Numerical Algorithms

https://doi.org/10.1007/s11075-024-01870-4

ORIGINAL PAPER

A least-change secant algorithm for solving generalized

complementarity problem

H. Vivas1 · R. Pérez1 · C. Arias1

Received: 19 February 2024 / Accepted: 21 June 2024

© The Author(s) 2024

Abstract
In this paper, we propose a least-change secant algorithm to solve the generalized
complementarity problem indirectly trough its reformulation as a nonsmooth system
of nonlinear equations using a one-parametric family of complementarity functions.
We present local and superlinear convergence results of new algorithm and analyze
its numerical performance.

Keywords Generalized complementarity · Least-change secant methods · Nonlinear

complementarity · Generalized jacobian · q-superlinear convergence.

Mathematics Subject Classification (2010) 78M99 · 90C33 · 49M15 · 90C30 · 65H10

1 Introduction
The generalized complementarity problem (GCP(F, G)) consists of finding a vector
x ∈ Rn , such that

F(x) ≥ 0, G(x) ≥ 0, F(x)T G(x) = 0, (1)

where F : Rn → Rn , (F1 (x), . . . , Fn (x)) and G : Rn → Rn , (G 1 (x) . . . , G n (x)),

are continuously differentiable functions.
The linear [1–4], nonlinear [5] and implicit [5, 6] complementarity problems are
particular cases of GCP(F, G).
R. Pérez and C. Arias These authors are contributed equally to this work

B H. Vivas
[email protected]
R. Pérez
[email protected]
C. Arias
[email protected]

1 Departamento de Matemáticas, Universidad del Cauca, Calle 5 No 4-70, Popayán, Cauca, Colombia

123
 Numerical Algorithms

To solve the GCP(F, G), we consider the one-parametric family of complementar-

ity functions ϕλ introduced in [7], defined by

ϕλ (a, b) = (a − b)2 + λab − a − b, λ ∈ (0, 4) , (2)

and we reformulate the problem as the following nonsmooth system of equations,

λ (x) = (ϕλ (F1 (x), G 1 (x)), . . . , ϕλ (Fn (x), G n (x))T = 0. (3)

From definition of complementarity function [7], we have that ϕλ satisfies that

ϕλ (a, b) = 0 ⇔ a ≥ 0, b ≥ 0, ab = 0, for all λ ∈ (0, 4) . This equivalence
guarantees that to solve the GCP(F, G) is equivalent to solve (3).
One motivation for using the family ϕλ in the reformulation of GCP(F, G) is that
it includes among its members the Fischer function (λ = 2) and multiples of the Min-
imum function (λ → 0+ ), which are the most popular complementarity functions.
Thus, an algorithm using ϕλ will allow not only to take advantage of these two com-
plementarity functions, but also to analyze the impact of varying the complementarity
function (varying λ).
Nonsmooth Newton-type methods, which use the concept of generalized Jacobian
[8], are popular for solving the nonlinear complementarity problem [5, 7, 9, 10]. Some
of them have been extended to generalized complementarity problem [11, 12].
In the last years, quasi-Newton methods such as the least change secant [1, 2] have
been proposed to solve the particular case of the nonlinear complementarity problem
[3, 4]. In [4], the reformulation is based on the Fischer and Minimum functions while
in [3], it is based on a family of complementarity functions ϕλ . It is important to
mention that, as far as we know, this type of methods has not been used in the general
case of the generalized complementarity problem.
In this work, we propose a least-change secant algorithm to solve the generalized
complementarity problem by its reformulation as a nonsmooth system of equations
using the one-parametric family of complementarity functions ϕλ . For this algorithm,
we prove local and superlinear convergence under suitable assumptions and present
numerical test that show its good local performance.
This document is organized as follows. In Section 2, we recall some definitions
and results which we use in the document. In Section 3 we present the algorithm
and convergence results. In Section 4 we prove, under suitable assumptions, local
and superlinear convergence. In Section 5 we present some additional convergence
results. In Section 6 we analyze the numerical performance of the algorithms. Finally,
In Section 7, we present some conclusions.

2 Preliminaries

In this section we present some basic concepts and results necessary for the develop-
ment of this work.

123
Numerical Algorithms

Definition 1 [4]. The generalized B-Jacobian of a function L : Rn → Rn locally

Lipschitz is defined, for all x ∈ Rn , by
 
∂ B L(x) = lim L  (x k ) ∈ Rn×n : lim x k = x, x k ∈ D L ,
k→∞ k→∞

with D L , the set where L is differentiable. The convex hull of ∂ B L(x) is called the
generalized Jacobian of L at x, denoted ∂ L(x).
To define the concepts of BD-regularity and R-regularity for generalized com-
plementarity, we consider a solution x ∗ of GCP(F, G) and the index sets, α =
{i ∈ I : Fi (x ∗ ) > 0 = G i (x ∗ )} and β = {i ∈ I : Fi (x ∗ ) = 0 = G i (x ∗ )} .
Definition 2 [13]. Let x ∗ be a solution of GCP(F, G).
1. If all matrices H ∈ ∂ B (x ∗ ) are nonsingular, x ∗ is called a BD-Regular solution.
2. Let F  (x ∗ ) be nonsingular and K = G  (x ∗ )F  (x ∗ )−1 . If the submatrix1 K αα is
nonsingular and the Schur’s complement of K ,
−1
K ββ − K βα K αα K αβ ∈ R|β|×|β|

is a P-matrix2 , x ∗ is called an R-regular solution.

The following two propositions characterize the matrices of ∂λ (x).
Proposition 1 [14] For all x ∈ Rn ,

∂λ (x) ⊆ Da (x)F(x) + Db (x)G  (x),

where Da (x) = diag (a1 (x) − 1, . . . , an (x) − 1) and Db (x) = diag(b1 (x) − 1,
. . . , bn (x) − 1) are diagonal matrices with ai (x) and bi (x) given by

2 (Fi (x) − G i (x)) + λG i (x)

ai (x) =  (4)
2 (Fi (x) − G i (x))2 + λFi (x)G i (x)

−2 (Fi (x) − G i (x)) + λFi (x)

bi (x) =  , (5)
2 (Fi (x) − G i (x))2 + λFi (x)G i (x)
if (Fi (x), G i (x)) = (0, 0),
√ and by ai (x) = ξi , bi (x) = χi , for all (ξi , χi ) ∈
R2 such that (ξi , χi ) ≤ 2, if (Fi (x), G i (x)) = (0, 0).
Proposition 2 [14] Any H ∈ ∂λ (x) can be written in the form
 
H = Da (x)F (x) + Db (x)G (x), (6)

where Da (x) y Db (x) are negative semidefinite diagonal matrices such that Da (x) +
Db (x) is negative definite.
1 Let A = (a ) ∈ Rm×n be. The matrix A
ij αβ is one whose elements ai j such that i ∈ α and j ∈ β.
2 A matrix M ∈ Rn×n is a P-matrix if for every nonzero vector z there is an index j ∈ {1, . . . , n} such
that z j [M z] j > 0.

