Abstract
In Chap. 8 we study discrete and continuous descent methods for minimizing a convex (Lipschitz) function on a general Banach space. We consider a space of vector fields V such that for any point x in the Banach space, the directional derivative in the direction Vx is nonpositive. This space of vector fields is equipped with a complete metric. Each vector field generates two gradient type algorithms (discrete descent methods) and a flow which consists of the solutions of the corresponding evolution equation (continuous descent method). We show that most (in the sense of Baire category) vector fields produce algorithms for which values of the objective function tend to its infimum as t tends to infinity. Actually, we introduce the subclass of regular vector fields, show that the convergence property stated above holds for them and that a generic vector field is regular. We also show that this convergence property is stable under small perturbations of a given regular vector field.
Access provided by Autonomous University of Puebla. Download chapter PDF
Keywords
These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.
8.1 Discrete Descent Methods for a Convex Objective Function
Given a Lipschitzian convex function f on a Banach space X, we consider a complete metric space \({\mathcal{A}}\) of vector fields V on X with the topology of uniform convergence on bounded subsets. With each such vector field we associate two iterative processes. We introduce the class of regular vector fields \(V \in{ \mathcal{A}}\) and prove (under two mild assumptions on f) that the complement of the set of regular vector fields is not only of the first category, but also σ-porous. We then show that for a locally uniformly continuous regular vector field V and a coercive function f, the values of f tend to its infimum for both processes. These results were obtained in [136].
Assume that (X,∥⋅∥) is a Banach space with norm ∥⋅∥, (X ∗,∥⋅∥∗) is its dual space with the norm ∥⋅∥∗, and f:X→R 1 is a convex continuous function which is bounded from below. Recall that for each pair of sets A,B⊂X ∗,
is the Hausdorff distance between A and B.
For each x∈X, let
be the subdifferential of f at x. It is well known that the set ∂f(x) is nonempty and bounded (in the norm topology). Set
Denote by \({\mathcal{A}}\) the set of all mappings V:X→X such that V is bounded on every bounded subset of X (i.e., for each K 0>0 there is K 1>0 such that ∥Vx∥≤K 1 if ∥x∥≤K 0), and for each x∈X and each l∈∂f(x), l(Vx)≤0. We denote by \({\mathcal{A}}_{c}\) the set of all continuous \(V \in{ \mathcal{A}}\), by \({\mathcal{A}}_{u}\) the set of all \(V \in{ \mathcal{A}}\) which are uniformly continuous on each bounded subset of X, and by \({\mathcal{A}}_{au}\) the set of all \(V\in{ \mathcal{A}}\) which are uniformly continuous on the subsets
for each integer n≥1. Finally, let \({\mathcal{A}}_{auc}={\mathcal{A}}_{au} \cap{ \mathcal{A}}_{c}\).
Next we endow the set \({\mathcal{A}}\) with a metric ρ: For each \(V_{1},V_{2} \in{ \mathcal{A}}\) and each integer i≥1, we first set
and then define
Clearly \(({\mathcal{A}},\rho)\) is a complete metric space. It is also not difficult to see that the collection of the sets
where N,ε>0, is a base for the uniformity generated by the metric ρ. Evidently \({\mathcal{A}}_{c}\), \({\mathcal{A}}_{u}\), \({\mathcal{A}}_{au}\) and \({\mathcal{A}}_{auc}\) are closed subsets of the metric space \(({\mathcal{A}},\rho)\). In the sequel we assign to all these spaces the same metric ρ.
To compute inf(f), we are going to associate with each vector field \(W \in{ \mathcal{A}}\) two gradient-like iterative processes (see (8.5) and (8.7) below).
The study of steepest descent and other minimization methods is a central topic in optimization theory. See, for example, [2, 19, 44, 47, 69, 73, 103] and the references mentioned therein. Note, in particular, that the counterexample studied in Sect. 2.2 of Chap. VIII of [73] shows that, even for two-dimensional problems, the simplest choice for a descent direction, namely the normalized steepest descent direction,
may produce sequences the functional values of which fail to converge to the infimum of f. This vector field V belongs to \(\mathcal{A}\) and the Lipschitzian function f attains its infimum. The steepest descent scheme (Algorithm 1.1.7) presented in Sect. 1.1 of Chap. VIII of [73] corresponds to any of the two iterative processes we consider below.
In infinite dimensions the problem is even more difficult and less understood. Moreover, positive results usually require special assumptions on the space and the functions. However, as shown in our paper [135] (under certain assumptions on the function f), for an arbitrary Banach space X and a generic vector field \(V \in \mathcal{A}\), the values of f tend to its infimum for both processes. In that paper, instead of considering a certain convergence property for a method generated by a single vector field V, we investigated it for the whole space \(\mathcal{A}\) and showed that this property held for most of the vector fields in \(\mathcal{A}\).
Here we introduce the class of regular vector fields \(V \in \mathcal{A}\). Our first result, Theorem 8.2, shows (under the two mild assumptions A(i) and A(ii) on f stated below) that the complement of the set of regular vector fields is not only of the first category, but also σ-porous in each of the spaces \(\mathcal{A}\), \({\mathcal{A}}_{c}\), \({\mathcal{A}}_{u}\), \({\mathcal{A}}_{au}\) and \({\mathcal{A}}_{auc}\). We then show (Theorem 8.3) that for any regular vector field \(V \in{ \mathcal{A}}_{au}\), if the constructed sequence \(\{x_{i}\}_{i=0}^{\infty} \subset X\) has a bounded subsequence (in the case of the first process) or is bounded (in the case of the second one), then the values of the function f tend to its infimum for both processes. If, in addition to A(i) and A(ii), f also satisfies the assumption A(iii), then this convergence result is valid for any regular \(V \in \mathcal{A}\). Note that if the function f is coercive, then the constructed sequences will always stay bounded. Thus we see, by Theorem 8.2, that for a coercive f the set of divergent descent methods is σ-porous. Our last result, Theorem 8.4, shows that in this case we obtain not only convergence, but also stability.
Our results are established in any Banach space and for those convex functions which satisfy the following two assumptions.
- A(i):
-
There exists a bounded (in the norm topology) set X 0⊂X such that
$$\inf(f)=\inf\bigl\{ f(x): x \in X\bigr\} =\inf\bigl\{ f(x): x \in X_0 \bigr\} ; $$ - A(ii):
-
for each r>0, the function f is Lipschitzian on the ball {x∈X:∥x∥≤r}.
Note that we may assume that the set X 0 in A(i) is closed and convex. It is clear that assumption A(i) holds if lim∥x∥→∞ f(x)=∞.
We say that a mapping \(V \in \mathcal{A}\) is regular if for any natural number n, there exists a positive number δ(n) such that for each x∈X satisfying
and each l∈∂f(x), we have
Denote by \(\mathcal{F}\) the set of all regular vector fields \(V \in \mathcal{A}\).
It is not difficult to verify the following property of regular vector fields. It means, in particular, that \({\mathcal{G}}={\mathcal{A}} \setminus {\mathcal{F}}\) is a face of the convex cone \({\mathcal{A}}\) in the sense that if a non-trivial convex combination of two vector fields in \(\mathcal{A}\) belongs to \(\mathcal{G}\), then both of them must belong to \(\mathcal{G}\).
Proposition 8.1
Assume that \(V_{1},V_{2} \in{ \mathcal{A}}\), V 1 is regular, ϕ:X→[0,1], and that for each integer n≥1,
Then the mapping x→ϕ(x)V 1 x+(1−ϕ(x))V 2 x, x∈X, also belongs to \(\mathcal{F}\).
Our first result shows that in a very strong sense most of the vector fields in \(\mathcal{A}\) are regular.
Theorem 8.2
Assume that both A(i) and A(ii) hold. Then \({\mathcal{A}} \setminus{ \mathcal{F}}\) (respectively, \({\mathcal{A}}_{c} \setminus{ \mathcal{F}}\), \({\mathcal{A}}_{au} \setminus{ \mathcal{F}}\) and \({\mathcal{A}}_{auc} \setminus{ \mathcal{F}}\)) is a σ-porous subset of the space \({\mathcal{A}}\) (respectively, \({\mathcal{A}}_{c}\), \({\mathcal{A}}_{au}\) and \({\mathcal{A}}_{auc}\)). Moreover, if f attains its infimum, then the set \({\mathcal{A}}_{u} \setminus{ \mathcal{F}}\) is also a σ-porous subset of the space \({\mathcal{A}}_{u}\).
Now let \(W \in \mathcal{A}\). We associate with W two iterative processes.
For x∈X we denote by P W (x) the set of all
such that
Given any initial point x 0∈X, one can construct a sequence \(\{x_{i}\}_{i=0}^{\infty} \subset X\) such that for all i=0,1,…,
This is our first iterative process.
Next we describe the second iterative process.
Given a sequence \(\mathbf{a}=\{a_{i}\}_{i=0}^{\infty} \subset(0,1]\) such that
we construct for each initial point x 0∈X, a sequence \(\{x_{i}\}_{i=0}^{\infty} \subset X\) according to the following rule:
where i=0,1,….
We will also make use of the following assumption:
- A(iii):
-
For each integer n≥1, there exists δ>0 such that for each x 1,x 2∈X satisfying
$$ \begin{gathered}\|x_1\|,\|x_2\| \le n,\quad\quad f(x_i) \ge \inf(f)+1/n,\quad i=1,2,\quad \hbox{and}\\ \|x_1-x_2\| \le\delta,\end{gathered} $$the following inequality holds:
$$H\bigl(\partial f(x_1),\partial f(x_2)\bigr) \le1/n. $$
This assumption is certainly satisfied if f is differentiable and its derivative is uniformly continuous on those bounded subsets of X over which the infimum of f is larger than inf(f).
Our next result is a convergence theorem for those iterative processes associated with regular vector fields. It is of interest to note that we obtain convergence when either the regular vector field W or the subdifferential ∂f enjoy a certain uniform continuity property.
Theorem 8.3
Assume that \(W \in \mathcal{A}\) is regular, A(i), A(ii) are valid and that at least one of the following conditions holds: 1. \(W \in{ \mathcal{A}}_{au}\); 2. A(iii) is valid. Then the following two assertions are true:
-
(i)
Let the sequence \(\{x_{i}\}_{i=0}^{\infty} \subset X\) satisfy (8.5) for all i=0,1,…. If
$$\liminf_{i \to\infty}\|x_i\|<\infty, $$then lim i→∞ f(x i )=inf(f).
