Low rank approximation for the numerical simulation of high dimensional Lindblad
equations
pierre.rouchon@mines-paristech.fr
Based on the arXiv preprint
http://arxiv.org/abs/1207.4580 with Claude Le Bris from CERMICS
PRACQSYS, September 2012, Tokyo
Outline
Simulation of high dimensional Lindblad equations
The low rank approximation
Numerical scheme
Numerical tests for oscillation revivals
Concluding remarks
High dimensional Lindblad equations
I
The Lindblad master equation governing open-quantum systems:
d
dtρ=−i[H, ρ]−12(L†Lρ+ρL†L) +LρL†,
where
ρis the density operator (ρ
†=ρ, Tr(ρ) =1,
ρ≥0),
His an Hermitian operator and
Lis any operator on the Hilbert space
Hof dimension
n=dim
H.I
Usually,
n=Qcj=1nj large
comes from
H=H1⊗ H2⊗. . .⊗ Hc
where each
Hjis of small or intermediate dimension
nj n. Moreover, the operatorsHand
Lare usually defined as sums with few terms of simple tensor products of operators acting only on some
Hj.
I
Typical situations of composite systems: coherent
feedback scheme, circuit/cavity QED, . . .
Quantum Monte-Carlo (QMC) simulations
1The Lindbald equation
dtdρ=−i[H, ρ]−12(L†Lρ+ρL†L) +LρL†,is the master equation of the
stochastic systemd|ψti= −iH−12L†L+hψt|L†L|ψti
|ψtidt+
L|ψti
√hψt|L†L|ψti− |ψti
dNt
with
dNt ∈ {0,1},
E(dNt) =hψt|L†L|ψtidt(Poisson process).
Monte-Carlo simulations: simulate
Nrealizations of such stochastic Schrödinger equation
[0,T]3t 7→ |ψkti,k =1, . . .
N: for
Nlarge (typically
N ∼1000)
ρt ≈ N1
N
X
k=1
|ψktihψkt|.
1J. Dalibard, Y. Castion, and K. Mølmer. Wave-function approach to dissipative processes in quantum optics.Phys. Rev. Lett., 68(5):580–583, 1992.
Approximation by projection methods
2 3Based on physical intuition, select an adaptedsub-set of density matrices, i.e. a sub-manifoldDof the vector space of Hermitian matrices equipped with Frobenius Euclidian metric. The approximate evolution is given by theorthogonal projectionΠρ(dρ/dt)ofdρ/dt onto the tangent space atρtoD:
forρ∈ D, d dtρ=
vector field onD
z }| {
Πρ
−i[H, ρ]−12(L†Lρ+ρL†L) +LρL†
.
In3, Mabuchi considers a reduced order model for a spin-spring system. The sub-manifoldDwas the (real) 5-dimensional manifold constructed with the tensor products of arbitrary two-level states and pure coherent states.
Computation ofΠρ(dρ/dt)in local coordinatesis not trivial and yields usually to nonlinear ODEs.
2R. van Handel and H. Mabuchi. Quantum projection filter for a highly nonlinear model in cavity qed.Journal of Optics B: Quantum and Semiclassical Optics, 7(10):S226, 2005.
3H. Mabuchi. Derivation of Maxwell-Bloch-type equations by projection of quantum models.Phys. Rev. A, 78:015801, Jul 2008.
Low rank Kalman filters
4Fordx=Ax dt+G dω,dy=C dx+H dη, computation of the best estimate ofx attknowing the past values of the outputy relies on the computation of theconditional error covariance matrixPsolution of the Riccati matrix equation
d
dtP=AP+PA0+GG0−PC0(HH0)−1CP.
WhenG=0, the Riccati equation is rank preserving. It defines then a vector field on thesub-manifold of rankm<ncovariance matrices (n=dimx here). This sub-manifold admits theover-parameterization
(U,R)7→URU0=P!
U
z}|{
R
z}|{
U0
z }| {=
P
z }| {
whereUbelongs to the set ofn×morthogonal matrices (U0U=Im) andRism×m, positive definite and symmetric.
