数学物理学报, 2021, 41(3): 882-891 doi:

论文

混水平列扩充设计的混偏差的下界

雷轶菊,1, 欧祖军,2, 赵国喜1

Lower Bounds on Mixture-Discrepancy of the Mixed-Level Column Augmented Design

Lei Yiju,1, Ou Zujun,2, Zhao Guoxi1

收稿日期: 2019-12-30  

基金资助: 国家自然科学基金.  11961027
国家自然科学基金.  11701213
国家自然科学基金.  11871237
国家自然科学基金.  11561025
河南省高等学校重点科研项目.  19A110032
湖南省教育厅重点科研项目.  18A284
湖南省教育厅重点科研项目.  19A403

Received: 2019-12-30  

Fund supported: the NSFC.  11961027
the NSFC.  11701213
the NSFC.  11871237
the NSFC.  11561025
the Key Scientific Research Project of Henan Colleges and Universities.  19A110032
the Key Scientific Research Project of Education Department of Hunan Province.  18A284
the Key Scientific Research Project of Education Department of Hunan Province.  19A403

作者简介 About authors

雷轶菊,E-mail:leiyiju2001@sina.com , E-mail:leiyiju2001@sina.com

欧祖军,E-mail:ozj9325@mail.ccnu.edu.cn , E-mail:ozj9325@mail.ccnu.edu.cn

Abstract

As a new type of experimental designs, the extended designs have been paid more and more attention in recent years. The extended design includes two parts: initial design and follow-up design. In many follow-up designs, some extra factors with two or three levels may be added in the follow-up stage since they are quite important but may be neglected in the initial stage. In this paper, under the uniformity criterion, we present the analytical expressions and the corresponding lower bounds on mixture-discrepancy of the mixed-level column augmented designs. In the sense of mixed discrepancy, mixed-level column augmented nearly uniform designs are proposed.

Keywords: Follow-up design ; Mixture-discrepancy ; Lower bound

PDF (402KB) 元数据 多维度评价 相关文章 导出 EndNote| Ris| Bibtex  收藏本文

本文引用格式

雷轶菊, 欧祖军, 赵国喜. 混水平列扩充设计的混偏差的下界. 数学物理学报[J], 2021, 41(3): 882-891 doi:

Lei Yiju, Ou Zujun, Zhao Guoxi. Lower Bounds on Mixture-Discrepancy of the Mixed-Level Column Augmented Design. Acta Mathematica Scientia[J], 2021, 41(3): 882-891 doi:

1 引言

部分因析设计是工业和科学研究中应用最广泛的试验设计, 其成功之处在于能够在较为经济的试验次数下有效的利用试验同时考查多个因子. 效应别名是部分因析设计不可避免的问题. 由于别名的因子效应会造成数据分析的困扰, 因此如何有效的解除别名效应的模糊性是部分因析设计中的重要问题. 解除效应别名的一般方法是添加更多处理到原设计中去, 即在原设计之后添加跟随设计. 处理的添加会增加试验的成本. 在许多工业试验中, 每个处理的成本都是非常高的, 因此试验者倾向于较小的处理个数. 通常使用包含线性的、二次的因子及因子间两两交互项的二阶模型来考察因子和响应间的非线性关系. 如果从初始设计来看, 解析结果满足了预期, 试验就可终止. 否则, 需要使用跟随设计来搜集进一步的信息. 有时, 一个二阶段试验可能仍然满足不了需要, 那么就需要逐步加入更多的设计点直到满足一定的要求. 这样的试验方法被称为多阶段扩充设计. 在许多跟随设计中, 在跟随阶段可以加入一些另外的2 -水平或3 -水平因子, 因为它们在初始阶段可能被忽略但又十分重要. 本文中, 我们在均匀性准则下, 以具有附加处理和附加多水平因子的2 -阶段的扩充设计为研究对象.

通常, 在跟随设计领域, 研究者仅仅考虑在初始设计中加一些处理而不会去添加任何一个因子(参见文献[1-2]). 他们通常选择一些感兴趣的因子, 因为处理大小的限制而忽略其它可能重要的因子. 然而, 另外的一些因子可以被加到初始设计之后. 初始设计和跟随设计通常在不同时间进行, 配置不同, 操作人员不同. 这就需要对那些讨厌因子测试在不同的阶段之间是否有系统差异, 那么在跟随设计中应加入一个区组因子. 基于上述原因, 试验者可以考虑在跟随阶段加入一些处理和一些因子.

Yang等[3]考虑了不仅加入了一些处理而且加入了一些只取2个不同值的2 -水平因子的扩充设计, 然而它的设计结构对附加的2 -水平因子来说非常特殊, 使得它们完全相关. 与Yang等[3]相比较, Yang等[4]对附加的2 -水平因子做了一些改进, 讨论了加入3 -水平因子的更复杂的情况, 这增加了技术难度. 3 -水平因子值得被考虑, 因为能用它们来考虑基于二阶模型的非线性关系.

区组因子可以被当作一般的2 -水平附加因子来处理,因为在下一阶段,2 -水平、3 -水平和区组因子可以被加入, 这样的跟随设计被叫作混水平列扩充设计, 它们有较广泛的应用. 另外, 所得到的列扩充设计没有任何两个完全相关的因子, 这克服了Yang等[3]中设计的缺陷.

最优设计(参见文献[5-6])被广泛地使用到昂贵且耗时的试验中. 然而, 最优设计需要事先明确潜在的模型. 当因子和响应间的关系未知时, Fang[7]提出的均匀设计是一个有效的试验方法. 均匀试验设计是一种全新的部分因子设计, 被广泛地应用于计算机仿真试验(参见文献[8])和国防、农业、工业、医药和高新技术创新等领域中, 已取得了显著的经济效益和社会效益. 均匀设计要求试验点均匀地分布在试验区域内, 是一种空间填充设计. 当真实的模型信息量比较少时,均匀设计对模型的变化具有稳健性(参见文献[9-10]). Xie和Fang[11]指出均匀设计具有可容许性和最小最大化的特性, 而且均匀设计的处理个数非常灵活, 可以是任一个整数. 均匀设计可以让我们节约试验成本. 进一步, 均匀性准则与广义最小混杂准则之间有密切的联系. 而广义最小混杂准则被广泛地应用在正交设计(参见文献[12-13])理论中. 因此, 考虑用均匀性准则来评价列扩充设计是合理的, 尤其是当真实模型未知时. Yang等[4]以可卷$ L_{2} $ -偏差为均匀性测度, 研究了最优的混水平列扩充设计, 称得到的设计为混水平列扩充均匀设计. Liu等[14]针对Lee偏差得到了2、3混水平列扩充设计的下界. 本文中, 我们讨论在混偏差(MD)测度下的最优混水平列扩充设计. 若混水平列扩充设计的偏差近似达到了偏差下界, 则称其为混水平列扩充近似均匀设计.

