def Af(u, v, alpha=1.0, beta=0.0, trans="N"):
# v := alpha * A(u) + beta * v if trans is 'N'
# v := alpha * A'(u) + beta * v if trans is 'T'
blas.scal(beta, v)
if trans == "N":
sgemv(A,
u,
v,
n=ns,
m=ms,
alpha=alpha,
beta=1.0,
trans="T",
offsetx=0)
elif trans == "T":
sgemv(A,
u,
v,
n=ns,
m=ms,
alpha=alpha,
beta=1.0,
trans="N",
offsetx=0)
def __cngrnc(r, x, alpha = 1.0, trans = 'N', offsetx = 0):
"""
In-place congruence transformation
x := alpha * r * x * r' if trans = 'N'.
x := alpha * r' * x * r if trans = 'T'.
r is a square n x n-matrix.
x is a square n x n-matrix or n^2-vector.
"""
n = r.size[0]
# Scale diagonal of x by 1/2.
blas.scal(0.5, x, inc = n+1, offset = offsetx)
# a := r*tril(x) if trans is 'N', a := tril(x)*r if trans is 'T'.
a = +r
if trans == 'N':
blas.trmm(x, a, side = 'R', m = n, n = n, ldA = n, ldB = n,
offsetA = offsetx)
else:
blas.trmm(x, a, side = 'L', m = n, n = n, ldA = n, ldB = n,
offsetA = offsetx)
# x := alpha * (a * r' + r * a')
# = alpha * r * (tril(x) + tril(x)') * r' if trans is 'N'.
# x := alpha * (a' * r + r' * a)
# = alpha * r' * (tril(x)' + tril(x)) * r if trans = 'T'.
blas.syr2k(r, a, x, trans = trans, alpha = alpha, n = n, k = n,
ldB = n, ldC = n, offsetC = offsetx)
def Af(u, v, alpha=1.0, beta=0.0, trans="N"):
# v := alpha * A(u) + beta * v if trans is 'N'
# v := alpha * A'(u) + beta * v if trans is 'T'
blas.scal(beta, v)
if trans == "N":
blas.axpy(u, v, alpha=alpha, n=nl / 2)
blas.axpy(u, v, alpha=-alpha, offsetx=nl / 2, n=nl / 2)
sgemv(A,
u,
v,
n=ns,
m=ms,
alpha=alpha,
beta=1.0,
trans="T",
offsetx=nl)
elif trans == "T":
blas.axpy(u, v, alpha=alpha, n=nl / 2)
blas.axpy(u, v, alpha=-alpha, offsety=nl / 2, n=nl / 2)
sgemv(A,
u,
v,
n=ns,
m=ms,
alpha=alpha,
beta=1.0,
trans="N",
offsety=nl)
def Gf(u, v, alpha=1.0, beta=0.0, trans='N'):
# v = -alpha*u + beta * v
# u and v are vectors representing N symmetric matrices in the
# cvxopt format.
blas.scal(beta, v)
blas.axpy(u, v, alpha=-alpha)
def classifier2(Y, soft=False):
M = Y.size[0]
# K = Y*X' / sigma
K = matrix(0.0, (width, M))
blas.gemm(X, Y, K, transB='T', alpha=1.0 / sigma, m=width)
# c[i] = ||Yi||^2 / sigma
ones = matrix(1.0, (max(width, n, M), 1))
c = Y**2 * ones[:n]
blas.scal(1.0 / sigma, c)
# Kij := Kij - 0.5 * (ci + aj)
# = || yi - xj ||^2 / (2*sigma)
blas.ger(ones[:width], c, K, alpha=-0.5)
blas.ger(a[:width], ones[:M], K, alpha=-0.5)
# Kij = exp(Kij)
K = exp(K)
# complete K
lapack.potrs(L11, K)
K = matrix([K, matrix(0., (N - width, M))], (N, M))
chompack.trsm(Lc, K, trans='N')
chompack.trsm(Lc, K, trans='T')
x = matrix(b, (M, 1))
blas.gemv(K, z, x, trans='T', beta=1.0)
if soft: return x
else: return matrix([2 * (xk > 0.0) - 1 for xk in x])
def G(u, v, alpha = 1.0, beta = 0.0, trans = 'N'):
"""
If trans is 'N':
v[:msq] := alpha * (A*u[0] - u[1][:]) + beta * v[:msq]
v[msq:] := -alpha * u[1][:] + beta * v[msq:].
If trans is 'T':
v[0] := alpha * A' * u[:msq] + beta * v[0]
v[1][:] := alpha * (-u[:msq] - u[msq:]) + beta * v[1][:].
"""
if trans == 'N':
blas.gemv(A, u[0], v, alpha = alpha, beta = beta)
blas.axpy(u[1], v, alpha = -alpha)
blas.scal(beta, v, offset = msq)
blas.axpy(u[1], v, alpha = -alpha, offsety = msq)
else:
misc.sgemv(A, u, v[0], dims = {'l': 0, 'q': [], 's': [m]},
alpha = alpha, beta = beta, trans = 'T')
blas.scal(beta, v[1])
blas.axpy(u, v[1], alpha = -alpha, n = msq)
blas.axpy(u, v[1], alpha = -alpha, n = msq, offsetx = msq)
def __imul__(self, a):
if isinstance(a,float):
blas.scal(a, self.blkval)
elif isinstance(a, int) or isinstance(a, long):
blas.scal(float(a), self.blkval)
else:
raise NotImplementedError("only scalar multiplication has been implemented")
return self
def __imul__(self, a):
if isinstance(a, float):
blas.scal(a, self.blkval)
elif isinstance(a, int) or isinstance(a, long):
blas.scal(float(a), self.blkval)
else:
raise NotImplementedError(
"only scalar multiplication has been implemented")
return self
def syr2(X, u, v, alpha=1.0, beta=1.0, reordered=False):
r"""
Computes the projected rank 2 update of a cspmatrix X
.. math::
X := \alpha*P(u v^T + v u^T) + \beta X.
"""
assert X.is_factor is False, "cspmatrix factor object"
symb = X.symb
n = symb.n
snptr = symb.snptr
snode = symb.snode
blkval = X.blkval
blkptr = symb.blkptr
relptr = symb.relptr
snrowidx = symb.snrowidx
sncolptr = symb.sncolptr
if symb.p is not None and reordered is False:
up = u[symb.p]
vp = v[symb.p]
else:
up = u
vp = v
for k in range(symb.Nsn):
nn = snptr[k + 1] - snptr[k]
na = relptr[k + 1] - relptr[k]
nj = na + nn
for i in range(nn):
blas.scal(beta, blkval, n=nj - i, offset=blkptr[k] + (nj + 1) * i)
uk = up[snrowidx[sncolptr[k]:sncolptr[k + 1]]]
vk = vp[snrowidx[sncolptr[k]:sncolptr[k + 1]]]
blas.syr2(uk, vk, blkval, n=nn, offsetA=blkptr[k], ldA=nj, alpha=alpha)
blas.ger(uk,
vk,
blkval,
m=na,
n=nn,
offsetx=nn,
offsetA=blkptr[k] + nn,
ldA=nj,
alpha=alpha)
blas.ger(vk,
uk,
blkval,
m=na,
n=nn,
offsetx=nn,
offsetA=blkptr[k] + nn,
ldA=nj,
alpha=alpha)
return
def G(u, vvV, alpha=1.0, beta=0.0, trans="N"):
"""
v := alpha*[I, -I; -I, -I] * u + beta * v (trans = 'N' or 'T')
"""
blas.scal(beta, vvV)
blas.axpy(u, vvV, n=iC, alpha=alpha)
blas.axpy(u, vvV, n=iC, alpha=-alpha, offsetx=iC)
blas.axpy(u, vvV, n=iC, alpha=-alpha, offsety=iC)
blas.axpy(u, vvV, n=iC, alpha=-alpha, offsetx=iC, offsety=iC)
def G(u, vvV, alpha=1.0, beta=0.0, trans='N'):
"""
v := alpha*[I, -I; -I, -I] * u + beta * v (trans = 'N' or 'T')
"""
blas.scal(beta, vvV)
blas.axpy(u, vvV, n=iC, alpha=alpha)
blas.axpy(u, vvV, n=iC, alpha=-alpha, offsetx=iC)
blas.axpy(u, vvV, n=iC, alpha=-alpha, offsety=iC)
blas.axpy(u, vvV, n=iC, alpha=-alpha, offsetx=iC, offsety=iC)
def G(u, v, alpha=1.0, beta=0.0, trans='N'):
"""
v := alpha*[I, -I; -I, -I] * u + beta * v (trans = 'N' or 'T')
"""
blas.scal(beta, v)
blas.axpy(u, v, n=n, alpha=alpha)
blas.axpy(u, v, n=n, alpha=-alpha, offsetx=n)
blas.axpy(u, v, n=n, alpha=-alpha, offsety=n)
blas.axpy(u, v, n=n, alpha=-alpha, offsetx=n, offsety=n)
def G(u, v, alpha = 1.0, beta = 0.0, trans = 'N'):
"""
v := alpha*[I, -I; -I, -I] * u + beta * v (trans = 'N' or 'T')
"""
blas.scal(beta, v)
blas.axpy(u, v, n = n, alpha = alpha)
blas.axpy(u, v, n = n, alpha = -alpha, offsetx = n)
blas.axpy(u, v, n = n, alpha = -alpha, offsety = n)
blas.axpy(u, v, n = n, alpha = -alpha, offsetx = n, offsety = n)
def factor(W, H = None, Df = None):
blas.scal(0.0, K)
if H is not None: K[:n, :n] = H
K[n:n+p, :n] = A
for k in range(n):
if mnl: g[:mnl] = Df[:,k]
g[mnl:] = G[:,k]
scale(g, W, trans = 'T', inverse = 'I')
pack(g, K, dims, mnl, offsety = k*ldK + n + p)
K[(ldK+1)*(p+n) :: ldK+1] = -1.0
# Add positive regularization in 1x1 block and negative in 2x2 block.
blas.copy(K, Ktilde)
Ktilde[0 : (ldK+1)*n : ldK+1] += EPSILON
Ktilde[(ldK+1)*n :: ldK+1] += -EPSILON
lapack.sytrf(Ktilde, ipiv)
def solve(x, y, z):
# Solve
#
# [ H A' GG'*W^{-1} ] [ ux ] [ bx ]
# [ A 0 0 ] * [ uy [ = [ by ]
# [ W^{-T}*GG 0 -I ] [ W*uz ] [ W^{-T}*bz ]
#
# and return ux, uy, W*uz.
#
# On entry, x, y, z contain bx, by, bz. On exit, they contain
# the solution ux, uy, W*uz.
blas.scal(0.0, sltn)
blas.copy(x, u)
blas.copy(y, u, offsety = n)
scale(z, W, trans = 'T', inverse = 'I')
pack(z, u, dims, mnl, offsety = n + p)
blas.copy(u, r)
# Iterative refinement algorithm:
# Init: sltn = 0, r_0 = [bx; by; W^{-T}*bz]
# 1. u_k = Ktilde^-1 * r_k
# 2. sltn += u_k
# 3. r_k+1 = r - K*sltn
# Repeat until exceed MAX_ITER iterations or ||r|| <= ERROR_BOUND
iteration = 0
resid_norm = 1
while iteration <= MAX_ITER and resid_norm > ERROR_BOUND:
lapack.sytrs(Ktilde, ipiv, u)
blas.axpy(u, sltn, alpha = 1.0)
blas.copy(r, u)
blas.symv(K, sltn, u, alpha = -1.0, beta = 1.0)
resid_norm = math.sqrt(blas.dot(u, u))
iteration += 1
blas.copy(sltn, x, n = n)
blas.copy(sltn, y, offsetx = n, n = p)
unpack(sltn, z, dims, mnl, offsetx = n + p)
return solve
def Fs(x, y, alpha = 1.0, beta = 0.0, trans = 'N'):
"""
y := alpha*(-diag(x)) + beta*y.
"""
if trans=='N':
# x is a vector; y is a matrix.
blas.scal(beta, y)
blas.axpy(x, y, alpha = -alpha, incy = n+1)
else:
# x is a matrix; y is a vector.
blas.scal(beta, y)
blas.axpy(x, y, alpha = -alpha, incx = n+1)
def Fs(x, y, alpha=1.0, beta=0.0, trans='N'):
"""
y := alpha*(-diag(x)) + beta*y.
"""
if trans == 'N':
# x is a vector; y is a matrix.
blas.scal(beta, y)
blas.axpy(x, y, alpha=-alpha, incy=n + 1)
else:
# x is a matrix; y is a vector.
blas.scal(beta, y)
blas.axpy(x, y, alpha=-alpha, incx=n + 1)
def Fgp(x = None, z = None):
if x is None: return mnl, matrix(0.0, (n,1))
f = matrix(0.0, (mnl+1,1))
Df = matrix(0.0, (mnl+1,n))
# y = F*x+g
blas.copy(g, y)
base.gemv(F, x, y, beta=1.0)
if z is not None: H = matrix(0.0, (n,n))
for i, start, stop in ind:
# yi := exp(yi) = exp(Fi*x+gi)
ymax = max(y[start:stop])
y[start:stop] = base.exp(y[start:stop] - ymax)
# fi = log sum yi = log sum exp(Fi*x+gi)
ysum = blas.asum(y, n=stop-start, offset=start)
f[i] = ymax + math.log(ysum)
# yi := yi / sum(yi) = exp(Fi*x+gi) / sum(exp(Fi*x+gi))
blas.scal(1.0/ysum, y, n=stop-start, offset=start)
# gradfi := Fi' * yi
# = Fi' * exp(Fi*x+gi) / sum(exp(Fi*x+gi))
base.gemv(F, y, Df, trans='T', m=stop-start, incy=mnl+1,
offsetA=start, offsetx=start, offsety=i)
if z is not None:
# Hi = Fi' * (diag(yi) - yi*yi') * Fi
# = Fisc' * Fisc
# where
# Fisc = diag(yi)^1/2 * (I - 1*yi') * Fi
# = diag(yi)^1/2 * (Fi - 1*gradfi')
Fsc[:K[i], :] = F[start:stop, :]
for k in range(start,stop):
blas.axpy(Df, Fsc, n=n, alpha=-1.0, incx=mnl+1,
incy=Fsc.size[0], offsetx=i, offsety=k-start)
blas.scal(math.sqrt(y[k]), Fsc, inc=Fsc.size[0],
offset=k-start)
# H += z[i]*Hi = z[i] * Fisc' * Fisc
blas.syrk(Fsc, H, trans='T', k=stop-start, alpha=z[i],
beta=1.0)
if z is None:
return f, Df
return f, Df, H
def G(x, y, alpha=1.0, beta=0.0, trans='N'):
"""
If trans is 'N', x is an N x m matrix, and y is an N*m-vector.
y := alpha * x[:] + beta * y.
If trans is 'T', x is an N*m vector, and y is an N x m matrix.
y[:] := alpha * x + beta * y[:].
