graphs.
Spectral expansion and uniformity for regular graphs
G= (V,E)ad-regular graph, with vertex setVand edge setE(dedges at each vertex).
Aits normalized adjacency matrix, i.e. the|V| × |V|matrix s.t.Auv=e(u,v)
| {z }
number of edges between verticesuandv
/dfor eachu,v∈V.
•Property (1) is related tospectral expansion:
λ1(A),λ2(A), . . .eigenvalues ofA, with|λ1(A)|>|λ2(A)|>· · ·. SinceGis regular,λ1(A) =1 (all-one vector as associated eigenvector), and the spectral expansion parameter ofGis
λ(G) :=|λ2(A)|; the smallestλ(G),the most spectrally expandingG.
•Property (2) is calleduniformity, and is an edge expansion property:
The uniformity parameterε(G)ofGis the smallestεs.t.
∀S,T⊂V,
∑
u∈S,v∈T
Auv−|S||T|
|V|
6ε|V|; the smallestε(G),the most uniformG.
−→Two ways of defining and quantifying expansion of a regular graphG: spectral perspective (throughλ(G)) and geometrical perspective (throughε(G)).
Known:If a graphGis spectrally expanding (i.e. ifλ(G)is small), then it is necessarily geometrically expanding (i.e.ε(G)is small as well). The converse (i.e.ε(G)small impliesλ(G) small) is known to hold only for some classes of graphs.
Spectral expansion and uniformity for regular graphs
G= (V,E)ad-regular graph, with vertex setVand edge setE(dedges at each vertex).
Aits normalized adjacency matrix, i.e. the|V| × |V|matrix s.t.Auv=e(u,v)
| {z }
number of edges between verticesuandv
/dfor eachu,v∈V.
•Property (1) is related tospectral expansion:
λ1(A),λ2(A), . . .eigenvalues ofA, with|λ1(A)|>|λ2(A)|>· · ·. SinceGis regular,λ1(A) =1 (all-one vector as associated eigenvector), and the spectral expansion parameter ofGis
λ(G) :=|λ2(A)|; the smallestλ(G),the most spectrally expandingG.
•Property (2) is calleduniformity, and is an edge expansion property:
The uniformity parameterε(G)ofGis the smallestεs.t.
∀S,T⊂V,
∑
u∈S,v∈T
Auv−|S||T|
|V|
6ε|V|; the smallestε(G),the most uniformG.
−→Two ways of defining and quantifying expansion of a regular graphG: spectral perspective (throughλ(G)) and geometrical perspective (throughε(G)).
Known:If a graphGis spectrally expanding (i.e. ifλ(G)is small), then it is necessarily geometrically expanding (i.e.ε(G)is small as well). The converse (i.e.ε(G)small impliesλ(G) small) is known to hold only for some classes of graphs.
Spectral expansion and uniformity for regular graphs
G= (V,E)ad-regular graph, with vertex setVand edge setE(dedges at each vertex).
Aits normalized adjacency matrix, i.e. the|V| × |V|matrix s.t.Auv=e(u,v)
| {z }
number of edges between verticesuandv
/dfor eachu,v∈V.
•Property (1) is related tospectral expansion:
λ1(A),λ2(A), . . .eigenvalues ofA, with|λ1(A)|>|λ2(A)|>· · ·. SinceGis regular,λ1(A) =1 (all-one vector as associated eigenvector), and the spectral expansion parameter ofGis
λ(G) :=|λ2(A)|; the smallestλ(G),the most spectrally expandingG.
•Property (2) is calleduniformity, and is an edge expansion property:
The uniformity parameterε(G)ofGis the smallestεs.t.
∀S,T⊂V,
∑
u∈S,v∈T
Auv−|S||T|
|V|
6ε|V|; the smallestε(G),the most uniformG.
−→Two ways of defining and quantifying expansion of a regular graphG: spectral perspective (throughλ(G)) and geometrical perspective (throughε(G)).
Known:If a graphGis spectrally expanding (i.e. ifλ(G)is small), then it is necessarily geometrically expanding (i.e.ε(G)is small as well). The converse (i.e.ε(G)small impliesλ(G) small) is known to hold only for some classes of graphs.
Quantum analogues
Adjacency matrixAof a graph: transition matrix (maps probability vectors to probability vectors).
−→Quantum analogue: quantum channelΦ(maps density operators to density operators).
Regularity condition onG:Aleaves the all-one vector invariant.
−→Quantum analogue:Φleaves the identity matrix invariant, i.e.Φis unital.
Φa unital quantum channel, acting on operators onCn. It hasspectral expansionparameter λ(Φ) :=|λ2(Φ)|,
anduniformityparameterε(Φ), the smallestεs.t.
∀V,W ⊂Cn,
Tr(PVΦ(PW))−Tr(PV)Tr(PW) n
6εn.