-
(ii)
Let a sequence \(\mathbf{a}=\{a_{i}\}_{i=0}^{\infty} \subset(0,1]\) satisfy (8.6) and let the sequence \(\{x_{i}\}_{i=0}^{\infty} \subset X\) satisfy (8.7) for all i=0,1,…. If \(\{x_{i}\}_{i=0}^{\infty}\) is bounded, then
$$\lim_{i \to\infty} f(x_i)=\inf(f). $$
Finally, we impose an additional coercivity condition on f and establish the following stability theorem. Note that this coercivity condition implies A(i).
Theorem 8.4
Assume that f(x)→∞ as ∥x∥→∞, \(V \in \mathcal{A}\) is regular, A(ii) is valid and that at least one of the following conditions holds: 1. \(V \in{ \mathcal{A}}_{au}\); 2. A(iii) is valid.
Let K,ε>0 be given. Then there exist a neighborhood \({\mathcal{U}} \) of V in \(\mathcal{A}\) and a natural number N 0 such that the following two assertions are true:
-
(i)
For each \(W \in \mathcal{U} \) and each sequence \(\{x_{i}\}_{i=0}^{N_{0}} \subset X\) which satisfies ∥x 0∥≤K and (8.5) for all i=0,…,N 0−1, the inequality \(f(x_{N_{0}}) \le\inf(f)+\varepsilon \) holds.
-
(ii)
For each sequence of numbers \(\mathbf{a}=\{a_{i}\}_{i=0}^{\infty} \subset(0,1]\) satisfying (8.6), there exists a natural number N such that for each \(W \in \mathcal{U}\) and each sequence \(\{x_{i}\}_{i=0}^{N} \subset X\) which satisfies ∥x 0∥≤K and (8.7) for all i=0,…,N−1, the inequality f(x N )≤inf(f)+ε holds.
8.2 An Auxiliary Result
Assume that \(\mathcal{K}\) is a nonempty, closed and convex subset of X. We consider the topological subspace \({\mathcal{K}} \subset X\) with the relative topology. For each function \(h: {\mathcal{K}} \to R^{1}\) define \(\inf(h):= \inf\{h(x): x \in{ \mathcal{K}}\}\).
Proposition 8.5
Let \(g: {\mathcal{K}} \to R^{1}\) be a convex, bounded from below, function which is uniformly continuous on bounded subsets of \({\mathcal{K}}\). Assume that there exists a bounded and convex set \({\mathcal{K}}_{0} \subset{ \mathcal{K}}\) such that for each \(x \in {\mathcal{K}}\), there exists \(y \in{ \mathcal{K}}_{0} \) for which g(y)≤g(x).
Then there exists a continuous mapping \(A_{g}: {\mathcal{K}} \to{ \mathcal{K}}_{0}\) which satisfies g(A g x)≤g(x) for all \(x \in{ \mathcal{K}}\) and has the following two properties:
- B(i):
-
For each integer n≥1, the mapping A g is uniformly continuous on the set
$$\bigl\{ x \in{ \mathcal{K}}: \|x\| \le n\ \textit{and}\ g(x) \ge\inf(g)+1/n\bigr\} ; $$ - B(ii):
-
if g(x)≥inf(g)+ε for some ε>0 and \(x \in {\mathcal{K}}\), then
$$g(A_gx) \le g(x)-\varepsilon /2. $$
Proof
If there exists \(x \in \mathcal{K}\) for which g(x)=inf(g), then there exists \(x^{*} \in{ \mathcal{K}}_{0}\) for which g(x ∗)=inf(g) and we can set A g (y)=x ∗ for all \(y \in{ \mathcal{K}}\). Therefore we may assume that
For each integer i≥0, there exists \(y_{i} \in{ \mathcal{K}}_{0}\) such that
Consider now the linear segments which join y 0,y 1,…,y n ,… (all contained in \({\mathcal{K}}_{0}\) by the convexity of \({\mathcal{K}}_{0}\)), represented as a continuous curve \(\gamma: [0,\infty) \to{ \mathcal{K}}_{0}\) and parametrized so that
The curve γ is Lipschitzian because the set \({\mathcal{K}}_{0}\) is bounded. Define
It is easy to see that \(A_{g}x \in{ \mathcal{K}}_{0}\) for all \(x \in{ \mathcal{K}}\), the mapping A g is continuous on \(\mathcal{K}\) and that it is uniformly continuous on the subsets
for each integer n≥1.
Assume that
There is an integer i≥0 such that
(here 0−1=∞). Then
and by (8.10), (8.9) and (8.13),
It follows from this relation, (8.8), (8.11), (8.12) and the convexity of g that
This completes the proof of Proposition 8.5. □
8.3 Proof of Theorem 8.2
We first note the following simple lemma.
Lemma 8.6
Assume that \(V_{1},V_{2} \in{ \mathcal{A}}\), ϕ:X→[0,1], and that
Then \(V \in{ \mathcal{A}}\). If \(V_{1},V_{2} \in{ \mathcal{A}}_{c}\) and ϕ is continuous on X, then \(V \in{ \mathcal{A}}_{c}\). If \(V_{1}, V_{2} \in{ \mathcal{A}}_{u}\) (respectively, \({\mathcal{A}}_{au}\), \({\mathcal{A}}_{auc}\)) and ϕ is uniformly continuous on bounded subsets of X, then \(V \in{ \mathcal{A}}_{u}\) (respectively, \({\mathcal{A}}_{au}\), \({\mathcal{A}}_{auc}\)).
For each pair of integers m,n≥1, denote by Ω mn the set of all \(V \in{ \mathcal{A}}\) such that
and
Clearly,
Therefore in order to prove Theorem 8.2 it is sufficient to show that for each pair of integers m,n≥1, the set Ω mn (respectively, \(\varOmega _{mn}\cap{ \mathcal{A}}_{c}\), \(\varOmega _{mn} \cap{ \mathcal{A}}_{au}\), \(\varOmega _{mn} \cap{ \mathcal{A}}_{auc}\)) is a porous subset of \({\mathcal{A}}\) (respectively, \({\mathcal{A}}_{c}\), \({\mathcal{A}}_{au}\), \({\mathcal{A}}_{auc}\)), and if f attains its minimum, then \(\varOmega _{mn}\cap{ \mathcal{A}}_{u}\) is a porous subset of \({\mathcal{A}}_{u}\).
By assumption A(i), there is a bounded and convex set X 0⊂X with the following property:
- C(i):
-
For each x∈X, there is x 0∈X 0 such that f(x 0)≤f(x). If f attains its minimum, then X 0 is a singleton.
By Proposition 8.5, there is a continuous mapping A f :X→X such that
and which has the following two properties:
- C(ii):
-
If x∈X, ε>0 and f(x)≥inf(f)+ε, then f(A f x)≤f(x)−ε/2;
- C(iii):
-
for any natural number n, the mapping A f is uniformly continuous on the set
$$\bigl\{ x \in X: \|x\| \le n \hbox{ and } f(x) \ge\inf(f)+1/n\bigr\} . $$
Let m,n≥1 be integers. In the sequel we will use the piecewise linear function ϕ:R 1→R 1 defined by
and
By assumption A(ii), there is c 0>1 such that
for all x,y∈X satisfying ∥x∥,∥y∥≤n+2. Choose α∈(0,1) such that
Assume that V∈Ω mn and r∈(0,1]. Let
and define V γ :X→X by
By Lemma 8.6, \(V_{\gamma}\in{ \mathcal{A}}\) and moreover, if \(V \in{ \mathcal{A}}_{c}\) (respectively, \({\mathcal{A}}_{au}\), \({\mathcal{A}}_{auc}\)), then \(V_{\gamma} \in {\mathcal{A}}_{c}\) (respectively, \({\mathcal{A}}_{au}\), \({\mathcal{A}}_{auc}\)), and if \(V \in{ \mathcal{A}}_{u}\) and f attains its minimum, then A f is constant (see C(i)) and \(V_{\gamma} \in{ \mathcal{A}}_{u}\).
Next we estimate the distance ρ(V γ ,V). It follows from (8.22) and the definition of ϕ (see (8.18)) that V γ x=Vx for all x∈X satisfying ∥x∥≥n+1 and
Since V∈Ω mn , the above equality, when combined with (8.2), (8.1), (8.22), (8.18) and (8.17), yields
Assume that \(W \in{ \mathcal{A}}\) with
Assume now that
Inequality (8.19) implies that
By (8.22), (8.26), the definition of ϕ (see (8.18)) and C(ii),
It follows from (8.26) and (8.1) that
By (8.24), (8.28) and the inequality ∥l∥∗≤c 0, we have
and
By (8.30), (8.27), (8.21) and (8.20),
Thus
In view of (8.25), we can conclude that Ω mn is porous in \({\mathcal{A}}\), \(\varOmega _{mn} \cap{ \mathcal{A}}_{c}\) is porous in \({\mathcal{A}}_{c}\), \(\varOmega _{mn} \cap{ \mathcal{A}}_{au}\) is porous in \({\mathcal{A}}_{au}\), \(\varOmega _{mn} \cap{ \mathcal{A}}_{auc}\) is porous in \({\mathcal{A}}_{auc}\), and if f attains its minimum, then \(\varOmega _{mn}\cap{ \mathcal{A}}_{u}\) is porous in \({\mathcal{A}}_{u}\). This completes the proof of Theorem 8.2.
8.4 A Basic Lemma
The following result is our key lemma.
Lemma 8.7
Assume that \(V \in{ \mathcal{A}}\) is regular, A(i), A(ii) are valid and that at least one of the following conditions holds: 1. \(V \in{ \mathcal{A}}_{au}\); 2. A(iii) is valid.
