AIMS Mathematics, 8(8): 18295–18313.

DOI:10.3934/math.2023930
Received: 06 March 2023
Revised: 16 May 2023
Accepted: 23 May 2023
http://www.aimspress.com/journal/Math Published: 29 May 2023

Research article
A prediction-correction based proximal method for monotone variational
inequalities with linear constraints

Feng Ma1,∗ , Bangjie Li1,∗ , Zeyan Wang2 , Yaxiong Li1 and Lefei Pan1
1
Xi’an Research institute of High-Tech, Xi’an 710025, China
2
Department of Basic Education, Army Engineering University of PLA, Nanjing 211101, China

* Correspondence: Email: fengma@outlook.com, rosoldier@163.com.

Abstract: The monotone variational inequalities are being widely used as mathematical tools for
studying optimal control problems and convex programming. In this paper, we propose a new
prediction-correction method for monotone variational inequalities with linear constraints. The method
consists of two procedures. The first procedure (prediction) utilizes projections to generate a predictor.
The second procedure (correction) produces the new iteration via some minor computations. The main
advantage of the method is that its main computational effort only depends on evaluating the resolvent
mapping of the monotone operator, and its primal and dual step sizes can be enlarged. We prove the
global convergence of the method. Numerical results are provided to demonstrate the efficiency of the
method.
Keywords: variational inequalities; linear constraints; proximal algorithm; prediction-correction
methods; contraction
Mathematics Subject Classification: 65K05, 65K10

1. Introduction

Let D ⊂ <n be a nonempty, closed and convex set and f : <n → <n be a given continuous
and monotone mapping. The variational inequality problem, denoted by VI(D, f ), is to find a vector
x∗ ∈ D such that
hx − x∗ , f (x∗ )i ≥ 0, ∀x ∈ D. (1.1)
The VI(D, f ) has found many important applications in areas such as nonlinear complementarity
problems (where D = <n+ ) [1], traffic equilibrium and economic problems [2, 3]. For recent
applications and numerical methods of the VI(D, f ), we refer the reader to [4–6] and the references
therein.
 18296

In this paper, we consider a special case of the general VI problem, where the set D is assumed to
have the following form:
D = {x ∈ <n |x ∈ X, Ax = b}, (1.2)
here A ∈ <m×n is a given matrix, b ∈ <m is a given vector, X ⊂ <n is a nonempty, closed and convex
subset. The solution set of (1.1) and (1.2), denoted by Ω∗ , is assumed to be nonempty. Note that the
VI problem (1.1) and (1.2) is closely related to the convex optimization problem with linear equality
constraints:
min θ(x)
(1.3)
s.t. Ax = b, x ∈ X.
To show this, we recall the first order optimization conditions of problem (1.3). Let x∗ be a minimum
point of the convex function θ(x) over D and ξ∗ be any vector in ∂θ(x∗ ), where ∂θ(·) denotes the
subdifferential operator of θ(·). Then for any feasible direction d ∈ <n at x∗ , we have (ξ∗ )T d > 0. It
means that, the following variational inequality problem is captured:

hx − x∗ , ξ∗ i ≥ 0, ∀x ∈ D. (1.4)

Thus, solving the VI problem (1.4) amounts to solving (1.3). This VI characterization is also used in
e.g., [7, 8].
The VI problem (1.1) and (1.2) or its equivalent form (1.3) is one of the fundamental problems in
convex optimization. In particular, it includes linear programming, conic and semidefinite optimization
as special cases. It can find many applications in compressed sensing, image processing and data
mining, see, e.g., [9–11]. We refer to [12] for recent examples and discussions. To solve the VI
problem (1.1) and (1.2), some proximal-like algorithms have been developed over the past years, see
e.g., [13] for a review. One benchmark method is the augmented Lagrangian Method (ALM) [14, 15]
for nonlinear problems. The ALM is applicable to solve VI(D, f ) (1.1) and (1.2). More specifically,
for a given uk = (xk , λk ), ALM uses the following procedure to carry out the new iteration uk+1 =
(xk+1 , λk+1 ) ∈ X × <m :
Find xk+1 ∈ X such that

hx − xk+1 , f (xk+1 ) − AT λk + βAT (Axk+1 − b)i ≥ 0, ∀x ∈ X, (1.5a)

then update λk+1 via

λk+1 = λk − β(Axk+1 − b), (1.5b)
where β ≥ 0 is a given penalty parameter for the violation of the linearly constraints. To make
ALM (1.5) more efficient and flexible, some alternative strategies can be used. For example, some
self-adaptive rules can be carried to adjust the parameter β based on certain strategies [16–19]. We can
also use some correction technologies to the output point [20,21]. Let ũk denote the predictor generated
by ALM. A simple and effective correction scheme is

uk+1 = uk − αk (uk − ũk ), (1.6)

where 0 < αk < 2 is the step size parameter; see, e.g., [22, 23] for details.
The main computational cost at each iteration of ALM is to solve the subproblem (1.5a), which
is still a variational inequality, structurally the same as the original problem. So in many cases the

AIMS Mathematics Volume 8, Issue 8, 18295–18313.

