1
Low-Rank Positive Semidefinite Matrix Recovery
from Corrupted Rank-One Measurements
Yuanxin Li, Student Member, IEEE, Yue Sun, and Yuejie Chi, Member, IEEE⋆
Abstract—We study the problem of estimating a low-rank
positive semidefinite (PSD) matrix from a set of rank-one
arXiv:1602.02737v2 [cs.IT] 31 Aug 2016
measurements using sensing vectors composed of i.i.d. standard
Gaussian entries, which are possibly corrupted by arbitrary
outliers. This problem arises from applications such as phase
retrieval, covariance sketching, quantum space tomography, and
power spectrum estimation. We first propose a convex optimiza-
tion algorithm that seeks the PSD matrix with the minimum ℓ1 -
norm of the observation residual. The advantage of our algorithm
is that it is free of parameters, therefore eliminating the need
for tuning parameters and allowing easy implementations. We
establish that with high probability, a low-rank PSD matrix can
be exactly recovered as soon as the number of measurements
is large enough, even when a fraction of the measurements are
corrupted by outliers with arbitrary magnitudes. Moreover, the
recovery is also stable against bounded noise. With the additional
information of an upper bound of the rank of the PSD matrix,
we propose another non-convex algorithm based on subgradient
descent that demonstrates excellent empirical performance in
terms of computational efficiency and accuracy.
Index Terms—rank-one measurements, low-rank PSD matrix
estimation, outliers
I. I NTRODUCTION
In many emerging applications of science and engineering,
we are interested in estimating a low-rank positive semidef-
inite (PSD) matrix X 0 ∈ Rn×n from a set of nonnegative
magnitude measurements:
zi = hZ i , X 0 i = hai aTi , X 0 i = aTi X 0 ai , (1)
for i = 1, . . . , m, where h·, ·i denotes the inner product
operator. The measurement zi is quadratic in the sensing
vector ai ∈ Rn , but linear in X 0 , where the sensing matrix
Z i = ai aTi is rank-one. On one hand, such magnitude
measurements could arise due to physical limitations, e.g.
incapability of capturing phases, such as in phase retrieval and
optical imaging from intensity measurements [1]–[6], where
only the squared intensity of linear measurements of a signal
x0 ∈ Rn is recorded:
2 In this paper, we focus on robust recovery of the low-
zi = |hai , x0 i| = aTi x0 xT0 ai = aTi X 0 ai ,


(2)
Y. Li and Y. Chi are with Department of Electrical and Computer Engineer-
ing, The Ohio State University, Columbus, OH 43210 USA (e-mails: {li.3822,
chi.97}@osu.edu). Y. Sun is with Department of Electronics Engineering,
Tsinghua University, Beijing, China. Part of the work was done while Y. Sun
was visiting The Ohio State University.
This work is supported in part by NSF under grant CCF-1422966, ECCS-
1462191 and AFOSR under grant FA9550-15-1-0205. Corresponding e-mail:
chi.97@osu.edu. Date: April 4, 2018.
Preliminary results of this paper were presented in part at the IEEE Inter-
national Conference on Acoustics, Speech and Signal Processing, Shanghai,
China, March 2016.
where X 0 = x0 xT0 is a lifted rank-one matrix from the signal
x0 of interest. On the other hand, they could arise by design,
such as from the covariance sketching scheme considered in
[7], where zi is aggregated from squared intensity measure-
ments of L data samples of a zero-mean ergodic data stream
{xl }∞l=1 as
L L
!
1X 2 1X
T
zi = |hai , xl i| = ai xl xl ai ≈ aTi X 0 ai .
T
L L
l=1 l=1
(3)
Here, X 0 = E[xl xTl ] corresponds to the covariance matrix of
the data when L is sufficiently large, and the goal of covariance
sketching is to recover the covariance matrix X 0 from the
set of measurements {zi }m i=1 . In many applications such as
array signal processing [8] and network traffic monitoring [9],
the covariance matrix of the data can be well approximated
by a low-rank PSD matrix, as most of its variance can be
explained by the few top principal components. Last but
not least, measurements of low-rank PSD matrices in the
form of (1) also occur in a number of applications such as
quantum state tomography [10], compressive power spectrum
estimation [11], non-coherent direction-of-arrival estimation
from magnitude measurements [12], synthetic aperture radar
imaging [13], and so on.
It is natural to ask if it is possible to recover the low-rank
PSD matrix X 0 in (1) from an information-theoretically op-
timal number of measurements in a computationally efficient
manner. A popular approach is based on convex relaxation
[7], which seeks the PSD matrix with the smallest trace norm
while satisfying the observation constraint. It is shown in [7]
that this algorithm exactly recovers all rank-r PSD matrices
as soon as the number of measurements exceeds the order of
nr in the absence of noise, and the recovery is stable against
bounded noise as well.
A. Our Goal and Contributions
rank PSD matrix when the measurements in (1) are fur-
ther corrupted by outliers, possibly adversarial with arbitrary
amplitudes. In signal processing applications, outliers are
somewhat inevitable, which may be caused by sensor failures,
malicious attacks, or reading errors. In the application of
covariance sketching, as in (3), a sufficient aggregation length
L is necessary in order for each measurement zi to be well
approximated by (1). Measurements which are not aggregated
from a large enough L may be regarded as outliers. Therefore,
it becomes critical to address robust recovery of X 0 in the

