Abstract
We consider a parametric stochastic quasi-variational inequality problem (SQVIP for short) where the underlying normal cone is defined over the solution set of a parametric stochastic cone system. We investigate the impact of variation of the probability measure and the parameter on the solution of the SQVIP. By reformulating the SQVIP as a natural equation and treating the orthogonal projection over the solution set of the parametric stochastic cone system as an optimization problem, we effectively convert stability of the SQVIP into that of a one stage stochastic program with stochastic cone constraints. Under some moderate conditions, we derive Hölder outer semicontinuity and continuity of the solution set against the variation of the probability measure and the parameter. The stability results are applied to a mathematical program with stochastic semidefinite constraints and a mathematical program with SQVIP constraints.
Similar content being viewed by others
References
Arutyunov, A.V., Izmailov, A.F.: Sensitivity analysis for cone-constrained optimization problems under the relaxed constraint qualifications. Math. Oper. Res. 30, 333–353 (2005)
Aumann, R.J.: Integrals of set-valued functions. J. Math. Anal. Appl. 12, 1–12 (1965)
Aubin, J.P., Frankowska, H.: Set-Valued Analysis. Birkhäuser, Basel (1990)
Birbil, S., Gürkan, G., Listes, O.: Solving stochastic mathematical programs with complementarity constraints using simulation. Math. Oper. Res. 31, 739–760 (2006)
Bonnans, J.F., Shapiro, A.: Perturbation Analysis of Optimization Problems. Springer, New York (2000)
Chan, D., Pang, J.S.: The generalized quasi-variational inequality problem. Math. Oper. Res. 7, 211–222 (1982)
Chen, X., Fukushima, M.: Expected residual minimization method for stochastic linear complementarity problems. Math. Oper. Res. 30, 1022–1038 (2005)
Chen, X., Wets, R.B.-J., Zhang, Y.: Stochastic variational inequalities: residual minimization smoothing/sample average approximations. SIAM J. Optim. 22, 649–673 (2012)
Ding, C., Sun, D.F., Zhang, L.W.: Characterization of the robust isolated calmness for a class of conic programming problems. http://arxiv.org/abs/1601.07418v1 (2016)
Fiala, J., Kocvara, M., Stingl, M.: PENLAB: A MATLAB solver for nonlinear semidefinite optimization. http://arxiv.org/abs/1311.5240 (2013)
Facchinei, F., Pang, J.S.: Finite-Dimensional Variational Inequalities and Complementarity Problems I–II. Springer, New York (2003)
Izmailov, A.F., Kurennoy, A.S., Solodov, M.V.: A note on upper Liptschitz stability, error bounds and critical multipliers for Liptschitz-continuous KKT systems. Math. Program. 142, 591–604 (2013)
Jiang, H., Xu, H.: Stochastic approximation approaches to the stochastic variational inequality problem. IEEE Trans. Autom. Control 53, 1462–1475 (2008)
King, A.J., Rockafellar, R.T.: Sensitivity analysis for nonsmooth generalized equations. Math. Program. 55, 193–212 (1992)
King, A.J., Rockafellar, R.T.: Asymptotic theory for solutions in statistical estimation and stochastic programming. Math. Oper. Res. 18, 148–162 (1993)
Klatte, D.: On the stability of local and global optimal solutions in parametric problems of nonlinear programming, part I: basic results. In: Humboldt-Universit\(\ddot{a}\)t Sektion Mathematik, vol. 75, pp. 1–21 (1985)
Klatte, D., Kummer, B.: Nonsmooth Equations in Optimization: Regularity, Calculus, Methods and Applications. Kluwer, Dordrecht (2002)
Liu, Y., Römisch, W., Xu, H.: Quantitative stability analysis of stochastic generalized equations. SIAM J. Optim. 24, 467–497 (2014)
Liu, Y., Xu, H., Lin, G.H.: Stability analysis of two stage stochastic mathematical programs with complementarity constraints via NLP-regularization. SIAM J. Optim. 21, 609–705 (2011)
Lu, S., Budhiraja, A.: Confidence regions for stochastic variational inequalities. Math. Oper. Res. 38, 545–568 (2013)