Let \(\bar{K}\) and \(\bar{\varepsilon }\) be positive. Then there exist a neighborhood \({\mathcal{U}}\) of V in \({\mathcal{A}}\) and positive numbers \(\bar{\alpha}\) and γ such that for each \(W \in{ \mathcal{U}}\), each x∈X satisfying
and each \(\beta\in(0, \bar{\alpha}]\),
Proof
There exists \(K_{0} >\bar{K}+1\) such that
By Assumption A(ii), there exists a constant L 0>4 such that
for all x 1,x 2∈X satisfying ∥x 1∥,∥x 2∥≤2K 0+4. Since V is regular, there exists a positive number δ 0∈(0,1) such that
for each y∈X satisfying ∥y∥≤K 0+4, \(f(y) \ge\inf (f)+\bar{\varepsilon }/4\), and each ξ∈∂f(y). Choose δ 1∈(0,1) such that
There exists a positive number \(\bar{\alpha}\) such that the following conditions hold:
(a) if \(V \in{ \mathcal{A}}_{au}\), then for each x 1,x 2∈X satisfying
the following inequality is true:
(b) if A(iii) is valid, then for each x 1,x 2∈X satisfying (8.38), the following inequality is true:
Next choose a positive number δ 2 such that
Now choose a positive number γ such that
and define
Assume that \(W \in{ \mathcal{U}}\), x∈X satisfies (8.31), and that \(\beta\in(0,\bar{\alpha}]\). We intend to show that (8.32) holds. To this end, we first note that (8.31), (8.33), (8.37), (8.43) and (8.41) yield
and
By these inequalities, the definition of L 0 (see (8.34)) and (8.43),
Next we will estimate f(x)−f(x+βVx). There exist θ∈[0,β] and l∈∂f(x+θVx) such that
It follows from (8.46) and the definition of L 0 (see (8.34)) that
It follows from (8.46), the definition of L 0 (see (8.34)), (8.37) and (8.31) that
Consider the case where \(V \in{ \mathcal{A}}_{au}\). By (8.47), condition (a), (8.46), (8.31) and (8.48),
By (8.46), (8.48) and the definition of δ 0 (see (8.35)),
When combined with (8.49) and (8.36), this inequality implies that
By these inequalities and (8.45),
Assume now that A(iii) is valid. It then follows from condition (b), (8.46), (8.31) and (8.48) that
Therefore there exists \(\bar{l} \in\partial f(x)\) such that \(\|\bar{l}-l\|_{*} \le\delta_{1}\). When combined with (8.45) and (8.46), this fact implies that
It follows from the definition of δ 0 (see (8.35)) and (8.31) that \(\beta\bar{l}(Vx) \le-\beta\delta_{0}\). Combining this inequality with (8.51) and (8.36), we see that
Thus in both cases (8.50) is true. It now follows from (8.50), (8.44), (8.41) and (8.42) that
Thus (8.32) holds. Lemma 8.7 is proved. □
8.5 Proofs of Theorems 8.3 and 8.4
Proof of Theorem 8.3
To show that assertion (i) holds, suppose that
We will show that
Assume the contrary. Then there exists ε>0 such that
There exists a number S>0 and a strictly increasing sequence of natural numbers \(\{i_{k}\}_{k=1}^{\infty}\) such that
By Lemma 8.7, there exist numbers α,γ∈(0,1) such that for each x∈X satisfying
and each β∈(0,α],
It follows from (8.52), (8.4), (8.5), the definitions of α and γ, (8.55) and (8.54) that for each integer k≥1,
Since this inequality holds for all integers k≥1, we conclude that
This contradicts our assumption that f is bounded from below. Therefore (8.53) and assertion (i) are indeed true, as claimed.
We turn now to assertion (ii). Let \(\mathbf{a}=\{a_{i}\}_{i=0}^{\infty} \subset(0,1]\) satisfy (8.6) and let a bounded \(\{x_{i}\}_{i=0}^{\infty} \subset X\) satisfy (8.7) for all integers i≥0. We will show that (8.53) holds. Indeed, assume that (8.53) is not true. Then there exists ε>0 such that (8.54) holds. Since the sequence \(\{x_{i}\}_{1=0}^{\infty}\) is bounded, there exists a number S>0 such that
By Lemma 8.7, there exist numbers α,γ∈(0,1) such that for each x∈X satisfying (8.56) and each β∈(0,α], inequality (8.57) holds. Since a i →0 as i→∞, there exists a natural number i 0 such that
Let i≥i 0 be an integer. Then it follows from (8.58), (8.54), the definitions of α and γ, and (8.59) that
and
Since \(\sum_{i=0}^{\infty}a_{i}=\infty\), we conclude that
The contradiction we have reached shows that (8.53), assertion (ii) and Theorem 8.3 itself are all true. □
Proof of Theorem 8.4
Let
and set
Clearly, E 0 is bounded and closed. Choose
By Lemma 8.7, there exist a neighborhood \(\mathcal{U}\) of V in \(\mathcal{A}\) and numbers α,γ∈(0,1) such that for each \(W \in{ \mathcal{U}}\), each x∈X satisfying
and each β∈(0,α],
Now choose a natural number N 0 which satisfies
First we will show that assertion (i) is true. Assume that \(W \in{ \mathcal{U}}\), \(\{x_{i}\}_{i=0}^{N_{0}} \subset X\),
Our aim is to show that
Assume that (8.67) is not true. Then
By (8.66) and (8.60)–(8.62), we also have
Let i∈{0,…,N 0−1}. It follows from (8.69), (8.68) and the definitions of \(\mathcal{U}\), α and γ (see (8.63) and (8.64)) that
Summing up from i=0 to N 0−1, we conclude that
It follows from this inequality, (8.60), (8.65) and (8.66) that
Since we have reached a contradiction, we see that (8.67) must be true and assertion (i) is proved.
Now we will show that assertion (ii) is also valid. To this end, let a sequence \(\mathbf{a}=\{a_{i}\}_{i=0}^{\infty} \subset(0,1]\) satisfy
Evidently, there exists a natural number N 1 such that
Choose a natural number N>N 1+4 such that
Now assume that \(W \in{ \mathcal{U}}\), \(\{x_{i}\}_{i=0}^{N} \subset X\), ∥x 0∥≤K, and that (8.7) holds for all i=0,…,N−1. We claim that
Assume the contrary. Then
Since ∥x 0∥≤K, we see by (8.7) and (8.60)–(8.62) that
Let i∈{N 1,…,N−1}. It follows from (8.75), (8.74), (8.71) and the definitions of α and γ (see (8.63) and (8.64)) that
This implies that
By this inequality, (8.7), the inequality ∥x 0∥≤K, (8.60) and (8.72), we obtain
The contradiction we have reached proves (8.73) and assertion (ii). This completes the proof of Theorem 8.4. □
8.6 Methods for a Nonconvex Objective Function
Assume that (X,∥⋅∥) is a Banach space, (X ∗,∥⋅∥∗) is its dual space, and f:X→R 1 is a function which is bounded from below and Lipschitzian on bounded subsets of X. Recall that for each pair of sets A,B⊂X ∗,
is the Hausdorff distance between A and B. For each x∈X, let
be the Clarke derivative of f at the point x [41],
the Clarke subdifferential of f at x, and
It is well known that the set ∂f(x) is nonempty and bounded. It should be mentioned that the functional Ξ was introduced in [176] and used in [182] in order to study penalty methods in constrained optimization.
Set inf(f)=inf{f(x):x∈X}. Denote by \({\mathcal{A}}\) the set of all mappings V:X→X such that V is bounded on every bounded subset of X, and for each x∈X, f 0(x,Vx)≤0. We denote by \({\mathcal{A}}_{c}\) the set of all continuous \(V \in{ \mathcal{A}}\) and by \({\mathcal{A}}_{b}\) the set of all \(V \in{ \mathcal{A}}\) which are bounded on X. Finally, let \({\mathcal{A}}_{bc}={\mathcal{A}}_{b} \cap{ \mathcal{A}}_{c}\). Next we endow the set \({\mathcal{A}}\) with two metrics, ρ s and ρ w . To define ρ s , we set, for each \(V_{1},V_{2} \in{ \mathcal{A}}\), \(\tilde{\rho}_{s}(V_{1},V_{2})=\sup\{\|V_{1}x-V_{2}x\|: x \in X\} \) and
(Here we use the convention that ∞/∞=1.) It is clear that \(({\mathcal{A}},\rho_{s})\) is a complete metric space. To define ρ w , we set, for each \(V_{1},V_{2} \in{ \mathcal{A}}\) and each integer i≥1,
Clearly, \(({\mathcal{A}},\rho_{w})\) is a complete metric space. It is also not difficult to see that the collection of the sets
where N,ε>0, is a base for the uniformity generated by the metric ρ w . It is easy to see that ρ w (V 1,V 2)≤ρ s (V 1,V 2) for all \(V_{1}, V_{2} \in{ \mathcal{A}}\). The metric ρ w induces on \(\mathcal{A}\) a topology which is called the weak topology and the metric ρ s induces a topology which is called the strong topology. Clearly, \({\mathcal{A}}_{c}\) is a closed subset of \(\mathcal{A}\) with the weak topology while \({\mathcal{A}}_{b}\) and \({\mathcal{A}}_{bc}\) are closed subsets of \(\mathcal{A}\) with the strong topology. We consider the subspaces \({\mathcal{A}}_{c}\), \({\mathcal{A}}_{b}\) and \({\mathcal{A}}_{bc}\) with the metrics ρ s and ρ w which induce the strong and the weak topologies, respectively.
When the function f is convex, one usually looks for a sequence \(\{x_{i} \}_{i=1}^{\infty}\) which tends to a minimum point of f (if such a point exists) or at least such that lim i→∞ f(x i )=inf(f). If f is not necessarily convex, but X is finite-dimensional, then we expect to construct a sequence which tends to a critical point z of f, namely a point z for which 0∈∂f(z). If f is not necessarily convex and X is infinite-dimensional, then the problem is more difficult and less understood because we cannot guarantee, in general, the existence of a critical point and a convergent subsequence. To partially overcome this difficulty, we have introduced the function Ξ:X→R 1. Evidently, a point z is a critical point of f if and only if Ξ(z)≥0. Therefore we say that z is ε-critical for a given ε>0 if Ξ(z)≥−ε. We look for sequences \(\{x_{i}\}_{i=1}^{\infty}\) such that either lim inf i→∞ Ξ(x i )≥0 or at least lim sup i→∞ Ξ(x i )≥0. In the first case, given ε>0, all the points x i , except possibly a finite number of them, are ε-critical, while in the second case this holds for a subsequence of \(\{x_{i}\}_{i=1}^{\infty}\).
We show, under certain assumptions on f, that for most (in the sense of Baire’s categories) vector fields \(W \in{ \mathcal{A}}\), the iterative processes defined below (see (8.84) and (8.85)) yield sequences with the desirable properties. Moreover, we show that the complement of the set of “good” vector fields is not only of the first category, but also σ-porous. These results, which were obtained in [141], are stated in this section. Their proofs are relegated to subsequent sections.