presence of outliers. Fortunately, it is reasonable to assume
that the number of outliers is usually much smaller than the
number of total measurements, making it possible to leverage
the sparsity of the outliers to faithfully recover the low-rank
PSD matrix of interest.
We first propose a convex optimization algorithm that seeks
the PSD matrix that minimizes the ℓ1 -norm of the mea-
surement residual, where the ℓ1 -norm is adopted to promote
outlier sparsity. The proposed convex program is free of tuning
parameters and eliminates the need for trace minimization,
a popular convex surrogate for low-rank matrix recovery, by
only enforcing the PSD constraint. Neither does it require
the knowledge of the outliers, even their existence. When
the sensing vectors are composed of i.i.d. standard Gaussian
entries, we establish that for a fixed n × n rank-r PSD matrix,
as long as the number of measurements exceeds the order
of nr2 , the proposed convex program can exactly recover it
with high probability, even when a fraction of an order of
1/r measurements are arbitrarily corrupted. Our measurement
complexity is order-wisely near-optimal up to a factor of r,
and is near-optimal in the rank-one case up to a constant
factor. Furthermore, the recovery is also stable against additive
bounded noise. While the proposed convex program coincides
with a version of the PhaseLift algorithm [14]–[16] studied
in the literature for phase retrieval, our work provides its first
theoretical performance guarantee to recover low-rank PSD
matrices in the presence of arbitrary outliers. Moreover, we
show the proposed approach can be easily extended to recover
low-rank Toeplitz PSD matrices via numerical simulations.
To further reduce the computational burden when facing
large-scale problems, we next develop a non-convex algorithm
based on subgradient descent when the rank of the PSD
matrix, or an upper bound of it, is known a priori. Since any
rank-r PSD matrix can be uniquely decomposed as X 0 =
U 0 U T0 , where U 0 ∈ Rn×r up to some orthonormal trans-
formations, it is sufficient to recover U 0 without constructing
the PSD matrix explicitly. The subgradient descent algorithm
then iteratively updates the estimate by descending along the
subgradient of the ℓ1 -norm of the measurement residual using
a properly selected step size and spectral initialization. We
conduct extensive numerical experiments to demonstrate its
excellent empirical performance, and compare it against the
convex program proposed above as well as other alternative
approaches in the literature.
B. Organization
The rest of the paper is organized as below. Section II
presents the proposed convex optimization algorithm and its
corresponding performance guarantee, where detailed com-
parisons to related work are presented. Section III describes
the proposed non-convex subgradient descent algorithm that
is computationally efficient with excellent empirical perfor-
mance. Numerical examples are provided in Section IV. The
proof of the main theorem is given in Section V. Finally, we
conclude in Section VI.
2
II. PARAMETER -F REE C ONVEX R ELAXATION
A. Problem Formulation
Let X 0 ∈ Rn×n be a rank-r PSD matrix, then the set of
m measurements, which may be corrupted by either arbitrary
outliers or bounded noise, can be represented as
z = A(X 0 ) + β + w, (4)
where z, β, w ∈ Rm . The mapping A: Rn×n → Rm is
 Tlinear m
defined as A (X 0 ) = ai X 0 ai i=1 , where ai ∈ Rn is the
ith sensing vector composed of i.i.d. standard Gaussian entries,
i = 1, . . . , m. The vector β denotes the outlier vector, which
is assumed to be sparse whose entries can be arbitrarily large.
The fraction of nonzero entries is defined as s := kβk0 /m.
Moreover, the vector w denotes the additive noise, which is
assumed bounded as kwk1 ≤ ǫ. Our goal is to robustly recover
X 0 from the measurements z.
B. Recovery via Convex Relaxation
To motivate our algorithm, consider the case when only
the outlier vector β is present in (4) and the rank of X 0
is known. One may seek a rank-r PSD matrix that minimizes
the cardinality of the measurement residual to motivate outlier
sparsity, given as
X̂ = argminX0 kz − A(X)k0 , s.t. rank(X) = r. (5)
However, both the cardinality minimization and the rank
constraint are NP-hard in general, making this method compu-
tationally infeasible. A common approach is to resort to convex
relaxation, where we relax the cardinality minimization by its
convex relaxation, i.e. the ℓ1 -norm, and meanwhile, drop the
rank constraint, yielding:
(Robust-PhaseLift:) X̂ = argminX0 kz − A(X)k1 . (6)
We denote the above convex program as the Robust-PhaseLift
algorithm, since it coincides with the PhaseLift algorithm
studied in [14]–[16] for phase retrieval1 . The advantage of
Robust-PhaseLift in (6) is that it does not require any prior
knowledge of the noise bound, the rank of X 0 , nor the sparsity
level of the outliers, and is free of any regularization parameter.
It is also worth emphasizing that due to the special rank-
one measurement operator, in (6) it is possible to only honor
the PSD constraint but not motivate the low-rank structure
explicitly, via for example, trace minimization2.
Encouragingly, we demonstrate that the algorithm (6) admits
robust recovery of a rank-r PSD matrix as soon as the number
of measurements is large enough, even with a fraction of arbi-
trary outliers in Theorem 1. To the best of our knowledge, this
is the first theoretical performance guarantee of the robustness
of (6) with respect to arbitrary outliers in the low-rank setting.
Our main theorem is given as below.
1 Note that there are a few different versions of PhaseLift in the literature
which are not outlier-robust, therefore we rename (6) to Robust-PhaseLift for
emphasis.
2 The interested readers are invited to look up Fig. 1 in [2] for an intuitive
geometric interpretation in the noise-free and outlier-free case.

Theorem 1. Suppose that kwk1 ≤ ǫ and s = kβk0 /m.
Assume the support of β is selected uniformly at random with
the signs of its nonzero entries generated from the Rademacher
distribution as P {sgn (βi ) = −1} = P {sgn (βi ) = 1} = 1/2
for each i ∈ supp(β). Then for a fixed rank-r PSD matrix
X 0 ∈ Rn×n , there exist some absolute constants c1 > 0 and
0 < s0 < 1 such that as long as
s0
m ≥ c1 nr2 , s ≤ ,
r
the solution to (6) satisfies
rǫ variant of PhaseLift for phase retrieval, corresponding to the
X̂ − X 0
≤ c2 , case where X 0 = x0 xT0 is rank-one. It is shown in [14], [15]
F m
with probability exceeding 1 − exp(−γm/r2 ) for some con-
stants c2 and γ.
Theorem 1 has the following consequences.
• Exact Recovery with Outliers: When ǫ = 0, Theorem 1
suggests the recovery is exact using Robust-PhaseLift (6),
i.e. X̂ = X 0 even when a fraction of measurements are
arbitrarily corrupted, as long as the number of measure-
ments m is on the order of nr2 . Given there are at least nr
unknowns, our measurement complexity is near-optimal
up to a factor of r.
• Stable Recovery with Bounded Noise: In the presence
of bounded noise, Theorem 1 suggests that the recovery
performance decreases gracefully with the increase of ǫ,
where the Frobenius norm of the reconstruction error is
proportional to the per-entry noise level of the measure-
ments.
• Phase Retrieval: When r = 1, the problem degenerates
to the case of phase retrieval, and Theorem 1 recovers
existing results in [16] for outlier-robust phase retrieval,
where the measurement complexity is on the order of n,
which is optimal up to a scaling factor.
Let us denote X̂ r = argminrank(Z)=r,Z0 kX̂ − ZkF as
the best rank-r PSD matrix approximation of X̂, the solution
to (6). Then Theorem 1 suggests that the estimate X̂ can be
well approximated by a rank-r PSD matrix since
rǫ
kX̂ − X̂ r kF ≤ kX̂ − X 0 kF ≤ c2 ,
m A LGORITHM
as long as the number of measurements is sufficiently large.
Furthermore, we have
kX̂ r − X 0 kF ≤ kX̂ r − X̂kF + kX̂ − X 0 kF
rǫ
≤ 2kX̂ − X 0 kF ≤ 2c2 ,
m decompose X 0 as X 0 = U 0 U T0 where U 0 ∈ Rn×r . Instead
indicating that X̂ r provides an accurate estimate of X 0 that
is exactly rank-r and PSD.
C. Comparisons to Related Work
In the absence of outliers, the PhaseLift algorithm in the
following form
min Tr(X) s.t.
X0
kz − A(X)k1 ≤ ǫ, (7)
where Tr(X) denotes the trace of X, has been proposed
to solve the phase retrieval problem [2], [3], [14]. Later
3
the same algorithm has been employed to recover low-rank
PSD matrices in [7], where an order of nr measurements
obtained from i.i.d. sub-Gaussian sensing vectors are shown
to guarantee exact recovery in the noise-free case and stable
recovery with bounded noise. One problem with the algorithm
(7) is that the noise bound ǫ is assumed known. Furthermore,
it is not amenable to handle outliers, since kz − A(X 0 )k1 can
be arbitrarily large with outliers and consequently the ground
truth X 0 quickly becomes infeasible for (7).
The proposed algorithm (6) is studied in [14]–[16] as a
that with O(n) i.i.d. Gaussian sensing vectors, the algorithm
succeeds with high probability. Compared with (7), the algo-
rithm (6) eliminates trace minimization and leads to easier
algorithm implementations. We note that [17] also considers
a regularization-free algorithm for PSD matrix estimation that
minimizes the ℓ2 -norm of the residual, which unfortunately,
cannot handle outliers as Robust-PhaseLift (6). Hand [16] first
considered the robustness of the Robust-PhaseLift algorithm
(6) in the presence of outliers for phase retrieval, establishing
that the same guarantee holds even with a constant fraction of
outliers. Our work extends the performance guarantee in [16]
to the general low-rank PSD matrix case.
Broadly speaking, our problem is related to low-rank matrix
recovery from an under-determined linear system [18]–[20],
where the linear measurements are drawn from inner products
with rank-one sensing matrices. It is due to this special
structure of the sensing matrices that we can eliminate the trace
minimization, and only consider the feasibility constraint for
PSD matrices. Standard approaches for separating low-rank
and sparse components [21]–[25] via convex optimization are
given as
min Tr(X) + λkβk1 , s.t. kz − A(X) − βk1 ≤ ǫ,
X0, β
where λ is a regularization parameter that requires to be tuned
properly. In contrast, the formulation (6) is parameter-free.
III. A N ON -C ONVEX S UBGRADIENT D ESCENT
In this section, we propose another algorithm for robust
low-rank PSD matrix recovery from corrupted rank-one mea-
surements assuming the rank (or its upper bound) of the PSD
matrix X 0 is known a priori as r. In this case, we can
of directly recovering X 0 , we may aim at recovering U 0 up
to orthogonal transforms, since (U 0 Q)(U 0 Q)T = U 0 U 0 for
any orthonormal matrix Q ∈ Rr×r . Consider relaxing of the
loss function in (5) but keeping the rank constraint, we obtain
the following problem:
X̂ = argminX0 kz − A(X)k1 , s.t. rank(X) = r. (8)
Since any rank-r PSD matrix X can be written as X = U U T
for some U ∈ Rn×r , (8) can be equivalently reformulated as
Û = argminU ∈Rn×r f (U ), (9)
 4

