Preprint
Article

This version is not peer-reviewed.

SQP-Type Methods for Quaternion Negative Semidefinite Matrix Cone Constrained Optimization Problems

Submitted:

03 November 2026

Posted:

05 November 2026

You are already at the latest version

Abstract
This paper studies the nonlinear programming problem with quaternion self-conjugate negative semidefinite matrix constraints(QNSDP). A theoretical framework for QNSDP is constructed, and an efficient numerical algorithm is designed. First, a projection operator on the space of quaternion self-conjugate matrices is defined, and constraint qualifications such as the Robinson constraint qualification, the constraint nondegeneracy condition, and the strict complementarity condition are provided. It is proved that the Lagrange multiplier set is nonempty and bounded if and only if the Robinson constraint qualification holds, among other results. Second, the KKT conditions and second-order optimality conditions are established, and an explicit expression for the critical cone is given, then SQP-type methods for solving QNSDP is proposed. Under the different assumptions, the convergence of the algorithm is proved. Finally, numerical experiments verify the effectiveness of the algorithm.
Keywords: 
;  ;  ;  ;  

1. Introduction

Quaternion matrices have been widely used in various applications, including color image denoising and restoration [1] and color face recognition [2]. Quaternions are also widely used in other engineering fields. In most of these applications, the core mathematical models use quaternion matrices as decision variables, such as the low-rank quaternion decomposition (LRQD) model [3], the robust quaternion matrix completion (RQMC) model [4], and the low-rank quaternion approximation (LRQA) model [5]. In sparse optimization, the generalized subdifferential of the l 0 norm and the rank function of real matrix variables have been used; in previous studies, there have also been studies on the derivatives of real-valued functions of quaternion variables and the subgradients of norms of quaternion matrix variables [6]. With the continuous expansion of applications of quaternion matrices in signal processing, image processing, and other engineering fields, how to establish optimization models suitable for quaternion matrices and solve them effectively has gradually become a problem worthy of study. In particular, when the optimization problem involves quaternion self-conjugate semidefinite matrix constraints, the corresponding constrained optimization theory and numerical algorithms need to be further studied.
This paper studies the following Quaternion Negative Semidefinite Matrix Cone Constrained Optimization Problems (QNSDP) :
min x ∈ R n f ( x ) s . t . h ( x ) = 0 , G ( x ) ⪯ 0 .
where f : R n → R , h : R n → R l , G : R n × n → H m × m .
The QNSDP problem can be regarded as a generalization of semidefinite programming problems. Quaternion self-conjugate matrices have good properties. By extending semidefinite constraints to the space of quaternion self-conjugate matrices, a nonlinear semidefinite programming problem under quaternion matrix constraints can be established. Compared with traditional semidefinite programming over real symmetric matrices, QNSDP extends the constraints to the space of quaternion self-conjugate negative semidefinite matrices, and can provide a corresponding mathematical description for optimization models involving quaternion matrix variables. Therefore, studying the optimality conditions, constraint qualifications, and numerical solution methods for QNSDP not only helps to improve the theoretical system of quaternion matrix constrained optimization, but also has certain practical application significance [7].
An Sequential Quadratic Programming (SQP) is a classic method for effectively solving nonlinear constrained optimization problems. It is an extension of SQP-type methods in nonlinear programming to the semidefinite cone space. The main idea of an SQP method is to transform the original problem into a series of quadratic programming subproblems. If x * is a local minimum point, then the set of Lagrange multipliers corresponding to x * is non-empty, and there is a quadratic programming problem of the following form:
min d ∈ R n ∇ f ( x ) T d + 1 2 d T M d s . t . h ( x ) + D h ( x ) d = 0 , G ( x ) + D G ( x ) d ⪯ 0 .
Nowadays, there have been studies on a kind of SQP-type methods. Chen and Miao [8] proposed a trust-region-based penalty-free Sequential Semidefinite Programming Algorithm. This method does not require an additional merit function and can converge to a feasible point where the constraint qualification fails. However, the proof of its global convergence requires an additional feasibility restoration phase. Zhao and Chen [9] gave a penalty-free Sequential Semidefinite Programming algorithm, which overcomes the difficulty of failing to achieve a full step size near the optimal solution by solving a second-order correction step, and has a superlinear convergence rate. When the initial point is far from the optimal solution, the linearized constraints may be incompatible. Zhao and Chen [10] designed a two-phase algorithm: the first phase minimizes the violation of the linearized constraints to determine whether the subproblem constraints are compatible; the second phase solves a quadratic semidefinite programming subproblem. When the linearized constraints are compatible, this subproblem is consistent with the classical problem; if they are incompatible, the search direction is obtained by minimizing a quadratic linear approximation model of the l 1 exact penalty function, thereby solving the incompatibility problem of the quadratic semidefinite programming subproblem in the classical algorithm. These studies show that a kind of SQP-type methods provide an effective approach for solving semidefinite programming problems, and also provide a reference for studying nonlinear semidefinite programming algorithms under quaternion matrix constraints.
Based on the above research background, this paper mainly investigates the theory and algorithms for optimization problems under quaternion self-conjugate negative semidefinite matrix constraints (QNSDP). The mathematical theoretical foundation of quaternion matrix constrained optimization problems is constructed, a constrained optimization theoretical framework suitable for quaternion matrices is established, and an SQP algorithm suitable for quaternion matrix constraints is designed. A corresponding quadratic programming subproblem is constructed to ensure the feasibility of the algorithm in the quaternion field, and the convergence theory of the algorithm is established together with numerical verification.
The structure of this paper is organized as follows. Section 2 introduces the basic knowledge of quaternions and quaternion matrices, including quaternion self-conjugate matrices, related spectral properties. Section 3 studies the optimization theory of QNSDP, introduces the Robinson constraint qualification, the constraint nondegeneracy condition, and the strict complementarity condition, establishes the corresponding KKT conditions and second-order optimality conditions, and gives an expression for the critical cone. Section 4 introduces the classical SQP method, constructs an SQP algorithm, and analyzes the local convergence properties of the algorithm, and presents numerical experiments, in which the numerical solution performance of the algorithm is verified through specific examples. Section 5 summarizes the full text.

2. Preliminaries

2.1. Quaternion

Throughout this paper, R , C , Q represents the sets of real numbers, complex numbers, and quaternions, respectively. The notation and results in this section follow those in [11], to which interested readers are referred for further details.
Definition 1.
Suppose
q = a + b i + c j + d k , a , b , c , d ∈ R
A quaternion q consists of a real unit a and three imaginary units, denoted by i , j , k , which satisfy the following relations:
i 2 = j 2 = k 2 = i j k = − 1 ,
i j = − j i = k , j k = − k j = i , k i = − i k = j .
Let q 1 = a 1 + b 1 i + c 1 j + d 1 k , q 2 = a 2 + b 2 i + c 2 j + d 2 k ,
Performing addition and subtraction yields:
q 1 ± q 2 = ( a 1 ± a 2 ) + ( b 1 ± b 2 ) i + ( c 1 ± c 2 ) j + ( d 1 ± d 2 ) k .
Multiplication is noncommutative, and performing multiplication yields:
q 1 q 2 = ( a 1 a 2 − b 1 b 2 − c 1 c 2 − d 1 d 2 ) + ( a 1 b 2 + b 1 a 2 + c 1 d 2 − d 1 c 2 ) i + ( a 1 c 2 + c 1 a 2 + d 1 b 2 − b 1 d 2 ) j + ( a 1 d 2 + d 1 a 2 + b 1 c 2 − c 1 b 2 ) k , q 2 q 1 = ( a 1 a 2 − b 1 b 2 − c 1 c 2 − d 1 d 2 ) + ( b 2 a 1 + a 2 b 1 − d 2 c 1 + c 2 d 1 ) i + ( c 2 a 1 + a 2 c 1 + d 2 b 2 − b 1 d 1 ) j + ( a 1 d 2 + d 1 a 2 + b 1 c 2 − c 1 b 2 ) k .
Definition 2.
The conjugate q * of the quaternion q = a + b i + c j + d k is defined as follows:
q * = a − b i − c j − d k , a , b , c , d ∈ R

