稀疏信号处理简介.pptVIP

稀疏信号处理简介 “Signal information processing is · · · an Art ” — Petre Stoica 一所大学，两个战场 科学：一个是和其他世界一流大学共同面对的国际学术前沿战场 技术：另一个是为我们国家经济、社会、国防、发展战略需要服务的战场 U E S T C 内容 从几个问题开始 稀疏重建理论 几个例子 阵列信号处理的例子 实孔径超分辨、阵列稀疏布阵 无线定位的例子 MDS、MC 信道估计的例子 一、从几个问题开始 高斯分布凭什么无所不在？ MMSE是最优的？ 吝啬原则：免费的午餐？ 分辨率受孔径限制？ 机器学习：支持向量是稀疏的？ 什么是多维标度问题? 高斯分布：An equation is for eternity Gauss’s Question (1809) What would be a distribution density f(x ;θ) for which the maximum likelihood estimate ofθ is the sample mean we use the modern terminology adopted by the scientific community more than a century later （the method of maximum likelihood was proposed by Fisher in 1921） DERIVATION OF GAUSS (1809) Using i.i.d. observations, the maximum likelihood estimate of parameter of location Derivation any real number can be arbitrarily accurately approximated by rational numbers Result Gauss assumed the sample mean due to its computational convenience and derived the Gaussian law. This line of reasoning is quite the opposite to the modern exposition in textbooks on statistics and signal processing where the LS method is derived from the assumed Gaussianity. 为什么要折衷？ 性能最优 计算最简单 跑题了？ 1.1 高斯分布凭什么无所不在？ The role of Gaussian models in signal processing is based on the optimal property of the Gaussian distribution minimizing Fisher information over the class of distributions with a bounded variance. The central limit theorem (CLT) is not only a unique reason but perhaps it is even not the main reason Fisher information 1.2 MMSE是最优的？ NP-Hard ？现代最小二乘 1.3 吝啬原则：免费的午餐？ 多成分混合（合成，正问题） 分离各个成分（感知，反问题） 贪婪的谱估计 = 滤波： 1.4 分辨率受孔径限制？ DFT BWE 1GHz (S) + 1GHz (X) = 10GHz ？ 超分辨是一个欠定问题 在线测量 + 先验模型 稀疏 1.5 机器学习：支持向量是稀疏的？ the training sample hyperplane that does the separation primal formulation of the problem convex quadratic programming problem dual formulation Great watershedin optimization It is not between linearity and nonlinearity, but convexity and non-convexity — R. Rockafellar, SIAM Review 1993 1.6 什么是多维标