"""
blas.scal(beta, y)
blas.axpy(x, y, alpha)
def A(x, y, alpha=1.0, beta=0.0, trans='N'):
"""
If trans is 'N', x is an N x m matrix and y an N-vector.
y := alpha * x * 1_m + beta y.
If trans is 'T', x is an N vector and y an N x m matrix.
y := alpha * x * 1_m' + beta y.
"""
if trans == 'N':
blas.gemv(x, ones, y, alpha=alpha, beta=beta)
else:
blas.scal(beta, y)
blas.ger(x, ones, y, alpha=alpha)
def g(x, y, z):
# Solve
#
# [ K d3 ] [ ux_y ]
# [ ] [ ] =
# [ d3' 1'*d3 ] [ ux_b ]
#
# [ bx_y ] [ D ]
# [ ] - [ ] * D3 * (D2 * bx_v + bx_z - bx_w).
# [ bx_b ] [ d' ]
x[:N] -= mul(d, mul(d3, mul(d2, x[-N:]) + z[:N] - z[-N:]))
x[N] -= blas.dot(d, mul(d3, mul(d2, x[-N:]) + z[:N] - z[-N:]))
# Solve dy1 := K^-1 * x[:N]
blas.copy(x, dy1, n=N)
chompack.trsm(L, dy1, trans='N')
chompack.trsm(L, dy1, trans='T')
# Find ux_y = dy1 - ux_b * dy2 s.t
#
# d3' * ( dy1 - ux_b * dy2 + ux_b ) = x[N]
#
# i.e. x[N] := ( x[N] - d3'* dy1 ) / ( d3'* ( 1 - dy2 ) ).
x[N] = ( x[N] - blas.dot(d3, dy1) ) / \
( blas.asum(d3) - blas.dot(d3, dy2) )
x[:N] = dy1 - x[N] * dy2
# ux_v = D4 * ( bx_v - D1^-1 (bz_z + D * (ux_y + ux_b))
# - D2^-1 * bz_w )
x[-N:] = mul(
d4, x[-N:] - div(z[:N] + mul(d, x[:N] + x[N]), d1) -
div(z[N:], d2))
# uz_z = - D1^-1 * ( bx_z - D * ( ux_y + ux_b ) - ux_v )
# uz_w = - D2^-1 * ( bx_w - uz_w )
z[:N] += base.mul(d, x[:N] + x[N]) + x[-N:]
z[-N:] += x[-N:]
blas.scal(-1.0, z)
# Return W['di'] * uz
blas.tbmv(W['di'], z, n=2 * N, k=0, ldA=1)
def f(x, y, z):
# z := mat(z)
# = mat(bz)
z.size = N, m
# x := x + D .* z
# = bx + mat(bz) ./ d.^2
x += mul(D, z)
# y := y - sum_k (Q + Dk)^-1 * X[:,k]
# = by - sum_k (Q + Dk)^-1 * (bxk + Dk * bzk)
for k in range(m):
blas.symv(H[k], x[:, k], y, alpha=-1.0, beta=1.0)
# y := H^-1 * y
# = -uy
lapack.potrs(S, y)
# x[:,k] := H[k] * (x[:,k] + y)
# = (Q + Dk)^-1 * (bxk + bzk ./ d.^2 + y)
# = ux[:,k]
w = matrix(0.0, (N, 1))
for k in range(m):
# x[:,k] := x[:,k] + y
blas.axpy(y, x, offsety=N * k, n=N)
# w := H[k] * x[:,k]
# = (Q + Dk)^-1 * (bxk + bzk ./ d.^2 + y)
blas.symv(H[k], x, w, offsetx=N * k)
# x[:,k] := w
# = ux[:,k]
blas.copy(w, x, offsety=N * k)
# y := -y
# = uy
blas.scal(-1.0, y)
# z := (x - z) ./ d
blas.axpy(x, z, -1.0)
blas.tbsv(W['d'], z, n=m * N, k=0, ldA=1)
blas.scal(-1.0, z)
z.size = N * m, 1
def classifier(Y, soft=False):
M = Y.size[0]
# K = Y*X' / sigma
K = matrix(0.0, (M, Nr))
blas.gemm(Y, Xr, K, transB='T', alpha=1.0 / sigma)
# c[i] = ||Yi||^2 / sigma
ones = matrix(1.0, (max([M, Nr, n]), 1))
c = Y**2 * ones[:n]
blas.scal(1.0 / sigma, c)
# Kij := Kij - 0.5 * (ci + aj)
# = || yi - xj ||^2 / (2*sigma)
blas.ger(c, ones, K, alpha=-0.5)
blas.ger(ones, a[sv], K, alpha=-0.5)
x = exp(K) * zr + b
if soft: return x
else: return matrix([2 * (xk > 0.0) - 1 for xk in x])
def Af(u, v, alpha=1.0, beta=0.0, trans='N'):
# v := alpha * A(u) + beta * v if trans is 'N'
# v := alpha * A'(u) + beta * v if trans is 'T'
blas.scal(beta, v)
if trans == 'N':
offseti = 0
for i in xrange(M):
offsetj = 0
for j in xrange(N):
if type(A[i][j]) is matrix or type(A[i][j]) is spmatrix:
sgemv(A[i][j],
u,
v,
n=ns[j],
m=ms[i],
trans='T',
alpha=alpha,
beta=1.0,
offsetx=offsetj,
offsety=offseti)
offsetj += ns[j]**2
offseti += ms[i]
else:
offsetj = 0
for j in xrange(N):
offseti = 0
for i in xrange(M):
if type(A[i][j]) is matrix or type(A[i][j]) is spmatrix:
sgemv(A[i][j],
u,
v,
n=ns[j],
m=ms[i],
trans='N',
alpha=alpha,
beta=1.0,
offsetx=offseti,
offsety=offsetj)
offseti += ms[i]
offsetj += ns[j]**2
def Gf(u, v, alpha = 1.0, beta = 0.0, trans = 'N'):
if trans == 'N':
# v[:m] := alpha * G * u[0] + beta * v[:m]
base.gemv(G, u[0], v, alpha = alpha, beta = beta)
# v[m:] := alpha * [-u[1], -A(u[0])'; -A(u[0]), -u[2]]
# + beta * v[m:]
blas.scal(beta, v, offset = m)
v[m + I11] -= alpha * u[1][:]
v[m + I21] -= alpha * A * u[0]
v[m + I22] -= alpha * u[2][:]
else:
# v[0] := alpha * ( G.T * u[:m] - 2.0 * A.T * u[m + I21] )
# + beta v[1]
base.gemv(G, u, v[0], trans = 'T', alpha = alpha, beta = beta)
base.gemv(A, u[m + I21], v[0], trans = 'T', alpha = -2.0*alpha,
beta = 1.0)
# v[1] := -alpha * u[m + I11] + beta * v[1]
blas.scal(beta, v[1])
blas.axpy(u[m + I11], v[1], alpha = -alpha)
# v[2] := -alpha * u[m + I22] + beta * v[2]
blas.scal(beta, v[2])
blas.axpy(u[m + I22], v[2], alpha = -alpha)
def Af(u, v, alpha = 1.0, beta = 0.0, trans = 'N'):
if trans == 'N':
pass
else:
blas.scal(beta, v[0])
blas.scal(beta, v[1])
blas.scal(beta, v[2])
def syr2(X, u, v, alpha = 1.0, beta = 1.0, reordered=False):
r"""
Computes the projected rank 2 update of a cspmatrix X
.. math::
X := \alpha*P(u v^T + v u^T) + \beta X.
"""
assert X.is_factor is False, "cspmatrix factor object"
symb = X.symb
n = symb.n
snptr = symb.snptr
snode = symb.snode
blkval = X.blkval
blkptr = symb.blkptr
relptr = symb.relptr
snrowidx = symb.snrowidx
sncolptr = symb.sncolptr
if symb.p is not None and reordered is False:
up = u[symb.p]
vp = v[symb.p]
else:
up = u
vp = v
for k in range(symb.Nsn):
nn = snptr[k+1]-snptr[k]
na = relptr[k+1]-relptr[k]
nj = na + nn
for i in range(nn): blas.scal(beta, blkval, n = nj-i, offset = blkptr[k]+(nj+1)*i)
uk = up[snrowidx[sncolptr[k]:sncolptr[k+1]]]
vk = vp[snrowidx[sncolptr[k]:sncolptr[k+1]]]
blas.syr2(uk, vk, blkval, n = nn, offsetA = blkptr[k], ldA = nj, alpha = alpha)
blas.ger(uk, vk, blkval, m = na, n = nn, offsetx = nn, offsetA = blkptr[k]+nn, ldA = nj, alpha = alpha)
blas.ger(vk, uk, blkval, m = na, n = nn, offsetx = nn, offsetA = blkptr[k]+nn, ldA = nj, alpha = alpha)
return
def Gf(u, v, alpha=1.0, beta=0.0, trans='N'):
if trans == 'N':
# v[:m] := alpha * G * u[0] + beta * v[:m]
base.gemv(G, u[0], v, alpha=alpha, beta=beta)
# v[m:] := alpha * [-u[1], -A(u[0])'; -A(u[0]), -u[2]]
# + beta * v[m:]
blas.scal(beta, v, offset=m)
v[m + I11] -= alpha * u[1][:]
v[m + I21] -= alpha * A * u[0]
v[m + I22] -= alpha * u[2][:]
else:
# v[0] := alpha * ( G.T * u[:m] - 2.0 * A.T * u[m + I21] )
# + beta v[1]
base.gemv(G, u, v[0], trans='T', alpha=alpha, beta=beta)
base.gemv(A,
u[m + I21],
v[0],
trans='T',
alpha=-2.0 * alpha,
beta=1.0)
# v[1] := -alpha * u[m + I11] + beta * v[1]
blas.scal(beta, v[1])
blas.axpy(u[m + I11], v[1], alpha=-alpha)
# v[2] := -alpha * u[m + I22] + beta * v[2]
blas.scal(beta, v[2])
blas.axpy(u[m + I22], v[2], alpha=-alpha)
def Af(u, v, alpha=1.0, beta=0.0, trans='N'):
if trans == 'N':
pass
else:
blas.scal(beta, v[0])
blas.scal(beta, v[1])
blas.scal(beta, v[2])
def balanced_cut_conelp(A, N, delta=1.0, cut='min'):
"""
Generates problem data for semidefinite relaxation of the
following balance-constrained min/max cut problem
min./max. -Tr(AX)
subject to X{ii} == 1, i = 1,...,n
sum(X) + x == delta^2
x >= 0, X p.s.d.
"""
c = matrix([matrix(-1.0, (N, 1)), -delta**2])
h = matrix([0.0, matrix(A.toarray(), (N**2, 1))], (N**2 + 1, 1))
if cut == 'min':
blas.scal(-1.0, h)
elif not cut == 'max':
raise ValueError('cut must be "min" or "max"')
G = sparse([[
spmatrix(1.0, [1 + i * (N + 1) for i in range(N)], range(N),
(N**2 + 1, N))
], [matrix(1.0, (N**2 + 1, 1))]])
dims = {'l': 1, 'q': [], 's': [N]}
return c, G, h, dims
def factor(W, H=None, Df=None):
blas.scal(0.0, K)
if H is not None:
K[:n, :n] = H
K[n:n+p, :n] = A
for k in range(n):
if mnl:
g[:mnl] = Df[:, k]
g[mnl:] = G[:, k]
scale(g, W, trans='T', inverse='I')
pack(g, K, dims, mnl, offsety=k*ldK + n + p)
K[(ldK+1)*(p+n):: ldK+1] = -1.0
# Add positive regularization in 1x1 block and negative in 2x2 block.
K[0: (ldK+1)*n: ldK+1] += REG_EPS
K[(ldK+1)*n:: ldK+1] += -REG_EPS
lapack.sytrf(K, ipiv)
def solve(x, y, z):
# Solve
#
# [ H A' GG'*W^{-1} ] [ ux ] [ bx ]
# [ A 0 0 ] * [ uy [ = [ by ]
# [ W^{-T}*GG 0 -I ] [ W*uz ] [ W^{-T}*bz ]
#
# and return ux, uy, W*uz.
#
# On entry, x, y, z contain bx, by, bz. On exit, they contain
# the solution ux, uy, W*uz.
blas.copy(x, u)
blas.copy(y, u, offsety=n)
scale(z, W, trans='T', inverse='I')
pack(z, u, dims, mnl, offsety=n + p)
lapack.sytrs(K, ipiv, u)
blas.copy(u, x, n=n)
blas.copy(u, y, offsetx=n, n=p)
unpack(u, z, dims, mnl, offsetx=n + p)
return solve
def factor(W, H=None, Df=None):
blas.scal(0.0, K)
if H is not None:
K[:n, :n] = H
K[n:n+p, :n] = A
for k in range(n):
if mnl:
g[:mnl] = Df[:, k]
g[mnl:] = G[:, k]
scale(g, W, trans='T', inverse='I')
pack(g, K, dims, mnl, offsety=k*ldK + n + p)
K[(ldK+1)*(p+n):: ldK+1] = -1.0
# Add positive regularization in 1x1 block and negative in 2x2 block.
K[0: (ldK+1)*n: ldK+1] += REG_EPS
K[(ldK+1)*n:: ldK+1] += -REG_EPS
lapack.sytrf(K, ipiv)
def solve(x, y, z):
# Solve
#
# [ H A' GG'*W^{-1} ] [ ux ] [ bx ]
# [ A 0 0 ] * [ uy [ = [ by ]
# [ W^{-T}*GG 0 -I ] [ W*uz ] [ W^{-T}*bz ]
#
# and return ux, uy, W*uz.
#
# On entry, x, y, z contain bx, by, bz. On exit, they contain
# the solution ux, uy, W*uz.
blas.copy(x, u)
blas.copy(y, u, offsety=n)
scale(z, W, trans='T', inverse='I')
pack(z, u, dims, mnl, offsety=n + p)
lapack.sytrs(K, ipiv, u)
blas.copy(u, x, n=n)
blas.copy(u, y, offsetx=n, n=p)
unpack(u, z, dims, mnl, offsetx=n + p)
return solve
def __cngrnc(r, x, alpha=1.0, trans='N', offsetx=0):
"""
In-place congruence transformation
x := alpha * r * x * r' if trans = 'N'.
x := alpha * r' * x * r if trans = 'T'.
r is a square n x n-matrix.
x is a square n x n-matrix or n^2-vector.
"""
n = r.size[0]
# Scale diagonal of x by 1/2.
blas.scal(0.5, x, inc=n + 1, offset=offsetx)
# a := r*tril(x) if trans is 'N', a := tril(x)*r if trans is 'T'.
a = +r
if trans == 'N':
blas.trmm(x, a, side='R', m=n, n=n, ldA=n, ldB=n, offsetA=offsetx)
else:
blas.trmm(x, a, side='L', m=n, n=n, ldA=n, ldB=n, offsetA=offsetx)
# x := alpha * (a * r' + r * a')
# = alpha * r * (tril(x) + tril(x)') * r' if trans is 'N'.
# x := alpha * (a' * r + r' * a)
# = alpha * r' * (tril(x)' + tril(x)) * r if trans = 'T'.
blas.syr2k(r,
a,
x,
trans=trans,
alpha=alpha,
n=n,
k=n,
ldB=n,
ldC=n,
offsetC=offsetx)
def G(u, v, alpha=1.0, beta=0.0, trans='N'):
"""
If trans is 'N':
v[:msq] := alpha * (A*u[0] - u[1][:]) + beta * v[:msq]
v[msq:] := -alpha * u[1][:] + beta * v[msq:].
