Design and Stability of Quantum Filters with Measurement Imperfections:
discrete-time and continuous-time cases
Pierre Rouchon
Mines-ParisTech,Centre Automatique et Systèmes Mathématiques et Systèmes
pierre.rouchon@mines-paristech.fr
Yale, April 26, 2013.
Based on collaborations with
Hadis Amini, Michel Brune, Igor Dotsenko, Serge Haroche, Zaki Leghtas, Mazyar Mirrahimi, Jean-Michel Raimond,
Clément Sayrin and Ram Somaraju
Quantum filtering: some references . . .
I Quantum stochastic equations: R. L. Hudson and K. R.
Parthasarathy. Quantum Itô’s formula and stochastic evolutions.
Commun. Math. Phys., 93:301-323, 1984.
I Infinite dimensional analysis, noncommutative probability, quantum information: V. Belavkin. Quantum stochastic calculus and quantum nonlinear filtering. Journal of Multivariate Analysis, 1992, 42, 171-201.
I Control theory: Bouten, L.; R. van Handel & James, M. R. An introduction to quantum filtering. SIAM J. Control Optim., 2007, 46, 2199-224.
I Discrete-time approximation: Bouten, L. & Van Handel, R.
Discrete approximation of quantum stochastic models. J. Math.
Phys., AIP, 2008, 49, 102109-19.
I Quantum Monte Carlo trajectories:Dalibard, J.; Castin, Y. &
Mølmer, K. Wave-function approach to dissipative processes in quantum optics. Phys. Rev. Lett., 1992, 68, 580-583.
The first experimental realization of a quantum
state feedback21
The LKB photon box: sampling time (∼100µs) long enough to estimate in real-time the quantum-stateρand to compute the control u=AeıΦas a function ofρ(quantum state feedback).
1Courtesy of Igor Dotsenko
2C. Sayrin et al., Nature, 1-September 2011
Experimental data
Anopen-loop trajectory starting from coherent state with an average of 3 photons relaxes towards vacuum (decoherence due to finite photon life time around 70 ms)Detection efficiency 40%
Detection error rate 10%
Delay 4 sampling periods
The quantum filter takes into account cavity decoherence, measure imperfections and delays (Bayes law).
Truncation to 9 photons
Stabilization around 3-photon state
Several "quantum states":
|ψki,ρbkand
ρbestk.
QuantumFilter_Controller PhotonBox
u
y y
Detector Coherent
pulse
u
est est
Thestate estimationρbestk used in the feedback law takes into account, measure imperfections, delays and cavity decoherence:
I Derived from Bayes law: depends on past detector outcomes between 0 andk; computed recursively from an initial valueρbest0 ;
I Stable and tends to converge towardsρbk,the expectation value ofρk =|ψkihψk|knowing its initial valueρ0=ρb0and the past detector outcomes between 0 andk.
Outline
Quantum filter of the LKB Photon Box Markov chain in the ideal case
The Markov chain with detection errors The Markov chain with cavity decoherence Structure of the complete quantum filter Discrete-time quantum systems
Markov chains in the ideal case
Markov chains with imperfections and decoherence Stability and convergence issues
Bayesian parameter estimations Continuous-time quantum filters
From discrete-time to continuous-time models
SDE driven by Poisson and Wiener processes
SDE with imperfections and decoherence
Conclusion
Markov chain in the ideal case (1)
I System
S corresponds to a quantized cavity mode:
HS
=
( ∞X
n=0
ψn|ni |
(ψ
n)
∞n=0∈l
2(
C)
),
where
|nirepresents the Fock state associated to exactly n photons inside the cavity
I Meter
M is associated to atoms :
HM=
C2, each atom admits two energy levels and is described by a wave function c
g|gi+ c
e|eiwith
|cg|2+
|ce|2= 1; atoms leaving B are all in state
|giI
When an atom comes out B, the state
|ΨiB ∈ HS⊗ HMof the composite system atom/field is
separable|ΨiB
=
|ψi ⊗ |gi.Markov chain in the ideal case (2) C
B
D R
1R
2I When an atom comes outB: |ΨiB =|ψi ⊗ |gi.
