《算法导论》学习心得(二)—— 矩阵乘法之Strassen算法

来源:互联网 发布:获取键值对java 编辑:程序博客网 时间:2024/05/01 02:34

个人blog迁移到www.forwell.me           

    在开始之前,请点击下载源码。提起矩阵乘法,你也许会说不就是三次循环就解决问题了吗,这有什么好说的。是啊,三个循环确实是完事了,时间效率是O(n^3),这是我们上第一节线代老师就清清楚楚的告诉我们的,但是他没有告诉你还有比这更好的矩阵乘法,时间效率为O(n^{log_2 7}) = O(n^{2.807}),也许你觉得这没有什么,就提高了0.2几,没啥,但是你想过没有,当N=100,10000的时候呢,Strassen算法和传统方法又有多少差别呢,让我们来看一下Strassen算法和传统方法的效率对比图:


通过图我们会发现Strassen算法在N超过50的时候就开始表现出明显的优势,然而现实生产中矩阵都是上百阶的,那Strassen算法更是占有绝对的优势,所以我们今天就很有必要学习Strassen算法,那下面就开始进入正题。

Strassen算法

1969年,德国的一位数学家Strassen证明O(N^3)的解法并不是矩阵乘法的最优算法,他做了一系列工作使得最终的时间复杂度降低到了O(n^2.80)。那他是怎么做到的呢?对于矩阵乘法  C =  A × B,通常的做法是将矩阵进行分块相乘,如下图所示:

从上图可以看出这种分块相乘总共用了8次乘法,要改进算法计算时间的复杂度,必须减少乘法运算次数。按分治法的思想,Strassen提出一种新的方法,用7次乘法完成2阶矩阵的乘法,算法如下:

M1 = A11(B12 - B12)
M2 = (A11 + A12)B22
M3 = (A21 + A22)B11
M4 = A22(B21 - B11)
M5 = (A11 + A22)(B11 + B22)
M6 = (A12 - A22)(B21 + B22)
M7 = (A11 - A21)(B11 + B12)
完成了7次乘法,再做如下加法:
C11 = M5 + M4 - M2 + M6
C12 = M1 + M2
C21 = M3 + M4
C22 = M5 + M1 - M3 - M7
全部计算使用了7次乘法和18次加减法,计算时间降低到O(nE2.81)。计算复杂性得到较大改进。

具体代码实现如下:

//Strassen二阶矩阵的乘法static int[][] twostrassenMatrixMultiply(int [][]x,int [][]y) //阶数为2的矩阵乘法    {   int matrixXColumnLength = x[0].length;int matrixYRowLength = x.length;//获取矩阵的行长度if(matrixXColumnLength!=matrixYRowLength){throw new RuntimeException("matrixXColumnLength!=matrixYRowLength,无法进行乘法计算!");}int p1,p2,p3,p4,p5,p6,p7;//这些都是按照算法定义进行的int [][]result = new int[2][2];p1=(y[0][1] - y[1][1]) * x[0][0];   p2=y[1][1] * (x[0][0] + x[0][1]);   p3=(x[1][0] + x[1][1]) * y[0][0];   p4=x[1][1] * (y[1][0] - y[0][0]);   p5=(x[0][0] + x[1][1]) * (y[0][0]+y[1][1]);   p6=(x[0][1] - x[1][1]) * (y[1][0]+y[1][1]);   p7=(x[0][0] - x[1][0]) * (y[0][0]+y[0][1]);   result[0][0] = p5 + p4 - p2 + p6;   result[0][1] = p1 + p2;   result[1][0] = p3 + p4;result[1][1] = p5 + p1 - p3 - p7;return result;   }   
整个计算过程为:

static int[][] strassenMatrixMultiply(int [][]x,int [][]y) //矩阵乘法方法    {   if(x.length==2)   {   return twostrassenMatrixMultiply(x,y);}   else   {   int matrixLength = x.length/2;int[][] a11,a12,a21,a22;a11 = new int[matrixLength][matrixLength];a12 = new int[matrixLength][matrixLength];a21 = new int[matrixLength][matrixLength];a22 = new int[matrixLength][matrixLength];int[][] b11,b12,b21,b22;b11 = new int[matrixLength][matrixLength];b12 = new int[matrixLength][matrixLength];b21 = new int[matrixLength][matrixLength];b22 = new int[matrixLength][matrixLength];int[][] c11,c12,c21,c22,c;   c11 = new int[matrixLength][matrixLength];c12 = new int[matrixLength][matrixLength];c21 = new int[matrixLength][matrixLength];c22 = new int[matrixLength][matrixLength];c = new int[2*matrixLength][2*matrixLength];int[][] m1,m2,m3,m4,m5,m6,m7;divide(x,a11,a12,a21,a22); //拆分A、B、C矩阵    divide(y,b11,b12,b21,b22);   divide(c,c11,c12,c21,c22);m1=strassenMatrixMultiply(a11,matrixMinus(b12,b22));   m2=strassenMatrixMultiply(matrixPlus(a11,a12),b22);m3=strassenMatrixMultiply(matrixPlus(a21,a22),b11);m4=strassenMatrixMultiply(a22,matrixMinus(b21,b11));   m5=strassenMatrixMultiply(matrixPlus(a11,a22),matrixPlus(b11,b22));   m6=strassenMatrixMultiply(matrixMinus(a12,a22),matrixPlus(b21,b22));   m7=strassenMatrixMultiply(matrixMinus(a11,a21),matrixPlus(b11,b12));   c11=matrixPlus(matrixMinus(matrixPlus(m5,m4),m2),m6);   c12=matrixPlus(m1,m2);   c21=matrixPlus(m3,m4);   c22=matrixMinus(matrixMinus(matrixPlus(m5,m1),m3),m7);c=merge(c11,c12,c21,c22); //合并C矩阵    return c;   }    }
上面就是整个算法的实现过程,欢迎大家前来讨论tangboneu@foxmail.com