Mordukhovich, B.S., Outrata, J.V.: Coderivative analysis of quasi-variational inequalities with applications to stability and optimization. SIAM J. Optim. 18, 389–412 (2007)
Pang, J.S., Fukushima, M.: Quasi-variational inequalities, generalized Nash equilibria, and multi-leader-follower games. CMS 2, 21–56 (2005)
Pflug, G.Ch.: Stochastic optimization and statistical inference. In: Ruszczynski, A., Shapiro, A. (eds.) Stochastic Programming (Handbooks in Operations Research and Management Science), vol. 10, pp. 427–480. Elsevier, Amsterdam (2003)
Pflug, G.Ch., Pichler, A.: Approximations for probability distributions and stochastic optimization problems. In: Bertocchi, M., Consigli, G., Dempster, M.A. (eds.) Stochastic Optimization Methods in Finance and Energy, vol. 163, pp. 343–387. Springer, New York (2011)
Ravat, U., Shanbhag, U.V.: On the characterization of solution sets of smooth and nonsmooth convex stochastic Nash games. SIAM J. Optim. 21, 1168–1199 (2011)
Rockafellar, R.T., Wets, R.J.B.: Variational Analysis. Springer, Berlin (1998)
Robinson, S.M.: An application of error bounds for convex programming in linear space. SIAM J. Control Optim. 13, 271–273 (1975)
Robinson, S.M.: Stability theory for systems of inequalities, part II: differentiable nonlinear systems. SIAM J. Numer. Anal. 13, 497–513 (1976)
Robinson, S.M.: Generalized equations and their solutions, part I: basic theory. Math. Progr. Study 10, 128–141 (1979)
Robinson, S.M.: Strongly regular generalized equations. Math. Oper. Res. 5, 43–62 (1980)
Robinson, S.M.: Generalized equations and their solutions, part II: applications to nonlinear programming. Math. Progr. Study 19, 200–221 (1982)
Rockafellar, R.T., Wets, R.J.B.: Stochastic variational inequalities: single-stage to multistage. Math. Program. (2016). doi:10.1007/s10107-016-0995-5
Römisch, W.: Stability of stochastic programming problems. In: Ruszczynski, A., Shapiro, A. (eds.) Stochastic Programming. Handbooks in Operations Research and Management Science, vol. 10, pp. 483–554. Elsevier, Amsterdam (2003)
Rusczynski, A., Shapiro, A.: Stochastic Programming, Handbooks in OR and MS. Elsevier, Amsterdam (2003)
Shapiro, A.: Sensitivity analysis of parameterized variational inequalities. Math. Oper. Res. 30, 109–126 (2005)
Shapiro, A.: Sensitivity analysis of generalized equations. J. Math. Sci. 115, 2554–2565 (2003)
Shapiro, A., Dentcheva, D., Ruszczynski, A.: Lectures on Stochastic Programming: Modeling and Theory. SIAM, Philadelphia (2009)
Wolfowitz, J.: Generalization of the theorem of Glivenko–Cantelli. Ann. Math. Stat. 25, 131–138 (1954)
Xu, H.: Sample average approximation methods for a class of stochastic variational inequality problems. Asian Pac. J. Oper. Res. 27, 103–119 (2010)
Xu, H., Ye, J.J.: Approximating stationary points of stochastic mathematical programs with variational inequality constraints via sample averaging. Set-Valued Var. Anal. 19, 283–309 (2011)
Yousefian, F., Nedić, A., Shanbhag, U. V.: A regularized smoothing stochastic approximation (RSSA) algorithm for stochastic variational inequality problems. In: Proceeding WSC ’13 Proceedings of the 2013 Winter Simulation Conference: Simulation: Making Decisions in a Complex World, pp. 933–944 (2013)
Yousefian, F., Nedich, A., Shanbhag, U.: Self-tuned stochastic approximation schemes for non-Lipschitzian stochastic multi-user optimization and Nash games. IEEE Trans. Autom. Control 61, 1753–1766 (2016)
Acknowledgements
We would like to thank the editor-in-chief and two anonymous referees for valuable comments which help us significantly consolidate the paper.
Author information
Authors and Affiliations
Corresponding author
Additional information
J. Zhang: The research of this author is supported by the NSFC under Project Nos. 11671183 and 11201210, CPSF under Project No. 2014M560200 and Program for Liaoning Excellent Talents in University under Project No. LJQ2015059.