123
 Numerical Algorithms

The following results are useful to guarantee the nonsingularity of the elements of
the generalized Jacobian of λ in a solution x ∗ of the GCP(F, G).
Lemma 1 [5] If F is strongly BD-regular on x then there exists a neighborhood N of
x and a constant c such that for any y ∈ N and V ∈ ∂ B F( y), V is nonsingular and

V −1 ≤ c. (7)

Theorem 1 [14] If x ∗ is an R-regular solution of the GCP(F, G) then all matrices in

the generalized Jacobian ∂λ (x ∗ ) are nonsingular.

The next lemma gives an upper bound for the partial derivatives of (4) and (5),
when (Fi (x), G i (x)) = (0, 0). This bound which will be used in later results.
Lemma 2 [7]. Let f λ : R2 → R be defined by

f λ (a, b) = (a − b)2 + λab, λ ∈ (0, 4). (8)
√
then ∇ f λ (a, b) ≤ 2, for all nonzero vector (a, b) ∈ R2 .
The following result will be useful to prove that the new algorithmic proposal is well
defined.

Lemma 3 [15] Let · be an induced matrix norm, A and C ∈ Rn×n . If C is nonsin-

gular and In − C −1 A < 1 then A is nonsingular and furthermore,

C −1
A−1 ≤ ·
1 − In − C −1 A

3 Algorithm and convergence results

In this section, we propose a least-change secant algorithm to solve the nonsmooth

system of equations (3) and thus, to solve indirectly the GCP(F, G). We also present
local convergence results for this algorithm.
We start with a generic quasi-Newton algorithm where the matrix Bk is an approx-
imation to Hk ∈ ∂λ (x k ), which does not specify how to update Bk , for this reason
the algorithm is called generic. In the second part of this section, we will deal with
the updating of these matrices.
From Proposition 2, the matrices H ∈ ∂λ (x) are of the form

H = Da (x)F  (x) + Db (x)G  (x).

Thus, following the philosophy of quasi-Newton methods, we propose to approxi-

mate the matrices Hk ∈ ∂λ (x k ) by matrices Bk given by

Bk = Da (x k )Ak + Db (x k )Ck , (9)

123
Numerical Algorithms

where Ak and Ck are approximations of the Jacobian matrices F  (xk ) and G  (x k ),

respectively. Equivalently, ⎛ ⎞
[Bk ]1
⎜ ⎟
Bk = ⎝ ... ⎠ , (10)
[Bk ]n
with 
(ai (x) − 1) [ Ak ]i + (bi (x) − 1) [Ck ]i , i ∈
/ β̄
[Bk ]i = (11)
(χi − 1) [Ak ]i + (ξi − 1) [Ck ]i , i ∈ β̄,
with ⎛ ⎞ ⎛ ⎞
[Ak ]1 [Ck ]1
⎜ ⎟ ⎜ ⎟
Ak = ⎝ ... ⎠ and Ck = ⎝ ... ⎠ ,
[Ak ]n [Ck ]n
where the terms ai (x), bi (x), χi and ξi are defined in Proposition 1 and β̄ =
{i ∈ I : Fi (x) = 0 = G i (x)} .
The generic quasi-Newton algorithm that we propose is the following.

Algorithm 1 Generic quasi-Newton algorithm.

Require: x 0 ∈ Rn and λ ∈ (0, 4), A0 , C0 ∈ Rn×n nonsingular, > 0 and k = 0.
Ensure: Approximation to the solution of (3).
1: while λ (x k ) > do
2: Let Bk be an approximation to Hk given by (10)-(11). Find d k such that

Bk d k = −λ (x k ). (12)
3: Do
x k+1 = x k + d k ,
4: Update Bk , such that Bk+1 is nonsingular. Go to Step 1.
5: end while
6: return x ∗

Remark 1 If λ = 2 or λ → 0+ and G is the identity function, the Algorithm 1 reduces

to those presented in [4] for the nonlinear complementarity problem.

Remark 2 If G is the identity function, the Algorithm 1 reduces to those presented in

[3] for the nonlinear complementarity problem.

Remark 3 when λ → 0+ , the Algorithm 1 reduces to that presented in [16], for the
generalized complementarity problem.

4 Assumptions and convergence results

Under the following Assumptions, we prove that the sequence {x k } generated by the
Algorithm 1 is well defined and converges linearly to a solution of (3).

123
 Numerical Algorithms

A1. There exists x ∗ ∈ Rn such that λ (x ∗ ) = 0.

A2. The Jacobian matrices of F and G are Lipschitz continuous (with the same
constant) in a neighborhood of x ∗ ∈ Rn . That is, there exists a positive constant γ ,
such that

F  (x) − F  (x ∗ ) ≤ γ x − x ∗ and G  (x) − G  (x ∗ ) ≤ γ x − x ∗ ,

for all x ∈ B(x ∗ ; δ), where · denotes any norm in Rn and also its respective induced
matrix norm.
A3. x ∗ is an R-regular solution of GCP(F, G).
From Theorem 1, Lemma 1 and the Assumption A3, we can deduce that the gener-
alized Jacobian matrices are nonsingular and their inverses are bounded. That is, for
H∗ ∈ ∂λ (x ∗ ), there exists a positive constant μ, such that H∗−1 ≤ μ.
Let H ∈ ∂λ (x) and H∗ ∈ ∂λ (x ∗ ) be defined by (6) and the matrix B define by
(9). The following two lemmas guarantee that, for every x close enough to a solution
of GCP(F, G), H − B ∞ and H − H∗ ∞ are bounded from above.

Lemma 4 Let F, G : Rn −→ Rn be continuously differentiable functions whose Jaco-

bian matrices satisfy the Assumption A2; , δ F , and δG , positive constants; H and B
defined by (6) and (10), respectively. Then for every x ∈ B(x ∗ ; ), A ∈ B(F  (x ∗ ); δ F )
and C ∈ B(G  (x ∗ ); δG ), there exists a positive constant θ, such that

H−B ∞ ≤ θ.

Proof Using the infinity norm, we have that

H−B ∞ = max { [H ]i − [B]i 1 }.

1≤i≤n

From (6) and (10),


(ai − 1)(∇ Fi (x)T − [A]i ) + (bi − 1)(∇G i (x)T − [C]i ), i ∈ / β̄
[H − B]i =
(χi − 1) (∇ Fi (x)T − [A]i ) + (ξi − 1) (∇G i (x)T − [C]i ), i ∈ β̄.

Suppose that the maximum is reached in the row j. We consider two cases.
If j ∈
/ β̄, then

H−B ∞ = [H ] j − [B] j 1
= (a j − 1)(∇ F j (x)T − [A] j ) + (b j − 1)(∇G j (x)T − [C] j ) 1
≤ |a j − 1| ∇ F j (x)T − [A] j 1 + |b j − 1| ∇G j (x)T − [C] j ) 1
≤ (|a j | + 1) ∇ F j (x)T − [A] j 1 + (|b j | + 1) ∇G j (x)T − [C] j ) 1
.
√ √
From Lemma 2 and Proposition 1, |a j | < 2 and |b j | < 2, so
√ 
H−B ∞ ≤ (1 + 2) ∇ F j (x)T − [A] j 1
+ ∇G j (x)T − [C] j ) 1
. (13)

123
Numerical Algorithms

On the other hand, using equivalence between the norms · ∞ and · 1 , the Assump-
tion A2 and the fact that A ∈ B(F  (x ∗ ) ; δ F ), we get
 
∇ F j (x) − [A] j 1
≤ n  ∇ F j (x) − ∇ F j (x ∗ ) ∞ + ∇ F j (x ∗) − [A] j ∞
≤ n γ x − x ∗ ∞ + ∇ F j (x ∗ ) − [A] j ∞
≤ n (γ x − x ∗ ∞ + δF ) ≤ n ( γ + δF ) ,

whereby ∇ F j (x) − [A] j 1 ≤ n( γ + δ F ).