For each set E⊂X, we denote by \(\operatorname{cl}(E)\) the closure of E in the norm topology. Our results hold for any Banach space and for those functions which satisfy the following two assumptions.
- A(i):
-
For each ε>0, there exists δ∈(0,ε) such that
$$\operatorname{cl}\bigl(\bigl\{ x \in X: \varXi (x) <-\varepsilon \bigr\} \bigr) \subset\bigl\{ x \in X: \varXi (x)<-\delta\bigr\} ; $$ - A(ii):
-
for each r>0, the function f is Lipschitzian on the ball {x∈X:∥x∥≤r}.
We say that a mapping \(V \in{ \mathcal{A}}\) is regular if for any natural number n, there exists a positive number δ(n) such that for each x∈X satisfying ∥x∥≤n and Ξ(x)<−1/n, we have f 0(x,Vx)≤−δ(n).
This concept of regularity is a non-convex analog of the regular vector fields introduced in [136]. We denote by \(\mathcal{F}\) the set of all regular vector fields \(V \in \mathcal{A}\).
Theorem 8.8
Assume that both A(i) and A(ii) hold. Then \({\mathcal{A}} \setminus{ \mathcal{F}}\) (respectively, \({\mathcal{A}}_{c} \setminus {\mathcal{F}}\), \({\mathcal{A}}_{b} \setminus{ \mathcal{F}}\) and \({\mathcal{A}}_{bc} \setminus{ \mathcal{F}}\)) is a σ-porous subset of the space \({\mathcal{A}}\) (respectively, \({\mathcal{A}}_{c}\), \({\mathcal{A}}_{b}\) and \({\mathcal{A}}_{bc}\)) with respect to the pair (ρ w ,ρ s ).
Now let \(W \in{ \mathcal{A}}\). We associate with W two iterative processes. For x∈X we denote by P W (x) the set of all y∈{x+αWx:α∈[0,1]} such that
Given any initial point x 0∈X, one can construct a sequence \(\{x_{i}\}_{i=0}^{\infty} \subset X\) such that for all i=0,1,…,
This is our first iterative process. Next we describe the second iterative process. Given a sequence \(\mathbf{a}=\{a_{i}\}_{i=0}^{\infty} \subset(0,1)\) such that
we construct for each initial point x 0∈X, a sequence \(\{x_{i}\}_{i=0}^{\infty} \subset X\) according to the following rule:
In the sequel we will also make use of the following assumption:
- A(iii):
-
For each integer n≥1, there exists δ>0 such that for each x 1,x 2∈X satisfying ∥x 1∥,∥x 2∥≤n, min{Ξ(x i ):i=1,2}≤−1/n, and ∥x 1−x 2∥≤δ, the following inequality holds: H(∂f(x 1),∂f(x 2))≤1/n.
We denote by \(\operatorname{Card}(B)\) the cardinality of a set B.
Theorem 8.9
Assume that \(W \in \mathcal{A}\) is regular, and that A(i), A(ii) and A(iii) are all valid. Then the following two assertions are true:
-
(i)
Let the sequence \(\{x_{i}\}_{i=0}^{\infty} \subset X\) satisfy (8.83) for all i=0,1,…. If \(\{x_{i}\}_{i=0}^{\infty}\) is bounded, then lim inf i→∞ Ξ(x i )≥0.
-
(ii)
Let a sequence \(\mathbf{a}=\{a_{i}\}_{i=0}^{\infty} \subset(0,1)\) satisfy (8.84) and let the sequence \(\{x_{i}\}_{i=0}^{\infty} \subset X\) satisfy (8.85) for all i=0,1,…. If \(\{x_{i}\}_{i=0}^{\infty}\) is bounded, then
$$\limsup_{i \to\infty} \varXi (x_i)\ge0. $$
Theorem 8.10
Assume that f(x)→∞ as ∥x∥→∞, \(V \in{ \mathcal{A}}\) is regular, and that A(i), A(ii) and A(iii) are all valid. Let K,ε>0 be given. Then there exist a neighborhood \(\mathcal{U} \) of V in \(\mathcal{A}\) with the weak topology and a natural number N 0 such that the following two assertions are true:
-
(i)
For each \(W \in \mathcal{U} \), each integer n≥N 0 and each sequence \(\{x_{i}\}_{i=0}^{n} \subset X\) which satisfies ∥x 0∥≤K and (8.83) for all i=0,…,n−1, we have
$$\operatorname{Card}\bigl\{ i \in\{0,\dots, N-1\}: \varXi (x_i) \le-\varepsilon \bigr\} \le N_0. $$ -
(ii)
For each sequence of numbers \(\mathbf{a}=\{a_{i}\}_{i=0}^{\infty} \subset(0,1)\) satisfying (8.84), there exists a natural number N such that for each \(W \in{ \mathcal{U}}\) and each sequence \(\{x_{i}\}_{i=0}^{N} \subset X\) which satisfies ∥x 0∥≤K and (8.85) for all i=0,…,N−1, we have
$$\max\bigl\{ \varXi (x_i): i=0,\dots,N\bigr\} \ge-\varepsilon . $$
8.7 An Auxiliary Result
For each positive number λ, set
Proposition 8.11
Let ε>0 be given. Suppose that
for some δ(ε)∈(0,ε). Then there exists a locally Lipschitzian vector field \(V \in{ \mathcal{A}}_{b}\) such that f 0(y,Vy)<−δ(ε) for all y∈X satisfying Ξ(y)<−ε.
Proof
It easily follows from definitions (8.76) and (8.78) that E λ is an open set for all λ>0. Let x∈E δ(ε). Then there exist h x ∈X such that ∥h x ∥=1 and f 0(x,h x )<−δ(ε), and (see (8.76)) an open neighborhood U x of x in X such that
For x∈X∖E δ(ε), set
Clearly, {U x } x∈X is an open covering of X. Since any metric space is paracompact, there is a locally finite refinement {Q α :α∈A} of {U x :x∈X}, i.e., an open covering of X such that each x∈X has a neighborhood Q(x) with Q(x)∩Q α ≠∅ only for finitely many α∈A, and such that for each α∈A, there exists x α ∈X with Q α ⊂U(x α ). Let α∈A. Define μ α :X→[0,∞) by μ α (x)=0 if x∉Q α and by μ α (x)=inf{∥x−y∥:y∈∂Q α } otherwise. (Here ∂B is the boundary of a set B⊂X.) The function μ α is clearly Lipschitzian on all of X with Lipschitz constant 1. Let ω α (x)=μ α (x)(∑ β∈A μ β (x))−1, x∈X. Since {Q α :α∈A} is locally finite, each ω α is well defined and locally Lipschitzian on X. Define a locally Lipschitzian, bounded mapping V:X→X by
Let y∈X. There are a neighborhood Q of y in X and α 1,…,α n ∈A such that
We have
Let i∈{1,…,n} with \(\omega_{\alpha_{i}}(y)>0\). Then
If \(x_{\alpha_{i}} \in X \setminus E_{\delta(\varepsilon )}\), then by (8.89), \(h_{x_{\alpha_{i}}}= 0\) and \(f^{0}(y,h_{x_{\alpha_{i}}})=0\). If \(x_{\alpha_{i}} \in E_{\delta(\varepsilon )}\), then by (8.88) and (8.94), \(f^{0}(y,h_{x_{\alpha_{i}}})<0\). Therefore \(f^{0}(y,h_{x_{\alpha_{i}}}) \le0\) in both cases and f 0(y,Vy)≤0. Thus \(V \in{ \mathcal{A}}\). Assume that y∈E ε , i∈{1,…,n} and \(\omega_{\alpha_{i}}(y)>0\). Then (8.94) holds. We assert that \(x_{\alpha_{i}} \in E_{\delta(\varepsilon )}\). Assume the contrary. Then \(x_{\alpha_{i}} \in X \setminus E_{\delta(\varepsilon )}\) and by (8.89), \(U_{x_{\alpha_{i}}}=X \setminus \operatorname{cl}(E_{\varepsilon })\). When combined with (8.94), this implies that \(y \in E_{\varepsilon } \cap U_{x_{\alpha_{i}}}=\) \(E_{\varepsilon } \cap(X \setminus \operatorname{cl}(E_{\varepsilon }))\), a contradiction. Thus \(x_{\alpha_{i}} \in E_{\delta(\varepsilon )}\), as asserted. By the definition of \(U_{x_{\alpha_{i}}}\) (see (8.88)) and (8.94), \(f^{0}(y,h_{x_{\alpha_{i}}})<-\delta(\varepsilon )\). When combined with (8.93), this implies that f 0(y,Vy)<−δ(ε). □
8.8 Proof of Theorem 8.8
For each pair of integers m,n≥1, denote by Ω mn the set of all \(V \in{ \mathcal{A}}\) such that
Clearly,
Therefore in order to prove Theorem 8.8 it is sufficient to show that for each pair of integers m,n≥1, the set Ω mn (respectively, \(\varOmega _{mn}\cap{ \mathcal{A}}_{c}\), \(\varOmega _{mn} \cap{ \mathcal{A}}_{b}\), \(\varOmega _{mn} \cap{ \mathcal{A}}_{bc}\)) is a porous subset of \({\mathcal{A}}\) (respectively, \({\mathcal{A}}_{c}\), \({\mathcal{A}}_{b}\), \({\mathcal{A}}_{bc}\)) with respect to the pair (ρ w ,ρ s ). Let m,n≥1 be integers. By Proposition 8.11, there exists a vector field \(V_{*} \in \mathcal{A}\) such that (i) V ∗ is bounded on X and V ∗ is locally Lipschitzian on X; (ii) there exists δ ∗∈(0,1) such that
By assumption A(ii), there is c 0>1 such that
for all x,y∈X satisfying ∥x∥,∥y∥≤n+2. Choose α∈(0,1) such that
Assume that \(V \in{ \mathcal{A}}\) and r∈(0,1]. There are two cases: (a) sup{∥Vx∥:x∈X,∥x∥≤n+1}≤m+1; (b) sup{∥Vx∥:x∈X,∥x∥≤n+1}>m+1. We first assume that (b) holds. Let \(W \in{ \mathcal{A}}\) with ρ w (W,V)≤2−n−4. Then ρ n+1(W,V)(1+ρ n+1(V,W))−1≤8−1, ρ n+1(W,V)≤1/7, and sup{∥Wx∥:x∈X,∥x∥≤n+1}>m. Thus \(\{W \in{ \mathcal{A}}: \rho_{w}(W,V) \le2^{-n-4}\}\cap \varOmega _{mn}=\emptyset\). Assume now that (a) holds. Let
and define \(V_{\gamma} \in{ \mathcal{A}}\) by
If \(V \in{ \mathcal{A}}_{c}\) (respectively, \({\mathcal{A}}_{b}\), \({\mathcal{A}}_{bc}\)), then \(V_{\gamma} \in {\mathcal{A}}_{c}\) (respectively, \({\mathcal{A}}_{b}\), \({\mathcal{A}}_{bc}\)). Next we estimate the distance ρ s (V γ ,V). It follows from (8.102), (8.101) and (8.76) that
Assume that \(W \in{ \mathcal{A}}\) with
Assume now that
Inequality (8.99) implies that
By (8.102), (8.98) and (8.106),
It follows from (8.106) and (8.80) that
By (8.104) and (8.81), we have 2−n ρ n (W,V γ )(1+ρ n (W,V γ ))−1≤ ρ w (W,V γ )≤αr, ρ n (W,V γ )(1+ρ n (W,V γ ))−1≤2n αr, and ρ n (W,V γ )(1−2n αr)≤2n αr. When combined with (8.109), the last inequality implies that ∥Wx−V γ x∥≤2n αr(1−2n αr)−1, and when combined with (8.107), this implies that
By (8.110), (8.108), (8.101) and (8.100),
Since l is an arbitrary element of ∂f(x), we conclude that f 0(x,Wx)≤−2rc 02n α. Thus \(\{W \in{ \mathcal{A}}: \rho_{w}(W,V_{\gamma})\le\alpha r\} \cap \varOmega _{mn}=\emptyset\). Recall that in case (b), \(\{W \in{ \mathcal{A}}: \rho_{w}(W,V) \le2^{-n-4}\} \cap \varOmega _{mn}=\emptyset\). Therefore Ω mn is porous in \({\mathcal{A}}\), \(\varOmega _{mn} \cap{ \mathcal{A}}_{c}\) is porous in \({\mathcal{A}}_{c}\), \(\varOmega _{mn} \cap{ \mathcal{A}}_{b}\) is porous in \({\mathcal{A}}_{b}\), and \(\varOmega _{mn} \cap{ \mathcal{A}}_{bc}\) is porous in \({\mathcal{A}}_{bc}\), as asserted.