 18297

ALM (1.5a) is easy to iterate only if A is an identity matrix. This is because, for this case, the
solution of the subproblem (1.5a) corresponds to the proximal mappings of f and it usually has closed-
form solutions or can be efficiently solved up to a high precision. However, in many applications in
sparse optimization, we often encounter the case where A is not an identity matrix, and the resulting
subproblem (1.5a) no longer has closed-form solutions. For this case, solving the subproblem (1.5a)
could be computationally intensive because of the costly evaluation of (AT A + β1 f )−1 (Aυ). Therefore,
efficient numerical methods with implementable iterative scheme are highly desired.
Recently, several techniques attempting to overcome this difficulty have been proposed. In the
framework of the proximal point algorithm (PPA), there are two relevant approaches. The first
one is regularization. By adding a customized regularization term to the saddle-point reformulation
of (1.1) and (1.2), the primal subproblem becomes easy as it only involves a simple evaluation, see
the customized PPA algorithms proposed in e.g., [24–27]. We refer the reader to e.g., [28, 29] of
the linearized regularization term for the separable case of problem (1.2). The second one employs
prediction-correction technology which adds an asymmetrical proximal term to make a prediction,
and then introduces a simple correction step to guarantee the convergence, see e.g., [30–32]. In this
paper, we propose a new prediction-correction method for the VI(D, f ). At each iteration, the method
first makes a simple prediction step to obtain a point, and then generates a new iteration via a minor
correction to the predictor. The reduced subproblems are easy to solve when the resolvent mapping of
f can be efficiently evaluated. As can be seen in the Section 5, the proposed method converges faster
with less iterations to achieve the same accuracy on most numerical cases.
The rest of this paper is organized as follows. In Section 2, we review some preliminaries which
are useful for further analysis. In Section 1, we present the implementable prediction-correction
method for VI(D, f ). In Section 4, we establish the global convergence of the proposed method.
The computational experiment is presented in Section 5. Finally, we make a conclusion in Section 6.

2. Preliminaries

2.1. Equivalent VI
In this section, we reformulate the VI problem (1.1) and (1.2) in succinct form. Let Ω = X×<m . By
attaching a Lagrange multiplier vector λ ∈ <m to the linear constraints Ax = b, the VI problem (1.1)
and (1.2) can be converted to the following form:

x − x∗ , f (x∗ ) − AT λ∗
* +
≥ 0, ∀ (x, λ)T ∈ Ω. (2.1)
λ − λ∗ , Ax∗ − b

By denoting
f (x) − AT λ
! !
x
u := , F(u) := , (2.2)
λ Ax − b
we can rewrite (2.1) into the following more compact form

hu − u∗ , F(u∗ )i ≥ 0, ∀u ∈ Ω. (2.3)

Henceforth, we will denote the VI problem (2.2) and (2.3) by VI(Ω, F). Now, we make some basic
assumptions and summarize some well known results of VI, which will be used in subsequent analysis.

AIMS Mathematics Volume 8, Issue 8, 18295–18313.

 18298

2.2. Assumption
(A1) X is a simple closed convex set.
A set is said to be simple if the projection onto it can be easily obtained. Here, the projection of
a point a onto the closed convex set X, denoted by PX (a), is defined as the nearest point x ∈ X to
a, i.e.,
PX (a) = argmin{kx − ak |x ∈ X}.

(A2) The mapping f is point-to-point, monotone and continuous.

A mapping F : <n → <n is said to be monotone on Ω if

hu − υ, F(u) − F(υ)i ≥ 0, ∀u, υ ∈ Ω. (2.4)

(A3) The solution set of (2.2) and (2.3), denoted by Ω∗ , is nonempty.

Remark 1. The mapping F(·) defined in (2.3) is monotone with respect to Ω since

(F(u) − F(ũ))T (u − ũ) ≥ 0, ∀u, ũ ∈ Ω. (2.5)

Proof.
!T
f (x) − f ( x̃) − AT (λ − λ̃)
!
x − x̃
(F(u) − F(ũ)) (u − ũ) =
T
A(x − x̃) λ − λ̃
= ( f (x) − f ( x̃))T (x − x̃)
≥ 0, (2.6)

where the last inequality follows from the monotone property of f . 

2.3. Properties

√ The G-norm of the vector u is denoted by kukG :=

√ Let G be a symmetric positive definite matrix.
hu, Gui. In particular, when G = I, kuk := hu, ui is the Euclidean norm of u. For a matrix A, kAk
denotes its norm kAk := max{kAxk : kxk ≤ 1}.
The following well-known properties of the projection operator will be used in the coming analysis.
The proofs can be found in textbooks, e.g., [2, 33].
Lemma 1. Let X ∈ <n be a nonempty closed convex set, PX (·) be the projection operator onto X under
the G-norm. Then

hy − PX (y), G(x − PX (y))i ≤ 0, ∀y ∈ <n , ∀x ∈ X. (2.7)

kPX (x) − PX (y)kG ≤ kx − ykG , ∀x, y ∈ <n . (2.8)
kx − PX (y)kG2 ≤ kx − ykG2 − ky − PX (y)kG2 , ∀y ∈ <n , ∀x ∈ X. (2.9)

For any arbitrary positive scalar β and u ∈ Ω, let e(u, β) denote the residual function associated with
the mapping F, i.e.,
e(u, β) = u − PΩ [u − βF(u)]. (2.10)

AIMS Mathematics Volume 8, Issue 8, 18295–18313.

 18299

Lemma 2. Solving VI(Ω, F) is equivalent to finding the zero point of the mapping

e1 (u, β) x − PX {x − β[ f (x) − AT λ]}

! !
e(u, β) := = . (2.11)
e2 (u, β) β(Ax − b)

Proof. See [2], pp 267. 

3. The prediction-correction method

We now formally present the new prediction-correction method for the VI(Ω, F) (Algorithm 1). The
method can be viewed as a generalization of [31] in relaxed case.

Algorithm 1: A prediction-correction based method (PCM) for the VI(Ω, F).

Step 0: Initialization step.
Given a small number  > 0. Take γ ∈ (0, 2), u0 ∈ <n+m ; set k = 0. Choose the parameters
r > 0, s > 0, such that
1
rs > kAT Ak. (3.1)
4

Step 1: Prediction step.

Generate the predictor x̃k via solving the following projection equation:
1
x̃k = PX [xk − ( f ( x̃k ) − AT λk )]. (3.2a)
r
Then update λ̃k ∈ <m via
1
λ̃k = λk − (A x̃k − b). (3.2b)
s

Step 2: Correction step.

Generate the new iteration uk+1 = (xk+1 , λk+1 ) via
γ T k
xk+1 = xk − γ(xk − x̃k ) − A (λ − λ̃k ), (3.3a)
2r
and
γ 1
λk+1 = λk + A(xk − x̃k ) − γ(I − AAT )(λk − λ̃k ). (3.3b)
2s 2rs

Step 3: Convergence verification.

If kuk+1 − uk k ≤ , stop; otherwise set k := k + 1; go to Step 1.