no outliers modest outlier amplitudes large outlier amplitudes

f (U ) = 1
2m kz − A(U U T )k1

g(U ) = 1
4m kz − A(U U T )k22

Fig. 1: Illustrations of the objective function − log f (U ) and its ℓ2 -norm counterpart − log g(U ) (in negative logarithmic
scales) under different corruption scenarios when U ∈ R2×1 . The number of measurements is m = 100 with i.i.d. Gaussian
sensing vectors, and the fraction of outliers is s = 0.2 with uniformly selected support and amplitudes drawn from Unif[0, 10]
or Unif[0, 100]. It is interesting to observe that while large outliers completely distort g(U ), the proposed objective is quite
robust with the ground truth being the only global optima of f (U ).

with following matrix with respect to Frobenius norm as

m   2
1 1 X  T 2  m
T
f (U ) = z − A(U U T ) =

z − a i  . (0)

(0)
1 X T

i U U = argminrank(X)=r X − zi ai ai .

2m 2m i=1 
U
1 2 m i=1
F
(11)
Clearly, (9) is no longer convex. To illustrate, the first row of
Secondly, at the (t+1)th iteration, t ≥ 0, we apply subgradient
Fig. 1 plots the value of the objective function in the negative
descent to refine the estimate as
logarithmic scale, i.e. − log f (U ), under different corruption
scenarios when U ∈ R2×1 . For comparison, the second row of U (t+1) = U (t) − µt · ∂f (U (t) ), (12)
Fig. 1 shows the loss function evaluated in ℓ2 -norm: g(U ) =
1 T 2 where the step size µt is adaptively set as
4m kz − A(U U )k2 , which is not robust to outliers.
Motivated by the recent non-convex approaches [26]–[28] of n o
solving quadratic systems, we propose a subgradient descent µt = 0.05 × max 2−t/1000 , 10−6 ,
algorithm to solve (9) effectively, working with a non-smooth
which provide more accurate estimates using fewer iterations
function f (U ). Note that a subgradient of f (U ) with respect
in the numerical simulations. The procedure is summarized in
to U can be given as
Alg. 1, where the stopping rule in Alg. 1 is simply put as a
m   maximum number of iterations.
1 X T 2

∂f (U ) = − sgn zi − U ai ai aTi U , (10)
m i=1 2
Algorithm 1: Subgradient descent for solving (9)
where the sign function sgn(·) is defined as Parameters: Rank r, number of iterations Tmax , and
 step size µt ;
 +1, x > 0 Input: Measurements z, and sensing vectors {ai }m i=1 ;
sgn(x) = 0, x = 0 . Initialization: Initialize U (0) ∈ Rn×r via (11);
−1, x < 0 for t = 0 : Tmax − 1 do

update U (t+1) via (12);
Our subgradient descent algorithm proceeds as below. De- end for
note the estimate in the tth iteration by U (t) ∈ Rn×r . First, Output: Û = U (Tmax ) .
U (0) is initialized as the best rank-r approximation of the
 5

The main advantage of Alg. 1 is its low memory and 12

1

0.9
12
1

0.9

computational complexity. Given that it does not construct 10 0.8 10 0.8

0.7 0.7

the full PSD matrix, the memory complexity is simply the 8

0.6
8
0.6

Rank (r)

Rank (r)
size of U (t) , which is on the order of nr. The computational 6
0.5

0.4
6
0.5

0.4

complexity per iteration is also low, which is on the order of 4 0.3 4 0.3

mnr, that is linear in all the parameters. We demonstrate the

0.2 0.2
2 2
0.1 0.1

excellent empirical performance of Alg. 1 in Section IV-C. 100 200 300 400
Number of measurements (m)
500 600
0
0 0.05 0.1
Percent of outliers (s)
0.15 0.2
0

(a) (b)
IV. N UMERICAL E XAMPLES
Fig. 3: Phase transitions of low-rank PSD matrix recovery
A. Performance of Convex Relaxation with respect to (a) the number of measurements and the rank,
with 5% of measurements corrupted by standard Gaussian
We first examine the performance of Robust-PhaseLift in variables; (b) the percent of outliers and the rank, when the
(6). Let n = 40. We randomly generate a low-rank PSD matrix number of measurements is m = 600, when n = 40.
of rank-r as X 0 = U 0 U T0 , where U 0 ∈ Rn×r is composed of
i.i.d. standard Gaussian variables. The sensing vectors are also
composed of i.i.d. standard Gaussian variables. Each Monte B. Convex Relaxation with additional Toeplitz Structure
Carlo simulation is called successful if the normalized estimate
We next consider robust recovery of low-rank Toeplitz PSD
error satisfies kX̂ −X 0 kF /kX 0 kF ≤ 10−6 , where X̂ denotes
matrices, where we allow complex-valued sensing vectors
the solution to (6). For each cell, the success rate is calculated
A(X) = {aH m
i Xai }i=1 and complex-valued Toeplitz PSD
by averaging over 100 Monte Carlo simulations.
matrices X. Estimating low-rank Toeplitz PSD matrices is of
great interests for array signal processing [29]. We modify (6)
12
1

0.9
12
1

0.9
by incorporating the Toeplitz constraint as:
10 0.8 10 0.8

0.7 0.7 X̂ = argminX0 kz − A(X)k1 , s.t. X is Toeplitz. (13)

8 8
0.6 0.6
Rank (r)

Rank (r)

6
0.5

0.4
6
0.5

0.4
Let n = 64, the Toeplitz PSD matrix X 0 is generated as X 0 =
4 0.3 4 0.3
V ΣV H , where V = [v(f1 ), . . . , v(fr )] ∈ Cn×r is a Vander-
monde matrix with v(fi ) = [1, ej2πfi , . . . , ej2π(n−1)fi ]T , fi ∼
0.2 0.2
2 2
0.1 0.1

100 200 300 400 500

Number of measurements (m)
600
0
100
(a)
Fig. 2: Phase transitions for low-rank PSD matrix recovery
with respect to the number of measurements and the rank, (a)
with trace minimization; and (b) without trace minimization
of noise-free measurements, when n = 40.
Fig. 2 shows the success rates of algorithms with respect
to the number of measurements and the rank, with the trace
minimization as in (7) in (a); and without the trace minimiza-
tion as proposed in Robust-PhaseLift (6) in (b) for noise-free
measurements. It can be seen that the performance of these
two algorithms are almost equivalent, confirming a similar
200 300 400 500
Number of measurements (m)
600
0
Unif[0, 1], and Σ = diag[σ12 , . . . , σr2 ], with σi2 ∼ Unif[0, 1].
Fig. 4 shows the phase transitions of Toeplitz PSD matrix
(b)
recovery with respect to the number of measurements and the
rank without outliers in (a), and when 5% of measurements
are selected uniformly at random and corrupted by standard
Gaussian variables in (b). It can be seen that the low-rank
Toeplitz PSD matrix can be robustly recovered from a sub-
linear number of measurements due to the additional Toeplitz
structure. We note that a different covariance sketching scheme
is considered in [30]–[32] for estimating low-rank Toeplitz
covariance matrices. Though not directly comparable to our
measurement scheme, it may benefit from a similar parameter-
free convex optimization to handle outliers.
12
1
12
1

