J Optim Theory Appl

DOI 10.1007/s10957-010-9671-8

Constant-Rank Condition and Second-Order

Constraint Qualification

R. Andreani · C.E. Echagüe · M.L. Schuverdt

© Springer Science+Business Media, LLC 2010

Abstract The constant-rank condition for feasible points of nonlinear programming

problems was defined by Janin (Math. Program. Study 21:127–138, 1984). In that
paper, the author proved that the constant-rank condition is a first-order constraint
qualification. In this work, we prove that the constant-rank condition is also a second-
order constraint qualification. We define other second-order constraint qualifications.

Keywords Nonlinear programming · Constraint qualifications

1 Introduction

We are concerned with the general nonlinear programming problems with equality
and inequality constraints in the form

min f (x), s.t. h(x) = 0, g(x) ≤ 0, (1)

Communicated by F. Giannessi.
This research was supported by PRONEX-Optimization 76.79.1008-00, FAPESP (Grant
01-04597-4) and CNPq.
R. Andreani
Department of Applied Mathematics, IMECC-UNICAMP, University of Campinas, CP 6065,
13081-970 Campinas, SP, Brazil
e-mail: [email protected]

C.E. Echagüe
Department of Mathematics, FCE, University of La Plata, CP 172, 1900 La Plata, Bs. As., Argentina
e-mail: [email protected]

M.L. Schuverdt ()

CONICET, Department of Mathematics, FCE, University of La Plata, CP 172, 1900 La Plata,
Bs. As., Argentina
e-mail: [email protected]
 J Optim Theory Appl

where f : Rn → R, h : Rn → Rm , g : Rn → Rp are twice continuously differentiable

on Rn . For a feasible point x, we define the set of active inequality constraints:

A(x) := {i ∈ {1, . . . , p} : gi (x) = 0}.

In constrained optimization, one is generally interested in finding global minimiz-

ers, but we realize that this is very difficult. This is the main reason why we study
optimality conditions and constraint qualifications in constrained optimization. The
idea behind this is to find good necessary optimality conditions for a minimum point
of the problem (1).
The desirable and most important first-order condition that combines the objective
function with the constraints is the well-known Karush/Kuhn-Tucker condition (KKT
condition, [2, 3]); given a feasible point x̂ of the problem (1), there are vectors λ ∈
Rm , μ ∈ Rp such that


m 
p
∇f (x̂) + λi ∇hi (x̂) + μi ∇gi (x̂) = 0,
i=1 i=1

μi ≥ 0, μi gi (x̂) = 0, i = 1, . . . , p. (2)

The vectors λ ∈ Rm , μ ∈ Rp are known as Karush/Kuhn-Tucker (KKT) multipliers

and we say that x̂ is a stationary point of the problem (1). See for example [4, 5].
A point that verifies (2) is a stationary point of the Lagrangian function associated to
the problem (1),


m 
p
l(x, λ, μ) = f (x) + λi hi (x) + μi gi (x).
i=1 i=1

Unfortunately, condition (2) is not a first-order necessary optimality condition for

a minimum point. For example, the solution of the problem of finding a minimizer
of the function f (x) = x, subject to h(x) = x 2 , does not verify the KKT condition.
First-order constraint qualifications are conditions over the constraints under which
it can be claimed that, if x is a feasible minimum point, then x is a stationary point
of the Lagrangian function associated to each objective function which it minimizes.
The most widely used first-order constraint qualification is the linear indepen-
dence of the gradients of the equality and inequality active constraints at a given
feasible point (LICQ). It is well known that LICQ is a first-order constraint quali-
fication and it implies the existence and uniqueness of KKT multipliers for a given
solution. There are weaker first-order constraint qualifications in the literature. The
Mangasarian-Fromovitz condition (MFCQ), defined in [6], establishes the positive
linear independence of the gradients of the equality and inequality active constraints
at a given feasible point and it is weaker than LICQ. Another first-order constraint
qualification weaker than LICQ is the constant-rank constraint qualification (CRCQ)
defined in [1]. We say that, a feasible point verifies the constant-rank constraint qual-
ification if there exists a neighbourhood of the feasible point in which the rank of
any subset of the gradients of the equality and inequality active constraints does
J Optim Theory Appl