•Quantum spectral expansion: first introduced to study entanglement in quantum many-body 1D systems(Hastings)and complexity of quantum entropy estimation(Ben-Aroya/Schwartz/Ta-Shma). Since then, great interest forquantum expandersin various fields: randomness reduction, quantum cryptography, quantum error-correction, quantum many-body physics etc.
•Quantum uniformity: formalized only recently(Bannink/Briët/Labib/Maassen).
−→Proving equivalence between these two notions of expansion could open the way to new ways of constructing quantum expanders.
Quantum analogues
Adjacency matrixAof a graph: transition matrix (maps probability vectors to probability vectors).
−→Quantum analogue: quantum channelΦ(maps density operators to density operators).
Regularity condition onG:Aleaves the all-one vector invariant.
−→Quantum analogue:Φleaves the identity matrix invariant, i.e.Φis unital.
Φa unital quantum channel, acting on operators onCn. It hasspectral expansionparameter λ(Φ) :=|λ2(Φ)|,
anduniformityparameterε(Φ), the smallestεs.t.
∀V,W ⊂Cn,
Tr(PVΦ(PW))−Tr(PV)Tr(PW) n
6εn.
•Quantum spectral expansion: first introduced to study entanglement in quantum many-body 1D systems(Hastings)and complexity of quantum entropy estimation(Ben-Aroya/Schwartz/Ta-Shma). Since then, great interest forquantum expandersin various fields: randomness reduction, quantum cryptography, quantum error-correction, quantum many-body physics etc.
•Quantum uniformity: formalized only recently(Bannink/Briët/Labib/Maassen).
−→Proving equivalence between these two notions of expansion could open the way to new ways of constructing quantum expanders.
Quantum analogues
Adjacency matrixAof a graph: transition matrix (maps probability vectors to probability vectors).
−→Quantum analogue: quantum channelΦ(maps density operators to density operators).
Regularity condition onG:Aleaves the all-one vector invariant.
−→Quantum analogue:Φleaves the identity matrix invariant, i.e.Φis unital.
Φa unital quantum channel, acting on operators onCn. It hasspectral expansionparameter λ(Φ) :=|λ2(Φ)|,
anduniformityparameterε(Φ), the smallestεs.t.
∀V,W ⊂Cn,
Tr(PVΦ(PW))−Tr(PV)Tr(PW) n
6εn.
•Quantum spectral expansion: first introduced to study entanglement in quantum many-body 1D systems(Hastings)and complexity of quantum entropy estimation(Ben-Aroya/Schwartz/Ta-Shma). Since then, great interest forquantum expandersin various fields: randomness reduction, quantum cryptography, quantum error-correction, quantum many-body physics etc.
•Quantum uniformity: formalized only recently(Bannink/Briët/Labib/Maassen).
−→Proving equivalence between these two notions of expansion could open the way to new ways of constructing quantum expanders.
Tensor norms on commutative and non-commutative Banach spaces
•Given a vectorx∈Cn, its (normalized)`pnorm is:kxk`p:= 1 n
n
∑
i=1
|xi|p
!1/p
. Given a linear operatorM:Cn→Cn, its`p→`qnorm is:
kMk`p→`q:= max 1
n|hy|M|xi|,x,y∈Cn,kxk`p=1,kyk`
q0 =1
. And the cut norm ofMis:kMkcut:= max
1
n|hy|M|xi|,x,y∈ {0,1}n
.
•Given a matrixX∈
M
n(C), its (normalized)Spnorm is:kXkSp:= 1nTr(|X|p) 1/p
. Given a linear operatorΨ :
M
n(C)→M
n(C), itsSp→Sqnorm is:kΨkSp→Sq:= max 1
n|Tr(Y∗Ψ(X))|,X,Y∈
M
n(C),kXkSp=1,kYkSq0 =1. And the cut norm ofΨiskΨkcut:= max
1
n|Tr(Y∗Ψ(X))|,X,Yprojectors onCn
. Fact:The commutative, resp. non-commutative, cut-norm is equivalent to the`∞→`1, resp.S∞→S1, norm, with domination constants independent from the dimension. Specifically:
∀M:Cn→Cn,kMkcut6kMk`∞→`16π2kMkcut(Conlon/Zhao).
∀Ψ :
M
n(C)→M
n(C),kΨkcut6kΨkS∞→S16π2kΨkcut(Bannink/Briët/Labib/Maassen).Tensor norms on commutative and non-commutative Banach spaces
•Given a vectorx∈Cn, its (normalized)`pnorm is:kxk`p:= 1 n
n
∑
i=1
|xi|p
!1/p
. Given a linear operatorM:Cn→Cn, its`p→`qnorm is:
kMk`p→`q:= max 1
n|hy|M|xi|,x,y∈Cn,kxk`p=1,kyk`
q0 =1
. And the cut norm ofMis:kMkcut:= max
1
n|hy|M|xi|,x,y∈ {0,1}n
.