Analogously, since C ∈ (BF  (x ∗ ) ; δG ), we obtain the other inequality ∇G j (x)
− [C] j 1 ≤ n( γ + δG ).
Therefore, from (13) we have that
√
H−B ∞ ≤ 2(1 + 2)n ( γ + δ) = θ1 , (14)

where δ = max {δ F , δG } .
Now, if j ∈ β, we have that

H−B ∞ = [H ] j − [B] j 1
≤ |χ j − 1| ∇ F j (x)T − [A] j 1 + |ξ j − 1| ∇G j (x)T − [C] j ) 1
≤ (|χ j | + 1) ∇ F j (x)T − [A] j 1 + (|ξ j | + 1) ∇G j (x)T − [C] j ) 1
.

From Proposition √ √ between the norms · 2 and · 1 , we have that

√ 1 and the equivalence
(χ , ξ ) 1 ≤ 2 (χ , ξ ) 2 ≤ 2 2 = 2, where |χ j | < 2 and |ξ j | < 2. Proceeding
as in the previous case, we have that

H−B ∞ ≤ 6n ( γ + δ) = θ2 , (15)

where δ = max {δ F , δG } .
Finally, from (14) and (15), since θ1 < θ2 and defining θ = θ2 = 6n ( γ + δ) , we
conclude that
H − B ∞ ≤ θ. (16)



Lemma 5 Assume that the Assumptions A1 and A2 are verified. Then for each
H ∈ ∂λ (x) and H∗ ∈ ∂λ (x ∗ ), there exists a positive constant α such that if
x − x ∗ ∞ < then H − H∗ ∞ < α.

Proof Suppose that x ∈ Rn , x ∗ ∈ Rn satisfies the Assumption A1 and consider the

matrix H − H∗ . Adding and subtracting Da F  (x ∗ ) + Db G  (x ∗ ), we get
   
H − H∗ = (Da − I )F (x) + (Db − I )G (x) − (Dâ − I )F (x ∗ ) − (Db̂ − I )G (x ∗ )
      
= −[F (x) − F (x ∗ ) + G (x) − G (x ∗ )] + Da [F (x) − F (x ∗ )] + Da F (x ∗ )
    
+Db [G (x) − G (x ∗ )] + Db G (x ∗ ) − Dâ F (x ∗ ) − Db̂ G (x ∗ ).

123
 Numerical Algorithms

Since F  and G  satisfy the Assumption A2, from Lemma 2 and Proposition 1, we
have that
√ √
H − H∗ ∞ ≤ 2γ x − x ∗ ∞ + 2 2γ x − x ∗ ∞ + 4 2M
√ √
≤ (2 + 2 2)γ + 4 2M = θ̂ ,
 
where M = max F  (x ∗ ) ∞ , G  (x ∗ ) ; that is, there exists the positive constant
√ √ ∞
θ̂ := (2 + 2 2)γ + 4 2M, such that

H − H∗ ∞ < θ̂ , (17)

which proves the lemma. 



Observe that θ is function of and δ, and θ̂ is function of ; that is, θ = θ ( , δ)

and θ̂ = θ̂( ).
The following lemma gives sufficient conditions for the linear convergence of
Algorithm 1. It will be useful in the proof of Theorem 2.

Lemma 6 Let r ∈ (0, 1). For each x ∈ Rn and A, C ∈ Rn×n , we define

Q(x, A, C) = x − B −1 λ (x), (18)

where the matrix B is given by (10). If x ∗ satisfies the Assumption A3 then there exist
constants , δ > 0, such that

if x − x∗ ∞ ≤ , C − G  (x ∗ ) ∞
≤ δ and A − F  (x ∗ ) ∞
≤ δ,

the function Q is well defined and

Q(x, A, C) − x ∗ ∞ ≤ r x − x∗ ∞. (19)

r r
Proof Let r ∈ (0, 1), 1 < and δ < , where γ is the Lipschitz constant
96nμγ 96nμ
in Assumption A2, and μ is the constant of Lemma 1.
To prove that Q is well defined, we must guarantee that B −1 exists. For this, let us
consider the inequality

B − H∗ ∞ ≤ B−H ∞ + H − H∗ ∞. (20)

If x − x ∗ ∞ < 1 , C − G  (x ∗ ) ∞ ≤ δ and A − F  (x ∗ ) ∞ ≤ δ, since the first

term on the right side of (20) was bounded in (16) by θ ( 1 , δ), we have
 
rγ r r
B−H ∞ ≤ θ ( 1 , δ) = 6n ( 1 γ + δ) < 6n + = ·
96nμγ 96nμ 8μ

123
Numerical Algorithms

Now, the second term of (20) was bounded by θ̂ ( 2 ) in (17). Taking

r √
√ − 4 2M
8(2 + 2 2)μ
2 < ,
γ
√ √ r
we get H − H∗ ∞ ≤ θ̂ ( 2 ) = (2 + 2 2)γ 2+ 4 2M < ·
8μ
Consequently, if we define = min { 1 , 2 } , we have that

r r
B−H ∞ < and H − H∗ ∞ < · (21)
8μ 8μ

Substituting (21) in (20), we obtain

r r r
B − H∗ ∞ ≤ B−H ∞ + H − H∗ ∞ < + = ·
8μ 8μ 4μ

Therefore, from (7) we have

r 1
B − H∗ ∞ ≤ < ·
4μ 4 H∗−1
∞

Then,
1
H∗−1 B − In ≤ H∗−1 B − H∗ ∞ < ,
∞ ∞ 4
therefore, H∗−1 B − In ∞
< 1 and, by Lemma 3, the function Q is well defined.
Moreover,
H −1 μ 4
B −1 ≤ < = μ.
1 − In − H −1 B 1 3
1−
4
Now, to prove (19), we subtract x∗ in (18) and after some algebraic manipulations,
we have

Q(x, A, C) − x ∗ ∞ = x − x ∗ − B −1 λ (x) ∞
= (x − x ∗ ) −B −1 λ ( x ) + B −1 H∗ (x − x ∗ ) −B −1 H∗ (x − x ∗ ) ∞
≤ B −1 ∞ (B − H∗ )(x − x ∗ ) − [λ (x) − λ (x ∗ ) + H∗ (x − x ∗ )] ∞
4 
≤ μ B − H ∗ ∞ x − x ∗ ∞ + λ ( x ) − λ ( x ∗ ) + H ∗ (x − x ∗ ) ∞ ·
3

We get the inequality,

 
4 r λ (x) − λ (x ∗ ) + H∗ (x − x ∗ ) ∞
Q(x, A, C) − x ∗ ∞ ≤ μ + x − x∗ ∞ . (22)
3 4μ x − x∗ ∞

123
 Numerical Algorithms

From the semismoothness of λ [7], we have

λ (x) − λ (x ∗ ) + H (x − x ∗ ) ∞
lim = 0,
x→x ∗ x − x∗ ∞

thus, for all ρ > 0, there exists 3 > 0, such that if x − x ∗ ∞ < 3 then

λ (x) − λ (x ∗ ) + H (x − x ∗ ) ∞
< ρ.
x − x∗ ∞
r
In particular, for ρ = , we have that
8μ

λ (x) − λ (x ∗ ) + H (x − x ∗ ) ∞ r
< · (23)
x − x∗ ∞ 8μ

From (21), (22) and (23), we obtain

 
4 r r r
Q(x, A) − x ∗ ∞ < μ + x − x∗ ∞ = x − x∗ ∞ < r x − x∗ ∞ ,
3 4μ 8μ 2

which complete the proof of Lemma 6.



