See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/49751450

Quantum State Tomography via Compressed Sensing

Article  in  Physical Review Letters · October 2010

DOI: 10.1103/PhysRevLett.105.150401 · Source: PubMed

CITATIONS READS
665 205

5 authors, including:

Yi-Kai Liu Stephen R. Becker

National Institute of Standards and Technology University of Colorado Boulder
31 PUBLICATIONS   2,036 CITATIONS    65 PUBLICATIONS   3,272 CITATIONS   

SEE PROFILE SEE PROFILE

Jens Eisert
Freie Universität Berlin
366 PUBLICATIONS   20,437 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

quantum thermodynamics View project

Multi-Document Summarization View project

All content following this page was uploaded by Jens Eisert on 31 May 2014.

The user has requested enhancement of the downloaded file.

 week ending
PRL 105, 150401 (2010) PHYSICAL REVIEW LETTERS 8 OCTOBER 2010

Quantum State Tomography via Compressed Sensing

David Gross,1 Yi-Kai Liu,2 Steven T. Flammia,3 Stephen Becker,4 and Jens Eisert5
1
Institute for Theoretical Physics, Leibniz University Hannover, 30167 Hannover, Germany
2
Institute for Quantum Information, California Institute of Technology, Pasadena, California, USA
3
Perimeter Institute for Theoretical Physics, Waterloo, Ontario, N2L 2Y5 Canada
4
Applied and Computational Mathematics, California Institute of Technology, Pasadena, California, USA
5
Institute of Physics und Astronomy, University of Potsdam, 14476 Potsdam, Germany
(Received 21 October 2009; published 4 October 2010)
We establish methods for quantum state tomography based on compressed sensing. These methods are
specialized for quantum states that are fairly pure, and they offer a significant performance improvement
on large quantum systems. In particular, they are able to reconstruct an unknown density matrix of
dimension d and rank r using Oðrdlog2 dÞ measurement settings, compared to standard methods that
require d2 settings. Our methods have several features that make them amenable to experimental
implementation: they require only simple Pauli measurements, use fast convex optimization, are stable
against noise, and can be applied to states that are only approximately low rank. The acquired data can be
used to certify that the state is indeed close to pure, so no a priori assumptions are needed.

DOI: 10.1103/PhysRevLett.105.150401 PACS numbers: 03.65.Wj, 03.67.
a

The tasks of reconstructing the quantum states and minimum-rank matrix subject to linear constraints is NP-
processes produced by physical systems—known respec- hard in general [8].
tively as quantum state and process tomography [1]—are In addition to a reduction in experimental complexity,
of increasing importance in physics and especially in one might hope that a postprocessing algorithm taking as
quantum information science. Tomography has been used input only OðrdÞ  d2 numbers could be tuned to run
to characterize the quantum state of trapped ions [2] and an considerably faster than standard methods. Since the out-
optical entangling gate [3] among many other implemen- put of the procedure is a low-rank approximation to the
tations. But a fundamental difficulty in performing tomo- density operator and only requires OðrdÞ numbers be
graphy on many-body systems is the exponential growth in specified, it becomes conceivable that the run time scales
the state space dimension. For example, to get a maximum- better than Oðd2 Þ, clearly impossible for naive approaches
likelihood estimate of a quantum state of 8 ions, Ref. [2] using dense matrices.
required hundreds of thousands of measurements and In this Letter, we introduce a method to achieve such
weeks of postprocessing. drastic reductions in measurement complexity, together
Still, one might hope to overcome this obstacle because with efficient algorithms for postprocessing. The approach
the vast majority of quantum states are not of physical further develops ideas that have recently been studied under
interest. Rather, one is often interested in states with special the label of ‘‘compressed sensing.’’ Compressed sensing [9]
properties: pure states, states with particular symmetries, provides techniques for recovering a sparse vector from a
ground states of local Hamiltonians, etc., and tomography small number of measurements [10]. Here, sparsity means
might be more efficient in such special cases [4]. that this vector contains only a few nonzero entries in a
In particular, consider pure or nearly pure quantum specified basis, and the measurements are linear functions
states, i.e., states with low entropy. More precisely, con- of its entries. When the measurements are chosen at random
sider a quantum state that is essentially supported on an (in a certain precise sense), then with high probability two
r-dimensional space, meaning the density matrix is close surprising things happen: the vector is uniquely determined
(in a given norm) to a matrix of rank r, where r is small. by a small number of measurements, and it can be recovered
Such states arise in very common physical settings, e.g., a by an efficient convex optimization algorithm [9].
pure state subject to a local noise process [5]. Matrix completion [11–13] is a generalization of com-
A standard implementation of tomography [6,7] would pressed sensing from vectors to matrices. Here, one recov-
use d2 or more measurement settings, where d ¼ 2n for an ers certain ‘‘incoherent’’ low-rank matrices X from a small
n-qubit system. But a simple parameter counting argument number of matrix elements Xi;j . The problem of low-rank
suggests that OðrdÞ settings could possibly suffice—a quantum state tomography bears a strong resemblance to
significant improvement. However, it is not clear how to matrix completion. However, there are important differ-
achieve this performance in practice, i.e., how to choose ences. We wish to use measurements that can be more
these measurements, or how to efficiently reconstruct the easily implemented in an experiment than obtaining ele-
density matrix. For instance, the problem of finding a ments
i;j of density matrices. Previous results [11–13]