0.9 0.9

numerical observation for the phase retrieval problem [4] also 10 0.8 10 0.8

holds in the low-rank setting, where trace minimization may 8

0.7

0.6
8
0.7

0.6
Rank (r)

Rank (r)

be eliminated for low-rank PSD matrix recovery using rank- 6

0.5
6
0.5

0.4 0.4
one measurements. 4 0.3 4 0.3

Fig. 3 further shows the success rates of the Robust- 2

0.2

0.1
2
0.2

0.1

PhaseLift algorithm (a) with respect to the number of measure- 10 20 30 40 50 60 70

0
10 20 30 40 50 60 70
0

Number of measurements (m) Number of measurements (m)

ments and the rank, when 5% of measurements are selected
uniformly at random and corrupted by standard Gaussian (a) (b)
variables; and (b) with respect to the percent of outliers and
the rank, for a fixed number of measurements m = 600. Fig. 4: Phase transitions of low-rank Toeplitz PSD matrix
This also suggests possible room for improvements of our recovery with respect to the number of measurements and the
theoretical guarantee, as the numerical results indicate that the rank, (a) without outliers, and (b) with 5% of measurements
required measurement complexity for successful recovery has corrupted by standard Gaussian variables, when n = 64.
a seemingly linear relationship with r.
 6

C. Performance of Non-Convex Subgradient Descent 12

1

0.9
12
1

0.9

We next examine the performance of the non-convex sub- 10 0.8

0.7
10 0.8

0.7

gradient descent algorithm in Alg. 1, where the number of 8

0.6
8
0.6

Rank (r)
Rank (r)
iterations is set as Tmax = 3 × 104 , which is a large
0.5 0.5
6 6
0.4 0.4

value to guarantee convergence when terminated. Denote the 4 0.3

0.2
4 0.3

0.2

solution to Alg. 1 by Û , and each Monte Carlo simulation is 2

0.1
2
0.1

deemed successful if the normalized estimate error Tsatisfies 0 0.05 0.1 0.15
Percent of outliers (s)
0.2
0
0 0.05 0.1 0.15
Percent of outliers (s)
0.2
0

kX̂ − X 0 kF /kX 0 kF ≤ 10−6 , where X̂ = Û Û is the (a) (b)

estimated low-rank PSD matrix. For each cell, the success rate
is calculated by averaging over 100 Monte Carlo simulations. Fig. 6: Phase transitions of low-rank PSD matrix recovery with
respect to the percent of outliers and the rank using (a) the
12
1
proposed Alg. 1, and (b) the WF algorithm, when n = 40 and
0.9
m = 600.
10 0.8

0.7
8
0.6
Rank (r)

0.5
in the bounded noise w is i.i.d. drawn from Unif[−4/m, 4/m],
6
0.4 thus kwk1 ≤ ǫ, where ǫ = 4. Fig. 7 depicts the mean
4 0.3 squared error kX̂ −X 0 k2F for different algorithms with respect
0.2 to the number of measurements, where X̂ is the estimated
2
0.1
PSD matrix. For the subgradient descent algorithm in Alg. 1,
0
100 200 300 400 500
Number of measurements (m)
600 various ranks are used as prior information, corresponding to
the correct rank r, its underestimate r −1, and its overestimate
r + 1. It can be seen that Alg. 1 works well as long as the
Fig. 5: Phase transitions of low-rank PSD matrix recovery with given rank provides an upper bound of the true rank, and it
respect to the number of measurements and the rank for the performs much better than the WF algorithm which is not
proposed Alg. 1 using noise-free measurements, when n = 40. outlier-robust. On the other hand, the PhaseLift algorithm (7)
does not admit favorable performance for various constraint
Fig. 5 shows the success rate of Alg. 1 with respect to the parameters (ǫ, 2ǫ, 4ǫ) as expected since the outliers do not fall
number of measurements and the rank under the same setup into the prescribed noise bound. In fact, it fails to return any
of Fig. 2 for noise-free measurements, when n = 40. Indeed, feasible solution when the number and amplitudes of outliers
empirically Alg. 1 performs similarly as the convex algorithms is too large in our simulation. In contrast, Robust-PhaseLift
but with a much lower computational cost. Moreover, the allows stable recovery even with an additional bounded noise,
proposed Alg. 1 allows perfect recovery even in the presence which performs comparably with Alg. 1 with the correct model
of outliers. For comparison, we implement the extension of the order.
Wirtinger Flow (WF) algorithm in [26], [28], [33] in the low-
rank case, that minimizes the squared ℓ2 -norm of the residual, 10 4
where the update rule per iteration becomes Robust-PhaseLift
Alg. 1 with r
m Alg. 1 with r+1
1 X 
10 2
Alg. 1 with r-1
U (t+1) = U (t) +µWF
t z i − k(U (t) T
) a k 2 T
i 2 ai ai U
(t)
, Wirtinger Flow
PhaseLift with ǫ
m i=1
Mean squared error

PhaseLift with 2ǫ
PhaseLift with 4ǫ
10 0
using the same initialization (11). The step size is set as
µWF
t = 0.1/ kU 0 k2F . Fig. 6 (a) shows the success rates of
Alg. 1 with respect to the percent of outliers and the rank, 10 -2

under the same setup of Fig. 3 (b), where the performance is

even better than the convex counterpart in (6). In contrast, the 10 -4
WF algorithm performs poorly even with very few outliers, as
shown in its success rate plot in Fig. 6 (b), as the loss function 10 -6
100 200 300 400 500 600
used for WF is not robust to outliers. Number of measurements (m)

D. Comparisons with Additional Bounded Noise
Finally, we compare the two proposed algorithms (Robust-
PhaseLift in (6) and Alg. 1), the WF algorithm and the
PhaseLift algorithm in (7) when the measurements are cor-
rupted by both outliers and bounded noise. Fix n = 40 and
r = 3. The rank-r PSD matrix X 0 , the sensing vectors, as
well as the outliers are generated similarly as earlier, where
the fraction of the outliers is set to 5%. Moreover, each entry
Fig. 7: Comparisons of mean squared errors using different
algorithms with respect to the number of measurements with
5% outliers and bounded noise, when n = 40 and r = 3.
V. P ROOF OF M AIN T HEOREM
In this section we prove Theorem 1, and the roadmap
of our proof is below. In Section V-A, we first provide
 7

the sufficient conditions for an approximate dual certificate the solution to (6) satisfies
that certifies the optimality of the proposed algorithm (6) in rǫ
X̂ − X 0 ≤ c ,

Lemma 1. Section V-B records a few lemmas that show A F m
satisfies the required restricted isometry properties. Then, a where c is a constant.
dual certificate is constructed and validated for a fixed low-
rank PSD matrix X 0 in Section V-C. Finally, the proof is Proof: Denote the solution to (6) by X̂ = X 0 +H 6= X 0 ,
concluded in Section V-D. then we have X̂  0, H T ⊥  0, and furthermore,
First we introduce some additional notations. Let S be a kA(H) − (β + w)k1 = kz − A(X 0 + H)k1
subset of {1, 2, . . . , m}, then S ⊥ is the complement of S with
respect to {1, 2, . . . , m}. AS is the mapping operator A con- = kz − A(X̂)k1
strained on S, which is defined as AS (X) = aTi Xai i∈S . ≤ kz − A(X 0 )k1 = kβ + wk1 ,