2.2. Quaternion Matrix

Definition 3.
A quaternion matrix A = ( a i j ) ∈ Q m × n can be represented as:
A = A 0 + A 1 i + A 2 j + A 3 k .
where A 0 , A 1 , A 2 , A 3 ∈ R m × n . The transpose of A is A T = a j i , the conjugate of A is A ¯ = ( a i j * ) , and the conjugate transpose of A is A * = ( a j i * ) = A ¯ T . For A , B ∈ Q m × n , the inner product is defined as:
〈 A , B 〉 = R e ( t r ( A * B ) ) .
Clearly obtained:
〈 A , B 〉 = 〈 B , A 〉 .
The Frobenius norm is defined as:
∥ A ∥ F = 〈 A , A 〉 = Tr ( A * A ) = ∑ i = 1 m ∑ j = 1 n | a i j | 2 .
Definition 4.
Any m × n quaternion matrix A over the quaternion set Q can be uniquely represented as:
A = A 1 + A 2 j , A 1 , A 2 ∈ C m × n .
This expression is the decomposition of the quaternion matrix over the complex field, called
A σ = A 1 − A 2 A ¯ 2 A ¯ 1 ∈ C 2 m × 2 n .
the complex representation matrix of the quaternion matrix A.
For a quaternion matrix A, the real representation is the following 4 m × 4 n matrix:
A R = A 0 − A 1 − A 2 − A 3 A 1 A 0 − A 3 A 2 A 2 A 3 A 0 − A 1 A 3 − A 2 A 1 A 0 .
Definition 5.
Suppose A ∈ Q n × n , a quaternion self-conjugate matrix A satisfies the following equation:
A * = A ,
The set of all n-order quaternion self-conjugate matrices is denoted by H n × n , if U ∈ Q n × n , and U U * = U * U = I , then U is called a generalized unitary matrix, the set of all n-order generalized unitary matrices is denoted by U n × n . On this basis, this paper gives the definition of quaternion negative semidefinite matrices:
Definition 6.
Suppose A ∈ H n × n , if for any 0 ≠ x = ( x 1 , … , x n ) T ∈ Q n × 1 , we can obtain
x * A x < 0 ( ≤ 0 ) ,
then A is called a quaternion negative (semi-)definite matrix. If A is a negative semidefinite matrix, then A ⪯ 0 . The n-order quaternion negative semidefinite matrix is denoted as Q − n × n , and the n-order quaternion self-conjugate negative semidefinite matrix is denoted as H − n × n .
If A ⪰ 0 , B ⪯ 0 , easy to obtain:
〈 A , B 〉 ≤ 0 .
Definition 7.
Suppose A ∈ Q n × n , if there exists λ ∈ Q and 0 ≠ α ∈ Q n × 1 , s.t.
A α = α λ ( or A α = λ α ) ,
then λ is called the right (or left) eigenvalue of A, and α is called the eigenvector of A corresponding to the right (or left) eigenvalue λ. If λ is both the right and left eigenvalue of A, then λ is called the eigenvalue of A. Generally, the left and right eigenvalues are not equal, but under the quaternion self-conjugate matrix, its left and right eigenvalues are equal.
Proposition 1.
[11] If A ∈ H n × n , then the right eigenvalues of A are equal to the left eigenvalues of A, and they must be real numbers.
We have eigenvalue decomposition theorem for quaternion self-conjugate matrices:
Proposition 2.
[12] Suppose A ∈ H n × n , then there exists U ∈ U n × n , such that
U * A U = λ 1 λ 2 ⋱ λ n .
where λ 1 ≥ λ 2 ≥ … ≥ λ n are the n eigenvalues of A, U is a quaternion unitary matrix.
According to the above proposition, it is easy to know:
Proposition 3.
H − n × n = Q n × n | λ Q ≤ 0 .