8.9 A Basic Lemma for Theorems 8.9 and 8.10
Lemma 8.12
Assume that \(V \in{ \mathcal{A}}\) is regular, and that A(i), A(ii) and A(iii) are all valid. Let \(\bar{K}\) and \(\bar{\varepsilon }\) be positive. Then there exist a neighborhood \({\mathcal{U}}\) of V in \({\mathcal{A}}\) with the weak topology and positive numbers \(\bar{\alpha}\) and γ such that for each \(W \in{ \mathcal{U}}\), each x∈X satisfying
and each \(\beta\in(0, \bar{\alpha}]\), we have
Proof
There exists \(K_{0} >\bar{K}+1\) such that
By Assumption A(ii), there exists a constant L 0>4 such that
for all x 1,x 2∈X satisfying ∥x 1∥,∥x 2∥≤2K 0+4. There is δ 0∈(0,1) such that
for each y∈X satisfying ∥y∥≤K 0+4 and \(\varXi (y) \le-\bar{\varepsilon }/4\). Choose δ 1∈(0,1) such that
By A(iii), there is a positive \(\bar{\alpha}\) such that the following conditions hold:
for each x 1,x 2∈X satisfying
the following inequality is true:
Next, choose a positive number δ 2 such that
Finally, choose a positive number γ and define a neighborhood \(\mathcal{U}\) such that
Assume that \(W \in{ \mathcal{U}}\), x∈X satisfies (8.111), and that \(\beta\in(0,\bar{\alpha}]\). We intend to show that (8.112)) holds. To this end, we first note that (8.111), (8.113), (8.117) and (8.122) yield
By these inequalities, the definition of L 0 (see (8.114)) and (8.122),
Next we estimate f(x)−f(x+βVx). By [89], there exist θ∈[0,β] and l∈∂f(x+θVx) such that
By (8.111), (8.114) and (8.117),
Note that (8.126) and the definition of L 0 (see (8.114)) imply that
It also follows from (8.111), (8.126) and the definition of \(\bar{\alpha}\) (see (8.118) and (8.119)) that H(∂f(x),∂f(x+θVx))<δ 1. Therefore there exists \(\bar{l} \in\partial f(x)\) such that \(\|\bar{l}-l\|_{*} \le \delta_{1}\). When combined with (8.125) and (8.126), this fact implies that
It follows from the definition of δ 0 (see (8.115)) and (8.111) that \(\beta\bar{l}(Vx) \le-\beta\delta_{0}\). Combining this inequality with (8.128) and (8.116), we see that f(x+βVx)−f(x)≤−βδ 0+βδ 1 K 0≤−βδ 0/2. It now follows from this inequality, (8.120), (8.124) and (8.121) that f(x+βWx)−f(x)≤ f(x+βVx)−f(x)+f(x+βWx)−f(x+βVx)≤−βδ 0/2+L 0 βδ 2≤−βδ 0/4≤−γβ. Thus (8.112) holds and Lemma 8.12 is proved. □
8.10 Proofs of Theorems 8.9 and 8.10
Proof of Theorem 8.9
To show that assertion (i) holds, suppose that
We claim that
Assume the contrary. Then there exist ε>0 and a strictly increasing sequence of natural numbers \(\{i_{k}\}_{k=1}^{\infty}\) such that
Choose a number S>0 such that
By Lemma 8.12, there exist numbers α,γ∈(0,1) such that for each x∈X satisfying
and each β∈(0,α], we have
It follows from (8.129), (8.82), (8.83), the definitions of α and γ, (8.132) and (8.131) that for each integer k≥1, \(f(x_{i_{k}})-f(x_{i_{k}+1}) \ge f(x_{i_{k}})-f(x_{i_{k}}+\alpha Wx_{i_{k}}) \ge \gamma\alpha\). Since this inequality holds for all integers k≥1, we conclude that lim n→∞(f(x 0)−f(x n ))=∞. This contradicts our assumption that f is bounded from below. Therefore (8.130) and assertion (i) are indeed true, as claimed.
We turn now to assertion (ii). Let \(\mathbf{a}=\{a_{i}\}_{i=0}^{\infty} \subset(0,1)\) satisfy (8.84) and let a bounded \(\{x_{i}\}_{i=0}^{\infty} \subset X\) satisfy (8.85) for all integers i≥0. We will show that
Indeed, assume that (8.135) is not true. Then there exist ε>0 and an integer i 1≥0 such that
Since the sequence \(\{x_{i}\}_{1=0}^{\infty}\) is bounded, there exists a number S>0 such that
By Lemma 8.12, there exist numbers α,γ∈(0,1) such that for each x∈X satisfying (8.133) and each β∈(0,α], inequality (8.134) holds. Since a i →0 as i→∞, there exists a natural number i 0≥i 1 such that
Let i≥i 0 be an integer. Then it follows from (8.137), (8.136), the definitions of α and γ, and (8.138) that f(x i )−f(x i +a i Wx i )≥γa i , x i+1=x i +a i Wx i , and f(x i )−f(x i+1)≥γa i . Since \(\sum_{i=0}^{\infty}a_{i}=\infty\), we conclude that lim n→∞(f(x 0)−f(x n ))=∞. The contradiction we have reached shows that (8.135), assertion (ii) and Theorem 8.9 itself are all true. □
Proof of Theorem 8.10
Let
It is clear that E 0 is bounded and closed. Choose
By Lemma 8.12, there exist a neighborhood \(\mathcal{U}\) of V in \(\mathcal{A}\) and numbers α,γ∈(0,1) such that for each \(W \in{ \mathcal{U}}\), each x∈X satisfying
and each β∈(0,α],
Now choose a natural number N 0 which satisfies
Let \(W \in{ \mathcal{U}}\), \(\{x_{i}\}_{i=0}^{n} \subset X\), where the integer n≥N 0,
By (8.145) and (8.139)–(8.141), we have
Let i∈B. It follows from (8.147), (8.146) and the definitions of \({\mathcal{U}}\), α and γ (see (8.142) and (8.143)) that f(x i )−f(x i+1)≥f(x i )−f(x i +αWx i )≥γα. Summing up from i=0 to n−1, we conclude that
It follows from this inequality, (8.139), (8.145) and (8.144) that
Thus we see that assertion (i) is proved.
To prove assertion (ii), let a sequence \(\mathbf{a}=\{a_{i}\}_{i=0}^{\infty} \subset(0,1)\) satisfy
Clearly, there exists a natural number N 1 such that
Choose a natural number N>N 1+4 such that
Now assume that \(W \in{ \mathcal{U}}\), \(\{x_{i}\}_{i=0}^{N} \subset X\), ∥x 0∥≤K, and that (8.85) holds for all i=0,…,N−1. We will show that
Assume the contrary. Then
Since ∥x 0∥≤K, we see by (8.85) and (8.139)–(8.141) that
Let i∈{N 1,…,N−1}. It follows from (8.153), (8.152), (8.149) and the definitions of α and γ (see (8.142)) and (8.143)) that
This implies that
By this inequality, (8.85), the inequality ∥x 0∥≤K, (8.139) and (8.150), we obtain that
The contradiction we have reached proves (8.151) and assertion (ii). □
8.11 Continuous Descent Methods
Let (X ∗,∥⋅∥∗) be the dual space of the Banach space (X,∥⋅∥), and let f:X→R 1 be a convex continuous function which is bounded from below. Recall that for each pair of sets A,B⊂X ∗,
is the Hausdorff distance between A and B.
