数学杂志  2018, Vol. 38 Issue (2): 375-380   PDF    
扩展功能
加入收藏夹
复制引文信息
加入引用管理器
Email Alert
RSS
本文作者相关文章
高雷阜
周庆
基于Lasserre松弛的紧约束多项式优化问题逼近界分析
高雷阜, 周庆    
辽宁工程技术大学优化与决策研究所, 辽宁 阜新 123000
摘要:本文研究了紧约束多项式优化问题(POP)的界.利用Lasserre提出的将原紧约束问题转化为多项式平方和(SOS)成立的条件,给出其条件推导SOS式子成立的证明.利用原有逼近界定理,将其进一步转化,获得了新的逼近界定理.新的逼近界定理较原有定理减少了参数,便于计算.
关键词紧约束    多项式优化问题    多项式平方和    逼近界    
APPROXIMATION BOUND ANALYSIS BASED ON THE TIGHT CONSTRAINTS POLYNOMIAL OPTIMIZATION PROBLEMS OF LASSERRE RELAXATION
GAO Lei-fu, ZHOU Qing    
Research Institute of Optimization and Decision, Liaoning Technical University, Fuxin 123000, China
Abstract: In this paper, the boundary of the tight constraints polynomial optimization problems (POP) is studied. The original tight constraints are transformed into sum of squares (SOS) by Lasserre, and its founded conditions are given. We prove that the founded conditions make the SOS feasible. By using the original approximation bound theorem, the new approximation bound theorem is obtained, and the new approximation bound theorem is reduced by the original theorem, which is easy to calculate.
Key words: tight constraints     polynomial optimization problems     sum of squares     approximation bound    
1 引言

多项式优化问题(POP)可描述为目标函数和约束条件都用多项式表示的一类优化问题, 是全局优化中的一个基本而重要的研究对象, 已被广泛应用于信号处理和系统控制等领域.近年来, 这类问题吸引了大批学者的关注, 产生了丰富的研究成果[1, 2]. POP的求解是一个颇具挑战性的问题, 传统的数值求解方法有最速下降法、牛顿法、拟牛顿法、共轭梯度法、信赖域法等[2-8].但POP一般是非凸的, 故用这些方法求出的值不一定是全局最优值, 大多情况下只是局部最优值.

POP的数学形式如下

$ \begin{array}{l} \min \;\;\;\;\;f(x)\\ {\rm{s.t.}}\;\;\;\;\;\;S = \left\{ {x \in {R^n}:{g_i}(x) \ge 0, i = 1, 2, \cdots, m} \right\}. \end{array} $

当约束集合是紧集时, 其退化为紧约束POP.此类问题是NP难的, 不易求解.经典求解采用Lagrange松弛方法, 将其松弛成一个凸优化问题.近年来, Lasserre提出了一种求解POP的全局优化算法[2], 该方法是基于半定规划(SDP)松弛, 能无限逼近问题的全局最优值.目前该方法已广泛应用于多个领域[2-6].此方法被相关学者称为Lasserre松弛方法.

本文主要研究紧约束POP, 基于上述松弛思想, 结合多项式平方和(SOS), 利用Lasserre将原紧约束POP转化为SOS形式及其成立条件, 给出其成立条件推导SOS式成立的证明.在进一步转化为求解另一SOS问题的基础上, 给出当其矩阵特征值的最小值大于0时, 上述条件成立的证明.从而找出其目标函数在约束集合的下界, 通过给出的理论估计出下界与最小值相差的距离, 即逼近界.最后, 文章将原有逼近界定理进一步转化, 给出新的逼近界定理.

2 预备知识
$ \left[x\right]_d^T = \left[{\begin{array}{*{20}{c}} 1&{{x_1}}&\cdots &{{x_n}}&{x_1^2}&\cdots &{x_1^d}& \cdots &{x_n^d} \end{array}}\right],\\ {\left[{{x^d}} \right]^T}= \left[{\begin{array}{*{20}{c}}{x_1^d}&{x_1^{d-1}{x_2}}& \cdots &{x_n^d} \end{array}} \right]. $

$\mathbb{N}$:非负整数集, 对于$x \in$ $\mathbb{R}^n$, $\alpha \in$ $\mathbb{N}^n$, ${x^\alpha } = x_1^{{\alpha _1}} \cdots x_n^{{\alpha _n}}$; $\mathbb{N}_k^{{n}}$:多指标集$\left\{ {} \right.$$\alpha \in$$\mathbb{N}^n$: $\left. {\left| \alpha \right| \le k} \right\}$.