Remark 2. Note that the regularized projection Eq (3.2a) amounts to solving the following VI problem

1
hx − x̃k , ( f ( x̃k ) − AT λk ) + ( x̃k − xk )i ≥ 0, ∀x ∈ X, (3.4)
r

AIMS Mathematics Volume 8, Issue 8, 18295–18313.

 18300

which represents to find x̃k ∈ X such that

1 1
0 ∈ x̃k + f ( x̃k ) − ( AT λk + xk ). (3.5)
r r
We can rewrite the above equation as
1 −1 1 T k
x̃k ∈ (I + f ) ( A λ + xk ). (3.6)
r r
Thus, the subproblem (3.2a) is equivalent to an evaluation of the resolvent operator (I + 1r f )−1 when
X is the finite-dimensional Euclidean spaces [34]. Notice that ALM (1.5a) needs to evaluate the
operator (AT A + β1 f )−1 . Therefore, the resulting subproblem (3.2a) could be much easier to solve than
ALM (1.5a). On the other hand, the correction step only consists of some elementary manipulations.
Thus, the resulting method (3.2a)–(3.3b) is easily implementable.
Remark 3. The parameters 1r and 1s in the prediction step can be viewed as two step sizes for the
projection step (3.2a) and dual step, respectively. Using the step size condition rs > 14 kAT Ak of this
algorithm, the parameters 1r and 1s can be chosen larger values compared to the condition rs > kAT Ak
of some other primal dual algorithms, e.g., linearized ALM, customized PPA algorithms. This larger
step size is usually beneficial to the effectiveness and robustness of the algorithm. In Section 5 we will
empirically see that our algorithm is significantly faster than some other primal dual algorithms by
allowing larger step sizes.

4. Convergence properties

In the following, we will focus our attention to solving VI(Ω, F). But at the beginning, to make the
notation of the proof more succinct, we define some matrices:

rI 12 AT
! !
rI AT
G= 1 , Q= . (4.1)
2
A sI 0 sI

Obviously, when rs > 14 kAT Ak, G is a positive definite matrix. Now, we start to prove the global
convergence of the sequence {uk }. Towards this end, we here follow the work [35] to reformulate the
algorithm into a prediction-correction method and establish its convergence results. We first prove
some lemmas. The first lemma quantifies the discrepancy between the point ũk and a solution point of
VI(Ω, F).
Lemma 3. Let {uk } be generated by the PCM and {ũk } be generated by PCM (3.2), and the matrix Q
be given in (4.1). Then we have

hu − ũk , F(ũk ) − Q(uk − ũk )i ≥ 0, ∀u ∈ Ω. (4.2)

Proof. Note that the sequence {ũk } generated by (3.2) is actually solutions of the following VIs:

hx − x̃k , f ( x̃k ) − AT λk − r(xk − x̃k )i ≥ 0, ∀x ∈ X, (4.3)

and
hλ − λ̃k , A x̃k − b − s(λk − λ̃k )i ≥ 0, ∀λ ∈ <m . (4.4)

AIMS Mathematics Volume 8, Issue 8, 18295–18313.

 18301

Combining (4.3) and (4.4) yields

x − x̃k , f ( x̃k ) − AT λ̃k − AT (λk − λ̃k ) − r(xk − x̃k )

* +
≥ 0, ∀u ∈ Ω. (4.5)
λ − λ̃, A x̃k − b − s(λk − λ̃k )

It can be rewritten as

f ( x̃k ) − AT λ̃k
* ! ! ! !+
x − x̃k rI AT xk − x̃k
, − ≥ 0, ∀u ∈ Ω. (4.6)
λ − λ̃k A x̃k − b 0 sI λk − λ̃k

Using the notation of F(u) (2.3) and Q (4.1), the assertion (4.2) is proved. 

The following lemma characterizes the correction step by a matrix form.

Lemma 4. Let {uk } be generated by the PCM and {ũk } be generated by PCM (3.2), Then we have

uk − uk+1 = γM(uk − ũk ), (4.7)

where
1 T
!
I A
M= 1
2r
AAT . (4.8)
− 2s A I− 2rs

Proof. From the correction Step (3.3a) and (3.3b), we have

T
(xk − x̃k ) + A2r (λk − λ̃k )
! ! !
xk+1 xk
= −γ , (4.9)
λk+1 λk − 2s1 A(xk − x̃k ) + (I − 2rs
1
AAT )(λk − λ̃k )

which can be written as

1 T
! ! !
xk − xk+1 I A xk − x̃k
=γ 2r . (4.10)
λk − λk+1 1
− 2s A I− AAT
2rs
λk − λ̃k

By noting the matrix M (4.8), the proof is completed. 

Using the matrices Q (4.1) and M (4.8), we define

H = QM −1 , (4.11)

then we have Q = HM. The inequality (4.2) can be written as

γhu − ũk , F(ũk ) − HM(uk − ũk )i ≥ 0, ∀u ∈ Ω. (4.12)

Substituting (4.7) into (4.12) and using the monotonicity of F (see (2.5)), we obtain

ũk ∈ Ω, hu − ũk , F(u)i ≥ hu − ũk , H(uk − uk+1 )i, ∀u ∈ Ω. (4.13)

Now, we prove a simple fact for the matrix H in the following lemma.
Lemma 5. The matrix H defined in (4.11) is positive definite for any r > 0, s > 0, rs > 41 kAT Ak.

AIMS Mathematics Volume 8, Issue 8, 18295–18313.

 18302

Proof. For the matrix Q defined by (4.1), we have

1
− rs1 AT
!
I
Q −1
= r
1 .
0 s
I

Thus, it follows from (4.1) that

1 T 1
− rs1 AT
! !
I A I
H −1
= MQ −1
= 2r
T
r
1
− 2s1 A I − AA2rs
0 s
I
1 1
AT 
 
r
I − 2rs
=   .