not change. There are simple examples showing that CRCQ neither implies nor is
implied by MFCQ; see [1]. More recently, the constant positive linear dependence
(CPLD) condition was defined in [7] and it was established as a first-order constraint
qualification in [8]. CPLD is weaker than CRCQ and MFCQ, see [8]. We can find
other first-order constraint qualifications in the literature. Some of them are, in order
of weakness: quasinormality defined by Hestenes in [9], pseudonormality [10, 11],
Abadie [12] and Guignard [13]. Guignard condition is the weakest first-order con-
straint qualification for differentiable problems as shown in [14]. When a first-order
constraint qualification holds, it is possible to think in terms of KKT multipliers and
efficient algorithms based on duality ideas can be defined. The discovery of new
and weaker first-order constraint qualifications and necessary optimality conditions
is an open issue in nonlinear optimization. Recently [15, 16], a necessary optimality
condition and a constraint qualification were defined by means of the image space
analysis [17].
Second-order necessary optimality conditions are important because they take into
account the curvature of the Lagrangian function over the set of feasible directions.
The desirable second-order condition that combines the objective function with the
constraints is the KKT condition plus the strong second-order necessary condition
(SSONC): given a feasible point x̂ of the problem (1), there are vectors λ ∈ Rm , μ ∈
Rp such that condition (2) holds at x̂ and SSONC:
 
m 
p 
d ∇ f (x̂) +
T 2
λi ∇ hi (x̂) +
2
μi ∇ gi (x̂) d ≥ 0,
2
(3)
i=1 i=1

for all direction d ∈ Rn in the following tangent subspace:

T̃ (x̂) := {d ∈ Rn : ∇hi (x̂)T d = 0, i = 1, . . . , m,

∇gj (x̂)T d = 0, j ∈ A+ (x̂),
∇gj (x̂)T d ≤ 0, j ∈ A0 (x̂)},

where
A+ (x̂) = {j ∈ A(x̂) : μj > 0}, A0 (x̂) = {j ∈ A(x̂) : μj = 0}.
Condition (3) says that the Hessian of the Lagrangian function at (x̂, λ, μ), restricted
to the tangent subspace T̃ (x̂), is positive semidefinite.
Unfortunately, the combination KKT + SSONC is not always a second-order nec-
essary optimality condition. It is true that, under an appropriate constraint qualifica-
tion, it is possible to show that a minimum point of the problem (1) verifies KKT +
SSONC.
Observe that, if a feasible point x̂ is a KKT point, then

T̃ (x̂) = {d ∈ Rn : ∇f (x̂)T d ≤ 0; ∇hi (x̂)T d = 0, i = 1, . . . , m;

∇gj (x̂)T d ≤ 0, j ∈ A(x̂)}.

Thus, the tangent subspace in SSONC depends on the objective function of the prob-
lem.
 J Optim Theory Appl

Some of the second-order practical algorithms (see for example [18]) take into ac-
count the analysis of the Hessian of the Lagrangian function in the following tangent
subspace:

T (x) := {d ∈ Rn : ∇hi (x)T d = 0, i = 1, . . . , m, ∇gj (x)T d = 0, j ∈ A(x)}.

Clearly, for a given feasible point x, T (x) ⊆ T̃ (x) and T (x) does not depend on the
objective function. By considering T (x), we can define another second-order condi-
tion that combines the objective function and the constraints—KKT condition plus
the weak second-order necessary condition (WSONC)—given a feasible point x̂ of
the problem (1), there are vectors λ ∈ Rm , μ ∈ Rp such that condition (2) holds at x̂
and WSONC:
 
m 
p 
d T ∇ 2 f (x̂) + λi ∇ 2 hi (x̂) + μi ∇ 2 gi (x̂) d ≥ 0,
i=1 i=1

for all d ∈ T (x̂).