$ f(x) = \sum\limits_{\alpha \in N_{2d}^n} {{f_\alpha }} {x^\alpha },{\left\| f \right\|_G} = {(\sum\limits_{\alpha \in N_{2d}^n} {\rho {{(\alpha )}^{ - 1}}f_\alpha ^2} )^{1/2}},\\ \rho (\alpha ) = \left| {\left\{ {(\beta ,\nu ) \in } \right.} \right._d^n \times _d^n:\left. {\left. {\beta + \nu = \alpha } \right\}} \right|. $

$\Delta \subset \left\{ {1, \cdots, n} \right\}$, $\left| \Delta \right| = 2d$, 若$\Delta = \left\{ {{i_1}, \cdots, {i_{2d}}} \right\}$, 则${x_\Delta } = ({x_{i1}}, \cdots, {x_{i2d}})$, ${\Omega _{2d}} = \left\{ {\Delta \subset \left[n \right]:} \right.$ $\left. {\left| \Delta \right| = 2d} \right\}$. ${L^2}$范数: ${\left\| f \right\|_{{L^2}(S)}}: = {\left( {\int\limits_S {f{{(x)}^2}d\mu (x)} } \right)^{1/2}}$, 与S有关的常量: ${\kappa _{2d}}(S) = \mathop {\min }\limits_{p \in R{{\left[x \right]}_{2d}}} $ $\left\{ {{{\left\| p \right\|}_{{L^2}(S)}}:{{\left\| p \right\|}_G} = 1} \right\}$.

引理2.1[3]  对于一个阶数为$2d$的多项式$f$, 存在一个对称矩阵$F$, 有$f(x) = \left[x \right]_d^TF{\left[x \right]_d}$.

引理2.2[3-5]   $f$是SOS $\Leftrightarrow f(x) = \left[x \right]_d^TF{\left[x \right]_d}$, 其中$F \ge 0$.

3 问题及主要结果

考虑如下POP

$ \begin{equation} \begin{array}{l} \min \;\;\;\;\;f(x),\\ {\rm{s.t.}}\;\;\;\;\;\;S = \left\{ {x \in {R^n}:{g_i}(x) \ge 0, i = 1, 2 \cdots m} \right\}, \end{array} \end{equation} $ (3.1)

其中$S$是紧集, $f, {g_1}, \cdots, {g_m}$的阶数不大于$2d$(称$d$为松弛阶数), (3.1)式不易求解,因此Lasserre将其转化成如下$SOS$形式去找(3.1)式中${f_{\min }}$的下界[2]

$ \begin{equation} \begin{array}{l} {\rm{max}} \;\;\;r,\\ {\rm{s.t.}}\;\;\;\;f - r = {\sigma _0} + {\sigma _1}{g_1} + \cdots + {\sigma _m}{g_m},\\ \;\;\;\;\;\;\;\deg ({\sigma _0}), \deg ({\sigma _1}{g_1}), \cdots, \deg ({\sigma _m}{g_m}) \le 2d,\\ \;\;\;\;\;\;\;\;{\sigma _0}, {\sigma _1}, \cdots, {\sigma _m}\text{是}{\rm{SOS}}, \end{array} \end{equation} $ (3.2)

其中, $\deg $表示多项式的阶数, ${f_{\min }}$(${f_{\max }}$)表示$f$$S$上的最小值(最大值), ${f_{{\rm{sos}}}}$表示(3.2)式中的最优值.由上述知, 对于固定的$d$, 有${f_{{\rm{sos}}}} \le {f_{\min }}$. 当${g_1}, \cdots ,{g_m}$在适当的条件下, 存在 $Q = Q({g_1}, \cdots ,{g_m})$

$ \begin{equation} \begin{array}{l} 1 \le \frac{{{f_{\max }} - {f_{{\rm{sos}}}}}}{{{f_{\max }} - {f_{\min }}}} \le Q. \end{array} \end{equation} $ (3.3)

为了保证(3.2)式成立, 需如下引理.

引理3.1[6]  存在一个对称的正定矩阵$E$及SOS多项式${\sigma _1}, \cdots, {\sigma _m}$, 且$\deg ({\sigma _i}{g_i}) \le 2d, i = 1, \cdots, m$${\sigma _1}{g_1} + \cdots + {\sigma _m}{g_m} = 1- \left[x \right]_d^TE{\left[x \right]_d}.$

定理3.1  这是定理内容.若$S$内部非空, $R\left[x \right]$是实多项式环, $R{\left[x \right]_k}$表示阶数最多是$k$的多项式的子空间, 那么下列叙述等价

(1) 引理3.1成立;

(2) 对于每个$f \in R{\left[x \right]_{2d}}$, (3.2)式都可求;

(3) 若(3.2)式可求, 有$f = - \left[x \right]_d^T{\left[x \right]_d}$.

