从动态平衡分布求转移矩阵-- 《气象随机场 -22 》 张学文,2014/8/28-31 2014,9,3注:本稿的第4段在9月3日发现错误,这也影响了后面的结论。我已经用一个新版代替本稿,见 http://blog.sciencenet.cn/blog-2024-824687.html 。但是此稿依然保留在这里。 经过前面的分析与讨论,我们逐步认识到,面积很大的气象场(这通常理解是全球、半球或者面积很大的一个空间、区域)中的某气象变量在某时刻的分布函数经常具有稳定性。而前面还说明可以存在一种转移矩阵,把分布函数与转移矩阵做乘法而获得的新分布函数(一个时间步长)依然与原分布函数相同。而此时的分布函数称为极限分布函数。 我们后面着重分析这种具有时间不变性的极限分布函数与其转移矩阵应当是什么关系。或者说我们是否可以从分布函数去推求它要求的转移矩阵。即可以使原分布函数在经过一个(以致充分多个)时间步长的转移矩阵作用以后依然没有变化。 上一讲我们把这归结为公式: xA=x 现在探索已经知道 x 求 A 。这里的 x 是一个矢量,它的各个分量是分布函数的各个函数值(自变量是规定的有限个离散状态的相空间),而 A 是待求的转移矩阵。 也许应当交代依据笔者的线性代数知识很差,以下的探索属于个人的经验摸索。承望高手指教。而如果这里有一定的创新性,也希望指出,而引用者请注明出处。 1. 矩阵中的未知数数量 :我们这里的 x 是已经知道的有 n 个离散值的 n 维矢量。而待求的转移矩阵 A 自然 n 2 个未知数(方阵中的元素数量)。我们要从 x 的 n 个已知数据中求得 n 2 个未知数数据。 2. 矩阵中的“0”很多 :注意我们在第16讲( 状态转移矩阵的时间步长问题 - 《气象随机场 -16 》 http://blog.sciencenet.cn/blog-2024-816969.html ),单独讨论过转移矩阵的时间步长问题,而特别指出我们可以把 时间步长取得很短 ,以使在这么短的时间内气象场内各点(有时是指气象场内的元面积,有时是指空气微团)的气象状态仅可能是依然留在本相格(状态区间)内或者移到它的左右两个临近相格,而转移到离它更远的相格的百分率是0。所以对于有n 2 个元素转移矩阵 A ,它 最多有3 n -2 个未知数 。即此转移矩阵至少有 个“0”。具体到矩阵的每一行,则其第1行和最后一行,仅有两个未知数,其他的行有三个未知数。 3. 转移矩阵的每行的合计值=1 :转移矩阵的每行的各个元素的合计值代表了从本状态转移到各个状态的百分比,它的合计值自然=1。这样我们就可以获得n个合计值=1的方程式,从而使未知数再减少n个。于是我们 未知数就仅有(2n-2)个 了。 4. 已知分布函数解决了(n-1)个未知数 :由于本问题中的极限分布函数矢量本身就代表着n个已知数,它们显然参与(限制)着转移矩阵的各个元素值。所以n维已知矢量本身就获得(n-1)个方程,从而可以获得(n-1)个未知数(注:由于分布函数的各个分量的合计值应当=1的限制,所以它仅具有(n-1)个独立已知数)。这样我们的未知数就从 (2n-2) 个减少为只有 (n-1) 个了 5. 细致平衡原则 :在已经达到极限分布的情况下(我们现在就是),在马尔科夫过程的理论中有一条所谓细致平衡原则。即这时各个相格中的存在量(也就是分布函数矢量的每个分量)与它转移到指定相格的转移率(矩阵的元素)的乘积,与对方转移回来的对应量是相等的。写成为公式就是 x i a i,j =x j a j,i ( i.j=1,2 , … , n ) 这里的 x i 或者 x j 表示分布函数的对这个矢量的对应分量,而 a i,j , a j,i 表示转移矩阵中的对应元素值。它体现在达到动态平衡时任何一个相格中从另外相格的流入量与自己流出到对方的量是相等的。对此在马尔科夫过程的理论中有论及。我们这里就不再多说了。 在转移矩阵的非 0 元素中,这个细致平衡关系又使未知数减少很多( 本段待完善 )。 6. 给一个规定速度值 :我们用限定一个时间步长相当短的规定已经使很多转移矩阵的元素值=0了。但是在没有具体规定时间步长的情况下,我们还是需要在转移矩阵的各个元素中,人为规定一个元素的值(一个时间步长的转移百分比),以具体体现转移的速度。在此我们一般规定转移矩阵中的 a 22 的值是我们事先给定的, a 22 的含义是处于第2个相格中的空气状态在下一个时间步长依然是留在本相格的百分比。而且我们一般令 a 22 =0.98 ,即一个时间步长内处于第2个相格的空气维持原状态的百分率是98%。这样我们就又减少了一个未知数。 7. 转移矩阵的“解”不是唯一的 :前面的分析指出如何在已知极限分布函数的情况下,并且在规定了 a 22 =0.98 的情况下,可以(?)求得对应的转移矩阵。显然,如果修改的 a 22 值,我们还会获得另外的转移矩阵。这说明转移矩阵并不是唯一的。即一个极限分布函数在不同转移速度下,对应这不同的转移矩阵。 8. 在极限分布函数已知的情况下、在限定一次转移最多是可以到达最近的邻态的假设下、在 a 22 =0.98 的规定下,在细致平衡原理的配合下,我们在理论上分析出可以从极限分布函数求得它要求的转移矩阵。极限分布函数也就是分布函数保持动态平衡(不变化)的分布函数。而它对应的转移矩阵则体现在动态平衡下的气象状态的变化规律。 以上是在代数学、马尔科夫过程知识、气象变量知识特点的基础上的从极限分布函数求其转移矩阵的一般分析。我们在下一讲中要以具体的例子具体计算其转移矩阵 。
对偶谢尔宾斯基分形概率转移矩阵的特征值谱及其应用 伍顺琪 章忠志 摘要 :网络概率转移矩阵的特征值谱,包含了网络许多重要的结构性质,同时也与网路的众多动力学行为密切相关 。本文研究了 d 维对偶 谢尔宾斯基 分 形(即汉诺塔图)的概率转移矩阵。通过重整化群的方法,得到了该类分形的所有特征值及其重数。在此基础上,进一步根据所得的特征值,求得了此类分形网络的生成树数目以及网络上随机游走特征时间的解析表达式。 相关结果已在《 Journal of Physics A 》正式发表。 文章发表的 PDF 版本: Eigenvalue spectrum of transition matrix of dual Sierpinski gaskets and its appl.pdf