论文余下的部分安排如下. 第2节介绍设计$ d $的混偏差(MD)的表达式, 第3节给出具有3 -水平附加因子列扩充设计混偏差的表达式及相应的下界. 第4节给出一个近似达到偏差下界的数值例子, 以说明我们理论结果的正确性.

2 设计d的混偏差的表达式

根据一维的投影均匀性准则, 我们更喜欢限制设计是平衡的, 即限制设计是$ U $ -型设计. 一个非对称的$ U $ -型设计对应于一个$ n\times m $的矩阵$ {\cal X} = (x_1, \cdots, x_m) $, 使得每列$ x_{i} $$ q_{i} $个整数组成的集合$ \{0, 1, 2, \cdots, q_{i}-1\} $中等频率取值, 如果某些$ q_{i} $是相等的, 我们将这个非对称设计用$ U(n;q^{m_{1}}_{1}, \cdots, q^{m_{s}}_{s}) $表示, 这里$ m = \sum\limits_{i = 1}^{s}m_{i} $. 将所有的$ U(n;q^{m_{1}}_{1}, \cdots, q^{m_{s}}_{s}) $$ {\cal U}(n;q^{m_{1}}_{1}, \cdots, q^{m_{s}}_{s}) $表示, 对每个$ d\in{\cal U}(n;q^{m_{1}}_{1}, \cdots, q^{m_{s}}_{s}) $, 将$ d $$ n $个处理通过映射$ f:x_{il}\rightarrow \frac{2x_{il}+1}{2q_{l}} $转化成$ C^{m} = [0, 1]^{m} $中的$ n $个点. 测度均匀性有许多不同的准则, 诸如$ CD, LD, MD $等. 当设计$ d\in{\cal U}(n;q^{m_{1}}_{1}, \cdots, q^{m_{s}}_{s}) $时, $ d $$ MD $ -值的平方为

$ \begin{eqnarray} [MD(d)]^2 & = &\left(\frac{19}{12}\right)^m-\frac{2}{n}\sum\limits_{i = 1}^{n}\prod\limits_{l = 1}^{m} \left(\frac{5}{3}-\frac{1}{4}\left|u_{il}-\frac{1}{2}\right|-\frac{1}{4}\left|u_{il}-\frac{1}{2}\right|^2\right){}\\ &&+\frac{1}{n^2}\sum\limits_{i = 1}^{n}\sum\limits_{j = 1}^{n}\prod\limits_{l = 1}^{m} \bigg(\frac{15}{8}-\frac{1}{4}\left|u_{il}-\frac{1}{2}\right|-\frac{1}{4}\left|u_{jl}-\frac{1}{2}\right| {}\\ &&-\frac{3}{4}\left|u_{il}-u_{jl}\right|+\frac{1}{2}\left|u_{il}-u_{jl}\right|^{2}\bigg), \end{eqnarray} $

其中$ u_{il} = \frac{2x_{il}+1}{2q_{l}} $.

3 具有附加3 -水平因子的列扩充均匀设计

设初始设计$ d_{0}\in{\cal U}(n;2^{m_{1}}3^{m_{2}}) $, $ m = m_{1}+m_{2} $. 在下面的阶段, 我们想加入$ n_{1} $个附加处理和$ r $个附加3 -水平因子. 可以合理地假定下面的每一个阶段不需要扩充太多的处理. 不失一般性, 设$ n_{1}\leq n $, $ 0_{t\times k} $$ 1_{t\times k} $分别是元素全为0和1的矩阵. 一个$ t\times k $的设计矩阵$ d\in\{0, 1, 2, \cdots, q-1\}^{t\times k} $意味着$ d $的每一个元素选自集合$ \{0, 1, 2, \cdots, q-1\} $.

文献[4]中, Yang等定义设计$ D_{3} = \left( \begin{array}{cc} d_{0} & 0_{n\times r} \\ d_{1}& d_{2}\\ \end{array} \right) $是一个具有附加3 -水平因子的列扩充设计, 如果初始设计$ d_{0}\in{\cal U}(n;2^{m_{1}}3^{m_{2}}) $$ d_{1}\in{\cal U}(n_{1};2^{m_{1}}3^{m_{2}}) $$ d_{2}\in\{1, 2\}^{n_{1}\times r} $扩充. 将所有这样的列扩充设计用$ C_{3}(n+n_{1};2^{m_{1}}3^{m_{2}}\cdot3^{r}) $表示. $ r $个附加3 -水平因子可以是量化因子或质量因子, 甚至可以是区组因子. 如果$ r = 0 $, 列扩充设计就退化为文献[3]中定义的行扩充设计.

在初始设计中, 最初被忽略的因子$ \{x_{m+1}, \cdots, x_{m+r}\} $的水平通常是固定的. 不失一般性, 在初始设计中, 假定附加因子$ \{x_{m+1}, \cdots, x_{m+r}\} $的所有水平为0, 那么初始设计可以用具有$ m+r $个因子的$ (d_{0}, 0_{n\times r}) $表示, 下一阶段的设计矩阵是$ (d_{1}, d_{2}) $, $ d_{1} $$ d_{2} $分别是对应于初始的$ m $个因子和附加的$ r $个因子的第二阶段的设计矩阵. 当$ n_{1}\leq n $时, 为了使列扩充设计尽可能均匀, 附加部分$ d_{1} $应该被限制为$ U $ -型设计. $ d_{2} $$ r $列的每列取1和2的个数相同, 因为每列的元素越平衡, 设计就能越均匀. 文献[4]中指出附加处理$ n_{1} $的数值应满足以下条件.