  (1)$ \Rightarrow $(2)由引理2.1, 有$f(x) = \left[x \right]_d^TF{\left[x \right]_d}$, 其中$F$是对称矩阵.由引理3.1知$E$是正定矩阵, 因此存在足够大的$\lambda $, 且$\lambda > 0$, 使

$ {\sigma _0}(x): = f(x) + \lambda \left[x \right]_d^TE{\left[x \right]_d} = \left[x \right]_d^T(F + \lambda E){\left[x \right]_d} $

是SOS.令$r = - \lambda $, 有${\sigma _0}(x): = f(x) - r\left[x \right]_d^TE{\left[x \right]_d}$.而${\sigma _1}{g_1} + \cdots + {\sigma _m}{g_m} = 1 - \left[x \right]_d^TE{\left[x \right]_d}$.故$1 - ({\sigma _1}{g_1} + \cdots + {\sigma _m}{g_m}) = \left[x \right]_d^TE{\left[x \right]_d}$, 从而${\sigma _0}(x): = f(x) - r\left[{1-({\sigma _1}{g_1} + \cdots + {\sigma _m}{g_m})} \right]$.有$ f-r = {\sigma _0} + \lambda {\sigma _1}{g_1} + \cdots + \lambda {\sigma _m}{g_m}.$故(3.2)式可求.

(2)$ \Rightarrow $(3)显然成立.

(3)$ \Rightarrow $(1)令$f(x) = - \left[x \right]_d^T{\left[x \right]_d}$, 因(3.2)式对$f$可求, 故存在$\hat \gamma $及SOS多项式${\hat \sigma _{0, }}{\hat \sigma _1}, \cdots, {\hat \sigma _m}$, 当$\deg ({\hat \sigma _i}{g_i}) \le 2d, i = 1, \cdots, m$

$ - \left[x \right]_d^T{\left[x \right]_d} -\hat \gamma = {\hat \sigma _0} + {\hat \sigma _1}{g_1} + \cdots + {\hat \sigma _m}{g_m}. $

对于$u \in S$, 有$- \hat r \ge \left[u \right]_d^T{\left[u \right]_d} \ge 1$

$ \frac{1}{{- \hat \gamma }}({\hat \sigma _1}{g_1} + \cdots + {\hat \sigma _m}{g_m}) = 1- \frac{1}{{- \hat \gamma }}(\left[x \right]_d^T{\left[x \right]_d} + {\hat \sigma _0}). $

因此, 引理3.1成立.故定理3.1成立.

在引理3.1中, 多项式${\sigma _1}, \cdots, {\sigma _m}$和正定矩阵$E$可能并不唯一.当${\lambda _{\min }}(E)$越大, 获得的逼近界越好.其中$X$是对称矩阵, ${\lambda _{\max }}(X)$ (${\lambda _{\min }}(X)$)表示$X$的特征值的最大值(最小值), $X > 0$意味着${\lambda _{\min }}(X) > 0$.因此, 尽可能找到${\lambda _{\min }}(E)$的最大值, 进而解决如下SOS问题

$ \begin{equation} \begin{array}{l} \mathop {\max }\limits_{{\sigma _1}, \cdots, {\sigma _m}, E} \;\;{\lambda _{\min }}(E)\\ \;\;\;{\rm{s.t.}}\;\;\;\;{\sigma _1}{g_1} + \cdots + {\sigma _m}{g_m} = 1 - \left[x \right]_d^TE{\left[x \right]_d}\\ \;\;\;\;\;\;\;\;\;\;\;\;{\sigma _0}, {\sigma _1}, \cdots, {\sigma _m}\text{是}{\rm{SOS}}\\ \;\;\;\;\;\;\;\;\;\;\;\;deg ({\sigma _1}{g_1}), \cdots, \deg ({\sigma _m}{g_m}) \le 2d \end{array} \end{equation} $ (3.4)

定理3.2  假设$\left({\sigma _1^ *, \cdots, \sigma _m^ *, {E^ * }} \right)$是(3.4)式的最优解.当且仅当${\lambda _{\min }}({E^ * }) > 0$, 引理3.1才成立.

