You are on page 1of 9

5060 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO.

8, AUGUST 2023

Almost Tight Sample Complexity Analysis of


Quantum Identity Testing by
Pauli Measurements
Nengkun Yu

Abstract— This paper studies the quantum identity testing TABLE I


problem, a quantum analogue of distribution identity testing. T YPICAL Q UANTUM S AMPLING M ETHODS
The goal is to determine whether a quantum state is identi-
cal to another fixed quantum state, using as few state sam-
ples as possible. Rather than general entangled measurements,
we consider the less powerful but experimentally friendly Pauli
measurements. We follow the standard setting where Pauli
measurements are regarded as two-outcome measurements, i.e.,
each n-qubit Pauli measurement has a one-bit outcome. We prove [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22],
that for an n-qubit quantum system,n
the sample complexity of [24], [25], [26], [27], [28], [29], [30], [31], [32]. Here the goal
this problem is Θ(poly(n) · 4ϵ2 ) if only Pauli measurements
are allowed. In other words, we provide simple algorithms to is to settle the sample complexity (optimal number of samples)
determine whether two n-qubit quantum states, ρ and σ, are as a function of the parameters.
identical or ϵ-far in trace distance usingn Pauli measurements, The equality, or identity, of distributions with support size
using O(min{n4 , n3 + log3 (1/ϵ)} · 4ϵ2 ) copies of ρ and σ. d is a central problem in distribution testing. It is frequently
n
Interestingly, O( 4ϵ2 ) copies are not sufficient under this setting. revisited with different approaches due to its significance
Index Terms— Quantum identity testing, Pauli measurements, in the research. In an important work [3], Goldreich √ and
sample complexity. Ron found that the ℓ2 norm can be estimated from O( ϵ2d )
samples. This finding led to an algorithm for uniformity
testing, i.e., to determine
√ whether a probability distribution
I. I NTRODUCTION
is a uniform using O( ϵ4d ) samples. Paninski [33] and Valiant

T ESTING whether observed data conforms to an under-


lying model is a fundamental statistical problem known
as hypothesis testing [1], [2]. The most basic form of hypoth-
and√Valiant [34] showed that the complexity of uniformity is
Θ( ϵ2d ). If one distribution was an arbitrary known distribution,
Batu et al. [4], [35] presented an ℓ2 -identity
√ tester and used
esis testing is binary hypothesis testing, consisting of a null
hypothesis H0 and H1 . The goal is to design an algorithm it to build an ℓ1 estimator using O( dϵlog 2
d
) samples; later
that takes a random sample from some distribution P ∈ in [36], Valiant and√Valiant showed the sample complexity of
H0 ∪H1 and, with a high probability, determines whether P is this problem is Θ( ϵ2d ). If both distributions were unknown,
2/3
in H0 and H1 . Hypothesis tests formalize yes-or-no questions Batu et al. provided a tester in [4] using O( d ϵ8/3 log d
) samples;
about an underlying population, given a random sample from In 2014, Chan et al., in [37],√ showed the complexity of the
2/3
that population. This type of test is ubiquitous in physical, identity testing is Θ(max( ϵ2d , dϵ4/3 )).
life, and social science. Since the seminal work of Goldreich To achieve efficient and reliable control of quantum devices,
and Ron [3] and Batu et al. [4], significant efforts from the testing the properties of new devices is of fundamental impor-
theoretical computer science, information theory, and statistics tance. Testing for properties of distributions has been gener-
community have been devoted to the non-asymptotic (small- alized for quantum computing [38]. Here, general quantum
sample) regime, resulting in a better understanding in various property testing aims to derive a measurement scheme that
distribution testing problems [5], [6], [7], [8], [9], [10], [11], allows for learning non-trivial properties using the fewest
possible copies.
Manuscript received 18 May 2022; revised 15 November 2022; accepted
13 April 2023. Date of publication 27 April 2023; date of current version There is a significant difference between quantum property
14 July 2023. This work was supported in part by the Australian Research testing and distribution testing. Each sample is an output
Council (ARC) Discovery Early Career Researcher Award DE180100156 and with a classical index in distribution testing according to the
in part by the ARC Discovery Program under Grant DP210102449.
The author is with the Department of Computer Science, Stony Brook Uni- probability distribution. However, the sampling methods have
versity, Stony Brook, NY 11794 USA, and also with the Centre for Quantum much richer structures with quantum property testing. See
Software and Information, University of Technology Sydney, Sydney, NSW Table I for details.
2007, Australia (e-mail: nengkunyu@gmail.com).
Communicated by S. Beigi, Associate Editor for Quantum. Among the many available sampling methods for quantum
Digital Object Identifier 10.1109/TIT.2023.3271206 property testing, given a fixed number of copies (say m of
0018-9448 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:47:32 UTC from IEEE Xplore. Restrictions apply.
YU: ALMOST TIGHT SAMPLE COMPLEXITY ANALYSIS OF QUANTUM IDENTITY TESTING 5061