The tangent T (x) is associated in practice with some of the well known nonlinear
programming algorithms; see [19] and references therein. In [19], the authors have
shown a class of nonlinear optimization algorithms, using barrier functions, for which
the strong second-order necessary condition may not hold at the limit points, even
if the sequence of the subproblems minimizers satisfies the second-order sufficient
condition. This implies the practical importance of the weak second-order condition.
Observe that, if the strict complementarity condition holds at a stationary point x̂
(this means that μi − gi (x̂) > 0, i = 1, . . . , p), then T̃ (x̂) = T (x̂) and SSONC is
equivalent to WSONC. The strict complementarity condition is used for the conver-
gence analysis of many nonlinear programming algorithms.
Thus, we can say that, from a theoretical point of view, it is important to analyze
conditions that imply SSONC; also, from a practical point of view, it is important to
analyze conditions that imply WSONC.
It is well established in the literature that, if a minimum point of the problem (1)
verifies the linear independence of the equality and inequality active constraints
gradients, then there is a unique KKT multiplier vector for which SSONC holds
(see [20]). The important question is: is it possible to relax the linear independence
constraint qualification and still have the possibility that a minimum point verifies
KKT + SSONC (or KKT + WSONC)? All the first-order constraint qualifications
that are weaker than LICQ imply the existence of a set of KKT multipliers at the so-
lution. Thus, we do not have uniqueness of the multipliers at the solution. Strong
second-order constraint qualifications (respectively weak second-order constraint
qualifications) are conditions over the constraints under which it can be claimed that,
if x is a minimum point (for any objective function f ), then x verifies the KKT condi-
tion and there is, at least, a KKT multiplier vector that verifies SSONC (respectively
WSONC).
As the example in [21] (rediscovered by Anitescu in [22]) shows, the Mangasarian-
Fromovitz constraint qualification is neither a strong nor a weak second-order con-
straint qualification. Thus, from those first-order constraint qualifications mentioned
J Optim Theory Appl

before, the only condition that can be a strong or a weak second-order constraint
qualification is the constant-rank condition. Here, we prove that CRCQ is, in fact,
a strong second-order constraint qualification. This is the main goal of the present
work.
In the article where the CRCQ was defined [1], the condition was used in the con-
text of nonconvex mathematical programming problems under general perturbation.
In the last years, CRCQ has been used to obtain theoretical results with practical rel-
evance in the context of bilevel problems; see for example [23]. In [24], CRCQ has
been used to achieve global and superlinear convergence of an infeasible interior-
point algorithm for monotone variational inequality problems. In [25], an augmented
Lagrangian method with convergence under a weaker constraint qualification (CPLD)
was defined. This convergence result implies convergence under CRCQ. In [26], the
CRCQ was used to investigate the properties of the parametric set defined by the
equality and inequality constraints. The author of [26] shows that, in the absence of
parameters, the CRCQ implies that the Mangasarian-Fromovitz constraint qualifica-
tion holds in some alternative expression of the feasible set.
As mentioned before, it is important, from the practical point of view, to define
weak second-order constraint qualifications. We consider the weak constant-rank
condition introduced in [27] to define a new weak second-order constraint qualifi-
cation. We believe that this new second-order constraint qualification can be used in
the future in the study of problems having complementarity constraints.
This paper is organized as follows. In Sect. 2, we state the main definitions. In
Sect. 3, we prove that the constant-rank constraint qualification is a strong second-
order constraint qualification. Also in this section, we prove that a generalization of
the second-order constraint qualification introduced in [27] implies WSONC for any
KKT multiplier vector. The conclusions are given in Sect. 4.

2 Definitions

Let h : Rn → Rm and g : Rn → Rp be twice continuously differentiable functions.

Define the feasible set  as

 := {x ∈ Rn : h(x) = 0, g(x) ≤ 0}.

Definition 2.1 (Ref. [1]) Given a family of differentiable functions {fi (x) : i =
1, . . . , r}, fi : Rn → R, we say that the constant-rank condition holds at x ∗ if and
only if, for any subset K ⊂ {i ∈ {1, . . . , r} : fi (x ∗ ) = 0}, the family of gradients

{∇fi (x)}i∈K

remains of constant rank near the point x ∗ .