 因为$\left({\sigma _1^*, \cdots, \sigma _m^*, {E^*}} \right)$是(3.4)式的最优解, 且${\lambda _{\min }}({E^ * }) > 0$, 则$E$是正定矩阵.由${\hat \sigma _0}, {\hat \sigma _1}, \cdots, {\hat \sigma _m}$是SOS且$\deg ({\hat \sigma _i}{g_i}) \le 2d, i = 1, \cdots, m$, 有

$ {\hat \sigma _0} + {\hat \sigma _1}{g_1} + \cdots + {\hat \sigma _m}{g_m} = \left[x \right]_d^TE{\left[x \right]_d}. $

${\hat \sigma _0}(x): = 2\left[x \right]_d^TE{\left[x \right]_d} + \ell $, 其中$0 < - \ell < 2\left[x \right]_d^TE{\left[x \right]_d}$, 则有

$ - \frac{1}{\ell }({\hat \sigma _1}{g_1} + \cdots + {\hat \sigma _m}{g_m}) = 1- (- \frac{1}{\ell })\left[x \right]_d^TE{\left[x \right]_d}. $

故结论得证.

$d = \max \left\{ {\left\lceil {\frac{{\deg (f)}}{2}} \right\rceil, \left\lceil {\frac{{\deg (g)}}{2}} \right\rceil } \right\}$, 其中对于任意的$t \in R$, $\left\lceil t \right\rceil $表示不小于$t$的最小整数, $\psi $是包含$f$$\mathbb{R}{\left[x \right]_{2d}}$的子空间, $p$为一多项式, 定义与$\psi$$S$有关的常量为[9]

$ \begin{equation} \begin{array}{l} \chi \left( {\psi, S} \right): = \mathop {\max }\limits_{p \in \psi } \left\{ {{{\left\| p \right\|}_G}:\left| {p(x) \le 1} \right|, \forall x \in S} \right\} \end{array} \end{equation} $ (3.5)

引理3.2[9]  若$S$内部非空, 对于每个$\Delta \in {\Omega _{2d}}$, 且$0$$S$内部, 则${\kappa _{2d}}(S) > 0$.

引理3.3[10]  如果$f\in$$\mathbb{R}{\left[x \right]_{2d}}$, 那么存在一个对称矩阵$W$, 使得

$ f(x) = \left[x \right]_d^TW{\left[x \right]_d}, \;\;\;{\left\| W \right\|_F} = {\left\| f \right\|_G}. $

其中对于任意矩阵$A$, ${\left\| A \right\|_2}$是标准的二范数, ${\left\| A \right\|_F} = \sqrt {{\rm{trace}}({A^T}A)}$.注意: ${\left\| A \right\|_2} \le {\left\| A \right\|_F}$.

引理3.4[11]  假设$\psi$是包含1的$\mathbb{R}{\left[x \right]_{2d}}$的子空间, $\chi \left({\psi, S} \right) < \infty$, $\left({{\sigma _1}, \cdots, {\sigma _m}, E} \right)$满足引理3.1.令$f\in\psi$, ${f_{\min }}({f_{\max }})$是在$S$上的最小值(最大值).如果${f_{sos}}$是(3.2)式的最优值, 则有

$ 1 \le \frac{{{f_{\max }}-{f_{{\rm{sos}}}}}}{{{f_{\max }}-{f_{\min }}}} \le \frac{{\chi \left( {\psi, S} \right)}}{{{\lambda _{\min }}(E)}}. $

若(3.4)式中的最优解$\left({\sigma _1^ *, \cdots, \sigma _{_m}^ *, {E^ * }} \right)$用到上式, 获得的最优界更好.

为了得到引理3.4的精确的逼近界, 需要估计$\chi (\psi, S)$${\lambda _{\min }}(E)$的值.

