矩阵和向量积
numpy.dot(a, b[, out])
计算两个矩阵乘积或两个向量内积
import numpy as np
x = np.array([1, 2, 3, 4, 5])
y = np.array([2, 3, 4, 5, 6])
z = np.dot(x, y) # 计算矩阵乘积
print(z) # 70
x = np.array([[1, 2, 3], [3, 4, 5], [6, 7, 8]])
y = np.array([[5, 4, 2], [1, 7, 9], [0, 4, 5]])
z = np.dot(x, y)
print(z)
# [[ 7 30 35]
# [ 19 60 67]
# [ 37 105 115]]
z = np.dot(y, x)
print(z)
# [[ 29 40 51]
# [ 76 93 110]
# [ 42 51 60]]
注:线性代数和numpy数组中维数的含义不同
线性代数 | numpy |
---|---|
n维行向量 | 一维数组 |
n维列向量 | n×1的二维数组 |
矩阵特征值与特征向量
计算方阵的特征值
numpy.linalg.eigvals(a)
计算方阵的特征值和特征向量
numpy.linalg.eig(a)
import numpy as np
# 求方阵的特征值特征向量
x = np.diag((1, 2, 3)) # 创建一个对角矩阵
print(x)
# [[1 0 0]
# [0 2 0]
# [0 0 3]]
print(np.linalg.eigvals(x)) # 计算特征值
# [1. 2. 3.]
a, b = np.linalg.eig(x)# 特征值保存在a中,特征向量保存在b中
print(a)
# [1. 2. 3.]
print(b)
# [[1. 0. 0.]
# [0. 1. 0.]
# [0. 0. 1.]]
# 检验特征值与特征向量是否正确
for i in range(3):
# np.allclose():比较两个array是不是每一元素都相等,默认在1e-05的误差范围内
if np.allclose(a[i] * b[:, i], np.dot(x, b[:, i])):
print('Right')
else:
print('Error')
# Right
# Right
# Right
import numpy as np
# 判断对称阵是否为正定阵(特征值是否全部为正)。
A = np.arange(16).reshape(4, 4)
print(A)
# [[ 0 1 2 3]
# [ 4 5 6 7]
# [ 8 9 10 11]
# [12 13 14 15]]
A = A + A.T # 将方阵转换成对称阵
print(A)
# [[ 0 5 10 15]
# [ 5 10 15 20]
# [10 15 20 25]
# [15 20 25 30]]
B = np.linalg.eigvals(A) # 求A的特征值
print(B)
# [ 6.74165739e+01 ‐7.41657387e+00 1.82694656e‐15 ‐1.72637110e‐15]
# 判断是不是所有的特征值都大于0
if np.all(B > 0):
print('Yes')
else:
print('No')
# No
矩阵分解
奇异值分解
奇异值分解的相关原理:奇异值分解(SVD)及其应用
我们可以用最大的k个奇异值的矩阵和相乘来近似描述矩阵,从而实现了降维、减少数据存储、提升计算性能等效果。
u, s, v = numpy.linalg.svd(a, full_matrices=True, compute_uv=True, hermitian=False)
s为A的奇异值分解,只有对角线元素不为零,且按元素从小到大排序。s中有n个奇异值,一般仅保留较大的r个奇异值。
import numpy as np
# 奇异值分解 A=u*s*vh,vh为v的转置
A = np.array([[4, 11, 14], [8, 7, -2]])
print(A)
# [[ 4 11 14]
# [ 8 7 -2]]
u, s, vh = np.linalg.svd(A, full_matrices=False) # full_matrices=False:u为M×K,v为K×N,K=min(M,N)
print(u.shape) # (2, 2)
print(u)
# [[‐0.9486833 ‐0.31622777]
# [‐0.31622777 0.9486833 ]]
print(s.shape) # (2,) 一维
print(np.diag(s)) # 一维数组时,np.diag()返回一个以一维数组为对角线元素的矩阵
# [[18.97366596 0. ]
# [ 0. 9.48683298]]
print(vh.shape) # (2, 3)
print(vh)
# [[‐0.33333333 ‐0.66666667 ‐0.66666667]
# [ 0.66666667 0.33333333 ‐0.66666667]]
a = np.dot(u, np.diag(s))
a = np.dot(a, vh)
print(a)
# [[ 4. 11. 14.]
# [ 8. 7. ‐2.]]
import numpy as np
# 奇异值分解2
A = np.array([[1, 1], [1, -2], [2, 1]])
print(A)
# [[ 1 1]
# [ 1 ‐2]
# [ 2 1]]
u, s, vh = np.linalg.svd(A, full_matrices=False)
print(u.shape) # (3, 2)
print(u)
# [[‐5.34522484e‐01 ‐1.11022302e‐16]
# [ 2.67261242e‐01 ‐9.48683298e‐01]
# [‐8.01783726e‐01 ‐3.16227766e‐01]]
print(s.shape) # (2,)
print(np.diag(s))
# [[2.64575131 0. ]
# [0. 2.23606798]]
print(vh.shape) # (2, 2)
print(vh)
# [[‐0.70710678 ‐0.70710678]
# [‐0.70710678 0.70710678]]
a = np.dot(u, np.diag(s))
a = np.dot(a, vh)
print(a)