Definition 2.2 (Ref. [1]) Given the family of differentiable functions {hi (x) : i =
1, . . . , m; gi (x) : i = 1, . . . , p} associated with problem (1), we say that a feasible
point x ∗ ∈  satisfies the constant-rank constraint qualification (CRCQ) if and only
if the constant-rank condition holds at x ∗ .
 J Optim Theory Appl

As we mentioned before, in [1] it was proved that CRCQ is a first-order constraint

qualification.
Several well-known constraint qualifications with practical relevance imply the
CRCQ. For example, the CRCQ is clearly implied by the LICQ. If all the constraints
are defined by affine functions, the CRCQ is obviously fulfilled. Moreover, if x ∈
 satisfies the CRCQ and some equality constraint hi (x) = 0 is replaced by two
inequality constraints (hi (x) ≤ 0 and −hi (x) ≤ 0) the CRCQ still holds with the new
description of the feasible set. We observe that the Mangasarian-Fromovitz constraint
qualification does not accomplish those properties.

3 Main Results

The proof that constant-rank is a strong second-order constraint qualification needs

the following two technical results. Both propositions were established in [1] in the
context of nonconvex mathematical programming problems under general perturba-
tions. For the sake of completeness, we will rewrite the results differently but in an
equivalent form.

Proposition 3.1 (Ref. [1]) Let {fi (x); i ∈ K} be a family of differentiable functions
on Rn such that the Jacobian matrix (∇fi (x))i∈K is of constant rank in a neighbor-
hood of x ∗ . Define the linear subspace

E = {y ∈ Rn : ∇fi (x ∗ )T y = 0, i ∈ K}.

Then, there exists some local diffeomorphism φ : V1 → V2 , where V1 , V2 are neigh-

borhoods of x ∗ , such that:
(i) φ(x ∗ ) = x ∗ ,
(ii) the Jacobian matrix of φ at x ∗ is the identity matrix,
(iii) the functions fi ◦ φ −1 (i ∈ K) are of constant value for all y ∈ E.

Remark 3.1 As it was mentioned in [1], this result is a special case of the constant-
rank theorem in Malliavin [28]. Observe that the hypothesis required is not the
constant-rank condition but the constant rank of the Jacobian matrix in a neighbor-
hood of the point. This is the hypothesis that appears in the constant-rank theorem in
Malliavin.

Proposition 3.2 is a consequence of Proposition 3.1 and it was used in [1] to prove
that CRCQ is a first-order constraint qualification. It will be rewritten in a new form
that will be more useful for us to prove the strong second-order necessary condi-
tion. The second part of the proposition was not stated in [1] and it can be deduced
straightforwardly from the proof made in it. We will explain it here for completeness.

Proposition 3.2 Assume that the constant-rank constraint qualification holds at

x ∗ ∈ . Then, for each y ∈ Rn such that

∇hi (x ∗ )T y = 0, i = 1, . . . , m; ∇gi (x ∗ )T y ≤ 0, i ∈ A(x ∗ ), (4)

J Optim Theory Appl

there exists some arc t → ξ(t), t ∈ (0, t¯), t¯ > 0, such that ξ(t) ⊂  and
∗
(a) limt→0+ ξ(t) = x ∗ , limt→0+ ξ(t)−x
t = y,
(b) for all j ∈ A(x ∗ ) such that ∇gj (x ∗ )T y = 0 then gj (ξ(t)) = 0.

Proof Let us suppose that A(x ∗ ) = {1, . . . , r} and define fi (x) = hi (x),
∀i = 1, . . . , m, fm+i (x) = gi (x), ∀i = 1, . . . , r. Take K = {i ∈ {1, . . . , m + r} :
∇fi (x ∗ )T y = 0} and consider the linear subspace E = {d ∈ Rn :
∇fi (x ∗ )T d = 0, i ∈ K}. Since the constant-rank constraint qualification holds at x ∗ ,
according to the hypothesis, we can use Proposition 3.1. Thus, there exists some local
diffeomorphism φ verifying items (i), (ii) and (iii) of Proposition 3.1.
Define the arc ξ(t) by
ξ(t) = φ −1 (x ∗ + ty),
for t > 0 sufficiently small. Then, by continuity and conditions (i) and (ii) of Propo-
sition 3.1, we have that

ξ(t) − x ∗
lim ξ(t) = x ∗ , lim = y.
t→0+ t→0+ t

Thus, the arc ξ(t) verifies (a).