3. Optimization Theory

3.1. Constrained Norms and KKT Conditions

For the QNSDP problem:
min x ∈ R n f ( x ) s . t . h ( x ) = 0 , G ( x ) ⪯ 0 .
Its Lagrangian function is:
L ( x , μ , Y ) = f ( x ) + μ T h ( x ) + 〈 Y , G ( x ) 〉 , μ ∈ R l , Y ∈ H m × m
where f : R n → R , h : R n → R l , G : R n × n → H m × m .   f , h , G are continuously differentiable, and there exist ( μ , Y ) ∈ R l × H m × m , where μ , Y are the Lagrange multipliers for the equality constraints and the negative semidefinite matrix constraint, respectively.
Proposition 4.
For the nonconvex QNSDP problem (1), T H − m × m ( A ) is the tangent cone to the negative semidefinite cone H − m × m at the point A:
T H − m × m ( A ) = Z ∈ H m × m | U 0 * Z U 0 ⪯ 0 .
where U 0 denotes the matrix whose columns are the eigenvectors corresponding to the zero eigenvalues of A.
Proof 
(Proof of Proposition 4). The tangent cone at A ∈ H − m × m is defined as:
T H − m × m ( A ) = Z ∈ H m × m : ∃ t k ↓ 0 , Z k → Z , s . t . A + t k Z k ∈ H − m × m .
i.e. Z ∈ T H − m × m ( A ) if and only if there exists t k ↓ 0 and Z k → Z such that A + t k Z k ⪯ 0 .
Perform eigenvalue decomposition on A, and there exists a unitary matrix U such that
A = U Λ U * , Λ = d i a g ( λ 1 , … , λ m ) ,
where λ i ∈ R , U 0 ∈ H m × q is a matrix U 0 consisting of feature vectors corresponding to zero eigenvalues.
Suppose Z ∈ T H − m × m ( A ) , take any y ∈ H q , set x = U 0 y , then
x * ( A + t k Z k ) x ≤ 0 ,
from A U 0 = 0 , we can obtain
x * A x = y * U 0 * A U 0 y = 0 ,
so
t k y * U 0 * Z k U 0 y ≤ 0 .
from t k > 0 , we can obtain
y * U 0 * A U 0 y ≤ 0 .
that is, U 0 * Z k U 0 ⪯ 0 holds for all k, set k → ∞ , from Z K → Z and the closedness of the semi-negative definite cone, we have: U 0 * Z U 0 ⪯ 0 . □
Definition 8.(Robinson constraint qualification)[13] f , h , G are continuously differentiable functions, the Robinson constraint qualification for the QNSDP problem (1) is said to hold at x if:
D h ( x ) D G ( x ) R n + 0 l T H − m × m ( G ( x ) ) = R l H m × m .
Proposition 5.
Suppose f , h , G be continuously differentiable functions, and x * be a local optimal solution to the QNSDP (1) problem. Denote M ( x * ) as the set of corresponding multipliers ( μ , Y ) at x * . M ( x * ) is nonempty and bounded if and only if the Robinson constraint qualification holds at x * , meaning there exists a pair of Lagrange multipliers ( μ , Y ) such that:
∇ f ( x ) + D h ( x ) T μ + D G ( x ) * Y = 0 , h ( x ) = 0 , G ( x ) ⪯ 0 , Y ⪰ 0 , 〈 Y , G ( x ) 〉 = 0 .
Proof 
(Proof of Proposition 5). Denote z = ( μ , Y ) ∈ R l × H m × m . The KKT system of the QNSDP problem (1) can be written as a generalized equation:
0 ∈ F ( x , z ) + N K ( z ) ,
where F ( x , z ) = ∇ f ( x ) + D h ( x ) T μ + D G ( x ) * Y − h ( x ) − G ( x ) , K = { 0 } l × H − m × m is a convex cone, N K ( z ) is the normal cone to z at K .
First prove sufficiency: assume that the Robinson constraint qualification holds at x * , and prove that M ( x * ) is nonempty and bounded.
By Definition 8, the Robinson constraint qualification holds, and the linearized mapping is surjective. By the Robinson constraint qualification theorem in variational analysis [13], if the above surjectivity condition holds, then the solution mapping M ( x ) of the KKT system is upper semicontinuous in some neighborhood of x * , and M ( x * ) is a nonempty compact set. Moreover, by the classical Robinson multiplier existence theorem, as long as x * is a local optimal solution and satisfies the Robinson constraint qualification, the corresponding set M ( x * ) is nonempty.
Assuming that M ( x * ) is unbounded, there exists a sequence of pairs ( μ k , Y k ) ∈ M ( x * ) such that:
ρ k = | | μ k | | 2 + | | Y k | | F 2 → + ∞ ,
set μ ^ k = μ k ρ k , Y ^ k = Y k ρ k ,
Due to
μ ^ k 2 + Y ^ k F 2 = 1 ,
and by the compactness of bounded sequences in finite-dimensional spaces, we can take a convergent subsequence such that:
( μ ^ k , Y ^ k ) → ( μ ¯ , Y ¯ ) , ∥ μ ¯ ∥ 2 + ∥ Y ¯ ∥ F 2 = 1 .
Based on the KKT stationary condition, divide both sides by ρ k , and then let k → ∞ to obtain:
D h ( x * ) T μ ¯ + D G ( x * ) * Y ¯ = 0 .
and from the complementary condition satisfied by the KKT multipliers, we have:
Y ¯ ⪰ 0 , 〈 Y ¯ , G ( x * ) 〉 = 0 .
For ∀ W ∈ T H − m × m ( G ( x * ) ) , it holds that:
〈 Y ¯ , W 〉 ≤ 0 .
According to the definition of tangent cone, there exist t j ↓ 0 and W j → W , such that G ( x * ) + t j W j ⪯ 0 . Therefore
0 ≥ 〈 Y ¯ , G ( x * ) + t j W j 〉 = t j 〈 Y ¯ , W j 〉 ,
according to Definition 8, there exist d ∈ R n and W ∈ T H − m × m ( G ( x * ) ) such that
μ ¯ = D h x * d , Y ¯ = D G x * d + W .
From equations (3) and (4), we get:
∥ μ ¯ ∥ 2 + ∥ Y ¯ ∥ F 2 = 〈 μ ¯ , D h ( x * ) d 〉 + 〈 Y ¯ , D G ( x * ) d + W 〉 = 〈 D h ( x * ) T μ ¯ + D G ( x * ) * Y ¯ , d 〉 + 〈 Y ¯ , W 〉 = 〈 Y ¯ , W 〉 ≤ 0 .
this contradicts equation (2); hence M ( x * ) is bounded.
Now prove necessity: assume that M ( x * ) is nonempty and bounded, and prove that the Robinson constraint qualification holds at x * .
If M ( x * ) is nonempty and bounded, then by the Ioffe–Tikhomirov theorem [13], the solution mapping of the KKT system at x * is stable, which is equivalent to the metric regularity of the constraint system [13]; this is an equivalent characterization of the Robinson constraint qualification, namely the surjectivity of the linearized system:
D h ( x ) D G ( x ) R n + 0 l T H − m × m ( G ( x ) ) = R l H m × m .
thus the Robinson constraint qualification holds, and the proof is complete. □
Definition 9.(Constraint Nondegeneracy Condition)[13]Let f , h , G be continuously differentiable functions. If the following equation holds, then the constraint nondegeneracy condition is said to hold:
D h ( x ) D G ( x ) R n + { 0 } l lin T H − m × m ( G ( x ) ) = R l H m × m .
Definition 10.(Strict Complementarity Condition)[13]If
rank ( G ( x * ) ) + rank ( Y * ) = m ,
then the QNSDP problem (1) is said to satisfy the strict complementarity condition at the point x * .
Proposition 6.
[13]Let f , h , G be continuously differentiable functions, and x * be the local optimal solution to the problem. The following conclusion is correct:
(1) If the constraint nondegeneracy condition holds at the point x * , then M ( x ) is a singleton, that is, the Lagrange multiplier exists and is unique;
(2) If M ( x ) is a singleton and the strict complementarity condition holds at the point x * , then the QNSDP problem (1) satisfies the constraint nondegeneracy condition at x * .
Proposition 7.
[15] Define the projection operator from the symmetric matrix space C m to the negative semidefinite matrix space C − m as C m → C − m , due to the strong semi-smoothness of the projection operator, at G 0 ∈ C m , there exists an operator γ 0 ∈ ∂ Π C − m ( G 0 + △ G ) , △ G → 0 such that:
Π C − m ( G 0 + Δ G ) = Π C − m ( G 0 ) + γ 0 ( Δ G ) + O ( ∥ Δ G ∥ F 2 ) .
Theorem 1.
For the projection operator:
Π : H m × m → H − m × m ,
at G 0 ∈ H m × m , there exists an operator γ 0 ∈ Π H − m × m ( G 0 + △ G ) , △ G → 0 such that
Π H − m × m ( G 0 + △ G ) = Π H − m × m ( G 0 ) + γ 0 ( △ G ) + O ( | | △ G | | F 2 ) .
Proof 
(Proof of Theorem 1). Let a matrix A ∈ H m × m have the eigenvalue decomposition:
A = P λ 1 ( A ) λ 2 ( A ) ⋱ λ n ( A ) P * .
where P is a quaternion unitary matrix, P P * = I , λ i ( A ) ∈ R , suppose there exist two integers s , s 1 such that
λ 1 ( x ) ≤ … ≤ λ s ( x ) < 0 = λ s + 1 ( x ) = … = λ s 1 ( x ) < λ s 1 + 1 ( x ) ≤ … ≤ λ r ( x ) .
then the projection Π H − m × m of A on H − m × m satisfies:
Π H − m × m ( A ) = P [ λ 1 ( A ) ] − [ λ 2 ( A ) ] − ⋱ [ λ n ( A ) ] − P * = P λ 1 ⋱ λ s 0 ⋱ 0 P * .
The mapping of A ∈ H m × m to the complex matrix A ˜ in C 2 m × 2 m corresponds to the Hermitian cone H − m × m of quaternion self-conjugate negative semidefinite matrices, which is equivalent to the Hermitian cone H − 2 m × 2 m of Hermitian negative semidefinite matrices in C 2 m × 2 m . Therefore, it is equivalent to proving that the projection operator Π H − 2 m × 2 m from the complex Hermitian matrix space to the Hermitian cone of negative semidefinite matrices is strongly semi-smooth.
Suppose X ∈ H n × n ( n = 2 m ) , performing the eigenvalue decomposition, we can obtain:
X = U Λ U * , Λ = d i a g ( λ 1 , … , λ n ) .
The projection of X onto the negative semidefinite cone H − n × n is:
Π H − n ( X ) = U diag min λ 1 , 0 , … , min λ n , 0 U * .
The quaternion self-conjugate matrix space H m × m is linearly isomorphic to the complex Hermitian matrix space H n × n , and the projection operators are equivalent. According to Proposition 7, and since strong semi-smoothness remains unchanged under isomorphism, it follows that Π H − m × m is strongly semi-smooth. □