I Just before the measurement inD, the state is in general entangled(not separable):
|ΨiR2 =USM |ψi ⊗ |gi
= Mg|ψi
⊗ |gi+ Me|ψi
⊗ |ei whereUSMis the total unitary transformation (Schrödinger propagator) defining the linear measurement operatorsMgand MeonHS. SinceUSMis unitary,Mg†Mg+Me†Me=I.
Markov chain in the ideal case (3)
Just before the measurement inD, the atom/field state is:
Mg|ψi ⊗ |gi+Me|ψi ⊗ |ei
Denote byµ∈ {g,e}the measurement outcome in detectorD: with probabilitypµ=hψ|Mµ†Mµ|ψiwe getµ. Just after the measurement outcomeµ,the state becomes separable:
|ΨiD= √1p
µ (Mµ|ψi)⊗ |µi= (Mµ|ψi)⊗ |µi q
hψ|Mµ†Mµ|ψi .
Markov process(density matrix formulationρ∼ |ψihψ|)
ρ+=
Mg(ρ) = MgρM
† g
Tr(MgρMg†), with probabilitypg = Tr
MgρMg†
; Me(ρ) = MeρMe†
Tr(MeρMe†), with probabilitype= Tr
MeρMe† .
Kraus map:
E
(ρ+/ρ) =K(ρ) =MgρMg†+MeρMe†.Markov chain with detection errors (1)
I ρ+= 1
Tr(MµρMµ†)MµρMµ† when the atom collapses inµ=g,e.
This happens with probability Tr MµρMµ† .
I Detection error rates: P(y =e/µ=g) =ηg∈[0,1]the probability of erroneous assignation toewhen the atom collapses ing;P(y =g/µ=e) =ηe∈[0,1](given by the contrast of the Ramsey fringes).
Bayes law gives the probability that the atom collapses inµ=g knowing the detector outcomey =g:
P(µ=g/y =g) =
(1−ηg)Tr
MgρMg† (1−ηg)Tr
MgρMg†
+ηeTr
MeρMe†
sinceP(y =g/µ=g) = (1−ηg)andP(y =g/µ=e) =ηe.
Markov chain with detection errors (2)
The expectation valueρb+ofρ+knowingρand the imperfect detection y =g is given by
ρb+=P(µ=g/y =g) MgρM
† g
Tr(MgρM†g)+P(µ=e/y =g) MeρMe†
Tr(MeρM†e) Since
P(µ=g/y =g) = (1−ηg)Tr(MgρMg†)
(1−ηg)Tr(MgρM†g)+ηeTr(MeρMe†)
andP(µ=e/y =g) =1−P(µ=g/y =g), this expectation value ρb+is given by
ρb+= 1
Tr((1−ηg)MgρMg†+ηeMeρMe†)
(1−ηg)MgρMg†+ηeMeρMe†
Similarly wheny =e, the expectation valueρb+is given by
ρb+= 1
Tr((1−ηe)MeρMe†+ηgMgρMg†)
(1−ηe)MeρMe†+ηgMgρMg†
Markov chain with detection errors (3)
We getρb+=
(1−ηg)MgρM†g+ηeMeρMe†
Tr((1−ηg)MgρMg†+ηeMeρMe†), with prob. Tr
(1−ηg)MgρMg†+ηeMeρMe†
;
ηgMgρMg†+(1−ηe)MeρMe†
Tr(ηgMgρM†g+(1−ηe)MeρMe†) with prob. Tr
ηgMgρMg†+ (1−ηe)MeρMe† .
Key point:
Tr
(1−ηg)MgρMg†+ηeMeρMe†
and Tr
ηgMgρMg†+ (1−ηe)MeρMe†
are the probabilities to detecty =g ande, knowingρ.
Withηµ0,µbeing the probability to detecty =µ0knowing that the atom collapses inµ, we have
ρb+= P
µηµ0,µMµρMµ† Tr
P
µηµ0,µMµρMµ†
when we detecty =µ0.
The probability to detecty =µ0knowingρis Tr P
µηµ0,µMµρMµ† .