0031-9007=10=105(15)=150401(4) 150401-1 Ó 2010 The American Physical Society

 week ending
PRL 105, 150401 (2010) PHYSICAL REVIEW LETTERS 8 OCTOBER 2010

cannot be applied to this more general situation. We would the solution is unique if for all deviations
:¼ 


also like to avoid any unnatural incoherence assumptions away from
either R
Þ 0 or k
þ
ktr > k
ktr .
crucial in prior work [11]. We will ascertain this by using a basic idea from convex
Our first result is a protocol for tomography that over- optimization: constructing a strict subgradient Y for the
comes both of these difficulties: it uses Pauli measurements norm. A matrix Y is a strict subgradient if k
þ
ktr >
only, and it works for arbitrary density matrices. We prove k
ktr þ trY
for all
Þ 0. The main contribution below
that only Oðrdlog2 dÞ measurement settings suffice. What is a method for constructing such a Y which is also in the
is more, our proof introduces some new techniques, which range of R. For then R
¼ 0 implies that
is orthogonal
both generalize and vastly simplify the previous work on to the range of R, thus trY
¼ 0 and the subgradient
matrix completion. condition reads k
þ
ktr > k
ktr . This implies unique-
In a real experiment, the measurements are noisy, and ness. (In fact, it suffices to approximate the condition.)
the true state is only approximately low rank. We show that Let E be the projection onto the range of
, let T be the
our method is robust to these sources of error. We also space spanned by those operators whose row or column
describe ways to certify that a state is nearly pure without space is contained in range
. Let P T be the projection
any a priori assumptions. onto T, P ?T onto the orthogonal complement. Decompose
Finally, we present fast algorithms for reconstructing the
¼
T þ
? T , the parts of
that lie in the subspaces T
density matrix from the measurement statistics based on and T ? . We distinguish two cases: (i) k
T k2 > d2 k
? T k2 ,
semidefinite programming—a feature not present in earlier and (ii) k
T k2  d2 k
? k
T 2 [16].
methods for pure-state tomography [4,6,7]. Reconstructing Case (i) is easier. In this case,
is well approximated by
a low-rank density matrix for 8 qubits takes about 1 min on
T and essentially we only have to show that the restriction
an ordinary laptop computer. A :¼ P T RP T of R to T is invertible. Using a noncom-
While our methods do not overcome the exponential mutative large-deviation bound (see Refs. [18,19]),
growth in measurement complexity (which is provably 2
Pr½kA
1T jj > t < 4dre
t =8 (2)
impossible for any protocol capable of handling generic
pure states), they do significantly push the boundary of where  ¼ m=ðdrÞ [16]. Hence the probability that kA