表 1   n1m1, m2r应满足的条件

n1m1m2r
2|n1m2=0r≥0
3|n1m1=0r=0
6|n1m1=0r>0
m1>0m2>0r0

新窗口打开| 下载CSV


与其它类型的跟随设计相比较, 比如折叠与半折叠设计, 分别限制$ n_{1} = n $$ n_{1} = \frac{n}{2} $, 我们对附加处理个数的要求更灵活. 而且, 注意到, $ d_{1} $是一个$ U $ -型设计的限制也可去掉, 即基于初始设计, 可以扩充任意$ n_{1} $个处理. 本文中, 我们考虑$ d_{1} $是一个$ U $ -型设计的情况.

对一个给定的初始设计$ d_{0} $, 有很多可选择的列扩充设计. 在均匀性准则下, 人们更趋向于添加跟随部分$ (d_{1}, d_{2}) $, 使得具有$ n+n_{1} $个点的列扩充设计$ D_{3} $尽可能均匀. 根据$ MD $ -值的平方的表达式, 我们可以得到列扩充设计$ D_{3}\in C_{3}(n+n_{1};2^{m_{1}}3^{m_{2}}\cdot3^{r}) $$ MD $ -值的平方为

$ \begin{eqnarray} [MD(D_{3})]^2 & = &\left(\frac{19}{12}\right)^{m+r}-\frac{2}{n+n_{1}}\sum\limits_{i = 1}^{n+n_{1}}\prod\limits_{l = 1}^{m+r} \left(\frac{5}{3}-\frac{1}{4}\left|u_{il}-\frac{1}{2}\right|-\frac{1}{4}\left|u_{il}-\frac{1}{2}\right|^2\right){}\\ &&+\frac{1}{(n+n_{1})^2}\sum\limits_{i = 1}^{n+n_{1}}\sum\limits_{j = 1}^{n+n_{1}}\prod\limits_{l = 1}^{m+r} \bigg(\frac{15}{8}-\frac{1}{4}\left|u_{il}-\frac{1}{2}\right|-\frac{1}{4}\left|u_{jl}-\frac{1}{2}\right| {}\\ && -\frac{3}{4}\left|u_{il}-u_{jl}\right|+\frac{1}{2}\left|u_{il}-u_{jl}\right|^{2}\bigg), \end{eqnarray} $

这里$ n+n_{1} $是列扩充设计总共的处理个数, $ m+r $是所有的因子个数.

在实践中, 当因子和响应之间的关系未知时, 我们经常选择一个均匀设计$ d_{0} $作初始设计, 而且可以在第一阶段以数据分析为基础终止试验. 如果更多的处理和因子应该被加到$ d_{0} $后, 我们应该搜寻跟随部分$ (d_{1}, d_{2}) $, 使得列扩充设计$ D_{3} = \Bigg( \begin{array}{cc} d_{0}\ & 0_{n\times r} \\ d_{1}\ & d_{2}\\ \end{array} \Bigg) $尽可能均匀. 而且初始设计在跟随阶段中被假定是已知的, 因此, 有必要求出以$ d_{0} $为基础的列扩充设计的$ MD $ -值的平方的表达式. 根据任意两行之间的相同数的个数, 我们可以把(3.1)式中的$ MD $ -值的平方重新表示.

定理 3.1  对于一个给定的初始设计$ d_{0}\in{\cal U}(n;2^{m_{1}}3^{m_{2}}) $, 它的列扩充设计$ D_{3}\in C_{3}(n+n_{1};2^{m_{1}}3^{m_{2}}\cdot3^{r}) $$ MD $ -值的平方为

$ \begin{eqnarray} [MD(D_{3})]^2 & = &C(r)+\frac{n^{2}}{(n+n_{1})^2}\left(\frac{41}{24}\right)^{r}[MD(d_{0})]^2{}\\ &&+\left[\frac{2n}{(n+n_{1})^2}\left(\frac{41}{24}\right)^{r}-\frac{2}{n+n_{1}}\left(\frac{14}{9}\right)^{r}\right]\left(\frac{305}{192}\right)^{m_{1}}\left(\frac{14}{9}\right)^{m_{2}} \sum\limits_{i = 1}^{n}\left(\frac{15}{14}\right)^{\lambda_{ii}^{m_{2}}}{}\\ &&+\frac{1}{(n+n_{1})^2}\left(\frac{7}{4}\right)^{m_{1}}\left(\frac{41}{24}\right)^{m_{2}+r}\sum\limits_{i = n+1}^{n+n_{1}} \left(\frac{45}{41}\right)^{\lambda_{ii}^{m_{2}}+\lambda_{ii}^{r}}{}\\ &&-\frac{2}{n+n_{1}}\left(\frac{305}{192}\right)^{m_{1}}\left(\frac{14}{9}\right)^{m_{2}+r}\sum\limits_{i = n+1}^{n+n_{1}}\left(\frac{15}{14}\right)^{\lambda_{ii}^{m_{2}}+\lambda_{ii}^{r}} {}\\ &&+\frac{2}{(n+n_{1})^2}\left(\frac{3}{2}\right)^{m_{1}}\left(\frac{41}{24}\right)^{m_{2}}\left(\frac{115}{72}\right)^{r}\sum\limits_{i = 1}^{n} \sum\limits_{j = n+1}^{n+n_{1}}\left(\frac{7}{6}\right)^{C_{ij}^{m_{1}}}{}\\ &&\times\left(\frac{45}{41}\right)^{\lambda_{ij}^{m_{2}}}\left(\frac{115}{123}\right)^{Q_{ij}^{m_{2}}}\left(\frac{103}{123}\right)^{\sigma_{ij}^{m_{2}}}\left(\frac{103}{115}\right)^{\sigma_{ij}^{r}}{}\\ &&+\frac{1}{(n+n_{1})^2}\left(\frac{3}{2}\right)^{m_{1}}\left(\frac{41}{24}\right)^{m_{2}+r} \sum\limits_{i = n+1}^{n+n_{1}}\sum\limits_{j(\neq i) = n+1}^{n+n_{1}}\left(\frac{7}{6}\right)^{C_{ij}^{m_{1}}} {}\\ && \times\left(\frac{45}{41}\right)^{\lambda_{ij}^{m_{2}}+\lambda_{ij}^{r}}\left(\frac{115}{123}\right)^{Q_{ij}^{m_{2}}+Q_{ij}^{r}}\left(\frac{103}{123}\right)^{\sigma_{ij}^{m_{2}}}{}\\ &&+\frac{1}{(n+n_{1})^2}\left(\frac{3}{2}\right)^{m_{1}}\left(\frac{41}{24}\right)^{m_{2}+r} \sum\limits_{i = n+1}^{n+n_{1}}\sum\limits_{j(\neq i) = n+1}^{n+n_{1}}\left(\frac{7}{6}\right)^{C_{ij}^{m_{1}}}{}\\ &&\times \left(\frac{45}{41}\right)^{\lambda_{ij}^{m_{2}}+\lambda_{ij}^{r}}\left(\frac{115}{123}\right)^{Q_{ij}^{m_{2}}+Q_{ij}^{r}}\left(\frac{103}{123}\right)^{\sigma_{ij}^{m_{2}}}, \end{eqnarray} $