If trans is 'T':
v[0] := alpha * A' * u[:msq] + beta * v[0]
v[1][:] := alpha * (-u[:msq] - u[msq:]) + beta * v[1][:].
"""
if trans == 'N':
blas.gemv(A, u[0], v, alpha=alpha, beta=beta)
blas.axpy(u[1], v, alpha=-alpha)
blas.scal(beta, v, offset=msq)
blas.axpy(u[1], v, alpha=-alpha, offsety=msq)
else:
misc.sgemv(A,
u,
v[0],
dims={
'l': 0,
'q': [],
's': [m]
},
alpha=alpha,
beta=beta,
trans='T')
blas.scal(beta, v[1])
blas.axpy(u, v[1], alpha=-alpha, n=msq)
blas.axpy(u, v[1], alpha=-alpha, n=msq, offsetx=msq)
def fsolve(x, y, z):
"""
Solves the system of equations
[ 0 G'*W^{-1} ] [ ux ] = [ bx ]
[ G -W' ] [ uz ] [ bz ]
"""
# Compute bx := bx + G'*W^{-1}*W^{-T}*bz
v = matrix(0., (N, 1))
for i in range(N):
blas.symv(z, rr, v, ldA=N, offsetA=1, n=N, offsetx=N * i)
x[i] += blas.dot(rr, v, n=N, offsetx=N * i)
blas.symv(z, q, v, ldA=N, offsetA=1, n=N)
x[N] += blas.dot(q, v) + z[0] * W['di'][0]**2
# Solve G'*W^{-1}*W^{-T}*G*ux = bx
lapack.potrs(H, x)
# Compute bz := -W^{-T}*(bz-G*ux)
# z -= G*x
z[1::N + 1] -= x[:-1]
z -= x[-1]
# Apply scaling
z[0] *= -W['di'][0]
blas.scal(0.5, z, n=N, offset=1, inc=N + 1)
tmp = +r
blas.trmm(z, tmp, ldA=N, offsetA=1, n=N, m=N)
blas.syr2k(r,
tmp,
z,
trans='T',
offsetC=1,
ldC=N,
n=N,
k=N,
alpha=-1.0)
def solve(x, y, z):
# Solve
#
# [ H A' GG'*W^{-1} ] [ ux ] [ bx ]
# [ A 0 0 ] * [ uy [ = [ by ]
# [ W^{-T}*GG 0 -I ] [ W*uz ] [ W^{-T}*bz ]
#
# and return ux, uy, W*uz.
#
# On entry, x, y, z contain bx, by, bz. On exit, they contain
# the solution ux, uy, W*uz.
blas.scal(0.0, sltn)
blas.copy(x, u)
blas.copy(y, u, offsety = n)
scale(z, W, trans = 'T', inverse = 'I')
pack(z, u, dims, mnl, offsety = n + p)
blas.copy(u, r)
# Iterative refinement algorithm:
# Init: sltn = 0, r_0 = [bx; by; W^{-T}*bz]
# 1. u_k = Ktilde^-1 * r_k
# 2. sltn += u_k
# 3. r_k+1 = r - K*sltn
# Repeat until exceed MAX_ITER iterations or ||r|| <= ERROR_BOUND
iteration = 0
resid_norm = 1
while iteration <= MAX_ITER and resid_norm > ERROR_BOUND:
lapack.sytrs(Ktilde, ipiv, u)
blas.axpy(u, sltn, alpha = 1.0)
blas.copy(r, u)
blas.symv(K, sltn, u, alpha = -1.0, beta = 1.0)
resid_norm = math.sqrt(blas.dot(u, u))
iteration += 1
blas.copy(sltn, x, n = n)
blas.copy(sltn, y, offsetx = n, n = p)
unpack(sltn, z, dims, mnl, offsetx = n + p)
def kernel_matrix(X,
kernel,
sigma=1.0,
theta=1.0,
degree=1,
V=None,
width=None):
"""
Computes the kernel matrix or a partial kernel matrix.
Input arguments.
X is an N x n matrix.
kernel is a string with values 'linear', 'rfb', 'poly', or 'tanh'.
'linear': k(u,v) = u'*v/sigma.
'rbf': k(u,v) = exp(-||u - v||^2 / (2*sigma)).
'poly': k(u,v) = (u'*v/sigma)**degree.
'tanh': k(u,v) = tanh(u'*v/sigma - theta). kernel is a
sigma and theta are positive numbers.
degree is a positive integer.
V is an N x N sparse matrix (default is None).
width is a positive integer (default is None).
Output.
Q, an N x N matrix or sparse matrix.
If V is a sparse matrix, a partial kernel matrix with the sparsity
pattern V is returned.
If width is specified and V = 'band', a partial kernel matrix
with band sparsity is returned (width is the half-bandwidth).
a, an N x 1 matrix with the products <xi,xi>/sigma.
"""
N, n = X.size
#### dense (full) kernel matrix
if V is None:
if verbose: print("building kernel matrix ..")
# Qij = xi'*xj / sigma
Q = matrix(0.0, (N, N))
blas.syrk(X, Q, alpha=1.0 / sigma)
a = Q[::N + 1] # ai = ||xi||**2 / sigma
if kernel == 'linear':
pass
elif kernel == 'rbf':
# Qij := Qij - 0.5 * ( ai + aj )
# = -||xi - xj||^2 / (2*sigma)
ones = matrix(1.0, (N, 1))
blas.syr2(a, ones, Q, alpha=-0.5)
Q = exp(Q)
elif kernel == 'tanh':
Q = exp(Q - theta)
Q = div(Q - Q**-1, Q + Q**-1)
elif kernel == 'poly':
Q = Q**degree
else:
raise ValueError('invalid kernel type')
#### general sparse partial kernel matrix
elif type(V) is cvxopt.base.spmatrix:
if verbose: print("building projected kernel matrix ...")
Q = +V
base.syrk(X, Q, partial=True, alpha=1.0 / sigma)
# ai = ||xi||**2 / sigma
a = matrix(Q[::N + 1], (N, 1))
if kernel == 'linear':
pass
elif kernel == 'rbf':
ones = matrix(1.0, (N, 1))
# Qij := Qij - 0.5 * ( ai + aj )
# = -||xi - xj||^2 / (2*sigma)
p = chompack.maxcardsearch(V)
symb = chompack.symbolic(Q, p)
Qc = chompack.cspmatrix(symb) + Q
chompack.syr2(Qc, a, ones, alpha=-0.5)
Q = Qc.spmatrix(reordered=False)
Q.V = exp(Q.V)
elif kernel == 'tanh':
v = +Q.V
v = exp(v - theta)
v = div(v - v**-1, v + v**-1)
Q.V = v
elif kernel == 'poly':
Q.V = Q.V**degree
else:
raise ValueError('invalid kernel type')
#### banded partial kernel matrix
elif V == 'band' and width is not None:
# Lower triangular part of band matrix with bandwidth 2*w+1.
if verbose: print("building projected kernel matrix ...")
I = [i for k in range(N) for i in range(k, min(width + k + 1, N))]
J = [k for k in range(N) for i in range(min(width + 1, N - k))]
V = matrix(0.0, (len(I), 1))
oy = 0
for k in range(N): # V[:,k] = Xtrain[k:k+w, :] * Xtrain[k,:].T
m = min(width + 1, N - k)
blas.gemv(X,
X,
V,
m=m,
ldA=N,
incx=N,
offsetA=k,
offsetx=k,
offsety=oy)
oy += m
blas.scal(1.0 / sigma, V)
# ai = ||xi||**2 / sigma
a = matrix(V[[i for i in range(len(I)) if I[i] == J[i]]], (N, 1))
if kernel == 'linear':
Q = spmatrix(V, I, J, (N, N))
elif kernel == 'rbf':
Q = spmatrix(V, I, J, (N, N))
ones = matrix(1.0, (N, 1))
# Qij := Qij - 0.5 * ( ai + aj )
# = -||xi - xj||^2 / (2*sigma)
symb = chompack.symbolic(Q)
Qc = chompack.cspmatrix(symb) + Q
chompack.syr2(Qc, a, ones, alpha=-0.5)
Q = Qc.spmatrix(reordered=False)
Q.V = exp(Q.V)
elif kernel == 'tanh':
V = exp(V - theta)
V = div(V - V**-1, V + V**-1)
Q = spmatrix(V, I, J, (N, N))
elif kernel == 'poly':
Q = spmatrix(V**degree, I, J, (N, N))
else:
raise ValueError('invalid kernel type')
else:
raise TypeError('invalid type V')
return Q, a
def f(x, y, z):
"""
Solve
C * ux + G' * uzl - 2*A'(uzs21) = bx
-uzs11 = bX1
-uzs22 = bX2
G * ux - D^2 * uzl = bzl
[ -uX1 -A(ux)' ] [ uzs11 uzs21' ]
[ ] - T * [ ] * T = bzs.
[ -A(ux) -uX2 ] [ uzs21 uzs22 ]
On entry, x = (bx, bX1, bX2) and z = [ bzl; bzs[:] ].
On exit, x = (ux, uX1, uX2) and z = [ D*uzl; (r'*uzs*r)[:] ].
Define X = uzs21, Z = T * uzs * T:
C * ux + G' * uzl - 2*A'(X) = bx
[ 0 X' ] [ bX1 0 ]
T * [ ] * T - Z = T * [ ] * T
[ X 0 ] [ 0 bX2 ]
G * ux - D^2 * uzl = bzl
[ -uX1 -A(ux)' ] [ Z11 Z21' ]
[ ] - [ ] = bzs
[ -A(ux) -uX2 ] [ Z21 Z22 ]
Return x = (ux, uX1, uX2), z = [ D*uzl; (rti'*Z*rti)[:] ].
We use the congruence transformation
[ V1 0 ] [ T11 T21' ] [ V1' 0 ] [ I S' ]
[ ] [ ] [ ] = [ ]
[ 0 V2' ] [ T21 T22 ] [ 0 V2 ] [ S I ]
and the factorization
X + S * X' * S = L( L'(X) )
to write this as
C * ux + G' * uzl - 2*A'(X) = bx
L'(V2^-1 * X * V1^-1) - L^-1(V2' * Z21 * V1') = bX
G * ux - D^2 * uzl = bzl
[ -uX1 -A(ux)' ] [ Z11 Z21' ]
[ ] - [ ] = bzs,
[ -A(ux) -uX2 ] [ Z21 Z22 ]
or
C * ux + Gs' * uuzl - 2*As'(XX) = bx
XX - ZZ21 = bX
Gs * ux - uuzl = D^-1 * bzl
-As(ux) - ZZ21 = bbzs_21
-uX1 - Z11 = bzs_11
-uX2 - Z22 = bzs_22
if we introduce scaled variables
uuzl = D * uzl
XX = L'(V2^-1 * X * V1^-1)
= L'(V2^-1 * uzs21 * V1^-1)
ZZ21 = L^-1(V2' * Z21 * V1')
and define
bbzs_21 = L^-1(V2' * bzs_21 * V1')
[ bX1 0 ]
bX = L^-1( V2' * (T * [ ] * T)_21 * V1').
[ 0 bX2 ]
Eliminating Z21 gives
C * ux + Gs' * uuzl - 2*As'(XX) = bx
Gs * ux - uuzl = D^-1 * bzl
-As(ux) - XX = bbzs_21 - bX
-uX1 - Z11 = bzs_11
-uX2 - Z22 = bzs_22
and eliminating uuzl and XX gives
H * ux = bx + Gs' * D^-1 * bzl + 2*As'(bX - bbzs_21)
Gs * ux - uuzl = D^-1 * bzl
-As(ux) - XX = bbzs_21 - bX
-uX1 - Z11 = bzs_11
-uX2 - Z22 = bzs_22.
In summary, we can use the following algorithm:
1. bXX := bX - bbzs21
[ bX1 0 ]
= L^-1( V2' * ((T * [ ] * T)_21 - bzs_21) * V1')
[ 0 bX2 ]
2. Solve H * ux = bx + Gs' * D^-1 * bzl + 2*As'(bXX).
3. From ux, compute
uuzl = Gs*ux - D^-1 * bzl and
X = V2 * L^-T(-As(ux) + bXX) * V1.
4. Return ux, uuzl,
rti' * Z * rti = r' * [ -bX1, X'; X, -bX2 ] * r
and uX1 = -Z11 - bzs_11, uX2 = -Z22 - bzs_22.
