View source for Fritz John condition

{{TEX|part}}

A necessary condition for local optimality in problems with inequality constraints. It is closely related to the classical problem of minimizing a function $f$ on a constraint set $F=\left\{x\in\mathbb{R}^n\colon h^i(x)=0, i\in P\right\}$. The functions are defined on $\mathbb{R}^n$, the objective $f$ is assumed to be [[Differentiable function|differentiable]] and the constraints $h^i$, $i\in P$, are continuously differentiable at a feasible point $x^*\in F$ tested for optimality; $P$ is a finite index set. If $f(x^*)\leq f(x)$ for every $x\in F\cap N(x^*)$, where $N(x^*)$ is a neighbourhood of $x^*$, then $x^*$ is said to be a constrained local minimum. At such a point the gradients of the objective function and the constraints are linearly dependent, i.e., there exist multipliers $\lambda_i\in\mathbb{R}$, $i\in\{0\}\cup P$, not all zero, such that $\lambda_0\nabla f(x^*) + \sum_{i\in P}\lambda_i\nabla h^i(x^*)=0$. If the gradients of the constraints are linearly independent, then one can specify $\lambda_0=1$. This result, usually proved by the [[Implicit function|implicit function theorem]], is used to formulate the [[Lagrange method]] in calculus.

Now assume that the feasible set is determined by inequality constraints, i.e., consider
\begin{equation} \label{eq:1}
\begin{array}{ll}
\mbox{minimize} & f(x), \\
\mbox{subject to} & x\in F=\left\{x\in\mathbb{R}^n\colon g^i(x)\leq 0,\; i\in P\right\}.
\end{array}
\end{equation}
The Fritz John condition describes local optimality of a feasible point $x^*$ using the gradients of the objective function and the  "active"  constraints $P(x^*)=\left\{i\in P\colon g^i(x^*)=0\right\}$, [[#References|[a10]]]. The basic Fritz John condition is as follows. Consider the problem \eqref{eq:1} where all functions are differentiable at some feasible $x^*$. If $x^*$ is a constrained local minimum, then there exist multipliers $u_i\geq 0$, $i\in\{0\}\cup P(x^*)$, not all zero, such that
\begin{equation} u_0\nabla f(x^*) + \sum_{i\in P(x^*)}u_i\nabla g^i(x^*)=0. \end{equation}
By Gordan's theorem of the alternative (e.g., [[#References|[a14]]]), or the Dubovitskii–Milyutin theorem (e.g., [[#References|[a7]]]), the Fritz John condition is equivalent to the inconsistency of the system
\begin{equation}
\nabla f(x^*)d<0,\quad\nabla g^i(x^*)d<0,\quad i\in P(x^*),
\end{equation}
which is a  "primal"  optimality condition. For problems with both equality and inequality constraints, the Fritz John condition requires that the equality constraints be continuously differentiable while the active inequality constraints need only be differentiable. The multipliers that correspond to the equality constraints are unrestricted in sign. This result, referred to as the Mangasarian–Fromovitz condition, does not follow from the Fritz John condition; e.g., [[#References|[a14]]], [[#References|[a15]]]. Conditions that guarantee that the leading multiplier in the Fritz John condition is positive, in which case it can be set $u_0=1$, are called regularization conditions or constraint qualifications. One of these is that the gradients of active constraints are linearly independent. This condition is typically violated in multi-level and [[multi-objective optimization]] problems; e.g., [[#References|[a4]]], [[#References|[a23]]].



==Convex programs==

The Fritz John condition is not sufficient for optimality even for [[Linear programming|linear programs]]; e.g., [[#References|[a3]]], p. 150, [[#References|[a4]]], [[#References|[a14]]]. However, a reformulation of it is both necessary and sufficient for optimality of a feasible point $x^*$ for a [[convex programming|convex program]], i.e., for the problem \eqref{eq:1} when all functions are [[Convex function (of a real variable)|convex]]. First, using properties of feasible directions, optimality of $x^*$ is characterized by the inconsistency of the system
\begin{equation}
\nabla f(x^*)d<0,\quad\nabla g^i(x^*)d<0,\quad i\in P(x^*)\setminus P^=,\quad d\in D(x^*),
\end{equation}
where $P^=$ is the set of the constraints that are equal to zero on the entire feasible set and $D(x^*)$ is the intersection of the cones of directions of constancy at $x^*$ of all such constraints. This is equivalent to the Fritz John formulation $u_0\nabla f(x^*) + \sum_{i\in P(x^*)\setminus P^=}u_i\nabla g^i(x^*)\in \{D(x^*)\}^+$, where the multipliers are non-negative and not all equal to zero, and $\{D(x^*)\}^+$ is the [[polar set]] of $D(x^*)$, [[#References|[a4]]]. Moreover, feasibility of $x^*$ guarantees here $u_0=1$. If the gradients of active constraints are linearly independent, or, more generally, if $P^=$ is an empty set (this is known as Slater's condition, cf. [[Mathematical programming]]), then the Fritz John condition becomes
\begin{equation} \nabla f(x^*) + \sum_{i\in P(x^*)}u_i\nabla g^i(x^*)=0,\quad u_i\geq 0,\quad i\in P(x^*) \end{equation}
Consistency of this system is known as the [[Karush–Kuhn–Tucker conditions]].