这里$ C(r) = \left(\left(\frac{19}{12}\right)^{r}-\frac{n^{2}}{(n+n_{1})^2}\left(\frac{41}{24}\right)^{r}\right) \left(\frac{19}{12}\right)^{m} $, $ m = m_{1}+m_{2} $, $ \lambda_{ij}^{r} = \sharp\{l:x_{il} = x_{jl} = 1, $$ m+1\leq l\leq m+r\} $, $ \lambda_{ij}^{m_{2}} = \sharp\{l:x_{il} = x_{jl} = 1, m_{1}+1\leq l\leq m\} $, $ Q_{ij}^{r} = \sharp\{l:(x_{il}, x_{jl}) = (1, 2), (2, 1), m+1\leq l\leq m+r\} $, $ \sigma_{ij}^{r} = \sharp\{l:(x_{il}, x_{jl}) = (0, 2), m+1\leq l\leq m+r\} $, $ C_{ij}^{m_{1}} = \sharp\{l:x_{il} = x_{jl}, 1\leq l\leq m_{1}\} $, $ Q_{ij}^{m_{2}} = \sharp\{l:(x_{il}, x_{jl}) = (0, 1), (1, 0), (1, 2)(2, 1), m_{1}+1\leq l\leq m\} $, $ \sigma_{ij}^{m_{2}} = \sharp\{l:(x_{il}, x_{jl}) = (0, 2), (2, 0), m_{1}+1\leq l\leq m\} $, $ \sharp\{S\} $表示集合$ S $中元素的个数.