1 1
(4.14)
− 2rs A s
I

For any any r > 0, s > 0 satisfying rs > 41 kAT Ak, H −1 is positive definite, and the positive definiteness
of H is followed directly. 
Then we show how to deal with the right-hand side of (4.13). The following lemma gives an
equivalent expression of it in terms of ku − uk kH and ku − uk+1 kH .
Lemma 6. Let {uk } be generated by the PCM and {ũk } be generated by PCM (3.2). Then we have
1 1
hu − ũk , H(uk − uk+1 )i = (ku − uk+1 k2H − ku − uk k2H ) + kuk − ũk kG2 , ∀u ∈ Ω, (4.15)
2 2
where the matrix G = γ(QT + Q) − γ2 M T HM.
Proof. Applying the identity
1 1
ha − b, H(c − d)i = (ka − dk2H − ka − ck2H ) + (kc − bk2H − kd − bk2H ), (4.16)
2 2
to the right term of (4.13) with a = u, b = ũk , c = uk , d = uk+1 , we obtain
1 1
hu − ũk , H(uk − uk+1 )i = (ku − uk+1 k2H − ku − uk k2H ) + (kuk − ũk k2H − kuk+1 − ũk k2H ). (4.17)
2 2
For the last term of (4.17), we have

kuk − ũk k2H − kuk+1 − ũk k2H

= kuk − ũk k2H − k(uk − ũk ) − (uk − uk+1 )k2H
(4.7)
= kuk − ũk k2H − k(uk − ũk ) − γM(uk − ũk )k2H (4.18)
= 2γhu − ũ , HM(u − ũ )i − γ hu − ũ , M HM(u − ũ )i
k k k k 2 k k T k k

(4.11)
= huk − ũk , (γ(QT + Q) − γ2 M T HM)(uk − ũk )i.

the assertion is proved. 

Now, we investigate the positive definiteness for the matrix G. Using (4.11), we have

G = γ(QT + Q) − γ2 M T HM

AIMS Mathematics Volume 8, Issue 8, 18295–18313.

 18303

= γ(QT + Q) − γ2 M T Q
− 2s1 AT
! ! !
2rI AT I rI AT
= γ − γ2 1 T
A 2sI 2r
A I − AA 2rs
0 sI
1 T
! !
2rI AT rI A
= γ − γ2 1
2
A 2sI 2
A sI
1 T
!
rI A
= (2γ − γ2 ) 1
2 . (4.19)
2
A sI

Thus, when rs > 14 kAT Ak and γ ∈ (0, 2), the matrix G is guaranteed to be positive definite, and we can
easily obtained the contraction property of the algorithm. This is given by the following theorem.
Theorem 1. Suppose the condition
1
rs > kAT Ak (4.20)
4
holds. Let the relaxation factor γ ∈ (0, 2). Then, for any u∗ = (x∗ , λ∗ )T ∈ Ω∗ , the sequence uk+1 =
(xk+1 , λk+1 )T generated by PCM satisfies the following inequality:

kuk+1 − u∗ k2H ≤ kuk − u∗ k2H − kuk − ũk kG2 . (4.21)

Proof. Combining (4.13) and (4.15), we have

ũk ∈ Ω, hu − ũk , F(u)i

1 1
≥ (ku − uk+1 k2H − ku − uk k2H ) + kuk − ũk kG2 , ∀u ∈ Ω. (4.22)
2 2
Note that u∗ ∈ Ω. We get

kuk − u∗ k2H − kuk+1 − u∗ k2H ≥ kuk − ũk kG2 + 2hũk − u∗ , F(u∗ )i.

Since u∗ is a solution of (2.3) and ũk ∈ Ω, we have

hũk − u∗ , F(u∗ )i ≥ 0, (4.23)

and thus
kuk − u∗ k2H − kuk+1 − u∗ k2H ≥ kuk − ũk kG2 .
The assertion (4.21) follows directly. 
Based on the above results, we are now ready to prove the global convergence of the algorithm.
Theorem 2. Let {uk } be the sequence generated by PCM for the VI problem (2.2) and (2.3). Then, for
any r > 0, s > 0 satisfying rs > 14 kAT Ak and γ ∈ (0, 2), the sequence {uk } converges to a solution of
VI(Ω, F).
Proof. We follows the standard analytic framework of contraction-type methods to prove the
convergence of the proposed algorithm. It follows from Theorem 1 that {uk } is bounded. Then we
have that
lim kuk − ũk kG = 0. (4.24)
k→∞

AIMS Mathematics Volume 8, Issue 8, 18295–18313.

 18304

Consequently,
lim kxk − x̃k k = 0, (4.25)
k→∞

and
lim kA x̃k − bk = lim ks(λk − λ̃k )k = 0. (4.26)
k→∞ k→∞

Since
1 1
x̃k = PX [ x̃k − ( f ( x̃k ) − AT λ̃k ) + (xk − x̃k ) + AT (λk − λ̃k )], (4.27)
r r
and
1
λ̃k = λk − (A x̃k − b), (4.28)
s
it follows from (4.25) and (4.26) that

1

lim x̃k − PX [ x̃k − ( f ( x̃k ) − AT λ̃k )] = 0,



 (4.29a)
 k→∞

 r


 lim A x̃k − b = 0.



 (4.29b)
k→∞

Because ũk is also bounded, it has at least one cluster point. Let u∞ be a cluster point of ũk and let ũkj
be the subsequence converges to u∞ . It follows from (4.29) that
 1
lim x̃ k
− P [ x̃ k
− ( f ( x̃kj ) − AT λ̃kj )] = 0, (4.30a)


 j X j
r


 j→∞



 lim A x̃kj − b = 0.



 (4.30b)
j→∞

Consequently, we have
 1
 x − PX [x − r ( f (x ) − A λ )] = 0,
∞ ∞ ∞ T ∞
(4.31a)







Ax∞ − b = 0.

(4.31b)


Using the continuity of F and the projection operator PΩ (·), we have that u∞ is a solution of
VI(Ω, F). On the other hand, by taking limits over the subsequences in (4.24) and using lim j→∞ ũkj =
u∞ . we have that, for any k > k j ,
kuk − u∞ kH ≤ kuk j − u∞ kH .
Thus, the sequence {uk } converges to u∞ , which is a solution of VI(Ω, F). 