Alternatively, in convex programming the Fritz John condition is equivalent to the existence of a [[saddle point]] of the Lagrangian $L(x,u)=f(x)+\sum_{i\in P\setminus P^=}u_i g^i(x)$. First, define $F^==\left\{x\colon g^i(x)=0,\; i\in P^=\right\}$ and let $c$ be the cardinality of $P\setminus P^=$. Consider the convex program \eqref{eq:1}. Then a point $x^*\in F^=$ is optimal if and only if there exists a $u^*\geq 0$ in $\mathbb{R}^c$ such that $L(x^*,u)\leq L(x^*,u^*)\leq L(x,u^*)$ for every $u\geq 0$ and every $x\in F^=$ (saddle-point characterization of optimality).

This formulation is useful when some functions are not differentiable. Using non-smooth analysis, one can replace derivatives by other objects such as subgradients (see, e.g., [[#References|[a3]]]).



==Optimality and stability==

Descriptions of optimality often require stability in the convex model

<table class="eq" style="width:100%;"> <tr><td valign="top" style="width:94%;text-align:center;"><img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025052.png" /></td> </tr></table>

all functions are continuous, <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025053.png" /> is considered as a  "parameter" , and <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025054.png" />, <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025055.png" /> are convex for every <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025056.png" />. (The functions need not be convex in <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025057.png" />.) Assume that the set of optimal solutions <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025058.png" /> of the corresponding convex program is non-empty and bounded at some <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025059.png" />. Perturbations of <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025060.png" /> from <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025061.png" /> that locally preserve lower semi-continuity of the feasible set mapping <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025062.png" /> form a  "region of stability"  at <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025063.png" />, denoted by <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025064.png" />. Denote the constraints that are equal to zero on <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025065.png" /> by <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025066.png" />. Let <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025067.png" />, let <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025068.png" /> be the cardinality of the set <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025069.png" />, and consider the Lagrangian <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025070.png" />. Then the following condition characterizes local optimality of <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025071.png" />, relative to stable perturbations, for the optimal value function <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025072.png" /> (cf., e.g., [[#References|[a23]]]).

Consider the convex model (NP,<img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025073.png" />) around some <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025074.png" />. Then <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025075.png" /> locally minimizes <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025076.png" /> relative to perturbations in <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025077.png" /> if and only if there exists a non-negative function <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025078.png" /> such that, whenever <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025079.png" />,

<table class="eq" style="width:100%;"> <tr><td valign="top" style="width:94%;text-align:center;"><img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025080.png" /></td> </tr></table>

<table class="eq" style="width:100%;"> <tr><td valign="top" style="width:94%;text-align:center;"><img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025081.png" /></td> </tr></table>

for every non-negative <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025082.png" /> and every <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025083.png" /> (a characterization of local optimality on a region of stability).

The above claim does not generally hold if the stability requirement is omitted. As an example, consider the convex model <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025084.png" />. Here <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025085.png" /> for any <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025086.png" />, hence <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025087.png" /> is a global minimum. The saddle-point condition requires <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025088.png" /> such that <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025089.png" /> for all <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025090.png" />. But such a multiplier function does not exist outside the region of stability, e.g., for <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025091.png" />, as a sequence on <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025092.png" /> with <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025093.png" /> shows.

The necessary conditions for local optimality of parameters, subject to stable perturbations, are simplified under  "input constraint qualifications"  [[#References|[a21]]]. Global optimality of a parameter <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025094.png" /> for the convex model can be characterized by a saddle-point condition on the region of cooperation of <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025095.png" />. This is the set of all <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025096.png" /> for which <img align="absmiddle" border="0" src="https://www.encyclopediaofmath.org/legacyimages/f/f130/f130250/f13025097.png" />. No stability is required in this case [[#References|[a22]]]. Programs that assume the form of a convex model, after  "freezing"  some variables, are called partly convex. Every program with twice continuously differentiable functions can be transformed into an equivalent partly convex program (cf., e.g., [[#References|[a8]]], [[#References|[a12]]]).