  设

$ \begin{eqnarray} [MD(D_{3})]^2 & = &\left(\frac{19}{12}\right)^{m+r}-\frac{2}{n+n_{1}}\bigg(\sum\limits_{i = 1}^{n}\prod\limits_{l = 1}^{m+r} \triangle_{il}+\sum\limits_{i = n+1}^{n+n_{1}}\prod\limits_{l = 1}^{m+r}\triangle_{il}\bigg) {}\\ &&+\frac{1}{(n+n_{1})^2}\bigg(\sum\limits_{i = 1}^{n}\sum\limits_{j = 1}^{n}\prod\limits_{l = 1}^{m+r}\triangle_{ijl}+2\sum\limits_{i = 1}^{n}\sum\limits_{j = n+1}^{n+n_{1}}\prod\limits_{l = 1}^{m+r}\triangle_{ijl} +\sum\limits_{i = n+1}^{n+n_{1}}\sum\limits_{j = n+1}^{n+n_{1}}\prod\limits_{l = 1}^{m+r}\triangle_{ijl}\bigg){}\\ & = &\left(\frac{19}{12}\right)^{m+r}-\frac{2}{n+n_{1}}\left(\frac{305}{192}\right)^{m_{1}}\left(\frac{14}{9}\right)^{r}\sum\limits_{i = 1}^{n}\left(\frac{14}{9}\right)^{m_{2}-\lambda_{ii}^{m_{2}}}\left(\frac{5}{3}\right)^{\lambda_{ii}^{m_{2}}} {}\\ &&-\frac{2}{n+n_{1}}\left(\frac{305}{192}\right)^{m_{1}}\sum\limits_{i = n+1}^{n+n_{1}}\left(\frac{14}{9}\right)^{m_{2} -\lambda_{ii}^{m_{2}}}\left(\frac{5}{3}\right)^{\lambda_{ii}^{m_{2}}} \left(\frac{14}{9}\right)^{r-\lambda_{ii}^{r}}\left(\frac{5}{3}\right)^{\lambda_{ii}^{r}} {}\\ &&+\frac{1}{(n+n_{1})^2}\sum\limits_{i = 1}^{n}\sum\limits_{j = 1}^{n}\left(\frac{41}{24}\right)^{r}\prod\limits_{l = 1}^{m}\triangle_{ijl} {}\\ &&+\frac{2}{(n+n_{1})^2}\sum\limits_{i = 1}^{n}\sum\limits_{j = n+1}^{n+n_{1}}\left(\frac{7}{4}\right)^{C_{ij}^{m_{1}}}\left(\frac{3}{2}\right)^{m_{1}-C_{ij}^{m_{1}}} \left(\frac{41}{24}\right)^{m_{2}-\lambda_{ij}^{m_{2}}-Q_{ij}^{m_{2}}-\sigma_{ij}^{m_{2}}} {}\\ &&\times\left(\frac{15}{8}\right)^{\lambda_{ij}^{m_{2}}} \left(\frac{115}{72}\right)^{ Q_{ij}^{m_{2}}+r-\sigma_{ij}^{r}}\left(\frac{103}{72}\right)^{\sigma_{ij}^{m_{2}}+\sigma_{ij}^{r}} {}\\ &&+\frac{1}{(n+n_{1})^2}\sum\limits_{i = n+1}^{n+n_{1}}\sum\limits_{j(\neq i) = n+1}^{n+n_{1}}\left(\frac{7}{4}\right)^{C_{ij}^{m_{1}}}\left(\frac{3}{2}\right)^{m_{1}-C_{ij}^{m_{1}}}\left(\frac{41}{24}\right)^{m_{2}-\lambda_{ij}^{m_{2}}-Q_{ij}^{m_{2}}-\sigma_{ij}^{m_{2}}} {}\\ &&\times\left(\frac{115}{72}\right)^{Q_{ij}^{m_{2}}} \left(\frac{103}{72}\right)^{\sigma_{ij}^{m_{2}}}\left(\frac{15}{8}\right)^{\lambda_{ij}^{m_{2}}}\left(\frac{41}{24}\right)^ {r-Q_{ij}^{r}-\lambda_{ij}^{r}}\left(\frac{115}{72}\right)^{Q_{ij}^{r}}\left(\frac{15}{8}\right)^{\lambda_{ij}^{r}}{}\\ &&+\frac{1}{(n+n_{1})^2}\left(\frac{7}{4}\right)^{m_{1}}\left(\frac{15}{8}\right)^{m_{2}+r}\sum\limits_{i = n+1}^{n+n_{1}}\left(\frac{41}{45}\right)^{m_{2}+r-\lambda_{ii}^{m_{2}}-\lambda_{ii}^{r}} {}\\ &&\times \left(\frac{23}{27}\right)^{Q_{ii}^{m_{2}}+Q_{ii}^{r}}\left(\frac{103}{135}\right)^{\sigma_{ii}^{m_{2}}}{}\\ & = &C(r)+\frac{n^{2}}{(n+n_{1})^2}\left(\frac{41}{24}\right)^{r}[MD(d_{0})]^2{}\\ &&+\left[\frac{2n}{(n+n_{1})^2}\left(\frac{41}{24}\right)^{r}-\frac{2}{n+n_{1}}\left(\frac{14}{9}\right)^{r}\right] \left(\frac{305}{192}\right)^{m_{1}}\left(\frac{14}{9}\right)^{m_{2}} \sum\limits_{i = 1}^{n}\left(\frac{15}{14}\right)^{\lambda_{ii}^{m_{2}}}{}\\ &&+\frac{1}{(n+n_{1})^2}\left(\frac{7}{4}\right)^{m_{1}}\left(\frac{41}{24}\right)^{m_{2}+r}\sum\limits_{i = n+1}^{n+n_{1}} \left(\frac{45}{41}\right)^{\lambda_{ii}^{m_{2}}+\lambda_{ii}^{r}}{}\\ &&-\frac{2}{n+n_{1}}\left(\frac{305}{192}\right)^{m_{1}}\left(\frac{14}{9}\right)^{m_{2}+r}\sum\limits_{i = n+1}^{n+n_{1}}\left(\frac{15}{14}\right)^{\lambda_{ii}^{m_{2}}+\lambda_{ii}^{r}}{}\\ &&+\frac{2}{(n+n_{1})^2}\left(\frac{3}{2}\right)^{m_{1}}\left(\frac{41}{24}\right)^{m_{2}}\left(\frac{115}{72}\right)^{r}\sum\limits_{i = 1}^{n} \sum\limits_{j = n+1}^{n+n_{1}}\left(\frac{7}{6}\right)^{C_{ij}^{m_{1}}}{}\\ &&\times\left(\frac{45}{41}\right)^{\lambda_{ij}^{m_{2}}}\left(\frac{115}{123}\right)^{Q_{ij}^{m_{2}}}\left(\frac{103}{123}\right)^{\sigma_{ij}^{m_{2}}}\left(\frac{103}{115}\right)^{\sigma_{ij}^{r}}{}\\ &&+\frac{1}{(n+n_{1})^2}\left(\frac{3}{2}\right)^{m_{1}}\left(\frac{41}{24}\right)^{m_{2}+r} \sum\limits_{i = n+1}^{n+n_{1}}\sum\limits_{j(\neq i) = n+1}^{n+n_{1}}\left(\frac{7}{6}\right)^{C_{ij}^{m_{1}}} {}\\ &&\times\left(\frac{45}{41}\right)^{\lambda_{ij}^{m_{2}}+\lambda_{ij}^{r}}\left(\frac{115}{123}\right)^{Q_{ij}^{m_{2}}+Q_{ij}^{r}} \left(\frac{103}{123}\right)^{\sigma_{ij}^{m_{2}}}. \end{eqnarray} $

证毕.

为了得到列扩充设计$ D_{3} $的混偏差的下界, 我们首先给出2个引理.

引理 3.1  对任一个列扩充设计$ D_{3}\in C_{3}(n+n_{1};2^{m_{1}}3^{m_{2}}\cdot3^{r}), $

$ \rm(1) $$ \sum\limits_{i = 1}^{n}\sum\limits_{j(\neq i) = 1}^{n}\lambda_{ij}^{m_{2}} = \frac{m_{2}n(n-3)}{9} $, $ \sum\limits_{i = 1}^{n}\lambda_{ii}^{m_{2}} = \frac{m_{2}n}{3} $;

$ \rm(2) $$ \sum\limits_{i = 1}^{n}\sum\limits_{j = n+1}^{n+n_{1}}\lambda_{ij}^{m_{2}} = \frac{m_{2}nn_{1}}{9} $;

$ \rm(3) $$ \sum\limits_{i = n+1}^{n+n_{1}}\sum\limits_{j(\neq i) = n+1}^{n+n_{1}}\lambda_{ij}^{m_{2}} = \frac{m_{2}n_{1}(n_{1}-3)}{9} $, $ \sum\limits_{i = n+1}^{n+n_{1}}\lambda_{ii}^{m_{2}} = \frac{m_{2}n_{1}}{3} $;

$ \rm(4) $$ \sum\limits_{i = n+1}^{n+n_{1}}\lambda_{ii}^{r} = \frac{n_{1}r}{2} $, $ \sum\limits_{i = n+1}^{n+n_{1}}\sum\limits_{j(\neq i) = n+1}^{n+n_{1}}\lambda_{ij}^{r} = \frac{rn_{1}(n_{1}-2)}{4} $;

$ \rm(5) $$ \sum\limits_{i = 1}^{n}\sum\limits_{j = n+1}^{n+n_{1}}Q_{ij}^{m_{2}} = \frac{4m_{2}nn_{1}}{9} $;

$ \rm(6) $$ \sum\limits_{i = n+1}^{n+n_{1}}\sum\limits_{j(\neq i) = n+1}^{n+n_{1}}Q_{ij}^{m_{2}} = \frac{4m_{2}n_{1}^{2}}{9} $, $ Q_{ii}^{m_{2}} = 0 $;