My professional goals and experiences for the blog. I would like to open a web forum to discuss Chinese Ecosystem Research Network (CERN) Data Analysis and possibly develop this into a graduate level seminar/course. CERN data supposed to be three-subscript data: they comprise numerous samples, multiple variables, and repeated collections, that is D (i,j,k) , i=1,2,m, j=1,2,n, and k=1,2,o . I have extensive experience with these three-subscript data and designed a model, so called Multi-Dimensional Sphere Model, MDSM,specifically for analyzing these data. The forum/seminar would contain two parts. The first part of the forum will discuss the application of clustering analysis using cosine values as similarity coefficients. After the clustering analysis, the CERN three-subscript data would be converted into two-subscript data, D (i,k) , i=1,2,m, k=1,2,o . These two-subscript data can be considered as multi-component vector ( m -vector) time series. The second part of the forum/seminar would consider the temporal dynamic analysis for the m -vector time series. The temporal dynamic analysis would include Trend Analysis to discover the changing trends of the CERN data. After the trends were discovered, they would be used to Project the future states; furthermore, the projection would be updated by actual sampling using the so called Kalman Filter (Jameson, 1993); and the prediction error would be estimated. This whole procedure needs to be further tested with CERN data and to be published. However, this System Monitoring and Trend Analysis can be conducted using stock market data, asthe stock market data is a m -vecot time series too. My second professional goal is to organize/write a text book entitled Vegetation Sciences and Vector Analysis aimed at students whose majors are in Forestry, Rangelands, Ecology, Computer Science, and Quantitative Ecology. Vegetation is the primary producer and main body of the ecosystem, as it accounts for about 90% of the biomass of the ecosystem. Based on my research experiences, the vegetation can be considered as a resource sharing, exponential growth, time succession, multi-components system, and defined as m -system. An m -system can be expressed as an m -vector, an identity in multi-species space, m- space. A vector has both magnitude and direction. The magnitude expresses the amount of the m -system, such as biomass, energy, and the information, while the direction expresses the composition of the m -system, i.e., how the biomass, the energy, and the information are distributed among the different species. The m -vector analysis extends the traditional 3 -vector analysis to multi-dimension, and defined m -vector division (Bai, et. al., 1997) so that it can analyze exponential growth, as well as linear growth. The composition changes of an m -system can be imagined as rotation of an m -vector in m -space. The exciting aspect of these analyses is that the results can be projected on to a two dimensional plane and visually shown to students. With research colleagues from different fields, I would like to organize a research institute (hyper-bang) for the m -system approach. Presently, I have explored two fields that can adopt the m -system approach. The two are vegetation and the financial