5. Numerical experiments

In this paper, we test the performance of PCM (3.2a)–(3.3b) for solving the basis pursuit (BP) and
matrix completion problem. All the simulations are performed on a Lenovo Laptop with CPU Intel
with 2.81GHz and 16GB RAM memory, using Matlab R2015b.

AIMS Mathematics Volume 8, Issue 8, 18295–18313.

 18305

5.1. Basis pursuit problem

The BP problem arises from areas such as the communities of information theory, signal processing,
statistics, machine learning. it seeks to encode a large sparse signal representations through a relatively
small number of linear equations. The BP problem can be cast as the following equality-constrained l1
minimization problem
min kxk1
(5.1)
s.t. Ax = b,
where x ∈ <n , data A ∈ <m×n with m < n, b ∈ <m . Here we assume that A has full row rank. By
invoking the first-order optimality condition, BP is equivalent to VI (1.1) and (1.2) with f (x) = ∂kxk1 .
Applying PCM with γ = 1 for this problem, we get the following iterative scheme:

1
x̃k ∈ P<n [xk − (∂k x̃k k1 − AT λk )],



 (5.2a)



 r
1


λ̃k = λk − (A x̃k − b),

(5.2b)



s



 1
xk+1 = x̃k − AT (λk − λ̃k ),




 (5.2c)



 2r
1 AAT k


 λ = λ + A(x − x̃ ) − (I − )(λ − λ̃k ).
 k+1 k k k
(5.2d)



2s 2rs
Note that the projection (5.2a) is equivalent to the l1 shrinkage operation:
1 1
x̃k := Shrink(xk + AT λk , ),
r r
where the l1 shrinkage operator, denoted by Shrink(M, ξ), is defined as
Mi − ξ, if Mi > ξ,







[Shrink(M, ξ)]i :=  Mi + ξ, if Mi < −ξ, i = 1, 2, . . . , n.






if |Mi | ≤ ξ,

 0,


In our experiments, we focus on comparing our algorithm with the linearized ALM [36] (L-ALM)
and the customized PPA (C-PPA) in [27] and verifying its efficiency. Similar with PCM, L-ALM and
C-PPA also depend on Shrink operation, which have the same easiness of implementation.
The data used in this experiment is similar to the one employed in [37]. The basic setup of the
problem is as follows. The data matrix A is a i.i.d. standard Gaussian matrix generated by the randn(m,
n) function in MATLAB with m = n/2. The original sparse signal xtrue is sampled from i.i.d. standard
Gaussian distributions with m/5 nonzero values. The output b is then created as the signs of b = Ax. It
is desirable to test problems that have a precisely known solution. In fact, when the original signal xtrue
is very sparse, it reduces to a minimization problem. The parameters used in the numerical experiments
are similar to that in [19, 38]: we set the relaxation factor γ = 1, s ∈ (25, 100), r = 1.01kAT Ak/(4s)
for PCA and s ∈ (25, 100), r = 1.01kAT Ak/s for C-PPA. (In order to ensure the convergence of
L-ALM and C-PPA, the parameters r, s should satisfy rs > kAT Ak.) We set the criterion error as
min{relL = kxk − xk−1 k, relS = Axk − b}, and declare successful recovery when this error is less than
T ol = 10−3 . In all the tests, the initial iteration is (x0 , λ0 ) = (0, 1).

AIMS Mathematics Volume 8, Issue 8, 18295–18313.

 18306

We first test the sensitivity of γ of PCM. We fix s = 50 and choose different values of γ in the
interval [0.6, 1.8]. The number of iterations required are reported in Figure 1. The curves in Figure 1
indicate that γ ∈ (1, 1.4) is preferable when we implement Algorithm PCM in practice.

350

m=500
n=1000
300 n=1500
n=2000
n=2500

250
Iter (k)

200

150

100
0.6 0.8 1 1.2 1.4 1.6 1.8 2
The value of .

Figure 1. Sensitivity test on the relaxation factor of γ.

In order to investigate the stability and efficiency of our algorithms, we test 8 groups of problems
with different n and we generated the model by 10 times and reported the average results. The
comparisons of these algorithms for small BP problems are presented in Tables 1–3.

Table 1. Numerical Results for Basis Pursuit problem s = 25.

PCM L-ALM C-PPA
n Iter. relL relS CPU(s) Iter. relL relS CPU(s) Iter. relL relS CPU(s)
100 81 4.2e-04 1.9e-04 0.03 161 1.6e-04 7.5e-04 0.03 232 8.2e-05 8.6e-04 0.04
300 88 5.1e-04 7.3e-04 0.05 207 7.5e-05 6.3e-04 0.07 311 5.3e-05 9.4e-04 0.10
600 110 9.2e-05 7.9e-04 0.54 262 7.2e-05 8.9e-04 0.62 374 6.0e-05 9.4e-04 1.01
1000 141 4.4e-05 9.8e-04 1.81 291 3.3e-05 7.9e-04 2.21 411 3.2e-05 9.4e-04 2.86
1500 139 6.7e-05 6.6e-04 3.95 359 7.0e-05 9.4e-04 5.04 484 6.9e-05 9.0e-04 6.54
2000 151 6.3e-05 7.9e-04 6.55 406 1.2e-05 9.7e-04 9.24 536 1.3e-05 9.7e-04 12.08
2500 171 3.9e-05 8.5e-04 11.57 519 1.3e-05 8.9e-04 18.60 635 1.8e-05 9.8e-04 22.69
3000 199 2.9e-05 8.8e-04 18.82 585 1.7e-05 9.8e-04 29.46 709 1.7e-05 9.0e-04 35.61

AIMS Mathematics Volume 8, Issue 8, 18295–18313.

 18307

Table 2. Numerical Results for Basis Pursuit problem s = 50.