The following result, analogous to the Theorem of two neighborhoods [15], guar-
antees q-linear convergence of Algorithm 1 in the infinity norm. Here, we have three
neighborhoods: one of the solution, another one of the Jacobian matrix of F at the
solution and the third one of the Jacobian matrix of G at the solution.
Theorem 2 Given r ∈ (0, 1), there exist > 0 and δ > 0 such that, if

x0 − x∗ ∞ ≤ , Ak − F  (x ∗ ) ∞
≤ δ and Ck − G  (x ∗ ) ∞
≤δ

then the sequence {x k } generated by

x k+1 = Q(x k , Ak , Ck ) = x k − Bk−1 λ (x k ), (24)

where Bk , defined by (10), is well defined, converges to x ∗ and satisfies

x k+1 − x ∗ ∞ ≤ r xk − x∗ ∞ ,k = 0, 1, . . .

Proof Consider the function Q defined by (18). Thus, x k+1 is given by (24). Let
r ∈ (0, 1), 1 ∈ (0, ) and δ1 ∈ (0, δ), where and δ are the constants of Lemma 6.
We will use mathematical induction on k.
• For k = 0. If x 0 − x ∗ ∞ ≤ 1 ≤ , since

A0 − F  (x ∗ ) ∞
≤ δ1 ≤ δ and C0 − G  (x ∗ ) ∞
≤ δ1 ≤ δ,

123
Numerical Algorithms

by the Lemma 6, Q(x 0 , A0 , C0 ) = x 1 = x 0 − B0−1 λ (x 0 ) is well defined and

x1 − x∗ ∞ ≤ r x0 − x∗ ∞.

• Inductive hypothesis. Suppose that the Theorem 2 is true for k = 0, . . . , m − 1.

That is, given r ∈ (0, 1) there exist > 0 and δ > 0 such that, if x 0 − x ∗ ∞ ≤
, Ak − F  (x ∗ ) ∞ ≤ δ and Ck − G  (x ∗ ) ∞ ≤ δ then

x k+1 = Q(x k , Ak , Ck ) = x k − Bk−1 λ (x k ),

is well defined and satisfies x k+1 − x ∗ ∞ ≤ r xk − x∗ ∞. Then,

xm − x∗ ∞ ≤ r x m−1 − x ∗ ∞· (25)

From (25), the inductive hypothesis and since r < 1, we have

xm − x∗ ∞ ≤ r x m−1 − x ∗ ∞ ≤ r m x0 − x∗ ∞ ≤ rm ≤ .

Thus, by the Lemma 6,

x m+1 = Q(x m , Am , Cm ) = x m − Bm−1 λ (x m )

is well defined and satisfies

x m+1 − x ∗ ∞ ≤ r xm − x∗ ∞

which concludes the proof of Theorem 2.




The following theorem gives a sufficient condition for q-superlinear convergence

of the Algorithm 1. The proof is an extension to generalized complementarity of the
one done in [3] for the particular case of nonlinear complementarity. In the proof, we
use infinity norm even though, we recall that the q-superlinear convergence results are
independent of the norm.

Theorem 3 Let Bk be defined by (10). Assume the Assumptions A1 to A3 and that the
sequence {x k } generated by

x k+1 = x k − Bk−1 λ (x k ) (26)

satisfies that limk→∞ x k = x ∗ . If

(Bk − H∗ ) d k
lim = 0, (27)
k→∞ dk

where d k = x k+1 − x k then {x k } converges q-superlinearly to x ∗ .

123
 Numerical Algorithms

Proof Since λ is semismooth and λ (x ∗ ) = 0, we have

λ (x k ) − H∗ (x k − x ∗ )
lim ∗ = 0. (28)
x k →x xk − x∗

On the other hand,

 
x k − x ∗ ≤ H∗−1 H∗ x k − x ∗ , (29)

furthermore,
     
 λ (x k ) − H∗ x k − x ∗  ≤ λ (x k ) − H∗ x k − x ∗ . (30)

From (28) to (30), it follows that

| λ (x k ) − H∗ (x k − x ∗ ) |
lim ∗ = 0,
x k →x
H∗−1 H∗ (x k − x ∗ )

then
| λ (x k ) − H∗ (x k − x ∗ ) |
lim ∗ = 0.
x k →x H∗ (x k − x ∗ )
1
This implies that for ρ = , there exists > 0, such that if x k − x ∗ < then
2

1 λ (x k ) − H∗ (x k − x ∗ ) 1
− < < ,
2 H∗ (x k − x ∗ ) 2

thus,
1   3  
H∗ x k − x ∗ < λ (x k ) < H∗ x k − x ∗ .
2 2
From the left inequality and by (29), we get

1   1
λ (x k ) > H∗ x k − x ∗ ≥ xk − x∗ . (31)
2 2 H∗−1

On the other hand, from (26), we have the equation

0 = Bk d k + λ (x k ) , (32)

where d k = x k+1 − x k . Adding and subtracting H∗ d k − λ (x k+1 ) in (32) and

rearranging terms, we obtain

−λ (x k+1 ) = Bk d k − H∗ d k + λ (x k ) + H∗ d k − λ (x k+1 ) .

123
Numerical Algorithms

Applying norm to both sides and the triangular inequality, we deduce that

λ (x k+1 ) = Bk d k − H∗ d k + λ (x k ) + H∗ d k − λ (x k+1 )
≤ (Bk − H∗ ) d k + λ (x k ) + H∗ d k − λ (x k+1 )

then

λ (x k+1 ) (Bk − H∗ ) d k λ (x k ) + H∗ x k − λ (x k+1 )

≤ + ·
dk dk dk

The first term of the right side converges to zero by the hypothesis (27). The second
one converges to zero by semismoothness of λ . So,

λ (x k+1 )
lim = 0. (33)
x k →x ∗ dk

Now, from (33), (31) and (7), we have that

λ (x k+1 ) 1 x k+1 − x ∗
0 = lim ∗ ≥ lim ∗
x k →x xk 2 H∗−1 x k →x sk
1 x k+1 − x ∗
= lim ∗ ∗ ∗
2 H∗−1 x k →x x k+1 − x + x − x k
1 x k+1 − x ∗
≥ lim ∗
2 μ x k →x x k+1 − x ∗ + x k − x ∗
x k+1 − x ∗
1 xk − x∗
= lim ∗ ·
2 μ x k →x x k+1 − x ∗
+1
xk − x∗

Hence,
x k+1 − x ∗
lim ∗ = 0,
x k →x xk − x∗
whereby the sequence {x k } converges q-superlinearly to x ∗ .