==Abstract programs==

Various extensions of the basic Fritz John condition have been studied in abstract spaces with finite or infinite number of constraints; e.g., [[#References|[a2]]], [[#References|[a7]]], [[#References|[a11]]], [[#References|[a13]]], [[#References|[a24]]]. The weakest derivative for which this condition holds appears to be the Gil de Lamadrid and Sova compact derivative [[#References|[a16]]] in non-sequential locally convex [[Hausdorff space]]s. A saddle-point condition that characterizes locally optimal parameters relative to their stable perturbations for the abstract convex model $\min_{x}f(x,\theta)$ on the set $x\in F(\theta)=\left\{x\colon g^\tau(x,\theta)\leq 0,\; \tau\in T\right\}$, where $x$ and the parameter $\theta$ belong to some suitable abstract spaces and $T$ is a compact subset of $\mathbb{R}$, can be given in terms of the existence of a finite [[Borel measure]] $\xi$ for the Lagrangian $L(x,\xi;\;\theta)=f(x,\theta)+\int_T g^\tau(x,\theta)\xi (dt)$ (cf. e.g. [[#References|[a1]]]). The abstract versions of the Fritz John and Karush–Kuhn–Tucker conditions are applicable to a wide range of problems, from optimal control problems in engineering [[#References|[a5]]], [[#References|[a6]]], [[#References|[a17]]] and management [[#References|[a18]]], [[#References|[a20]]] to the prior selection problem in Bayesian statistical inference [[#References|[a1]]]. These conditions are used for identification of optimal solutions, parameters, controls, and states, and for formulations of duality theories and numerical methods. The multipliers describe sensitivity of the optimal value function subject to the right-hand side perturbations and can be interpreted as values of the constraints ( "shadow prices"  in linear programming). Some optimization problems can alternatively be studied by calculus of variations which, fundamentally, is the same method; e.g., [[#References|[a6]]], [[#References|[a9]]], [[#References|[a18]]], [[#References|[a19]]], [[#References|[a20]]], p. 17.



