随着大数据时代的到来,越来越多的学者关注稀疏优化问题,如带有零模约束的优化问题。
| $ \mathop {\min }\limits_{x \in {\mathbb{R}^n}} \;\{ f({\boldsymbol{x}}):{\boldsymbol{x}} \in {{\varOmega}} ,\;{\left\| {\boldsymbol{x}} \right\|_0} \leqslant \kappa {\text{\} }} $ | (1) |
式中:
| $ \langle {\boldsymbol{e}},{\boldsymbol{y}}\rangle \geqslant n - \kappa ,\langle \left| {\boldsymbol{x}} \right|,{\boldsymbol{y}}\rangle = 0,0\leqslant{\boldsymbol{y}} \leqslant{\boldsymbol{e}},{\boldsymbol{x}} \in \mathit{{ \varOmega }},{\boldsymbol{y}} \in {\mathbb{R}^n} $ |
式中:
| $ \{ ({\boldsymbol{x}},{\boldsymbol{y}}) \in {\mathbb{R}^n}:\left\langle {\left| {\boldsymbol{x}} \right|,{\boldsymbol{y}}} \right\rangle = 0,\;{\boldsymbol{y}} \geqslant 0\} $ | (2) |
另外,文献[12]引理3.1,给出了零模函数的一个等价刻画,即零模函数可看成是一个特定带有互补约束集(2)的参数优化的最优值函数。利用零模函数的这个变分刻画,文献[13-14]研究了零模正则问题和零模极小化问题的一些等价Lipschitz代理,并讨论了它们的稳定点之间的关系。
众所周知,刻画约束集合的相依切锥和极限法锥对描述约束优化问题的最优性条件是至关重要的。文献[15]刻画了集合
任意给定正整数
| $ {C^ \circ }: = \{ {\boldsymbol{v}} \in {\mathbb{R}^n}:\;\left\langle {{\boldsymbol{v}},{\boldsymbol{x}}} \right\rangle \leqslant 0,\;\forall {\boldsymbol{x}} \in C\} $ | (3) |
设
定义1
定义2 向量
| $ \mathop {\lim \sup }\limits_{{\boldsymbol{x}} \to \overline {\boldsymbol{x}} ,\;{\boldsymbol{x}} \in \varOmega } \frac{{\left\langle {{\boldsymbol{v}_n},{\boldsymbol{x}} - \overline {\boldsymbol{x}} } \right\rangle }}{{\| {{\boldsymbol{x}} - \overline {\boldsymbol{x}} } \|}} \leqslant 0 $ |
成立。所有在点
定义3 向量
显然,根据定义3可知式(4)的包含关系成立。
| $ N_\mathit{\Omega } ^f(\overline {\boldsymbol{x}} ) \subseteq {N_\mathit{\Omega } }(\overline {\boldsymbol{x}} ) $ | (4) |
根据文献[16]中的命题6.41可知,半互补集合(2)的相依切锥、正则法锥和极限法锥的刻画主要依赖
| $ D: = \{ (t,{{\tau }} ) \in \mathbb{R} \times {\mathbb{R}_ + }:\;\left| t \right|{ {\tau}} = 0\} $ |
的相依切锥、正则法锥和极限法锥的刻画。下文刻画了半互补集合
定理1 假设
| $ {T_D}\left( {\overline t,\overline \tau } \right) = \left\{ {\begin{array}{*{20}{l}} {\left\{ 0 \right\} \times \mathbb{R},}&{\overline t = 0,\overline \tau > 0}\\ {\left( {\mathbb{R} \times \left\{ 0 \right\}} \right) \cup \left( {\left\{ 0 \right\} \times {\mathbb{R}_ + }} \right),}&{\overline t = 0,\overline \tau = 0}\\ {\mathbb{R} \times \left\{ 0 \right\},}&{\overline t \ne 0,\overline \tau = 0} \end{array}} \right. $ |
证明 根据半互补集合
| $ D = (\{ 0\} \times {\mathbb{R}_ + }) \cup (\mathbb{R} \times \{ 0\} ) $ |
由于
情况1
| $ (\overline t ,\overline \tau ) + {\lambda _k}(0,{{{d}}^k}) = (0,\overline \tau + \frac{1}{k}{d^k}) \in D $ |
成立。根据定义1,可知
| $ \{ 0\} \times \mathbb{R} \subseteq {T_D}(\overline t ,\overline {{ {\tau}}} ) $ | (5) |
反过来,任意从集合