Pm
Denote the adjoint operator of A by A∗ (µ) = i=1 µi ai aTi , where the inequality follows from the optimality of X̂ since
where µi is the ith entry of µ, 1 ≤ i ≤ m. We use kXk, kXkF both X̂ and X 0 are feasible to (6). Since
and kXk1 to denote the spectral norm, the Frobenius norm
and the nuclear norm of the matrix X, respectively, and use kA(H)−(β+w)k1 = kAS (H)−β−wS k1 +kAS ⊥ (H)−wS ⊥ k1 ,
kxkp to denote the ℓp -norm of the vector x. Let the singular and
value decomposition of the fixed rank-r PSD matrix X 0 be kβ + wk1 = kβ + wS k1 + kwS ⊥ k1 ,
X 0 = U ΛU T , then the symmetric tangent space T at X 0 is
denoted by we have
kAS ⊥ (H)k1 ≤ kAS ⊥ (H) − wS ⊥ k1 + kwS ⊥ k1
n o
T := U Z T + ZU T | Z ∈ Rn×r .
≤ kβ + wk1 − kAS (H) − β − wS k1 + kwS ⊥ k1
We denote by PT and PT ⊥ the orthogonal projection onto ≤ kβ + wS k1 − kAS (H) − β − wS k1 + 2kwS ⊥ k1
T and its orthogonal complement, respectively. And for no-
≤ kAS (H)k1 + 2kwS ⊥ k1 ,
tational simplicity, we denote H T := PT (H) and H T ⊥ :=
H −PT (H) for any symmetric matrix H ∈ Rn×n . Moreover, where the last inequality follows from the triangle inequality.
γ, c, c1 and c2 represent absolute constants, whose values may We could further bound
change according to context.
kAS ⊥ (H T )k1 ≤ kAS ⊥ (H)k1 + kAS ⊥ (H T ⊥ )k1
≤ kAS (H)k1 + kAS ⊥ (H T ⊥ )k1 + 2kwS ⊥ k1
A. Approximate Dual Certificate
≤ kAS (H T )k1 + kAS (H T ⊥ )k1
The following lemma suggests that under certain appropriate + kAS ⊥ (H T ⊥ )k1 + 2kwS ⊥ k1
restricted isometry preserving properties of A, a properly
constructed dual certificate can guarantee faithful recovery of = kAS (H T )k1 + kA(H T ⊥ )k1 + 2kwS ⊥ k1 .
the proposed algorithm (6). (19)
Our assumptions on A imply that
Lemma 1 (Approximate Dual Certificate for (6)). Denote a  
subset S with |S| s√0 1
m := ⌈ 13 2r ⌉, where 0 < s0 < 1 is some 1+ Tr (H T ⊥ )
constant, and the support of β satisfies supp(β) ⊆ S. Suppose 10
that the mapping A obeys that for all symmetric matrices X, 1
≥ kA (H T ⊥ )k1
  m
1 1 1
kA (X)k1 ≤ 1 + kXk1 , (14) ≥ (kA ⊥ (H T )k1 − kAS (H T )k1 − 2kwS ⊥ k1 )
m 10 m S
|S ⊥ |
  
and 1 |S| 1 2ǫ
  ≥ 1− kH T kF − 1+ kH T k1 − ,
1 1 5m 12 m 10 m
kAS (X)k1 ≤ 1+ kXk1 , (15)
|S| 10 where the first inequality follows from (14) due to kH T ⊥ k1 =
and for all matrices X ∈ T , Tr (H T ⊥ ), as H T ⊥  0, the second inequality follows from
  (19), and the last inequality follows from (15) and (16). This
1 1 1 gives
kAS ⊥ (X)k1 > 1− kXkF , (16)
|S ⊥ | 5 12
|S | |S| √
 ⊥ 
2ǫ
Tr (H T ⊥ ) ≥ − 2r kH T kF − , (20)
where AS and AS ⊥ is the operator constrained on S and S ⊥ 6m m m
respectively. Then if there exists a matrix Y = A∗ (µ) that √
satisfies where we use the inequality kH T k1 ≤ 2rkH T kF .
1 1 On the other hand, since µ/(9/m) is a subgradient of the
Y T ⊥  − I T ⊥ , kY T kF ≤ , (17) ℓ1 -norm at β from (18), we have
r 13r Dm E
and kβk1 + µ, w − A(H) ≤ kw + β − A(H)k1

µi = m 9
sgn(βi ), i ∈ supp(β) 9
9 , (18) ≤ kβ + wk1 ≤ kβk1 + kwk1 ,
|µi | ≤ m , i∈/ supp(β)
 8

which, by a simple transformation, is Lemma 3 ( [7]). Suppose the sensing vectors ai ’s are
9 composed of i.i.d. sub-Gaussian entries, then there exist pos-
hµ, A(H)i ≥ hµ, wi − kwk1 itive universal constants c1 , c2 , c3 such that, provided that
m
  m > c3 nr, for all matrices X of rank at most r, one has
9 18ǫ
≥ − kµk∞ + kwk1 ≥ − . 2
m m 1 − δrlb kXkF ≤


kB (X)k1 ≤ 1 + δrub kXkF ,


m
Then with
hH, Y i = hA(H), µi, with probability exceeding 1 − c1 e−c2 m , where δrlb and δrub
are defined as the RIP-ℓ2/ℓ1 constants. And the operator B
we can get represents the linear transformation that maps X ∈ Rn×n
to {Bi (X)}m/2i=1 ∈ R
m/2
, where Bi (X) := ha2i−1 aT2i−1 −
18ǫ T
− ≤ hA(H), µi = hH, Y i a2i a2i , Xi.
m
= hH T , Y T i + hH T ⊥ , Y T ⊥ i The third condition (16) can be easily validated from the
2
1 lower bound by setting δrlb appropriately, since m kB (X)k1 ≤
≤ kY T kF kH T kF − hH T ⊥ , I T ⊥ i 2 Pm/2 
 T
  T


r m i=1 ha 2i−1 a 2i−1 , Xi +  ha 2i a 2i , Xi =
1 1 2kA (X) k1 .
≤ kH T kF − Tr(H T ⊥ ),
13r r
which gives C. Construction of Dual Certificate
1 18rǫ For notational simplicity, let α0 := EZ 2 1{|Z|≤3} ≈ 0.9707,
Tr(H T ⊥ ) ≤ kH T kF + . (21)
13 m β0 := EZ 4 1{|Z|≤3} ≈ 2.6728 and θ0 := EZ 6 1{|Z|≤3} ≈
Combining with (20), we know 11.2102 for a standard Gaussian random variable Z, where
1E is an indicator function with respect to an event E.
|S | |S| √
 ⊥ 
2ǫ 1 18rǫ Consider that the singular value decomposition of a PSD
− 2r kH T kF − ≤ kH T kF + .
6m m m 13 m matrix
Pr X 0 of rank at most r can be represented as X 0 =
T
⊥ √ λ
i=1 i ui ui , then inspired by [14], [16], we construct Y as
Since |S6m| − |S|
m
1
2r − 13 > 0 under the assumption on |S|
m r
in Lemma 1, we have 1 X h 1 X  T 2
Y := a ui 1{|aT ui |≤3}
m r i=1 j j
20rǫ rǫ j∈S ⊥
kH T kF ≤ √  ≤ c1 ,
β0 − α0 i 9 X

m |S ⊥ |
− |S| 2r − 1 m
6m m 13 − (α0 + ) · aj aTj + χj aj aTj
r m
j∈S
where c1 is some fixed constant. Finally, we have (0) (1) (2)
:= Y −Y +Y ,
kX̂ − X 0 kF ≤ kH T kF + kH T ⊥ kF
where
≤ kH T kF + Tr(H T ⊥ ) " r #