the d-dimensional state ρ), the three measurements in Table I entangled measurements. The downside of these entangled
are of particular interest, i.e., joint measurement, independent measurements is the difficulty in their implementations. Let
measurement, and the local measurement. Joint measurement, us consider an n-qubit system, i.e., d = 2n . The algorithm
the most general, allows arbitrary measurements of dimension in [51] requires that all Θ( ϵd2 ) copies be stored in a noiseless
dm . Independent measurement only allows measurements on environment. Furthermore, for an n-qubit system that is d =
d
each copy of ρ, which results in n measurements of size d. 2n , then measurement needed is within dimension dΘ( ϵ2 ) =
For an n-qubit system, d = 2n , local measurements apply to n·2
2Θ( ϵ2 ) , a double exponential function of the number of
each qubit, i.e., a dimension of 2. qubits. The independent measurement result of [49] requires
Quantum state tomography solves the general quantum the lab to perform an n-qubit random unitary, although unitary
property testing problem immediately. The problem of quan- 4-design is enough, which is also highly entangled.
tum state tomography asks how many copies of an unknown As a significant class of local measurements, Pauli measure-
d-dimensional quantum state ρ are necessary and sufficient to ments are the most widely implemented among all quantum
generate an output of a good approximation of ρ in trace dis- measurements, partly because they are the most experimentally
tance, with high probability. A sequence of work that involved friendly. The standard setting of Pauli measurements is as
performing non-adaptive measurements on each copy of ρ follows: each Pauli measurement, described as a Pauli matrix
3
independently showed that Θ( dϵ2 ) copies are both sufficient ⊗n ⊗n
P , is defined as { I 2+P , I 2−P }. In other words, Pauli mea-
and necessary for quantum state tomography in a trace distance surements are two-outcome measurements. Significant efforts
of no more than ϵ [39], [40]. If it is possible to measure many have been devoted to studying quantum systems under this
copies of ρ simultaneously (joint measurements), the sample setting, including the quantum state tomography problem
2
complexity of state tomography is Θ( dϵ2 ) in [40], [41], and via compressed sensing [52], [53], [54], and quantum state
[42]. [43] proves that the sample complexity of quantum state certification [55], [56], [57].
tomography using adaptive measurements on each copy of ρ is
3
Θ( dϵ2 ). Reference [44] provides information-theoretic bounds
A. Our Results
on quantum advantage in machine learning. Reference [45]
studies the power of quantum memory for learning properties Given the importance of Pauli measurements, we study
of quantum systems and dynamics. Reference [47] studies the the following quantum identity testing problem using Pauli
quantum advantage in learning from experiments. measurements.
The “quantum identity testing” problem is of considerable Problem 1: Given two unknown n-qubit mixed states ρ and
interest. This problem, a quantum analogue of the identity σ, and ϵ > 0, the goal is to distinguish between the two cases
testing of probabilistic distributions, is as follows: Given many ||ρ − σ||1 > ϵ (1)
copies of unknown quantum states ρ and σ, the goal is to
distinguish the case that they are identical and the case that and
they are ϵ-far in trace distance. If one state is known, this
problem is called “quantum state certification”. If one state is ρ = σ. (2)
known to be a maximally mixed state, a quantum analogue of How many copies of ρ and σ are needed to achieve this goal
uniform distribution, this is known as the quantum mixedness with high probability?
problem. We show that:
For the more general mixed states case, [48] solved the Theorem 1: The quantum identity testing problem can be
n
problem in the joint measurement setting where σ is a max- solved using O(min{n4 , n3 + log3 (1/ϵ)} · 4ϵ2 ) copies of ρ
imally mixed state case by showing that Θ( ϵd2 ) copies are and σ via Pauli measurements. Under the nonadaptive setting,
n
necessary and sufficient where an arbitrary measurement is at least Ω( n·4
ϵ2 ) copies of ρ are needed for solving the
allowed to apply on the collection of all copies of the states. quantum identity testing problem.
Interestingly, work by [49] showed the sample complexity of Remark: The upper bound can be improved to
n
this problem in the independent measurement setting, where σ O(min{n3+µ , n3 + log3 (1/ϵ)} · 4ϵ2 ) for any µ > 0. See
3/2
is a maximally mixed state case, is Θ( dϵ2 ) with nonadaptive Subsection III-A for details.
4/3
measurements, and Ω( dϵ2 ) with adaptive measurements. [46] Problem 2: Given an unknown n-qubit mixed state σ and
3/2
showed Θ( dϵ2 ) is also tight with adaptive measurements. ϵ > 0, the goal is to distinguish between the two cases
Reference [23] also studied quantum state certification in I2n
the independent measurements setting. This idea implies an ||σ − ||1 > ϵ (3)
3/2
2n
independent measurement scheme using Θ( dϵ2 ) copies for and
the general ρ and σ. Moreover, the results of “quantum identity
testing” were used to study “quantum independence testing”, I2n
σ=
. (4)
i.e., whether a given state is in tensor product form, as shown 2n
in [50]. The sample complexity of the “quantum identity How many copies of σ are needed to achieve this goal, with
testing” problem was proven to be Θ( ϵd2 ) in [51]. high probability?
To achieve the optimal complexity in the generally mixed The sample complexity of Problem 2 is no more than the
state scenario, [40], [41], [42], [48], [51] employ highly sample complexity of Problem 1.

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:47:32 UTC from IEEE Xplore. Restrictions apply.
5062 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 8, AUGUST 2023

TABLE II If the state of a quantum system was ρ immediately before


R ESULTS OF Q UANTUM M IXEDNESS AND Q UANTUM I DENTITY T ESTING measurement {Mi } was performed, the probability of that
result i occurring is
p(i) = Tr(Mi ρ).

P
A Hermitian O = j λj |ψj ⟩⟨ψj | (that is O = O ) on H
with orthonormal basis |ψj ⟩ and λj ∈ R always corresponds
to the following measurement protocol. Suppose the state of
a quantum system is ρ when measured in basis |ψj ⟩⟨ψj |.
We list the known results (joint and independent mea- If the outcome is j, the observed result is λj , and the output
surements) and our results (Pauli measurements) in Table II. corresponds directly to a random variable X such that
To our knowledge, this problem has not been studied under
the Pauli measurement setting. A related question is studied p(X = λj ) = Tr(ρ|ψj ⟩⟨ψj |).
in [44], [45], and [47]: predicting the absolute value of just For U ∈ {X, Y, Z} with U = |ψ0 ⟩⟨ψ0 | − |ψ1 ⟩⟨ψ1 |,
a single observable requires exponentially many copies in the we use the Pauli measurement to denote the corresponding
⊗n ⊗n
conventional scenario where ρ = I2n or ρ = I 2±0.9P n and measurement,
the observable is a Pauli matrix.
M0 = |ψ0 ⟩⟨ψ0 |, M1 = |ψ1 ⟩⟨ψ1 |.

II. P RELIMINARIES For an n-qubit Pauli matrix P = P1 P2 · · · Pn ∈


{I, X, Y, Z}⊗n with P ̸= I ⊗n , we can always find orthogonal
A. Basic Quantum Mechanics projections Q+ and Q− , i.e., Q2+ = Q+ and Q2− = Q− , such
An isolated physical system is associated with a Hilbert that P = Q+ − Q− and Q+ + Q− = I ⊗n . Furthermore, the
⊗n ⊗n
space is called the state space. A pure state of a quantum two outcome measurement {Q+ , Q− } = { I 2+P , I 2−P }
system is a normalized vector in its state space, denoted by can be implemented through the one-qubit Pauli measurements
the Dirac notation |φ⟩. A mixed state is represented by a corresponding to P1 , P2 , · · · , Pn .
density operator on the state space. Here, a density operator ρ
on d-dimensional Hilbert space Cd is a semi-definite positive D. ℓ1 Distance
linear operator such that Tr(ρ) = 1. We let We use the ℓ1 distance to describe the difference between
d
D(C ) = {ρ : ρ is d−dimensional density operator of C } d quantum states. The ℓ1 distance between quantum states ρ and
σ is defined as
denote the set of quantum states.
||ρ − σ||1 ≡ Tr|ρ − σ|
We use I, X, Y , and Z to denote the Pauli matrices, the √
most important one-qubit matrix, where |A| ≡ A† A is the positive square root of A† A.