==References==

<table>
<TR><TD valign="top">[a1]</TD> <TD valign="top">  M. Asgharian,   S. Zlobec,   "Abstract parametric programming"  ''Preprint McGill Univ.'' , '''March'''  (2000)</TD></TR>
<TR><TD valign="top">[a2]</TD> <TD valign="top">  V. Barbu,   Th. Precupanu,   "Convexity and optimization in Banach spaces" , Sijthoff &amp; Noordhoff  (1978)</TD></TR>
<TR><TD valign="top">[a3]</TD> <TD valign="top">  M.S. Bazaraa,   H.D. Sherali,   C.M. Shetty,   "Nonlinear programming: Theory and algorithms" , Wiley  (1993)  (Edition: Second)</TD></TR>
<TR><TD valign="top">[a4]</TD> <TD valign="top">  A. Ben-Israel,   A. Ben-Tal,   S. Zlobec,   "Optimality in nonlinear programming: A feasible directions approach" , Wiley/Interscience  (1981)</TD></TR>
<TR><TD valign="top">[a5]</TD> <TD valign="top">  E. Bryson, Jr.,   Yu-Chi Ho,   "Applied optimal control" , Blaisdell  (1969)</TD></TR>
<TR><TD valign="top">[a6]</TD> <TD valign="top">  M. Canon,   C. Cullum,   E. Polak,   "Theory of optimal control and mathematical programming" , McGraw-Hill  (1970)</TD></TR>
<TR><TD valign="top">[a7]</TD> <TD valign="top">  I.V. Girsanov,   "Lectures on mathematical theory of extremum problems" , ''Lecture Notes in economics and math. systems'' , '''67''' , Springer  (1972)</TD></TR>
<TR><TD valign="top">[a8]</TD> <TD valign="top">  J. Guddat,   H.Th. Jongen,   "On global optimization based on parametric optimization"  J. Guddat (ed.)  et al. (ed.) , ''Advances in Mathematical Optimization'' , Akad. Berlin  (1988)  pp. 63–79</TD></TR>
<TR><TD valign="top">[a9]</TD> <TD valign="top">  H. Halkin,   "Maximum principle of the Pontryagin type for systems described by nonlinear difference equations"  ''SIAM J. Control'' , '''4'''  (1966)  pp. 90–111</TD></TR>
<TR><TD valign="top">[a10]</TD> <TD valign="top">  F. John,   "Extremum problems with inequalities as subsidiary conditions"  K.O. Friedrichs (ed.)  et al. (ed.) , ''Studies and Essays, Courant Anniversary Volume'' , Wiley/Interscience  (1948)  (Reprinted in: J. Moser (ed.): Fritz John Collected Papers 2, Birkhäuser, 1985, pp. 543-560)</TD></TR>
<TR><TD valign="top">[a11]</TD> <TD valign="top">  M.B. Lignola,   J. Morgan,   "Existence of solutions to generalized bilevel programming problem"  A. Migdalas (ed.)  et al. (ed.) , ''Multilevel Optimization: Algorithms and Applications'' , Kluwer Acad. Publ.  (1998)  pp. 315–332</TD></TR>
<TR><TD valign="top">[a12]</TD> <TD valign="top">  W.B. Liu,   C.A. Floudas,   "A remark on the GOP algorithm for global optimization"  ''J. Global Optim.'' , '''3'''  (1993)  pp. 519–521</TD></TR>
<TR><TD valign="top">[a13]</TD> <TD valign="top">  D.G. Luenberger,   "Optimization by vector space methods" , Wiley  (1969)</TD></TR>
<TR><TD valign="top">[a14]</TD> <TD valign="top">  O.L. Mangasarian,   "Nonlinear programming" , McGraw-Hill  (1969)</TD></TR>
<TR><TD valign="top">[a15]</TD> <TD valign="top">  O.L. Mangasarian,   S. Fromovitz,   "The Fritz John optimality conditions in the presence of equality and inequality constraints"  ''J. Math. Anal. Appl.'' , '''17'''  (1967)  pp. 37–47</TD></TR>
<TR><TD valign="top">[a16]</TD> <TD valign="top">  H. Massam,   S. Zlobec,   "Various definitions of the derivative in mathematical programming"  ''Math. Programming'' , '''7'''  (1974)  pp. 144–161  (Addendum: ibid 14 (1978), 108-111)</TD></TR>
<TR><TD valign="top">[a17]</TD> <TD valign="top">  L.S. Pontryagin,   V.G. Boltyanski,   R.V. Gamkrelidze,   E.F. Mishchenko,   "The mathematical theory of optimal processes" , Wiley  (1962)</TD></TR>
<TR><TD valign="top">[a18]</TD> <TD valign="top">  S.P. Sethi,   "A survey of management science applications of the deterministic maximum principle" , ''TIMS Studies in the Management Sci.'' , '''9''' , North-Holland  (1978)  pp. 33–67</TD></TR>
<TR><TD valign="top">[a19]</TD> <TD valign="top">  D.R. Smith,   "Variational methods in optimization" , Prentice-Hall  (1974)</TD></TR>
<TR><TD valign="top">[a20]</TD> <TD valign="top">  C.S. Tapiero,   "Time, dynamics and the process of management modeling" , ''TIMS Studies in the Management Sci.'' , '''9''' , North-Holland  (1978)  pp. 7–31</TD></TR>
<TR><TD valign="top">[a21]</TD> <TD valign="top">  M. van Rooyen,   M. Sears,   S. Zlobec,   "Constraint qualifications in input optimization"  ''J. Austral. Math. Soc. Ser. B'' , '''30'''  (1989)  pp. 326–342</TD></TR>
<TR><TD valign="top">[a22]</TD> <TD valign="top">  S. Zlobec,   "Partly convex programming and Zermelo's navigation problems"  ''J. Global Optim.'' , '''7'''  (1995)  pp. 229–259</TD></TR>
<TR><TD valign="top">[a23]</TD> <TD valign="top">  S. Zlobec,   "Stable parametric programming"  ''Optimization'' , '''45'''  (1999)  pp. 387–416  ((Augmented version forthcoming as research monograph, Kluwer Acad. Publ., Applied Optim. Series.))</TD></TR>
<TR><TD valign="top">[a24]</TD> <TD valign="top">  S. Zlobec,   B.D. Craven,   "Stabilization and determination of the set of minimal binding constraints in convex programming"  ''Math. Operationsforschung und Statistik, Ser. Optim.'' , '''12'''  (1981)  pp. 203–220</TD></TR>
</table>