3.2. Second-order Optimality Conditions

Definition 11.(Critical Cone)[13]Suppose f , h , G be continuously differentiable functions, and let x * be a local optimal solution of the QNSDP problem (1). The critical cone is defined as:
C ( x * ) = d ∈ R n | D h ( x * ) d = 0 , D G ( x * ) d ∈ T H − m × m ( G ( x * ) ) , ∇ f ( x * ) T d ≤ 0 .
Proposition 8.
[13] If M ( x * ) is nonempty, then we have:
C ( x * ) = d ∈ R n | D h ( x * ) d = 0 , D G ( x * ) d ∈ T H − m × m ( G ( x * ) ) , ∇ f ( x * ) T d = 0 .
Proposition 9.(Second-Order Sufficient Condition)[16]Suppose f , h , G be second-order continuously differentiable functions, and ( x * , μ , Y ) be a point satisfying the KKT conditions. If the Robinson constraint qualification holds and
d T ∇ x x 2 L ( x * , μ , Y ) d > 0 , for all d ∈ C ( x * ) ∖ { 0 } .
then x * is a strict local minimizer of the QNSDP problem (1).
Definition 12.(Second-Order Growth Condition)[13]If x * is a feasible point of the QNSDP(1) problem, and there exist a constant b > 0 and a neighborhood B ( x * ) of x * , such that:
f ( x ) ≥ f ( x * ) + b | | x − x * | | 2 , x ∈ B ( x * ) .
then the second-order growth condition is said to hold at the point x * .

4. An SQP-type Method

4.1. Classic SQP Algorithm

If x * is a local minimizer, then the set of Lagrange multipliers corresponding to x * is nonempty, and the quadratic programming subproblem with equality and inequality constraints takes the following form:
min d ∈ R n ∇ f ( x ) T d + 1 2 d T M d s . t . h ( x ) + D h ( x ) d = 0 , G ( x ) + D G ( x ) d ⪯ 0 .
where M ∈ S n is a real symmetric positive definite matrix. The Lagrangian function of this subproblem is:
L ( d , μ , Y ) = ∇ f ( x ) T d + 1 2 d T M d + μ T ( h ( x ) + D h ( x ) d ) + Y , G ( x ) + D G ( x ) d .
Under some constraint qualification, the KKT conditions are satisfied:
∇ f ( x ) + M d * + D h ( x ) T μ + D G ( x ) * Y h ( x ) + D h ( x ) d Π H − m × m ( G ( x ) + D G ( x ) d + Y ) − ( G ( x ) + D G ( x ) d ) = 0 .
This paper studies the SQP method based on the QNSDP problem (1). This method is a numerical solution method for nonconvex matrix optimization problems. Its core idea is to transform the complex nonconvex problem into a series of efficiently solvable convex programming subproblems (5) through successive convexification, and to approximate the optimal solution of the original problem by these subproblems. At each iteration point x k , the trial step d k is obtained by minimizing the quadratic programming subproblem:
min d ∈ R n ∇ f x k T d + 1 2 d T M k d s . t . h x k + D h x k d = 0 , G x k + D G x k d ⪯ 0 .
where M k is the Hessian matrix of the above Lagrangian function. The specific algorithm is stated as follows:
step 0: Given an initial point x 0 ∈ R n , an initial Hessian approximation M 0 , a maximum number of iterations n m a x , convergence tolerances t o l d k , t o l v , and a condition number threshold c o n d ¯ .
step 1: While k ≤ n m a x , execute the following steps:
step 2: Compute the gradient ∇ f ( x k ) of the objective function at x k .
step 3: Solve the subproblem to obtain the search direction d k , and the Lagrange multipliers corresponding to the equality constraints and the matrix inequality constraint, respectively μ k + 1 , Y k + 1 .
step 4: Preliminary convergence check: if | | d k | | ≤ t o l d k , then compute the constraint violation v ( x k ) at the current point, if v ( x k ) ≤ t o l v , then Break.
step 5: Determine the step size α k by line search.
step 6: Let x k + 1 = x k + α k d k . If x k + 1 is invalid, or if | | x k + 1 − x k | | < 10 − 6 , then perturb the search direction d k and perform a line search again to obtain a new x k + 1 .
step 7: Compute the gradient difference y k = ∇ f ( x k + 1 ) − ∇ f k , the iterate difference s k = x k + 1 − x k , and update the Hessian approximation M k + 1 .
step 8: M k + 1 is updated by the modified BFGS method [18], that is,
M k + 1 = M k − M k s k s k T M k s k T M k s k + y ^ k y ^ k T s k T y ^ k .
where:
s k = x k + 1 − x k , y ^ k = ψ k y k + ( 1 − ψ k ) M k s k , y k = ∇ x L ( x k + 1 , μ k + 1 , Y k + 1 ) − ∇ z L ( x k , μ k + 1 , Y k + 1 ) , ψ k = 1 , if s k T y k ≥ 0.2 s k T M k s k , 0.8 s k T M k s k s k T ( M k s k − y k ) , otherwise .
step 9: If c o n d ( M k + 1 ) > c o n d ¯ , then Break.
step 10: k:=k+1.
step 11: End while.
( x k , μ k , Y k ) is the iterative sequence generated by the above algorithm, the convergence of this sequence will be analyzed below.