"""
# Save bzs_11, bzs_22, bzs_21.
lapack.lacpy(z, bz11, uplo = 'L', m = q, n = q, ldA = p+q,
offsetA = m)
lapack.lacpy(z, bz21, m = p, n = q, ldA = p+q, offsetA = m+q)
lapack.lacpy(z, bz22, uplo = 'L', m = p, n = p, ldA = p+q,
offsetA = m + (p+q+1)*q)
# zl := D^-1 * zl
# = D^-1 * bzl
blas.tbmv(W['di'], z, n = m, k = 0, ldA = 1)
# zs := r' * [ bX1, 0; 0, bX2 ] * r.
# zs := [ bX1, 0; 0, bX2 ]
blas.scal(0.0, z, offset = m)
lapack.lacpy(x[1], z, uplo = 'L', m = q, n = q, ldB = p+q,
offsetB = m)
lapack.lacpy(x[2], z, uplo = 'L', m = p, n = p, ldB = p+q,
offsetB = m + (p+q+1)*q)
# scale diagonal of zs by 1/2
blas.scal(0.5, z, inc = p+q+1, offset = m)
# a := tril(zs)*r
blas.copy(r, a)
blas.trmm(z, a, side = 'L', m = p+q, n = p+q, ldA = p+q, ldB =
p+q, offsetA = m)
# zs := a'*r + r'*a
blas.syr2k(r, a, z, trans = 'T', n = p+q, k = p+q, ldB = p+q,
ldC = p+q, offsetC = m)
# bz21 := L^-1( V2' * ((r * zs * r')_21 - bz21) * V1')
#
# [ bX1 0 ]
# = L^-1( V2' * ((T * [ ] * T)_21 - bz21) * V1').
# [ 0 bX2 ]
# a = [ r21 r22 ] * z
# = [ r21 r22 ] * r' * [ bX1, 0; 0, bX2 ] * r
# = [ T21 T22 ] * [ bX1, 0; 0, bX2 ] * r
blas.symm(z, r, a, side = 'R', m = p, n = p+q, ldA = p+q,
ldC = p+q, offsetB = q)
# bz21 := -bz21 + a * [ r11, r12 ]'
# = -bz21 + (T * [ bX1, 0; 0, bX2 ] * T)_21
blas.gemm(a, r, bz21, transB = 'T', m = p, n = q, k = p+q,
beta = -1.0, ldA = p+q, ldC = p)
# bz21 := V2' * bz21 * V1'
# = V2' * (-bz21 + (T*[bX1, 0; 0, bX2]*T)_21) * V1'
blas.gemm(V2, bz21, tmp, transA = 'T', m = p, n = q, k = p,
ldB = p)
blas.gemm(tmp, V1, bz21, transB = 'T', m = p, n = q, k = q,
ldC = p)
# bz21[:] := D * (I-P) * bz21[:]
# = L^-1 * bz21[:]
# = bXX[:]
blas.copy(bz21, tmp)
base.gemv(P, bz21, tmp, alpha = -1.0, beta = 1.0)
base.gemv(D, tmp, bz21)
# Solve H * ux = bx + Gs' * D^-1 * bzl + 2*As'(bXX).
# x[0] := x[0] + Gs'*zl + 2*As'(bz21)
# = bx + G' * D^-1 * bzl + 2 * As'(bXX)
blas.gemv(Gs, z, x[0], trans = 'T', alpha = 1.0, beta = 1.0)
blas.gemv(As, bz21, x[0], trans = 'T', alpha = 2.0, beta = 1.0)
# x[0] := H \ x[0]
# = ux
lapack.potrs(H, x[0])
# uuzl = Gs*ux - D^-1 * bzl
blas.gemv(Gs, x[0], z, alpha = 1.0, beta = -1.0)
# bz21 := V2 * L^-T(-As(ux) + bz21) * V1
# = X
blas.gemv(As, x[0], bz21, alpha = -1.0, beta = 1.0)
blas.tbsv(DV, bz21, n = p*q, k = 0, ldA = 1)
blas.copy(bz21, tmp)
base.gemv(P, tmp, bz21, alpha = -1.0, beta = 1.0, trans = 'T')
blas.gemm(V2, bz21, tmp)
blas.gemm(tmp, V1, bz21)
# zs := -zs + r' * [ 0, X'; X, 0 ] * r
# = r' * [ -bX1, X'; X, -bX2 ] * r.
# a := bz21 * [ r11, r12 ]
# = X * [ r11, r12 ]
blas.gemm(bz21, r, a, m = p, n = p+q, k = q, ldA = p, ldC = p+q)
# z := -z + [ r21, r22 ]' * a + a' * [ r21, r22 ]
# = rti' * uzs * rti
blas.syr2k(r, a, z, trans = 'T', beta = -1.0, n = p+q, k = p,
offsetA = q, offsetC = m, ldB = p+q, ldC = p+q)
# uX1 = -Z11 - bzs_11
# = -(r*zs*r')_11 - bzs_11
# uX2 = -Z22 - bzs_22
# = -(r*zs*r')_22 - bzs_22
blas.copy(bz11, x[1])
blas.copy(bz22, x[2])
# scale diagonal of zs by 1/2
blas.scal(0.5, z, inc = p+q+1, offset = m)
# a := r*tril(zs)
blas.copy(r, a)
blas.trmm(z, a, side = 'R', m = p+q, n = p+q, ldA = p+q, ldB =
p+q, offsetA = m)
# x[1] := -x[1] - a[:q,:] * r[:q, :]' - r[:q,:] * a[:q,:]'
# = -bzs_11 - (r*zs*r')_11
blas.syr2k(a, r, x[1], n = q, alpha = -1.0, beta = -1.0)
# x[2] := -x[2] - a[q:,:] * r[q:, :]' - r[q:,:] * a[q:,:]'
# = -bzs_22 - (r*zs*r')_22
blas.syr2k(a, r, x[2], n = p, alpha = -1.0, beta = -1.0,
offsetA = q, offsetB = q)
# scale diagonal of zs by 1/2
blas.scal(2.0, z, inc = p+q+1, offset = m)
def F(W):
"""
Create a solver for the linear equations
C * ux + G' * uzl - 2*A'(uzs21) = bx
-uzs11 = bX1
-uzs22 = bX2
G * ux - Dl^2 * uzl = bzl
[ -uX1 -A(ux)' ] [ uzs11 uzs21' ]
[ ] - r*r' * [ ] * r*r' = bzs
[ -A(ux) -uX2 ] [ uzs21 uzs22 ]
where Dl = diag(W['l']), r = W['r'][0].
On entry, x = (bx, bX1, bX2) and z = [ bzl; bzs[:] ].
On exit, x = (ux, uX1, uX2) and z = [ Dl*uzl; (r'*uzs*r)[:] ].
1. Compute matrices V1, V2 such that (with T = r*r')
[ V1 0 ] [ T11 T21' ] [ V1' 0 ] [ I S' ]
[ ] [ ] [ ] = [ ]
[ 0 V2' ] [ T21 T22 ] [ 0 V2 ] [ S I ]
and S = [ diag(s); 0 ], s a positive q-vector.
2. Factor the mapping X -> X + S * X' * S:
X + S * X' * S = L( L'( X )).
3. Compute scaled mappings: a matrix As with as its columns the
coefficients of the scaled mapping
L^-1( V2' * A() * V1' )
and the matrix Gs = Dl^-1 * G.
4. Cholesky factorization of H = C + Gs'*Gs + 2*As'*As.
"""
# 1. Compute V1, V2, s.
r = W['r'][0]
# LQ factorization R[:q, :] = L1 * Q1.
lapack.lacpy(r, Q1, m = q)
lapack.gelqf(Q1, tau1)
lapack.lacpy(Q1, L1, n = q, uplo = 'L')
lapack.orglq(Q1, tau1)
# LQ factorization R[q:, :] = L2 * Q2.
lapack.lacpy(r, Q2, m = p, offsetA = q)
lapack.gelqf(Q2, tau2)
lapack.lacpy(Q2, L2, n = p, uplo = 'L')
lapack.orglq(Q2, tau2)
# V2, V1, s are computed from an SVD: if
#
# Q2 * Q1' = U * diag(s) * V',
#
# then V1 = V' * L1^-1 and V2 = L2^-T * U.
# T21 = Q2 * Q1.T
blas.gemm(Q2, Q1, T21, transB = 'T')
# SVD T21 = U * diag(s) * V'. Store U in V2 and V' in V1.
lapack.gesvd(T21, s, jobu = 'A', jobvt = 'A', U = V2, Vt = V1)
# # Q2 := Q2 * Q1' without extracting Q1; store T21 in Q2
# this will requires lapack.ormlq or lapack.unmlq
# V2 = L2^-T * U
blas.trsm(L2, V2, transA = 'T')
# V1 = V' * L1^-1
blas.trsm(L1, V1, side = 'R')
# 2. Factorization X + S * X' * S = L( L'( X )).
#
# The factor L is stored as a diagonal matrix D and a sparse lower
# triangular matrix P, such that
#
# L(X)[:] = D**-1 * (I + P) * X[:]
# L^-1(X)[:] = D * (I - P) * X[:].
# SS is q x q with SS[i,j] = si*sj.
blas.scal(0.0, SS)
blas.syr(s, SS)
# For a p x q matrix X, P*X[:] is Y[:] where
#
# Yij = si * sj * Xji if i < j
# = 0 otherwise.
#
P.V = SS[Itril2]
# For a p x q matrix X, D*X[:] is Y[:] where
#
# Yij = Xij / sqrt( 1 - si^2 * sj^2 ) if i < j
# = Xii / sqrt( 1 + si^2 ) if i = j
# = Xij otherwise.
#
DV[Idiag] = sqrt(1.0 + SS[::q+1])
DV[Itriu] = sqrt(1.0 - SS[Itril3]**2)
D.V = DV**-1
# 3. Scaled linear mappings
# Ask := V2' * Ask * V1'
blas.scal(0.0, As)
base.axpy(A, As)
for i in xrange(n):
# tmp := V2' * As[i, :]
blas.gemm(V2, As, tmp, transA = 'T', m = p, n = q, k = p,
ldB = p, offsetB = i*p*q)
# As[:,i] := tmp * V1'
blas.gemm(tmp, V1, As, transB = 'T', m = p, n = q, k = q,
ldC = p, offsetC = i*p*q)
# As := D * (I - P) * As
# = L^-1 * As.
blas.copy(As, As2)
base.gemm(P, As, As2, alpha = -1.0, beta = 1.0)
base.gemm(D, As2, As)
# Gs := Dl^-1 * G
blas.scal(0.0, Gs)
base.axpy(G, Gs)
for k in xrange(n):
blas.tbmv(W['di'], Gs, n = m, k = 0, ldA = 1, offsetx = k*m)
# 4. Cholesky factorization of H = C + Gs' * Gs + 2 * As' * As.
blas.syrk(As, H, trans = 'T', alpha = 2.0)
blas.syrk(Gs, H, trans = 'T', beta = 1.0)
base.axpy(C, H)
lapack.potrf(H)
def f(x, y, z):
"""
Solve
C * ux + G' * uzl - 2*A'(uzs21) = bx
-uzs11 = bX1
-uzs22 = bX2
G * ux - D^2 * uzl = bzl
[ -uX1 -A(ux)' ] [ uzs11 uzs21' ]
[ ] - T * [ ] * T = bzs.
[ -A(ux) -uX2 ] [ uzs21 uzs22 ]
On entry, x = (bx, bX1, bX2) and z = [ bzl; bzs[:] ].
On exit, x = (ux, uX1, uX2) and z = [ D*uzl; (r'*uzs*r)[:] ].
Define X = uzs21, Z = T * uzs * T:
C * ux + G' * uzl - 2*A'(X) = bx
[ 0 X' ] [ bX1 0 ]
T * [ ] * T - Z = T * [ ] * T
[ X 0 ] [ 0 bX2 ]
G * ux - D^2 * uzl = bzl
[ -uX1 -A(ux)' ] [ Z11 Z21' ]
[ ] - [ ] = bzs
[ -A(ux) -uX2 ] [ Z21 Z22 ]
Return x = (ux, uX1, uX2), z = [ D*uzl; (rti'*Z*rti)[:] ].
We use the congruence transformation
[ V1 0 ] [ T11 T21' ] [ V1' 0 ] [ I S' ]
[ ] [ ] [ ] = [ ]
[ 0 V2' ] [ T21 T22 ] [ 0 V2 ] [ S I ]
and the factorization
X + S * X' * S = L( L'(X) )
to write this as
C * ux + G' * uzl - 2*A'(X) = bx
L'(V2^-1 * X * V1^-1) - L^-1(V2' * Z21 * V1') = bX
G * ux - D^2 * uzl = bzl
[ -uX1 -A(ux)' ] [ Z11 Z21' ]
[ ] - [ ] = bzs,
[ -A(ux) -uX2 ] [ Z21 Z22 ]
or
C * ux + Gs' * uuzl - 2*As'(XX) = bx
XX - ZZ21 = bX
Gs * ux - uuzl = D^-1 * bzl
-As(ux) - ZZ21 = bbzs_21
-uX1 - Z11 = bzs_11
-uX2 - Z22 = bzs_22
if we introduce scaled variables
uuzl = D * uzl
XX = L'(V2^-1 * X * V1^-1)
= L'(V2^-1 * uzs21 * V1^-1)
ZZ21 = L^-1(V2' * Z21 * V1')
and define
bbzs_21 = L^-1(V2' * bzs_21 * V1')
[ bX1 0 ]
bX = L^-1( V2' * (T * [ ] * T)_21 * V1').
[ 0 bX2 ]
Eliminating Z21 gives
C * ux + Gs' * uuzl - 2*As'(XX) = bx
Gs * ux - uuzl = D^-1 * bzl
-As(ux) - XX = bbzs_21 - bX
-uX1 - Z11 = bzs_11
-uX2 - Z22 = bzs_22
and eliminating uuzl and XX gives
H * ux = bx + Gs' * D^-1 * bzl + 2*As'(bX - bbzs_21)
Gs * ux - uuzl = D^-1 * bzl
-As(ux) - XX = bbzs_21 - bX
-uX1 - Z11 = bzs_11
-uX2 - Z22 = bzs_22.
In summary, we can use the following algorithm:
1. bXX := bX - bbzs21
[ bX1 0 ]
= L^-1( V2' * ((T * [ ] * T)_21 - bzs_21) * V1')
[ 0 bX2 ]
2. Solve H * ux = bx + Gs' * D^-1 * bzl + 2*As'(bXX).
3. From ux, compute
uuzl = Gs*ux - D^-1 * bzl and
X = V2 * L^-T(-As(ux) + bXX) * V1.
4. Return ux, uuzl,
rti' * Z * rti = r' * [ -bX1, X'; X, -bX2 ] * r
and uX1 = -Z11 - bzs_11, uX2 = -Z22 - bzs_22.
"""
# Save bzs_11, bzs_22, bzs_21.
lapack.lacpy(z, bz11, uplo = 'L', m = q, n = q, ldA = p+q,
offsetA = m)
lapack.lacpy(z, bz21, m = p, n = q, ldA = p+q, offsetA = m+q)
lapack.lacpy(z, bz22, uplo = 'L', m = p, n = p, ldA = p+q,
offsetA = m + (p+q+1)*q)
# zl := D^-1 * zl
# = D^-1 * bzl
blas.tbmv(W['di'], z, n = m, k = 0, ldA = 1)
# zs := r' * [ bX1, 0; 0, bX2 ] * r.
# zs := [ bX1, 0; 0, bX2 ]
blas.scal(0.0, z, offset = m)
lapack.lacpy(x[1], z, uplo = 'L', m = q, n = q, ldB = p+q,
offsetB = m)
lapack.lacpy(x[2], z, uplo = 'L', m = p, n = p, ldB = p+q,
offsetB = m + (p+q+1)*q)
# scale diagonal of zs by 1/2
blas.scal(0.5, z, inc = p+q+1, offset = m)
# a := tril(zs)*r
blas.copy(r, a)
blas.trmm(z, a, side = 'L', m = p+q, n = p+q, ldA = p+q, ldB =
p+q, offsetA = m)
# zs := a'*r + r'*a
blas.syr2k(r, a, z, trans = 'T', n = p+q, k = p+q, ldB = p+q,
ldC = p+q, offsetC = m)
# bz21 := L^-1( V2' * ((r * zs * r')_21 - bz21) * V1')
#
# [ bX1 0 ]
# = L^-1( V2' * ((T * [ ] * T)_21 - bz21) * V1').
# [ 0 bX2 ]
# a = [ r21 r22 ] * z
# = [ r21 r22 ] * r' * [ bX1, 0; 0, bX2 ] * r
# = [ T21 T22 ] * [ bX1, 0; 0, bX2 ] * r
blas.symm(z, r, a, side = 'R', m = p, n = p+q, ldA = p+q,
ldC = p+q, offsetB = q)
# bz21 := -bz21 + a * [ r11, r12 ]'
# = -bz21 + (T * [ bX1, 0; 0, bX2 ] * T)_21
blas.gemm(a, r, bz21, transB = 'T', m = p, n = q, k = p+q,
beta = -1.0, ldA = p+q, ldC = p)
# bz21 := V2' * bz21 * V1'
# = V2' * (-bz21 + (T*[bX1, 0; 0, bX2]*T)_21) * V1'
blas.gemm(V2, bz21, tmp, transA = 'T', m = p, n = q, k = p,
ldB = p)
blas.gemm(tmp, V1, bz21, transB = 'T', m = p, n = q, k = q,
ldC = p)
# bz21[:] := D * (I-P) * bz21[:]
# = L^-1 * bz21[:]
# = bXX[:]
blas.copy(bz21, tmp)
base.gemv(P, bz21, tmp, alpha = -1.0, beta = 1.0)
base.gemv(D, tmp, bz21)