$ \rm(7) $$ \sum\limits_{i = n+1}^{n+n_{1}}\sum\limits_{j(\neq i) = n+1}^{n+n_{1}}Q_{ij}^{r} = \frac{rn_{1}^{2}}{2} $, $ Q_{ii}^{r} = 0 $;

$ \rm(8) $$ \sum\limits_{i = 1}^{n}\sum\limits_{j = n+1}^{n+n_{1}}\sigma_{ij}^{m_{2}} = \frac{2m_{2}nn_{1}}{9} $;

$ \rm(9) $$ \sum\limits_{i = n+1}^{n+n_{1}}\sum\limits_{j(\neq i) = n+1}^{n+n_{1}}\sigma_{ij}^{m_{2}} = \frac{2m_{2}n_{1}^{2}}{9} $, $ \sigma_{ii}^{m_{2}} = 0 $;

$ \rm(10) $$ \sum\limits_{i = 1}^{n}\sum\limits_{j = n+1}^{n+n_{1}}\sigma_{ij}^{r} = \frac{rnn_{1}}{2} $.

引理 3.2[15]  设$ x_{1}, \cdots, x_{n} $$ n $个非负整数, 且$ \sum\limits_{i = 1}^{n}x_{i} = c $, 则对任一个正数$ t $, 有

这里$ \omega = \lfloor\frac{c}{n}\rfloor $, $ p', q' $是使得$ p'+q' = n $, $ p'\omega+q'(\omega+1) = c $的整数.

列扩充设计$ D_{3} $$ MD $值是$ MD(d_{0}) $的一个函数. 可以利用方程(3)获得列扩充设计的$ MD $值的下界, 这个下界可以作为判断一个设计是否均匀的基准. 如果一个设计的$ MD $值达到下界, 那么在设计空间中这个设计必须有最小的$ MD $值, 它是一个均匀设计. 若混水平列扩充设计的偏差近似达到了偏差下界, 则称其为混水平列扩充近似均匀设计.

定理 3.2  对一个给定的初始设计$ d_{0}\in {\cal U}(n;2^{m_{1}}3^{m_{2}}) $, 它的列扩充设计$ D_{3}\in C_{3}(n+n_{1};2^{m_{1}}3^{m_{2}}\cdot3^{r}) $, 令$ f(x) = \left(\frac{7}{4}\right)^{m_{1}}\left(\frac{41}{24}\right)^{m_{2}+r}\frac{4}{41}(\frac{45}{41})^{x} -\frac{(n+n_{1})}{7}\left(\frac{305}{192}\right)^{m_{1}}(\frac{14}{9})^{m_{2}+r}(\frac{15}{14})^{x} $, 如果$ f(\mu_{1})\geq f(0) $, 则有$ \left[MD(D_{3})\right]^{2}\geq LBM_{1} $, 这里

$ \begin{eqnarray} LBM_{1}& = &C(r)+\frac{n^{2}}{(n+n_{1})^2}\left(\frac{41}{24}\right)^{r}[MD(d_{0})]^2{}\\ &&+\left[\frac{2n}{(n+n_{1})^2}\left(\frac{41}{24}\right)^{r}-\frac{2}{n+n_{1}}\left(\frac{14}{9}\right)^{r}\right] {}\\ &&\times\left(\frac{305}{192}\right)^{m_{1}} \left(\frac{14}{9}\right)^{m_{2}}\left[p_{2}\left(\frac{15}{14}\right)^{\mu}+q_{2}\left(\frac{15}{14}\right)^{\mu+1}\right]{}\\ &&+\frac{1}{(n+n_{1})^2}\left(\frac{7}{4}\right)^{m_{1}}\left(\frac{41}{24}\right)^{m_{2}+r}\left[p_{3}\left(\frac{45}{41}\right)^{\mu_{1}}+q_{3}\left(\frac{45}{41}\right)^{\mu_{1}+1}\right]{}\\ &&-\frac{2}{n+n_{1}}\left(\frac{305}{192}\right)^{m_{1}}\left(\frac{14}{9}\right)^{m_{2}+r}\left[p_{3}\left(\frac{15}{14}\right)^{\mu_{1}}+q_{3}\left(\frac{15}{14}\right)^{\mu_{1}+1}\right]{}\\ &&+\frac{2}{(n+n_{1})^2}\left(\frac{3}{2}\right)^{m_{1}}\left(\frac{41}{24}\right)^{m_{2}}\left(\frac{115}{72}\right)^{r}T_{1}+\frac{1}{(n+n_{1})^2}\left(\frac{3}{2}\right)^{m_{1}}\left(\frac{41}{24}\right)^{m_{2}+r}T_{2}, {\qquad} \end{eqnarray} $

其中, $ \mu = \lfloor\frac{m_{2}}{3}\rfloor $, $ \mu_{1} = \lfloor\frac{m_{2}}{3}+\frac{r}{2}\rfloor $, $ p_{2}, q_{2} $是使得$ p_{2}+q_{2} = n $, $ p_{2}\mu+q_{2}(\mu+1) = \frac{m_{2}n}{3} $的整数, $ p_{3}, q_{3} $是使得$ p_{3}+q_{3} = n_{1} $, $ p_{3}\mu_{1}+q_{3}(\mu_{1}+1) = \frac{m_{2}n_{1}}{3}+\frac{rn_{1}}{2} $的整数, $ a_{2} = \ln\frac{7}{6} $, $ b_{2} = \ln\frac{45}{41} $, $ g = \ln\frac{115}{123} $, $ h = \ln\frac{103}{123} $, $ k = \ln\frac{103}{115} $, $ \varphi_{1} = a_{2}\cdot\frac{m_{1}}{2}+b_{2}\cdot\frac{m_{2}}{9}+g\cdot\frac{4m_{2}}{9}+h\cdot\frac{2m_{2}}{9}+k\cdot\frac{r}{2} $, $ T_{1} = nn_{1}e^{\varphi_{1}} $, $ \varphi_{2} = a_{2}\cdot\frac{m_{1}(n_{1}-2)}{2(n_{1}-1)}+b_{2}\left(\frac{m_{2}(n_{1}-3)}{9(n_{1}-1)} +\frac{r(n_{1}-2)}{4(n_{1}-1)}\right) +g\left(\frac{4m_{2}n_{1}}{9(n_{1}-1)}+\frac{rn_{1}}{2(n_{1}-1)}\right)+h\cdot\frac{2m_{2}n_{1}}{9(n_{1}-1)} $, $ T_{2} = n_{1}(n_{1}-1)e^{\varphi_{2}} $.