PCM L-ALM C-PPA
n Iter. relL relS CPU(s) Iter. relL relS CPU(s) Iter. relL relS CPU(s)
100 80 5.9e-04 7.0e-04 0.03 151 3.7e-04 9.6e-04 0.03 224 1.4e-04 9.4e-04 0.04
300 87 6.1e-04 8.1e-04 0.05 203 1.2e-04 8.5e-04 0.09 299 3.3e-05 9.0e-04 0.10
600 110 1.1e-04 6.1e-04 0.49 249 3.5e-05 9.9e-04 0.63 331 5.7e-05 8.6e-04 0.87
1000 137 4.0e-05 8.5e-04 1.75 266 3.2e-05 7.9e-04 1.80 370 3.1e-05 8.3e-04 2.62
1500 135 3.5e-05 8.6e-04 3.54 312 3.6e-05 8.5e-04 4.35 424 3.4e-05 8.2e-04 5.84
2000 146 5.4e-05 7.6e-04 6.38 331 1.3e-05 9.8e-04 7.57 445 1.5e-05 9.5e-04 10.19
2500 143 4.1e-05 9.7e-04 9.67 375 1.4e-05 9.3e-04 13.41 491 1.5e-05 9.1e-04 17.63
3000 170 3.1e-05 9.7e-04 16.00 418 1.5e-05 9.5e-04 20.70 533 1.6e-05 8.9e-04 26.42

Table 3. Numerical Results for Basis Pursuit problem s = 75.

PCM L-ALM C-PPA
n Iter. relL relS CPU(s) Iter. relL relS CPU(s) Iter. relL relS CPU(s)
100 82 5.3e-04 5.7e-04 0.02 164 2.7e-04 6.2e-04 0.03 236 6.3e-05 9.2e-04 0.04
300 99 1.9e-04 7.9e-04 0.05 213 9.4e-05 5.9e-04 0.07 284 4.2e-05 9.2e-04 0.09
600 117 6.7e-05 8.9e-04 0.55 226 1.1e-04 9.9e-04 0.58 316 8.5e-05 9.3e-04 0.84
1000 179 4.2e-05 8.7e-04 2.35 253 5.4e-05 9.5e-04 1.68 369 1.9e-05 9.5e-04 2.47
1500 159 3.4e-05 9.3e-04 4.20 299 3.4e-05 8.1e-04 4.13 399 3.5e-05 7.9e-04 5.39
2000 142 6.6e-05 9.4e-04 6.18 313 1.3e-05 9.0e-04 7.32 410 1.9e-05 9.7e-04 9.56
2500 139 2.9e-05 9.8e-04 9.41 337 1.4e-05 1.0e-03 12.20 444 1.1e-05 9.8e-04 16.00
3000 168 3.7e-05 9.0e-04 15.92 370 1.6e-05 8.2e-04 18.54 464 2.2e-05 9.0e-04 23.22

From Tables 1–3, it can be seen that the PCM performs the best, both in terms of number of
iterations and CPU time for all test cases. These numerical results illustrate that if the step size
condition relaxed, can indeed be beneficial to yield larger step sizes, which could accelerate the
convergence of algorithm.

To verify the performance results of our algorithm, we plotted the approximation error relL =
kx − xk−1 k, relS = kAxk − bk achieved for n = 1500, s = 50 by each of the algorithms versus the
k

iteration number k in Figures 2 and 3, respectively. It is clear to see that PCM outperforms all the other
algorithms significantly in terms of number of iterations.

AIMS Mathematics Volume 8, Issue 8, 18295–18313.

 18308

104
PCM
L-ALM
C-PPA
102

100
||r||2

10-2

10-4

10-6
0 50 100 150 200 250 300 350 400 450
iter (k)

Figure 2. Relative errors of relL = kxk − xk−1 k.

106
PCM
L-ALM
C-PPA
104

102
||s||2

100

10-2

10-4
0 50 100 150 200 250 300 350 400 450
iter (k)

Figure 3. Relative errors of relS = kAxk − bk.

5.2. Matrix completion problem

Matrix completion problem (MC) comes from many fields such as signal processing, statistics,
machine learning communities. It tries to recover the low-rank matrix X from its incomplete known

AIMS Mathematics Volume 8, Issue 8, 18295–18313.

 18309

entries. Mathematically, its convex formula is as follows:

min kXk∗
(5.3)
s.t. Xi j = Mi j , (i, j) ∈ Ω,

where kXk∗ is the nuclear norm of X, M is the unknown matrix with p available sampled entries and
Ω is a set of pairs of indices of cardinality p. By invoking the first-order optimality condition, MC can
also be equivalent to VI (1.1) and (1.2) with f (x) = ∂kXk∗ .
The basic setup of the problem is as follows. We first generate two random matrices ML ∈ <n×ra and
MR ∈ <n×ra , all with i.i.d. standard Gaussian entries, and then set the low-rank matrix M = ML MRT .
The available index set Ω is randomly uniformly sampled in all cardinality sets |Ω|. We denote the
oversampling factor (OS) by OS = |Ω|/ra (2n − ra ), i.e., the ratio of sample size to degrees of freedom
in an asymmetric matrix of rank ra . The relative error of the approximation X is defined as

kXΩ − MΩ kF
relative error = . (5.4)
kMΩ kF

We set the relative error T ol = 10−5 as the tolerance for all algorithms. In all tests, the initial iteration
is (X 0 , Λ0 ) = (0, 0). The parameters used in the numerical experiments are set as follows: we set
r = 0.006, s = 1.01/(4r), γ = 1 for PCM, s = 1.01/r for L-ALM and C-PPA.
Tables 4–6 list the comparison between these algorithms with three different OS values. The results
confirm that PCM outperforms other methods in terms of computation time and number of iterations
in all cases.

Table 4. Comparison results of PCM, L-ALM, C-PPA (OS = 5).