# [[ 1. 1.]
# [ 1. ‐2.]
# [ 2. 1.]]
QR分解
QR(正交三角)分解法是求一般矩阵全部特征值的最有效并广泛应用的方法,一般矩阵先经过正交相似变化成为Hessenberg矩阵,然后再应用QR方法求特征值和特征向量。
A=QR
A为M×N的待分解矩阵,Q为正交矩阵,R为上三角形矩阵。
q,r = numpy.linalg.qr(a, mode=‘reduced’)
mode = reduced :返回(M, N)的列向量两两正交的矩阵q,和(N, N)的三角阵r (Reduced QR分解)。
mode = complete :返回(M, M)的正交矩阵q ,和(M, N)的三角阵r (Full QR分解)。
import numpy as np
# QR分解
A = np.array([[2, -2, 3], [1, 1, 1], [1, 3, -1]])
print(A)
# [[ 2 ‐2 3]
# [ 1 1 1]
# [ 1 3 ‐1]]
q, r = np.linalg.qr(A) # QR分解
print(q.shape) # (3, 3)
print(q)
# [[‐0.81649658 0.53452248 0.21821789]
# [‐0.40824829 ‐0.26726124 ‐0.87287156]
# [‐0.40824829 ‐0.80178373 0.43643578]]
print(r.shape) # (3, 3)
print(r)
# [[‐2.44948974 0. ‐2.44948974]
# [ 0. ‐3.74165739 2.13808994]
# [ 0. 0. ‐0.65465367]]
print(np.dot(q, r))
# [[ 2. ‐2. 3.]
# [ 1. 1. 1.]
# [ 1. 3. ‐1.]]
a = np.allclose(np.dot(q.T, q), np.eye(3)) # 检验q是否为正交阵
print(a) # True
import numpy as np
# QR分解
A = np.array([[1, 1], [1, -2], [2, 1]])
print(A)
# [[ 1 1]
# [ 1 ‐2]
# [ 2 1]]
q, r = np.linalg.qr(A, mode='complete') # q为正交阵,Full QR分解
print(q.shape) # (3, 3)
print(q)
# [[‐0.40824829 0.34503278 ‐0.84515425]
# [‐0.40824829 ‐0.89708523 ‐0.16903085]
# [‐0.81649658 0.27602622 0.50709255]]
print(r.shape) # (3, 2)
print(r)
# [[‐2.44948974 ‐0.40824829]
# [ 0. 2.41522946]
# [ 0. 0. ]]
print(np.dot(q, r))
# [[ 1. 1.]
# [ 1. ‐2.]
# [ 2. 1.]]
a = np.allclose(np.dot(q, q.T), np.eye(3))
print(a) # True
print('****************************')
qq, r = np.linalg.qr(A) # mode='reduced' qq的列向量两两正交,Reduced QR分解
print(qq.shape) # (3, 2)
print(qq)
# [[-0.40824829 0.34503278]
# [-0.40824829 -0.89708523]
# [-0.81649658 0.27602622]]
print(r.shape) # (2, 2)
print(r)
# [[-2.44948974 -0.40824829]
# [ 0. 2.41522946]]
print(np.dot(qq, r))
# [[ 1. 1.]
# [ 1. ‐2.]
# [ 2. 1.]]
a = np.allclose(np.dot(qq.T,qq), np.eye(2))
print(a) # True
Cholesky分解
Cholesky 分解是把一个对称正定的矩阵表示成一个下三角矩阵L和其转置的乘积的分解。它要求矩阵的所有特征值必须大于零,故分解的下三角的对角元也是大于零的。
A=LL’
L = numpy.linalg.cholesky(a)
import numpy as np
# Cholesky分解
A = np.array([[1, 1, 1, 1],
[1, 3, 3, 3],
[1, 3, 5, 5],
[1, 3, 5, 7]])
print(A)
print(np.linalg.eigvals(A)) # A的特征值
# [13.13707118 1.6199144 0.51978306 0.72323135]
L = np.linalg.cholesky(A) # Cholesky分解
print(L)
# [[1. 0. 0. 0. ]
# [1. 1.41421356 0. 0. ]
# [1. 1.41421356 1.41421356 0. ]
# [1. 1.41421356 1.41421356 1.41421356]]
print(np.dot(L, L.T))