| $ \overline \tau + {\lambda _k}{\eta ^k} \geqslant 0 {,}\; \;\; {\lambda _k}| {{\zeta ^k}} |(\overline \tau + {\lambda _k}{\eta ^k}) = 0 $ |
由于
| $ {T_D}(\overline t ,\overline {{ {\tau}}} ) \subseteq \{ 0\} \times \mathbb{R} $ | (6) |
根据式(5)和式(6)可知,当
| $ {T_D}(\overline t ,\overline \tau ){\text{ = }}\{ 0\} \times \mathbb{R} $ |
情况2
| $ (\mathbb{R} \times \{ 0\} ) \cup (\{ 0\} \times {\mathbb{R}_ + }) \subseteq {T_D}(\overline t ,\overline \tau ) $ |
成立。任意取定
| $ (\overline t ,\overline \tau ) + {\lambda _k}({d^k},0) = (\frac{1}{k}{d^k},0) \in D $ |
成立。根据定义1,可知
| $ \mathbb{R} \times {\text{\{ }}0{\text{\} }} \subseteq {T_D}(\overline t ,\overline \tau ) $ | (7) |
类似地,任意取定
| $ (\overline t ,\overline \tau ) + {\lambda _k}(0,{v^k}) = (0,\frac{1}{k}{v^k}) \in D $ |
成立。根据定义1,可知
| $ \{ 0\} \times {\mathbb{R}_{\text{ + }}} \subseteq {T_D}(\overline t ,\overline \tau ) $ | (8) |
联合式(7)和式(8),可得
| $ (\mathbb{R} \times \{ 0\} ) \cup (\{ 0\} \times {\mathbb{R}_{\text{ + }}}) \subseteq {T_D}(\overline t ,\overline \tau ) $ | (9) |
下文主要来论证式(10)的包含关系成立。
| $ {T_D}(\overline t ,\overline \tau ) \subseteq (\mathbb{R} \times \{ 0\} ) \cup (\{ 0\} \times {\mathbb{R}_ + }) $ | (10) |
任意从集合
| $ {{\lambda _k}}{\eta ^k}\geqslant 0,\; ({\lambda _k})^2| {{\zeta ^k}} |{\eta ^k} = 0 $ | (11) |
注意到
| $ {\eta ^k} \geqslant 0 和 | {{\zeta ^k}} |{\eta ^k} = 0 $ |
如果存在正整数
| $ (\zeta ,\eta ) \in \mathbb{R} \times \{ 0\} $ | (12) |
否则,对于序列
| $ (\zeta ,\eta ) \in \{ 0\} \times \mathbb{R} $ | (13) |
由
| $ {T_D}(\overline t ,\overline \tau ){\text{ = }}(\mathbb{R} \times \{ 0\} ) \cup (\{ 0\} \times {\mathbb{R}_ + }) $ |
情况3
(1)
| $ {T_D}(\overline t ,\overline \tau ){\text{ = }}\mathbb{R} \times \{ 0\} $ | (14) |
任意取定
| $ (\overline{t},\overline{\tau })+{\lambda }_{k}({d}^{k},0)=(\overline{t}+\frac{1}{k}{d}^{k},0)\in D $ |
成立。根据定义1,可知
| $ \mathbb{R} \times {\text{\{ }}0{\text{\} }} \subseteq {T_D}(\overline t ,\overline \tau ) $ | (15) |
反过来,任意从集合
| $ {\lambda _k}{\eta ^k} \geqslant 0$ |
| $ {({\lambda _k})^2}\left| {\overline t + {\lambda _k}{\zeta ^k}} \right|{\eta ^k} = 0 $ | (16) |
注意到
| $ {T_D}(\overline t ,\overline \tau ) \subseteq \mathbb{R} \times \{ 0\} $ | (17) |
结合式(15)和式(17),可以得知式(14)成立。
(2)
| $ {T_D}(\overline t ,\overline \tau ){\text{ = }}\mathbb{R} \times \{ 0\} $ |
综上所述,当
定理2 假设
| $ N_D^f(\overline t ,\overline \tau ) = \left\{ {\begin{array}{*{20}{c}} {\mathbb{R} \times \{ 0\} ,{\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} }&{\overline t = 0,\overline \tau > 0} \\ {\{ 0\} \times {\mathbb{R}\_ },}&{\overline t = 0,\overline \tau = 0} \\ {\{ 0\} \times \mathbb{R},{\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} }&{\overline t \ne 0,\overline \tau = 0} \end{array}} \right. $ | (18) |