L.-W. Zhang: The research of this author was supported by the NSFC under Project Nos. 91330206 and 11571059.
Appendix
Appendix
Proof of Lemma 2.2
Part (i). Observe that (2.4) implies that \(\Psi (\bar{u}) \ne \emptyset \), and under uniform continuity w.r.t. u further that \(\Psi (u) \ne \emptyset \) when u is sufficiently close to \(\bar{u}\).
It is well known (see [3, Definition 1.4.6]) that
Since \(\Psi (\bar{u})\) is closed, to show continuity of \(\Psi (\cdot )\) at \(\bar{u}\), it suffices to show that the set-valued mapping is both upper and lower semicontinuous at \(\bar{u}\), that is,
The upper semicontinuity can be easily verified under the continuity of \(\psi \) and closeness of \(\mathcal{K}\). Hence it suffices to show the lower semicontinuity.
Observe that condition (2.4) and continuity of \(\psi (\cdot ,\bar{u})\) imply that \({\mathrm{int}} \; \Psi (\bar{u})\ne \emptyset \). Moreover, since \(\psi (z,u)\) is uniformly continuous w.r.t. u under condition (b), it is easy to show that for any \(z \in {\mathrm{int}} \; \Psi (\bar{u})\), we can set u sufficiently close to \(\bar{u}\) such that \(z\in \Psi (u)\). This allows us to claim that
Moreover, since \(\Psi (\bar{u})\) is convex and closed and by Rockafellar and Wets [26, Proposition 4.4], \(\liminf _{u\rightarrow \bar{u}} \Psi (u)\) is closed, we have
which yields the second inclusion of (6.1).
Part (ii). Since \(\Psi (\bar{u})\) is compact and \(\Psi (\cdot )\) is closed at \(\bar{u}\), the continuity of \(\Psi (\cdot )\) at \(\bar{u}\) means that for any number \(\delta >0\), there exists \(\epsilon >0\) such that when \(\Vert u-\bar{u}\Vert _U\le \epsilon \),
or equivalently
In what follows, we use Lemma 2.1 (Robinson–Ursescu’s theorem) to derive an error bound for set-valued mapping \(\Psi (\cdot )\) at \(\bar{u}\) under condition (d) and Slater constraint qualification (2.4).
Let
Condition (a) ensures that \(\mathcal {F}_{u}(z)\) is closed and convex set-valued over its domain \(\mathcal{Z}\). Moreover,
and \(z\in \Psi (u)\) if and only if \((z,0)\in \text{ gph }\; \mathcal {F}_{u}\). Furthermore, it follows from (2.4)
Let \(z \in \Psi (\bar{u})\) be fixed, that is, \((z,0)\in \text{ gph }\; \mathcal {F}_{\bar{u}}\). By Lemma 2.1, there exist positive numbers \(\delta _{z}\), \(\eta _z\) and \(c_{z}\) (depending on z) such that
for each \(z'\in B_m(z,\delta _{z})\cap \mathcal{Z}\) and \(q\in B_W(0,\eta _z)\). In particular,
Under condition (d), we have
for \(z'\in \Psi (u)\).
On the other hand, since z is arbitrarily chosen from \(\Psi (\bar{u})\), the set \(\Psi (\bar{u})\) may be covered by the union of a collection of \(\delta \)-balls, i.e.,
Since \(\Psi (\bar{u})\) is a compact set, by the finite covering theorem, there exist a finite number of points \(z_1, z_2,\dots , z_k \in \Psi (\bar{u})\) and positive constants \(\delta _{z_i}, c_{z_i}, i=1,\dots ,k\) such that
and
for any \(z'\in B_m(z_i,\delta _{z_i})\cap \mathcal{Z}\). Since \(\Psi (\bar{u})\) is a compact set and \(\bigcup _{i=1}^k\text{ int }\,\,B_m(z_i,\delta _{z_i})\) is an open set, there exists a positive constant \(\delta \) such that
Let \(C:=\max \{c_{z_1},\dots ,c_{z_k}\}\). Combining (6.11) and (6.10), we obtain
for all \(z'\in \left( \Psi (\bar{u})+\delta \mathcal{B}_m\right) \cap \mathcal{Z}\). Through (6.3) and (6.8), we arrive at
for all \(z'\in \Psi (u)\).