1 0
 
0 1
 
1 0
 
0 i
 Given a general operator A, the ℓ1 norm is defined as
I= ,X = ,Z = ,Y = .
0 1 1 0 0 −1 −i 0 ||A||1 = Tr|A|.
d Id Their ℓ2 distance is defined as
Id denotes the identity operator of D(C ), and d denotes the
maximally mixed states of D(Cd ).
p
||ρ − σ||2 = Tr(ρ − σ)2 .
For ρ, σ ∈ D(Cd ), we have the following relation between
B. The Tensor Product of Hilbert Space ℓ1 and ℓ2 distances,
The state space of a composed quantum system is the tensor √
dk ||ρ − σ||2 ≤ ||ρ − σ||1 ≤ d||ρ − σ||2 .
product of the state spaces of its component systems.
NnLet Cdk
be a Hilbert space. One can define a Hilbert space k=1 C The observation outlined below is well-known,
as the tensor product of Hilbert spaces Cdk . The following Observation 1: For two unknown binary distributions p and
notation
Nn denotes the quantum state on the multipartite system q, the sample complexity of distinguishing the two cases p = q
dk
k=1 C , and ||p−q||2 > ϵ with probability at least 1−δ is Θ( log(1/δ) ).
ϵ2
Πn If we replace the ℓ2 norm (the Euclidean norm) by the ℓ1 norm
D(⊗ni=1 Cdi ) = D(C i=1 di ).
(the total variation) in the above observation, the complexity
will not change because the √ ℓ2 norm and the ℓ1 norm are
C. Quantum Measurement equivalent up to a constant 2.
A positive-operator valued measure (POVM) is a measure III. Q UANTUM I DENTITY T ESTING
where the values are non-negative self-adjoint operators in
In the first part of this section, we provide a measurement
a Hilbert space Cd , which is described by a collection of 4 n n 3
scheme that uses O( n ϵ·4 ) or O( (4 ·[n+log(1/ϵ)] ) copies to
matrices {Mi } with Mi ≥ 0 and 2 ϵ2
X solve Problem 1. The main tool is an identity tester for
Mi = Id . probability collections. In the second part, we show a lower
n
i bound Ω( n·4ϵ2 ) for the same problem.

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:47:32 UTC from IEEE Xplore. Restrictions apply.
YU: ALMOST TIGHT SAMPLE COMPLEXITY ANALYSIS OF QUANTUM IDENTITY TESTING 5063

A. Upper Bound In other words, the probability of success becomes very high
The following property testing problem of collections of for a sufficiently
Pmlarge constant L.
1 2
discrete distributions in the query model is investigated as a Case 2: m i=1 ||p i − qi ||2 > µ2 .
means to solving Problem 1. In this case, the correct output is “No”. We can bound the
Problem 3: Suppose we are given two collections of binary error probability, i.e., the probability of output “Yes”.
distributions in a query model, with m > 1 binary distributions In the following, we use |A| to denote the cardinality of
p1 , . . . , pm for the first collection and q1 , . . . , qm for the set A.
second collection, such that for any given index, we can First, we observe that
choose to access pi and qi and obtain samples. The goal is to 1 X
m
distinguish the case of pi = qi for all i from the case ||pi − qi ||22
m i=1
m
1 X 2 2
||pi − qi ||22 > µ2 . |{i : ||pi − qi ||22 < µ2 }| µ2
m i=1 <
m
4 X |{i : 2k−1 µ2 ≤ ||pi − qi ||2 < 2k µ2 }|2k µ2
2
Lemma 1: Problem 3 can be solved using O( logµ2 m ) +
m
queries. k≥0
Proof: We derive Algorithm 1 to solve this problem. 2 X |{i : 2k−1 µ2 ≤ ||pi − qi ||2 }|2k µ2
µ 2
< + .
2 m
k≥0
Algorithm 1 An Identity Test for Collections with a
Query Model Thus,
1 Access to binary distributions p1 , . . . , pm and q1 , . . . , µ2 X |{i : 2k−1 µ2 ≤ ||pi − qi ||22 }|2k µ2
qm with µ > 0; +
2 m
2 Output “Yes” with a probability of at least 32 if k≥0
2
pi = qi for all 1 ≤ i ≤ m; andP“No” with a >µ
m
probability of at least 23 if m 1 2
i=1 ||pi − qi ||2 > µ ;
2
µ2 X 1
Let L be a sufficiently large constant; > + · 2k µ2 ,
3 2 100(k + 1)2k
2
k≥0
4 for k ← 0 to ⌈log2 m⌉ do
5 Select 2k · (k 2 + 1) · L uniformly random elements, where we use the following
without replacement, in [m];
6 For each selected i, distinguish between pi = qi
X 1 1 X 1 1
< + < .
and ||pi − qi ||22 > 2k−1 µ2 with a failure 100(k 2 + 1) 100 100k 2 2
k≥0 k≥1
probability not greater than L−2 6−k ;
7 If any of these testers returned “No”, return “No”; Therefore, there exists some k ≥ 0 such that
Return “Yes”; 1
8
|{i : 2k−1 µ2 ≤ ||pi − qi ||22 }| > m.
2k · 100(k 2 + 1)

We first show that this algorithm solves Problem 3 before There is always such a k ≤ k0 = log2 m. This follows from
counting its sample complexity. the ensuing arguments. If we find some k > k0 with the above
Case 1: pi = qi for all 1 ≤ i ≤ m. In this case, the correct property, then the above property is also true for k0 :
output is “Yes”. We can bound the error probability, i.e., the 1
probability of output “No”. |{i : 2k−1 µ2 ≤ ||pi − qi ||22 }| > m>0
2k · 100(k 2 + 1)
Using union bound, the probability of output “No” is at
⇒|{i : 2k−1 µ2 ≤ ||pi − qi ||22 }| ≥ 1
most
⌈log2 m⌉
⇒|{i : 2k0 −1 µ2 ≤ ||pi − qi ||22 }| ≥ |{i : 2k−1 µ2
≤ ||pi − qi ||22 }| ≥ 1.
X
qk ,
k=0
The last inequality is due to k > k0 , then each
where qk denotes the probability of output “No” in the k-th
iteration of the loop. {i : 2k−1 µ2 ≤ ||pi − qi ||22 } ⊆ {i : 2k0 −1 µ2 ≤ ||pi − qi ||22 }.
For any k, using union bound, qk is at most
According to the definition of k0 ,
k 2 k2 + 1
−2 −k
qk ≤ 2 · (k + 1) · L · L 6 = k . 1
3 L 1≥ m.
2k0 · 100(k02 + 1)
Therefore, the error probability of this case does not exceed
That is
⌈log2 m⌉ ⌈log2 m⌉
X X k2 + 1 1 1
qk ≤ k
≤ O( ). |{i : 2k0 −1 µ2 ≤ ||pi − qi ||22 }| ≥ m.
3 L L 2k0 · 100(k02 + 1)
k=0 k=0