what can be done in a realistic setting [14]. 1T k > 12 is smaller than 4dre
=32 ¼: p1 . If that is not the
Matrix recovery using Pauli measurements.—We con- case, one easily sees that kR
k2 > 0, concluding the
sider the case of n spin-1=2 systems in an unknown Nstate
proof for this case.
[15]. An n-qubit Pauli matrix is of the form w ¼ ni¼1 wi , Case (ii) is more involved. A matrix Y 2
where wi 2 f1; x ; y ; z g. There are d2 such matrices, spanðwðA1 Þ; . . . ; wðAm ÞÞ is an almost subgradient [20] if
labeled wðaÞ, a 2 ½1; d2 . The protocol proceeds as fol-
lows: choose m integers A1 ; . . . ; Am 2 ½1; d2  at random kP T Y
Ek2  1=ð2d2 Þ; kP ?
T Yk < 1=2: (3)
and measure the expectation values tr
wðAi Þ. One then First, suppose such a Y exists. Then a simple calculation
solves a convex optimization problem: minimize kktr (see Ref. [18]) using the condition (ii) shows that R
¼ 0
[16] subject to indeed implies k
þ
ktr > k
ktr as hinted at above. This
tr  ¼ 1; trwðAi Þ ¼ trwðAi Þ
: (1) proves uniqueness in case (ii). The difficult part consists in
showing that an almost subgradient exists.
Theorem 1 (low-rank tomography)—Let
be an arbi-
To this end, we design a recursive process (the ‘‘golfing
trary state of rank r. If m ¼ cdrlog2 d randomly chosen
scheme’’ [17]) which converges to a subgradient exponen-
Pauli expectations are known, then
can be uniquely
tially fast. Assume we draw l batches of 0 rd Pauli
reconstructed by solving the convex optimization problem
observables independently at random (0 will be chosen
(1) with probability of failure exponentially small in c.
later). Define recursively X0 ¼ E,
The proof is inspired by, but technically very different
from, earlier work on matrix completion [11]. Our methods Xi
Yi ¼ Rj Xj
1 ; Xi ¼ E
P T Yi ; (4)
are more general, can be tuned to give tighter bounds, and j¼1
are much more compact, allowing us to present a fairly
complete argument in this Letter. A more detailed presen- Y ¼ Yl . Let Ri be the sampling operator associated with
tation will be published elsewhere [17]. the ith batch, and Ai its restriction to T. Assume that in
Proof.—Here we sketch the argument and explain the each run kAi
1T k2 < 1=2. Denote the probability of
main ideas; detailed calculations are in the supplementary this event not occurring by p2 . Then
material [18]. kXi k2 ¼ kXi
1
P T Ri Xi
1 k2 ¼ kð1T
Ai ÞXi
1 k2
Note that the linear constraints (1) depend only on the
projection of
onto the span of the measured observables  1=2kXi
1 k2 ;
pffiffiffi
wðA1 Þ; . . . ; wðAm Þ. This is preciselyPthe range of the so that kXi k2  2
i kX0 k ¼ 2
i r. Hence,pYffiffiffi ¼ Yl fulfills
‘‘sampling operator’’ R:
° md m i¼1 wðAi Þtr
wðAi Þ. the first part of (3), as soon as l  log2 ð2d2 rÞ. We turn to
(Note that E½Rð
Þ ¼
.) Indeed, the convex program the second part. Again using large-deviation
pffiffiffi techniques
can be written as min kktr s.t. R ¼ R
. Evidently, [18] we find kP ? T Ri Xi
1 k  1=ð4 rÞkXi
1 k2 with some
150401-2
 week ending
PRL 105, 150401 (2010) PHYSICAL REVIEW LETTERS 8 OCTOBER 2010

(high) probability (1
p3 ). Therefore: expression in terms of Pauli expectation values. See