# Solve H * ux = bx + Gs' * D^-1 * bzl + 2*As'(bXX).
# x[0] := x[0] + Gs'*zl + 2*As'(bz21)
# = bx + G' * D^-1 * bzl + 2 * As'(bXX)
blas.gemv(Gs, z, x[0], trans = 'T', alpha = 1.0, beta = 1.0)
blas.gemv(As, bz21, x[0], trans = 'T', alpha = 2.0, beta = 1.0)
# x[0] := H \ x[0]
# = ux
lapack.potrs(H, x[0])
# uuzl = Gs*ux - D^-1 * bzl
blas.gemv(Gs, x[0], z, alpha = 1.0, beta = -1.0)
# bz21 := V2 * L^-T(-As(ux) + bz21) * V1
# = X
blas.gemv(As, x[0], bz21, alpha = -1.0, beta = 1.0)
blas.tbsv(DV, bz21, n = p*q, k = 0, ldA = 1)
blas.copy(bz21, tmp)
base.gemv(P, tmp, bz21, alpha = -1.0, beta = 1.0, trans = 'T')
blas.gemm(V2, bz21, tmp)
blas.gemm(tmp, V1, bz21)
# zs := -zs + r' * [ 0, X'; X, 0 ] * r
# = r' * [ -bX1, X'; X, -bX2 ] * r.
# a := bz21 * [ r11, r12 ]
# = X * [ r11, r12 ]
blas.gemm(bz21, r, a, m = p, n = p+q, k = q, ldA = p, ldC = p+q)
# z := -z + [ r21, r22 ]' * a + a' * [ r21, r22 ]
# = rti' * uzs * rti
blas.syr2k(r, a, z, trans = 'T', beta = -1.0, n = p+q, k = p,
offsetA = q, offsetC = m, ldB = p+q, ldC = p+q)
# uX1 = -Z11 - bzs_11
# = -(r*zs*r')_11 - bzs_11
# uX2 = -Z22 - bzs_22
# = -(r*zs*r')_22 - bzs_22
blas.copy(bz11, x[1])
blas.copy(bz22, x[2])
# scale diagonal of zs by 1/2
blas.scal(0.5, z, inc = p+q+1, offset = m)
# a := r*tril(zs)
blas.copy(r, a)
blas.trmm(z, a, side = 'R', m = p+q, n = p+q, ldA = p+q, ldB =
p+q, offsetA = m)
# x[1] := -x[1] - a[:q,:] * r[:q, :]' - r[:q,:] * a[:q,:]'
# = -bzs_11 - (r*zs*r')_11
blas.syr2k(a, r, x[1], n = q, alpha = -1.0, beta = -1.0)
# x[2] := -x[2] - a[q:,:] * r[q:, :]' - r[q:,:] * a[q:,:]'
# = -bzs_22 - (r*zs*r')_22
blas.syr2k(a, r, x[2], n = p, alpha = -1.0, beta = -1.0,
offsetA = q, offsetB = q)
# scale diagonal of zs by 1/2
blas.scal(2.0, z, inc = p+q+1, offset = m)
return f
def F(W):
"""
Generate a solver for
A'(uz0) = bx[0]
-uz0 - uz1 = bx[1]
A(ux[0]) - ux[1] - r0*r0' * uz0 * r0*r0' = bz0
- ux[1] - r1*r1' * uz1 * r1*r1' = bz1.
uz0, uz1, bz0, bz1 are symmetric m x m-matrices.
ux[0], bx[0] are n-vectors.
ux[1], bx[1] are symmetric m x m-matrices.
We first calculate a congruence that diagonalizes r0*r0' and r1*r1':
U' * r0 * r0' * U = I, U' * r1 * r1' * U = S.
We then make a change of variables
usx[0] = ux[0],
usx[1] = U' * ux[1] * U
usz0 = U^-1 * uz0 * U^-T
usz1 = U^-1 * uz1 * U^-T
and define
As() = U' * A() * U'
bsx[1] = U^-1 * bx[1] * U^-T
bsz0 = U' * bz0 * U
bsz1 = U' * bz1 * U.
This gives
As'(usz0) = bx[0]
-usz0 - usz1 = bsx[1]
As(usx[0]) - usx[1] - usz0 = bsz0
-usx[1] - S * usz1 * S = bsz1.
1. Eliminate usz0, usz1 using equations 3 and 4,
usz0 = As(usx[0]) - usx[1] - bsz0
usz1 = -S^-1 * (usx[1] + bsz1) * S^-1.
This gives two equations in usx[0] an usx[1].
As'(As(usx[0]) - usx[1]) = bx[0] + As'(bsz0)
-As(usx[0]) + usx[1] + S^-1 * usx[1] * S^-1
= bsx[1] - bsz0 - S^-1 * bsz1 * S^-1.
2. Eliminate usx[1] using equation 2:
usx[1] + S * usx[1] * S
= S * ( As(usx[0]) + bsx[1] - bsz0 ) * S - bsz1
i.e., with Gamma[i,j] = 1.0 + S[i,i] * S[j,j],
usx[1] = ( S * As(usx[0]) * S ) ./ Gamma
+ ( S * ( bsx[1] - bsz0 ) * S - bsz1 ) ./ Gamma.
This gives an equation in usx[0].
As'( As(usx[0]) ./ Gamma )
= bx0 + As'(bsz0) +
As'( (S * ( bsx[1] - bsz0 ) * S - bsz1) ./ Gamma )
= bx0 + As'( ( bsz0 - bsz1 + S * bsx[1] * S ) ./ Gamma ).
"""
# Calculate U s.t.
#
# U' * r0*r0' * U = I, U' * r1*r1' * U = diag(s).
# Cholesky factorization r0 * r0' = L * L'
blas.syrk(W['r'][0], L)
lapack.potrf(L)
# SVD L^-1 * r1 = U * diag(s) * V'
blas.copy(W['r'][1], U)
blas.trsm(L, U)
lapack.gesvd(U, s, jobu='O')
# s := s**2
s[:] = s**2
# Uti := U
blas.copy(U, Uti)
# U := L^-T * U
blas.trsm(L, U, transA='T')
# Uti := L * Uti = U^-T
blas.trmm(L, Uti)
# Us := U * diag(s)^-1
blas.copy(U, Us)
for i in range(m):
blas.tbsv(s, Us, n=m, k=0, ldA=1, incx=m, offsetx=i)
# S is m x m with lower triangular entries s[i] * s[j]
# sqrtG is m x m with lower triangular entries sqrt(1.0 + s[i]*s[j])
# Upper triangular entries are undefined but nonzero.
blas.scal(0.0, S)
blas.syrk(s, S)
Gamma = 1.0 + S
sqrtG = sqrt(Gamma)
# Asc[i] = (U' * Ai * * U ) ./ sqrtG, for i = 1, ..., n
# = Asi ./ sqrt(Gamma)
blas.copy(A, Asc)
misc.scale(
Asc, # only 'r' part of the dictionary is used
{
'dnl': matrix(0.0, (0, 1)),
'dnli': matrix(0.0, (0, 1)),
'd': matrix(0.0, (0, 1)),
'di': matrix(0.0, (0, 1)),
'v': [],
'beta': [],
'r': [U],
'rti': [U]
})
for i in range(n):
blas.tbsv(sqrtG, Asc, n=msq, k=0, ldA=1, offsetx=i * msq)
# Convert columns of Asc to packed storage
misc.pack2(Asc, {'l': 0, 'q': [], 's': [m]})
# Cholesky factorization of Asc' * Asc.
H = matrix(0.0, (n, n))
blas.syrk(Asc, H, trans='T', k=mpckd)
lapack.potrf(H)
def solve(x, y, z):
"""
1. Solve for usx[0]:
Asc'(Asc(usx[0]))
= bx0 + Asc'( ( bsz0 - bsz1 + S * bsx[1] * S ) ./ sqrtG)
= bx0 + Asc'( ( bsz0 + S * ( bsx[1] - bssz1) S )
./ sqrtG)
where bsx[1] = U^-1 * bx[1] * U^-T, bsz0 = U' * bz0 * U,
bsz1 = U' * bz1 * U, bssz1 = S^-1 * bsz1 * S^-1
2. Solve for usx[1]:
usx[1] + S * usx[1] * S
= S * ( As(usx[0]) + bsx[1] - bsz0 ) * S - bsz1
usx[1]
= ( S * (As(usx[0]) + bsx[1] - bsz0) * S - bsz1) ./ Gamma
= -bsz0 + (S * As(usx[0]) * S) ./ Gamma
+ (bsz0 - bsz1 + S * bsx[1] * S ) . / Gamma
= -bsz0 + (S * As(usx[0]) * S) ./ Gamma
+ (bsz0 + S * ( bsx[1] - bssz1 ) * S ) . / Gamma
Unscale ux[1] = Uti * usx[1] * Uti'
3. Compute usz0, usz1
r0' * uz0 * r0 = r0^-1 * ( A(ux[0]) - ux[1] - bz0 ) * r0^-T
r1' * uz1 * r1 = r1^-1 * ( -ux[1] - bz1 ) * r1^-T
"""
# z0 := U' * z0 * U
# = bsz0
__cngrnc(U, z, trans='T')
# z1 := Us' * bz1 * Us
# = S^-1 * U' * bz1 * U * S^-1
# = S^-1 * bsz1 * S^-1
__cngrnc(Us, z, trans='T', offsetx=msq)
# x[1] := Uti' * x[1] * Uti
# = bsx[1]
__cngrnc(Uti, x[1], trans='T')
# x[1] := x[1] - z[msq:]
# = bsx[1] - S^-1 * bsz1 * S^-1
blas.axpy(z, x[1], alpha=-1.0, offsetx=msq)
# x1 = (S * x[1] * S + z[:msq] ) ./ sqrtG
# = (S * ( bsx[1] - S^-1 * bsz1 * S^-1) * S + bsz0 ) ./ sqrtG
# = (S * bsx[1] * S - bsz1 + bsz0 ) ./ sqrtG
# in packed storage
blas.copy(x[1], x1)
blas.tbmv(S, x1, n=msq, k=0, ldA=1)
blas.axpy(z, x1, n=msq)
blas.tbsv(sqrtG, x1, n=msq, k=0, ldA=1)
misc.pack2(x1, {'l': 0, 'q': [], 's': [m]})
# x[0] := x[0] + Asc'*x1
# = bx0 + Asc'( ( bsz0 - bsz1 + S * bsx[1] * S ) ./ sqrtG)
# = bx0 + As'( ( bz0 - bz1 + S * bx[1] * S ) ./ Gamma )
blas.gemv(Asc, x1, x[0], m=mpckd, trans='T', beta=1.0)
# x[0] := H^-1 * x[0]
# = ux[0]
lapack.potrs(H, x[0])
# x1 = Asc(x[0]) .* sqrtG (unpacked)
# = As(x[0])
blas.gemv(Asc, x[0], tmp, m=mpckd)
misc.unpack(tmp, x1, {'l': 0, 'q': [], 's': [m]})
blas.tbmv(sqrtG, x1, n=msq, k=0, ldA=1)
# usx[1] = (x1 + (x[1] - z[:msq])) ./ sqrtG**2
# = (As(ux[0]) + bsx[1] - bsz0 - S^-1 * bsz1 * S^-1)
# ./ Gamma
# x[1] := x[1] - z[:msq]
# = bsx[1] - bsz0 - S^-1 * bsz1 * S^-1
blas.axpy(z, x[1], -1.0, n=msq)
# x[1] := x[1] + x1
# = As(ux) + bsx[1] - bsz0 - S^-1 * bsz1 * S^-1
blas.axpy(x1, x[1])
# x[1] := x[1] / Gammma
# = (As(ux) + bsx[1] - bsz0 + S^-1 * bsz1 * S^-1 ) / Gamma
# = S^-1 * usx[1] * S^-1
blas.tbsv(Gamma, x[1], n=msq, k=0, ldA=1)
# z[msq:] := r1' * U * (-z[msq:] - x[1]) * U * r1
# := -r1' * U * S^-1 * (bsz1 + ux[1]) * S^-1 * U * r1
# := -r1' * uz1 * r1
blas.axpy(x[1], z, n=msq, offsety=msq)
blas.scal(-1.0, z, offset=msq)
__cngrnc(U, z, offsetx=msq)
__cngrnc(W['r'][1], z, trans='T', offsetx=msq)
# x[1] := S * x[1] * S
# = usx1
blas.tbmv(S, x[1], n=msq, k=0, ldA=1)
# z[:msq] = r0' * U' * ( x1 - x[1] - z[:msq] ) * U * r0
# = r0' * U' * ( As(ux) - usx1 - bsz0 ) * U * r0
# = r0' * U' * usz0 * U * r0
# = r0' * uz0 * r0
blas.axpy(x1, z, -1.0, n=msq)
blas.scal(-1.0, z, n=msq)
blas.axpy(x[1], z, -1.0, n=msq)
__cngrnc(U, z)
__cngrnc(W['r'][0], z, trans='T')
# x[1] := Uti * x[1] * Uti'
# = ux[1]
__cngrnc(Uti, x[1])
return solve
def xscal(alpha, u):
blas.scal(alpha, u[0])
blas.scal(alpha, u[1])
blas.scal(alpha, u[2])
def f(x, y, z):
"""
Solve
C * ux + G' * uzl - 2*A'(uzs21) = bx
-uzs11 = bX1
-uzs22 = bX2
G * ux - D^2 * uzl = bzl
[ -uX1 -A(ux)' ] [ uzs11 uzs21' ]
[ ] - T * [ ] * T = bzs.
[ -A(ux) -uX2 ] [ uzs21 uzs22 ]
On entry, x = (bx, bX1, bX2) and z = [ bzl; bzs[:] ].
On exit, x = (ux, uX1, uX2) and z = [ D*uzl; (r'*uzs*r)[:] ].
Define X = uzs21, Z = T * uzs * T:
C * ux + G' * uzl - 2*A'(X) = bx
[ 0 X' ] [ bX1 0 ]
T * [ ] * T - Z = T * [ ] * T
[ X 0 ] [ 0 bX2 ]
G * ux - D^2 * uzl = bzl
[ -uX1 -A(ux)' ] [ Z11 Z21' ]
[ ] - [ ] = bzs
[ -A(ux) -uX2 ] [ Z21 Z22 ]
Return x = (ux, uX1, uX2), z = [ D*uzl; (rti'*Z*rti)[:] ].