证明 根据文献[1]中的定理6.28可知
| $ N_D^f(\overline t ,\overline \tau ) = {({T_D}(\overline t ,\overline \tau ))^ \circ } $ | (19) |
结合式(3),容易验证当
| $ N_D^f(\overline t ,\overline \tau ) = \mathbb{R} \times {\text{\{ }} 0{\text{\} }} $ |
当
| $ N_D^f(\overline t ,\overline \tau ) = {\text{\{ }} 0{\text{\} }} \times \mathbb{R} $ |
因此,只需要验证当
| $ N_D^f(\overline t ,\overline \tau ) = {\text{\{ }} 0{\text{\} }} \times {\mathbb{R}\_ } $ | (20) |
从集合
| $ \langle (u,v),(\zeta ,\eta )\rangle \leqslant 0,\forall (\zeta ,\eta )\in (\mathbb{R}\times \{0\})\cup (\{0\}\times {\mathbb{R}}_+) $ |
成立。根据式(3)和定理1相依切锥的刻画,可知
| $ (u,v) \in {({T_D}(\overline t ,\overline \tau ))^ \circ } $ |
再根据式(19)和
| $ {\text{\{ }} 0{\text{\} }} \times {\mathbb{R}\_ } \subseteq N_D^f(\overline t ,\overline \tau ) $ | (21) |
反过来,从集合
| $ \left\langle {(u,v),(\zeta ,\eta )} \right\rangle \leqslant 0 $ | (22) |
根据定理1,在式(22)中分别取
| $ N_D^f(\overline t ,\overline \tau ) \subseteq \{0\} \times {\mathbb{R}\_ } $ | (23) |
最后根据式(21)和式(23),可知式(20)成立。
定理3 假设
| $ {N_D}(\overline t ,\overline \tau ) = \left\{ {\begin{array}{*{20}{c}} {\mathbb{R} \times \{ 0\} ,{\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;{\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} }&{\overline t = 0,\overline \tau > 0} \\ {(\mathbb{R} \times \{ 0\} ) \cup (\{ 0\} \times \mathbb{R}),}&{\overline t = 0,\overline \tau = 0} \\ {\{ 0\} \times \mathbb{R},{\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;{\kern 1pt} {\kern 1pt} }&{\overline t \ne 0,\overline \tau = 0} \end{array}} \right. $ |
证明 由于
情况1
| $ {N_D}(0,\overline \tau ) \subseteq \mathbb{R} \times \{ 0\} $ | (24) |
现从
| $ ({u^k},{v^k}) \in N_D^f({t^k},{\tau ^k}) $ | (25) |
由
情况2
| $ {N_D}(0,0) \subseteq (\mathbb{R} \times \{ 0\} ) \cup (\{ 0\} \times \mathbb{R}) $ | (26) |
现从集合
| $ ({u^k},{v^k}) \in N_D^f({t^k},{\tau ^k}) $ | (27) |
注意到
(1) 假设存在正整数
| $ {u^k} = 0,\;\;\forall \;k > {k_3} $ | (28) |
否则一定存在一个子列
| $ {u}^{k}=0,\forall k\in {K}_{1} $ | (29) |
注意到
(2) 如果上述情况2中的(1)不成立,则一定存在一个子列
下文说明反向包含成立。任取一点
情况3
| $ {N_D}(\overline t ,0) \subseteq \{ 0\} \times \mathbb{R} $ | (30) |
现从集合
| $ ({u^k},{v^k}) \in N_D^f({t^k},{\tau ^k}) $ | (31) |
注意到
稀疏约束使得优化问题(1)产生了组合特征,是一个NP难问题。一般来讲,连续优化的理论通常不能用于处理该类问题,但是稀疏约束集合的特殊结构也为研究提供了一个新机遇。切锥、法锥是刻画优化问题的最优性条件的重要工具之一。本文利用了切锥、法锥的工具,刻画了半互补集合D的相依切锥、正则法锥和极限法锥。后续将提出相应的约束规范并建立最优性条件。这些结果在某种程度上丰富了非连续、非凸规划问题的最优性理论。
| [1] |
BIENTOCK D. Computational study of a family of mixed-integer quadratic programming problems[J].