For each x∈X, let
be the subdifferential of f at x. It is well known that the set ∂f(x) is nonempty and norm-bounded. Set
Denote by \(\mathcal{A}\) the set of all mappings V:X→X such that V is bounded on every bounded subset of X (that is, for each K 0>0, there is K 1>0 such that ∥Vx∥≤K 1 if ∥x∥≤K 0), and for each x∈X and each l∈∂f(x), l(Vx)≤0. We denote by \({\mathcal{A}}_{c}\) the set of all continuous \(V \in \mathcal{A}\), by \({\mathcal{A}}_{u}\) the set of all \(V \in \mathcal{A}\) which are uniformly continuous on each bounded subset of X, and by \({\mathcal{A}}_{au}\) the set of all \(V \in{ \mathcal{A}}\) which are uniformly continuous on the subsets
for each integer n≥1. Finally, let \({\mathcal{A}}_{auc}={\mathcal{A}}_{au} \cap{ \mathcal{A}}_{c}\).
Our results are valid in any Banach space and for those convex functions which satisfy the following two assumptions.
- A(i):
-
There exists a bounded set X 0⊂X such that
$$\inf(f)=\inf\bigl\{ f(x): x \in X\bigr\} =\inf\bigl\{ f(x): x \in X_0 \bigr\} ; $$ - A(ii):
-
for each r>0, the function f is Lipschitzian on the ball {x∈X:∥x∥≤r}.
Note that assumption A(i) clearly holds if lim∥x∥→∞ f(x)=∞.
We recall that a mapping \(V \in{ \mathcal{A}}\) is regular if for any natural number n, there exists a positive number δ(n) such that for each x∈X satisfying
and for each l∈∂f(x), we have
Denote by \({\mathcal{F}}\) the set of all regular vector fields \(V \in{ \mathcal{A}}\).
Let T>0, x 0∈X and let u:[0,T]→X be a Bochner integrable function. Set
Then x:[0,T]→X is differentiable and x′(t)=u(t) for almost every t∈[0,T]. Recall that the function f:X→R 1 is assumed to be convex and continuous, and therefore it is, in fact, locally Lipschitzian. It follows that its restriction to the set {x(t):t∈[0,T]} is Lipschitzian. Indeed, since the set {x(t):t∈[0,T]} is compact, the closure of its convex hull C is both compact and convex, and so the restriction of f to C is Lipschitzian. Hence the function (f⋅x)(t):=f(x(t)), t∈[0,T], is absolutely continuous. It follows that for almost every t∈[0,T], both the derivatives x′(t) and (f⋅x)′(t) exist:
We continue with the following fact.
Proposition 8.13
Assume that t∈[0,T] and that both the derivatives x′(t) and (f⋅x)′(t) exist. Then
Proof
There exist a neighborhood \(\mathcal{U}\) of x(t) in X and a constant L>0 such that
Let ε>0 be given. There exists δ>0 such that
and such that for each h∈[(−δ,δ)∖{0}]∩[−t,T−t],
Let
It follows from (8.156), (8.155) and (8.157) that
Clearly,
Relations (8.159) and (8.160) imply that
Since ε is an arbitrary positive number, we conclude that (8.154) holds. □
Assume now that \(V \in{ \mathcal{A}}\) and that the differentiable function x:[0,T]→X satisfies
Then by Proposition 8.13, (f⋅x)′(t)≤0 for a.e. t∈[0,T], and f(x(t)) is decreasing on [0,T].
In the sequel we denote by μ(E) the Lebesgue measure of E⊂R 1.
In the next two sections, we prove the following two results which were obtained in [148].
Theorem 8.14
Let \(V \in{ \mathcal{A}}\) be regular, let x:[0,∞)→X be differentiable and suppose that
Assume that there exists a positive number r such that
Then lim t→∞ f(x(t))=inf(f).
Theorem 8.15
Let \(V \in{ \mathcal{A}}\) be regular, let f be Lipschitzian on bounded subsets of X, and assume that lim∥x∥→∞ f(x)=∞. Let K 0 and ε>0 be positive. Then there exist N 0>0 and δ>0 such that for each T≥N 0 and each differentiable mapping x:[0,T]→X satisfying
the following inequality holds for all t∈[N 0,T]:
8.12 Proof of Theorem 8.14
Assume the contrary. Since f(x(t)) is decreasing on [0,∞), this means that there exists ε>0 such that
Then by Proposition 8.13 and (8.162), we have for each T>0,
where
Since V is regular, there exists δ>0 such that for each x∈X satisfying
and each l∈∂f(x), we have
It follows from (8.165), (8.166), (8.164), the definition of δ (see (8.167) and (8.168)) and (8.163) that for each T>0,
as T→∞, a contradiction. The contradiction we have reached proves Theorem 8.14.
8.13 Proof of Theorem 8.15
We may assume without loss of generality that ε<1/2. Choose
The set
is bounded. Therefore there exists
such that
There exists a number K 3>K 2+1 such that
There exists a number L 0>0 such that
for each x 1,x 2∈X satisfying
Fix an integer
There exists a positive number δ(n)<1 such that:
- (P1):
-
for each x∈X satisfying
$$\|x\| \le n \quad\hbox{and}\quad f(x) \ge\inf(f)+1/n, $$and each l∈∂f(x), we have
$$l(Vx) \le-\delta(n). $$
Choose a natural number N 0>8 such that
and a positive number δ which satisfies
Let T≥N 0 and let x:[0,T]→X be a differentiable function such that
and
We claim that
Assume the contrary. Then there exists t 0∈(0,min{2N 0,T}] such that
It follows from Proposition 8.13, the convexity of directional derivatives, the inequality f 0(x(t),Vx(t))≤0, which holds for all t∈[0,T], (8.181), the definition of L 0 (see (8.173), (8.174) and (8.179)) that
Thus by (8.177),
Since ∥x(0)∥≤K 2 (see (8.178)) and ∥x(t 0)∥=K 3, the inequality just obtained contradicts (8.172). The contradiction we have reached proves (8.180).
We now claim that there exists a number
such that
Assume the contrary. Then
It follows from (8.184), Property (P1) and (8.175) that
By (8.185), (8.184), (8.179), (8.177), the convexity of the directional derivatives of f, and the definition of L 0 (see (8.173) and (8.174)), we have, for almost every t∈[1,N 0],
It follows from the convexity of the directional derivatives of f, the inclusion \(V \in{ \mathcal{A}}\), (8.179), (8.180) and the definition of L 0 (see (8.173) and (8.174)), that for almost every t∈[0,1],
Inequalities (8.178), (8.186) and (8.187) imply that
This contradicts (8.176). The contradiction we have reached yields the existence of a point t 0 which satisfies both (8.182) and (8.183). Clearly, ∥x(t 0)|≤K 2. Having established (8.180) and the existence of such a point t 0 for an arbitrary mapping x satisfying both (8.178) and (8.179), we now consider the mapping x 0(t)=x(t+t 0), t∈[0,T−t 0]. Evidently, (8.178) and (8.179) hold true with x replaced by x 0 and T replaced by T−t 0. Hence, if T−t 0≥N 0, then we have
and there exists
for which
Repeating this procedure, we obtain by induction a finite sequence of points \(\{t_{i}\}_{i=0}^{q}\) such that
Let i∈{0,…,q}, t≤T, and 0<t−t i ≤N 0. Then by Proposition 8.13, the convexity of the directional derivative of f, the inclusion \(V \in{ \mathcal{A}}\), the definition of L 0 (see (8.173) and (8.174)), (7.177) and (8.179), we have
and hence
This completes the proof of Theorem 8.15.
8.14 Regular Vector-Fields
In the previous sections of this chapter, given a continuous convex function f on a Banach space X, we associate with f a complete metric space \({\mathcal{A}}\) of mappings V:X→X such that f 0(x,Vx)≤0 for all x∈X. Here f 0(x,u) is the right-hand derivative of f at x in the direction of u∈X. We call such mappings descent vector-fields (with respect to f). We identified a regularity property of such vector-fields and showed that regular vector-fields generate convergent discrete descent methods. This has turned out to be true for continuous descent methods as well. Such results are significant because most of the elements in \({\mathcal{A}}\) are, in fact, regular. Here by “most” we mean an everywhere dense G δ subset of \(\mathcal{A}\). Thus it is important to know when a given descent vector-field V:X→X is regular. In [163] we established necessary and sufficient conditions for regularity: see Theorems 8.18–8.21 below.
More precisely, let (X,∥⋅∥) be a Banach space and let (X ∗,∥⋅∥∗) be its dual.
For each h:X→R 1, set inf(h)={h(z):z∈X}.
Let U be a nonempty, open subset of X and let f:U→R 1 be a locally Lipschitzian function.
For each x∈U, let
be the Clarke derivative of f at the point x, and let
be the Clarke subdifferential of f at x.
For each x∈U, set
Clearly, Ξ f (x)≤0 for all x∈X and Ξ f (x)=0 if and only if 0∈∂f(x).
For each x∈U, set
Let x∈U. Clearly, \(\tilde{\varXi }_{f}(x)\ge \varXi _{f}(x)\) and 0∈∂f(x) if and only if \(\tilde{\varXi }_{f}(x) \ge0\).
In the next section we prove the following two propositions.
Proposition 8.16
Let x∈U. If \(\tilde{\varXi }_{f}(x) \ge0\), then Ξ f (x)=0. If \(\tilde{\varXi }_{f}(x)<0\), then \(\varXi _{f}(x)= \tilde{\varXi }_{f}(x)\).
Proposition 8.17
For each x∈U,
Assume now that f:X→R 1 is a continuous and convex function which is bounded from below. It is known that f is locally Lipschitzian. It is also known (see Chap. 2, Sect. 2 of [41]) that in this case
Recall that a mapping V:X→X is called regular if V is bounded on every bounded subset of X, f 0(x,Vx)≤0 for all x∈X, and if for any natural number n, there exists a positive number δ(n) such that for each x∈X satisfying ∥x∥≤n and f(x)≥inf(f)+1/n, we have
We now present four results which were established in [163]. Their proofs are given in subsequent sections.
Theorem 8.18
Let f:X→R 1 be a convex and continuous function which is bounded from below, let \(\bar{x} \in X\) satisfy
and let the following property hold:
- (P1):
-
for every sequence \(\{y_{i}\}_{i=1}^{\infty} \subset X\) satisfying \(\lim_{i \to\infty}f(y_{i})=f(\bar{x})\), \(\lim_{i \to\infty}y_{i} =\bar{x}\) in the norm topology.
For each natural number n, let ϕ n :[0,∞)→[0,∞) be an increasing function such that ϕ n (0)=0 and the following property holds:
- (P2):
-
for each ε>0, there exists δ:=δ(ε,n)>0 such that for each t≥0 satisfying ϕ n (t)≤δ, the inequality t≤ε holds.