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:47:32 UTC from IEEE Xplore. Restrictions apply.
5064 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 8, AUGUST 2023

In the k-th iteration of the loop, the probability of selecting Then,


some i with this property is at least
− βP ) 2 ||ρ − σ||21
P
1 P (αP
1 − (1 −
k 2
)2 ·(k +1)·L ≥ 1 − O(e−L/100 ). ||ρ − σ||22 = ≥ .
100 · 2k · (k 2 + 1) 2n 2n
For this case, the probability of output “Yes” is, The last inequality is identified according to the relation
at most, the probability of not returning “No” in the between ℓ1 distance and ℓ2 distance.
k-th iteration of the loop. By union bound, this probability is Therefore, if ρ = σ, αi = βi for all i.
not greater than the summation of If ||ρ − σ||1 > ϵ, we have
• the probability that the selected is in the k-th iteration do X
not satisfy (αP − βP )2 > ϵ2 .
1 i
|{i : 2k−1 µ2 ≤ ||pi − qi ||22 }| > k m.
2 · 100(k 2 + 1) If we measure ρ and σ for the Pauli operator P , the
and output of ρ would be a sample of probability distribution
• the failure probability of distinguishing between pi = qi pP = ( 1+α
2 ,
P 1−αP
2 ), and the output of σ would be a sample
and ||pi − qi ||22 > 2k−1 µ2 for selected i(s) in the k-th of probability distribution qP = ( 1+β
2 ,
P 1−βP
2 ). That is
iteration which satisfy
1 (αP − βP )2
|{i : 2k−1 µ2 ≤ ||pi − qi ||22 }| > k m. ||pP − qP ||22 = .
2 · 100(k 2 + 1) 2
That is, the error probability is at most Therefore, to solve the quantum identity testing problem,
ϵ2
O(e−L/100 ) + L−2 6−k ≪ 1. we only need to solve Problem 3 with m = 4n and µ2 = 2·4n.

According to Lemma 1, it can be solved using


Therefore, the corresponding tester will return “No” with a
high probability. n4 · 4n
Using observation 1, we can count the sample complexity O( )
ϵ2
of this algorithm as
k0 copies. □
X 1 Remark: In Lemma 1, we have not attempted to optimize
2k · (k 2 + 1) · L · O( · log(L2 · 6k ))
2k−1 µ2 the exponent of log m. We can easily obtain an improved
k=0 3+γ
k0 sample complexity of O( log µ2 m ) for any γ > 0 by selecting
X 1
= 2 · (k 2 + 1) · L · O( · (k + log L)) 2k · k 1+γ · L uniformly random elements 1 ≤ i ≤ m in
µ2
k=0 Algorithm 1.
k04 As a direct consequence, Problem 1 can be solved using
=O( ) 3+γ n
µ2 O( n ϵ2·4 ) copies.
log4 m
=O( ).
µ2
B. An Alternative Approach
Although, in the beginning, it seems that 2k0 · k02 >> m, this
term 2k0 · k02 appears only if 2k−1 µ2 < 1. □ It is interesting to know whether it is possible to obtain
Using Lemma 1, we can prove the upper bound of a bound independent of m and just dependent on γ for
Theorem 1 by a standard reduction of quantum identity testing Problem 3. We provide an alternative or more straightforward
to the problem of Bernoulli testing (Problem 3). Such a solution to Problem 3.
reduction can be found as Proposition 1, 2 in [58]: For The proof of this proposition employs the following lemma
(Fact A.1 in [59]).
P
P α P
a quantum state ρ = P
2n , the application of Pauli
Lemma 2: Let D be a probability distribution, q :
measurement P on ρ generates a Bernoulli distribution (p0 , p1 )
Supp(D) 7→ [0, 1], and µ ∈ (0, 1]. Suppose that Es∼D [q(s)] >
with p0 = 1+α2 , p1 =
P 1−αP
2 . µ, and let ℓ = ⌈log(2/µ)⌉. Then, there exists j ∈ [ℓ] such that
Proof: Suppose we are given two n-qubit quantum states
P rs∼D [q(s) > 2−j ] > 2j µ/4ℓ.
ρ and σ, and we guarantee that either ρ = σ or ||ρ − σ||1 > ϵ,
Proposition 1: Problem 3 can be solved using
and the goal is to distinguish between these two cases. 3

Without loss of generality, we can let O( (log µ(1/µ)


2 ) queries.
P Proof: Let D be a uniform distribution over the num-
αP P bers (pi − qi )2 : i = 1, . . . , m. Then Es∼D [s] = µ2 . Let
ρ= P n ,
P 2 ℓ = ⌈log(2/µ2 )⌉. By Lemma 2, we know that there exists
βP P j ∈ [ℓ] such that P ri∼D [||pi − qi ||22 > 2−j ] > 2j µ2 /4ℓ.
σ= Pn ,
2 Then we only need to go up to ⌈log(2/µ2 )⌉ levels in the
where in the summation P ranges over {I, X, Y, Z}⊗n , and iteration of Algorithm 1, and not log m, see the following
αP , βP ∈ C. algorithm.

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:47:32 UTC from IEEE Xplore. Restrictions apply.
YU: ALMOST TIGHT SAMPLE COMPLEXITY ANALYSIS OF QUANTUM IDENTITY TESTING 5065

Algorithm 2 An alternative Identity Test for Collec- According to Observation 1, to distinguish σP from I22nn ,
tions with a Query Model Ω( log(1/δ)
4ϵ2 ) copies are needed using Pauli measurement corre-
1 Access to binary distributions p1 , . . . , pm and q1 , . . . , sponding to P , to reach the probability of success of at least
qm with µ > 0; 1 − δ. In other words, the probability of success is at most
2 Output “Yes” with a probability of at least 32 if 1 − exp(−4mϵ2 ) using m copies.
pi = qi for all 1 ≤ i ≤ m; andP“No” with a In total, we have 4n − 1 different σP s with P ranging over
probability of at least 23 if m1 m 2 2 P ∈ {I, X, Y, Z}⊗n \ I2n .
i=1 ||pi − qi ||2 > µ ;
3 Let L be a sufficiently large constant; One crucial observation is that for Paulis P ̸= P ′ , the
4 for k ← 0 to ⌈log(2/µ2 )⌉ do measurement corresponding to P reveals no information about
5 Select 2k · (k 2 + 1) · L uniformly random elements P ′ . In other words, Pauli measurement that corresponds to P
in [m]; can not distinguish I22nn and σP .
6 For each selected i, distinguish between pi = qi Assume we are given σP with unknown P , we want to
and ||pi − qi ||22 > 2−k with a failure probability identify whether σP = I22nn .
not greater than L−2 6−k ; For a scheme to accomplish this task with a probability of
7 If any of these testers returned “No”, return “No”; at least 1 − δ by measuring each Pauli P ̸= I2n nP copies,
we know that
8 Return “Yes”;
ΠP (1 − exp(−4nP ϵ2 )) ≥ 1 − δ.
According to the inequality of arithmetic and geometric means,
The correctness analysis is similar to Lemma 1. We count we have
the sample complexity of this algorithm as