•Given a matrixX∈
M
n(C), its (normalized)Spnorm is:kXkSp:=1 nTr(|X|p)
1/p
. Given a linear operatorΨ :
M
n(C)→M
n(C), itsSp→Sqnorm is:kΨkSp→Sq:= max 1
n|Tr(Y∗Ψ(X))|,X,Y∈
M
n(C),kXkSp=1,kYkSq0 =1. And the cut norm ofΨiskΨkcut:= max
1
n|Tr(Y∗Ψ(X))|,X,Yprojectors onCn
.
Fact:The commutative, resp. non-commutative, cut-norm is equivalent to the`∞→`1, resp.S∞→S1, norm, with domination constants independent from the dimension. Specifically:
∀M:Cn→Cn,kMkcut6kMk`∞→`16π2kMkcut(Conlon/Zhao).
∀Ψ :
M
n(C)→M
n(C),kΨkcut6kΨkS∞→S16π2kΨkcut(Bannink/Briët/Labib/Maassen).Tensor norms on commutative and non-commutative Banach spaces
•Given a vectorx∈Cn, its (normalized)`pnorm is:kxk`p:= 1 n
n
∑
i=1
|xi|p
!1/p
. Given a linear operatorM:Cn→Cn, its`p→`qnorm is:
kMk`p→`q:= max 1
n|hy|M|xi|,x,y∈Cn,kxk`p=1,kyk`
q0 =1
. And the cut norm ofMis:kMkcut:= max
1
n|hy|M|xi|,x,y∈ {0,1}n
.
•Given a matrixX∈
M
n(C), its (normalized)Spnorm is:kXkSp:=1 nTr(|X|p)
1/p
. Given a linear operatorΨ :
M
n(C)→M
n(C), itsSp→Sqnorm is:kΨkSp→Sq:= max 1
n|Tr(Y∗Ψ(X))|,X,Y∈
M
n(C),kXkSp=1,kYkSq0 =1. And the cut norm ofΨiskΨkcut:= max
1
n|Tr(Y∗Ψ(X))|,X,Yprojectors onCn
. Fact:The commutative, resp. non-commutative, cut-norm is equivalent to the`∞→`1, resp.S∞→S1, norm, with domination constants independent from the dimension. Specifically:
∀M:Cn→Cn,kMkcut6kMk`∞→`16π2kMkcut(Conlon/Zhao).
∀Ψ :
M
n(C)→M
n(C),kΨkcut6kΨkS∞→S16π2kΨkcut(Bannink/Briët/Labib/Maassen).Proof of the inequality relating the cut norm and the
S∞→
S1norm
•Since projectors haveS∞norm equal to 1, we have:
max 1
n|Tr(Y∗Ψ(X))|,kXkS∞=kYkS∞=1
>
1
n|Tr(Y∗Ψ(X))|,X,Yprojectors
. By definition, this means thatkΨkS∞→S1>kΨkcut.
•The set of matrices withS∞norm at most 1 is the convex hull of unitary matrices. So there exist unitariesX,Ys.t.kΨkS∞→S1=1
n|Tr(Y∗Ψ(X))|. WriteX=UAU∗andY=VBV∗, withA,Bdiagonal andU,Vunitary. For anyw∈Cs.t.|w|=1, setA0(w)ii=1{R(A
iiw¯)},B0(w)ii=1{R(B
iiw¯)}, 16i6n. Observe that, for anyz∈Cs.t.|z|=1,z=πEw∈C,|w|=1
w1{R(zw¯)}
. Hence,X=πEu(u UA0(u)U∗)andY=πEv(v VB0(v)V∗).
Now,UA0(u)U∗andVB0(u)V∗are projectors for allu,v∈Cs.t.|u|=|v|=1, and therefore kΨkS∞→S1=1
nTr(Y∗Ψ(X))
=π2 n
Eu,v
¯
v uTr(V∗B0(v)∗VΨ(UA0(u)U∗))
6π2 nEu,v
Tr(V∗B0(v)∗VΨ(UA0(u)U∗))
6π2kΨkcut
Proof of the inequality relating the cut norm and the
S∞→
S1norm
•Since projectors haveS∞norm equal to 1, we have:
max 1
n|Tr(Y∗Ψ(X))|,kXkS∞=kYkS∞=1
>
1
n|Tr(Y∗Ψ(X))|,X,Yprojectors
. By definition, this means thatkΨkS∞→S1>kΨkcut.
•The set of matrices withS∞norm at most 1 is the convex hull of unitary matrices.
So there exist unitariesX,Ys.t.kΨkS∞→S1=1
n|Tr(Y∗Ψ(X))|. WriteX=UAU∗andY=VBV∗, withA,Bdiagonal andU,Vunitary.
For anyw∈Cs.t.|w|=1, setA0(w)ii=1{R(A
iiw¯)},B0(w)ii=1{R(B
iiw¯)}, 16i6n.
Observe that, for anyz∈Cs.t.|z|=1,z=πEw∈C,|w|=1
w1{R(zw¯)}
. Hence,X=πEu(u UA0(u)U∗)andY=πEv(v VB0(v)V∗).