5 Least-change secant update algorithm

It is clear that Algorithm 1 is generic because it does specify how to update the matrices
Bk . In this section, we propose least-change secant updates for those matrices following
the theory developed by [1], which was used in [4] for nonlinear complementarity. In

123
 Numerical Algorithms

this sense, we will see that the proves are practically analogous to those presented in
that document.
Observe that to update Bk in the Algorithm 1 implies to update the matrices Ak and
Ck , respectively. Following [1], the updates Ak+1 and Ck+1 must satisfy the secant
equations:

Ak+1 (x k+1 − x k ) = F(x k+1 ) − F(x k ) (34)

Ck+1 (x k+1 − x k ) = G(x k+1 ) − G(x k ) (35)

and the change between Ak and Ak+1 , Ck and Ck+1 , must be minimum; i.e.,

Ak+1 − Ak = min A − Ak , Ck+1 − Ck = min C − Ck ,

A∈VF C∈VG

where the sets VF = VF (x, y) and VG = VG (x, y) contain all matrices that satisfy
the secant equations (34) and (35), respectively. Thus, it is natural to think in the
orthogonal projection of these matrices onto VF and VG . That is Ak+1 = PVF (Ak )
and Ck+1 = PVG (Ck ). Some examples de least-secant updates are the follows,

• Broyden’s “good” update [17].

( yk − Ak sk )skT ( ŷk − Ck sk )skT

Ak+1 = Ak + Ck+1 = Ck +
skT sk skT sk

• Broyden’s “bad” update [18].

( yk − Ak sk ) ykT Ak ( ŷk − Ck sk ) ŷkT Ck

Ak+1 = Ak + Ck+1 = Ck + ·
ykT Ak sk ŷkT Ck sk

• Schubert’s update [19].

F(x k+1 )skT G(x k+1 )skT

Ak+1 = Ak + Ck+1 = Ck + ·
skT sk skT sk

Next, we present a least-change secant update algorithm to solve the GCP(F, G).
In order to develop the convergence theory of Algorithm 2, we assume an additional
Assumption to the previous three.
A4. Suppose that for all x, z ∈ Rn there exist matrices A ∈ VF (x, z) and C ∈
VG (x, z) such that

A − F  (x ∗ ) ≤ α1 σ (x, z) and C − G  (x ∗ ) ≤ α2 σ (x, z),

where σ (x, z) = max{ x − x ∗ , z − x ∗ }.

123
Numerical Algorithms

Algorithm 2 Least-change secant algorithm.

Require: x 0 ∈ Rn , λ ∈ (0, 4), A0 and C0 ∈ Rn×n , > 0.
Ensure: Approximation to the , solution of (3).
1: while λ (x k ) > do
2: Let Bk be an approximation to Hk given by (10)-(11). Find d k such that

Bk d k = −λ (x k ). (36)
3: Do
x k+1 = x k + d k ,
4: Update Ak and Ck such that
Ak+1 = PVF (x k ,x k+1 ) (Ak )
Ck+1 = PVG (x k ,x k+1 ) (Ck ).

Go back to Step 1.
5: end while
6: return x ∗

6 Additional convergence results

Lemma 7 Assume that the Assumptions A1 to A4 are verified. Let x, y ∈ Rn be close

enough to x ∗ , A, C ∈ Rn×n . Then there exists a constant γ̄ > 0, such that

PVF (x, y) (A) − F  (x ∗ ) ≤ A − F  (x ∗ ) + γ̄ x − x ∗ (37)

and
PVG (x, y) (C) − G  (x ∗ ) ≤ C − G  (x ∗ ) + γ̄ x − x ∗ ,
for all x, y ∈ B(x ∗ ; ), A ∈ B(F  (x ∗ ); δ), C ∈ B(G  (x ∗ ); δ) and y − x ∗ ≤
x − x∗ .

Proof It is analogous to the proof of Lemma 3.1 in [2].



In the following lemma, we show that the updates to matrices Ak+1 and Ck+1 , can
deteriorate but in a controlled way.
Lemma 8 Suppose that the Assumptions A1 to A4 are verified. Let A+ be the orthog-
onal projection of A onto the set VF (x, z), C+ the orthogonal projection of C onto
 the orthogonal projection of F  (x) onto VF (x, z) and C
VG (x, z), A  the orthogonal

projection of G (x) onto VG (x, z) then

A+ − F  (x) ≤ A − F  (x) + β1 σ (x, z)

and
C+ − G  (x) ≤ C − G  (x) + β2 σ (x, z), (38)
where α1 > 0, α2 > 0 and σ (x, z) = max{ x − x ∗ , z − x ∗ }.

123
 Numerical Algorithms

Proof By the triangle inequality, we have

A+ − F  (x) ≤ A+ − A  − F  (x) .
 + A (39)

Since the matrix A+ satisfies that

 =
A+ − A min  ,
A− A
A∈VF (x,z)

we have the inequality:

A+ − A  ≤ A − F  (x) + A
 ≤ A− A  − F  (x) . (40)

Replacing (40) in (39),

 − F  (x) .
A+ − F  (x) ≤ A − F  (x) + 2 A

 − F  (x) ≤ α1 σ (x, z) then

By the Assumption A4, A

A+ − F  (x) ≤ A − F  (x) + β1 σ (x, z),

where β1 = 2α1 . Analogously, taking β2 = 2α2 , proves (38).



The following result guarantees that the Algorithm 2 is well defined and converges.
Its proof is analogous to that of a similar result for nonlinear complementarity demon-
strated in [4].

Theorem 4 Suppose the Assumptions A1 to A4. Let {Ak } and {Ck } be the sequences
generated by the Algorithm 2. Given r ∈ (0, 1), there exist positive constants and δ
such that, if

x0 − x∗ ≤ , A0 − F  (x ∗ ) ≤ δ and C0 − G  (x ∗ ) ≤ δ,

the succession {x k } generated by

x k+1 = x k − Bk−1 (x k ),

is well defined, it converges to x ∗ and furthermore,

x k+1 − x ∗ ≤ r x k − x ∗ .

Proof Let and δ be the constants of the Lemma 6, δ ∈ (0, δ) and

< min { , δ} and < min { , δr } , (41)

123
Numerical Algorithms

where
δ−δ 1−r
δ= and δr = (δ − δ) ,
γ γ
with γ the constant of (37).
From (41), we can deduce that

γ (δ − δ)
δ+γ 0 < δ + γ δ0 = δ + =δ
γ

and
γ δr γ γ (δ − δ)(1 − r )
δ+ <δ+ <δ+ = δ· (42)
1−r 1−r γ (1 − r )
We use induction on k.