To complete the proof of (2.5), it suffices to show
For any \(z\in \mathcal{Z}\), let \(q=\psi (z,\bar{u})-\psi (z,u)\). Under condition (d), \( \Vert q\Vert _W\le \sigma \Vert u-\bar{u}\Vert _U^\nu . \) Moreover, it is easy to see that \( \Psi (u) =\mathcal{F}^{-1}_{\bar{u}}(q). \) Let \(z\in \Psi (\bar{u})\), \(\delta _z\) be defined as in (6.6) and \(\epsilon \) be defined at the beginning of the proof of this part with \( \sigma \epsilon ^\nu \le \delta _z. \) Under condition (d), we have \(\Vert q\Vert \le \delta _z\). It follows from (6.6) (by setting \(z'=z\)) that
Since \(z\in \Psi (\bar{u})\), then \(0\in \mathcal{F}_{\bar{u}}(z)\) and hence
Combining the two inequalities above, we obtain
for any \(z\in \Psi (\bar{u})\) and \(\Vert u-\bar{u}\Vert _U\le \epsilon \). Utilizing the second inclusion in (6.3), we can apply the finite covering theorem to set \(\Psi (\bar{u})\) (similar to previous discussion) and find a positive constant C such that (6.13) holds. The proof is complete. \(\square \)
In the case when \(\psi \) is continuously differentiable, Lemma 2.2 coincides with [28, Corollary 1]. Thus we may regard our result as an extension of the Robinson’s result.
Proof of Lemma 2.3
Let \(\hat{\phi }(z,u):=\phi (z,u)+\delta _{\Psi (u)}(z)\), where \(\delta _C(\cdot )\) denotes the indicator function of a set \(C\subseteq \mathbb {R}^m\) with \(\delta _C(x)=0\) for \(x\in C\) and \(\delta _C(x)=+\infty \) for \(x\notin C\). Problem (2.3) can be equivalently written as
The objective function \(\hat{\phi }\) has the following properties:
- (i):
-
Since \(\phi \) and \(\psi \) are continuous, \(\mathcal{K}\) and \(\mathcal{Z}\) are closed and convex, we have graph\(\Psi \) is closed, which means that \(\mathrm{epi}\,\delta _{\Psi (\cdot )}(\cdot )=\{(u,z, \alpha ): z\in \Psi (u), \alpha \ge 0\}=\mathrm{graph} \Psi \times \mathbb {R}_+\) is closed. Hence \(\hat{\phi }\) is proper lower semicontinuous;
- (ii):
-
Since \(\mathcal{Z}\) is compact, \(\hat{\phi }\) is level-bounded in z locally uniformly in u;
- (iii):
-
Under the Slater condition (2.4), one has from Lemma 2.2 that
$$\begin{aligned} \lim _{u\rightarrow \bar{u}}\Psi (u)=\Psi (\bar{u}), \end{aligned}$$which means that
$$\begin{aligned}\limsup _{u\rightarrow \bar{u}}\mathrm {epi}\, [\delta _{\Psi (u)}(\cdot )]=\mathrm {epi}\, [\delta _{\Psi (\bar{u})}(\cdot )],\end{aligned}$$where
$$\begin{aligned}\mathrm {epi}\, [\delta _{\Psi (u)}(\cdot )]=\{(z,\alpha ): z\in \Psi (u), \alpha \ge 0\}=\Psi (u)\times \mathbb {R}_+.\end{aligned}$$Then we obtain from Definition 2.1 that
$$\begin{aligned}e\!-\!\!\!\lim _{u\rightarrow \bar{u}}\delta _{\Psi (u)}(\cdot )=\delta _{\Psi (\bar{u})}(\cdot ),\end{aligned}$$which, by Proposition 2.1, means that for any sequence \(u^{\nu }\rightarrow \bar{u}\) and point \(\bar{z}\in \mathbb {R}^m\),
$$\begin{aligned}\liminf _{\nu \rightarrow \infty }\delta _{\Psi (u^{\nu })}(z^{\nu })\ge \delta _{\Psi (\bar{u})}(\bar{z})\end{aligned}$$for every sequence \(z^{\nu }\rightarrow \bar{z}\) and there exists some \(z^{\nu }\rightarrow \bar{z}\) such that
