A Classroom Note on Twice Continuously Di ff erentiable Strictly Convex and Strongly Quasiconvex Functions

We provide some remarks and clarifications for twice continuously differentiable strictly convex and strongly quasiconvex functions. Characterizations of these classes and their relationships with other classes of generalized convex functions are also examined.


Introduction
It is well-known that if f : R n −→ R is a twice continuously differentiable function on the open convex set X ⊆ R n , then the following results hold: a) The function f is convex on X if and only if its Hessian matrix H f (x) is positive semidefinite at every point x ∈ X. b) If H f (x) is positive definite at every point x ∈ X, then f is strictly convex on X.
Condition b) is therefore only sufficient for the strict convexity of a twice continuously differentiable function on an open convex set.Consider, e. g., the function f (x 1 , x 2 ) = (x 1 ) 4 + (x 2 ) 4 , which is obviously strictly convex on the whole R 2 , but for which H f (0, 0) is the zero matrix.We have to note that the above results a) and b) can be better precised in the following results.See, e. g., Bertsekas (2009), Hiriart-Urruty and Lemaréchal (1993).
Theorem 1.Let X ⊆ R n be a nonempty convex set and let f : R n −→ R be twice continuously differentiable over an open set that contains X. i) If H f (x) is positive semidefinite for all x ∈ X, then f is convex on X.
ii) If H f (x) is positive definite for all x ∈ X, then f is strictly convex on X.
iii) If X is open and f is convex on X, then H f (x) is positive semidefinite for all x ∈ X.

Remark 1.
Theorem 4.28 in Güler ( 2010) is not entirely correct, on the grounds of Theorem 1. Indeed, if f is convex over a convex set that is not open, H f (x) may not be positive semidefinite at any point of X : take for example X = {(x 1 , 0) : x 1 ∈ R} and f (x 1 , x 2 ) = (x 1 ) 2 − (x 2 ) 2 .However, it can be shown that the conclusions of Theorem 1 also holds if X has a nonempty interior instead of being open (i.e. X is a solid convex set).
Remark 2. Theorem 1 holds also under the assumption that f is twice Fréchet differentiable.A further weakening by means of twice Gâteaux differentiability is made by Borwein and Vanderwerff (2010).This paper is organized as follows.
In Section 2 we shall make some comments on the above classical results a) and b), in order to get a simple characterization of twice continuously differentiable strictly convex functions of several variables on an open convex set X ⊆ R n and of twice differentiable strictly convex functions of one variable on an open interval I ⊆ R.
In Section 3 we shall make some comments on twice continuously differentiable strongly quasiconvex functions on an open convex set X ⊆ R n .

Twice Continuously Differentiable Strictly Convex Functions
For the reader's convenience we recall the basic characterizations of strictly convex functions.
Theorem 2. Let f : X −→ R be defined on the convex set X ⊆ R n .Then f is strictly convex on X if and only if: (i) For each x ∈ X, for each y ∈ R n , y 0, the function Equivalently, if and only if: (ii) For each x 1 , x 2 ∈ X, x 1 x 2 , the function is strictly convex on (0, 1).
Let X ⊆ R n be open and convex and let f be differentiable on X.Then f is strictly convex on X if and only if: Equivalently, if and only if: Equivalently, if and only if: For what concerns convex functions, their characterizations corresponding to (i) and (ii) of Theorem 2 are, respectively: f is convex on the convex set X ⊆ R n if and only if: Equivalently, if and only if: Characterizations (i) and (ii) of Theorem 2 and the above characterizations (i) ′ and (ii) ′ show that the concept of convex and strictly convex functions is genuinely unidimensional: f is convex (strictly convex) on a convex set X ⊆ R n if and only if the restriction of f to each line segment contained in X is a convex (strictly convex) function.It is therefore useful to recall the main properties concerning convex and strictly convex functions of one single variable.
Theorem 3. Let φ : 1.If φ is differentiable on I, the function φ is convex on I if and only if its derivative φ ′ is increasing on I.
2. If φ is differentiable on I, the function φ is strictly convex on I if and only if its derivative φ ′ is strictly increasing on I.
It is well-known that: The following result is less known; for the reader's convenience we give a proof.
Theorem 4. Let f : I −→ R be differentiable on the open interval I ⊆ R. Then f is strictly increasing (resp.: strictly decreasing) on I if and only if f ′ (x) 0 (resp.: f ′ (x) 0), ∀x ∈ I, and there exists no subinterval of I where f ′ (x) = 0 in all points of the said subinterval.In other words, it must hold f ′ (x) 0 (resp.: f ′ (x) 0), ∀x ∈ I and the set