Problems PCM L-ALM C-PPA
n ra Rel.err. Iter. Rel.err. Iter. Rel.err. Iter.
100 5 9.71e-06 60 6.22e-05 101 5.83e-05 101
100 10 9.74e-06 18 9.17e-06 38 9.15e-06 37
200 5 9.40e-06 69 1.41e-04 101 1.33e-04 101
200 10 8.88e-06 27 9.14e-06 56 9.15e-06 55
500 10 9.67e-06 43 9.29e-06 72 8.96e-06 71
500 15 8.93e-06 31 9.87e-06 50 9.06e-06 49
1000 10 9.21e-06 71 9.23e-06 119 9.52e-06 117
1500 10 9.92e-06 95 9.67e-06 166 9.41e-06 165

AIMS Mathematics Volume 8, Issue 8, 18295–18313.

 18310

Table 5. Comparison results of PCM, L-ALM, C-PPA (OS = 10).

Problems PCM L-ALM C-PPA
n ra Rel.err. Iter. Rel.err. Iter. Rel.err. Iter.
100 5 8.67e-06 14 7.66e-06 29 7.67e-06 28
100 10 8.39e-06 13 9.10e-06 27 9.12e-06 26
200 5 6.84e-06 21 8.85e-06 36 8.67e-06 35
200 10 5.78e-06 7 6.17e-06 13 8.32e-06 10
500 10 9.13e-06 24 8.22e-06 39 8.86e-06 38
500 15 8.34e-06 17 9.13e-06 26 9.74e-06 25
1000 10 8.22e-06 48 9.25e-06 77 9.33e-06 76
1500 10 9.41e-06 66 9.39e-06 112 9.40e-06 111

Table 6. Comparison results of PCM, L-ALM, C-PPA (OS = 15).

Problems PCM L-ALM C-PPA
n ra Rel.err. Iter. Rel.err. Iter. Rel.err. Iter.
100 5 9.62e-06 11 8.92e-06 23 8.99e-06 22
100 10 9.62e-06 13 9.15e-06 27 9.20e-06 26
200 5 4.73e-06 13 9.27e-06 21 8.34e-06 20
200 10 7.98e-06 6 6.15e-06 13 8.34e-06 10
500 10 8.15e-06 17 8.51e-06 25 9.93e-06 24
500 15 4.58e-06 10 4.22e-06 14 7.19e-06 13
1000 10 9.35e-06 35 9.38e-06 56 9.81e-06 55
1500 10 9.75e-06 50 9.20e-06 84 9.46e-06 83

6. Conclusions

This paper proposes a new prediction-correction method for solving the monotone variational
inequalities with linear constraints. At the prediction step, the implementation is carried out by a
simple projection. At the correction step, the method introduces a simple updating step to generate
the new iteration. We establish the global convergence of the method. The convergence condition of
the method also allows larger step sizes that can in potential make the algorithm numerically converge
faster. The numerical experiments approve the efficiency of the proposed methods. The future work
is to explore combining self adaptive technique for the method. Besides, further applications of our
method are expected.

Use of AI tools declaration

The authors declare they have not used Artificial Intelligence (AI) tools in the creation of this article.

AIMS Mathematics Volume 8, Issue 8, 18295–18313.

 18311

Acknowledgments

This research was funded by NSF of Shaanxi Province grant number 2020-JQ-485.

Conflict of interest

The authors declare no conflict of interest.

References

1. R. Rockafellar, J. Sun, Solving monotone stochastic variational inequalities and

complementarity problems by progressive hedging, Math. Program., 174 (2019), 453–471.
https://doi.org/10.1007/s10107-018-1251-y
2. D. Bertsekas, E. Gafni, Projection methods for variational inequalities with application to the
traffic assignment problem, In: Nondifferential and variational techniques in optimization, Berlin:
Springer, 1982, 39–159. https://doi.org/10.1007/BFb0120965
3. S. Dafermos, Traffic equilibrium and variational inequalities, Transport. Sci., 14 (1980), 42–54.
https://doi.org/10.1287/trsc.14.1.42
4. E. Gorbunov, N. Loizou, G. Gidel, Extragradient method: O (1/k) last-iterate convergence for
monotone variational inequalities and connections with cocoercivity, Proceedings of The 25th
International Conference on Artificial Intelligence and Statistics, 2022, 366–402.
5. G. Gu, J. Yang, Tight ergodic sublinear convergence rate of the relaxed proximal point algorithm for
monotone variational inequalities, J. Optim. Theory Appl., in press. https://doi.org/10.1007/s10957-
022-02058-3
6. B. He, W. Xu, H. Yang, X. Yuan, Solving over-production and supply-guarantee problems in
economic equilibria, Netw. Spat. Econ., 11 (2011), 127–138. https://doi.org/10.1007/s11067-008-
9095-2
7. Y. Gao, W. Zhang, An alternative extrapolation scheme of PDHGM for saddle point problem with
nonlinear function, Comput. Optim. Appl., 85 (2023), 263–291. https://doi.org/10.1007/s10589-
023-00453-8
8. X. Zhao, J. Yao, Y. Yao, A nonmonotone gradient method for constrained
multiobjective optimization problems, J. Nonlinear Var. Anal., 6 (2022), 693–706.
https://doi.org/10.23952/jnva.6.2022.6.07
9. S. Bubeck, Convex optimization: algorithms and complexity, Found. Trends Mach. Le., 8 (2015),
231–357. https://doi.org/10.1561/2200000050
10. T. Chan, S. Esedoglu, F. Park, A. Yip, Total variation image restoration: overview and recent
developments, In: Handbook of mathematical models in computer vision, Boston: Springer, 2006,
17–31. https://doi.org/10.1007/0-387-28831-7 2
11. G. Kutyniok, Theory and applications of compressed sensing, GAMM-Mitteilungen, 36 (2013),
79–101. https://doi.org/abs/10.1002/gamm.201310005

AIMS Mathematics Volume 8, Issue 8, 18295–18313.