完整代码:

package com.tangbo;import java.util.Random;import java.util.Scanner;/* * @Author:唐波 * Strassen矩阵乘法 * 2014.10.31 * 程序对比了传统方法和Strassen算法计算的结果是否相等 * 算法来源:1969年,德国的一位数学家Strassen证明O(N^3)的解法并不是矩阵乘法的最优算法,他做了一系列工作使得最终的时间复杂度降低到了O(n^2.80) */public class SquareMatrixMultiply {static Random random = new Random();static Scanner in;public static void main(String[] args) { int matrixLength=0;in = new Scanner(System.in);   System.out.print("输入矩阵的阶数: ");   matrixLength = in.nextInt();if(isEven(matrixLength)==0){int [][]x=productMatrix(matrixLength);int [][]y=productMatrix(matrixLength);System.out.println("x矩阵:");printMatrix(x);System.out.println("y矩阵:");printMatrix(y);int [][]strassenResult =strassenMatrixMultiply(x,y);//Strassen计算结果System.out.println("Strassen计算结果:");printMatrix(strassenResult);int [][] forceResult = forceMatrixMultiply(x, y);//传统方法计算结果System.out.println("传统计算结果:");printMatrix(forceResult);boolean isEqual = isEqual(forceResult, strassenResult);//比较两种计算结果是否相等if(isEqual){System.out.println("两个计算结果相等!");}else{System.err.println("两个计算结果不相等!");System.exit(0);//程序退出}}else{System.out.println("矩阵不是2^k方阵,无法计算!");}}static boolean isEqual(int [][]x,int [][]y)//遍历判断两个矩阵是否相等{boolean equal=true;for(int i =0;i<x.length;i++){for(int j=0;j<x[0].length;j++){if(x[i][j]!=y[i][j]){equal=false;}}}return equal;}static int isEven(int n)//判断输入矩阵阶数是否为2^k{   int a = 1,temp=n;   while(temp%2==0)   {   if(temp%2==0)    temp/=2; }  if(temp==1)    a=0;   return a;}   static int[][] productMatrix(int matrixLength)//自动生成矩阵{int matrix[][] = new int[matrixLength][matrixLength];//初始化矩阵for(int i=0;i<matrixLength;i++){for(int j=0;j<matrixLength;j++){matrix[i][j] = (int)(Math.random()*10);}}System.out.println();return matrix;}static void printMatrix(int matrix[][])//矩阵打印函数{int matrixRowLength = matrix.length;//获取矩阵的行数int matrixColumnLength = matrix[0].length;//获取矩阵的列数for(int i=0;i<matrixRowLength;i++){for(int j=0;j<matrixColumnLength;j++){System.out.print(matrix[i][j]+" ");}System.out.println();}}static int[][] matrixPlus(int[][] x,int[][] y) //矩阵加法    {   int matrixXRowLength = x.length;//获取矩阵的行长度int matrixXColumnLength = x[0].length;int matrixYRowLength = x.length;//获取矩阵的行长度int matrixYColumnLength = x[0].length;if(matrixXColumnLength!=matrixYColumnLength || matrixXRowLength!=matrixYRowLength)//判断矩阵是否同型{throw new RuntimeException("矩阵不同型,无法进行加法计算!");}int[][] result = new int[matrixXRowLength][matrixXColumnLength];for(int i=0;i<matrixXColumnLength;i++){for (int j = 0; j < matrixXColumnLength; j++) {result[i][j] = x[i][j]+y[i][j]; }}return result;}   static int[][] matrixMinus(int[][] x,int[][] y)//矩阵减法{int matrixXRowLength = x.length;//获取矩阵的行长度int matrixXColumnLength = x[0].length;int matrixYRowLength = x.length;//获取矩阵的行长度int matrixYColumnLength = x[0].length;if(matrixXColumnLength!=matrixYColumnLength || matrixXRowLength!=matrixYRowLength){throw new RuntimeException("矩阵不同型,无法进行减法计算!");}int[][] result = new int[matrixXRowLength][matrixXColumnLength];for(int i=0;i<matrixXColumnLength;i++){for (int j = 0; j < matrixXColumnLength; j++) {result[i][j] = x[i][j]-y[i][j]; }}return result;}//Strassen二阶矩阵的乘法static int[][] twostrassenMatrixMultiply(int [][]x,int [][]y) //阶数为2的矩阵乘法    {   int matrixXColumnLength = x[0].length;int matrixYRowLength = x.length;//获取矩阵的行长度if(matrixXColumnLength!=matrixYRowLength){throw new RuntimeException("matrixXColumnLength!