4.2. Convergence Analysis

To discuss the convergence rate of the sequence ( x k , μ k , Y k ) , first make the following assumptions:
A f , h , G are twice continuously differentiable functions.
B The constraint nondegeneracy condition for problem (1) holds at the point x * .
C Let ( x * , μ , Y ) ∈ R n × R l × H m × m be a KKT point of problem (1). For any d ∈ C ( x * ) ∖ 0 , we can have:
d T ∇ x x 2 L ( x * , μ , Y ) d > 0 .
D The strict complementarity condition for problem (1) holds at the point x * .
Theorem 2.
Assuming that A, B, and C hold, if M k − ∇ x x 2 L ( x * , μ , Y ) = o ( 1 ) , then there exists δ > 0 such that when | | x 0 − x * | | < δ , | | μ 0 − μ | | < δ , | | Y 0 − Y | | < δ , the algorithm is defined, and the sequence ( x k , μ k , Y k ) in the algorithm converges superlinearly to ( x * , μ , Y ) . Specifically, when M k − ∇ x x 2 L ( x * , μ , Y ) = O ( | | x k − x * | | + | | ( μ k , Y k ) − ( μ , Y ) | | ) and the second derivatives of f , h , G are locally Lipschitz continuous at point x * , the sequence ( x k , μ k , Y k ) converges second-order to ( x * , μ , Y ) .
Proof 
(Proof of Theorem 2). Define z = ( x , μ , Y ) ∈ ( R n × R l × H m × m ) , the KKT conditions can be written as: 0 ∈ F ( z ) + N ( z ) , where:
F ( z ) = ∇ x L ( x , μ , Y ) h ( x ) G ( x ) + Y , N ( z ) = 0 × 0 × N H − m × m ( G ( x ) ) .
In H m × m , the projection Π H − m × m is strongly semismooth, and from the Robinson constraint qualification and the second-order sufficient condition it follows that the KKT system is strongly regular at z * , consequently, there exists a neighborhood U ( z * ) such that the generalized equation
0 ∈ F ( z * ) + ∇ F ( z * ) ( z − z * ) + N ( z ) ,
has a unique solution z.
The KKT conditions for the QNSDP subproblem (5) are:
0 ∈ F ( z k ) + J k ( z k + 1 − z k ) + N ( z k + 1 ) .
where
J k = M k D h ( x k ) T D G ( x k ) * D h ( x k ) 0 0 D G ( x k ) 0 I ,
D G ( x k ) * is the adjoint operator:
D G ( x k ) * Y = ( Y , ∂ G ( x ) ∂ x 1 , Y , ∂ G ( x ) ∂ x 2 , … , Y , ∂ G ( x ) ∂ x n ) T .
Set ▵ k = J k − ∇ F ( z * ) , suppose | | ▵ k | | = o ( 1 ) , by strong regularity, there exists L > 0 such that:
∥ z k + 1 − z * ∥ ≤ L ( ∥ F ( z * ) − F ( z k ) − ∇ F ( z * ) ( z k − z * ) ∥ + ∥ ▵ k ∥ ∥ z k + 1 − z * ∥ ) = o ( ∥ z k − z * ∥ ) + L ∥ ▵ k ∥ ∥ z k + 1 − z * ∥ = o ( ∥ z k − z * ∥ ) + L o ( ∥ z k + 1 − z k ∥ ) .
thus the sequence is proved to converge superlinearly.
If | | ▵ k | | = O ( | | z k − z * | | ) , since ∇ F is Lipschitz continuous, then
∥ F ( z * ) − F ( z k ) − ∇ F ( z * ) ( z k − z * ) ∥ = O ( ∥ z k − z * ∥ 2 ) .
substituting into (6) yields: ∥ z k + 1 − z * ∥ = O ( ∥ z k − z * ∥ 2 ) , which proves that the sequence converges quadratically. □
Lemma 1.
[19]Assuming that A, B, and C hold, the second derivatives of f , h , G are locally Lipschitz continuous near the point x * , the sequence ( x k , μ k , Y k ) is generated by an algorithm, M k is uniformly positive definite and uniformly bounded, and satisfies M k − ∇ x x 2 L ( x * , μ , Y ) = O ( | | x k − x * | | + | | ( μ k , Y k ) − ( μ , Y ) | | ) , then there exists δ > 0 such that when the initial point satisfies | | x k − x * | | < δ , | | μ k − μ | | < δ , | | Y k − Y | | < δ , we can obtain:
| | d k | | + | | μ k + 1 − μ | | + | | Y k + 1 − Y | | F = O ( | | x k − x * | | + | | ( μ k , Y k ) − ( μ , Y ) | | ) .
Theorem 3.
Assuming that A, B, C, and D hold, the second derivatives of f , h , G are locally Lipschitz continuous near the point x * , the sequence ( x k , μ k , Y k ) is generated by an algorithm, and M k is uniformly positive definite and uniformly bounded, and satisfies
M k − ∇ x x 2 L ( x * , μ , Y ) = O ( | | x k − x * | | + | | ( μ k , Y k ) − ( μ , Y ) | | F ) ,
it can be deduced that ∃ δ > 0 , and when the initial point satisfies | | x k − x * | | < δ , | | μ k − μ | | < δ , | | Y k − Y | | < δ , the sequence ( x k , μ k , Y k ) converges quadratically to ( x * , μ , Y ) , and x k converges to x * superlinearly.
Proof 
(Proof of Theorem 3). The quadratic convergence of the sequence is first proved. From Lemma 1, it follows that:
| | d k | | = O ( ε k ) , | | μ k + 1 − μ | | = O ( ε k ) , | | Y k + 1 − Y | | F = O ( ε k ) ,
where ε k = | | ( x k , μ k , Y k ) − ( x * , μ , Y ) | | .
Expanding the equality constraints by Taylor’s theorem, it obtained that:
0 = h ( x k ) + D h ( x k ) d k = h ( x * ) + D h ( x * ) ( x k − x * ) + O ( | | x k − x * | | 2 ) + D h ( x * ) d k + O ( | | x k − x * | | | | d k | | ) = D h ( x * ) ( x k + 1 − x * ) + O ( ε k 2 ) .
Using the properties of the projection operator, the following complementarity conditions hold:
Π H − m × m ( G ( x * ) + Y ) = G ( x * ) ,
Π H − m × m ( G ( x k ) + D G ( x k ) d k + Y k + 1 ) = G ( x k ) + D G ( x k ) d k ,
Since the projection Π H − m × m is strongly semismooth, there exists a generalized Jacobian matrix J * ∈ ∂ Π H − m × m ( G ( x * ) + Y ) such that:
Π H − m × m ( G ( x * ) + Y ) = Π H − m × m ( G ( x k ) + D G ( x k ) d k + Y k + 1 ) + J * ( G ( x * ) − G ( x k ) − D G ( x k ) d k − Y k + 1 + Y ) + O ( | | ( G ( x * ) − G ( x k ) − D G ( x k ) d k − Y k + 1 + Y ) | | F 2 ) .
Analogously to (7), we can obtain the matrix constraint:
G ( x k ) + D G ( x k ) d k − G ( x * ) = D G ( x * ) ( x k + 1 − x * ) + O ( ε k 2 ) ,
Therefore, we can obtain:
| | G ( x * ) − G ( x k ) − D G ( x k ) d k − Y k + 1 + Y | | F = O ( ε k + 1 ) + O ( ε k 2 ) .
From equations (8), (9), and (10), and using the complementarity property of the projection, we can obtain:
( J * − I ) D G ( x * ) ( x k + 1 − x * ) + J * ( Y k + 1 − Y ) = O ( ε k 2 ) + O ( ε k + 1 2 ) ,
The KKT conditions for the subproblem (5) of the QNSDP problem are:
∇ f ( x k ) + M k d k + D h ( x k ) T μ k + 1 + D G ( x k ) * Y k + 1 = 0 .