The Markov chain with cavity decoherence
When the sampling time∆T is much smaller than the photon life time Tcav, cavity decoherence (at zero temperature) can be described approximatively by the Kraus map
ρ7→M0ρM0†+M−ρM−† withM0= (1−2T∆T
cav)I−2T∆T
cavNandM−= q∆T
Tcava
M0andM−can be seen as "measurement" operators corresponding to information catched by the "environment", information unknown in the real life but known in "Matlab/Simulink world":
I M0corresponds to no photon destruction during the sampling interval∆T; probability Tr
M0ρM0† .
I M−corresponds to one photon destruction during the sampling interval∆T; probability Tr
M−ρM−† .
The fact that we do not have access to this information can be interpreted as a detection error of 50%forM0andM−. We get
ρb+=M0ρM0†+M−ρM−†.
Photon-box quantum filter parameterized by left stochastic matrix
ηµ0,µ 3 ρbestk+1=
1Tr P
µηµ0,µMµρbestk Mµ†
P
µηµ0,µ
M
µρbestkM
µ†
with
I
we have a total of m = 3
×7 = 21 Kraus operators M
µ. The "jumps" are labeled by
µ= (µ
a, µc) with
µa∈ {no,
g, e, gg, ge, eg, ee} labeling atom related jumps and
µc ∈ {o,+,−}cavity decoherence jumps.I
we have only m
0= 6 real detection possibilities
µ0 ∈ {no,
g, e, gg, ge, ee} corresponding respectively to no detection, a single detection in g, a single detection in e, a double detection both in g, a double detection one in g and the other in e, and a double detection both in e.
µ0\µ (no, µc) (g, µc) (e, µc) (gg, µc) (ee, µc) (ge, µc)or(eg, µc)
no 1 1−d 1−d (1−d)2 (1−d)2 (1−d)2
g 0 d(1−ηg) dηe 2d(1−d)(1−ηg) 2d(1−d)ηe d(1−d)(1−ηg+ηe) e 0 dηg d(1−ηe) 2d(1−d)ηg 2d(1−d)(1−ηe) d(1−d)(1−ηe+ηg)
gg 0 0 0 2d(1−ηg)2 2dη2e 2dηe(1−ηg)
ge 0 0 0 22dηg(1−ηg) 22dηe(1−ηe) 2d((1−ηg)(1−ηe) +ηgηe)
ee 0 0 0 2dη2g 2d(1−ηe)2 2dηg(1−ηe)
3Somaraju, A.; Dotsenko, I.; Sayrin, C. & PR. Design and Stability of Discrete-Time Quantum Filters with Measurement Imperfections. American Control Conference, 2012, 5084-5089.
Markov chain in ideal life (e.g. Matlab/Simulink world): pure state
ρkρk+1
=
Mµk(ρ
k) =: M
µkρkM
µ†kTr
M
µkρkM
µ†kI To each measurement outcomeµis attached the Kraus operator Mµdepending onµand also on time (not explicitly recalled here, Mµ=Mµ,k could depend onk).
I µk is a random variable taking valuesµin{1,· · ·,m}with probabilitypµ,ρk = Tr MµρkMµ†
. Conservation of probability (P
µpµ,ρ=1 for allρ) is guarantied byPm
µ=1Mµ†Mµ=I.
I The Kraus mapK(ρ) =Pm
µ=1MµρMµ† provides
E
(ρk+1/ρk) =K(ρk)The Markov chain in real life: mixed states,
ρbkand
ρbestk(1)
4 Takeρk+1= 1Tr(MµkρkMµ†k)
MµkρkMµ†k
with measure imperfections and decoherence described by theleft stochastic matrixη:
ηµ0,µ∈ |0,1]is the probability of having the imperfect outcome µ0∈ {1, . . . ,m0}knowing that the perfect one isµ∈ {1, . . . ,m}.
I ρbk =
E
ρk|ρ0, µ00, . . . , µ0k−1can be computed efficiently via the following recurrence
ρbk+1= 1
TrPm µ=1ηµ0
k,µMµρbkMµ†
m
X
µ=1
ηµ0
k,µMµρbkMµ†
where the detector outcomeµ0k takes valuesµ0 in{1,· · ·,m0} with probabilitypµ0,ρbk = Tr
Pm µ=1ηµ0
k,µMµρbkMµ† .