1

18rǫ rǫ (0) 1 X 1 X  T 2
≤ 1+ kH T kF + ≤c , Y = aj ui 1{|aT ui |≤3} aj aTj ,
13 m m m r i=1
j
j∈S ⊥
  X
for some constant c. 1 β0 − α0
Y (1) = α0 + aj aTj ,
m r
j∈S ⊥
B. Restricted Isometry of A 9 X
Y (2) = χj aj aTj .
The first two conditions (14) and (15) in Lemma 1 are m
j∈S
supplied straightforwardly in the following lemma as long as
We set χj = sgn (βj ) if j ∈ supp(β), otherwise χj ’s
m ≥ cnr and |S| = c1 m/r ≥ c2 n for some constants c, c1
are i.i.d. Rademacher random variables with P {χj = 1} =
and c2 .
P {χj = −1} = 1/2.
Lemma 2 ( [2]). Fix any δ ∈ (0, 12 ) and assume m ≥ 20δ −2 n. The construction immediately indicates that Y satisfies (18).
Then for all PSD matrices X, one has We will show that Y satisfies (17) with high proability. In what
follows, we separate the constructed Y into two parts and
1
(1 − δ) kXk1 ≤ kA (X)k1 ≤ (1 + δ) kXk1 consider the bounds on Y (0) − Y (1) and Y (2) respectively.
m
1) Proof of Y T ⊥ + r1 I T ⊥  0: First, by standard results
2
with probability exceeding 1 − 2e−mǫ /2 , where ǫ2 + ǫ = 4δ . in random matrix theory [34, Corollary 5.35], we have
The right hand side holds for all symmetric matrices.
m

β0 − α0

β0
(1)


|S ⊥ | Y − α0 + I ≤ 40r ,
The third condition (16) in Lemma 1 can be obtained using r
the mixed-norm RIP-ℓ2 /ℓ1 provided in [7] as long as m ≥ cnr
1 − 2e−γ |S |/r for some constant γ
⊥ 2

and |S| ≤ c1 m for some constants c and c1 . with probability

 ⊥  at least
provided S  ≥ cnr2 for some constant c. In particular, this
 9

 
gives 2) Proof of kY T kF ≤ 13r1
: Let Ỹ = Y (0) − Y (1) U ,
′
 
and Ỹ = I − U U T Ỹ be the projection of Ỹ onto the
 
m (1) β0 − α0 β0
|S ⊥ | Y T ⊥ − α0 + IT ⊥ ≤ . (22)

r 40r orthogonal complement of U , then we have
2 2 ′ 2
Let a′j = (I − U U T )aj be the projection of aj onto the (0) (1)
Y T − Y T = U T Ỹ + 2 Ỹ .

(26)
orthogonal complement of the column space of U , then we F F F
T
have First consider the term kU Ỹ k2F
in (26), where the kth
(0) 1 X column of U T Ỹ can be expressed explicitly as
Y T⊥ = ǫj ǫTj ,
m ⊥
 
j∈S
U T Ỹ
 P 2 1/2 k "
1 r r #
where ǫj = r i=1 aTj ui  1{|aT ui |≤3} a′j are i.i.d.
 
1 X 1 X  T 2 β0 − α0
j
= a ui 1{|aT ui |≤3} − α0 +
copies of a zero-mean, isotropic and sub-Gaussian random m r i=1 j j r
j∈S ⊥
vector ǫ, which satisfies E[ǫǫT ] = α0 I T ⊥ . Following [34,  
· aTj uk U T aj


Theorem 5.39], we have
1

m (0) α0
Y − α 0 T ≤
I ⊥
, (23) := Φck ,
|S ⊥ | T ⊥ 40r m
where Φ ∈ Rr×|S | is constructed by U T aj ’s, and ck ∈
⊥
−γ |S ⊥ |/r 2
 ⊥  at least2 1 − 2e
with probability for some constant
R|S | is composed of ck,j ’s, each one expressed as
⊥

γ provided S  ≥ cnr for some constant c. As a result, if

m ≥ cnr2 for some large constant c and |S| ≤ c1 m for some ck,j =
2
constant c1 small enough, with probability at least 1−e−γm/r , " r  #
there exists 1 X  T 2 β0 − α0
aTj uk ,


aj ui 1{|aT ui |≤3} − α0 +
r i=1 j r
Y ⊥ − Y (1)⊥ + β0 − α0 I T ⊥
(0)
T T r with

1
 

(0) (1) β0 − α0 S ⊥ 
E[c2k,j ] = θ0 + (r − 1) β0 − β02 − (r − 1) α20


≤ Y T ⊥ − Y T ⊥ + IT ⊥

r 2
r m
1 1
4.07

β0 − α20 + 2 θ0 + α20 − β02 − β0 ≤


 ⊥ !
S  β0 − α0 = .
r r r
+ 1−
m r Note that c2k,j ’s are i.i.d. sub-exponential random variables

m (0) m (1) β0 − α0 |S| β0 − α0 with c2k,j ≤ K, for some constant K, then according

≤ ⊥ Y T⊥ − ⊥ Y T⊥ +
IT ⊥ + m ψ1
|S | |S | r r to [34, Corollary 5.17],
β0 α0
≤ + .
  
(24)  X   !
2  ⊥
30r 60r ǫ  ǫ S
c2 − Ec2k,j  ≥ S ⊥  ≤ 2exp −c 2 2

  
P  ,

(2)  ⊥ k,j r K r
let’s check Y T ⊥ . Since Y (2) =

Next, j∈S 
1 P T T
m j∈S 9χj aj aj , where E[9χj aj aj ] = 0, by [34, which shows that as long as |S| ≤ c1 m, for some constants c
Theorem 5.39] we have and c1 ,
2 4.07 + c 4.1m
kck k2 ≤ m≤

|S|
r r
1 1
X
(2) T

Y = 9χ j j j ≤
a a ,
m |S|
10r 2
holds with probability at least 1 − e−γm/r . Furthermore, for
j∈S
a fixed vector x ∈ R|S | obeying kxk2 = 1, kΦxk2 is
⊥ 2
with probability at least 1 − 2 exp(−γm/r) as long as m ≥ distributed as a chi-square random variable with r degrees
cnr2 and |S| = c1 m/r ≥ c2 nr, for some constants c, c1 and of freedom. From [35, Lemma 1], we have
c2 . In particular, this gives
2 m
kΦxk2 ≤ ,

(2)
1 12000r2
Y T ⊥ ≤ . (25)
10r 2
with probability at least 1 − e−γm/r , provided m ≥ cnr2 for
Putting this together with (24), we can obtain that if m ≥ some sufficiently large constant c. Therefore, we can obtain
cnr2 and |S| = c1 m/r ≥ c2 nr for some constants c, c1 and 2
2
c2 , with probability at least 1 − e−γm/r ,

T
 2 1 ck kck k22 ≤ 1
U Ỹ = Φ ,

m2 kck k2 2 2700r3

k 2
Y T ⊥ + 1.7 I T ⊥ = Y (0)⊥ − Y (1)⊥ + Y (2)⊥ + 1.7 I T ⊥

T T T which yields
r r
  r 
α0 β0 1 0.25 X  2 1
≤ + + 0.11 ≤ . kU T Ỹ k2F = U T Ỹ ≤ , (27)