For any index j ∈ K, using item (iii) of Proposition 3.1, we have that, for t > 0
sufficiently small,

fj (ξ(t)) = fj (φ −1 (x ∗ + ty)) = fj (φ −1 (x ∗ )) = fj (x ∗ ) = 0.

Thus, if j ∈ A(x ∗ ) is such that ∇gj (x ∗ )T y = 0, then we have that gj (ξ(t)) = 0 for
t > 0 small enough. This proves that (b) holds.
Let us prove that the arc is feasible. For i ∈ A(x ∗ )\K, we have that ∇fi (x ∗ )T d <
0. Thus,

fi (ξ(t)) = t∇fi (x ∗ )T y + tε(t), with ε(t) such that ε(t) → 0.

Then, for t¯1 > 0 sufficiently small, we have that

fi (ξ(t)) ≤ 0, 0 < t < t¯.

A similar argument, based on the continuity of each fi (x) would show that, for i ∈
{1, . . . , p}\A(x ∗ ), there exists t¯2 > 0 sufficiently small such that

fi (ξ(t)) ≤ 0, 0 < t < t¯2 .

Thus, for t¯ = min{t¯1 , t¯2 }, we have that ξ(t) ⊂ , 0 < t < t¯. 

In the next theorem, we prove that, if a minimum point verifies the constant-rank
constraint qualification, then the strong second-order necessary condition holds for
any KKT multiplier.
 J Optim Theory Appl

Theorem 3.1 Let x ∗ be a minimum point of problem (1) that verifies the constant-
rank constraint qualification. Then, for any KKT multiplier vector (λ∗ , μ∗ ) ∈ Rm+p ,
x ∗ verifies the strong second-order necessary optimality condition.

Proof The existence of a set of KKT multipliers at x ∗ was clearly established in [1].
Let us prove that the strong second-order necessary condition is verified for any of
those multipliers. Let (λ, μ) ∈ Rm+p be any fixed KKT multiplier.
Take a direction y ∈ T̃ (x ∗ ). Then, y verifies

∇hi (x ∗ )T y = 0, i = 1, . . . , m,
∇gj (x ∗ )T y = 0, j ∈ A+ (x ∗ ),
∇gj (x ∗ )T y ≤ 0, j ∈ A0 (x ∗ ).

Thus,
 
m  T
∇f (x ∗ ) + λi ∇hi (x ∗ ) + μi ∇gi (x ∗ ) y = 0
i=1 i∈A+ (x ∗ )

and we have that

∇f (x ∗ )T y = 0.
By Proposition 3.2, there exist a feasible arc ξ(t), t ∈ [0, t¯) such that

ξ(t) − x ∗
ξ(0) = x ∗ , ξ (0) = lim = y, gi (ξ(t)) = 0, ∀i ∈ A+ (x ∗ ).
t→0+ t
Define ϕ(t) = f (ξ(t)). Since x ∗ is a minimum point of f in  and since

ϕ (0) = ∇f (x ∗ )T y = 0,

then
d2
ϕ (0) = f (ξ(t))|t=0 = y T ∇ 2 f (x ∗ )y + ∇f (x ∗ )T ξ (0) ≥ 0. (5)
dt 2
Furthermore, by the feasibility of the arc and using gi (ξ(t)) = 0, ∀i ∈ A+ (x ∗ ), 0 ≤
t < t¯, we have that

m 
R(t) = λi hi (ξ(t)) + μi gi (ξ(t)) = 0, 0 ≤ t < t¯.
i=1 i∈A(x ∗ )