对初始的混水平设计, 列扩充设计$ D_{3} $可以根据实际要求被简化成初始设计为对称的2 -水平或3 -水平设计的列扩充设计, 即$ C_{3}(n+n_{1};2^{m_{1}}3^{m_{2}}\cdot3^{r}) $变成$ C_{3}(n+n_{1};2^{m}\cdot3^{r}) $$ m_{2} = 0 $时; 或$ C_{3}(n+n_{1};3^{m}\cdot3^{r}) $$ m_{1} = 0 $时. 对这种情况, 我们可以得到更精确的下界.

定理 3.3  对一个给定的初始设计$ d_{0}\in {\cal U}(n;2^{m}) $ (此时$ m_{2} = 0, m_{1} = m $), 它的列扩充设计$ D_{3}\in C_{3}(n+n_{1};2^{m}\cdot3^{r}) $, 令$ f^{*}(x) = \left(\frac{7}{4}\right)^{m}\left(\frac{41}{24}\right)^{r}\frac{4}{41}(\frac{45}{41})^{x} -\frac{(n+n_{1})}{7}\left(\frac{305}{192}\right)^{m}(\frac{14}{9})^{r}(\frac{15}{14})^{x} $, 如果$ f^{*}(\omega_{1}')\geq f^{*}(0) $, 则有$ \left[MD(D_{3})\right]^{2}\geq LBM_{2} $, 这里

$ \begin{eqnarray} LBM_{2}& = &C(r)+\frac{n^{2}}{(n+n_{1})^2}\left(\frac{41}{24}\right)^{r}[MD(d_{0})]^2{}\\ &&+\left[\frac{2n^{2}}{(n+n_{1})^2}\left(\frac{41}{24}\right)^{r}-\frac{2n}{n+n_{1}}\left(\frac{14}{9}\right)^{r}\right]\left(\frac{305}{192}\right)^{m}{}\\ &&+\frac{1}{(n+n_{1})^2}\left(\frac{7}{4}\right)^{m}\left(\frac{41}{24}\right)^{r}\left[p_{1}'\left(\frac{45}{41}\right)^{\omega_{1}'}+q_{1}'\left(\frac{45}{41}\right)^{\omega_{1}'+1}\right]{}\\ &&-\frac{2}{n+n_{1}}\left(\frac{305}{192}\right)^{m}\left(\frac{14}{9}\right)^{r}\left[p_{1}'\left(\frac{15}{14}\right)^{\omega_{1}'}+q_{1}'\left(\frac{15}{14}\right)^{\omega_{1}'+1}\right]{}\\ &&+\frac{2}{(n+n_{1})^2}\left(\frac{3}{2}\right)^{m}\left(\frac{115}{72}\right)^{r}T_{1}'+\frac{1}{(n+n_{1})^2}\left(\frac{3}{2}\right)^{m}\left(\frac{41}{24}\right)^{r}T_{2}', \end{eqnarray} $

其中, $ \omega_{1}' = \lfloor\frac{r}{2}\rfloor $, $ p_{1}', q_{1}' $是使得$ p_{1}'+q_{1}' = n_{1} $, $ p_{1}'\omega_{1}'+q_{1}'(\omega_{1}'+1) = \frac{rn_{1}}{2} $的整数, $ a_{2} = \ln\frac{7}{6} $, $ b_{2} = \ln\frac{45}{41} $, $ g = \ln\frac{115}{123} $, $ k = \ln\frac{103}{115} $, $ \varphi_{1}' = a_{2}\cdot\frac{m}{2}+k\cdot\frac{r}{2} $, $ T_{1}' = nn_{1}e^{\varphi_{1}'} $, $ \varphi_{2}' = a_{2}\cdot\frac{m(n_{1}-2)}{2(n_{1}-1)}+b_{2}\cdot\frac{r(n_{1}-2)}{4(n_{1}-1)}+g\cdot\frac{rn_{1}}{2(n_{1}-1)} $, $ T_{2}' = n_{1}(n_{1}-1)e^{\varphi_{2}'} $.

定理 3.4  对一个给定的初始设计$ d_{0}\in {\cal U}(n;3^{m}) $(此时$ m_{1} = 0, m_{2} = m $), 它的列扩充设计$ D_{3}\in C_{3}(n+n_{1};3^{m}\cdot3^{r}) $, 令$ f^{**}(x) = \left(\frac{41}{24}\right)^{m+r}\frac{4}{41}(\frac{45}{41})^{x} -\frac{(n+n_{1})}{7}(\frac{14}{9})^{m+r}(\frac{15}{14})^{x} $, 如果$ f^{**}(\mu_{1})\geq f^{**}(0) $, 则有$ \left[MD(D_{3})\right]^{2}\geq LBM_{3} $, 这里

$ \begin{eqnarray} LBM_{3}& = &C(r)+\frac{n^{2}}{(n+n_{1})^2}\left(\frac{41}{24}\right)^{r}[MD(d_{0})]^2{}\\ &&+\left[\frac{2n}{(n+n_{1})^2}\left(\frac{41}{24}\right)^{r}-\frac{2}{n+n_{1}}\left(\frac{14}{9}\right)^{r}\right] \left(\frac{14}{9}\right)^{m}\left[p_{2}\left(\frac{15}{14}\right)^{\mu}+q_{2}\left(\frac{15}{14}\right)^{\mu+1}\right]{}\\ &&+\frac{1}{(n+n_{1})^2}\left(\frac{41}{24}\right)^{m+r}\left[p_{3}\left(\frac{45}{41}\right)^{\mu_{1}}+q_{3}\left(\frac{45}{41}\right)^{\mu_{1}+1}\right]{}\\ &&-\frac{2}{n+n_{1}}\left(\frac{14}{9}\right)^{m+r}\left[p_{3}\left(\frac{15}{14}\right)^{\mu_{1}}+q_{3}\left(\frac{15}{14}\right)^{\mu_{1}+1}\right]{}\\ &&+\frac{2}{(n+n_{1})^2}\left(\frac{41}{24}\right)^{m}\left(\frac{115}{72}\right)^{r}T_{1}''+\frac{1}{(n+n_{1})^2}\left(\frac{41}{24}\right)^{m+r}T_{2}'', \end{eqnarray} $

