1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

60

61

62

63

64

65

66

67

68

69

70

71

72

73

74

75

76

77

78

79

80

81

82

83

84

85

86

87

88

89

90

91

92

93

94

95

96

97

98

99

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

116

117

118

119

120

121

122

123

124

125

126

127

128

129

130

131

132

133

134

135

136

137

138

139

140

141

142

143

144

145

146

147

148

149

150

151

152

153

154

155

156

157

158

159

160

161

162

163

164

165

166

167

168

169

170

171

172

173

174

175

176

177

178

179

180

181

182

183

184

185

186

187

188

189

190

191

192

193

194

195

196

""" 

Matrix square root for general matrices and for upper triangular matrices. 

 

This module exists to avoid cyclic imports. 

 

""" 

from __future__ import division, print_function, absolute_import 

 

__all__ = ['sqrtm'] 

 

import numpy as np 

 

from scipy._lib._util import _asarray_validated 

 

 

# Local imports 

from .misc import norm 

from .lapack import ztrsyl, dtrsyl 

from .decomp_schur import schur, rsf2csf 

 

 

class SqrtmError(np.linalg.LinAlgError): 

pass 

 

 

def _sqrtm_triu(T, blocksize=64): 

""" 

Matrix square root of an upper triangular matrix. 

 

This is a helper function for `sqrtm` and `logm`. 

 

Parameters 

---------- 

T : (N, N) array_like upper triangular 

Matrix whose square root to evaluate 

blocksize : int, optional 

If the blocksize is not degenerate with respect to the 

size of the input array, then use a blocked algorithm. (Default: 64) 

 

Returns 

------- 

sqrtm : (N, N) ndarray 

Value of the sqrt function at `T` 

 

References 

---------- 

.. [1] Edvin Deadman, Nicholas J. Higham, Rui Ralha (2013) 

"Blocked Schur Algorithms for Computing the Matrix Square Root, 

Lecture Notes in Computer Science, 7782. pp. 171-182. 

 

""" 

T_diag = np.diag(T) 

keep_it_real = np.isrealobj(T) and np.min(T_diag) >= 0 

if not keep_it_real: 

T_diag = T_diag.astype(complex) 

R = np.diag(np.sqrt(T_diag)) 

 

# Compute the number of blocks to use; use at least one block. 

n, n = T.shape 

nblocks = max(n // blocksize, 1) 

 

# Compute the smaller of the two sizes of blocks that 

# we will actually use, and compute the number of large blocks. 

bsmall, nlarge = divmod(n, nblocks) 

blarge = bsmall + 1 

nsmall = nblocks - nlarge 

if nsmall * bsmall + nlarge * blarge != n: 

raise Exception('internal inconsistency') 

 

# Define the index range covered by each block. 

start_stop_pairs = [] 

start = 0 

for count, size in ((nsmall, bsmall), (nlarge, blarge)): 

for i in range(count): 

start_stop_pairs.append((start, start + size)) 

start += size 

 

# Within-block interactions. 

for start, stop in start_stop_pairs: 

for j in range(start, stop): 

for i in range(j-1, start-1, -1): 

s = 0 

if j - i > 1: 

s = R[i, i+1:j].dot(R[i+1:j, j]) 

denom = R[i, i] + R[j, j] 

num = T[i, j] - s 

if denom != 0: 

R[i, j] = (T[i, j] - s) / denom 

elif denom == 0 and num == 0: 

R[i, j] = 0 

else: 

raise SqrtmError('failed to find the matrix square root') 

 

# Between-block interactions. 

for j in range(nblocks): 

jstart, jstop = start_stop_pairs[j] 

for i in range(j-1, -1, -1): 

istart, istop = start_stop_pairs[i] 

S = T[istart:istop, jstart:jstop] 

if j - i > 1: 

S = S - R[istart:istop, istop:jstart].dot(R[istop:jstart, 

jstart:jstop]) 

 

# Invoke LAPACK. 

# For more details, see the solve_sylvester implemention 

# and the fortran dtrsyl and ztrsyl docs. 

Rii = R[istart:istop, istart:istop] 

Rjj = R[jstart:jstop, jstart:jstop] 

if keep_it_real: 

x, scale, info = dtrsyl(Rii, Rjj, S) 

else: 

x, scale, info = ztrsyl(Rii, Rjj, S) 

R[istart:istop, jstart:jstop] = x * scale 

 

# Return the matrix square root. 

return R 

 

 

def sqrtm(A, disp=True, blocksize=64): 

""" 

Matrix square root. 

 

Parameters 

---------- 

A : (N, N) array_like 

Matrix whose square root to evaluate 

disp : bool, optional 

Print warning if error in the result is estimated large 

instead of returning estimated error. (Default: True) 

blocksize : integer, optional 

If the blocksize is not degenerate with respect to the 

size of the input array, then use a blocked algorithm. (Default: 64) 

 

Returns 

------- 

sqrtm : (N, N) ndarray 

Value of the sqrt function at `A` 

 

errest : float 

(if disp == False) 

 

Frobenius norm of the estimated error, ||err||_F / ||A||_F 

 

References 

---------- 

.. [1] Edvin Deadman, Nicholas J. Higham, Rui Ralha (2013) 

"Blocked Schur Algorithms for Computing the Matrix Square Root, 

Lecture Notes in Computer Science, 7782. pp. 171-182. 

 

Examples 

-------- 

>>> from scipy.linalg import sqrtm 

>>> a = np.array([[1.0, 3.0], [1.0, 4.0]]) 

>>> r = sqrtm(a) 

>>> r 

array([[ 0.75592895, 1.13389342], 

[ 0.37796447, 1.88982237]]) 

>>> r.dot(r) 

array([[ 1., 3.], 

[ 1., 4.]]) 

 

""" 

A = _asarray_validated(A, check_finite=True, as_inexact=True) 

if len(A.shape) != 2: 

raise ValueError("Non-matrix input to matrix function.") 

if blocksize < 1: 

raise ValueError("The blocksize should be at least 1.") 

keep_it_real = np.isrealobj(A) 

if keep_it_real: 

T, Z = schur(A) 

if not np.array_equal(T, np.triu(T)): 

T, Z = rsf2csf(T, Z) 

else: 

T, Z = schur(A, output='complex') 

failflag = False 

try: 

R = _sqrtm_triu(T, blocksize=blocksize) 

ZH = np.conjugate(Z).T 

X = Z.dot(R).dot(ZH) 

except SqrtmError: 

failflag = True 

X = np.empty_like(A) 

X.fill(np.nan) 

 

if disp: 

if failflag: 

print("Failed to find a square root.") 

return X 

else: 

try: 

arg2 = norm(X.dot(X) - A, 'fro')**2 / norm(A, 'fro') 

except ValueError: 

# NaNs in matrix 

arg2 = np.inf 

 

return X, arg2