So, differentiating that relation twice and taking t = 0, we obtain


m  
∗ ∗
R (0) = y T
λi ∇ hi (x ) +
2
μi ∇ gi (x ) y
2

i=1 i∈A(x ∗ )

m  T
+ λi ∇hi (x ∗ ) + μi ∇gi (x ∗ ) ξ (0) = 0. (6)
i=1 i∈A(x ∗ )
J Optim Theory Appl

Adding (6) to (5), we obtain

 
m  
∗ ∗ ∗
y T
∇ f (x ) +
2
λi ∇ hi (x ) +
2
μi ∇ gi (x ) y ≥ 0.
2

i=1 i∈A(x ∗ )

Thus, the strong second-order necessary condition holds at x ∗ as we wanted to

prove. 

Remark 3.2 Observe that, given a direction y ∈ T̃ (x ∗ ), it can happens that not all the
gradients of the inequality active constraints verify ∇gi (x ∗ )T y = 0. Because of this,
the hypothesis of constant-rank constraint qualification is necessary.

The existence of the feasible arc ξ(t) shows that the constant-rank constraint qual-
ification can be seen as a natural generalization of the linear independence constraint
qualification. Theorem 3.1 shows that CRCQ captures correctly the geometry of the
tangent subspace T̃ (x ∗ ), just as LICQ does. Observe that, under the CRCQ, the
SSONC condition holds for any KKT multiplier.
Due to the counterexample in [21], most of the new second-order constraint qual-
ifications have the form: MFCQ + some condition, which gives the validity of the
(strong or weak) second-order necessary condition. In this scope, we can mention the
constraint qualifications defined in [29] and [27]. Both conditions imply the existence
of some KKT multiplier vector for which a second-order necessary condition holds.
In [29], the authors proved that there is a KKT multiplier for which the SSONC
condition holds at x ∗ , if x ∗ is a minimum point of (1) verifying the following second-
order constraint qualification:
(i) MFCQ holds at x ∗ .
(ii) The rank of the gradients of the equality and inequality active constraints at x ∗
is m + q − 1, where q is the number of active inequality constraints at x ∗ .
(iii) There exists at most only one index i0 ∈ A(x ∗ ) such that, if (λ, μ) is a KKT
multiplier, then μi0 = 0.
The third condition is a complementarity condition and the authors conjecture that it
is not necessary.
In [27], the authors proved that there is a KKT multiplier for which the WSONC
condition holds at x ∗ , if x ∗ is a minimum point of (1) verifying the following second-
order constraint qualification:
(i) MFCQ holds at x ∗ .
(ii) The weak constant-rank condition (WCR) holds at x ∗ : The rank of the Jacobian
matrix made of the gradients {∇hi (x)}i=1,...,m ∪ {∇gi (x)}i∈A(x ∗ ) does not change
in a neighborhood of x ∗ .
The proof that the last condition is a weak second-order constraint qualification was
achieved using penalty ideas and this constraint qualification was used in the conver-
gence analysis of the second-order augmented Lagrangian method defined in [27].
Using Proposition 3.2, we are able to prove the following theorem that generalizes
the result obtained in [27].
 J Optim Theory Appl

Theorem 3.2 Let x ∗ be a minimum point of problem (1) that verifies the KKT con-
dition and the weak constant-rank condition. Then, for any KKT multiplier vector
(λ∗ , μ∗ ) ∈ Rm+p , the weak second-order necessary condition is verified.

Proof Let take y ∈ T (x ∗ ). Then, y verifies ∇hi (x ∗ )T y = 0, i = 1, . . . , m,

∇gj (x ∗ )T y = 0, j ∈ A(x ∗ ). Thus, y verifies (4) and we can prove the existence
of a feasible arc as in Proposition 3.2. The proof follows the idea of Theorem 3.1. 

Remark 3.3 In Theorem 3.2, we prove that the weak second-order necessary con-
dition holds at a minimum point, under any first-order constraint qualification and
the weak constant-rank condition. The result of this theorem is interesting when we
consider the following special class of mathematical programming problems with
complementarity constraints:

min f (x, y, z), s.t. x ≥ 0, y ≥ 0, xy = 0.