Lift ofdP/dt (P=URU0solution the above Riccati equation):
d
dtU= (In−UU0)AU, d
dtR=U0AUR+RU0AU−RU0C(HH0)−1CUR
4S. Bonnabel and R. Sepulchre. The geometry of low-rank Kalman filters.
preprint arXiv:1203.4049v1, March 2012.
Projection and lift for rank-m density operators of
Cn×nThe sub-manifoldDmofdensity matricesρof rankm<nis over-parameterizedvia
ρ=UσU† !
ρ
z }| {
=
U
z}|{
σ
z}|{
U†
z }| {
whereσis am×mstrictly positive Hermitian matrix,Uan×m matrix withU†U=Im.
The family of liftsfordρ/dt =−i[H, ρ]−12(L†Lρ+ρL†L) +LρL†
d
dtU=−iAU+ (In−UU†) −i(H−A)−12L†L+LUσU†L†Uσ−1U† U, d
dtσ=−i[U†(H−A)U, σ]−12(U†L†LUσ+σU†L†LU) +U†LUσU†L†U +m1Tr (L†(In−UU†)L UσU†
Im.
wherethe gage degree of freedomAis any time varyingn×n Hermitian matrix.
The computation of the lifted dynamics
Tangent mapof the submersion:(U, σ)7→UσU†=ρ!
U
z}|{
σ
z}|{
U†
z }| {=
ρ
z }| {
with theinfinitesimal variationsδU=ıηUandδσ=ς: (η, ς)7→i[η, ρ] +UςU†
whereηis anyn×nHermitian matrix,ςis anym×mHermitian matrix with zero trace.
An×nHermitian matrixξin the tangent space atρ=UσU†toDm admits the parameterizationξ=i[η, ρ] +UςU†.
The projectionΠρm(dtdρ)corresponds to thetangent vectorξ associated toηandς minimizing
Tr
−i[H, ρ]−(L†Lρ+ρL†L)/2+LρL†−i[η, ρ]−UςU†2 ,
First order stationary conditionsgiveηandς as function ofρ=UσU†: the lifted evolution is given bydtdU=iηUand dtdσ=ς where the arbitrary matrixAappears.
Gage
A=Hadapted to weak dissipation
In d
dtU=−iAU+ (In−UU†) −i(H−A)−12L†L+LUσU†L†Uσ−1U† U, d
dtσ=−i[U†(H−A)U, σ]−12(U†L†LUσ+σU†L†LU) +U†LUσU†L†U +m1Tr (L†(In−UU†)L UσU†
Im.
setA=H:
d
dtU=−iHU+ (In−UU†) −12L†L+LUσU†L†Uσ−1U† U, d
dtσ=−12(U†L†LUσ+σU†L†LU) +U†LUσU†L†U +m1Tr (L†(In−UU†)L UσU†
Im,
Honly appears in the dynamics ofUand not in the dynamics ofσ.