If V:X→X is bounded on bounded subsets of X,
and if for each natural number n and each x∈X satisfying ∥x∥≤n, we have
then V is regular.
Theorem 8.19
Assume that f:X→R 1 is a convex and continuous function, \(\bar{x} \in X\),
property (P1) holds and the following property also holds:
- (P3):
-
if \(\{x_{i}\}_{i=1}^{\infty} \subset X\) converges to \(\bar{x}\) in the norm topology, then
$$\lim_{i \to\infty} \varXi _f(x_i)=0. $$
Assume that V:X→X is regular and let n≥1 be an integer. Then there exists an increasing function ϕ n :[0,∞)→[0,∞) such that ϕ n (0)=0, property (P2) holds, and for each x∈X satisfying ∥x∥≤n, we have
Assume now that f:X→R 1 is merely locally Lipschitzian. Recall that in this case a mapping V:X→X is called regular if V is bounded on every bounded subset of X,
and for any natural number n, there exists δ(n)>0 such that for each x∈X satisfying ∥x∥≤n and Ξ f (x)≤−1/n, we have f 0(x,Vx)≤−δ(n).
Theorem 8.20
Let f:X→R 1 be a locally Lipschitzian function. For each natural number n, let ϕ n :[0,∞)→[0,∞) be an increasing function such that ϕ n (0)=0 and property (P2) holds.
Assume that V:X→X is bounded on every bounded subset of X,
and for each natural number n and each x∈X satisfying ∥x∥≤n, we have
Then V is regular.
Theorem 8.21
Assume that the function f:X→R 1 is locally Lipschitzian and that V:X→X is regular.
Then for each natural number n, there exists an increasing function ϕ n :[0,∞)→[0,∞) such that (P2) holds and for each natural number n and each x∈X satisfying ∥x∥≤n, (8.197) holds.
8.15 Proofs of Propositions 8.16 and 8.17
Proof of Proposition 8.16
Assume that \(\tilde{\varXi }_{f}(x)\ge0\). Then 0∈∂f(x) and Ξ f (x)=0. Assume that \(\tilde{\varXi }_{f}(x) <0\). Then by definition (see (8.191)),
By (8.198) and the homogeneity of f 0(x,⋅),
By (8.198), (8.191), (8.190) and (8.199),
This implies that
as claimed. Proposition 8.16 is proved. □
We precede the proof of Proposition 8.17 with the following lemma.
Lemma 8.22
Let x∈U and c>0 be given. Then the following statements are equivalent:
-
(i)
Ξ f (x)≥−c;
-
(ii)
\(\tilde{\varXi }_{f}(x) \ge-c\);
-
(iii)
there is l∈∂f(x) such that ∥l∥∗≤c.
Proof
By Proposition 8.16,
It follows from (8.191) that \(\tilde{\varXi }_{f}(x)\ge-c\) if and only if
which is, in its turn, equivalent to the following relation:
Rewriting this last inequality as
we see that it is equivalent to the inclusion
Thus we have proved that (ii) is equivalent to (iii). This completes the proof of Lemma 8.22. □
Proof of Proposition 8.17
Clearly, equality (8.192) holds if either one of its sides equals zero. Therefore we only need to prove (8.192) in the case where
Assume that (8.200) holds. By Lemma 8.22, there is \(\bar{l}\) such that
Hence
Let ε be any positive number. There is l ε ∈∂f(x) such that
By (8.203) and Lemma 8.22,
Since ε is any positive number, we conclude that
When combined with (8.202), this inequality completes the proof of Proposition 8.17. □
8.16 An Auxiliary Result
Proposition 8.23
Let g:X→R 1 be a convex and continuous function, \(\bar{x} \in X\),
and let the following property hold:
- (P4):
-
for any sequence \(\{y_{i}\}_{i=1}^{\infty} \subset X\) satisfying \(\lim_{i \to\infty}g(y_{i})=g(\bar{x})\), we have \(\lim_{i \to \infty}\|y_{i}-\bar{x}\|=0\).
Assume that \(\{x_{i}\}_{i=1}^{\infty} \subset X\),
Then \(\lim_{i \to \infty}\|x_{i}-\bar{x}\|=0\).
Proof
By (8.205) and Proposition 8.17, there exists a sequence \(\{l_{i}\}_{i=1}^{\infty} \subset X^{*}\) such that
Choose a number M>0 such that
and let i≥1 be an integer. By (8.206),
It follows from (8.208), (8.207) and (8.206) that
and therefore
Together with (P4) this implies that \(\lim_{i \to\infty}\|x_{i}-\bar{x}\|=0\). Proposition 8.23 is proved. □
8.17 Proof of Theorem 8.18
To show that V is regular, let n be a natural number. We have to find a positive number δ=δ(n) such that for each x∈X satisfying ∥x∥≤n and f(x)≥inf(f)+1/n,
Assume the contrary. Then for each natural number k, there exists x k ∈X satisfying
and
It follows from (8.210), (8.209) and (8.195) that for each natural number k,
and hence ϕ n (−Ξ f (x k ))<k −1.
Together with (P2) this inequality implies that lim k→∞ Ξ f (x k )=0. When combined with Proposition 8.23 and (8.209), this implies \(\lim_{k \to\infty}\|x_{k}-\bar{x}\|=0\). Since f is continuous,
This, however, contradicts (8.209). The contradiction we have reached proves that V is indeed regular, as asserted.
8.18 Proof of Theorem 8.19
In what follows we make the convention that the infimum over the empty set is infinity. Set ϕ n (0)=0 and let t>0. Put
Clearly, ϕ n :[0,∞)→[0,1] is well defined and increasing.
We show that for each x∈X satisfying ∥x∥≤n,
Let x∈X with ∥x∥≤n. If Ξ f (x)=0, then it is obvious that (8.212) holds.
Assume now that
Then by (8.211)), (8.213) and the inequality ∥x∥≤n,
and hence
Thus (8.212) holds for each x∈X satisfying ∥x∥≤n.
Next we show that (P2) holds. To this end, let ε>0 be given. We claim that there is δ>0 such that for each t≥0 satisfying ϕ n (t)≤δ, the inequality t≤ε holds.
Assume the contrary. Then for each natural number i, there exists t i ≥0 such that
By (8.214) and (8.211), for each natural number i, there exists a point x i ∈X such that
and
Now it follows from (8.215), (8.216) and the definition of regularity that
Together with (P1) this implies that \(\lim_{i \to\infty}\|x_{i}-\bar{x}\|=0\). When combined with (P3), this inequality implies that lim i→∞ Ξ f (x i )=0. This, however, contradicts (8.215). The contradiction we have reached proves Theorem 8.19.
8.19 Proof of Theorem 8.20
Let n be a given natural number. We need to show that there exists δ>0 such that for each x∈X satisfying
we have
Assume the contrary. Then for each natural number k, there exists x k ∈X such that
and
and
It now follows from (8.219) and property (P2) that
and
The last equality contradicts (8.218) and this contradiction proves Theorem 8.20.
8.20 Proof of Theorem 8.21
Set ϕ n (0)=0 and let t>0. Define
Clearly, ϕ:[0,∞)→[0,1] is well defined and increasing.
We show that for each x∈X satisfying ∥x∥≤n,
Consider x∈X with
If Ξ f (x)=0, then (8.221) clearly holds. Assume that
Then by (8.220), (8.221), (8.222) and (8.223),
and hence (8.221) holds for all x∈X satisfying ∥x∥≤n, as claimed.
Now we show that property (P2) also holds. To this end, let ε be positive.
We claim that there is δ>0 such that for each t≥0 satisfying ϕ n (t)≤δ, the inequality t≤ε holds.
Assume the contrary. Then for each natural number i, there exists t i ≥0 such that
Let i be a natural number. By (8.224) and (8.220), there exists x i ∈X such that
and
Clearly,
Choose a natural number p such that
Since V is regular, there is δ>0 such that
Choose a natural number j such that
Then for all integers i≥j, it follows from (8.225) and (8.227) that
Together with (8.228) and (8.229), this implies that for all integers i≥j,
Since this contradicts (8.226), the proof of Theorem 8.21 is complete.
8.21 Most Continuous Descent Methods Converge
Let (X,∥⋅∥) be a Banach space and let f:X→R 1 be a convex continuous function which satisfies the following conditions:
- C(i):
-
lim∥x∥→∞ f(x)=∞;
- C(ii):
-
there is \(\bar{x} \in X\) such that \(f(\bar{x}) \le f(x)\) for all x∈X;
- C(iii):
-
if \(\{x_{n}\}_{n=1}^{\infty} \subset X\) and \(\lim_{n \to\infty} f(x_{n})=f(\bar{x})\), then
$$\lim_{n \to\infty}\|x_n-\bar{x} \|=0. $$
By C(iii), the point \(\bar{x}\), where the minimum of f is attained, is unique.
For each x∈X, let
Let (X ∗,∥⋅∥∗) be the dual space of (X,∥⋅∥).
For each x∈X, let
be the subdifferential of f at x. It is well known that the set ∂f(x) is nonempty and norm-bounded.
For each x∈X and r>0, set
For each mapping A:X→X and each r>0, put
Denote by \({\mathcal{A}}_{l}\) the set of all mappings V:X→X such that \(\operatorname{Lip}(V,r)<\infty\) for each positive r (this means that the restriction of V to any bounded subset of X is Lipschitzian) and f 0(x,Vx)≤0 for all x∈X.
For the set \({\mathcal{A}}_{l}\) we consider the uniformity determined by the base
Clearly, this uniform space \({\mathcal{A}}_{l}\) is metrizable and complete. The topology induced by this uniformity in \({\mathcal{A}}_{l}\) will be called the strong topology.
We also equip the space \({\mathcal{A}}_{l}\) with the uniformity determined by the base
where n,ε>0. The topology induced by this uniformity will be called the weak topology.
The following existence result is proved in the next section.
Proposition 8.24
Let x 0∈X and \(V \in{ \mathcal{A}}_{l}\). Then there exists a unique continuously differentiable mapping x:[0,∞)→X such that
In the subsequent sections we prove the following result which was obtained in [1].