In this kind of problems, the origin (0, 0, 0) is a problematic point for most of the non-
linear programming algorithms, since it only verifies the Guignard constraint qualifi-
cation [13]. We observe that this point verifies the new weak second-order constraint
qualification defined by Guignard + WRC.

In [27], there is an example showing that the weak constant-rank condition is

not a first-order constraint qualification. In that example, the problem has equality
and inequality constraints. The KKT condition does not hold in the example due to
the presence of an inequality constraint. The WCR condition was previously defined
in [30], for problems with just equality constraints. In that work, the author proved
that, in this case, WCR implies the Abadie constraint qualification. In the following
theorem, we prove that the WCR condition is also a second-order constraint qualifi-
cation whenever the problem has only equality constraints.

Theorem 3.3 Let x ∗ be a minimum point of the following problem with only equality
constraints:
min f (x), s.t. h(x) = 0.
Suppose that x ∗ verifies the weak constant-rank condition. Then, x ∗ verifies the
KKT condition and the second-order necessary condition for any KKT multiplier.

Proof This theorem is a consequence of Proposition 3.1 and Proposition 3.2. 

Remark 3.4 Observe that, when the problem has only equality constraints, T̃ (x ∗ ) =
T (x ∗ ) and the weak second-order necessary condition is equivalent to the strong
one. The weak constant-rank condition for equality problems is weaker than the
Mangasarian-Fromovitz constraint qualification (which is, in this case, equivalent to
the linear independence constraint qualification) and weaker than the constant-rank
constraint qualification. Therefore, the constant-rank constraint qualification can be
seen as a generalization of those conditions.
J Optim Theory Appl

4 Conclusions

The constant-rank constraint qualification seems to be a useful tool for the analysis of
convergence of some nonlinear programming methods; see for example [23–25]. The
status of the constant-rank as a first-order constraint qualification was clearly proved
in [1]. In this work, we proved that CRCQ is in fact a strong second-order constraint
qualification. We proved that, under this constraint qualification, a minimum point
verifies the strong second-order necessary condition for any KKT multiplier.
Given the practical importance of the weak second-order necessary condition,
we generalized the second-order constraint qualification defined in [27]. We proved
that, on having a minimum point that verifies any first-order constraint qualification
and the weak constant-rank condition, then the WSONC holds for any KKT mul-
tiplier. WCR as well as MFCQ, were useful for obtaining the convergence of the
second-order augmented Lagrangian method defined in [27]. Throughout this work,
we proved that WCR is a first-order and a second-order constraint qualification when
we consider equality constraint problems. We believe that this fact can be useful
in the future, since WCR can be seen as a generalization of the regularity and the
constant-rank constraint qualification for this kind of problems.
It is still an open issue the challenge of finding new and weaker constraint qual-
ifications, not only of first-order but also of second-order. It is desirable that those
conditions be easily verifiable and possibly be associated from a practical point of
view with the convergence analysis of nonlinear optimization algorithms.

Acknowledgements We are grateful to the Associate Editor Franco Giannessi whose comments helped
us to improve the first version of this paper.

References

1. Janin, R.: Direction derivative of the marginal function in nonlinear programming. Math. Program.
Stud. 21, 127–138 (1984)
2. Karush, W.: Minima of functions of several variables with inequalities as side conditions. Master’s
Thesis, University of Chicago (1939)
3. Kuhn, H.W., Tucker, A.W.: Nonlinear programming. In: Neyman, J. (ed.) Proceedings of the Second
Berkeley Symposium on Mathematical Statistics and Probability, pp. 481–492. University of Califor-
nia Press, Berkeley (1951)
4. Bertsekas, D.P.: Nonlinear Programming, 2nd edn. Athena Scientific, Belmont (1999)
5. Mangasarian, O.L.: Nonlinear Programming. McGraw Hill, Bombay (1969)
6. Mangasarian, O.L., Fromovitz, S.: The Fritz-John necessary optimality conditions in presence of
equality and inequality constraints. J. Math. Analysis Appl. 17, 37–47 (1967)
7. Qi, L., Wei, Z.: On the constant positive linear dependence condition and its application to SQP
methods. SIAM J. Optim. 10, 963–981 (2000)
8. Andreani, R., Martínez, J.M., Schuverdt, M.L.: On the relation between the Constant Positive Linear
Dependence condition and quasinormality constraint qualification. J. Optim. Theory Appl. 125, 473–
485 (2005)
9. Hestenes, M.R.: Optimization Theory—The finite-dimensional case. Wiley, New York (1975)
10. Bertsekas, D.P., Ozdaglar, A.E.: Pseudonormality and a Lagrange multiplier theory for constrained
optimization. J. Optim. Theory Appl. 114, 287–343 (2002)
11. Bertsekas, D.P., Ozdaglar, A.E.: The relation between pseudonormality and quasiregularity in con-
strained optimization. Optim. Methods Softw. 19, 493–506 (2004)
12. Abadie, J.: On the Kuhn-Tucker Theorem. In: Abadie, J. (ed.) Nonlinear Programming. North-
Holland, Amsterdam (1967)
 J Optim Theory Appl