ΠP (1 − exp(−4nP ϵ2 ))
X 4ℓ 1 P
exp(−4nP ϵ2 ) 4n −1
· L × O( −k × log(L2 · 6k )) ≤[1 − P n ]
2k µ2 2 4 −1
k=0
ℓ3 X 4nP ϵ2 n
=O( ) ≤[1 − exp(− )]4 −1 .
µ2 4n − 1
P
log3 (1/µ)
=O( ). That is,
µ2 X 4nP ϵ2 n
□ [1−exp(− n
)] ≥ (1 − δ)1/(4 −1) ≈ 1 − δ/(4n − 1).
4 −1
This bound is not always better than Lemma 1, i.e., in the P
regime log4 m ≤ log3 (1/µ). Thus, we have
ϵ2
If we apply this proposition to Problem 1 with µ2 = 2×4 n, X 4nP ϵ2
we have the following theorem. ≥ Ω(n + log(1/δ)).
Theorem 2: Problem 1 can be solved using 4n − 1
P
(4n ·[n+log(1/ϵ)]3 (4n ·[n3 +log3 (1/ϵ)]
O( ϵ2 ) = O( ϵ2 ) samples by Therefore,
Pauli measurements. X (n + log(1/δ)) · 4n
Compared with the bound of the last subsection, this bound nP ≥ Ω( ).
ϵ2
has better dependence of n but worse dependence of ϵ. P


C. Lower Bound
The lower bound of Theorem 1 follows from the lemma D. Discussion on the Optimality of Lemma 1
below.
Lemma 1 is tight up to the poly(log m) factor: Even for the
Lemma 3: The sample complexity of Problem 2 is at least
n log( 1 )
Ω( (n+log(1/δ))·4
ϵ2 ) using nonadaptive Pauli measurements, case m = 2, ϵ2 δ samples are needed to achieve a probability
to reach the probability of success of at least 1 − δ of success 1 − δ.
Proof: By regarding Pauli measurement corresponding Corollary 1: Θ( µ12 ) queries are not sufficient to solve
to P = Q+ − Q− ∈ {I, X, Y, Z}⊗n with P ̸= I2n as a Problem 3.
two-outcome measurement {Q+ , Q− } with Q+ = I2n2+P and Proof: Assume Problem 3 can be solved using
n
Q− = I2n2−P , we observe that Ω( log(1/δ)4
ϵ2 ) copies of σ are 1
necessary to solve the quantum mixedness testing. To see this, Θ( )
µ2
for any P ∈ {I, X, Y, Z}⊗n , we let
queries. According to our proof of Theorem 1, Problem 1
n
I2n + 2ϵP can be solved using Θ( 4ϵ2 ) copies by choosing m = 4n and
σP =
2n ϵ2
µ2 = 2·4 n . This violates Lemma 3. □
It is clear that It is also interesting to compare Problem 3 with the fol-
I2n lowing property testing problem about collections of discrete
||σP − n ||1 = 2ϵ.
2 distributions in the query model studied in [60] and [61]: For

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:47:32 UTC from IEEE Xplore. Restrictions apply.
5066 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 8, AUGUST 2023

two collections of d-dimensional distributions with p1 , . . . , pm • If σ is ϵ-far from being independent, then ||σ − σ1 ⊗
for the first collection and q1 , . . . , qm for the second collection σ2 ||1 > ϵ/3. This can be proved by triangle inequality.
in a query model, distinguish the case of pi = qi for all i from We use induction to prove the upper bound for n-qubit inde-
4 k
the case pendence. Assume that L k ϵ·4 2 copies are enough for testing
m the independence of k-qubit system using Pauli measurements
1 X
||pi − qi ||1 > ϵ. for k < n.
m i=1
For n-qubit state σ, we choose k = n/2. We first test
√ 1/2 whether σ1,··· ,k and σk+1,··· ,n are both independent or ϵ/5-
The sample complexity is proved to be Θ(max{ ϵ2n , nϵ3/4 }),
far from independent with at least a 0.999 probability of
which does not depend on m. However, the sample complexity
successful. Then we test whether σ = σ1,··· ,k ⊗ σk+1,··· ,n
of Problem 3 does depend on m, as illustrated in Corollary 1.
or ϵ/5-far from independent with at least a 0.999 probability
of successful.
IV. A PPLICATION : I NDEPENDENCE T ESTING One can verify that
Recognizing the vital role of Problem 1 in quantum property • If σ is independent, we will pass all the tests with at least
testing, we study the applications of Theorem 1 in this section. a 2/3 probability of success.
General quantum property testing problems can be formu- • If σ is ϵ-far from being independent, then we will fail at
lated as follows: Given a subset P ⊊ D(Cd ), the ℓ1 distance least one of the tests with at least a 2/3 probability of
between ρ and P is defined as success.
4 n