$$\begin{aligned}\limsup _{\nu \rightarrow \infty }\delta _{\Psi (u^{\nu })}(z^{\nu })\le \delta _{\Psi (\bar{u})}(\bar{z}).\end{aligned}$$We know from the continuity of \(\phi \) that
$$\begin{aligned}\lim _{\nu \rightarrow \infty } \phi (z^{\nu }, u^{\nu })=\phi (\bar{z},\bar{u})\,\,\text{ as }\,\, \nu \rightarrow \infty ,\end{aligned}$$which leads to
$$\begin{aligned}\liminf _{\nu \rightarrow \infty }\,[\phi (z^{\nu }, u^{\nu })+\delta _{\Psi (u^{\nu })}(z^{\nu })]\ge \phi (\bar{z}, \bar{u})+\delta _{\Psi (\bar{u})}(\bar{z})\end{aligned}$$and there exists \(z^{\nu }\rightarrow \bar{z}\) such that
$$\begin{aligned} \limsup _{\nu \rightarrow \infty }\,[\phi (z^{\nu }, u^{\nu })+\delta _{\Psi (u^{\nu })}(z^{\nu })]\le \phi (\bar{z}, \bar{u})+\delta _{\Psi (\bar{u})}(\bar{z}).\end{aligned}$$Then also by Proposition 2.1, we have that \(u\mapsto \hat{\phi }(\cdot ,u)\) is epi-continuous at \(\bar{u}\).
Hence all conditions in [26, Theorem 7.41] are satisfied, and from this theorem we obtain that \({\vartheta }(\cdot )\) is continuous at \(\bar{u}\) and \(\mathcal{Z}(\cdot )\) is outer semicontinuous at \(\bar{u}\). Noting that \(\mathcal{Z}(u)\subset \mathcal{Z}\) and it actually is locally bounded around \(\bar{u}\), we have from [26, Theorem 5.19] that \(\mathcal{Z}(\cdot )\) is upper semicontinuous at \(\bar{u}\). \(\square \)
Proof of Theorem 2.1
Part (i). First we show that
where \(\Pi _{\Psi (\bar{u})}(z)\) denotes orthogonal projection of z on \(\Psi (\bar{u})\),
Let \(z(u) \in \mathcal{Z}(u)\) and \(z(\bar{u}) \in \mathcal{Z}(\bar{u})\). Under conditions (b) and (c)
for \(u\in B(\bar{u},\bar{\delta }_u)\). On the other hand, since \(\Pi _{\Psi (u)}(z(\bar{u}))\in \Psi (u)\) and z(u) is an optimal solution to \(\mathcal{P}_{u}\), we have
Using this inequality, (2.8) and the growth condition (c), we have
for \(u\in B(\bar{u},\bar{\delta }_u)\). Combining (6.17) and (6.18), we obtain
which yields (6.15) through the triangle inequality below
because z(u) and \(z(\bar{u})\) are arbitrarily taken from the set of optimal solutions. Since \(\mathcal {Z}(u)\subseteq \Psi (u)\), it follows by Lemma 2.2 that there exists a constant \(\epsilon >0\) such that
when \(\Vert u-\bar{u}\Vert _U\le \epsilon \). Combining (6.15) and (6.21), we obtain (2.9).
Part (ii). We only need to show that
From the proof of Part (i), we can see that we can swap u with \(\bar{u}\) except that the first inequality in formulae (6.18) requires second order growth condition of \(\phi (\cdot ,u)\) over the optimal solution set \(\mathcal{Z}(u)\). \(\square \)
Rights and permissions
About this article
Cite this article
Zhang, J., Xu, H. & Zhang, L. Quantitative stability analysis of stochastic quasi-variational inequality problems and applications. Math. Program. 165, 433–470 (2017). https://doi.org/10.1007/s10107-017-1116-9
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10107-017-1116-9
Keywords
- Stochastic quasi-variational inequality
- Quantitative stability analysis
- Mathematical program with stochastic semidefinite constraints
- Mathematical program with SQVIP constraints