Now,UA0(u)U∗andVB0(u)V∗are projectors for allu,v∈Cs.t.|u|=|v|=1, and therefore kΨkS∞→S1=1
nTr(Y∗Ψ(X))
=π2 n Eu,v
¯
v uTr(V∗B0(v)∗VΨ(UA0(u)U∗))
6π2 nEu,v
Tr(V∗B0(v)∗VΨ(UA0(u)U∗))
6π2kΨkcut
Characterizing spectral expansion and uniformity through tensor norms
Observation:The parametersλandεcan be re-expressed in terms oftensor norms.•For any regular graphGonnvertices, with adjacency matrixA, λ(G) =kA−Jk`
2→`2 andε(G) =kA−Jkcut,
whereJis the adjacency matrix of the complete graph, i.e. the matrix with all entries equal to 1/n.
•For any unital quantum channelΦonn×nmatrices, λ(Φ) =kΦ−ΠkS
2→S2 andε(Φ) =kΦ−Πkcut, whereΠis the completely randomizing quantum channel, i.e.Π :X7→Tr(X)I/n.
−→Comparing parametersλandε, for graphs or quantum channels, boils down to comparing
`2→`2and`∞→`1norms orS2→S2andS∞→S1norms.
Fact:By ordering of the`pandSpnorms,k · k`2→`2>k · k`∞→`1andk · kS2→S2>k · kS∞→S1. So for any regular graphGand unital quantum channelΦ, spectral expansion implies uniformity:
λ(G) =kA−Jk`
2→`2>kA−Jk`
∞→`1>kA−Jkcut=ε(G), λ(Φ) =kΦ−ΠkS
2→S2>kΦ−ΠkS
∞→S1>kΦ−Πkcut=ε(Φ).
Characterizing spectral expansion and uniformity through tensor norms
Observation:The parametersλandεcan be re-expressed in terms oftensor norms.
•For any regular graphGonnvertices, with adjacency matrixA, λ(G) =kA−Jk`
2→`2 andε(G) =kA−Jkcut,
whereJis the adjacency matrix of the complete graph, i.e. the matrix with all entries equal to 1/n.
•For any unital quantum channelΦonn×nmatrices, λ(Φ) =kΦ−ΠkS
2→S2 andε(Φ) =kΦ−Πkcut, whereΠis the completely randomizing quantum channel, i.e.Π :X7→Tr(X)I/n.
−→Comparing parametersλandε, for graphs or quantum channels, boils down to comparing
`2→`2and`∞→`1norms orS2→S2andS∞→S1norms.
Fact:By ordering of the`pandSpnorms,k · k`2→`2>k · k`∞→`1andk · kS2→S2>k · kS∞→S1. So for any regular graphGand unital quantum channelΦ, spectral expansion implies uniformity:
λ(G) =kA−Jk`
2→`2>kA−Jk`
∞→`1>kA−Jkcut=ε(G), λ(Φ) =kΦ−ΠkS
2→S2>kΦ−ΠkS
∞→S1>kΦ−Πkcut=ε(Φ).
Characterizing spectral expansion and uniformity through tensor norms
Observation:The parametersλandεcan be re-expressed in terms oftensor norms.
•For any regular graphGonnvertices, with adjacency matrixA, λ(G) =kA−Jk`
2→`2 andε(G) =kA−Jkcut,
whereJis the adjacency matrix of the complete graph, i.e. the matrix with all entries equal to 1/n.
•For any unital quantum channelΦonn×nmatrices, λ(Φ) =kΦ−ΠkS
2→S2 andε(Φ) =kΦ−Πkcut, whereΠis the completely randomizing quantum channel, i.e.Π :X7→Tr(X)I/n.
−→Comparing parametersλandε, for graphs or quantum channels, boils down to comparing
`2→`2and`∞→`1norms orS2→S2andS∞→S1norms.
Fact:By ordering of the`pandSpnorms,k · k`2→`2>k · k`∞→`1andk · kS2→S2>k · kS∞→S1. So for any regular graphGand unital quantum channelΦ, spectral expansion implies uniformity:
λ(G) =kA−Jk`
2→`2>kA−Jk`
∞→`1>kA−Jkcut=ε(G), λ(Φ) =kΦ−ΠkS
2→S2>kΦ−ΠkS
∞→S1>kΦ−Πkcut=ε(Φ).
Characterizing spectral expansion and uniformity through tensor norms
Observation:The parametersλandεcan be re-expressed in terms oftensor norms.
•For any regular graphGonnvertices, with adjacency matrixA, λ(G) =kA−Jk`
2→`2 andε(G) =kA−Jkcut,
whereJis the adjacency matrix of the complete graph, i.e. the matrix with all entries equal to 1/n.
•For any unital quantum channelΦonn×nmatrices, λ(Φ) =kΦ−ΠkS
2→S2 andε(Φ) =kΦ−Πkcut, whereΠis the completely randomizing quantum channel, i.e.Π :X7→Tr(X)I/n.