We use the congruence transformation
[ V1 0 ] [ T11 T21' ] [ V1' 0 ] [ I S' ]
[ ] [ ] [ ] = [ ]
[ 0 V2' ] [ T21 T22 ] [ 0 V2 ] [ S I ]
and the factorization
X + S * X' * S = L( L'(X) )
to write this as
C * ux + G' * uzl - 2*A'(X) = bx
L'(V2^-1 * X * V1^-1) - L^-1(V2' * Z21 * V1') = bX
G * ux - D^2 * uzl = bzl
[ -uX1 -A(ux)' ] [ Z11 Z21' ]
[ ] - [ ] = bzs,
[ -A(ux) -uX2 ] [ Z21 Z22 ]
or
C * ux + Gs' * uuzl - 2*As'(XX) = bx
XX - ZZ21 = bX
Gs * ux - uuzl = D^-1 * bzl
-As(ux) - ZZ21 = bbzs_21
-uX1 - Z11 = bzs_11
-uX2 - Z22 = bzs_22
if we introduce scaled variables
uuzl = D * uzl
XX = L'(V2^-1 * X * V1^-1)
= L'(V2^-1 * uzs21 * V1^-1)
ZZ21 = L^-1(V2' * Z21 * V1')
and define
bbzs_21 = L^-1(V2' * bzs_21 * V1')
[ bX1 0 ]
bX = L^-1( V2' * (T * [ ] * T)_21 * V1').
[ 0 bX2 ]
Eliminating Z21 gives
C * ux + Gs' * uuzl - 2*As'(XX) = bx
Gs * ux - uuzl = D^-1 * bzl
-As(ux) - XX = bbzs_21 - bX
-uX1 - Z11 = bzs_11
-uX2 - Z22 = bzs_22
and eliminating uuzl and XX gives
H * ux = bx + Gs' * D^-1 * bzl + 2*As'(bX - bbzs_21)
Gs * ux - uuzl = D^-1 * bzl
-As(ux) - XX = bbzs_21 - bX
-uX1 - Z11 = bzs_11
-uX2 - Z22 = bzs_22.
In summary, we can use the following algorithm:
1. bXX := bX - bbzs21
[ bX1 0 ]
= L^-1( V2' * ((T * [ ] * T)_21 - bzs_21) * V1')
[ 0 bX2 ]
2. Solve H * ux = bx + Gs' * D^-1 * bzl + 2*As'(bXX).
3. From ux, compute
uuzl = Gs*ux - D^-1 * bzl and
X = V2 * L^-T(-As(ux) + bXX) * V1.
4. Return ux, uuzl,
rti' * Z * rti = r' * [ -bX1, X'; X, -bX2 ] * r
and uX1 = -Z11 - bzs_11, uX2 = -Z22 - bzs_22.
"""
# Save bzs_11, bzs_22, bzs_21.
lapack.lacpy(z, bz11, uplo='L', m=q, n=q, ldA=p + q, offsetA=m)
lapack.lacpy(z, bz21, m=p, n=q, ldA=p + q, offsetA=m + q)
lapack.lacpy(z,
bz22,
uplo='L',
m=p,
n=p,
ldA=p + q,
offsetA=m + (p + q + 1) * q)
# zl := D^-1 * zl
# = D^-1 * bzl
blas.tbmv(W['di'], z, n=m, k=0, ldA=1)
# zs := r' * [ bX1, 0; 0, bX2 ] * r.
# zs := [ bX1, 0; 0, bX2 ]
blas.scal(0.0, z, offset=m)
lapack.lacpy(x[1], z, uplo='L', m=q, n=q, ldB=p + q, offsetB=m)
lapack.lacpy(x[2],
z,
uplo='L',
m=p,
n=p,
ldB=p + q,
offsetB=m + (p + q + 1) * q)
# scale diagonal of zs by 1/2
blas.scal(0.5, z, inc=p + q + 1, offset=m)
# a := tril(zs)*r
blas.copy(r, a)
blas.trmm(z,
a,
side='L',
m=p + q,
n=p + q,
ldA=p + q,
ldB=p + q,
offsetA=m)
# zs := a'*r + r'*a
blas.syr2k(r,
a,
z,
trans='T',
n=p + q,
k=p + q,
ldB=p + q,
ldC=p + q,
offsetC=m)
# bz21 := L^-1( V2' * ((r * zs * r')_21 - bz21) * V1')
#
# [ bX1 0 ]
# = L^-1( V2' * ((T * [ ] * T)_21 - bz21) * V1').
# [ 0 bX2 ]
# a = [ r21 r22 ] * z
# = [ r21 r22 ] * r' * [ bX1, 0; 0, bX2 ] * r
# = [ T21 T22 ] * [ bX1, 0; 0, bX2 ] * r
blas.symm(z,
r,
a,
side='R',
m=p,
n=p + q,
ldA=p + q,
ldC=p + q,
offsetB=q)
# bz21 := -bz21 + a * [ r11, r12 ]'
# = -bz21 + (T * [ bX1, 0; 0, bX2 ] * T)_21
blas.gemm(a,
r,
bz21,
transB='T',
m=p,
n=q,
k=p + q,
beta=-1.0,
ldA=p + q,
ldC=p)
# bz21 := V2' * bz21 * V1'
# = V2' * (-bz21 + (T*[bX1, 0; 0, bX2]*T)_21) * V1'
blas.gemm(V2, bz21, tmp, transA='T', m=p, n=q, k=p, ldB=p)
blas.gemm(tmp, V1, bz21, transB='T', m=p, n=q, k=q, ldC=p)
# bz21[:] := D * (I-P) * bz21[:]
# = L^-1 * bz21[:]
# = bXX[:]
blas.copy(bz21, tmp)
base.gemv(P, bz21, tmp, alpha=-1.0, beta=1.0)
base.gemv(D, tmp, bz21)
# Solve H * ux = bx + Gs' * D^-1 * bzl + 2*As'(bXX).
# x[0] := x[0] + Gs'*zl + 2*As'(bz21)
# = bx + G' * D^-1 * bzl + 2 * As'(bXX)
blas.gemv(Gs, z, x[0], trans='T', alpha=1.0, beta=1.0)
blas.gemv(As, bz21, x[0], trans='T', alpha=2.0, beta=1.0)
# x[0] := H \ x[0]
# = ux
lapack.potrs(H, x[0])
# uuzl = Gs*ux - D^-1 * bzl
blas.gemv(Gs, x[0], z, alpha=1.0, beta=-1.0)
# bz21 := V2 * L^-T(-As(ux) + bz21) * V1
# = X
blas.gemv(As, x[0], bz21, alpha=-1.0, beta=1.0)
blas.tbsv(DV, bz21, n=p * q, k=0, ldA=1)
blas.copy(bz21, tmp)
base.gemv(P, tmp, bz21, alpha=-1.0, beta=1.0, trans='T')
blas.gemm(V2, bz21, tmp)
blas.gemm(tmp, V1, bz21)
# zs := -zs + r' * [ 0, X'; X, 0 ] * r
# = r' * [ -bX1, X'; X, -bX2 ] * r.
# a := bz21 * [ r11, r12 ]
# = X * [ r11, r12 ]
blas.gemm(bz21, r, a, m=p, n=p + q, k=q, ldA=p, ldC=p + q)
# z := -z + [ r21, r22 ]' * a + a' * [ r21, r22 ]
# = rti' * uzs * rti
blas.syr2k(r,
a,
z,
trans='T',
beta=-1.0,
n=p + q,
k=p,
offsetA=q,
offsetC=m,
ldB=p + q,
ldC=p + q)
# uX1 = -Z11 - bzs_11
# = -(r*zs*r')_11 - bzs_11
# uX2 = -Z22 - bzs_22
# = -(r*zs*r')_22 - bzs_22
blas.copy(bz11, x[1])
blas.copy(bz22, x[2])
# scale diagonal of zs by 1/2
blas.scal(0.5, z, inc=p + q + 1, offset=m)
# a := r*tril(zs)
blas.copy(r, a)
blas.trmm(z,
a,
side='R',
m=p + q,
n=p + q,
ldA=p + q,
ldB=p + q,
offsetA=m)
# x[1] := -x[1] - a[:q,:] * r[:q, :]' - r[:q,:] * a[:q,:]'
# = -bzs_11 - (r*zs*r')_11
blas.syr2k(a, r, x[1], n=q, alpha=-1.0, beta=-1.0)
# x[2] := -x[2] - a[q:,:] * r[q:, :]' - r[q:,:] * a[q:,:]'
# = -bzs_22 - (r*zs*r')_22
blas.syr2k(a,
r,
x[2],
n=p,
alpha=-1.0,
beta=-1.0,
offsetA=q,
offsetB=q)
# scale diagonal of zs by 1/2
blas.scal(2.0, z, inc=p + q + 1, offset=m)
def F(W):
"""
Generate a solver for
A'(uz0) = bx[0]
-uz0 - uz1 = bx[1]
A(ux[0]) - ux[1] - r0*r0' * uz0 * r0*r0' = bz0
- ux[1] - r1*r1' * uz1 * r1*r1' = bz1.
uz0, uz1, bz0, bz1 are symmetric m x m-matrices.
ux[0], bx[0] are n-vectors.
ux[1], bx[1] are symmetric m x m-matrices.
We first calculate a congruence that diagonalizes r0*r0' and r1*r1':
U' * r0 * r0' * U = I, U' * r1 * r1' * U = S.
We then make a change of variables
usx[0] = ux[0],
usx[1] = U' * ux[1] * U
usz0 = U^-1 * uz0 * U^-T
usz1 = U^-1 * uz1 * U^-T
and define
As() = U' * A() * U'
bsx[1] = U^-1 * bx[1] * U^-T
bsz0 = U' * bz0 * U
bsz1 = U' * bz1 * U.
This gives
As'(usz0) = bx[0]
-usz0 - usz1 = bsx[1]
As(usx[0]) - usx[1] - usz0 = bsz0
-usx[1] - S * usz1 * S = bsz1.
1. Eliminate usz0, usz1 using equations 3 and 4,
usz0 = As(usx[0]) - usx[1] - bsz0
usz1 = -S^-1 * (usx[1] + bsz1) * S^-1.
This gives two equations in usx[0] an usx[1].
As'(As(usx[0]) - usx[1]) = bx[0] + As'(bsz0)
-As(usx[0]) + usx[1] + S^-1 * usx[1] * S^-1
= bsx[1] - bsz0 - S^-1 * bsz1 * S^-1.
2. Eliminate usx[1] using equation 2:
usx[1] + S * usx[1] * S
= S * ( As(usx[0]) + bsx[1] - bsz0 ) * S - bsz1
i.e., with Gamma[i,j] = 1.0 + S[i,i] * S[j,j],
usx[1] = ( S * As(usx[0]) * S ) ./ Gamma
+ ( S * ( bsx[1] - bsz0 ) * S - bsz1 ) ./ Gamma.
This gives an equation in usx[0].
As'( As(usx[0]) ./ Gamma )
= bx0 + As'(bsz0) +
As'( (S * ( bsx[1] - bsz0 ) * S - bsz1) ./ Gamma )
= bx0 + As'( ( bsz0 - bsz1 + S * bsx[1] * S ) ./ Gamma ).
"""
# Calculate U s.t.
#
# U' * r0*r0' * U = I, U' * r1*r1' * U = diag(s).
# Cholesky factorization r0 * r0' = L * L'
blas.syrk(W['r'][0], L)
lapack.potrf(L)
# SVD L^-1 * r1 = U * diag(s) * V'
blas.copy(W['r'][1], U)
blas.trsm(L, U)
lapack.gesvd(U, s, jobu = 'O')
# s := s**2
s[:] = s**2
# Uti := U
blas.copy(U, Uti)
# U := L^-T * U
blas.trsm(L, U, transA = 'T')
# Uti := L * Uti = U^-T
blas.trmm(L, Uti)
# Us := U * diag(s)^-1
blas.copy(U, Us)
for i in range(m):
blas.tbsv(s, Us, n = m, k = 0, ldA = 1, incx = m, offsetx = i)
# S is m x m with lower triangular entries s[i] * s[j]
# sqrtG is m x m with lower triangular entries sqrt(1.0 + s[i]*s[j])
# Upper triangular entries are undefined but nonzero.
blas.scal(0.0, S)
blas.syrk(s, S)
Gamma = 1.0 + S
sqrtG = sqrt(Gamma)
# Asc[i] = (U' * Ai * * U ) ./ sqrtG, for i = 1, ..., n
# = Asi ./ sqrt(Gamma)
blas.copy(A, Asc)
misc.scale(Asc, # only 'r' part of the dictionary is used
{'dnl': matrix(0.0, (0, 1)), 'dnli': matrix(0.0, (0, 1)),
'd': matrix(0.0, (0, 1)), 'di': matrix(0.0, (0, 1)),
'v': [], 'beta': [], 'r': [ U ], 'rti': [ U ]})
for i in range(n):
blas.tbsv(sqrtG, Asc, n = msq, k = 0, ldA = 1, offsetx = i*msq)
# Convert columns of Asc to packed storage
misc.pack2(Asc, {'l': 0, 'q': [], 's': [ m ]})