=matrixYRowLength,无法进行乘法计算!");}int p1,p2,p3,p4,p5,p6,p7;//这些都是按照算法定义进行的int [][]result = new int[2][2];p1=(y[0][1] - y[1][1]) * x[0][0];   p2=y[1][1] * (x[0][0] + x[0][1]);   p3=(x[1][0] + x[1][1]) * y[0][0];   p4=x[1][1] * (y[1][0] - y[0][0]);   p5=(x[0][0] + x[1][1]) * (y[0][0]+y[1][1]);   p6=(x[0][1] - x[1][1]) * (y[1][0]+y[1][1]);   p7=(x[0][0] - x[1][0]) * (y[0][0]+y[0][1]);   result[0][0] = p5 + p4 - p2 + p6;   result[0][1] = p1 + p2;   result[1][0] = p3 + p4;result[1][1] = p5 + p1 - p3 - p7;return result;   }   static void divide(int[][] a,int[][] a11,int[][] a12,int[][] a21,int[][] a22)//分解矩阵{   int matrixLength = a.length/2;for(int i=0;i<matrixLength;i++)   for(int j=0;j<matrixLength;j++)   {a11[i][j]=a[i][j];a12[i][j]=a[i][j+matrixLength];   a21[i][j]=a[i+matrixLength][j];   a22[i][j]=a[i+matrixLength][j+matrixLength];   }   }static int[][] merge(int [][]a11,int [][]a12,int [][]a21,int [][]a22)//合并矩阵    {   int n=a11.length;int [][] result = new int[2*n][2*n];for(int i=0;i<n;i++){for(int j=0;j<n;j++){result[i][j]=a11[i][j];   result[i][j+n]=a12[i][j];   result[i+n][j]=a21[i][j];   result[i+n][j+n]=a22[i][j];   }}return result;}static int[][] strassenMatrixMultiply(int [][]x,int [][]y) //矩阵乘法方法    {   if(x.length==2)   {   return twostrassenMatrixMultiply(x,y);}   else   {   int matrixLength = x.length/2;int[][] a11,a12,a21,a22;a11 = new int[matrixLength][matrixLength];a12 = new int[matrixLength][matrixLength];a21 = new int[matrixLength][matrixLength];a22 = new int[matrixLength][matrixLength];int[][] b11,b12,b21,b22;b11 = new int[matrixLength][matrixLength];b12 = new int[matrixLength][matrixLength];b21 = new int[matrixLength][matrixLength];b22 = new int[matrixLength][matrixLength];int[][] c11,c12,c21,c22,c;   c11 = new int[matrixLength][matrixLength];c12 = new int[matrixLength][matrixLength];c21 = new int[matrixLength][matrixLength];c22 = new int[matrixLength][matrixLength];c = new int[2*matrixLength][2*matrixLength];int[][] m1,m2,m3,m4,m5,m6,m7;divide(x,a11,a12,a21,a22); //拆分A、B、C矩阵    divide(y,b11,b12,b21,b22);   divide(c,c11,c12,c21,c22);m1=strassenMatrixMultiply(a11,matrixMinus(b12,b22));   m2=strassenMatrixMultiply(matrixPlus(a11,a12),b22);m3=strassenMatrixMultiply(matrixPlus(a21,a22),b11);m4=strassenMatrixMultiply(a22,matrixMinus(b21,b11));   m5=strassenMatrixMultiply(matrixPlus(a11,a22),matrixPlus(b11,b22));   m6=strassenMatrixMultiply(matrixMinus(a12,a22),matrixPlus(b21,b22));   m7=strassenMatrixMultiply(matrixMinus(a11,a21),matrixPlus(b11,b12));   c11=matrixPlus(matrixMinus(matrixPlus(m5,m4),m2),m6);   c12=matrixPlus(m1,m2);   c21=matrixPlus(m3,m4);   c22=matrixMinus(matrixMinus(matrixPlus(m5,m1),m3),m7);c=merge(c11,c12,c21,c22); //合并C矩阵    return c;   }    }static int[][] forceMatrixMultiply(int [][]x,int [][]y){int matrixXRowLength = x.length;//获取矩阵的行长度int matrixXColumnLength = x[0].length;int matrixYRowLength = x.length;//获取矩阵的行长度int matrixYColumnLength = x[0].length;if(matrixXColumnLength!=matrixYRowLength){throw new RuntimeException("matrixXColumnLength!=matrixYRowLength,无法进行乘法计算!");}int [][] result = new int[matrixXRowLength][matrixYColumnLength];for(int i=0;i<matrixXRowLength;i++){for(int j=0;j<matrixYColumnLength;j++){result[i][j]=0;for(int k=0;k<matrixYRowLength;k++){result[i][j] = result[i][j]+x[i][k]*y[k][j];}}}return result;}}





0 0