−→Comparing parametersλandε, for graphs or quantum channels, boils down to comparing
`2→`2and`∞→`1norms orS2→S2andS∞→S1norms.
Fact:By ordering of the`pandSpnorms,k · k`2→`2>k · k`∞→`1andk · kS2→S2>k · kS∞→S1. So for any regular graphGand unital quantum channelΦ, spectral expansion implies uniformity:
λ(G) =kA−Jk`
2→`2>kA−Jk`
∞→`1>kA−Jkcut=ε(G), λ(Φ) =kΦ−ΠkS
2→S2>kΦ−ΠkS
∞→S1>kΦ−Πkcut=ε(Φ).
Proof of tensor norm expressions for λ and ε
Note that theS2→S2norm ofΦ−Πis simply its operator norm, i.e. its largest eigenvalue.
Πhas eigenvalues 1,0, . . . ,0, withIas eigenvector associated to 1.
Φhas eigenvalues 1,λ2(Φ), . . . ,λn(Φ), withIas eigenvector associated to 1.
So we indeed have:|λ1(Φ−Π)|=|λ2(Φ)|=λ(Φ).
For any projectorsX,YonCn, we have: 1
nTr(Y∗(Φ−Π)(X)) = 1
nTr(Y∗Φ(X))− 1
n2Tr(Y∗)Tr(X).
By definition, the maximum of the l.h.s. iskΦ−Πkcutand the maximum of the r.h.s. isε(Φ).
Proof of tensor norm expressions for λ and ε
Note that theS2→S2norm ofΦ−Πis simply its operator norm, i.e. its largest eigenvalue.
Πhas eigenvalues 1,0, . . . ,0, withIas eigenvector associated to 1.
Φhas eigenvalues 1,λ2(Φ), . . . ,λn(Φ), withIas eigenvector associated to 1.
So we indeed have:|λ1(Φ−Π)|=|λ2(Φ)|=λ(Φ). For any projectorsX,YonCn, we have:
1
nTr(Y∗(Φ−Π)(X)) = 1
nTr(Y∗Φ(X))− 1
n2Tr(Y∗)Tr(X).
By definition, the maximum of the l.h.s. iskΦ−Πkcutand the maximum of the r.h.s. isε(Φ).
Equivalence between spectral expansion and uniformity in the dense case
•Classical case:Gad-regular graph onnvertices,δ:=d/nitsedge density.
λ(G)6 2ε(G)
δ2 1/4
, so ifδ= Ω(1), thenλ(G)6Cε(G)1/4(Chung/Graham/Wilson).
−→For a dense regular graph, uniformity implies spectral expansion.
•Quantum case:Φa unital quantum channel,η:=kΦ−ΠkS1→S∞ itsrandomizing parameter. λ(Φ)6 π2η3ε(G)1/4
, so ifη=O(1), thenλ(G)6Cε(G)1/4(Bannink/Briët/Labib/Maassen).
−→For a randomizing unital channel, uniformity implies spectral expansion.
Observation:In the classical case,η:=kA−Jk`1→`∞=1/δ−1, so the conditionδ= Ω(1)is equivalent to the conditionη=O(1).
−→The quantum statement can be seen as similar to the classical one.
In fact, the inequality involvingηrather thanδholds classically exactly as quantumly, and is better than the original one in the regime whereηis small enough (i.e.δis large enough).
Main technical tool in the proof:Clever use of Cauchy-Schwartz inequality (thus valid for`p norms as forSpnorms).
Equivalence between spectral expansion and uniformity in the dense case
•Classical case:Gad-regular graph onnvertices,δ:=d/nitsedge density.
λ(G)6 2ε(G)
δ2 1/4
, so ifδ= Ω(1), thenλ(G)6Cε(G)1/4(Chung/Graham/Wilson).
−→For a dense regular graph, uniformity implies spectral expansion.
•Quantum case:Φa unital quantum channel,η:=kΦ−ΠkS1→S∞ itsrandomizing parameter.
λ(Φ)6 π2η3ε(G)1/4
, so ifη=O(1), thenλ(G)6Cε(G)1/4(Bannink/Briët/Labib/Maassen).
−→For a randomizing unital channel, uniformity implies spectral expansion.
Observation:In the classical case,η:=kA−Jk`1→`∞=1/δ−1, so the conditionδ= Ω(1)is equivalent to the conditionη=O(1).
−→The quantum statement can be seen as similar to the classical one.
In fact, the inequality involvingηrather thanδholds classically exactly as quantumly, and is better than the original one in the regime whereηis small enough (i.e.δis large enough).
Main technical tool in the proof:Clever use of Cauchy-Schwartz inequality (thus valid for`p norms as forSpnorms).
Equivalence between spectral expansion and uniformity in the dense case
•Classical case:Gad-regular graph onnvertices,δ:=d/nitsedge density.