Since ( M k − ∇ x x 2 L ( x * , μ , Y ) ) d k = O ( ε k 2 ) , expanding, we have:
0 = ∇ f ( x k ) + D h ( x k ) T μ + D G ( x k ) * Y + ∇ x x 2 L ( x * , μ , Y ) d k + O ( ε k 2 ) + D h ( x k ) T ( μ k + 1 − μ ) + D G ( x k ) * ( Y k + 1 − Y ) .
Substitute ∇ x L ( x k , μ , Y ) − ∇ x L ( x * , μ , Y ) = ∇ x x 2 L ( x * , μ , Y ) ( x k − x * ) + O ( ε k 2 ) into (11), and from
D h ( x k ) T ( μ k + 1 − μ ) + D G ( x k ) * ( Y k + 1 − Y ) = D h ( x * ) T ( μ k + 1 − μ ) + D G ( x * ) * ( Y k + 1 − Y ) + O ( ε k ε k + 1 ) ,
holds:
∇ x x 2 L ( x * , μ , Y ) ( x k + 1 − x * ) + D h ( x * ) T ( μ k + 1 − μ ) + D G ( x * ) * ( Y k + 1 − Y ) = O ( ε k 2 ) + O ( ε k ε k + 1 ) .
Written in matrix form:
∇ x x 2 L ( x * , μ , Y ) D h ( x * ) T D G ( x * ) * D h ( x * ) 0 0 ( J * − I ) D G ( x * ) 0 J * x k + 1 − x * μ k + 1 − μ Y k + 1 − Y = O ( ε k 2 ) + O ( ε k + 1 ) .
By Definition 10, G ( x * ) + Y is nonsingular, so the projection operator is differentiable at this point, and the matrix operator A be written as:
A = ∇ x x 2 L ( x * , μ , Y ) D h ( x * ) T D G ( x * ) * D h ( x * ) 0 0 ( J * − I ) D G ( x * ) 0 J * .
further prove k e r ( A ) = 0 :
Suppose ( d , u , V ) satisfies:
A d u V = 0 ,
i.e.
∇ x x 2 L ( x * , μ , Y ) d + D h ( x * ) T u + D G ( x * ) * V = 0 , D h ( x * ) d = 0 , ( J * − I ) D G ( x * ) d + J * V = 0 .
From the KKT complementarity conditions and the strict complementarity condition, there exists a quaternion unitary matrix U such that:
U * G ( x * ) U = Λ − 0 0 0 , U * Y U = 0 0 0 Λ + ,
where Λ − ≺ 0 , Λ + ≻ 0 .
For the cross block, we can obtain:
( J * Z ) − + = Θ ∘ Z − + , θ i j = λ i λ i − η j ∈ ( 0 , 1 ) ,
where λ i < 0 < η j , ∘ denotes the elementwise product, under strict complementarity conditions, for any matrix Z, transform it to the above coordinate system, denoted as: Z ˜ = U * Z U , and write it in blocks as:
Z ˜ = Z − − Z − + Z + − Z + + .
Let C = D G ( x * ) d , from the third equation in (12), we can obtain:
V − − = 0 , C + + = 0 , V − + = ( Θ − 1 − I ) ∘ C − + ,
so
〈 D G ( x * ) d , V 〉 = 2 ∑ i , j 1 − θ i j θ i j | C i j | 2 ≥ 0 .
Since D h ( x * ) d = 0 , C + + = 0 , and the expression for the tangent cone of the negative semidefinite cone, we have obtain:
D G ( x * ) d ∈ T H − m × m ( G ( x * ) ) .
and from the KKT stationary point condition:
∇ f ( x * ) T d = − 〈 Y , D G ( x * ) d 〉 = 0 .
so
d ∈ C ( x * ) .
Take the inner product of the first equation in (12) with d, and use D h ( x * ) d = 0 to obtain:
d T ∇ x x 2 L ( x * , μ , Y ) d + 〈 D G ( x * ) d , V 〉 = 0 .
from equation (13), we can obtain d T ∇ x x 2 L ( x * , μ , Y ) d ≤ 0 . According to Proposition 9, if d ≠ 0 , then d T ∇ x x 2 L ( x * , μ , Y ) d > 0 , which is a contradiction. Therefore
d = 0 ,
Substituting into (12), we can obtain:
D h ( x * ) T u + D G ( x * ) * V = 0 , J * V = 0 .
From the eigenvalue decomposition of J * , V can only be nonzero on the diagonal block corresponding to the positive eigenvalues of Y, hence
〈 V , W 〉 = 0 , ∀ W ∈ lin T H − m × m ( G ( x * ) ) .
and from Definition 9 and D h ( x * ) T u + D G ( x * ) * V = 0 , we can obtain: u = 0 , V = 0 .
Therefore, if the homogeneous system of equations has only a zero solution, i.e., k e r ( A ) = 0 , then the operator A is nonsingular. Hence, ε k + 1 = O ( ε k 2 ) , meaning that the sequence ( x k , μ k , Y k ) converges to ( x * , μ , Y ) and x k at a quadratic convergence rate.
Further prove the superlinear convergence:
To prove that x k converges superlinearly to x * , it suffices to prove that:
| | x k + 1 − x * | | | | x k − x * | | → 0 ,
by the triangle inequality and the previous estimates | | d k | | = O ( ε k ) , where ε k = | | ( x k , μ k , Y k ) − ( x * , μ , Y ) | | .
Therefore, it suffices to prove that:
ε k + 1 = O ( | | x k − x * | | ε k ) .
Since h ( x k ) + D h ( x k ) d k = 0 , by Taylor expansion, we can obtain:
0 = h ( x * ) + D h ( x * ) ( x k − x * ) + O ( | | x k − x * | | 2 ) + D h ( x * ) d k + O ( | | x k − x * | | | | d k | | ) = h ( x * ) + D h ( x * ) ( x k − x * ) + D h ( x * ) d k + O ( | | x k − x * | | ε k ) = D h ( x * ) ( x k + 1 − x * ) + O ( | | x k − x * | | ε k ) ,
From G ( x k ) + D G ( x k ) d k ⪯ 0 , we get:
G ( x k ) + D G ( x k ) d k − G ( x * ) = D G ( x * ) ( x k − x * ) + O ( | | x k − x * | | 2 ) + D G ( x * ) d k + O ( | | x k − x * | | | | d k | | ) = D G ( x * ) ( x k + 1 − x * ) + O ( | | x k − x * | | ε k ) ,
Thus, this expression is established:
| | G ( x * ) − G ( x k ) − D G ( x k ) d k − Y k + 1 + Y | | F = O ( | | x k − x * | | ε k ) + O ( ε k + 1 ) .
Analogously to the proof of quadratic convergence, we can get that the following expression holds:
( J * − I ) D G ( x * ) ( x k + 1 − x * ) + J * ( Y k + 1 − Y ) = O ( | | x k − x * | | ε k ) + O ( ε k + 1 2 ) .
From the KKT conditions, we can obtain:
( M k − ∇ x x 2 L ( x * , μ , Y ) ) d k = O ( | | d k | | ε k ) = O ( | | x k − x * | | ε k ) + O ( | | x k + 1 − x * | | ε k ) = O ( | | x k − x * | | ε k ) + o ( ε k + 1 ) ,
Therefore, we can obtain a conclusion analogous to that in the proof of quadratic convergence:
∇ x x 2 L ( x * , μ , Y ) ( x k + 1 − x * ) + D h ( x * ) T ( μ k + 1 − μ ) + D G ( x * ) * ( Y k + 1 − Y ) = O ( | | x k − x * | | ε k ) + o ( ε k + 1 ) .
Written in matrix form, this is:
∇ x x 2 L ( x * , μ , Y ) D h ( x * ) T D G ( x * ) * D h ( x * ) 0 0 ( J * − I ) D G ( x * ) 0 J * x k + 1 − x * μ k + 1 − μ Y k + 1 − Y = O ( | | x k − x * | | ε k ) + o ( ε k + 1 ) .
Similarly, it can be proven that the 3 × 3 matrix operator is also nonsingular, leading to ε k + 1 = O ( | | x k − x * | | ε k ) , that is:
| | x k + 1 − x * | | | | x k − x * | | = O ( ε k + 1 ) | | x k − x * | | = O ( ε k ) → 0 .
therefore, it is proven that the sequence x k converges to x * with superlinear convergence rate. □