1. For k = 0. If

x 0 − x ∗ ≤ , A0 − F  (x ∗ ) ≤ δ and C0 − G  (x ∗ ) ≤ δ,

then by Lemma 6, x 1 is well defined and satisfies that

x1 − x∗ ≤ r x0 − x∗ ,

then, the Theorem 4 is true for k = 0.

2. Inductive hypothesis. Suppose that the Theorem 4 is true for k = 0, 1, . . . , m − 1
then, if x m−1 − x ∗ ≤ ,

−1
x m = x m−1 − Bm−1 (x m−1 ),

is well defined and x m − x ∗ ≤ r x m−1 − x ∗ .

3. By Inductive hypothesis and given that 0 < r < 1, we have that

x m − x ∗ ≤ r x m−1 − x ∗ ≤ r m x 0 − x ∗ < r m < .

From (37) and 0 < r < 1,

Am − F  (x ∗ ) ≤ Am−1 − F  (x ∗ ) + γ x m−1 − x ∗

m−1
≤ δ+γ rj
j=0
∞

≤ δ+γ rj <δ+γ ≤ δ,
1−r
j=0

where the last inequality is obtained from (42). So,

Am − F  (x ∗ ) < δ.

123
 Numerical Algorithms

In an analogous way, it is shown that Cm − G  (x ∗ ) < δ. Thus, by Lemma 6,

x m is well defined and satisfies

x m+1 − x ∗ ≤ r x m − x ∗ ,

then the Theorem 4 is true for k = m.




Lemma 9 Under the Assumptions of Theorem 4,

lim Ak+1 − Ak = 0 and lim Ck+1 − Ck = 0. (43)

k→∞ k→∞

Proof The proof in a similar to the one in Theorem 3.3 of [1].



Lemma 10 Under the hypotheses of Theorem 4,

lim Bk+1 − Bk = 0.
k→∞

Proof From (10) and (11), we have that if [Bk ]i = Bk+1 i − [Bk ]i then
⎧  
⎪
⎪ (a i − 1) Ak+1 i + (bi − 1) Ck+1 
⎨ i
− (ai − 1) [Ak ]i + (b
 i − 1)[C k ] ,
 i ∈
/β
[Bk ]i = i
(44)
⎪ (χi − 1) Ak+1 i + ξi − 1 Ck+1
⎪  i
⎩
− (χi − 1) [Ak ]i + (ξi − 1) [Ck ]i , i ∈ β.

Now, proceeding analogously to the proof of Lemma 4, we have


Bk+1 − Bk ∞ = max { Bk+1 i − [Bk ]i 1
},
1≤i≤n

whereby
√ √
Bk+1 − Bk ∞ ≤ (1 + 2) Ak+1 − Ak ∞ + (1 + 2) Ck+1 − Ck ∞,

thus, using (43) and equivalence between norms that

lim Bk+1 − Bk = 0,
k→∞

which completes the proof.



The following theorem gives a sufficient condition for q-superlinear convergence
of Algorithm 2.

123
Numerical Algorithms

Theorem 5 Let us suppose the Assumptions A1 to A4, {Ak } and {Ck } are the sequences
defined by the Algorithm 2 and lim x k = x ∗ . If
k→∞

(Bk+1 − H∗ ) d k
lim =0 (45)
k→∞ dk

then the sequence {x k } converges q-superlinearly.

Proof The proof is a straightforward application of the Theorems 3 and 4. In fact,

(Bk − H∗ )) d k (Bk − Bk+1 ) d k (Bk+1 − H∗ )) d k

lim ≤ lim + lim
k→∞ dk k→∞ dk k→∞ dk
(Bk+1 − H∗ ) d k
≤ lim Bk − Bk+1 + lim ·
k→∞ k→∞ dk

By Lemma 10 and (45),

(Bk − H∗ ) d k
lim = 0.
k→∞ dk
Therefore, we conclude that the sequence {x k } converges q-superlinearly.


In the following theorem we summarize the results obtained in this section.
Theorem 6 Suppose the Assumptions A1 to A4. Let {Ak } , {Ck } and {x k } be the
sequences defined by the Algorithm 2. Given r ∈ (0, 1) there are positive constants
and δ such that, if

x0 − x∗ ≤ , Ak − F  (x ∗ ) ≤ δ and Ck − G  (x ∗ ) ≤ δ,

then the sequence {x k } converges local and q-linearly to x ∗ . If

(Bk+1 − H∗ ) sk
lim =0
k→∞ sk

then {x k } converges q-superlinearly to x ∗ .

Proof It is a direct application of Theorems 3, 4 and 5.



7 Numerical experimentation

In this section, we analyze the numerical performance of Algorithm 2. We update the

matrices Ak and Ck using Broyden’s good, Broyden’s bad and Schubert’s updates,
respectively. Thus, we have three versions of the algorithm.
For the experiments, we consider ten problems. The first seven problems with their
initial points and solution(s) are those used in [14]. We describe below the functions

123
 Numerical Algorithms

that define the remaining three problems, with their starting points and solution(s).
For simplicity, we denote en as the vector of ones of order n.
Problem8 [6]. The function F is the same as in Problem
 4 in [14] and G is defined
by G(x) = 2.5 − x12 , 2.5 − x22 , 2.5 − x32 , 2.5 − x42 .

Starting points: x 1 = (0, 0, 0, 0), x 2 = −5e4 , x 3 = 2e4 and x 4 = 4e4 . Solution:

x ∗ = (−0.9, −1.2, −1.2, −0.9) .
Problem 9 [6] Let F, G : R2 → R2 defined by
   
F(x) = e x1 − x2 , x1 + x2 − 1 and G(x) = x12 − 2x1 + x2 − 2, x2 − e−x1 .

Starting points: x 1 = (2, 3) , x 2 = (3, 1) , x 3 = (3.5, 2) and x 4 = (4, 0.5) . Solution:

x ∗ = (2.7128, 0.0664) .

Problem 10. Let F, G : R3 → R3 defined by

⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞
x y2 + 1 x z
F ⎝ y ⎠ = ⎝ z − x 2 − y 2 + 1 ⎠ and G ⎝ y ⎠ = ⎝ z + x 2 + y 2 − 1 ⎠ .
z z z x2 + 1

Starting points: x 1 = (0, 0, 0), x 2 = −0.5e3 , x 3 = −e3 and x 4 = −2e3 . Solutions

set: {(x, y, z) ∈ R3 : x 2 + y 2 = 1, z = 0}; that is, the circle intersection of the
paraboloids z = −1 + x 2 + y 2 and z = 1 − x 2 − y 2 .
We implement the algorithms and test functions in Matlab R2020a, on a computer
with a 2.21 GHz AMD Sempron (tm) processor. The parameters used were = 10−4
and N = 100, the latter indicates the maximum number of iterations allowed in the
algorithms. To study the effect of the parameter λ in the algorithms, we use three ways
of choose it: λ = 10−5 , λ = 2 and the dynamic strategy proposed in [7].
In the first experiment, we estimate the convergence radius of the variants of the
Algorithm 2 for three choices of the parameter λ. We will call these variants: good
Broyden (gB), bad Broyden (bB) and Schubert (Sch) (for λ = 10−5 ); Fischer-good
Broyden (FgB), Fischer-bad Broyden (FbB) and Fischer-Schubert (FSch) (for λ = 2);
Dynamic-good Broyden (DgB), Dynamic-bad Broyden (DbB) and Dynamic-Schubert
(DSch) (when λ is chosen using the dynamic strategy proposed in [7]).
The radius of convergence is the maximum radius (r ) of a neighborhood centered
on a solution x ∗ (Br (x ∗ )), such that if we take an initial point in this neighborhood,
the algorithm converges. For the infinity norm, this region is the set
 
Br (x ∗ ) = x ∈ Rn : x − x ∗ ∞ ≤r .