λ(G)6 2ε(G)
δ2 1/4
, so ifδ= Ω(1), thenλ(G)6Cε(G)1/4(Chung/Graham/Wilson).
−→For a dense regular graph, uniformity implies spectral expansion.
•Quantum case:Φa unital quantum channel,η:=kΦ−ΠkS1→S∞ itsrandomizing parameter.
λ(Φ)6 π2η3ε(G)1/4
, so ifη=O(1), thenλ(G)6Cε(G)1/4(Bannink/Briët/Labib/Maassen).
−→For a randomizing unital channel, uniformity implies spectral expansion.
Observation:In the classical case,η:=kA−Jk`1→`∞=1/δ−1, so the conditionδ= Ω(1)is equivalent to the conditionη=O(1).
−→The quantum statement can be seen as similar to the classical one.
In fact, the inequality involvingηrather thanδholds classically exactly as quantumly, and is better than the original one in the regime whereηis small enough (i.e.δis large enough).
Main technical tool in the proof:Clever use of Cauchy-Schwartz inequality (thus valid for`p norms as forSpnorms).
Equivalence between spectral expansion and uniformity in the dense case
•Classical case:Gad-regular graph onnvertices,δ:=d/nitsedge density.
λ(G)6 2ε(G)
δ2 1/4
, so ifδ= Ω(1), thenλ(G)6Cε(G)1/4(Chung/Graham/Wilson).
−→For a dense regular graph, uniformity implies spectral expansion.
•Quantum case:Φa unital quantum channel,η:=kΦ−ΠkS1→S∞ itsrandomizing parameter.
λ(Φ)6 π2η3ε(G)1/4
, so ifη=O(1), thenλ(G)6Cε(G)1/4(Bannink/Briët/Labib/Maassen).
−→For a randomizing unital channel, uniformity implies spectral expansion.
Observation:In the classical case,η:=kA−Jk`1→`∞=1/δ−1, so the conditionδ= Ω(1)is equivalent to the conditionη=O(1).
−→The quantum statement can be seen as similar to the classical one.
In fact, the inequality involvingηrather thanδholds classically exactly as quantumly, and is better than the original one in the regime whereηis small enough (i.e.δis large enough).
Main technical tool in the proof:Clever use of Cauchy-Schwartz inequality (thus valid for`p norms as forSpnorms).
Equivalence between spectral expansion and uniformity in the symmetric case
•Classical case:Gavertex-transitiveregular graph onnvertices.
This means: There existsΓa transitive subgroup of the group of permutations ofnelements, s.t.∀γ∈Γ,MγAMγ−1=A, whereMγis then×npermutation matrix associated toγ. By a factorization version of thecommutative Grothendieck inequality(Pisier):
kA−Jk`2→`26KkA−Jk`∞→`1,with 1.33<K<1.41 the complex Grothendieck constant.
−→For any vertex-transitive regular graphG,
λ(G) =kA−Jk`2→`26KkA−Jk`∞→`16Kπ2kA−Jkcut=Kπ2ε(G).
•Quantum case:Φanirreducibly covariantunital quantum channel.
This means: There exists a compact groupΓwith irreducible unitary representations γ∈Γ7→Uγ∈U(n)andγ∈Γ7→Vγ∈U(n)s.t.∀γ∈Γ,Φ(Uγ·Uγ−1) =VγΦ(·)Vγ−1. By a factorization version of thenon-commutative Grothendieck inequality(Haagerup):
kΦ−ΠkS2→S262kΦ−ΠkS∞→S1.
−→For any irreducibly covariant unital quantum channelΦ,
λ(Φ) =kΦ−ΠkS2→S262kΦ−ΠkS∞→S162π2kΦ−Πkcut=2π2ε(Φ).
Equivalence between spectral expansion and uniformity in the symmetric case
•Classical case:Gavertex-transitiveregular graph onnvertices.
This means: There existsΓa transitive subgroup of the group of permutations ofnelements, s.t.∀γ∈Γ,MγAMγ−1=A, whereMγis then×npermutation matrix associated toγ. By a factorization version of thecommutative Grothendieck inequality(Pisier):
kA−Jk`2→`26KkA−Jk`∞→`1,with 1.33<K<1.41 the complex Grothendieck constant.
−→For any vertex-transitive regular graphG,
λ(G) =kA−Jk`2→`26KkA−Jk`∞→`16Kπ2kA−Jkcut=Kπ2ε(G).
•Quantum case:Φanirreducibly covariantunital quantum channel.
This means: There exists a compact groupΓwith irreducible unitary representations γ∈Γ7→Uγ∈U(n)andγ∈Γ7→Vγ∈U(n)s.t.∀γ∈Γ,Φ(Uγ·Uγ−1) =VγΦ(·)Vγ−1. By a factorization version of thenon-commutative Grothendieck inequality(Haagerup):
kΦ−ΠkS2→S262kΦ−ΠkS∞→S1.