 18312

12. E. Ryu, W. Yin, Large-scale convex optimization: algorithms & analyses via monotone operators,
Cambridge: Cambridge University Press, 2022. https://doi.org/10.1017/9781009160865
13. N. Parikh, S. Boyd, Proximal algorithms, Foundations and Trends R in Optimization, 1 (2014),
127–239. https://doi.org/10.1561/2400000003
14. M. Hestenes, Multiplier and gradient methods, J. Optim. Theory Appl., 4 (1969), 303–320.
https://doi.org/10.1007/BF00927673
15. M. Powell, A method for nonlinear constraints in minimization problems, In: Optimization, New
York: Academic Press, 1969, 283–298.
16. B. He, H. Yang, S. Wang, Alternating direction method with self-adaptive penalty parameters
for monotone variational inequalities, J. Optim. Theory Appl., 106 (2000), 337–356.
https://doi.org/10.1023/A:1004603514434
17. G. Qian, D. Han, L. Xu, H. Yang, Solving nonadditive traffic assignment problems: a self-adaptive
projection-auxiliary problem method for variational inequalities, J. Ind. Manag. Optim., 9 (2013),
255–274. https://doi.org/10.3934/jimo.2013.9.255
18. X. Fu, B. He, Self-adaptive projection-based prediction-correction method for constrained
variational inequalities, Front. Math. China, 5 (2010), 3–21. https://doi.org/10.1007/s11464-009-
0045-1
19. X. Fu, A general self-adaptive relaxed-PPA method for convex programming with linear
constraints, Abstr. Appl. Anal., 2013 (2013), 1–13. https://doi.org/10.1155/2013/492305
20. F. Ma, M. Ni, W. Tong, X. Wu, Matrix completion via extended linearized augmented Lagrangian
method of multipliers, Proceedings of International Conference on Informative and Cybernetics
for Computational Social Systems, 2015, 45–49. https://doi.org/10.1109/ICCSS.2015.7281147
21. Y. Shen, H. Wang, New augmented Lagrangian-based proximal point algorithm for convex
optimization with equality constraints, J. Optim. Theory Appl., 171 (2016), 251–261.
https://doi.org/10.1007/s10957-016-0991-1
22. B. He, L. Liao, X. Wang, Proximal-like contraction methods for monotone variational inequalities
in a unified framework II: general methods and numerical experiments, Comput. Optim. Appl., 51
(2012), 681–708. https://doi.org/10.1007/s10589-010-9373-z
23. B. He, F. Ma, X. Yuan, Optimal proximal augmented Lagrangian method and its application to full
jacobian splitting for multi-block separable convex minimization problems, IMA J. Numer. Anal.,
40 (2020), 1188–1216. https://doi.org/10.1093/imanum/dry092
24. G. Gu, B. He, X. Yuan, Customized proximal point algorithms for linearly constrained convex
minimization and saddle-point problems: a unified approach, Comput. Optim. Appl., 59 (2014),
135–161. https://doi.org/10.1007/s10589-013-9616-x
25. B. He, X. Yuan, W. Zhang, A customized proximal point algorithm for convex minimization with
linear constraints, Comput. Optim. Appl., 56 (2013), 559–572. https://doi.org/10.1007/s10589-013-
9564-5
26. F. Ma, On relaxation of some customized proximal point algorithms for convex minimization:
from variational inequality perspective, Comput. Optim. Appl., 73 (2019), 871–901.
https://doi.org/10.1007/s10589-019-00091-z

AIMS Mathematics Volume 8, Issue 8, 18295–18313.

 18313

27. F. Ma, M. Ni, A class of customized proximal point algorithms for linearly constrained convex
optimization, Comput. Appl. Math., 37 (2018), 896–911. https://doi.org/10.1007/s40314-016-
0371-3
28. B. He, F. Ma, X. Yuan, Optimally linearizing the alternating direction method of multipliers for
convex programming, Comput. Optim. Appl., 75 (2020), 361–388. https://doi.org/10.1007/s10589-
019-00152-3
29. J. Liu, J. Chen, J. Zheng, X. Zhang, Z. Wan, A new accelerated positive-indefinite proximal
ADMM for constrained separable convex optimization problems, J. Nonlinear Var. Anal., 6 (2022),
707–723. https://doi.org/10.23952/jnva.6.2022.6.08
30. B. He, F. Ma, S. Xu, X. Yuan, A generalized primal-dual algorithm with improved
convergence condition for saddle point problems, SIAM J. Imaging Sci., 15 (2022), 1157–1183.
https://doi.org/10.1137/21M1453463
31. F. Ma, Y. Bi, B. Gao, A prediction-correction-based primal-dual hybrid gradient method
for linearly constrained convex minimization, Numer. Algor., 82 (2019), 641–662
https://doi.org/10.1007/s11075-018-0618-8
32. S. Xu, A dual-primal balanced augmented Lagrangian method for linearly constrained convex
programming, J. Appl. Math. Comput., 69 (2023), 1015–1035. https://doi.org/10.1007/s12190-022-
01779-y
33. E. Blum, W. Oettli, Mathematische optimierung: grundlagen und verfahren, Berlin: Springer,
1975.
34. P. Mahey, A. Lenoir, A survey on operator splitting and decomposition of convex programs,
RAIRO-Oper. Res., 51 (2017), 17–41. https://doi.org/10.1051/ro/2015065
35. B. He, A uniform framework of contraction methods for convex optimization and
monotone variational inequality, Scientia Sinica Mathematica, 48 (2018), 255.
https://doi.org/10.1360/N012017-00034
36. J. Yang, X. Yuan, Linearized augmented Lagrangian and alternating direction methods for nuclear
norm minimization, Math. Comput., 82 (2013), 301–329. https://doi.org/10.1090/S0025-5718-
2012-02598-1
37. S. Boyd, N. Parikh, E. Chu, B. Peleato, J. Eckstein, Distributed optimization and statistical learning
via the alternating direction method of multipliers, Found. Trends Mach. Le., 3 (2011), 1–122.
https://doi.org/10.1561/2200000016
38. G. Ye, Y. Chen, X. Xie, Efficient variable selection in support vector machines via the alternating
direction method of multipliers, Proceedings of the 14th International Conference on Artificial
Intelligence and Statistics, 2011, 832–840.

c 2023 the Author(s), licensee AIMS Press. This

is an open access article distributed under the
terms of the Creative Commons Attribution License
(http://creativecommons.org/licenses/by/4.0)

AIMS Mathematics Volume 8, Issue 8, 18295–18313.