I Thus
E
(ρbk+1|ρbk) =K(ρbk) =Pmµ=1MµρbkMµ†.4Somaraju, A.; Dotsenko, I.; Sayrin, C. & PR. Design and Stability of Discrete-Time Quantum Filters with Measurement Imperfections. American Control Conference, 2012, 5084-5089.
The Markov chain in real life: mixed states,
ρbkand
ρbestk(2)
ρbk =E
ρk|ρ0, µ00, . . . , µ0k−1is given by
ρbk+1= 1
TrPm µ=1ηµ0
k,µMµρbkMµ†
m
X
µ=1
ηµ0
k,µMµρbkMµ†
with theperfect initialization:ρb0=ρ0. ρbestk+1= 1
Tr Pm
µ=1ηµ0
k,µMµρbestk Mµ†
Pm µ=1ηµ0
k,µMµρbestk Mµ†
but with imperfect initializationρbest0 6=ρ0.
This filtering process is stable5: the fidelityF(ρbk,ρbestk )is a sub-martingale for anyηandMµ:
E
F(ρbk+1,ρbestk+1)/ρbk≥F(ρbk,ρbestk )
Convergence ofρbestk towardsρbk whenk 7→+∞is an open problem.6
5PR. Fidelity is a Sub-Martingale for Discrete-Time Quantum Filters IEEE Transactions on Automatic Control, 2011, 56, 2743-2747 .
6A partial result (continuous-time): R. van Handel. The stability of quantum Markov filters. Infin. Dimens. Anal. Quantum Probab. Relat. Top. , 2009, 12, 153-172.
Bayesian parameter estimations
Consider detector outcomesµ0k corresponding to a parameter valuep¯ poorly known. Assume to simplify that eitherp¯=aor¯p=b, with a6=b. We can discriminate betweenaandband recover¯pvia the following Bayesian scheme using information contained in theµ0k’s:
ρbesta,k+1=
P
µηaµ0
k,µMµaρbesta,kMµa† Tr
P
p
P
µηp
µ0
k,µMµpρbestp,kMµp
†, ρbestb,k+1=
P
µηbµ0
k,µMbµρbestb,kMbµ† Tr
P
p
P
µηp
µ0
k,µMµpρbestp,kMµp
†
with initializationρbesta,k+1=ρbestb,k+1=ρbest0 /2 whereρbest0 is some guess ofρb0assuming initial probability of 12 to havep¯=aand¯p=b.
This estimation/filtering process is also stable:
•F(ρbk,ρbesta,k) +F(ρbk,ρbestb,k)is a sub-martingale
• Tr ρbesta,k
, Tr ρbestb,k
) estim. of proba. to havep¯=a,¯p=b.
Direct generalization to a continuumof choices forp¯∈[pmin,pmax] (see7for a first experimental use)
7Brakhane, S.; Alt, W.; Kampschulte, T.; Martinez-Dorantes, M.; Reimann, R.; Yoon, S.; Widera, A. & Meschede, D. Bayesian Feedback Control of a Two-Atom Spin-State in an Atom-Cavity System. Phys. Rev. Lett., 2012, 109, 173601-
Dynamical models with a precise structure
Discrete-time modelsareMarkov chainsρk+1= 1
pµ(ρk)MµρkMµ† with proba. pµ(ρk) = Tr MµρkMµ† associated toKraus maps(ensemble average, open quantum channels)
E
(ρk+1/ρk) =K(ρk) =Xµ
MµρkMµ† with X
µ
Mµ†Mµ=I
Continuous-time modelsarestochastic differential systems dρ=
−i[H, ρ] +LρL†−1
2(L†Lρ+ρL†L) dt +
Lρ+ρL†− Tr (L+L†)ρ ρ
dw driven byWiener processes8dw =dy− Tr (L+L†)ρ
dt with measurey and associated toLindbaldmaster equations:
d dtρ=−i
~[H, ρ] +LρL†−1
2(L†Lρ+ρL†L)
8Another possibility: SDE driven by Poisson processes.