−→For any irreducibly covariant unital quantum channelΦ,
λ(Φ) =kΦ−ΠkS2→S262kΦ−ΠkS∞→S162π2kΦ−Πkcut=2π2ε(Φ).
Main ideas in the proof
Goal:Show that, ifψ:
M
n(C)→M
n(C)is irreducibly covariant, thenkΨkS2→S262kΨkS∞→S1.Fact 1 (Factorized non-commutative Grothendieck inequality):
Letψ:
M
n(C)→M
n(C). There exist statesρ,ρ0,σ,σ0onCns.t., for allX,Y∈M
n(C),1
nTr(Y∗Ψ(X))
6kΨkS∞→S1 Tr(ρXX∗) +Tr(ρ0X∗X)1/2
Tr(σYY∗) +Tr(σ0Y∗Y)1/2
.
Fact 2 (Characterization of irreducible unitary representations):
LetΓbe a compact group. A unitary representationγ∈Γ7→Uγ∈U(n)is irreducible iff
∀X∈
M
n(C), 1|Γ|
∑
γ∈Γ
UγXUγ∗=Tr(X)I n.
By assumption onΨ, 1
nTr(Y∗Ψ(X))
= 1
|Γ|
∑
γ∈Γ
1
nTr(Yγ∗Ψ(Xγ))
,Xγ:=UγXUγ∗,Yγ:=VγXVγ∗. By Fact 1 and the AM-GM inequality, the r.h.s. above is upper bounded by:
1
2kΨkS∞→S1 1
|Γ|
∑
γ∈Γ
Tr(ρXγXγ∗) +Tr(ρ0Xγ∗Xγ) +Tr(σYγYγ∗) +Tr(σ0Yγ∗Yγ) . By Fact 2, 1
|Γ|
∑
γ∈Γ
Tr(ρXγXγ∗) =kXk2S
2, and similarly for the other terms. Hence, for anyX,Y∈
M
n(C)s.t.kXkS2=kYkS2=1,1
nTr(Y∗Ψ(X))
62kΨkS∞→S1.
Main ideas in the proof
Goal:Show that, ifψ:
M
n(C)→M
n(C)is irreducibly covariant, thenkΨkS2→S262kΨkS∞→S1. Fact 1 (Factorized non-commutative Grothendieck inequality):Letψ:
M
n(C)→M
n(C). There exist statesρ,ρ0,σ,σ0onCns.t., for allX,Y∈M
n(C),1
nTr(Y∗Ψ(X))
6kΨkS∞→S1 Tr(ρXX∗) +Tr(ρ0X∗X)1/2
Tr(σYY∗) +Tr(σ0Y∗Y)1/2
.
Fact 2 (Characterization of irreducible unitary representations):
LetΓbe a compact group. A unitary representationγ∈Γ7→Uγ∈U(n)is irreducible iff
∀X∈
M
n(C), 1|Γ|
∑
γ∈Γ
UγXUγ∗=Tr(X)I n.
By assumption onΨ, 1
nTr(Y∗Ψ(X))
= 1
|Γ|
∑
γ∈Γ
1
nTr(Yγ∗Ψ(Xγ))
,Xγ:=UγXUγ∗,Yγ:=VγXVγ∗. By Fact 1 and the AM-GM inequality, the r.h.s. above is upper bounded by:
1
2kΨkS∞→S1 1
|Γ|
∑
γ∈Γ
Tr(ρXγXγ∗) +Tr(ρ0Xγ∗Xγ) +Tr(σYγYγ∗) +Tr(σ0Yγ∗Yγ) . By Fact 2, 1
|Γ|
∑
γ∈Γ
Tr(ρXγXγ∗) =kXk2S
2, and similarly for the other terms. Hence, for anyX,Y∈
M
n(C)s.t.kXkS2=kYkS2=1,1
nTr(Y∗Ψ(X))
62kΨkS∞→S1.
Main ideas in the proof
Goal:Show that, ifψ:
M
n(C)→M
n(C)is irreducibly covariant, thenkΨkS2→S262kΨkS∞→S1. Fact 1 (Factorized non-commutative Grothendieck inequality):Letψ:
M
n(C)→M
n(C). There exist statesρ,ρ0,σ,σ0onCns.t., for allX,Y∈M
n(C),1
nTr(Y∗Ψ(X))
6kΨkS∞→S1 Tr(ρXX∗) +Tr(ρ0X∗X)1/2
Tr(σYY∗) +Tr(σ0Y∗Y)1/2
.
Fact 2 (Characterization of irreducible unitary representations):
LetΓbe a compact group. A unitary representationγ∈Γ7→Uγ∈U(n)is irreducible iff
∀X∈
M
n(C), 1|Γ|
∑
γ∈Γ
UγXUγ∗=Tr(X)I n.