定理3.3  假设0在$S$的内部, $({\sigma _1}, \cdots, {\sigma _m}, E)$满足引理3.1, 令$f\in$$\mathbb{R}{\left[x\right]_{2d}}$, 则有

$ 1 \le \frac{{{f_{\max }}-{f_{{\rm{sos}}}}}}{{{f_{\max }}-{f_{\min }}}} \le \frac{1}{{{\kappa _{2d}}(S){\lambda _{\min }}(E)}}. $

 因为$S$内部非空, 由引理3.2知${\kappa _{2d}}(S)>0$.令$\psi=$$\mathbb{R}{\left[x \right]_{2d}}$, 如果$p \in \psi $, 对所有的$x \in S $都有$\left|{p(x)}\right|\le 1$, 则${\left\|p\right\|_{{L^2}(S)}}\le1$.由${\kappa _{2d}}(S) = \mathop {\min }\limits_{p \in R{{\left[x \right]}_{2d}}} \left\{ {{{\left\| p \right\|}_{{L^2}(S)}}:{{\left\| p \right\|}_G} = 1} \right\}$, 有$1 \ge {\kappa _{2d}}(S){\left\| p \right\|_G}$, $\chi (\psi, S) \le \frac{1}{{{\kappa _{2d}}(S)}}$.由引理3.4得

$1 \le \frac{{{f_{\max }} - {f_{{\rm{sos}}}}}}{{{f_{\max }} - {f_{\min }}}} \le \frac{1}{{{\kappa _{2d}}(S){\lambda _{\min }}(E)}},$

故定理得证.

 如果(3.4)式中的最优解$\left({\sigma _1^ *, \cdots, \sigma _m^ *, {E^ * }} \right)$用到上式, 获得的最优界更好.(3.4)式中的最优值${\lambda _{\min }}({E^ * })$的选取依赖于$S$.已有文献给出$\sum\limits_{k = 0}^d {\frac{{{R^{2k}}}}{{k!}}} \le \frac{1}{{{\lambda _{\min }}({E^ * })}} \le d!\sum\limits_{k = 0}^d {\frac{{{R^{2k}}}}{{k!}}} $, 但未给$\sum\limits_{k = 0}^d {\frac{{{R^{2k}}}}{{k!}}} $收敛性, 下面给出具体过程并证明$\sum\limits_{k = 0}^d {\frac{{{R^{2k}}}}{{k!}}} $收敛.

$R = \mathop {\max }\limits_{x \in S} {\left\| x \right\|_2}$, 有

$ \left\| x \right\|_{_2}^{2k} = {(x_1^2 + \cdots + x_n^2)^k} = \sum\limits_{\alpha \in N_k^n} {{x^{2\alpha }}\frac{{k!}}{{{\alpha _1}! \cdots {\alpha _n}!}}} \le k!\sum\limits_{\alpha \in N_k^n} {{x^{2\alpha }}} = k!\left\| {\left[{{x^k}} \right]} \right\|_2^2, $
$ \left\| {{{\left[x \right]}_d}} \right\|_{_2}^2 = 1 + \left\| {\left[{{x^1}} \right]} \right\|_2^2 + \left\| {\left[{{x^2}} \right]} \right\|_2^2 + \cdots + \left\| {\left[{{x^d}} \right]} \right\|_2^2 \ge \sum\limits_{k = 0}^d {\frac{{\left\| x \right\|_2^{2k}}}{{k!}}}, $

对于所有的$x \in S$, 有$1 \ge \left[x \right]_d^T{E^ * }{\left[x \right]_d} \ge {\lambda _{\min }}({E^ * })\left\| {{{\left[x \right]}_d}} \right\|_2^2$, 则上式变为

$ \frac{1}{{{\lambda _{\min }}({E^ * })}} \ge \sum\limits_{k = 0}^d {\frac{{{R^{2k}}}}{{k!}}} . $

$\left\| {\left[{{x^k}} \right]} \right\|_{_2}^2 = \sum\limits_{\alpha \in N_k^n} {{x^{2\alpha }}} \le \sum\limits_{\alpha \in N_k^n} {{x^{2\alpha }}\frac{{k!}}{{{\alpha _1}! \cdots {\alpha _n}!}}} = {(x_1^2 + \cdots + x_n^2)^k} = \left\| x \right\|_2^{2k}$

$ r(x): = 1- {(1 + {R^2} + \cdots + {R^{2d}})^{- 1}}\left[x \right]_d^T{\left[x \right]_d} $

$S$上非负.若存在SOS多项式${s_0}, {s_1}, \cdots, {s_m}$, 对每个$\deg ({s_i}{g_i}) \le 2d$

$ \begin{equation} \begin{array}{l} r = {s_0} + {s_1}{g_1} + \cdots + {s_m}{g_m}. \end{array} \end{equation} $ (3.6)

可找到一正定矩阵$\hat E$满足

$ \left[x \right]_d^T\hat E{\left[x \right]_d} = 1 - ({s_1}{g_1} + \cdots + {s_m}{g_m}) = {s_0} + {(1 + {R^2} + \cdots + {R^{2d}})^{ - 1}}\left[x \right]_d^T{\left[x \right]_d}. $