||σ − P||1 = inf ||ρ − σ||1 . In total, we use O( n ϵ·4


2 ) copies of σ.
σ∈P To prove the lower bound, we assume there is an Algo-
If ||σ − P||1 > ϵ, we say that σ is ϵ-far from P. rithm A that uses f (n, ϵ) copies to decide, with at least a
P = {⊗ni=1 σi } with σi being single qubit states is of inter- 2/3 probability of success, whether a given n-qubit state σ
est here because it is the quantum counterpart of independent is independent or ϵ-far from being independent using Pauli
distributions. If ||σ − P||1 > ϵ, σ is called ϵ-far from being measurements. In the following, Algorithm A is applied as a
independent. If ||σ − P||1 = 0, σ = ⊗ni=1 σi is identified as black box to solve Problem 2.
independent.
This motivates us to investigate the following related Algorithm 4 A Mixedness Tester
problem. 1 Input: Θ[f (n, 6n ϵ
3
)] + Θ[ n ϵlog n
] copies of σ;
2
Problem 4: Given an unknown n-qubit mixed state σ and 2 Output “Yes” with a probability of at least 32 if
ϵ > 0, the goal is to distinguish between the case σ is ϵ-far σ = ⊗ni=1 I22 ; and “No” with a probability of at least
from being independent and the case σ is independent. How 2 n I2
many copies of σ are needed to achieve this goal, with high 3 if ||σ − ⊗i=1 2 ||1 > ϵ;
3 for i ← 1 to n do
probability?
4 Test whether σi = I22 or ||σi − I22 ||1 > 6n ϵ
with at
Using Theorem 1 and Lemma 3, we can show that 1
least a 1 − 6n probability of success;
Corollary 2: The sample complexity of Problem 4 is
4 n 5 if No then
O( n ϵ·4
2 ), using nonadaptive Pauli measurements. Moreover,
n 6 Return “No”;
the sample complexity is at Ω( n·4 1
ϵ2 ) for ϵ < O( ϵ ) using
nonadaptive Pauli measurements. 7 Run Algorithm A on σ to test whether it is
Proof: To prove the upper bound, we first use the ϵ
independent or it is 6n -far from being independent
Algorithm 3 for testing the independence of the bipartite 1
with at least a 1 − 1000 probability of success;
system. 8 if No then
9 Return “No”;
Algorithm 3 Independence Testing using Pauli 10 Return “Yes”;
Measurements
5 n
1 Input: O( n ϵ·4
2 ) copies of σ ∈ D(Cd1 ⊗ Cd2 );
2 Output “Yes” with a probability of at least 23 if σ is Algorithm 4 solves Problem 2 with a high probability
independent; and “No” with a probability of at least through the following arguments.
2 If σ = ⊗ni=1 I22 , the probability of outputting “Yes” is at
3 if σ is ϵ-far from being independent;
3
4 n
Use O( n ϵ·4 ) copies of σ to generate σ1 ; least
2
n4 ·4n 1 n 1 2
4 Use O( ϵ2 ) copies of σ to generate σ2 ;
(1 − ) · (1 − )> .
5 Use the upper bound of Theorem 1 to test whether 6n 1000 3
σ = σ1 ⊗ σ2 or ϵ/3-far from σ1 ⊗ σ2 ;
If ||σ − ⊗ni=1 I22 ||1 > ϵ, at least one of the following two
statements is true:
The correctness of this algorithm follows from the fact that I ϵ
• ||σi − 22 ||1 > 6n for some 1 ≤ i ≤ n; and/or
ϵ
• If σ is independent, σ = σ1 ⊗ σ2 . • σ is 6n -far from being independent.

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:47:32 UTC from IEEE Xplore. Restrictions apply.
YU: ALMOST TIGHT SAMPLE COMPLEXITY ANALYSIS OF QUANTUM IDENTITY TESTING 5067

