Abstract Generated abstract
This paper studies directional differentiability of maximin functions of the form defined by maximizing over one variable after minimizing over another, with dependence on a parameter. In finite-dimensional settings, it gives formulas for the one-sided directional derivative with respect to the parameter under several sets of hypotheses, including compactness and continuity of minimizer sets, strict concavity with respect to the maximizing variable on an unbounded domain, and a compact constrained case using cones of admissible directions. The results extend differentiability techniques used for minimax and approximation problems to maximin functions relevant to applications such as differential games, and include remarks on relaxing boundedness assumptions when minimizer sets remain bounded.
Full Text
UDC 519.8
MATHEMATICS
V. F. Dem’yanov
ON DIRECTIONAL DIFFERENTIATION OF A MAXIMIN FUNCTION
(Presented by Academician L. V. Kantorovich, 1 VI 1967)
In \((^{1,2})\), for the solution of minimax problems the maximum function was studied, and directional differentiability was proved for the function
\[ \varphi(X)=\min f(X,Y). \]
The use of directional differentiability of such functions makes it possible to develop effective methods for solving a number of problems (best-approximation problems, minimax problems, etc.).
For the study of some other problems (for example, pursuit problems in the theory of differential games) it may be useful to study a function of the form
\[ \varphi(Z)=\max_X \min_Y f(X,Y,Z). \]
Below we present some results connected with the question of the differentiability of the function \(\varphi(Z)\) in directions.
Although a finite-dimensional case is considered here, many of the results can be generalized to the infinite-dimensional case (just as the results obtained in \((^1)\) for the finite-dimensional case were extended in \((^2)\) to a more general case).
\(1^\circ\). Let \(f(X,Y,Z)\) be a function continuous in \(X\) and \(Y\) and continuously differentiable in \(Z\) on \(\Omega_X \times \Omega_Y \times \Omega_Z\), where \(\Omega_X \subset E_n,\ \Omega_Y \subset E_m;\ \Omega_Z \subset E_p\).
On \(\Omega_Z\) consider the function
\[ \varphi(Z)\equiv \max_{X\in\Omega_X}\ \min_{Y\in\Omega_Y} f(X,Y,Z). \tag{1} \]
The sets \(\Omega_X,\Omega_Y,\Omega_Z\) are bounded and closed.
Fix some \(Z\in\Omega_Z\). Let \(g\in E_p;\ g\ne 0\) be such that, for \(\alpha\in[0,\alpha_0]\) \((\alpha_0=\alpha_0(g)>0)\), the point \(Z_\alpha=Z+\alpha g\in\Omega_Z\). Such a \(g\) will be called an admissible direction.
It is required to find the derivative of \(\varphi(Z)\) in the direction \(g\)
\[ \varphi'_Z(g)\equiv \lim_{\alpha\to +0}\frac{\varphi(Z+\alpha g)-\varphi(Z)}{\alpha}. \tag{2} \]
Introduce into consideration the sets \(R(Z)\subset\Omega_X\) and \(Q(X,Z)\subset\Omega_Y\):
\[ R(Z)=\left\{X\mid X\in\Omega_X;\ \min_{Y\in\Omega_Y} f(X,Y,Z)=\max_{X\in\Omega_X} f(X,Y,Z)\right\}, \]
\[ Q(X,Z)=\left\{Y\mid Y\in\Omega_Y;\ f(X,Y,Z)=\min_{Y\in\Omega_Y} f(X,Y,Z)\right\}. \]
In the present section we shall assume that the sets \(Q(X,Z)\) satisfy the following rather stringent condition:
Condition A. For a given \(Z\), the set \(Q(X,Z)\) depends continuously on \(X\) on the set \(R(Z)\) \((Q(X',Z)\to Q(X,Z)\) as \(X'\to X;\ X'\in\Omega_X;\ X\in R(Z))\).
This means that
\[ \rho\bigl(Q(X',Z),Q(X,Z)\bigr)\equiv \]
\[ \equiv \sup_{Y\in Q(X',Z)}\inf_{V\in Q(X,Z)}\|V-Y\|+ \sup_{V\in Q(X,Z)}\inf_{Y\in Q(X',Z)}\|V-Y\| \longrightarrow 0,\qquad X'\to X. \tag{3} \]
We note that this condition is satisfied, for example, if \(Q(X,Z)\) consists, for every \(X\in R(Z)\), of a single point.