因为${s_0}$是SOS, 有

$ \frac{1}{{{\lambda _{\min }}({E^ * })}} \le \frac{1}{{{\lambda _{\min }}(\hat E)}} \le \sum\limits_{k = 0}^d {{R^{2k}}} \le d!\sum\limits_{k = 0}^d {\frac{{{R^{2k}}}}{{k!}}} . $

所以, 如果(3.6)式成立, 则有

$ \sum\limits_{k = 0}^d {\frac{{{R^{2k}}}}{{k!}}} \le \frac{1}{{{\lambda _{\min }}({E^ * })}} \le d!\sum\limits_{k = 0}^d {\frac{{{R^{2k}}}}{{k!}}} . $

如果(3.6)式不成立, 但已知$R$, 可以增加(3.1)式的约束条件$r(x) \ge 0$.故$\frac{1}{{{\lambda _{\min }}({E^ * })}}$可以通过$\sum\limits_{k = 0}^d {\frac{{{R^{2k}}}}{{k!}}} $来估计.下证其收敛性.

${U_k} = \frac{{{R^{2k}}}}{{k!}}$, 则${U_1} = \frac{{{R^2}}}{{1!}} = {R^2}$有界, 且

$ \frac{{{U_{k + 1}}}}{{{U_k}}} = \frac{{{\textstyle{{{R^{2(k + 1)}}} \over {(k + 1)!}}}}}{{{\textstyle{{{R^{2k}}} \over {k!}}}}} = \frac{{{R^{2(k + 1)}}}}{{(k + 1)!}} \cdot \frac{{k!}}{{{R^{2k}}}} = \frac{{{R^2}}}{{k + 1}}, $

$ \mathop {\lim }\limits_{k \to \infty } \frac{{{U_{k + 1}}}}{{{U_k}}} = \mathop {\lim }\limits_{k \to \infty } \frac{{{R^2}}}{{k + 1}} = 0 < 1. $

$\sum\limits_{k = 0}^d {{U_k}} $收敛, 即$\sum\limits_{k = 0}^d {\frac{{{R^{2k}}}}{{k!}}} $收敛.

4 结论

紧约束POP是NP难的, 不易求解, 故Lasserre将其转化为SOS形式求目标函数在约束集合的下界.为了使得Lasserre转化的SOS可求, 在已知其成立条件前提下, 进一步给出证明过程, 从而找出所求问题的下界.在已知下界的基础上, 通过理论估计出下界与最小值相差的距离, 即逼近界.将原有的逼近界定理进行转化, 给出另外一个逼近界定理, 从而解决原紧约束POP.

参考文献
[1] 陈德俊. 多项式优化方法在二次规划问题中的应用[D]. 北京: 清华大学, 2011.
[2] Lasserre J B. Global optimization with polynomials and the problem of moments[J]. Soc. Indus. Appl. Math., 2001, 11(3): 796–817.
[3] Prajna S, Papachristodoulou A, Wu F. Nonlinear control synthesis by sum of squares optimization: A Lyapunov-based approach[C]. Proc. Asian Contr. Conf., 2004, 1: 157-165.
[4] Parrilo P A. Semidefinite programming relaxations for semialgebraic problems[J]. Math. Prog., 2003, 96(2): 293–320. DOI:10.1007/s10107-003-0387-5
[5] Laurent M. Sum of squares, moment matrices and optimization over polynomials[J]. Emer. Appl. Alg. Geom., 2008, 149: 157–270.
[6] Lasserre J B. Moments, positive polynomials and their applications[M]. London: Imperial College Press, 2009.
[7] 董丽, 周金川. 无约束优化问题的一个下降方法[J]. 数学杂志, 2015, 35(1): 173–179.
[8] 张新华. 等式约束非凸优化问题的修正牛顿算法[J]. 数学杂志, 2015, 35(1): 1–11.
[9] Barvinok. Eastimating L norms by L2k norms for functions on orbits[J]. Found. Comp. Math., 2002, 2(4): 393–412. DOI:10.1007/s102080010031
[10] Nie J, Demmel J, Sturmfels B. Minimizing polynomials via sum of squares over the gradient ideal[J]. Math. Prog., 2005, 106(3): 587–606.
[11] Reznick B. Some concrete aspects of Hilbert's 17th problem[J]. Contem. Math., Amer. Math. Soc., 2000, 253: 251–272. DOI:10.1090/conm/253