By assumption onΨ, 1
nTr(Y∗Ψ(X))
= 1
|Γ|
∑
γ∈Γ
1
nTr(Yγ∗Ψ(Xγ))
,Xγ:=UγXUγ∗,Yγ:=VγXVγ∗. By Fact 1 and the AM-GM inequality, the r.h.s. above is upper bounded by:
1
2kΨkS∞→S1 1
|Γ|
∑
γ∈Γ
Tr(ρXγXγ∗) +Tr(ρ0Xγ∗Xγ) +Tr(σYγYγ∗) +Tr(σ0Yγ∗Yγ) . By Fact 2, 1
|Γ|
∑
γ∈Γ
Tr(ρXγXγ∗) =kXk2S
2, and similarly for the other terms.
Hence, for anyX,Y∈
M
n(C)s.t.kXkS2=kYkS2=1, 1nTr(Y∗Ψ(X))
62kΨkS∞→S1.
Embedding the classical setting in the quantum one
GivenA:Cn→Cn, defineΦA:X∈
M
n(C)7→n
∑
i,j=1
AijXjj|iihi| ∈
M
n(C).−→Embedding of adjacency matrices of regular graphs into unital quantum channels.
Sanity check:The adjacency matrix of the complete graph is embedded into the completely randomizing quantum channel, i.e.ΦJ= Π.
Moreover:kΦA−ΠkSp→Sq=kA−Jk`p→`qandkΦA−Πkcut=kA−Jkcut.
−→Spectral expansion and uniformity parameters are preserved under the embedding.
Preservation of density and symmetry properties under the embedding:
ΦAirreducibly covariant quantum channel⇔Aadjacency matrix of a vertex-transitive graph. ΦAO(1)-randomizing quantum channel⇔Aadjacency matrix of anΩ(1)-dense graph. Interest:There are examples of non-dense regular graphsGwhich are not vertex-transitive, and s.t.ε(G)is small butλ(G)is large. They provide examples of non-randomizing unital channelsΦ which are not irreducibly covariant, and s.t.ε(Φ)is small butλ(Φ)is large.
Embedding the classical setting in the quantum one
GivenA:Cn→Cn, defineΦA:X∈
M
n(C)7→n
∑
i,j=1
AijXjj|iihi| ∈
M
n(C).−→Embedding of adjacency matrices of regular graphs into unital quantum channels.
Sanity check:The adjacency matrix of the complete graph is embedded into the completely randomizing quantum channel, i.e.ΦJ= Π.
Moreover:kΦA−ΠkSp→Sq=kA−Jk`p→`qandkΦA−Πkcut=kA−Jkcut.
−→Spectral expansion and uniformity parameters are preserved under the embedding.
Preservation of density and symmetry properties under the embedding:
ΦAirreducibly covariant quantum channel⇔Aadjacency matrix of a vertex-transitive graph.
ΦAO(1)-randomizing quantum channel⇔Aadjacency matrix of anΩ(1)-dense graph.
Interest:There are examples of non-dense regular graphsGwhich are not vertex-transitive, and s.t.ε(G)is small butλ(G)is large. They provide examples of non-randomizing unital channelsΦ which are not irreducibly covariant, and s.t.ε(Φ)is small butλ(Φ)is large.
Final comments
Optimality 1:Being vertex-transitive, resp. irreducibly covariant, is a necessary condition for a non-dense regular graph, resp. non-randomizing unital channel, to be s.t. uniformity implies spectral expansion.
Optimality 2:In the classical case of a vertex-transitive regular graphG, the inequality λ(G)6Kπ2ε(G)is optimal. In the quantum case of an irreducibly covariant unital channel Φ, it is not known whether the inequalityλ(Φ)62π2ε(Φ)is optimal (each of the two steps in the proof is optimal, but their combination might not be).
Final comments
Optimality 1:Being vertex-transitive, resp. irreducibly covariant, is a necessary condition for a non-dense regular graph, resp. non-randomizing unital channel, to be s.t. uniformity implies spectral expansion.
Optimality 2:In the classical case of a vertex-transitive regular graphG, the inequality λ(G)6Kπ2ε(G)is optimal. In the quantum case of an irreducibly covariant unital channel Φ, it is not known whether the inequalityλ(Φ)62π2ε(Φ)is optimal (each of the two steps in the proof is optimal, but their combination might not be).
References
T. Bannink, J. Briët, F. Labib, H. Maassen.Quasirandom quantum channels. 2020.
A. Ben-Aroya, O. Schwartz, A. Ta-Shma.Quantum expanders: motivation and construction. 2010.
D. Conlon, Y. ZhaoQuasirandom Cayley graphs. 2017.
F.R.K. Chung, R.L. Graham, R.M. Wilson.Quasi-random graphs. 1988.
U. Haagerup.The Grothendieck inequality for bilinear forms on C*-algebras. 1985.
M.B. Hastings.Random unitaries give quantum expanders. 2007.
S. Hoory, N. Linial, A. Widgerson.Expander graphs and their application. 2006.
G. Pisier.Grothendieck’s theorem, past and present. 2011.