Theorem 8.25
There exists a set \({\mathcal{F}} \subset {\mathcal{A}}_{l}\) which is a countable intersection of open (in the weak topology) everywhere dense (in the strong topology) subsets of \({\mathcal{A}}_{l}\) such that for each \(V \in{ \mathcal{F}}\), the following property holds:
For each ε>0 and each n>0, there exist T εn >0 and a neighborhood \({\mathcal{U}}\) of V in \({\mathcal{A}}_{l}\) with the weak topology such that for each \(W \in{ \mathcal{U}}\) and each differentiable mapping y:[0,∞)→X satisfying
the inequality \(\|y(t)-\bar{x}\| \le \varepsilon \) holds for all t≥T εn .
8.22 Proof of Proposition 8.24
Since V is locally Lipschitzian, there exists a unique differentiable function x:I→X, where I is an interval of the form [0,b), b>0, such that
We may and will assume that I is the maximal interval of this form on which the solution exists.
We need to show that b=∞. Suppose, by contradiction, that b<∞.
By Proposition 8.13 and the relation \(V \in{ \mathcal{A}}_{l}\), the function f(x(t)) is decreasing on I. By C(i), the set {x(t):t∈[0,b)} is bounded. Thus there is K 0>0 such that
Since V is Lipschitzian on bounded subsets of X, there is K 1>0 such that
Let ε>0 be given. Then it follows from (8.235), (8.236) and (8.237) that for each t 1,t 2∈[0,b) such that 0<t 2−t 1<ε/K 1,
Hence there exists \(z_{0}=\lim_{t \to b^{-}}x(t)\) in the norm topology. It follows that there exists a unique solution of the initial value problem
defined on a neighborhood of b, and this implies that our solution x(⋅) can be extended to an open interval larger than I. The contradiction we have reached completes the proof of Proposition 8.24.
8.23 Proof of Theorem 8.25
For each \(V \in{ \mathcal{A}}_{l}\) and each γ∈(0,1), set
We first prove several lemmata.
Lemma 8.26
Let \(V \in{ \mathcal{A}}_{l}\) and γ∈(0,1). Then \(V_{\gamma} \in{ \mathcal{A}}_{l}\).
Proof
Clearly, V γ is Lipschitzian on any bounded subset of X. Let x∈X. Then by (8.238), the subadditivity and positive homogeneity of the directional derivative of a convex function, the relation \(V \in{ \mathcal{A}}_{l}\), and C(ii),
This completes the proof of Lemma 8.26. □
It is easy to see that the following lemma also holds.
Lemma 8.27
Let \(V \in{ \mathcal{A}}_{l}\). Then \(\lim_{\gamma \to0^{+}}V_{\gamma}=V\) in the strong topology.
Lemma 8.28
Let \(V \in{ \mathcal{A}}_{l}\), γ∈(0,1), ε>0, and let x∈X satisfy \(f(x) \ge f(\bar{x})+\varepsilon \). Then f 0(x,V γ x)≤−γε.
Proof
It follows from (8.238), the properties of the directional derivative of a convex function, and the relation \(V \in {\mathcal{A}}_{l}\) that
The lemma is proved. □
Lemma 8.29
Let \(V \in{ \mathcal{A}}_{l}\), γ∈(0,1), and let x∈C 1([0,∞);X) satisfy
Assume that T 0,ε>0 are such that
Then for each t≥T 0, \(f(x(t)) \le f(\bar{x})+\varepsilon \).
Proof
Since the function f(x(⋅)) is decreasing on [0,∞) (see Proposition 8.13, Lemma 8.26 and (8.239)), it is sufficient to show that
Assume the contrary. Then \(f(x(T_{0}))>f(\bar{x})+\varepsilon \), and since f(x(⋅)) is decreasing on [0,∞), we have
When combined with Lemma 8.28, inequality (8.242) implies that
It now follows from Proposition 8.13, (8.239) and (8.243) that
whence
This contradicts (8.240). The contradiction we have reached proves the lemma. □
Lemma 8.30
Let \(V \in{ \mathcal{A}}_{l}\), γ∈(0,1), ε>0 and n>0. Then there exist a neighborhood \(\mathcal{U}\) of V γ in \({\mathcal{A}}_{l}\) with the weak topology and τ>0 such that for each \(W \in{ \mathcal{U}}\) and each continuously differentiable mapping x:[0,∞)→X satisfying
and
the following inequality holds:
Proof
By C(i), there is n 1>n such that
By C(iii), there is δ 1>0 such that
Since f is continuous, there is ε 1>0 such that
In view of C(iii), there exists δ 0∈(0,1) such that
Since \(V_{\gamma} \in{ \mathcal{A}}_{l}\), there is L>0 such that
Fix
and choose a positive number Δ such that
Set
Assume that
and that x∈C 1([0,∞);X) satisfies (8.244) and (8.245). We have to prove (8.246). In view of (8.248), it is sufficient to show that
Since the function f(x(⋅)) is decreasing on [0,∞), in order to prove the lemma we only need to show that
By (8.249), this inequality will follow from the inequality
We now prove (8.256).
To this end, consider a continuously differentiable mapping y:[0,∞)→X which satisfies
and
Since the functions f(x(⋅)) and f(y(⋅)) are decreasing on [0,∞), we obtain by (8.258) and (8.245) that for each s≥0,
When combined with (8.247), this inequality implies that
It follows from Lemma 8.29 (with x=y, ε=δ 0), (8.258), (8.257), (8.252) and (8.245) that
This inequality and (8.250) imply that
Now we estimate ∥x(τ)−y(τ)∥. It follows from (8.257), (8.244) and (8.258) that for each s∈[0,τ],
By (8.259) and (8.254), for each s∈(0,τ], we have
By (8.259) and (8.251), for each s∈[0,τ],
It follows from (8.261), (8.262) and (8.263) that for each s∈[0,τ],
Applying Gronwall’s inequality, we obtain that
When combined with (8.253), this inequality implies that
Together with (8.260), this implies that \(\|x(\tau)-\bar{x}\| \le \varepsilon _{1}/2\). Lemma 8.30 is proved. □
Completion of the proof of Theorem 8.25
Let \(V \in{ \mathcal{A}}_{\gamma}\), γ∈(0,1), and let i be a natural number. By Lemma 8.30, there exist an open neighborhood \({\mathcal{U}}(V,\gamma,i)\) of V γ in \({\mathcal{A}}_{l}\) with the weak topology and a positive number τ(V,γ,i) such that the following property holds:
- (P):
-
For each \(W \in{ \mathcal{U}}(V,\gamma,i)\) and each continuously differentiable mapping x:[0,∞)→X satisfying
$$\begin{gathered} x'(t)=Wx(t), \quad t \in[0,\infty), \\ \bigl|f\bigl(x(0)\bigr)\bigr| \le i,\end{gathered} $$the following inequality holds:
$$\bigl\| x(t)-\bar{x}\bigr\| \le i^{-1} \quad\hbox{for all } t \ge\tau(V,\gamma,i). $$
Set
By Lemma 8.27, \({\mathcal{F}}\) is a countable intersection of open (in the weak topology) everywhere dense (in the strong topology) subsets of \({\mathcal{A}}_{l}\).
Let \(\tilde{V} \in{ \mathcal{F}}\) and let n,ε>0 be given. Choose a natural number i such that
By (8.265), there are \(V \in{ \mathcal{A}}_{l}\) and γ∈(0,1) such that
We claim show that the assertion of Theorem 8.15 holds with \({\mathcal{U}}={\mathcal{U}}(V,\gamma,i)\) and T εn =τ(V,γ,i).
Assume that \(W \in{ \mathcal{U}}(V,\gamma,i)\) and that the continuously differentiable mapping y:[0,∞)→X satisfies
Then by (8.268), (8.266) and property (P), it follows that
When combined with (8.266), this inequality implies that \(\|y(t)-\bar{x}\| \le \varepsilon \) for all t≥τ(V,γ,i). Theorem 8.25 is established. □
References
Aizicovici, S., Reich, S., & Zaslavski, A. J. (2005). Archiv der Mathematik, 85, 268–277.
Alber, Y. I., Iusem, A. N., & Solodov, M. V. (1997). Journal of Convex Analysis, 4, 235–255.
Ben-Tal, A., & Zibulevsky, M. (1997). SIAM Journal on Optimization, 7, 347–366.
Clarke, F. H. (1983). Optimization and nonsmooth analysis. New York: Wiley.
Correa, R., & Lemaréchal, C. (1993). Mathematical Programming, 62, 261–275.
Curry, H. B. (1944). Quarterly of Applied Mathematics, 2, 258–261.
Gowda, M. S., & Teboulle, M. (1990). SIAM Journal on Control and Optimization, 28, 925–935.
Hiriart-Urruty, J.-B., & Lemaréchal, C. (1993). Convex analysis and minimization algorithms. Berlin: Springer.
Lebourg, G. (1979). Transactions of the American Mathematical Society, 256, 125–144.
Neuberger, J. W. (2010). Lecture notes in mathematics: Vol. 1670. Sobolev gradients and differential equations (2nd ed.). Berlin: Springer.
Reich, S., & Zaslavski, A. J. (2000). Mathematics of Operations Research, 25, 231–242.
Reich, S., & Zaslavski, A. J. (2001). SIAM Journal on Optimization, 11, 1003–1018.
Reich, S., & Zaslavski, A. J. (2001). Nonlinear Analysis, 47, 3247–3258.
Reich, S., & Zaslavski, A. J. (2003). Electronic Journal of Differential Equations, 2003, 1–11.
Reich, S., & Zaslavski, A. J. (2008). Taiwanese Journal of Mathematics, 12, 1165–1176.
Zaslavski, A. J. (1981). Siberian Mathematical Journal, 22, 63–68.
Zaslavski, A. J. (2010). Optimization on metric and normed spaces. New York: Springer.
Author information
Authors and Affiliations
Rights and permissions
Copyright information
© 2014 Springer Science+Business Media New York
About this chapter
Cite this chapter
Reich, S., Zaslavski, A.J. (2014). Descent Methods. In: Genericity in Nonlinear Analysis. Developments in Mathematics, vol 34. Springer, New York, NY. https://doi.org/10.1007/978-1-4614-9533-8_8
Download citation
DOI: https://doi.org/10.1007/978-1-4614-9533-8_8
Publisher Name: Springer, New York, NY
Print ISBN: 978-1-4614-9532-1
Online ISBN: 978-1-4614-9533-8
eBook Packages: Mathematics and StatisticsMathematics and Statistics (R0)