Appropriate whenHdominatesL: aslow evolution ofσas compared to afast evolution ofU(important for the numerical procedure)
A numerical integration scheme adapted to weak dissipation
Uk andσk the numerical approximations ofU(kδt)andσ(kδt).The update from timekδtto time(k+1)δtis split into3 steps forU and2 steps forσ
Uk+1
3 =
In−iδt2H−δt82H2+iδt483H3 Uk
Uk+2
3 =Uk+1
3 +δt(In−Uk+1
3Uk+† 1 3
)
−12L†LUk+1
3 +LUk+1
3σkUk+† 1 3
L†Uk+1
3σk−1 Uk+1= Υ In−i2δtH−δt82H2+iδt483H3
Uk+2 3
(Υortho-normalization)
σk+1
2 =σk +δt Uk+† 1 3
LUk+1 3σkUk+† 1
3
L†Uk+1 3
+δt Tr
(U†
k+13L†LUk+1 3 −U†
k+13L†Uk+1 3U†
k+13LUk+1 3)σk
Im
m σk+1=
(Im−δt2Uk+† 1 3
L†LUk+1 3)σk+1
2(Im−δt2Uk+† 1 3
L†LUk+1 3) Tr
(Im−δt2U†
k+13L†LUk+1 3)σk+1
2(Im−δt2U†
k+13L†LUk+1 3). This scheme preservesU†U=Im,σ†=σ,
σ > 0
and Tr(σ) =1.Computational cost versus QMC procedure
d|ψti= −iH−12L†L+hψt|L†L|ψti
|ψtidt +
L|ψti
√
hψt|L†L|ψti− |ψti
dNt
Uk+1
3 =
In−iδt2H−δt82H2+iδt483H3 Uk
Uk+2 3 =Uk+1
3 +δt(In−Uk+1 3U†
k+13)
−12L†LUk+1
3 +LUk+1 3σkU†
k+13L†Uk+1 3σk−1 Uk+1= Υ In−iδt2H−δt82H2+iδt483H3
Uk+2
3
Both methods useessentially right multiplications ofH,L,L†byn×1 orn×mmatrices, as, for example, the productsH|ψi,L|ψiorHU, LU,L†(LU). No stringn×nmatrices sinceHandLare defined as tensor products of operators of small dimensions. Whennis very large andmis small, this point is crucial for an efficient numerical implementation:evaluations of products likeHUorLUcan be parallelized.
Empirical estimation
5of truncation error
I
Based on
Frobenius norms ofρ˙= dtdρandρ˙⊥ = ˙ρ−Πρm( ˙ρ)for
ρ=UσU†using:
˙
ρ=−i[H, ρ]−12(L†Lρ+ρL†L) +LρL†
˙
ρ⊥= (In−Pρ)LρL†(In−Pρ)−Tr(LρL†(In−Pρ))
m Pρ
where
Pρ=UU†.
I
Good approximation when Tr
ρ˙2⊥Tr
ρ˙2.
I
At each time step, Tr
ρ˙2and Tr
ρ˙2⊥may be numerically evaluated with a complexity similar to the complexity of the numerical scheme (no need to explicitly compute
ρ˙and
ρ˙⊥as
n×nmatrices before taking their Frobenius norms).
5Inspired from R. van Handel and H. Mabuchi. Quantum projection filter for a highly nonlinear model in cavity qed.Journal of Optics B: Quantum and Semiclassical Optics
Initialization procedure
σ0
and
U0need to be deduced from a given initial condition
ρ0:
I When the rank ofρ0≥m:σ0
diagonal matrix made of the largest
meigenvalues of
ρ0with sum normalized to one;
U0
the associated normalized eigenvectors.
I
When the rank of
ρ0=1 and
m>1:
ρ0=|ψ0ihψ0|. It isthen natural to take for
σ0a diagonal matrix where the first diagonal element is 1
−(m−1) and the over ones are equal to 1. Then
U0is constructed, up to an
ortho-normalization preserving the first column, with
|ψ0ias the first column,
H|ψ0ias the second column, . . . ,
Hm−1|ψ0ias the last column.
I
When the rank of
ρ0in
]1,m[: combine the aboveinitialization scheme . . .
Lindblad equation of oscillation revivals
Thecollective symmetric behavior ofNatwo-level atomsresonantly interacting with aquantized field:
d
dtρ= Ω0
2 [a†J−−aJ+, ρ]−κ(nρ/2+ρn/2−aρa†) Preliminary tests viatwo different type of simulationsincluding the first complete revival:
I Na=1 atom initially in the excited state, a field initially in a coherent state with¯n=15 photons (truncation to 30 photons):
comparisons between the full-rank and rank-2-4-6 solutionswith κ= Ω0/500:
I Na=50 atoms all initially in excited states, a field with¯n=200 (truncation to 300 photons):comparison of the analytic
approximate weak-damping modelproposed in6(predicts a reduction of a factorr =2 of the complete first revival between κ=0 andκ=log(r)Ω0/(4π¯n3/2)) withthe rank-8 approximation given by the above integration scheme withδt =1/(Ω0
√nN¯ a).