X Ref. [18] for details. We get:
l
1X 1
1
kP ?
T Yl k  kP ?
T Rj Xj
1 k  2
l < ; (5) Observation 2 (certified tomography).—Assume that the
j¼1 4 j¼0 2 unknown physical state is close to being pure. Then one
which is the second part of (3). can find a certificate for that assumption, and reconstruct
Lastly, we have to bound the total probability of failure the state with explicit guarantees on the reconstruction
pf  p1 þ p2 þ p3 . Set 0 ¼ 64½1 þ lnð8dlÞ, which error, from Oðcdlog2 dÞ Pauli expectation values. The
means that m ¼ drðlndÞ2 Oð1Þ coefficients will be sampled probability of failure is exponentially small in c.
in total. A simple calculation gives pf  e
 . h Finally, when the state is approximately low rank but not
In the remaining space, we address the important aspects nearly pure (r > 1), one may perform tomography using
of resilience against noise, certified tomography, and different numbers of random Pauli expectation values m.
numerical performance. Owing to space limitations, the When m is larger than necessary (corresponding to an over-
presentation will focus on conceptual issues, with the estimate of r), we are guaranteed to find the correct density
details in [21]. matrix. When m is too small, we find empirically that
Robustness to noise.—Realistic situations will differ the algorithms for reconstructing the density matrix fail to
from the previous case in two regards. First, the true state converge.

t may not be low rank, but only well approximated by a A hybrid approach to matrix recovery.—Here we de-
state
of rank r: k
t

k2  "1 . Second, due to system- scribe a variant of our tomography method that makes
atic and statistical noise, the available estimates for the the classical postprocessing step (i.e., solving the convex
Pauli expectations are not exactly tr
t wðaÞ, but of the form program (1) to reconstruct the density matrix) faster. This
tr!wðaÞ for some matrix !. Assume kR!
R
t k2  "2 method also uses random Pauli measurements, but they are
(in practical situations, "2 may be estimated from the error chosen in a structured
N way. Any Pauli matrix is of the
bars associated with the individual Pauli expectation val- form wðu; vÞ ¼ nk¼1 iuk vk ðx Þuk ðz Þvk for u, v 2 f0; 1gn .
ues). In order to get an estimate for
t , choose some   1 We choose a random subset S  f0; 1gn of size
pffiffiffiffiffiffiffiffiffiffiffiffi Oðrpoly logðdÞÞ, and then for all u 2 S and v 2 f0; 1gn ,
and "  ð d2 =mÞ"1 þ "2 , and solve the convex program
measure the Pauli matrix wðu; vÞ. We call this the ‘‘hybrid
minkktr ; subject to kR
R!k2  ": (6) method’’ because it is equivalent to a certain structured
matrix completion problem. This fact implies that certain
Observation 1 (robustness to noise)—Let
t be an ap-
proximately low-rank state as described above. Suppose key computations in solving the convex program (1) can be
m ¼ cdrlog2 d randomly chosen Pauli expectations are implemented in time OðdÞ rather than Oðd2 Þ [24].
known up to an error of " as in (6), and let ? be the However, the hybrid method is not covered by the strong
solution of (6). Then the difference k?