# [[1. 1. 1. 1.]
# [1. 3. 3. 3.]
# [1. 3. 5. 5.]
# [1. 3. 5. 7.]]
范数和其他数字
矩阵的范数
numpy.linalg.norm(x, ord=None, axis=None, keepdims=False)
计算向量和矩阵的范数(函数详情)
ord参数决定不同的范数:
常用ord参数及计算方法:
import numpy as np
# 向量的范数
x = np.array([1, 2, 3, 4])
print(np.linalg.norm(x, ord=1)) # ord=1:一范数
# 10.0
print(np.sum(np.abs(x)))
# 10
print(np.linalg.norm(x, ord=2)) # ord=2:二范数
# 5.477225575051661
print(np.sum(np.abs(x) ** 2) ** 0.5)
# 5.477225575051661
print(np.linalg.norm(x, ord=-np.inf)) # 负无穷范数
# 1.0
print(np.min(np.abs(x)))
# 1
print(np.linalg.norm(x, ord=np.inf)) # 无穷范数
# 4.0
print(np.max(np.abs(x)))
# 4
import numpy as np
# 矩阵的范数
A = np.array([[1, 2, 3, 4],
[2, 3, 5, 8],
[1, 3, 5, 7],
[3, 4, 7, 11]])
print(A)
print(np.linalg.norm(A, ord=1)) # 30.0 1-范数:列和的最大值
print(np.max(np.sum(A, axis=0))) # 30
print(np.linalg.norm(A, ord=2)) # 2-范数:矩阵A'A的最大特征值开平方根
# 20.24345358700576
print(np.max(np.linalg.svd(A, compute_uv=False))) # compute_uv=False:只计算s,A的奇异值是A'A的特征值的平方根
# 20.24345358700576
print(np.linalg.norm(A, ord=np.inf)) # 25.0 行和的最大值
print(np.max(np.sum(A, axis=1))) # 25
print(np.linalg.norm(A, ord='fro')) # F-范数:矩阵各元素平方和开根号
# 20.273134932713294
print(np.sqrt(np.trace(np.dot(A.T, A)))) # A'A主对角元素开根号
# 20.273134932713294
方阵的行列式
numpy.linalg.det(a)
返回方阵的行列式
import numpy as np
# 方阵的行列式
x = np.array([[1, 2], [3, 4]])
print(x)
print(np.linalg.det(x))
# ‐2.0000000000000004
矩阵的秩
numpy.linalg.matrix_rank(M, tol=None, hermitian=False)
返回矩阵的秩
import numpy as np
# 矩阵的秩
I = np.eye(3) # 创建一个单位阵
print(I)
# [[1. 0. 0.]
# [0. 1. 0.]
# [0. 0. 1.]]
r = np.linalg.matrix_rank(I) # 求矩阵的秩
print(r) # 3
I[1, 1] = 0
print(I)
# [[1. 0. 0.]
# [0. 0. 0.]
# [0. 0. 1.]]
r = np.linalg.matrix_rank(I)
print(r) # 2
方阵的迹
numpy.trace(a, offset=0, axis1=0, axis2=1, dtype=None, out=None)
返回方阵的迹(主对角元素之和)
import numpy as np
# 方阵的迹
x = np.array([[1, 2, 3], [3, 4, 5], [6, 7, 8]])
print(x)
y = np.array([[5, 4, 2], [1, 7, 9], [0, 4, 5]])
print(y)
print(np.trace(x)) # A的迹等于A.T的迹
# 13
print(np.trace(np.transpose(x)))
# 13
print(np.trace(y))
print(np.trace(x + y)) # 和的迹等于迹的和
# 30
print(np.trace(x) + np.trace(y))
# 30
解方程和逆矩阵
逆矩阵
AB=BA=E 则称B为A的逆矩阵
矩阵可逆的充要条件:det(A) != 0 ,或者A 满秩
numpy.linalg.inv(a)
返回矩阵A的逆矩阵
import numpy as np
# 计算矩阵的逆矩阵
A = np.array([[1, -2, 1], [0, 2, -1], [1, 1, -2]])
print(A)
# 求A的行列式,不为零则存在逆矩阵
A_det = np.linalg.det(A)
print(A_det)
# ‐2.9999999999999996
A_inverse = np.linalg.inv(A) # 求A的逆矩阵
print(A_inverse)
# [[ 1.00000000e+00 1.00000000e+00 ‐1.11022302e‐16]
# [ 3.33333333e‐01 1.00000000e+00 ‐3.33333333e‐01]
# [ 6.66666667e‐01 1.00000000e+00 ‐6.66666667e‐01]]
x = np.allclose(np.dot(A, A_inverse), np.eye(3))
print(x) # True
x = np.allclose(np.dot(A_inverse, A), np.eye(3))
print(x) # True
A_companion = A_inverse * A_det # 求A的伴随矩阵
print(A_companion)
# [[‐3.00000000e+00 ‐3.00000000e+00 3.33066907e‐16]
# [‐1.00000000e+00 ‐3.00000000e+00 1.00000000e+00]
# [‐2.00000000e+00 ‐3.00000000e+00 2.00000000e+00]]
求解线性方程组
numpy.linalg.solve(a, b)
返回线性方程组的解
# x + 2y + z = 7
# 2x ‐ y + 3z = 7
# 3x + y + 2z =18
import numpy as np
# 求解线性矩阵方程
A = np.array([[1, 2, 1], [2, -1, 3], [3, 1, 2]])
b = np.array([7, 7, 18])
x = np.linalg.solve(A, b)
print(x) # [ 7. 1. ‐2.]
x = np.linalg.inv(A).dot(b)
print(x) # [ 7. 1. ‐2.]
y = np.allclose(np.dot(A, x), b)
print(y) # True