# Cholesky factorization of Asc' * Asc.
H = matrix(0.0, (n, n))
blas.syrk(Asc, H, trans = 'T', k = mpckd)
lapack.potrf(H)
def solve(x, y, z):
"""
1. Solve for usx[0]:
Asc'(Asc(usx[0]))
= bx0 + Asc'( ( bsz0 - bsz1 + S * bsx[1] * S ) ./ sqrtG)
= bx0 + Asc'( ( bsz0 + S * ( bsx[1] - bssz1) S )
./ sqrtG)
where bsx[1] = U^-1 * bx[1] * U^-T, bsz0 = U' * bz0 * U,
bsz1 = U' * bz1 * U, bssz1 = S^-1 * bsz1 * S^-1
2. Solve for usx[1]:
usx[1] + S * usx[1] * S
= S * ( As(usx[0]) + bsx[1] - bsz0 ) * S - bsz1
usx[1]
= ( S * (As(usx[0]) + bsx[1] - bsz0) * S - bsz1) ./ Gamma
= -bsz0 + (S * As(usx[0]) * S) ./ Gamma
+ (bsz0 - bsz1 + S * bsx[1] * S ) . / Gamma
= -bsz0 + (S * As(usx[0]) * S) ./ Gamma
+ (bsz0 + S * ( bsx[1] - bssz1 ) * S ) . / Gamma
Unscale ux[1] = Uti * usx[1] * Uti'
3. Compute usz0, usz1
r0' * uz0 * r0 = r0^-1 * ( A(ux[0]) - ux[1] - bz0 ) * r0^-T
r1' * uz1 * r1 = r1^-1 * ( -ux[1] - bz1 ) * r1^-T
"""
# z0 := U' * z0 * U
# = bsz0
__cngrnc(U, z, trans = 'T')
# z1 := Us' * bz1 * Us
# = S^-1 * U' * bz1 * U * S^-1
# = S^-1 * bsz1 * S^-1
__cngrnc(Us, z, trans = 'T', offsetx = msq)
# x[1] := Uti' * x[1] * Uti
# = bsx[1]
__cngrnc(Uti, x[1], trans = 'T')
# x[1] := x[1] - z[msq:]
# = bsx[1] - S^-1 * bsz1 * S^-1
blas.axpy(z, x[1], alpha = -1.0, offsetx = msq)
# x1 = (S * x[1] * S + z[:msq] ) ./ sqrtG
# = (S * ( bsx[1] - S^-1 * bsz1 * S^-1) * S + bsz0 ) ./ sqrtG
# = (S * bsx[1] * S - bsz1 + bsz0 ) ./ sqrtG
# in packed storage
blas.copy(x[1], x1)
blas.tbmv(S, x1, n = msq, k = 0, ldA = 1)
blas.axpy(z, x1, n = msq)
blas.tbsv(sqrtG, x1, n = msq, k = 0, ldA = 1)
misc.pack2(x1, {'l': 0, 'q': [], 's': [m]})
# x[0] := x[0] + Asc'*x1
# = bx0 + Asc'( ( bsz0 - bsz1 + S * bsx[1] * S ) ./ sqrtG)
# = bx0 + As'( ( bz0 - bz1 + S * bx[1] * S ) ./ Gamma )
blas.gemv(Asc, x1, x[0], m = mpckd, trans = 'T', beta = 1.0)
# x[0] := H^-1 * x[0]
# = ux[0]
lapack.potrs(H, x[0])
# x1 = Asc(x[0]) .* sqrtG (unpacked)
# = As(x[0])
blas.gemv(Asc, x[0], tmp, m = mpckd)
misc.unpack(tmp, x1, {'l': 0, 'q': [], 's': [m]})
blas.tbmv(sqrtG, x1, n = msq, k = 0, ldA = 1)
# usx[1] = (x1 + (x[1] - z[:msq])) ./ sqrtG**2
# = (As(ux[0]) + bsx[1] - bsz0 - S^-1 * bsz1 * S^-1)
# ./ Gamma
# x[1] := x[1] - z[:msq]
# = bsx[1] - bsz0 - S^-1 * bsz1 * S^-1
blas.axpy(z, x[1], -1.0, n = msq)
# x[1] := x[1] + x1
# = As(ux) + bsx[1] - bsz0 - S^-1 * bsz1 * S^-1
blas.axpy(x1, x[1])
# x[1] := x[1] / Gammma
# = (As(ux) + bsx[1] - bsz0 + S^-1 * bsz1 * S^-1 ) / Gamma
# = S^-1 * usx[1] * S^-1
blas.tbsv(Gamma, x[1], n = msq, k = 0, ldA = 1)
# z[msq:] := r1' * U * (-z[msq:] - x[1]) * U * r1
# := -r1' * U * S^-1 * (bsz1 + ux[1]) * S^-1 * U * r1
# := -r1' * uz1 * r1
blas.axpy(x[1], z, n = msq, offsety = msq)
blas.scal(-1.0, z, offset = msq)
__cngrnc(U, z, offsetx = msq)
__cngrnc(W['r'][1], z, trans = 'T', offsetx = msq)
# x[1] := S * x[1] * S
# = usx1
blas.tbmv(S, x[1], n = msq, k = 0, ldA = 1)
# z[:msq] = r0' * U' * ( x1 - x[1] - z[:msq] ) * U * r0
# = r0' * U' * ( As(ux) - usx1 - bsz0 ) * U * r0
# = r0' * U' * usz0 * U * r0
# = r0' * uz0 * r0
blas.axpy(x1, z, -1.0, n = msq)
blas.scal(-1.0, z, n = msq)
blas.axpy(x[1], z, -1.0, n = msq)
__cngrnc(U, z)
__cngrnc(W['r'][0], z, trans = 'T')
# x[1] := Uti * x[1] * Uti'
# = ux[1]
__cngrnc(Uti, x[1])
return solve
def Pf(u, v, alpha=1.0, beta=0.0):
base.symv(C, u[0], v[0], alpha=alpha, beta=beta)
blas.scal(beta, v[1])
blas.scal(beta, v[2])
def sysid(y, u, vsig, svth = None):
"""
System identification using the subspace method and nuclear norm
optimization. Estimate a linear time-invariant state-space model
given inputs and outputs. The algorithm is described in [1].
INPUT
y 'd' matrix of size (p, N). y are the measured outputs, p is
the number of outputs, and N is the number of data points
measured.
u 'd' matrix of size (m, N). u are the inputs, m is the number
of inputs, and N is the number of data points.
vsig a weighting parameter in the nuclear norm optimization, its
value is approximately the 1-sigma output noise level
svth an optional parameter, if specified, the model order is
determined as the number of singular values greater than svth
times the maximum singular value. The default value is 1E-3
OUTPUT
sol a dictionary with the following words
-- 'A', 'B', 'C', 'D' are the state-space matrices
-- 'svN', the original singular values of the Hankel matrix
-- 'sv', the optimized singular values of the Hankel matrix
-- 'x0', the initial state x(0)
-- 'n', the model order
[1] Zhang Liu and Lieven Vandenberghe. "Interior-point method for
nuclear norm approximation with application to system
identification."
"""
m, N, p = u.size[0], u.size[1], y.size[0]
if y.size[1] != N:
raise ValueError, "y and u must have the same length"
# Y = G*X + H*U + V, Y has size a x b, U has size c x b, Un has b x d
r = min(int(30/p),int((N+1.0)/(p+m+1)+1.0))
a = r*p
c = r*m
b = N-r+1
d = b-c
# construct Hankel matrix Y
Y = Hankel(y,r,b,p=p,q=1)
# construct Hankel matrix U
U = Hankel(u,r,b,p=m,q=1)
# compute Un = null(U) and YUn = Y*Un
Vt = matrix(0.0,(b,b))
Stemp = matrix(0.0,(c,1))
Un = matrix(0.0,(b,d))
YUn = matrix(0.0,(a,d))
lapack.gesvd(U,Stemp,jobvt='A',Vt=Vt)
Un[:,:] = Vt.T[:,c:]
blas.gemm(Y,Un,YUn)
# compute original singular values
svN = matrix(0.0,(min(a,d),1))
lapack.gesvd(YUn,svN)
# variable, [y(1);...;y(N)]
# form the coefficient matrices for the nuclear norm optimization
# minimize | Yh * Un |_* + alpha * | y - yh |_F
AA = Hankel_basis(r,b,p=p,q=1)
A = matrix(0.0,(a*d,p*N))
temp = spmatrix([],[],[],(a,b),'d')
temp2 = matrix(0.0,(a,d))
for ii in xrange(p*N):
temp[:] = AA[:,ii]
base.gemm(temp,Un,temp2)
A[:,ii] = temp2[:]
B = matrix(0.0,(a,d))
# flip the matrix if columns is more than rows
if a < d:
Itrans = [i+j*a for i in xrange(a) for j in xrange(d)]
B[:] = B[Itrans]
B.size = (d,a)
for ii in xrange(p*N):
A[:,ii] = A[Itrans,ii]
# regularized term
x0 = y[:]
Qd = matrix(2.0*svN[0]/p/N/(vsig**2),(p*N,1))
# solve the nuclear norm optimization
sol = nrmapp(A, B, C = base.spdiag(Qd), d = -base.mul(x0, Qd))
status = sol['status']
x = sol['x']
# construct YhUn and take the svd
YhUn = matrix(B)
blas.gemv(A,x,YhUn,beta=1.0)
if a < d:
YhUn = YhUn.T
Uh = matrix(0.0,(a,d))
sv = matrix(0.0,(d,1))
lapack.gesvd(YhUn,sv,jobu='S',U=Uh)
# determine model order
if svth is None:
svth = 1E-3
svthn = sv[0]*svth
n=1
while sv[n] >= svthn and n < 10:
n=n+1
# estimate A, C
Uhn = Uh[:,:n]
for ii in xrange(n):
blas.scal(sv[ii],Uhn,n=a,offset=ii*a)
syseC = Uhn[:p,:]
Als = Uhn[:-p,:]
Bls = Uhn[p:,:]
lapack.gels(Als,Bls)
syseA = Bls[:n,:]
Als[:,:] = Uhn[:-p,:]
Bls[:,:] = Uhn[p:,:]
blas.gemm(Als,syseA,Bls,beta=-1.0)
Aerr = blas.nrm2(Bls)
# stabilize A
Sc = matrix(0.0,(n,n),'z')
w = matrix(0.0, (n,1), 'z')
Vs = matrix(0.0, (n,n), 'z')
def F(w):
return (abs(w) < 1.0)
Sc[:,:] = syseA
ns = lapack.gees(Sc, w, Vs, select = F)
while ns < n:
#print "stabilize matrix A"
w[ns:] = w[ns:]**-1
Sc[::n+1] = w
Sc = Vs*Sc*Vs.H
syseA[:,:] = Sc.real()
Sc[:,:] = syseA
ns = lapack.gees(Sc, w, Vs, select = F)
# estimate B,D,x0 stored in vector [x0; vec(D); vec(B)]
F1 = matrix(0.0,(p*N,n))
F1[:p,:] = syseC
for ii in xrange(1,N):
F1[ii*p:(ii+1)*p,:] = F1[(ii-1)*p:ii*p,:]*syseA
F2 = matrix(0.0,(p*N,p*m))
ut = u.T
for ii in xrange(p):
F2[ii::p,ii::p] = ut
F3 = matrix(0.0,(p*N,n*m))
F3t = matrix(0.0,(p*(N-1),n*m))
for ii in xrange(1,N):
for jj in xrange(p):
for kk in xrange(n):
F3t[jj:jj+(N-ii)*p:p,kk::n] = ut[:N-ii,:]*F1[(ii-1)*p+jj,kk]
F3[ii*p:,:] = F3[ii*p:,:] + F3t[:(N-ii)*p,:]
F = matrix([[F1],[F2],[F3]])
yls = y[:]
Sls = matrix(0.0,(F.size[1],1))
Uls = matrix(0.0,(F.size[0],F.size[1]))
Vtls = matrix(0.0,(F.size[1],F.size[1]))
lapack.gesvd(F, Sls, jobu='S', jobvt='S', U=Uls, Vt=Vtls)
Frank=len([ii for ii in xrange(Sls.size[0]) if Sls[ii] >= 1E-6])
#print 'Rank deficiency = ', F.size[1] - Frank
xx = matrix(0.0,(F.size[1],1))
xx[:Frank] = Uls.T[:Frank,:] * yls
xx[:Frank] = base.mul(xx[:Frank],Sls[:Frank]**-1)
xx[:] = Vtls.T[:,:Frank]*xx[:Frank]
blas.gemv(F,xx,yls,beta=-1.0)
xxerr = blas.nrm2(yls)
x0 = xx[:n]
syseD = xx[n:n+p*m]
syseD.size = (p,m)
syseB = xx[n+p*m:]
syseB.size = (n,m)
return {'A': syseA, 'B': syseB, 'C': syseC, 'D': syseD, 'svN': svN, 'sv': \
sv, 'x0': x0, 'n': n, 'Aerr': Aerr, 'xxerr': xxerr}
def Pf(u, v, alpha = 1.0, beta = 0.0):
base.symv(C, u[0], v[0], alpha = alpha, beta = beta)
blas.scal(beta, v[1])
blas.scal(beta, v[2])
def completion(X, factored_updates = True):
"""
Supernodal multifrontal maximum determinant positive definite
matrix completion. The routine computes the Cholesky factor
:math:`L` of the inverse of the maximum determinant positive
definite matrix completion of :math:`X`:, i.e.,
.. math::
P( S^{-1} ) = X
where :math:`S = LL^T`. On exit, the argument `X` contains the
lower-triangular Cholesky factor :math:`L`.
The optional argument `factored_updates` can be used to enable (if
True) or disable (if False) updating of intermediate
factorizations.
:param X: :py:class:`cspmatrix`
:param factored_updates: boolean
"""
assert isinstance(X, cspmatrix) and X.is_factor is False, "X must be a cspmatrix"
n = X.symb.n
snpost = X.symb.snpost
snptr = X.symb.snptr
chptr = X.symb.chptr
chidx = X.symb.chidx
relptr = X.symb.relptr
relidx = X.symb.relidx
blkptr = X.symb.blkptr
blkval = X.blkval
stack = []
for k in reversed(list(snpost)):
nn = snptr[k+1]-snptr[k] # |Nk|
na = relptr[k+1]-relptr[k] # |Ak|
nj = na + nn
# allocate F and copy X_{Jk,Nk} to leading columns of F
F = matrix(0.0, (nj,nj))
lapack.lacpy(blkval, F, offsetA = blkptr[k], ldA = nj, m = nj, n = nn, uplo = 'L')
# if supernode k is not a root node:
if na > 0:
# copy Vk to 2,2 block of F
Vk = stack.pop()
lapack.lacpy(Vk, F, offsetB = nn*nj+nn, m = na, n = na, uplo = 'L')
# if supernode k has any children:
for ii in range(chptr[k],chptr[k+1]):
i = chidx[ii]
if factored_updates:
r = relidx[relptr[i]:relptr[i+1]]
stack.append(frontal_get_update_factor(F,r,nn,na))
else:
stack.append(frontal_get_update(F,relidx,relptr,i))
# if supernode k is not a root node:
if na > 0:
if factored_updates:
# In this case we have Vk = Lk'*Lk
trL1 = 'T'
trL2 = 'N'
else:
# factorize Vk
lapack.potrf(F, offsetA = nj*nn+nn, n = na, ldA = nj)
# In this case we have Vk = Lk*Lk'
trL1 = 'N'
trL2 = 'T'
# compute L_{Ak,Nk} and inv(D_{Nk,Nk}) = S_{Nk,Nk} - S_{Ak,Nk}'*L_{Ak,Nk}
lapack.trtrs(F, blkval, offsetA = nj*nn+nn, trans = trL1,\
offsetB = blkptr[k]+nn, ldB = nj, n = na, nrhs = nn)
blas.syrk(blkval, blkval, n = nn, k = na, trans= 'T', alpha = -1.0, beta = 1.0,
offsetA = blkptr[k]+nn, offsetC = blkptr[k], ldA = nj, ldC = nj)
lapack.trtrs(F, blkval, offsetA = nj*nn+nn, trans = trL2,\
offsetB = blkptr[k]+nn, ldB = nj, n = na, nrhs = nn)
for i in range(nn):
blas.scal(-1.0, blkval, n = na, offset = blkptr[k] + i*nj + nn)
# factorize inv(D_{Nk,Nk}) as R*R' so that D_{Nk,Nk} = L*L' with L = inv(R)'
lapack.lacpy(blkval, F, offsetA = blkptr[k], ldA = nj,\
ldB = nj, m = nn, n = nn, uplo = 'L') # copy -- FIX!
F[:nn,:nn] = matrix(F[:nn,:nn][::-1],(nn,nn)) # reverse -- FIX!
lapack.potrf(F, ldA = nj, n = nn, uplo = 'U') # factorize
F[:nn,:nn] = matrix(F[:nn,:nn][::-1],(nn,nn)) # reverse -- FIX!
lapack.lacpy(F, blkval, offsetB = blkptr[k], ldA = nj,\
ldB = nj, m = nn, n = nn, uplo = 'L') # copy -- FIX!