60 30 r r k 2 2700r2
k=1
 10

2
with probability at least 1 − e−γm/r , when m ≥ cnr2 and with E[d2k,j ] = 81. Note 2
that dk,j ’s are i.i.d. sub-exponential
|S| ≤ c1 m. 2
random variables with dk,j ≤ K, for some constant K,
To bound the second term in (26), we could adopt the same ψ1
′ then based on [34, Corollary 5.17],
techniques as before. The kth column of Ỹ can be expressed   
explicitly as X   
ǫ2 |S|

2 2


P   dk,j − Edk,j  ≥ ǫ |S| ≤ 2exp −c1
 ,
K2
 ′
Ỹ
 
j∈S 
k
which indicates that if |S| = cm/r, for some constant c,
" r  #
1 X 1 X  T 2 β0 − α0
= a ui 1{|aT ui |≤3} − α0 + 2
m r i=1 j j r kdk k2 ≤ (81 + c1 ) |S| ≤ 82 |S| := δ0 |S|
j∈S ⊥
holds with probability at least 1 − e−γm/r
 
· (aTj uk ) I − U U T aj . And
2
for a fixed
vector x ∈ R|S| obeying kxk2 = 1, Φ̄x 2 is also a chi-
1 X 1 square random variable with r degrees of freedom, so
:= ck,j a′j := Ψck ,
m ⊥
m 2 m
j∈S Φ̄x ≤
2
,
2700δ0cr2
where Ψ ∈ Rn×|S | is constructed by a′j ’s, each of which,
⊥
2
with probability at least 1 − e−γm/r , provided m ≥ c1 nr2
as a reminder, is the projection of aj onto the orthog-
for some sufficiently large constant c1 . Thus we have
onal
 complement
 of the column space
 of U as a′j = 2
 2 1 dk 1
I − U U T aj . Equivalently, Ψ = I − U U T A, where

U T Ȳ = 2 Φ̄ kdk k22 ≤ ,

k 2 m kdk k2 2700r3
A ∈ Rn×|S | is constructed by aj ’s, j ∈ S ⊥ . For a fixed
⊥
2

vector x ∈ R|S | obeying kxk2 = 1, we have kΨxk2 =

⊥ 2 which gives

  2 r
 
2
2 2

I − U U T Ax
≤ kAxk2 , where kAxk2 is distributed
X 1
kU T Ȳ k2F =
U T Ȳ
≤



, (30)



2 2700r2
as a chi-square random variable with n degrees of freedom. k 2
k=1
Again [35, Lemma 1] tells us 2
with probability at least 1 − e−γm/r , when m ≥ c1 nr2 and
2 2 m |S| = cm/r, for some appropriate constants c and c1 .
kΨxk2 ≤ kAxk2 ≤ , ′
12000r2 (2)
Now consider the second term kȲ k2F in kY T k2F , where
′
2
with probability exceeding 1 − e−γm/r , provided m ≥ cnr2 the kth column of Ȳ can be expressed explicitly as
for a sufficiently large constant c. Hence,
 
′ 1 X  
Ȳ = 9χj (aTj uk ) I − U U T aj

2 k m
j∈S

 ′ 
2

1
ck

kck k2 ≤ 1

Ỹ
≤ 2
Ψ , 1 X 1




2
k 2 m kck k2
2 2700r3 := dk,j a′j := Ψ̄dk ,
m m
j∈S
which leads to
n×|S| ′

′
2 r
 
2 where Ψ̄ ∈ R  by aj ’s. Also, we can
is constructed
X ′ 1
Ỹ =
Ỹ
≤ . (28) decompose Ψ̄ as Ψ̄ = I − U U T Ā, where Ā ∈ Rn×|S| is





2700r2



F k 2
k=1 |S|
constructed by aj ’s, j ∈ S. For a fixed

vector x ∈ R obey-

2
Then, combining (27) and (28), we know that
2
ing kxk2 = 1, we have
Ψ̄x
2 =
I − U U T Āx
≤




2

Āx
2 , where
Āx
2 is a chi-square random variable with
r

2
′
2







(0)


(1)
1

Y T − Y T
=
U T Ỹ
+ 2
Ỹ
≤ . (29) 2 2





F F F 30r n degrees of freedom as well. Since we already know that
provided m ≥ c1 nr2 for a sufficiently large constant c1 ,
(2)
Next, let’s check kY T k2F , which can be written as m

Ψ̄x
2 ≤
Āx
2 ≤





2 2
,
(2) ′ 2700δ0cr2
kY T k2F = kU T Ȳ k2F + 2kȲ k2F , 2
with probability exceeding 1 − e−γm/r , we can have
′
where Ȳ = Y (2) U and Ȳ = (I − U U T )Ȳ . For the first

2
term kU T Ȳ k2F , the kth column of U T Ȳ can be formulated

 
2
′ 1
dk

kdk k22 ≤ 1

Ȳ
≤ 2
Ψ̄ ,



explicitly as k 2 m
kdk k2 2

2700r3
  1 X
  1 and a further result
U T Ȳ = 9χj aTj uk U T aj := Φ̄dk ,

2 r
 
2 1
m m
′
′
X
k

Ȳ
=
Ȳ
≤ , (31)



j∈S
F k 2 2700r2
k=1
where Φ̄ ∈ Rr×|S| is constructed by U T aj ’s, and dk ∈ R|S| which, combining with (30), leads to
is composed of dk,j ’s, each one expressed as r





(2)


2

2

′
1

Y T
=
U T Ȳ
+ 2
Ȳ
≤ . (32)