market. This research institute may include professionals from, but not limited to, Math, Computer Science, Geography, Forestry, Grasslands, Economy, and Finance. My third professional goal is to design a mathematically sound and easily accessed stock market index, for example, vector length, in addition to the existing old indices. My blog is entitled as m -Vector, m -Space, and m -System Monitoring. This blog and my service could greatly support and advance the Math, Ecology, Finance, and Economy, as well as the this ScienceNet. Information in ecology IS NOT one dimensional. Thus, a one dimensional real number, a scalar, may not be enough to fully explain our multi-component ecosystems. While studying ecosystems, we must explicitly treat them as systems, instead of a bunch of single variables. Thus, we may have to extend our view from one dimensional axis to multi-dimensional space, m -space, and to extend our research tools from scalars to multi-component vectors, m -vectors. In this blog I believe my articles would offer a view to analyze those inter-correlated and auto-correlated variables while they form a resource sharing, exponential growth, time succession, multi-component system, m -system. CERN, has functioned for many years, and has the goal of promoting temporal dynamic analysis to predict future states of the environment. However, discussions with my academic advisors, including later Academician Bo Li from the Chinese Academy , Prof. Song-ling Zhao at University of Lanzhou , and Dr. Donald Jameson at Colorado State University has convinced me that for ecological data the matrix inverse, or pseudo inverse, does not exist. Not finding a transition matrix may be one of the reasons that even with so many CERN stations, after so many years of data collection, researchers have been unable to accomplish a true temporal dynamic analysis. Based on my research, the transition matrix, if it exists, has to be diagonal matrix (Bai, 1998), or m -vector. On the other hand, the statistics that have been applied to these data were designed to work in sample space, n -space, to discover the relation among the variables, the so called R -analysis, but the temporal dynamic analysis that we propose is to work in variable space, m -space, to discover the relation among the samples, previous and present, the so called Q - and O -analysis (Legendre Legendre, 1998). Facing environmental and financial crises, scientists are required to investigate the world from different and multiple angles, to investigate the world as a whole, as a system. The importance of extending our view from scalar to m -vector, and switching our emphasis from R -analysis to Q - and O -analysis cannot be overstated in my opinion. Using Q - and O -analysis to perform temporal dynamic analysis would be a great breakthrough for CERN program. Thus, I am very optimistic that this blog will greatly advance the CERN research and benefit the Net and all the involved netters.