13. Guignard, M.: Generalized Kuhn-Tucker conditions for mathematical programming problems in a
Banach space. SIAM J. Control 7, 232–241 (1969)
14. Gould, F.J., Tolle, J.W.: A necessary and sufficient qualification for constrained optimization. SIAM
J. Appl. Math. 20, 164–172 (1971)
15. Moldovan, A., Pellegrini, L.: On regularity for Constrained Extremum Problem. Part 1: Sufficient
optimality conditions. J. Optim. Theory Appl. 142, 147–163 (2009)
16. Moldovan, A., Pellegrini, L.: On regularity for Constrained Extremum Problem. Part 2: Necessary
optimality conditions. J. Optim. Theory Appl. 142, 165–183 (2009)
17. Giannessi, F.: Constrained Optimization and Image Space Analysis, Separation of Sets and Optimality
Conditions, vol. 1. Springer, New York (2005)
18. Andreani, R., Birgin, E.G., Martínez, J.M., Schuverdt, M.L.: Second-order negative-curvature meth-
ods for box-constrained and general constrained optimization. Comput. Optim. Appl. (2009).
doi:10.1007/s10589-009-9240-y
19. Gould, N.I.M., Toint, Ph.L.: A note on the convergence of barrier algorithms for second-order neces-
sary points. Math. Program. 85, 433–438 (1999)
20. Nocedal, J., Wright, S.J.: Numerical Optimization. Springer, New York (1999)
21. Arutyunov, A.V.: Perturbations of extremum problems with constraints and necessary optimality con-
ditions. J. Sov. Math. 54, 1342–1400 (1991)
22. Anitescu, M.: Degenerate nonlinear programming with a quadratic growth condition. SIAM J. Optim.
10, 1116–1135 (2000)
23. Dempe, S.: Foundations of Bilevel Programming, Nonconvex Optimization and its Applications Se-
ries, vol. 61. Kluwer, Amsterdam (2002)
24. Ralph, D., Wright, S.: A superlinear infeasible-interior-point algorithm for monotone complementar-
ity problems. Math. Oper. Res. 21(4), 815–838 (1996)
25. Andreani, R., Birgin, E.G., Martínez, J.M., Schuverdt, M.L.: On augmented Lagrangian methods with
general lower-level constraints. SIAM J. Optim. 18, 1286–1309 (2007)
26. Lu, S.: Implications of the constant rank constraint qualification. Math. Program. A (2009).
doi:10.1007/s10107-009-0288-3
27. Andreani, R., Martínez, J.M., Schuverdt, M.L.: On second-order optimality conditions for nonlinear
programming. Optimization 56, 529–542 (2007)
28. Malliavin, P.: Geometrie differentielle intrinsique. Herman, Paris (1974)
29. Baccari, A., Trad, A.: On the classical necessary second-order optimality conditions in the presence
of equality and inequality constraints. SIAM J. Optim. 15, 394–408 (2004)
30. Penot, J.P.: A new constraint qualification condition. J. Optim. Theory Appl. 48(3), 459–468 (1986)