t ktr is smaller theoretical guarantees shown earlier, though it does give
pffiffiffiffiffi accurate results in practice (when combined with twirling
than Oð" rdÞ. This holds with probability of failure at
most 1=2 plus the probability of failure in Theorem 1. by unitary k-designs [25] to ensure incoherence). For a
The proof combines ideas from Ref. [13] with our argu- more complete discussion, see Ref. [18].
ment above [22]. The main difference from the noise-free Numerical results.—We numerically simulated both the
case is that, instead of using trY
¼ 0, we must now work random Pauli and hybrid approaches discussed above. For
with jtrY
j  2kYk2 . With this estimate, Observation 1 both approaches, we used singular value thresholding
follows from the noise-free proof, together with some (SVT) [24]. Instead of directly solving Eq. (6), SVT mini-
elementary calculations [18]. We remark that the above mizes kktr þ kk22 =2 subject to jtrð
!ÞwðAi Þj  ,
bound is likely to be quite loose; based on related work which is a good proxy to Eq. (6) when  dominates
[23] and more natural noise models, we conjecture that the the second term; the programs are equivalent in the limit
robustness is substantially stronger than what is shown here.  ! 1 ([provided Eq. (6) has a unique solution] [24].
Certified tomography of almost pure states.—The pre- Estimating the second term for typical states suggests
ceding results require an a priori promise: that the true choosing 2r  1; we use  ¼ 5. To simulate tomography,
state
t is 1 close to a rank-r state. However, when per- we chose a random state from the Haar measure on a d  r
forming tomography of an unknown state, neither r nor 1 dimensional system and traced out the r-dimensional an-
are known beforehand. There are a few solutions to this cilla, then applied depolarizing noise of strength . We
quandary. First, r and 1 may be estimated from other sampled expectation values associated with randomly
physical parameters of the system, such as the strength of chosen operators as above, and added additional statistical
the local noise [5]. noise which was i.i.d. Gaussian with variance 2 and mean
Another approach is to estimate r and 1 from the same zero. We used SVT and quantified the quality of the re-
data that are used to reconstruct the state. When r ¼ 1, this construction by the fidelity and the trace distance for
approach is particularly effective to facilitate entirely various values of m, each averaged over 5 simulations.
assumption-free tomography. This is because 1 is related This dependence is shown in Fig. 1. The reconstruction is
to the purity Tr
2 , which has a simple closed-form remarkably high fidelity, despite severe undersampling and
150401-3
 week ending
PRL 105, 150401 (2010) PHYSICAL REVIEW LETTERS 8 OCTOBER 2010

Oversampling Ratio: m 2dr r2 [1] In Quantum State Estimation, edited by M. Paris and
3 4 5 6 7 8 9 10 J. Řeháček, Lect. Notes Phys. Vol. 649 (Springer,
1. 1.
0.9 0.9 Heidelberg, 2004).
0.8 0.8 [2] H. Häffner et al., Nature (London) 438, 643 (2005).
Hybrid Random
0.7 0.7 [3] J. L. O’Brien et al., Nature (London) 426, 264 (2003).

Trace Distance
Fidelity [4] M. S. Kaznady and D. F. V. James, Phys. Rev. A 79,
0.6 0.6
Fidelity