Theorem 1. If at the point \(Z \in \Omega_Z\) condition A is satisfied, then the function \(\varphi(Z)\) is differentiable in any admissible direction, and
\[ \varphi'_Z(g)=\max_{X\in R(Z)}\min_{Y\in Q(X,Z)}(\partial f(X,Y,Z)/\partial Z,g). \]
\(2^\circ\). Let the function \(f(X,Y,Z)\) be given and continuous on \(E_n \times \Omega_X \times \Omega_Z\), where \(\Omega_X \subset E_m\), \(\Omega_Z \subset E_p\); \(\Omega_Y,\Omega_Z\) are bounded closed sets. Fix a point \(Z \in E_p\). Consider the sets \(R(Z)\subset E_n\) and \(Q(X,Z)\); they were defined in \(1^\circ\), only here already
\[ R(Z)=\{X\mid X\in E_n;\ \min_{Y\in\Omega_Y} f(X,Y,Z)=\max_{X\in E_n}\min_{Y\in\Omega_Y} f(X,Y,Z)\}. \]
We shall assume that:
-
For the point \(Z \in E_p\) there does not exist a sequence of points \(\{X_j\}\), \(\|X_j\|\to\infty\), such that \(\Phi(X_i,Z)\to \sup_{X\in E_n}\Phi(X,Z)\), where
\[ \Phi(X,Z)=\min_{Y\in\Omega_Y} f(X,Y,Z). \] -
The function \(f(X,Y,Z)\) is twice continuously differentiable with respect to \(X\) and \(Z\) on \(E_n\times\Omega_Y\times\Omega_Z\) and is strictly concave with respect to \(X\) for any fixed \(Y\in\Omega_Y,\ Z\in\Omega_Z\), and the components of the vector functions \(\partial f(X,Y,Z)/\partial X\), \(\partial f(X,Y,Z)/\partial Y\) and the elements of the matrices \(\partial^2 f(X,Y,Z)/\partial X^2\), \(\partial^2 f(X,Y,Z)/\partial Z^2\), \(\partial^2 f(X,Y,Z)/\partial X\partial Z\) are bounded on \(E_n\times\Omega_Y\times\Omega_Z\).
Let \(g\in E_p\) be an admissible direction.
Theorem 2. If conditions 1 and 2 are satisfied at the point \(Z\), then the function
\[ \varphi(Z)=\max_{X\in E_n}\min_{Y\in\Omega_Y} f(X,Y,Z) \]
is differentiable in any admissible direction \(g\), and
\[ \varphi'_Z(g)\equiv \lim_{\alpha\to+0}\frac{1}{\alpha}[\varphi(Z+\alpha g)-\varphi(Z)] = \]
\[ = \max_{X\in R(Z)}\max_{V\in E_n}\min_{Y\in Q(X,Z)} \left[(\partial f(X,Y,Z)/\partial X,V)+(\partial f(X,Y,Z)/\partial Z,g)\right]. \tag{4} \]
\(3^\circ\). Finally, let the function \(f(X,Y,Z)\) be given and continuous on \(\Omega_X\times\Omega_Y\times\Omega_Z\), where \(\Omega_X\subset E_n\), \(\Omega_Y\subset E_m\), \(\Omega_Z\subset E_p\) are bounded closed sets, and let \(f(X,Y,Z)\) be twice continuously differentiable on \(\Omega_X\times\Omega_Y\times\Omega_Z\) and strictly convex with respect to \(X\) for any fixed \(Y\in\Omega_Y,\ Z\in\Omega_Z\). The sets \(R(Z)\) and \(Q(X,Z)\) are defined as above.
Theorem 3. The function
\[ \varphi(Z)\equiv \max_{X\in\Omega_X}\min_{Y\in\Omega_Y} f(X,Y,Z) \]
is differentiable in any admissible direction \(g\), and
\[ \varphi'_Z(g)= \]
\[ = \max_{X\in R(Z)}\max_{V\in M_X(\Omega_X)}\min_{Y\in Q(X,Z)} \left[(\partial f(X,Y,Z)/\partial X,V)+(\partial f(X,Y,Z)/\partial Z,g)\right], \tag{5} \]
where \(M_X(\Omega_X)\) is the cone of admissible directions in the broad sense of the word, constructed at the point \(X\) with respect to the set \(\Omega_X\) (see (3)).
Remark 1. Formulas (3), (4), (5) are also valid in the case when the sets \(\Omega_Y\) and \(\Omega_Z\) are not bounded. In this case an additional condition must be imposed on the function \(f(X,Y,Z)\): the sets \(Q(X,Z)\) must be bounded, i.e., for some \(\varepsilon>0\) the set
\[
\bigcup_{X\in R(Z),\,Z'\in S_\varepsilon(Z)} Q(X,Z'),
\]
where \(S_\varepsilon(Z)\subset \Omega_Z\) is the sphere of radius \(\varepsilon\) with center at the point \(Z\), must be bounded.
Remark 2. In a similar way one can get rid of the boundedness of the set \(\Omega_X\) in \(1^\circ\).
Remark 3. In \(2^\circ\) and \(3^\circ\) the fulfillment of condition A was not required.
Leningrad State University
named after A. A. Zhdanov
Received
22 V 1967
REFERENCES
\(^{1}\) V. F. Demyanov, Cybernetics, 2, No. 6 (1966).
\(^{2}\) V. F. Demyanov, Vestnik Leningrad University, No. 7 (1966).
\(^{3}\) V. F. Demyanov, A. M. Rubinov, Economics and Mathematical Methods, No. 3 (1966).