Otherwise, ||σi − I22 ||1 ≤ 6n ϵ ϵ


and σ is 6n -close to being inde- [14] J. Jiao, K. Venkat, Y. Han, and T. Weissman, “Minimax estimation of
pendent. According to triangle inequality, one can conclude functionals of discrete distributions,” IEEE Trans. Inf. Theory, vol. 61,
no. 5, pp. 2835–2885, May 2015.
that ||σ − ⊗ni=1 I22 ||1 ≤ 3ϵ . [15] G. Valiant and P. Valiant, “Instance optimal learning of discrete dis-
Therefore, the algorithm outputs “No”’ with a probability tributions,” in Proc. 48th Annu. ACM Symp. Theory Comput. (STOC),
1
of at least 1 − 1000 in this case. Jun. 2016, pp. 142–155.
Invoking Lemma 3, we know that [16] Y. Wu and P. Yang, “Minimax rates of entropy estimation on large
alphabets via best polynomial approximation,” IEEE Trans. Inf. Theory,
ϵ n3 log n n · 4n vol. 62, no. 6, pp. 3702–3720, Jun. 2016.
Θ[f (n,
)] + Θ[ ] ≥ Ω( ) [17] I. Diakonikolas, G. Kamath, D. M. Kane, J. Li, A. Moitra, and
6n ϵ2 ϵ2 A. Stewart, “Robust estimators in high dimensions without the compu-
n
n·4 1 tational intractability,” in Proc. IEEE 57th Annu. Symp. Found. Comput.
⇒ f (n, ϵ) ≥ Ω( ) if ϵ < O( ). Sci. (FOCS), Oct. 2016, pp. 655–664.
n · ϵ2 n
[18] G. Valiant and P. Valiant, “Estimating the unseen: Improved estimators
This completes the proof. □ for entropy and other properties,” J. ACM, vol. 64, no. 6, pp. 1–41,
Dec. 2017.
[19] C. Daskalakis and Q. Pan, “Square Hellinger subadditivity for Bayesian
V. C ONCLUSION networks and its applications to identity testing,” in Proc. Conf. Learn.
This paper studies the quantum identity testing problem Theory, vol. 65, 2017, pp. 697–703.
using experimentally friendly Pauli measurements. We prove [20] O. Goldreich, Introduction to Property Testing. Cambridge, U.K.:
Cambridge Univ. Press, 2017.
that for an n-qubit quantum system, the sample complexity [21] I. Diakonikolas, D. M. Kane, and V. Nikishkin, “Near-optimal closeness
n
of this problem is between Θ(n · 4ϵ2 ) and Θ(min{n4 , n3 + testing of discrete histogram distributions,” in Proc. 44th Int. Colloq.
n
log3 (1/ϵ)} · 4ϵ2 ) for nonadaptive Pauli Measurements. Automata, Lang., Program. (ICALP), vol. 80, 2017, pp. 8:1–8:15.
[22] J. Acharya, I. Diakonikolas, J. Li, and L. Schmidt, “Sample-optimal
density estimation in nearly-linear time,” in Proc. 28th Annu. ACM-
ACKNOWLEDGMENT SIAM Symp. Discrete Algorithms (SODA), Jan. 2017, pp. 1278–1289.
[23] S. Chen, J. Li, B. Huang, and A. Liu, “Tight bounds for quantum state
The author would like to thank the reviewer for pointing certification with incoherent measurements,” in Proc. IEEE 63rd Annu.
out the mistake of the lower bound part on Lemma 3 and Symp. Found. Comput. Sci. (FOCS), Oct. 2022, pp. 2541–2596.
Corollary 2 in the original version. [24] C. L. Canonne, I. Diakonikolas, T. Gouleakis, and R. Rubinfeld, “Test-
ing shape restrictions of discrete distributions,” Theory Comput. Syst.,
vol. 62, no. 1, pp. 4–62, 2018.
R EFERENCES [25] C. Daskalakis, N. Dikkala, and G. Kamath, “Testing Ising models,” in
Proc. 29th Annu. ACM-SIAM Symp. Discrete Algorithms (SODA), 2018,
[1] J. Neyman and E. S. Pearson, “On the problem of the most efficient tests pp. 1989–2007.
of statistical hypotheses,” Phil. Trans. Roy. Soc. London A, Containing
[26] I. Diakonikolas, D. M. Kane, and A. Stewart, “Sharp bounds for
Papers Math. Phys. Character, vol. 231, nos. 694–706, pp. 289–337,
generalized uniformity testing,” in Proc. Adv. Neural Inf. Process. Syst.,
1933.
2018, pp. 6201–6210.
[2] E. L. Lehmann, Testing Statistical Hypotheses. New York, NY,
USA: Springer, 1986. [27] Y. Cheng, I. Diakonikolas, and R. Ge, “High-dimensional robust mean
[3] O. Goldreich and D. Ron, “On testing expansion in bounded-degree estimation in nearly-linear time,” in Proc. 30th Annu. ACM-SIAM Symp.
graphs,” in Studies in Complexity and Cryptography. Miscellanea on Discrete Algorithms, 2019, pp. 2755–2771.
the Interplay Between Randomness and Computation (Lecture Notes in [28] R. Rubinfeld, “Taming big probability distributions,” XRDS, vol. 19,
Computer Science), vol. 6650. Berlin, Germany: Springer, 2000. no. 1, pp. 24–28, 2012.
[4] T. Batu, L. Fortnow, R. Rubinfeld, W. D. Smith, and P. White, “Testing [29] C. L. Canonne, “A survey on distribution testing: Your data is big.
that distributions are close,” in Proc. 41st Annu. Symp. Found. Comput. But is it blue?” Electron. Colloq. Comput. Complex., vol. 22, p. 63,
Sci. (FOCS), 2000, pp. 259–269. Apr. 2015.
[5] T. Batu, S. Dasgupta, R. Kumar, and R. Rubinfeld, “The complexity [30] M. Aliakbarpour, I. Diakonikolas, and R. Rubinfeld, “Differentially
of approximating entropy,” in Proc. 34th Annu. ACM Symp. Theory private identity and equivalence testing of discrete distributions,” in
Comput. (STOC), May 2002, pp. 678–687. Proc. 35th Int. Conf. Mach. Learn., in Proceedings of Machine Learning
[6] T. Batu, R. Kumar, and R. Rubinfeld, “Sublinear algorithms for testing Research, J. Dy and A. Krause, Eds., vol. 80, Jul. 2018, pp. 169–178.
monotone and unimodal distributions,” in Proc. 36th Annu. ACM Symp. [31] J. Acharya, Z. Sun, and H. Zhang, “Differentially private testing of
Theory Comput. (STOC), Jun. 2004, pp. 381–390. identity and closeness of discrete distributions,” in Proc. 32nd Int.
[7] P. Valiant, “Testing symmetric properties of distributions,” in Proc. 40th Conf. Neural Inf. Process. Syst. (NIPS). Red Hook, NY, USA: Curran
Annu. ACM Symp. Theory Comput. (STOC), May 2008, pp. 383–392. Associates, 2018, pp. 6879–6891.
[8] J. Acharya, H. Das, A. Jafarpour, A. Orlitsky, and S. Pan, “Competitive [32] C. L. Canonne, G. Kamath, A. McMillan, J. Ullman, and
closeness testing,” in Proc. 24th Annu. Conf. Learn. Theory, vol. 19, L. Zakynthinou, “Private identity testing for high-dimensional distribu-
2011, pp. 47–68. tions,” in Advances in Neural Information Processing Systems, vol. 33,
[9] G. Valiant and P. Valiant, “The power of linear estimators,” in Proc. IEEE H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, Eds.
52nd Annu. Symp. Found. Comput. Sci. (FOCS), Oct. 2011, pp. 403–412. Red Hook, NY, USA: Curran Associates, 2020, pp. 10099–10111.
[10] R. Levi, D. Ron, and R. Rubinfeld, “Testing properties of collections [33] L. Paninski, “A coincidence-based test for uniformity given very sparsely
of distributions,” in Proc. 2nd Symp. Innov. Comput. Sci. (ICS), 2011, sampled discrete data,” IEEE Trans. Inf. Theory, vol. 54, no. 10,
pp. 179–194. pp. 4750–4755, Oct. 2008.
[11] P. Indyk, R. Levi, and R. Rubinfeld, “Approximating and testing k- [34] G. Valiant and P. Valiant, “Estimating the unseen: An n/log(n)-sample
histogram distributions in sub-linear time,” in Proc. 31st ACM SIGMOD- estimator for entropy and support size, shown optimal via new CLTs,”
SIGACT-SIGAI Symp. Princ. Database Syst. (PODS), May 2012, in Proc. 43rd Annu. ACM Symp. Theory Comput. (STOC), Jun. 2011,
pp. 15–22. pp. 685–694.
[12] C. Daskalakis, I. Diakonikolas, R. A. Servedio, G. Valiant, and P. Valiant, [35] T. Batu, E. Fischer, L. Fortnow, R. Kumar, R. Rubinfeld, and P.
“Testing k-modal distributions: Optimal algorithms via reductions,” White, “Testing random variables for independence and identity,” in
in Proc. 24th Annu. ACM-SIAM Symp. Discrete Algorithms (SODA), Proc. 42nd IEEE Symp. Found. Comput. Sci. (FOCS), Oct. 2001,
Jan. 2013, pp. 1833–1852. pp. 442–451.
[13] I. Diakonikolas, D. M. Kane, and V. Nikishkin, “Optimal algorithms and [36] G. Valiant and P. Valiant, “An automatic inequality prover and instance
lower bounds for testing closeness of structured distributions,” in Proc. optimal identity testing,” in Proc. IEEE 55th Annu. Symp. Found.
IEEE 56th Annu. Symp. Found. Comput. Sci., Oct. 2015, pp. 1183–1202. Comput. Sci. (FOCS), Oct. 2014, pp. 51–60.