0.5 Trace Dist. 0.5 022109 (2009).

0.4 0.4 [5] Consider a pure state of n qubits subject to local noise that
0.3 0.3 occurs with probability p on each site. Then the density matrix
0.2 0.2 is well approximated by a matrix of rank r ¼ 2nHðpÞ ¼ dHðpÞ ,
0.1 0.1 where HðpÞ is the binary entropy of p, and d ¼ 2n is the
0 0 Hilbert space dimension. When p is small, we have r  d.
0.05 0.1 0.15 0.2 [6] G. M. D’Ariano, L. Maccone, and M. Paini, J. Opt. B 5, 77
Sparsity: m d 2 (2003); V. V. Dodonov and V. I. Man’ko, Phys. Lett. A
229, 335 (1997).
FIG. 1 (color online). Average fidelity and trace distance vs [7] J.-P. Amiet and S. Weigert, J. Phys. A 32, 2777 (1999).
(scaled) number of measurement settings m for random states of [8] B. K. Natarajan, SIAM J. Comput. 24, 227 (1995).
n ¼ 8 qubits, so d ¼ 2n . As discussed in the text, the sampled [9] D. Donoho, IEEE Trans. Inf. Theory 52, 1289 (2006); E.
states had rank r ¼ 3, depolarizing noise of 5% and Gaussian Candès and T. Tao, IEEE Trans. Inf. Theory 52, 5406 (2006).
statistical noise with  ¼ 0:1=d. Both the random Pauli and [10] R. L. Kosut, arXiv:0812.4323.
hybrid approaches are shown. [11] E. J. Candès and B. Recht, Found. Comput. Math. 9, 717 (2009).
[12] E. J. Candès and T. Tao, arXiv:0903.1476 [IEEE Trans.
Inform. Th. (to be published)].
corruption by both depolarizing and statistical noise [26]. [13] E. J. Candès and Y. Plan, Proc. IEEE 98, 925 (2010).
Using the hybrid method with 8 qubits on a rank 3 state plus [14] Our techniques also apply to process tomography: to
 ¼ 5% depolarizing, and statistical noise strength characterize an unknown quantum process E, prepare the
d ¼ 0:1, we typically achieve 95% fidelity reconstruc- Jamiołkowski state
E , and perform state tomography on
tions in under 10 seconds on a modest laptop with 2 GB of
E . Our methods work when E can approximately be
RAM and a 2.2 GHz dual-core processor using MATLAB— written as a sum of only a few Kraus operators, because
even though 90% of the matrix elements remain un- this implies that
E has small rank.
sampled. Increasing the number of samples only improves [15] The techniques easily generalize to spin-j P particles [17].
2
our accuracy and speed, so long as sparsity is maintained. [16] We use theP 2 usual matrix norms kAk tr ¼ i i , kAk2 ¼
y
trA A ¼ i i , kAk ¼ maxi i , with i the singular val-
Using truly randomly chosen Pauli observables (instead
ues of A. The last definition extends to superoperators: if
of the hybrid method) slightly increases the processing A is a superoperator, then kAk is its largest singular
time due to the dense matrix multiplications involved: in value, or, equivalently kAk ¼ sup;kk2 ¼1 kAk2 (a.k.a.
our setup about 1 min. However, this method achieves even ‘‘2 ! 2’’-norm).
better performance with respect to errors, as seen in Fig. 1. [17] D. Gross, arxiv:0910.1879 [IEEE Trans. Inform. Th. (to be
The simulations above show that our method works for published)].
generic low-rank states. Lastly, we demonstrate the func- [18] See supplementary material at http://link.aps.org/
tioning of the approach in the experimental context of the supplemental/10.1103/PhysRevLett.105.150401.
state
found in the 8 ion experiment of Ref. [2]. To [19] R. Ahlswede and A. Winter, IEEE Trans. Inf. Theory 48,
exemplify the above results, we simulated physical mea- 569 (2002).
[20] If the term 1=ð2d2 Þ were zero, Y would be a strict
surements by sampling from the probability distribution
subgradient.
computed using the Born rule applied to the reconstructed [21] S. Becker, S. T. Flammia, D. Gross, Y.-K. Liu, and J. Eisert
state
. This state is approximately low rank, with 99% of (to be published).
the weight concentrated on the first 11 eigenvectors. The [22] Going beyond [13], we bound deviations in 1-norm, as
standard deviation per observable was 3=d. Fewer than opposed to 2-norm. This carries an operational meaning in
30% of all Pauli matrices were chosen randomly. From terms of statistical distinguishability.
this information, a rank ¼ 3 approximation  with fidelity [23] M. Fazel, E. Candès, B. Recht, and P. Parrilo, Proc.
of 90.5% with respect to
was found in about 3 min on the Asilomar Conf. CA, Nov 2008.
aforementioned laptop. [24] J.-F. Cai, E. J. Candès, and Z. Shen, arXiv:0810.3286.
We thank E. Candès and Y. Plan for useful discussions. [25] A. W. Harrow and R. A. Low, Proceedings of 13th Intl.
Research at PI is supported by the Government of Canada Workshop on Randomization and Computation (RANDOM
2009), Berkeley, CA, USA, Aug 2009, Lecture Notes in
through Industry Canada and by the Province of Ontario
Comput. Sci. Vol. 5687 (Springer, Berlin, 2009), p. 548;
through the Ministry of Research & Innovation. Y. L. is D. Gross, K. Audenaert, and J. Eisert, J. Math. Phys.
supported by the NSF, J. E. by the EU (QAP, QESSENCE, (N.Y.) 48, 052104 (2007).
MINOS, COMPAS) and the EURYI, D. G. by the EU [26] The estimate returned by SVT typically has a subnormal-
(CORNER). We thank the anonymous referees for many ized trace, which we handle by renormalizing or by
helpful suggestions. debiasing [21].

150401-4

View publication stats