4.3. Numerical Experiments

This paper designs two numerical experiments to test the convergence of the algorithm. The programs are written in MATLAB R2018a. In the following experiments, the initial parameter is M 0 = I n , the matrix M k is updated by the modified BFGS method [18], and the algorithm stops under the following conditions:
(1)The algorithm is considered successful if | | d k | | ≤ 1 e − 8 ;
(2)The algorithm reaches the maximum number of iterations, if k ≥ n m a x ;
(3)The update of the matrix M k fails if the condition number of M k is c o n d ( M k ) = λ 1 ( M k ) λ n ( M k ) > 1 e + 6 .
We select the following two examples, use the algorithm to solve them, and summarize the results in Tables 1 and 2. For these two examples, the algorithm successfully converges to a KKT point after multiple iterations. The c o n d ( M k ) fluctuates during the iterations but always remains within a bounded range without triggering the condition number threshold termination condition, which indicates that the BFGS update maintains good properties and also reflects the superlinear convergence of the algorithm.
Example 1.
f ( x ) = 0.1 ( x 1 4 + x 2 4 + x 3 4 ) − 0.5 ( x 1 2 + x 2 2 + x 3 2 ) + 0.2 x 1 x 2 x 3 + 0.1 ( x 1 2 x 2 2 + x 2 2 x 3 2 ) + 1 ,
h ( x ) = x 1 2 + x 2 2 + x 3 2 = 1 ,
G ( x ) = x 1 2 + x 2 x 3 + x 1 j x 3 − x 1 j x 2 + x 3 ⪯ 0 .
Example 2.
f ( x ) = 0.5 x 1 4 + 0.3 x 2 4 + 0.4 x 3 4 + sin ( 2 x 1 − x 2 ) + e ( 0.5 x 2 − 0.3 x 3 ) + 2 x 1 x 2 − 3 x 2 x 3 + 5 ,
h ( x ) = 6 x 1 2 − x 2 2 − x 3 2 = 0 ,
G ( x ) = x 1 2 + x 2 x 3 + x 1 j x 3 − x 1 j x 2 + x 3 ⪯ 0 .
Table 1. Algorithm 4.1.1: Results of solving Example 1.
Table 1. Algorithm 4.1.1: Results of solving Example 1.
Nitk x k , 1 x k , 2 x k , 3 f ( x k ) ∥ d k ∥ v ( x k ) cond( M k )
0 5.0000e-01 -5.0000e-01 5.0000e-01 6.3125e-01 4.5169e-01 5.9307e-01 1.0000e+00
1 4.5218e-01 -8.0228e-01 4.2054e-01 5.3028e-01 8.9517e-02 1.3711e-01 1.2584e+00
2 4.5162e-01 -8.0279e-01 4.1983e-01 5.3052e-01 8.8264e-02 1.3510e-01 2.6322e+00
3 4.5107e-01 -8.0329e-01 4.1913e-01 5.3075e-01 8.7030e-02 1.3312e-01 1.0000e+00
4 4.5052e-01 -8.0378e-01 4.1844e-01 5.3098e-01 8.5816e-02 1.3117e-01 3.4210e+00
5 4.5000e-01 -8.0426e-01 4.1775e-01 5.3121e-01 8.4590e-02 1.2925e-01 1.0000e+00
6 4.4947e-01 -8.0473e-01 4.1708e-01 5.3144e-01 8.3431e-02 1.2735e-01 3.3880e+00
7 4.4896e-01 -8.0519e-01 4.1640e-01 5.3166e-01 8.2216e-02 1.2548e-01 1.0000e+00
8 4.4845e-01 -8.0565e-01 4.1575e-01 5.3188e-01 8.1090e-02 1.2364e-01 3.3577e+00
9 4.4795e-01 -8.0610e-01 4.1509e-01 5.3210e-01 7.9907e-02 1.2182e-01 1.0000e+00
10 4.4745e-01 -8.0655e-01 4.1445e-01 5.3231e-01 7.8813e-02 1.2003e-01 3.3299e+00
… … … … … … … …
420 4.1718e-01 -8.3225e-01 3.6530e-01 4.8656e-01 6.1597e-04 1.1880e-04 1.0000e+00
421 4.1719e-01 -8.3225e-01 3.6529e-01 4.8657e-01 1.3296e-04 1.1657e-04 2.8482e+03
422 4.1718e-01 -8.3225e-01 3.6529e-01 4.8658e-01 1.0091e-04 1.1373e-04 1.0000e+00
423 4.1718e-01 -8.3225e-01 3.6529e-01 4.8659e-01 1.3787e-04 1.1103e-04 1.0000e+00
424 4.1705e-01 -8.3225e-01 3.6528e-01 4.8713e-01 0.0000e+00 0.0000e+00 2.3831e+00
Table 2. Algorithm 4.1.1: Results of solving Example 2.
Table 2. Algorithm 4.1.1: Results of solving Example 2.
Nitk x k , 1 x k , 2 x k , 3 f ( x k ) ∥ d k ∥ v ( x k ) cond( M k )
0 1.0000e+00 1.0000e+00 1.0000e+00 7.2629e+00 0.0000e+00 3.4142e+00 1.0000e+00
1 6.3889e-01 2.0991e-02 8.1236e-01 6.9763e+00 2.1204e+00 1.6843e+00 5.5320e+01
2 4.2502e-01 -4.7029e-02 3.5537e-01 6.6923e+00 1.0182e+00 7.8184e-01 5.5320e+01
3 3.0390e-01 -9.9391e-02 1.5138e-01 6.5482e+00 4.8590e-01 3.6328e-01 5.5320e+01
4 2.2313e-01 -1.1252e-01 3.0859e-02 6.4282e+00 2.9135e-01 1.5326e-01 5.5320e+01
5 1.6829e-01 -1.1352e-01 -4.2036e-02 6.3398e+00 1.8245e-01 5.6623e-02 5.5320e+01
6 1.3041e-01 -1.0653e-01 -7.4386e-02 6.2773e+00 1.0061e-01 2.1712e-02 5.5320e+01
7 1.0066e-01 -8.9174e-02 -7.2513e-02 6.2266e+00 6.8990e-02 1.0396e-02 5.5320e+01
8 7.5120e-02 -6.6959e-02 -5.1168e-02 6.1772e+00 8.0035e-02 5.5055e-03 5.5320e+01
9 5.5397e-02 -4.9532e-02 -3.0961e-02 6.1342e+00 6.6364e-02 2.2247e-03 5.5320e+01
10 4.1051e-02 -3.7156e-02 -1.7878e-02 6.1008e+00 4.6049e-02 5.7825e-04 5.5320e+01
… … … … … … … …
49 -4.8924e-02 -1.0658e-01 5.4791e-02 5.9694e+00 4.0046e-04 2.0062e-08 1.0489e+01
50 -4.9032e-02 -1.0682e-01 5.4911e-02 5.9694e+00 2.8628e-04 1.6196e-08 1.0489e+01
51 -4.8955e-02 -1.0665e-01 5.4825e-02 5.9694e+00 2.0462e-04 1.4405e-08 1.0489e+01
52 -4.8999e-02 -1.0674e-01 5.4874e-02 5.9694e+00 1.4634e-04 1.3415e-08 1.0489e+01
53 -4.8978e-02 -1.0670e-01 5.4852e-02 5.9694e+00 5.4589e-05 1.2502e-08 1.0489e+01