We consider an initial radius r0 , such that for 1000 random points in Br0 (x ∗ ), the
respective algorithm converges. Then, we increase the radius by 0.1 units and get a
radius r1 . Again, we generate 1000 random points in Br1 (x ∗ ). If there is no convergence
for any of those points then the radius of convergence will be r0 ; otherwise, we continue
the process until the radius is estimated.

123
Numerical Algorithms

The results of this experiment are shown in Table 1 for the three ways in which the
parameter λ was chosen. The table contains the following information: Problem (P),
convergence radius (r ) and average number of iterations (k) for which each variant of
the algorithm converged.
In general, we observe that in 50% of the cases (Problems P2, P4, P6, P7 and P9)
there are no significant variations in the radius of convergence for the choices of the
parameter λ. For Problems P3, P5 and P10 the radius of convergence of the algorithm
increases significantly for λ = 2 with the good Broyden and bad Broyden updates.
Also, for Problems P1 and P8 the radius of convergence increases when λ = 10−5
with the good Broyden update. This suggests that the size of the convergence radius
depends not only on the characteristics of each problem, but also on the value of λ.
It is important to mention that we carried out the experiments for others values of
λ (namely, λ = 3.9, 3.7, 3.5, 3.1, 2.7, 2.52, 1.5, 1, 0.5, 0.1, 10−3 , 10−5 ), the results
were similar to those shown in Table 1 without significant changes in the convergence
radii of the algorithm. For this reason we do not report it in the paper.
In the second experiment, in order to illustrate numerically the rate of convergence
of the Algorithm 2, we choose the Problem 2, the starting point x 1 and the nine
versions of the algorithm. We calculate the quotient Rk , defined by

x k+1 − x ∗
Rk = ·
xk − x∗
Table 1 Results of Experiment 1. Convergence radius varying λ

P gB bB Sch FgB FbB FSch DgB DbB Dsch

P1 r 0.8 0.2 0.6 0.2 0.2 0.2 0.6 0.5 0.8

k 10.2 18.1 14.2 12.8 5.1 7.4 13.5 9.6 8.2
P2 r 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 0.6
k 3.0 2.9 5.3 14.2 5.9 8.5 13.2 7.0 2.0
P3 r 1.1 0.9 0.9 10.0 10.0 1.0 1.2 1.3 0.6
k 9.1 7.9 10.3 5.1 5.1 7.4 14.2 11.5 7.1
P4 r 10.0 10.0 0.4 10.0 10.0 10.0 0.9 0.9 10.0
k 17.3 17.2 64.6 5.9 5.9 8.5 12.3 11.9 17
P5 r 1.7 1.4 1.6 10.0 5.4 0.8 1.9 1.5 1.7
k 6.5 5.8 13.1 27.2 11.7 12.6 9.1 7.5 6.1
P6 r 0.4 0.4 0.2 0.4 0.5 0.3 0.4 0.4 0.2
k 10.3 8.9 13.1 7.1 10.3 10.0 12.7 7.6 8.9
P7 r 0.4 0.4 0.2 0.5 0.5 0.2 0.5 0.4 0.2
k 10.7 9.0 13.9 10.2 10.5 8.8 10.4 7.6 6.8
P8 r 10.0 5.4 0.8 0.9 0.9 0.7 13 5.4 10.0
k 26.7 11.6 12.3 16.8 16.4 20.2 13.3 12.8 9.9
P9 r 10.0 10.0 10.0 10.0 10.0 1.5 10.0 10.0 1.9
k 2.9 2.8 6.7 7.4 6.5 12.0 9.4 6.6 2.0
P10 r 0.4 0.4 0.2 2.0 1.5 0.1 0.5 0.4 0.2
k 10.7 9.0 13.9 5.3 5.2 0 10.4 7.6 6.8

123
 Numerical Algorithms

Table 2 Rate of convergence of Algorithm 2 for Problem 2

Rk Rk Rk

k gB bB Sch FgB FbB FSch DgB DbB Dsch

l 0.472 0.472 0.444 0.472 0.472 0.444 0.472 0.472 −
2 0.163 0.163 0.342 0.163 0.163 0.342 0.163 0.163 −
3 0.149 0.149 0.328 0.149 0.149 0.328 0.149 0.149 −
4 0.028 0.028 0.119 0.028 0.028 0.119 0.257 0.257 −
5 0.001 0.001 0.167 0.001 0.001 0.167 0.546 0.546 −
6 0.000 0.000 0.133 0.133 0.001 0.001 −
7 0.103 0.098 0.000 0.000 −
8 0.083 0.015 0.000 0.000 −
. .
.. ..

13 0.008
14 0.000

The results are shown in the Table 2, which contains the iteration k and the value
of Rk , for each version of the algorithm. The symbol “−” indicates divergence.
We observe that in all cases where there is convergence, the quotient Rk goes to
zero, when k increases. This indicates at least superlinear convergence of the pro-
posed algorithm, which agrees with the theoretical results in previous section. This
experiment also suggests that the convergence rate is not affected by the choice of the
parameter λ.
In the third experiment, we illustrate the region of convergence of three versions of
Algorithm 2, for
 the problem P5, and the six versions for the
 problem P10. We associate
the set X i∗ = y = xi,∗ + 0.1t : t = −10, −9, . . . , 9, 10 to each of the components
of the solution x ∗ and we run the algorithms with all points of the Cartesian product
n ∗
i=1 X i as starting points.
The Figures 1, 2 and 3 present in blue color the points for which the algorithms
converged and in magenta those that diverged.

Fig. 1 Region and percent of convergence of Algorithm 2 for problem P5, with λ dynamic choice

123
Numerical Algorithms

Fig. 2 Region and percent convergence of Algorithm 2 for problem P10, with λ = 10−5

In these figures, we can see that for problem P5, the good Broyden update with the
dynamic choice of λ improves the region of convergence.
For Problem 10, the dynamic λ turns out to be the best option since for λ = 10−5 ,
the Schubert’s update converges only for 38.4% of the cases, compared to the other
updates. In general, we see the need to implement a globalization strategy in each of
these methods.
In the four experiment, we analyze the efficiency and robustness of the algorithms.
That is, its ability of solving a problem from different starting points. For this, we use
the efficiency (E), robustness (R) and combined robustness and efficiency (C) indices
[20–22]. An algorithm is more robust, efficient or balanced, if the indices are near to
1.
The Table 3 shows that the Algorithm 2 is more robust with the bad Broyden update
for the different λ choices; for λ = 10−5 and λ = 2 the algorithm was more efficient
when Bk was updated using the bad Broyden formula, meanwhile for λ dynamic, the
Algorithm was more efficient with the Schubert update. Finally, The most balanced
versions of the Algorithm 2 were obtained for λ = 10−5 and λ dynamic strategy,
respectively, with the Schubert update.