Proof.
i) The condition is necessary.It is well-known that it must hold f ′ (x) 0 (resp.: f ′ (x) 0), ∀x ∈ I.If there would exist an interval J ⊆ I such that, for each x ∈ J, we have f ′ (x) = 0, the restricftion of f to J would be constant on J, which is absurd.
ii) The condition is sufficient.Let us consider any pair of distinct points of I, say x ′ and x ′′ with x ′ < x ′′ .Thanks to the mean value theorem (or Lagrange mean value theorem), applied to the restriction of where ξ is a suitable point of (x ′ , x ′′ ).Therefore, taking the assumptions into account: i. e. f is increasing (resp.: decreasing) on I. From this, obviously we have also: x ′′ ] would be constant on [x ′ , x ′′ ] and this would imply that every x ∈ [x ′ , x ′′ ] , it holds f ′ (x) = 0, against the assumption that no subinterval of I exists such that the derivative of f is identically zero on the same subinterval.It results therefore i. e. the thesis.
We can therefore formulate the following result.
is strictly convex on R. We have indeed but the infinitely many points where f ′′ (x) = 0 are all isolated points.
The previous results provide a simple way to obtain the characterization of convex functions of several variables, which are twice continuously differentiable on some open convex set X ⊆ R n .First we restate a part of Theorem 1 in its "classical" version.
Theorem 1 ′ .Let f : X −→ R n be twice continuously differentiable on a nonempty open convex set X ⊆ R n .Then f is convex on X if and only if its Hessian matrix H f (x) is positive semidefinite at each point x ∈ X.
Proof.Let x 0 ∈ X, v ∈ R n , v 0 and consider the restriction We recall that a positive (negative) semidefinite quadratic form Q(x) = x ⊤ Ax (A symmetric) is positive (negative) definite if and only if A is a nonsingular matrix (see, e. g.Hestenes (1966), Theorem 6.3).In other words, for positive semidefinite matrices we may have x ⊤ Ax = 0 even if x 0, but in this case we neverthless have Ax = 0, and being x 0, it must hold |A| = 0. Indeed, if y ∈ R n and t ∈ R are arbitrary, then If x ⊤ Ax = 0, x 0, this implies y ⊤ Ax = 0. Now let y = Ax, so (Ax) ⊤ Ax = 0, whence finally Ax = 0. Therefore, being x 0, it must hold |A| = 0.
Contrary to the unidimensional case, the condition: • " H f (x) is positive semidefinite for every x ∈ X and |H f (x)| is not identically zero on any segment belonging to X" is only sufficient for the strict convexity of the twice continuously differentiable function f on the open convex set X ⊆ R n (see Fenchel (1953), Ortega and Rheinboldt (1970)).
i) See also Bernstein and Toupin (1962).Suppose n = 2, X = {(x 1 , x 2 ) : x 2 < 0} and The determinant of its Hessian matrix is which is positive throughout X except on the line x 1 = 0 where it vanishes.It is seen that f is convex on X but not strictly convex.
Necessary and sufficient conditions for the strict convexity of twice continuously differentiable functions have been established by Bernstein and Toupin (1962) and by Diewert and others (1981).Following these last authors, we "translate" the conditions of Bernstein and Toupin with a more convenient notation and statement.
Theorem 6.Let f : X −→ R be twice continuously differentiable on an open convex set X ⊆ R n .Then f is strictly convex on X if and only if: Remark 3. Diewert and others (1981) prove that another (equivalent) characterization of twice continuously differentiable strictly convex functions (on an open convex set X ⊆ R n ) is: The recent paper of Stein ( 2012) is concerned with the more general case of twice differentiable strictly convex functions defined on a convex set X ⊆ R n , not necessarily open.
Remark 4. Ginsberg (1973) defines the class of strongly convex functions, as those twice continuously differentiable functions on an open convex set X ⊆ R n for which all leading principal minors (see Section 3, after Theorem 7) of their Hessian matrix H f (x) are positive, for each x ∈ X. Obviously, this one is a sufficient condition for f to be a strictly convex function on X (Theorem 1), however the above definition is misleading, as in the current literature on Convex Analysis (see, e. g., Diewert and others (1981), Avriel and others (1981), Rockafellar (1976), Vial (1982)) twice continuously differentiable strongly convex functions (on an open convex set X ⊆ R n ) are characterized by the property: • There exists α > 0 such that x ∈ X =⇒ H f (x) − αI is positive semidefinite (I is the identity matrix).
Remark 5.For the special case of quadratic functions, i. e. of the functions where A is a symmetric matrix of order n, some of the previous results can be stated as follows.
a) Martos (1975) proved for φ(x) what already pointed out in Remark 1: φ(x) is convex on any solid convex set X ⊆ R n if and only if A is positive semidefinite.
b) The quadratic function φ(x) is strictly convex on R n if and only if A is positive definite (the same result holds with reference to a solid convex set X ⊆ R n ).Indeed, if φ is strictly convex and h ∈ R n , h 0, the first-order characterization of strictly convex functions gives We have h ⊤ Ah > 0 for all h ∈ R n , h 0, that is A is positive definite.Conversely, if A is positive definite and h 0, then which implies that φ is strictly convex.
c) Martos (1975) has proved that φ(x) is quasiconvex (see Section 3) on R n if and only if it is convex on R n .This shows that there is no reason to study quadratic functions that are quasiconvex, without being convex, on the whole R n .