dk,j = 9χj aTj uk ,


F F F 30r

Finally we can obtain that if m ≥ cnr2 and |S| = c1 m/r,
for some constants c and c1 , with probability at least 1 −
2 contemporary overview,” Signal Processing Magazine, IEEE, vol. 32,
e−γm/r ,
1 [7] Y. Chen, Y. Chi, and A. Goldsmith, “Exact and stable covariance esti-
(0) (1) (2)
kY T kF =
Y T − Y T + Y T
≤ . (33) mation from quadratic sampling via convex programming,” Information
F 15r
D. Proof of Theorem 1
The required restricted isometry properties of the linear
mapping A are supplied in Section V-B and a valid appropriate
dual certificate is constructed in Section V-C, therefore, The-
orem 1 can be straightforwardly obtained from the Lemma 1
in Section V-A.
VI. C ONCLUSION
In this paper, we address the problem of estimating a low-
rank PSD matrix X ∈ Rn×n from rank-one measurements
that are possibly corrupted by arbitrary outliers and bounded
noise. This problem has many applications in covariance
sketching, phase space tomography, and noncoherent detection
in communications. It is shown that with an order of nr2
random Gaussian sensing vectors, a PSD matrix of rank-r
can be robustly recovered by minimizing the ℓ1 -norm of the
observation residual within the semidefinite cone with high
probability, even when a fraction of the measurements are
adversarially corrupted. This convex formulation eliminates
the need for trace minimization and tuning of parameters
without prior knowledge of the outliers. Moreover, a non-
convex subgradient descent algorithm is proposed with excel-
lent empirical performance, when additional information of
the rank of the PSD matrix is available. For future work,
it would be interesting to theoretically justify the proposed
non-convex algorithm. Finally, we note that very recently one
of the authors proposed a median-truncated gradient descent
algorithm for phase retrieval under a constant proportion of
outliers with provable performance guarantees in [36], which
might be possible to extend to the problem of robust low-
rank PSD matrix recovery considered in this paper and will
be pursued elsewhere.
ACKNOWLEDGEMENT
We thank the anonymous reviewers for their valuable sug-
gestions that greatly improved the quality of this paper.
R EFERENCES
[1] J. R. Fienup, “Reconstruction of an object from the modulus of its fourier
transform,” Optics letters, vol. 3, no. 1, pp. 27–29, 1978.
[2] E. J. Candes, T. Strohmer, and V. Voroninski, “Phaselift: Exact and
stable signal recovery from magnitude measurements via convex pro-
gramming,” Communications on Pure and Applied Mathematics, vol. 66,
no. 8, pp. 1241–1274, 2013.
[3] E. J. Candes, Y. C. Eldar, T. Strohmer, and V. Voroninski, “Phase
retrieval via matrix completion,” SIAM Journal on Imaging Sciences,
vol. 6, no. 1, pp. 199–225, 2013.
[4] I. Waldspurger, A. d’Aspremont, and S. Mallat, “Phase recovery, maxcut
and complex semidefinite programming,” Mathematical Programming,
vol. 149, no. 1-2, pp. 47–81, 2015.
[5] P. Schniter and S. Rangan, “Compressive phase retrieval via generalized
approximate message passing,” Signal Processing, IEEE Transactions
on, vol. 63, no. 4, pp. 1043–1055, 2015.
11
[6] Y. Shechtman, Y. C. Eldar, O. Cohen, H. N. Chapman, J. Miao,
and M. Segev, “Phase retrieval with application to optical imaging: a
no. 3, pp. 87–109, 2015.
Theory, IEEE Transactions on, vol. 61, no. 7, pp. 4034–4059, July 2015.
[8] L. L. Scharf, Statistical signal processing. Addison-Wesley Reading,
1991, vol. 98.
[9] A. Lakhina, K. Papagiannaki, M. Crovella, C. Diot, E. D. Kolaczyk,
and N. Taft, “Structural analysis of network traffic flows,” in ACM
SIGMETRICS Performance evaluation review, vol. 32, no. 1. ACM,
2004, pp. 61–72.
[10] D. Gross, Y.-K. Liu, S. T. Flammia, S. Becker, and J. Eisert, “Quantum
state tomography via compressed sensing,” Physical review letters, vol.
105, no. 15, p. 150401, 2010.
[11] D. D. Ariananda and G. Leus, “Compressive wideband power spectrum
estimation,” Signal Processing, IEEE Transactions on, vol. 60, no. 9,
pp. 4775–4789, 2012.
[12] H. Kim, A. M. Haimovich, and Y. C. Eldar, “Non-coherent direction of
arrival estimation from magnitude-only measurements,” Signal Process-
ing Letters, IEEE, vol. 22, no. 7, pp. 925–929, 2015.
[13] E. Mason, I.-Y. Son, and B. Yazici, “Passive synthetic aperture radar
imaging using low-rank matrix recovery methods,” Selected Topics in
Signal Processing, IEEE Journal of, vol. 9, no. 8, pp. 1570–1582, 2015.
[14] E. J. Candes and X. Li, “Solving quadratic equations via phaselift
when there are about as many equations as unknowns,” Foundations
of Computational Mathematics, vol. 14, no. 5, pp. 1017–1026, 2014.
[15] L. Demanet and P. Hand, “Stable optimizationless recovery from phase-
less linear measurements,” Journal of Fourier Analysis and Applications,
vol. 20, no. 1, pp. 199–221, 2014.
[16] P. Hand, “Phaselift is robust to a constant fraction of arbitrary errors,”
Applied and Computational Harmonic Analysis, 2016.
[17] M. Kabanava, R. Kueng, H. Rauhut, and U. Terstiege, “Stable
low-rank matrix recovery via null space properties,” arXiv preprint
arXiv:1507.07184, 2015.
[18] B. Recht, M. Fazel, and P. A. Parrilo, “Guaranteed minimum-rank
solutions of linear matrix equations via nuclear norm minimization,”
SIAM review, vol. 52, no. 3, pp. 471–501, 2010.
[19] W. Dai, O. Milenkovic, and E. Kerman, “Subspace evolution and
transfer (set) for low-rank matrix completion,” Signal Processing, IEEE
Transactions on, vol. 59, no. 7, pp. 3120–3132, 2011.
[20] M. Wang, W. Xu, and A. Tang, “A unique “nonnegative” solution to an
underdetermined system: From vectors to matrices,” Signal Processing,
IEEE Transactions on, vol. 59, no. 3, pp. 1007–1016, 2011.
[21] E. J. Candès, X. Li, Y. Ma, and J. Wright, “Robust principal component
analysis?” Journal of ACM, vol. 58, no. 3, pp. 11:1–11:37, Jun 2011.
[22] V. Chandrasekaran, S. Sanghavi, P. A. Parrilo, and A. S. Willsky,
“Rank-sparsity incoherence for matrix decomposition,” SIAM Journal
on Optimization, vol. 21, no. 2, pp. 572–596, 2011.
[23] J. Wright, A. Ganesh, K. Min, and Y. Ma, “Compressive principal
component pursuit,” Information and Inference, vol. 2, no. 1, pp. 32–68,
2013.
[24] X. Li, “Compressed sensing and matrix completion with constant
proportion of corruptions,” Constructive Approximation, vol. 37, pp. 73–
99, 2013.
[25] G. Mateos and G. B. Giannakis, “Robust pca as bilinear decomposition
with outlier-sparsity regularization,” Signal Processing, IEEE Transac-
tions on, vol. 60, no. 10, pp. 5176–5190, 2012.
[26] E. J. Candès, X. Li, and M. Soltanolkotabi, “Phase retrieval via wirtinger
flow: Theory and algorithms,” Information Theory, IEEE Transactions
on, vol. 61, no. 4, pp. 1985–2007, 2015.
[27] Y. Chen and E. J. Candès, “Solving random quadratic systems of
equations is nearly as easy as solving linear systems,” arXiv:1505.05114,
May 2015.
[28] C. D. White, R. Ward, and S. Sanghavi, “The local convexity of solving
quadratic equations,” arXiv preprint arXiv:1506.07868, 2015.
[29] Y. I. Abramovich, D. A. Gray, A. Y. Gorokhov, and N. K. Spencer,
“Positive-definite toeplitz completion in doa estimation for nonuniform
linear antenna arrays. i. fully augmentable arrays,” Signal Processing,
IEEE Transactions on, vol. 46, no. 9, pp. 2458–2471, 1998.
[30] H. Qiao and P. Pal, “Generalized nested sampling for compressing low
rank toeplitz matrices,” IEEE Signal Processing Letters, vol. 22, no. 11,
pp. 1844–1848, 2015.
[31] D. Romero, D. D. Ariananda, Z. Tian, and G. Leus, “Compressive co-
variance sensing: Structure-based compressive sensing beyond sparsity,”
IEEE Signal Processing Magazine, vol. 33, no. 1, pp. 78–93, 2016.
 12

[32] D. Romero, R. López-Valcarce, and G. Leus, “Compression limits for

random vectors with linearly parameterized second-order statistics,”
IEEE Transactions on Information Theory, vol. 61, no. 3, pp. 1410–
1425, 2015.
[33] Q. Zheng and J. Lafferty, “A convergent gradient descent algorithm for
rank minimization and semidefinite programming from random linear
measurements,” in Advances in Neural Information Processing Systems,
2015, pp. 109–117.
[34] R. Vershynin, “Introduction to the non-asymptotic analysis of random
matrices,” Compressed Sensing, Theory and Applications, pp. 210 – 268,
2012.
[35] B. Laurent and P. Massart, “Adaptive estimation of a quadratic functional
by model selection,” Annals of Statistics, pp. 1302–1338, 2000.
[36] H. Zhang, Y. Chi, and Y. Liang, “Provable non-convex phase retrieval
with outliers: Median truncated wirtinger flow,” in International Con-
ference on Machine Learning (ICML), New York, NY, 2016.