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:47:32 UTC from IEEE Xplore. Restrictions apply.
5068 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 69, NO. 8, AUGUST 2023

[37] S.-O. Chan, I. Diakonikolas, P. Valiant, and G. Valiant, “Optimal [52] Y.-K. Liu, “Universal low-rank matrix recovery from Pauli measure-
algorithms for testing closeness of discrete distributions,” in Proc. ments,” in Proc. 24th Int. Conf. Neural Inf. Process. Syst. (NIPS).
25th Annu. ACM-SIAM Symp. Discrete Algorithms (SODA), Jan. 2014, Red Hook, NY, USA: Curran Associates, 2011, pp. 1638–1646.
pp. 1193–1203. [53] D. Gross, Y.-K. Liu, S. T. Flammia, S. Becker, and J. Eisert, “Quantum
[38] A. Montanaro and R. de Wolf, “A survey of quantum property testing,” state tomography via compressed sensing,” Phys. Rev. Lett., vol. 105,
Theory Comput. Graduate Surv., vol. 7, no. 7, p. 1081, 2016. no. 15, Oct. 2010, Art. no. 150401.
[39] R. Kueng, H. Rauhut, and U. Terstiege, “Low rank matrix recovery from [54] S. T. Flammia, D. Gross, Y.-K. Liu, and J. Eisert, “Quantum tomography
rank one measurements,” Appl. Comput. Harmon. Anal., vol. 42, no. 1, via compressed sensing: Error bounds, sample complexity, and efficient
pp. 88–116, Jan. 2017. estimators,” New J. Phys., vol. 14, May 2012, Art. no. 095022.
[40] J. Haah, A. W. Harrow, Z. Ji, X. Wu, and N. Yu, “Sample-optimal [55] S. T. Flammia and Y.-K. Liu, “Direct fidelity estimation from few Pauli
tomography of quantum states,” in Proc. 48th Annu. ACM Symp. Theory measurements,” Phys. Rev. Lett., vol. 106, Jun. 2011, Art. no. 230501.
Comput. (STOC), Jun. 2016, pp. 913–925. [56] M. P. da Silva, O. Landon-Cardinal, and D. Poulin, “Practical charac-
[41] R. O’Donnell and J. Wright, “Efficient quantum tomography,” in terization of quantum devices without tomography,” Phys. Rev. Lett.,
Proc. 48th Annu. ACM Symp. Theory Comput. (STOC), Jun. 2016, vol. 107, no. 21, Nov. 2011, Art. no. 210404.
pp. 899–912. [57] L. Aolita, C. Gogolin, M. Kliesch, and J. Eisert, “Reliable quantum
[42] R. O’Donnell and J. Wright, “Efficient quantum tomography II,” in Proc. certification of photonic state preparations,” Nature Commun., vol. 6,
47th Annu. ACM Symp. Theory Comput. (STOC), 2017, pp. 962–974. no. 1, p. 8498, Nov. 2015.
[43] S. Chen, B. Huang, J. Li, A. Liu, and M. Sellke, “Tight bounds for state [58] T. Cai, D. Kim, Y. Wang, M. Yuan, and H. H. Zhou, “Optimal large-
tomography with incoherent measurements,” 2022, arXiv:2206.05265. scale quantum state tomography with Pauli measurements,” Ann. Statist.,
[44] H.-Y. Huang, R. Kueng, and J. Preskill, “Information-theoretic bounds vol. 44, no. 2, pp. 682–712, Apr. 2016.
on quantum advantage in machine learning,” Phys. Rev. Lett., vol. 126, [59] O. Goldreich, “On multiple input problems in property testing,”
no. 19, May 2021, Art. no. 190505. in Approximation, Randomization, and Combinatorial Optimization.
[45] S. Chen, J. Cotler, H.-Y. Huang, and J. Li, “Exponential separa- Algorithms and Techniques (APPROX/RANDOM 2014) (Leibniz Inter-
tions between learning with and without quantum memory,” in Proc. national Proceedings in Informatics (LIPIcs)), vol. 28, K. Jansen,
IEEE 62nd Annu. Symp. Found. Comput. Sci. (FOCS), Feb. 2022, J. D. P. Rolim, N. R. Devanur, and C. Moore, Eds. Dagstuhl, Germany:
pp. 574–585. Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik, 2014, pp. 704–720.
[46] S. Chen, B. Huang, J. Li, and A. Liu, “Tight bounds for quantum state [Online]. Available: http://drops.dagstuhl.de/opus/volltexte/2014/4733
certification with incoherent measurements,” in Proc. IEEE 63rd Annu. [60] R. Levi, D. Ron, and R. Rubinfeld, “Testing properties of collections of
Symp. Found. Comput. Sci. (FOCS), Oct./Nov. 2022, pp. 1205–1213. distributions,” Theory Comput., vol. 9, no. 8, pp. 295–347, Mar. 2013.
[47] H.-Y. Huang et al., “Quantum advantage in learning from experiments,” [61] I. Diakonikolas and D. M. Kane, “A new approach for testing properties
Science, vol. 376, no. 6598, pp. 1182–1186, Jun. 2022. of discrete distributions,” in Proc. IEEE 57th Annu. Symp. Found.
[48] R. O’Donnell and J. Wright, “Quantum spectrum testing,” in Proc. 47th Comput. Sci. (FOCS), Oct. 2016, pp. 685–694.
Annu. ACM Symp. Theory Comput. (STOC), Jun. 2015, pp. 529–538.
[49] S. Bubeck, S. Chen, and J. Li, “Entanglement is necessary for optimal
quantum property testing,” in Proc. IEEE 61st Annu. Symp. Found.
Comput. Sci. (FOCS), Los Alamitos, CA, USA, Nov. 2020, pp. 692–703.
[50] N. Yu, “Sample efficient identity testing and independence test-
ing of quantum states,” in Proc. 12th Innov. Theor. Comput. Sci. Nengkun Yu received the B.S. and Ph.D. degrees from the Department
Conf. (ITCS) (Leibniz International Proceedings in Informatics), of Computer Science and Technology, Tsinghua University, Beijing, China,
vol. 185, J. R. Lee, Ed. Dagstuhl, Germany: Schloss Dagstuhl–Leibniz- in 2008 and 2013, respectively. He is currently an Associate Professor
Zentrum für Informatik, 2021, pp. 11:1–11:20. [Online]. Available: and a SUNY Empire Innovation Scholar with the Department of Computer
https://drops.dagstuhl.de/opus/volltexte/2021/13550 Science, Stony Brook University. His research interests include quantum
[51] C. Bădescu, R. O’Donnell, and J. Wright, “Quantum state certification,” computing. He won the ACM SIGPLAN Distinguished Paper Award at
in Proc. 51st Annu. ACM SIGACT Symp. Theory Comput. (STOC), OOPSLA 2020 and the ACM SIGPLAN Distinguished Paper Award at PLDI
Jun. 2019, pp. 503–514. 2021.

Authorized licensed use limited to: ANIRBAN kanungoe. Downloaded on August 22,2023 at 01:47:32 UTC from IEEE Xplore. Restrictions apply.

You might also like