From discrete-time to continuous-time: heuristic connection For Monte-Carlo simulations of
d
ρ=
−i[H, ρ] +
LρL
†−12
(L
†Lρ +
ρL†L)
dt +
Lρ +
ρL†−Tr
(L + L
†)ρ
ρdw
take a small sampling time dt, generate a random number
dwtaccording to a Gaussian law of
standard deviation√dt, and set ρt+dt
=
Mdyt(ρ
t) where the jump operator
Mdytis labelled by the measurement outcome
dyt= Tr (L + L
†)
ρtdt +
dwt:
Mdyt
(ρ
t) =
I+(−iH−1
2L†L)dt+dytL
ρt I+(iH−12L†L)dt+dytL†
Tr
I+(−iH−12L†L)dt+dytL
ρt I+(iH−12L†L)dt+dytL†.
Then
ρt+dtremains always a density operator and using the
Itô rules(dw of order
√dt and
dw2≡dt ) we get the good
dρ =
ρt+dt −ρtup to O((dt)
3/2) terms.
From discrete-time to continuous-time: heuristic connection (end)
For the Lindblad equation d
dt
ρ=
−i~
[H, ρ] + LρL
†−12
(L
†Lρ +
ρL†L) take a small sampling time
dtand set
ρt+dt
= I +
dt(−iH
−12L
†L)
ρt
I +
dt(iH
−12L
†L)
+
dtLρ
tL
†Tr I +
dt(−iH
− 12L
†L)
ρt
I +
dt(iH
−12L
†L)
+
dtLρ
tL
†.Then
ρt+dtremains always a density operator and
d
dtρ
= (ρ
t+dt−ρt)/dt up to O(dt ) terms.
SDE driven by Poisson and/or Wiener processes
dρ
t=
L(ρt) dt+
mw
X
ν=1
Λ
ν(ρ
t) dw
tν+
mP
X
µ=1
Υ
µ(ρ
t)
dN
tµ−Tr
C
µρtC
µ†dt
,
where
I L(ρt
) :=
−i[H, ρ
t] +
PmPµ=1LPµ
(ρ
t) +
Pmwν=1Lwν
(ρ
t),
LPµ
(ρ) :=
−12{Cµ†C
µ, ρ}+ C
µρCµ†, Lwν(ρ) :=
−12{L†νL
ν, ρ}+ L
νρL†ν; Υ
µ(ρ) :=
CµρC† µ
Tr
CµρCµ†−ρ,
Λ
ν(ρ) := L
νρ+
ρL†ν −Tr
(L
ν+ L
†ν)ρ
ρI
Detector click no
µis related to the Poisson process dN
tµ= N
µ(t + dt )
−N
µ(t) = 1 and happens with probability Tr
C
µρtC
µ†dt;
I
Continuous detector y
tνis related to the Wiener process dw
tνby dy
tν= dw
tν+ Tr
(L
ν+ L
†ν)ρ
t
dt.
Quantum filter in the ideal case
dρt =L(ρt)dt+
mw
X
ν=1
Λν(ρt)dwtν+
mP
X
µ=1
Υµ(ρt) dNtµ− Tr CµρtCµ† dt
,
and the associated quantum filter
dρbestt =L(ρbestt )dt+
mw
X
ν=1
Λν(ρbestt ) dytν− Tr (Lν+L†ν)ρbestt dt
+
mP
X
µ=1
Υµ(ρbestt ) dNtµ− Tr Cµρbestt Cµ† dt
.
It can be rewritten as follows
dρbestt =L(ρbestt )dt+
mw
X
ν=1
Λν(ρbestt ) Tr (Lν+L†ν)ρt
−Tr (Lν+L†ν)ρbestt dt
+
mw
X
ν=1
Λν(ρbestt )dwtν+
mP
X
µ=1
Υµ(ρbestt ) dNtµ− Tr Cµρbestt Cµ† dt
.
Quantum filters with imperfections and decoherence
9(1)
I
Imperfection model for the Poisson processes dN
tµ:
I real outcomesµ0∈ {0,1, . . . ,m0P}
I ideal outcomesµ∈ {0,1, . . . ,mP}.