其中, $ \mu = \lfloor\frac{m}{3}\rfloor $, $ \mu_{1} = \lfloor\frac{m}{3}+\frac{r}{2}\rfloor $, $ p_{2}, q_{2} $是使得$ p_{2}+q_{2} = n $, $ p_{2}\mu+q_{2}(\mu+1) = \frac{mn}{3} $的整数, $ p_{3}, q_{3} $是使得$ p_{3}+q_{3} = n_{1} $, $ p_{3}\mu_{1}+q_{3}(\mu_{1}+1) = \frac{mn_{1}}{3}+\frac{rn_{1}}{2} $的整数, $ a_{2} = \ln\frac{7}{6} $, $ b_{2} = \ln\frac{45}{41} $, $ g = \ln\frac{115}{123} $, $ h = \ln\frac{103}{123} $, $ k = \ln\frac{103}{115} $, $ \varphi_{1}'' = b_{2}\cdot\frac{m}{9}+g\cdot\frac{4m}{9}+h\cdot\frac{2m}{9}+k\cdot\frac{r}{2} $, $ T_{1}'' = nn_{1}e^{\varphi_{1}''} $, $ \varphi_{2}'' = b_{2}\left(\frac{m(n_{1}-3)}{9(n_{1}-1)}+\frac{r(n_{1}-2)}{4(n_{1}-1)}\right) +g\left(\frac{4mn_{1}}{9(n_{1}-1)}+\frac{rn_{1}}{2(n_{1}-1)}\right)+h\cdot\frac{2mn_{1}}{9(n_{1}-1)} $, $ T_{2}'' = n_{1}(n_{1}-1)e^{\varphi_{2}''} $.

4 数值例子

例 1  考虑如下的初始设计$ d_0\in{\cal U}(12; 2^{10}3^5) $, 其中$ n = 12, m_1 = 10, m_2 = 5 $. 表 2列出了设计$ d_0 $$ n_1 = 6, r = 1, \cdots , 6 $时的数值结果, 其中$ Eff = LBM_1/[MD(D_3)]^2 $. 为节省篇幅, 表 2中仅给出了三水平因子列扩充附加部分$ d_2 $分别在$ r = 1, \cdots , 6 $时的设计表.

表 2   例1的数值结果

n1r[MD(D3)]2LBM1Effd2
61651.2708630.20810.9677112122211121221222111
621.19E+031.15E+030.9707221112122111212111212
632.15E+032.09E+030.9728211221122212121122112
643.91E+033.82E+030.9771222221211112221212221
656.96E+036.83E+030.9810112111122222112211221
661.24E+041.22E+040.9859121212211221112121122

新窗口打开| 下载CSV


例1中设计的设计效率都达到了0.95以上,说明这些设计是近似均匀的列扩充设计. 这也说明我们的理论结果是正确的.

参考文献

Li W , Lin D K J .

Optimal foldover plans for two-level fractional factorial designs

Technometrics, 2003, 45, 142- 149

DOI:10.1198/004017003188618779      [本文引用: 1]

Elsawah M A , Qin H .

Lee discrepancy on symmetric three-level combined designs

Statistics & Probability Letters, 2015, 96, 273- 280

URL     [本文引用: 1]

Yang F , Zhou Y D , Zhang X R .

Augmented uniform designs

Journal of Statistical Planning and Inference, 2017, 182, 61- 73

DOI:10.1016/j.jspi.2016.09.007      [本文引用: 4]

Yang F , Zhou Y D , Zhang A J .

Mixed-level column augmented uniform designs

Journal of Complexity, 2019, 53, 23- 29

DOI:10.1016/j.jco.2018.10.006      [本文引用: 4]

Kiefer J , Wolfowitz J .

Optimum designs in regression problems

The Annals of Mathematical Statistics, 1959, 30, 271- 294

DOI:10.1214/aoms/1177706252      [本文引用: 1]

Fedorov V V . Theory of Optimal Experiments. Holland: Elsevier, 1972

[本文引用: 1]

Fang K T .

The uniform design: application of number-theoretic methods in experimental design

Acta Mathematicae Applicatae Sinica, 1980, 3, 363- 372

URL     [本文引用: 1]

Koehler J R, Owen A B. Computer experiments//Ghosh S, Rao C R. Handbook of Statistics. Amsterdam: Elsevier, 1996, 13: 261-308

[本文引用: 1]

Hickernell F J .

Goodness of fit statistics, discrepancy and robust designs

Statistics & Probability Letters, 1999, 44, 73- 78

URL     [本文引用: 1]

Yue R X , Hickernell F J .

Robust designs for fitting linear models with misspecification

Statistica Sinica, 1999, 9, 1053- 1069

[本文引用: 1]

Xie M Y , Fang K T .

Admissibility and minimaxity of the uniform design in nonparametric regression model

Journal of Statistical Planning and Inference, 2000, 83, 101- 111

DOI:10.1016/S0378-3758(99)00089-0      [本文引用: 1]

Fang K T , Mukerjee R .

A connection between uniformity and aberration in regular fractions of two-level factorials

Biometrika, 2000, 87, 193- 198

DOI:10.1093/biomet/87.1.193      [本文引用: 1]

Zhou Y D , Xu H .

Space-filling fractional factorial designs

Journal of the American Statistical Association, 2014, 109, 1134- 1144

DOI:10.1080/01621459.2013.873367      [本文引用: 1]

Liu J Q , Ou Z J , Hu L P , Wang K .

Lee discrepancy on mixed two-and three-level uniform augmented designs

Communications in Statistics-Theory and Methods, 2019, 48 (10): 2409- 2424

DOI:10.1080/03610926.2018.1465087      [本文引用: 1]

Chatterjee K , Li Z H , Qin H .

Some new lower bounds to centered and wrap-round L2-discrepancies

Statistics & Probability Letters, 2012, 82, 1367- 1373

URL     [本文引用: 1]

/