Mathematical Programming, 1996, 74(2): 121-140.
DOI: 10.1007/BF02592208. |
| [2] |
CANDES, E J, WAKIN M B. An introduction to compressive sampling[J].
IEEE Signal Processing Magazine, 2008, 25(2): 21-30.
DOI: 10.1109/MSP.2007.914731. |
| [3] |
GAO J, LI D. Optimal cardinality constrained portfolio selection[J].
Operations Research, 2013, 61(3): 745-761.
DOI: 10.1287/opre.2013.1170. |
| [4] |
GAO J, LI D. A polynomial case of the cardinality-constrained quadratic optimization problem[J].
Journal of Global Optimization, 2013, 56(4): 1441-1455.
|
| [5] |
LI D, SUN X, WANG J. Optimal lot solution to cardinality constrained mean-variance formulation for portfolio selection[J].
Mathematical Finance, 2006, 16(1): 83-101.
DOI: 10.1111/j.1467-9965.2006.00262.x. |
| [6] |
MILLER R, MILLER J A, MILLER J F A P. Subset selection in regression[M]. Boca Raton, FL: CRC Press, 2002.
|
| [7] |
LORENZO D D, LIUZZI G, RINALDI F, et al. A concave optimization-based approach for sparse portfolio selection[J].
Optimization Methods & Software, 2012, 27(6): 983-1000.
|
| [8] |
MURRAY W, SHEK H. A local relaxation method for the cardinality constrained portfolio optimization problem[J].
Computational Optimization and Applications, 2012, 53: 681-709.
DOI: 10.1007/s10589-012-9471-1. |
| [9] |
SUN X, ZHENG X, LI D. Recent advances in mathematical programming with semicontinuous variables and cardinality constraint[J].
Journal of the Operational Research Society, 2013, 77(1): 55-77.
|
| [10] |
FENG M, MITCHELL J E, PANG J S, et al. Complementary formulations of l0 norm optimization problems[J].
Industrial Engineering and Management Sciences, 2013, 1: 11-21.
|
| [11] |
BURDAKOV O P, KANZOW C, SCHWARTE A. Mathematical programs with cardinality constraints: reformulation by complementarity-type conditions and a regularization method[J].
SIAM Journal on Optimization, 2015, 26(1): 397-425.
|
| [12] |
BI S, PAN S. Multistage convex relaxation approach to rank regularized minimization problems based on equivalent mathematical program with a generalized complementarity constraint[J].
SIAM Journal on Control and Optimization, 2017, 55(4): 2493-2518.
DOI: 10.1137/15M1037160. |
| [13] |
LIU Y, BI S, PAN S. Equivalent lipschitz surrogates for zero-norm and rank optimization problems[J].
Journal of Global Optimization, 2018, 72(4): 679-704.
DOI: 10.1007/s10898-018-0675-5. |
| [14] |
LIU Y, BI S, PAN S. Several classes of stationary points for rank regularized minimization problems[J].
SIAM Journal on Optimization, 2020, 30(2): 1756-1775.
DOI: 10.1137/19M1270987. |
| [15] |
PAN L, XIU N, ZHOU S. On solutions of sparsity constrained optimization[J].
Journal of the Operations Research Society of China, 2015, 3(4): 421-439.
DOI: 10.1007/s40305-015-0101-3. |
| [16] |
ROCKAFELLAR R T, WET R J. Variational Analysis[M]. New York: Springer, 1998.
|
2022, Vol. 39