Fig. 3 Region and percent convergence of Algorithm 2 for problem P10, with λ dynamic

123
 Numerical Algorithms

Table 3 Robustness (R), efficiency (E) and combined (C) indices

gB bB Sch FgB FbB FSch DgB DbB DSch

R 0.92 0.97 0.79 0.87 0.87 0.79 0.84 0.97 0.89
E 0.64 0.67 0.44 0.58 0.62 0.35 0.51 0.44 0.73
C 0.59 0.65 0.34 0.50 0.54 0.28 0.43 0.43 0.65

8 Final remarks

In this work we propose a nonsmooth least-change secant algorithm to solve the gen-
eralized complementarity problem through its reformulation as a system of nonlinear
equations using a one-parametric family of complementarity functions. This algorithm
can be seen as an extension to generalized complementarity of the least-change secant
family of methods proposed in [4] for nonlinear complementarity. We prove, under
suitable assumptions, local and q-superlinear convergence.
The numerical tests show a good performance of the algorithmic proposal and allow
us to observe that most of the convergence radii of the algorithms are relatively small.
The numerical tests indicates that the choice of th parameter λ does not affect the
convergence rate of the algorithm, but it does affect its radius of convergence, robust-
ness and efficiency. In addition, it was observed that a small or tending to zero value of λ
can improve the performance of the algorithm. This is in agreement with observations
of other researchers who, from their numerical experiments, have concluded that for
local algorithms, the Minimum function performs better than other complementarity
functions.
On the other hand, since the methods are local, the choice of the starting point influ-
ences the region of convergence as evidenced in the first experiment. An alternative
may be to incorporate a globalization strategy and do the numerical analysis of the
globalized algorithm.
Acknowledgements We are grateful to the University of Cauca for the time allowed for this research.

Author Contributions All authors contributed equally to this work.

Funding Open Access funding provided by Colombia Consortium. This work was not funded by any entity.

Availability of Data and Material The data obtained in this research are available.

Code Availability This work introduce an algorithm completely available for the readers.

Declarations

Conflicts of Interest The authors declare that they have no conflict of interest.

Competing Interests The authors have no financial or proprietary interests in any material discussed in this
article.

Ethics Approval The authors declare that they followed all the rules of a good scientific practice.

123
Numerical Algorithms

Consent to Participate All authors approve their participation in this work.

Consent for Publication The authors approve the publication of this research.

Human and Animal Ethics Not applicable.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give
appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence,
and indicate if changes were made. The images or other third party material in this article are included
in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If
material is not included in the article’s Creative Commons licence and your intended use is not permitted
by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the
copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

References
1. Martínez, J.M.: On the relation between two local convergence theories of least-change secant update
methods. Math. Comput. 59(200), 457–481 (1992)
2. Martínez, J.M.: Local convergence theory of inexact newton methods based on structured least change
updates. Math. Comput. 55(191), 143–167 (1990)
3. Arenas, F., Martínez, H.J., Pérez, R.: Least change secant update methods for nonlinear complemen-
tarity problem. Ingeniería y Ciencia. 11(21), 11–36 (2015). https://doi.org/10.17230/ingciencia.11.21.
1
4. Lopes, V.L.R., Martínez, J.M., Pérez, R.: On the local convergence of quasi-Newton methods for
nonlinear complementarity problems. J. Comput. Appl. Math. 30(1), 3–22 (1999). https://doi.org/10.
1016/S0168-9274(98)00080-4
5. Qi, L.: Convergence analysis of some algorithms for solving nonsmooth equations. Math. Oper. Res.
18(1), 227–244 (1993)
6. Outrata, J.V., Zowe, J.: A newton method for a class of quasi-variational inequalities. Comput. Optim.
Appl. 4, 5–21 (1995). https://doi.org/10.1007/BF01299156
7. Kanzow, C., Kleinmichel, H.: A new class of semi-smooth Newton-type methods for nonlinear com-
plementarity problems. Comput. Optim. Appl. 11(3), 227–251 (1998)
8. Clarke, F.H.: Necessary conditions for nonsmooth problems in optimal control and the calculus of
variations. PhD thesis, University of Washington (1973)
9. Jiang, H., Fukushima, M., Qi, L., Sun, D.: A trust region method for solving generalized complemen-
tarity problems. SIAM J. Optim. 8(1), 140–157 (1998). https://doi.org/10.1137/S1052623495296541
10. De Luca, T., Facchinei, F., Kanzow, C.: A semi-smooth equation approach to the solution of nonlinear
complementarity problems. Math. Program. 75, 407–439 (1996)
11. Song, L., Gao, Y.: On the local convergence of a levenberg-marquardt method for nonsmooth nonlinear
complementarity problems. Sci. Asia. 43, 377 (2017). https://doi.org/10.2306/scienceasia1513-1874.
2017.43.377
12. Du, S.-q.: Generalized Newton method for a kind of complementarity problem. In: Abstr. Appl. Anal.,
vol. 2014 (2014). https://doi.org/10.1155/2014/745981
13. Arias, C.A., Martínez, H.J., Pérez, R.: A global quasi-Newton algorithms for nonlinear complementarity
problems. Pacific J. Optim. 13(1), 1–15 (2017)
14. Vivas, H., Pérez, R., Arias, C.A.: A nonsmooth newton method for solving the generalized comple-
mentarity problem. Numer. Algor (2023). https://doi.org/10.1007/s11075-023-01581-2
15. Dennis, J.E., Jr., Schnabel, R.B.: Numerical Methods for Unconstrained Optimization and Nonlinear
Equations. SIAM, New Jersey (1996)
16. Grueso, G.Z.T.: Un algoritmo local secante de cambio mínimo para complementariedad generalizada.
Master’s thesis, Universidad del Cauca (2021)
17. Broyden, C.G.: A class of methods for solving nonlinear simultaneous equations. Math. Comput.
19(92), 577–593 (1965)
18. Brown, P.N., Saad, Y.: Convergence theory of nonlinear newton-krylov algorithms. SIAM J. Optim.
4(2), 297–330 (1994). https://doi.org/10.1137/0804017

123
 Numerical Algorithms

19. Schubert, L.K.: Modification of a quasi-newton method for nonlinear equations with a sparse jacobian.
Math. Comput. 24(109), 27–30 (1970)
20. Bogle, I.D.L., Perkins, J.D.: A new sparsity preserving quasi-newton update for solving nonlinear
equations. SIAM J. Sci. Comput. 11(4), 621–630 (1990)
21. Ralevic, N.M., Cebic, D.: The properties of modifications of newton’s method for solving nonlinear
equations. In: 2013 IEEE 11th International Symposium on Intelligent Systems and Informatics (SISY),
pp. 341–345 (2013). https://doi.org/10.1109/SISY.2013.6662599
22. Buhmiler, S., Krejic, N.: A new smoothing quasi-newton method for nonlinear complementarity prob-
lems. J. Comput. Appl. Math. 211(2), 141–155 (2008)

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.

123