Twice Continuously Differentiable Strongly Quasiconvex Functions
First we recall some basic definitions.
Definition 2. Let f be defined on a convex set X ⊆ R n ; then f is said to be quasiconvex on X if } for every x 1 , x 2 ∈ X and for every λ ∈ [0, 1] or, equivalently, for every x 1 , x 2 ∈ X and for every λ ∈ [0, 1].
Definition 3. A function f be defined on a convex set X ⊆ R n is said to be strictly quasiconvex on X if } for every x 1 , x 2 ∈ X, x 1 x 2 , and for every λ ∈ (0, 1) or, equivalently, for every x 1 , x 2 ∈ X, x 1 x 2 , and for every λ ∈ (0, 1).
Definition 4. A function f be defined on a convex set X ⊆ R n is said to be semistrictly quasiconvex on X if } for every x 1 , x 2 ∈ X, with f (x 1 ) f (x 2 ), and for every λ ∈ (0, 1) or, equivalently, for every x 1 , x 2 ∈ X, and for every λ ∈ (0, 1).
Under lower semicontinuity of f we have the following inclusion diagram.
strictly convex =⇒ strictly quasiconvex In their pioneering paper on quasiconcave functions and quasiconcave programming, Arrow and Enthoven (1961) give the following necessary conditions for a twice continuously differentiable function to be quasiconvex on the open convex set X ⊆ R n . where See also Avriel (1972), Kemp and Kimura (1978) and, for characterizations of twice continuously differentiable quasiconvex functions, Crouzeix (1980), Crouzeix and Ferland (1982), Diewert and others (1981).Needless to say, condition (1) is trivially satisfied for r = 1.In the same paper Arrow and Enthoven show that the following condition is sufficient for the quasiconvexity of f on X.Indeed, this condition is even sufficient for strict quasiconvexity and more: see, e. g., Ginsberg (1973), Diewert and others (1981).
Relations (1) and ( 2) require a brief review on the main properties concerning quadratic forms subject to a system of homogeneous linear constraints.Given a (real) symmetric matrix A, of order n, and its associated quadratic form we are interested in the sign of (3), but when x ∈ S , S being the set of non trivial solutions of the homogeneous system of linear equations Bx = 0, where B is a (real) (m, n) matrix, with m < n.This problem has been treated by several authors; see, e. g., Chabrillac and Crouzeix (1984), Debreu (1952), Farebrother (1977), Giorgi (2003Giorgi ( , 2017)), Murata (1977), Samuelson (1947).The following result may be considered a generalization of the Sylvester criterion for unconstrained quadratic forms, to the ∇ f (x) 0, for x ∈ X, then H f (x) has to be positive definite on the subspace orthogonal to the gradient vector ∇ f (x).In this last case Theorem 8 can be usefully applied.First we recall that f : Pseudoconvex functions are semistrictly quasiconvex and therefore also quasiconvex, but not strictly quasiconvex and not strongly quasiconvex.Under differentiability assumption we have the following relationships.
strongly quasiconvex ⇒ pseudoconvex ⇒ semistrictly quasiconvex ⇒ quasiconvex ⇓ ⇑ ⇒ strictly quasiconvex ⇒ It turns out that ( 8) is a sufficient condition for a twice continuously differentiable function f : X −→ R to be pseudoconvex and also strictly quasiconvex on the open convex set X ⊆ R n .
Theorem 11.Let f : X −→ R be twice continuously differentiable on the open convex set X ⊆ R n and let ∇ f (x) 0 for all x ∈ X.Then f is quasiconvex on X and also pseudoconvex on X if and only if v ⊤ H f (x)v 0 for all x ∈ X and all v ∈ R n , v 0, such that ∇ f (x)v = 0.
In other words, under the assumptions of Theorem 11, the Hessian matrix of f has to be positive semidefinite on the subspace orthogonal to ∇ f (x), for x ∈ X.
Remark 10.Crouzeix and Ferland (1982) prove that, if ∇ f (x) 0 for all x ∈ X, then any one of the following conditions is equivalent to condition (9): I) Either H f (x) is positive semidefinite for every x ∈ X, or H f (x) has one simple negative eigenvalue, for every x ∈ X, and there exists a vector b ∈ R n such that H f (x)b = (∇ f (x)) ⊤ and ∇ f (x)b 0.
II) If we denote by M(x) the following bordered matrix (of order (n + 1)): ] , then M(x) has one simple negative eigenvalue, for every x ∈ X. III) All the principal minors of M(x) (and not only its leading principal minors) are less than or equal to zero, for every x ∈ X.
Taking the previous theorems into account, the following results are at hand.Theorem 12. Let f : X −→ R be twice continuously differentiable on the open convex set X ⊆ R n and let ∂ f ∂x 1 0 for all x ∈ X.Let us denote again by M(x) the bordered matrix of Remark 10.Then, f is strongly quasiconvex on X if and only if the leading principal minors of M(x) of order 3, 4, ..., n + 1, are all negative for all x ∈ X.
Theorem 13.Let f : X −→ R be twice continuously differentiable on the open convex set X ⊆ R n and let f be quasiconvex on X, with ∇ f (x) 0 for all x ∈ X.Then f is strongly quasiconvex on X if and only if |M(x)| 0, for all x ∈ X.