I (mP0 +1)×mPleft stochastic matrix ηP = (ηPµ0,µ)0≤µ0≤m0P,1≤µ≤mP
I positive vectorη¯P= (¯ηµP0)1≤µ0≤m0
P inRm
0 P
+ .
I
Imperfection model for the diffusion processes dw
tν:
I mw0 real continuous signalsytν0 withν0∈ {1, . . . ,mw0 },
I mw ideal continuous signalsytνwithν∈ {1, . . . ,mw}
I correlationmw0 ×mw matrixηw= (ηνw0,ν)1≤ν0≤m0w,1≤ν≤mw, with 0≤ηνw0,ν≤1 andPmrw
ν0=1ηνw0,ν≤1.
9see last chapter of H. Amini. Stabilization of discrete-time quantum systems and stability of continuous-time quantum filters. PhD thesis, Mines-ParisTech, September 2012.
Quantum filters with imperfections and decoherence (2)
dρbt = L(ρbt)dt+
m0w
X
ν0=1
pη¯wν0Λbν0(ρbt)dwbtν0
+
m0P
X
µ0=1
Υbµ0(ρbt)
dNbtµ0−η¯µP0dt−
mP
X
µ=1
ηPµ0,µTr CµρbtCµ† dt
I η¯wν0 =Pmw
ν=1ηwν0,ν ,Υbµ0(ρ) := η¯
P µ0ρ+PmP
µ=1ηµP0,µCµρC†µ
¯ ηP
µ0+PmP
µ=1ηP
µ0,µTr(CµρCµ†)−ρ, bΛν0(ρ) =bLν0ρ+ρbL†ν0− Tr
(bLν0+bL†ν0)ρ
ρ, bLν0 := (Pmw
ν=1ηνw0,νLν)/η¯νw0;
I the jump detectorµ0corresponds toNbµ0(t):
dNbtµ0 =Nbµ0(t+dt)−Nbµ0(t) =1 happens with probability Pbµ0(ρbt) = ¯ηPµ0dt+PmP
µ=1ηµP0,µTr CµρbtCµ† dt;
I the continuous detectorν0refers tobytν0 anddwbtν0: dbytν0 =dwbtν0+p
¯ ηνw0 Tr
(bLν0+bL†ν0)ρbt dt.
Quantum filters with imperfections and decoherence (3)
dρbt = L(ρbt)dt+
m0w
X
ν0=1
pη¯wν0Λbν0(ρbt)dwbtν0
+
m0P
X
µ0=1
Υbµ0(ρbt)
dNbtµ0−η¯µP0dt−
mP
X
µ=1
ηPµ0,µTr CµρbtCµ† dt
and the associated quantum filter
dρbestt = L(ρbestt )dt+
m0w
X
ν0=1
pη¯νw0bΛν0ρbestt )
dybtν0−p
¯ ηwν0 Tr
(bLν0+bL†ν0)ρbestt dt
+
mP0
X
µ0=1
Υbµ0(ρbestt )
dNbtµ0−η¯Pµ0dt −
mP
X
µ=1
ηµP0,µTr Cµρbestt Cµ† dt
Quantum filtering combines the following key points
1. Bayes law: P(µ0/µ) =P(µ/µ0)P(µ0)/ P
ν0P(µ/ν0)P(ν0) . 2. Schrödinger equationsdefining unitary transformations.
3. Partial collapse of the wave packet:irreversibility and
convergence are induced by the measure of observablesOwith degenerate spectra,O=P
µλµPµ:
I measure outcomeλµ with proba.pµ=hψ|Pµ|ψi= Tr(ρPµ) depending|ψi,ρjust before the measurement
I measure back-action if outcomeµ:
|ψi 7→ |ψi+ = Pµ|ψi
phψ|Pµ|ψi, ρ7→ρ+ = PµρPµ
Tr(ρPµ) 4. Tensor product for the description of composite systems(S,M):
I Hilbert spaceH=HS⊗ HM
I HamiltonianH=HS⊗IM+Hint+IS⊗HM I observable on sub-systemM only:O=IS⊗ OM.