5. Conclusions

This paper addresses the Quaternion Negative Semidefinite Matrix Cone Constrained Optimization Problems (QNSDP), constructs its theoretical framework, and designs an efficient numerical algorithm.
First, this paper defines the projection operator on the quaternion self-conjugate matrix space and rigorously proves that this projection operator is strongly semismooth. On this basis, it gives definitions of the Robinson constraint qualification, the constraint nondegeneracy condition, and the strict complementarity condition for QNSDP, and proves that the Lagrange multiplier set is nonempty and bounded if and only if the Robinson constraint qualification holds. Furthermore, it gives an explicit expression of the critical cone and establishes the first-order necessary optimality condition (KKT condition) and the second-order sufficient optimality condition.
Second, this paper proposes an SQP-type method for solving QNSDP. The algorithm obtains the search direction by solving a quadratic programming subproblem, updates the Hessian approximation matrix by the modified BFGS formula, and analyzes the convergence of the algorithm under the assumptions that different constraint qualification conditions hold. Finally, numerical experiments are provided, which are effective and stable.

Author Contributions

Yuzhuo Zhang conducted the main research, developed the theoretical derivations, performed the numerical experiments, and prepared the first draft of the manuscript. Yun Wang contributed to the theoretical derivations, revised the second version of the manuscript, and assisted in summarizing and refining the overall presentation of the work. Both authors read and approved the final manuscript.

Funding

This work is supported by the Natural Science Foundation of Shandong Province of China (No.ZR2023MA011).

Acknowledgments

The author(s) declare(s) they have used Artificial Intelligence (AI) tools in the creation of this article. AI tools used: ChatGPT. How were the AI tools used: ChatGPT was used for language polishing and translation assistance. It did not contribute to the mathematical results, theoretical derivations, or numerical experiments. Where in the article is the information located: The AI-assisted text is distributed throughout the manuscript, mainly in the Introduction, Theoretical Framework, and Conclusion sections.

Conflicts of Interest

The authors declare no conflicts of interest.

Abbreviations

The following abbreviations are used in this manuscript:
QNSDP quaternion self-conjugate negative semidefinite matrix constraints
KKT Karush-Kuhn-Tucker conditions
SQP Sequential Quadratic Programming

References

  1. Wu, T. T.; Huang, C. Y.; Ng, M. K. Total variation based pure quaternion dictionary learning method for color image denoising. Int. J. Numer. Anal. Model. 2022, 19, 709–737. [Google Scholar] [CrossRef]
  2. Sun, Y. F.; Chen, S. Y.; Yin, B. C. Color face recognition based on quaternion matrix representation. Pattern Recognit. Lett. 2011, 32, 597–605. [Google Scholar] [CrossRef]
  3. Wu, P. L.; Kou, K. I.; Yu, Z. Y. Efficient quaternion CUR method for low-rank approximation to quaternion matrix. Numer. Algorithms 2025, 99, 1531–1549. [Google Scholar] [CrossRef]
  4. Jia, Z. G.; Jin, Q. Y.; Zhao, X. L. Non-local robust quaternion matrix completion for large-scale color image and video inpainting. IEEE Trans. Image Process. 2022, 31, 3868–3883. [Google Scholar] [CrossRef] [PubMed]
  5. Miao, J. F.; Kou, K. I. Color image recovery using low-rank quaternion matrix completion algorithm. IEEE Trans. Image Process. 2022, 31, 190–201. [Google Scholar] [CrossRef] [PubMed]
  6. Wang, Y.; Yang, Y. N. Optimization on the quaternion Stiefel manifold with applications-part II: L1-QPCA. J. Optim. Theory Appl. 2026, 208, 76. [Google Scholar] [CrossRef]
  7. Ciardo, L.; Zivny, S. Semidefinite programming and linear equations vs. homomorphism problems. SIAM J. Comput. 2025, 54, 545–584. [Google Scholar] [CrossRef]
  8. Chen, Z. W.; Miao, S. C. A penalty-free method with trust region for nonlinear semidefinite programming. Asia-Pac. J. Oper. Res. 2015, 32, 1540006. [Google Scholar] [CrossRef]
  9. Zhao, Q.; Chen, Z. W. On the superlinear local convergence of a penalty-free method for nonlinear semidefinite programming. J. Comput. Appl. Math. 2016, 308, 1–19. [Google Scholar] [CrossRef]
  10. Zhao, Q.; Chen, Z. W. A line search exact penalty method for nonlinear semidefinite programming. Comput. Optim. Appl. 2020, 75, 467–491. [Google Scholar] [CrossRef]
  11. Ling, S. T.; Cheng, X. H.; Jiang, T. S. An algorithm for coneigenvalues and coneigenvectors of quaternion matrices. Adv. Appl. Clifford Algebr. 2015, 25, 377–384. [Google Scholar] [CrossRef]
  12. Wang, T.; Li, Y.; Zhang, M. C. The generalized eigenvalue decomposition of a dual quaternion regular matrix pencil. Appl. Math. Lett. 2026, 174, 109821. [Google Scholar] [CrossRef]
  13. Bonnans, J. F.; Shapiro, A. Perturbation Analysis of Optimization Problems; Springer: New York, 2000. [Google Scholar]
  14. Nocedal, J.; Wright, S. J. Numerical Optimization, 2nd edn; Springer: Berlin, 2006. [Google Scholar]
  15. Sun, D. F.; Sun, J. Semismooth matrix-valued functions. Math. Oper. Res. 2002, 27, 150–169. [Google Scholar] [CrossRef]
  16. Freund, R. W.; Jarre, F.; Vogelbusch, C. H. Nonlinear semidefinite programming: sensitivity, convergence, and an application in passive reduced-order modeling. Math. Program. 2007, 109, 581–611. [Google Scholar] [CrossRef]
  17. Sun, D. F. The strong second-order sufficient condition and constraint nondegeneracy in nonlinear semidefinite programming and their implications. Math. Oper. Res. 2006, 31, 761–776. [Google Scholar] [CrossRef]
  18. Yamashita, H.; Yabe, H.; Harada, K. A primal-dual interior point method for nonlinear semidefinite programming. Math. Program. 2012, 135, 89–121. [Google Scholar] [CrossRef]
  19. Zhao, Q.; Chen, Z. W. An SQP-type method with superlinear convergence for nonlinear semidefinite programming. Asia Pac. J. Oper. Res. 2018, 35, 1850009. [Google Scholar] [CrossRef]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.