6T. Meunier, A. Le Diffon, C. Ruef, P. Degiovanni, and J.-M. Raimond.
Entanglement and decoherence of N atoms and a mesoscopic field in a cavity.Phys. Rev. A, 74:033802, 2006.
Full rank (left) versus rank 2 (right) (N
a=1,
¯n=15,
φ= Ω0t/2√¯n)
0 2 4 6 8 10
0 0.5 1
population
(a)
0 2 4 6 8 10
0 0.5 1
low rank population
(b)
0 2 4 6 8 10
0 0.5 1
eigenvalues
(c)
0 2 4 6 8 10
0 0.5 1
low rank eigenvalues
(d)
0 2 4 6 8 10
0.9 0.95 1
fidelity
φ (e)
0 2 4 6 8 10
10−6 10−4 10−2 100 102
(f)
φ
Frobenius norm
Full rank (left) versus rank 4 (right) (N
a=1,
¯n=15,
φ= Ω0t/2√¯n)
0 2 4 6 8 10
0 0.5 1
population
(a)
0 2 4 6 8 10
0 0.5 1
low rank population
(b)
0 2 4 6 8 10
0 0.5 1
eigenvalues
(c)
0 2 4 6 8 10
0 0.5 1
low rank eigenvalues
(d)
0 2 4 6 8 10
0.98 0.99 1
fidelity
φ (e)
0 2 4 6 8 10
10−6 10−4 10−2 100 102
(f)
φ
Frobenius norm
Full rank (left) versus rank 6 (right) (N
a=1,
¯n=15,
φ= Ω0t/2√¯n)
0 2 4 6 8 10
0 0.5 1
population
(a)
0 2 4 6 8 10
0 0.5 1
low rank population
(b)
0 2 4 6 8 10
0 0.5 1
eigenvalues
(c)
0 2 4 6 8 10
0 0.5 1
low rank eigenvalues
(d)
0 2 4 6 8 10
0.996 0.998 1
fidelity
φ (e)
0 2 4 6 8 10
10−6 10−4 10−2 100 102
(f)
φ
Frobenius norm
Oscillation revival with
κ=0 (N
a=50,
n¯=200,
φ= Ω0t/2√ n)¯0 1 2 3 4 5 6 7 8
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
population
φ
6.2 6.4 6.6 6.8 7
0 0.05 0.1 0.15 0.2 0.25 0.3
Schrödinger simulation time 1h15 (Dell precision M4440 with Matlab)
Rank-8 solution with
κ=log(2)Ω
0/(4πn¯3/2)(N
a=50,
¯n=200)
0 1 2 3 4 5 6 7 8
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
population
φ
6.2 6.4 6.6 6.8 7
0 0.05 0.1 0.15 0.2 0.25 0.3
Rank-8 simulation time 17h00 (Dell precision M4440 with Matlab)
Concluding remarks
A single tuning parameter: the rankmn.
Extension to an arbitrary number of Lindblad operators:
d
dtρ=−i[H, ρ] +X
ν
LνρL†ν−12(L†νLνρ+ρL†νLν) d
dtU=−iHU+ (In−UU†) X
ν
−1
2L†νLν+LνUσU†L†νUσ−1U†
! U d
dtσ=X
ν
−1
2 (U†L†νLνUσ+σU†L†νLνU) +U†LνUσU†L†νU +m1Tr X
ν
(L†ν(In−UU†)LνUσU†
! Im.
Similar low-rank approximations could be done forcontinuous-time quantum filters. . .
Implemented in simulation packages such as QuTip7?
Adaptation whennis huge8?Low-rank quantum tomography ?
7J.R Johansson, P.D. Nation, F.Nori: QuTiP an open-source Python framework for dynamics of open quantum systems. Computers Physics Communications 183 (2012) 1760–1772.
8Ilya Kuprov: Spinach - software library for spin dynamics simulation of large spin systems. PRACQSYS 2010.