# compute L = inv(R')
lapack.trtri(blkval, offsetA = blkptr[k], ldA = nj, n = nn)
X._is_factor = True
return
def spmatrix(self, reordered = True, symmetric = False):
"""
Converts the :py:class:`cspmatrix` :math:`A` to a sparse matrix. A reordered
matrix is returned if the optional argument `reordered` is
`True` (default), and otherwise the inverse permutation is applied. Only the
default options are allowed if the :py:class:`cspmatrix` :math:`A` represents
a Cholesky factor.
:param reordered: boolean (default: True)
:param symmetric: boolean (default: False)
"""
n = self.symb.n
snptr = self.symb.snptr
snode = self.symb.snode
relptr = self.symb.relptr
snrowidx = self.symb.snrowidx
sncolptr = self.symb.sncolptr
blkptr = self.symb.blkptr
blkval = self.blkval
if self.is_factor:
if symmetric: raise ValueError("'symmetric = True' not implemented for Cholesky factors")
if not reordered: raise ValueError("'reordered = False' not implemented for Cholesky factors")
snpost = self.symb.snpost
blkval = +blkval
for k in snpost:
j = snode[snptr[k]] # representative vertex
nn = snptr[k+1]-snptr[k] # |Nk|
na = relptr[k+1]-relptr[k] # |Ak|
if na == 0: continue
nj = na + nn
if nn == 1:
blas.scal(blkval[blkptr[k]],blkval,offset = blkptr[k]+1,n=na)
else:
blas.trmm(blkval,blkval, transA = "N", diag = "N", side = "R",uplo = "L", \
m = na, n = nn, ldA = nj, ldB = nj, \
offsetA = blkptr[k],offsetB = blkptr[k] + nn)
cc = matrix(0,(n,1)) # count number of nonzeros in each col
for k in range(self.symb.Nsn):
nn = snptr[k+1]-snptr[k]
na = relptr[k+1]-relptr[k]
nj = nn + na
for i in range(nn):
j = snode[snptr[k]+i]
cc[j] = nj - i
# build col. ptr
cp = [0]
for i in range(n): cp.append(cp[-1] + cc[i])
cp = matrix(cp)
# copy data and row indices
val = matrix(0.0, (cp[-1],1))
ri = matrix(0, (cp[-1],1))
for k in range(self.symb.Nsn):
nn = snptr[k+1]-snptr[k]
na = relptr[k+1]-relptr[k]
nj = nn + na
for i in range(nn):
j = snode[snptr[k]+i]
blas.copy(blkval, val, offsetx = blkptr[k]+nj*i+i, offsety = cp[j], n = nj-i)
ri[cp[j]:cp[j+1]] = snrowidx[sncolptr[k]+i:sncolptr[k+1]]
I = []; J = []
for i in range(n):
I += list(ri[cp[i]:cp[i+1]])
J += (cp[i+1]-cp[i])*[i]
tmp = spmatrix(val, I, J, (n,n)) # tmp is reordered and lower tril.
if reordered or self.symb.p is None:
# reordered matrix (do not apply inverse permutation)
if not symmetric: return tmp
else: return symmetrize(tmp)
else:
# apply inverse permutation
tmp = perm(symmetrize(tmp), self.symb.ip)
if symmetric: return tmp
else: return tril(tmp)
def solve(x, y, z):
"""
1. Solve for usx[0]:
Asc'(Asc(usx[0]))
= bx0 + Asc'( ( bsz0 - bsz1 + S * bsx[1] * S ) ./ sqrtG)
= bx0 + Asc'( ( bsz0 + S * ( bsx[1] - bssz1) S )
./ sqrtG)
where bsx[1] = U^-1 * bx[1] * U^-T, bsz0 = U' * bz0 * U,
bsz1 = U' * bz1 * U, bssz1 = S^-1 * bsz1 * S^-1
2. Solve for usx[1]:
usx[1] + S * usx[1] * S
= S * ( As(usx[0]) + bsx[1] - bsz0 ) * S - bsz1
usx[1]
= ( S * (As(usx[0]) + bsx[1] - bsz0) * S - bsz1) ./ Gamma
= -bsz0 + (S * As(usx[0]) * S) ./ Gamma
+ (bsz0 - bsz1 + S * bsx[1] * S ) . / Gamma
= -bsz0 + (S * As(usx[0]) * S) ./ Gamma
+ (bsz0 + S * ( bsx[1] - bssz1 ) * S ) . / Gamma
Unscale ux[1] = Uti * usx[1] * Uti'
3. Compute usz0, usz1
r0' * uz0 * r0 = r0^-1 * ( A(ux[0]) - ux[1] - bz0 ) * r0^-T
r1' * uz1 * r1 = r1^-1 * ( -ux[1] - bz1 ) * r1^-T
"""
# z0 := U' * z0 * U
# = bsz0
__cngrnc(U, z, trans='T')
# z1 := Us' * bz1 * Us
# = S^-1 * U' * bz1 * U * S^-1
# = S^-1 * bsz1 * S^-1
__cngrnc(Us, z, trans='T', offsetx=msq)
# x[1] := Uti' * x[1] * Uti
# = bsx[1]
__cngrnc(Uti, x[1], trans='T')
# x[1] := x[1] - z[msq:]
# = bsx[1] - S^-1 * bsz1 * S^-1
blas.axpy(z, x[1], alpha=-1.0, offsetx=msq)
# x1 = (S * x[1] * S + z[:msq] ) ./ sqrtG
# = (S * ( bsx[1] - S^-1 * bsz1 * S^-1) * S + bsz0 ) ./ sqrtG
# = (S * bsx[1] * S - bsz1 + bsz0 ) ./ sqrtG
# in packed storage
blas.copy(x[1], x1)
blas.tbmv(S, x1, n=msq, k=0, ldA=1)
blas.axpy(z, x1, n=msq)
blas.tbsv(sqrtG, x1, n=msq, k=0, ldA=1)
misc.pack2(x1, {'l': 0, 'q': [], 's': [m]})
# x[0] := x[0] + Asc'*x1
# = bx0 + Asc'( ( bsz0 - bsz1 + S * bsx[1] * S ) ./ sqrtG)
# = bx0 + As'( ( bz0 - bz1 + S * bx[1] * S ) ./ Gamma )
blas.gemv(Asc, x1, x[0], m=mpckd, trans='T', beta=1.0)
# x[0] := H^-1 * x[0]
# = ux[0]
lapack.potrs(H, x[0])
# x1 = Asc(x[0]) .* sqrtG (unpacked)
# = As(x[0])
blas.gemv(Asc, x[0], tmp, m=mpckd)
misc.unpack(tmp, x1, {'l': 0, 'q': [], 's': [m]})
blas.tbmv(sqrtG, x1, n=msq, k=0, ldA=1)
# usx[1] = (x1 + (x[1] - z[:msq])) ./ sqrtG**2
# = (As(ux[0]) + bsx[1] - bsz0 - S^-1 * bsz1 * S^-1)
# ./ Gamma
# x[1] := x[1] - z[:msq]
# = bsx[1] - bsz0 - S^-1 * bsz1 * S^-1
blas.axpy(z, x[1], -1.0, n=msq)
# x[1] := x[1] + x1
# = As(ux) + bsx[1] - bsz0 - S^-1 * bsz1 * S^-1
blas.axpy(x1, x[1])
# x[1] := x[1] / Gammma
# = (As(ux) + bsx[1] - bsz0 + S^-1 * bsz1 * S^-1 ) / Gamma
# = S^-1 * usx[1] * S^-1
blas.tbsv(Gamma, x[1], n=msq, k=0, ldA=1)
# z[msq:] := r1' * U * (-z[msq:] - x[1]) * U * r1
# := -r1' * U * S^-1 * (bsz1 + ux[1]) * S^-1 * U * r1
# := -r1' * uz1 * r1
blas.axpy(x[1], z, n=msq, offsety=msq)
blas.scal(-1.0, z, offset=msq)
__cngrnc(U, z, offsetx=msq)
__cngrnc(W['r'][1], z, trans='T', offsetx=msq)
# x[1] := S * x[1] * S
# = usx1
blas.tbmv(S, x[1], n=msq, k=0, ldA=1)
# z[:msq] = r0' * U' * ( x1 - x[1] - z[:msq] ) * U * r0
# = r0' * U' * ( As(ux) - usx1 - bsz0 ) * U * r0
# = r0' * U' * usz0 * U * r0
# = r0' * uz0 * r0
blas.axpy(x1, z, -1.0, n=msq)
blas.scal(-1.0, z, n=msq)
blas.axpy(x[1], z, -1.0, n=msq)
__cngrnc(U, z)
__cngrnc(W['r'][0], z, trans='T')
# x[1] := Uti * x[1] * Uti'
# = ux[1]
__cngrnc(Uti, x[1])
def loop(P, q, G, h, A, b, round):
# coneqp
# P = np.outer(y, y) * K (373x373)
# q = vector of -1 (373x1)
# G = np.vstack( np.diag(np.ones(n_samples) * -1), np.diag(np.ones(n_samples)) )
# (746x373)
# h = np.vstack( np.zeros(n_samples), np.ones(n_samples) * c )
# (746x1)
# A = y
# b = 0.0
def fP(P, x, y, alpha=1.0, beta=0.0):
# base.symv
return alpha * np.dot(P, x) + beta * y
def fA(x, y, trans="N", alpha=1.0, beta=0.0):
# base.gemv
# general matrix-vector product
return cvxopt.gemv(x, y, trans, alpha, beta)
dims = None
kktsolver = "chol2"
if kktsolver == "chol":
factor = getattr(misc, "kkt_chol")(G, dims, A)
else:
factor = getattr(misc, "kkt_chol2")(G, dims, A)
def kktsolver(W):
return factor(W, P)
f3 = kktsolver({"d": matrix(0.0, (0, 1)), "di": matrix(0.0, (0, 1)), "beta": [], "v": [], "r": [], "rti": []})
resx0 = max(1.0, math.sqrt(blas.dot(q, q))) # 19.13
resy0 = max(1.0, math.sqrt(blas.dot(b, b))) # 1.0
resz0 = max(1.0, misc.snrm2(h, dims)) # 193.13
x = matrix(q)
blas.scal(-1.0, x)
y = matrix(b)
f3(x, y, matrix(0.0, (0, 1)))
# dres = || P*x + q + A'*y || / resx0
rx = matrix(q)
fP(x, rx, beta=1.0)
pcost = 0.5 * (blas.dot(x, rx) + blas.dot(x, q))
fA(y, rx, beta=1.0, trans="T")
dres = math.sqrt(blas.dot(rx, rx)) / resx0
# pres = || A*x - b || / resy0
ry = matrix(b)
fA(x, ry, alpha=1.0, beta=-1.0)
pres = math.sqrt(blas.dot(ry, ry)) / resy0
if pcost == 0.0:
relgap = None
else:
relgap = 0.0
cdim = G.size[0]
x, y = matrix(q), matrix(b)
s, z = matrix(0.0, (cdim, 1)), matrix(0.0, (cdim, 1))
pcost = -6.3339e03
dcost = -5.5410e05
gap = 2e06
pres = 2e00
dres = 2e-14
print_round(round, pcost, dcost, gap, pres, dres)
def xscal(alpha, u):
blas.scal(alpha, u[0])
blas.scal(alpha, u[1])
def sysid(y, u, vsig, svth=None):
"""
System identification using the subspace method and nuclear norm
optimization. Estimate a linear time-invariant state-space model
given inputs and outputs. The algorithm is described in [1].
INPUT
y 'd' matrix of size (p, N). y are the measured outputs, p is
the number of outputs, and N is the number of data points
measured.
u 'd' matrix of size (m, N). u are the inputs, m is the number
of inputs, and N is the number of data points.
vsig a weighting parameter in the nuclear norm optimization, its
value is approximately the 1-sigma output noise level
svth an optional parameter, if specified, the model order is
determined as the number of singular values greater than svth
times the maximum singular value. The default value is 1E-3
OUTPUT
sol a dictionary with the following words
-- 'A', 'B', 'C', 'D' are the state-space matrices
-- 'svN', the original singular values of the Hankel matrix
-- 'sv', the optimized singular values of the Hankel matrix
-- 'x0', the initial state x(0)
-- 'n', the model order
[1] Zhang Liu and Lieven Vandenberghe. "Interior-point method for
nuclear norm approximation with application to system
identification."
"""
m, N, p = u.size[0], u.size[1], y.size[0]
if y.size[1] != N:
raise ValueError, "y and u must have the same length"
# Y = G*X + H*U + V, Y has size a x b, U has size c x b, Un has b x d
r = min(int(30 / p), int((N + 1.0) / (p + m + 1) + 1.0))
a = r * p
c = r * m
b = N - r + 1
d = b - c
# construct Hankel matrix Y
Y = Hankel(y, r, b, p=p, q=1)
# construct Hankel matrix U
U = Hankel(u, r, b, p=m, q=1)
# compute Un = null(U) and YUn = Y*Un
Vt = matrix(0.0, (b, b))
Stemp = matrix(0.0, (c, 1))
Un = matrix(0.0, (b, d))
YUn = matrix(0.0, (a, d))
lapack.gesvd(U, Stemp, jobvt='A', Vt=Vt)
Un[:, :] = Vt.T[:, c:]
blas.gemm(Y, Un, YUn)
# compute original singular values
svN = matrix(0.0, (min(a, d), 1))
lapack.gesvd(YUn, svN)
# variable, [y(1);...;y(N)]
# form the coefficient matrices for the nuclear norm optimization
# minimize | Yh * Un |_* + alpha * | y - yh |_F
AA = Hankel_basis(r, b, p=p, q=1)
A = matrix(0.0, (a * d, p * N))
temp = spmatrix([], [], [], (a, b), 'd')
temp2 = matrix(0.0, (a, d))
for ii in xrange(p * N):
temp[:] = AA[:, ii]
base.gemm(temp, Un, temp2)
A[:, ii] = temp2[:]
B = matrix(0.0, (a, d))
# flip the matrix if columns is more than rows
if a < d:
Itrans = [i + j * a for i in xrange(a) for j in xrange(d)]
B[:] = B[Itrans]
B.size = (d, a)
for ii in xrange(p * N):
A[:, ii] = A[Itrans, ii]
# regularized term
x0 = y[:]
Qd = matrix(2.0 * svN[0] / p / N / (vsig**2), (p * N, 1))
# solve the nuclear norm optimization
sol = nrmapp(A, B, C=base.spdiag(Qd), d=-base.mul(x0, Qd))
status = sol['status']
x = sol['x']
# construct YhUn and take the svd
YhUn = matrix(B)
blas.gemv(A, x, YhUn, beta=1.0)
if a < d:
YhUn = YhUn.T
Uh = matrix(0.0, (a, d))
sv = matrix(0.0, (d, 1))
lapack.gesvd(YhUn, sv, jobu='S', U=Uh)
# determine model order
if svth is None:
svth = 1E-3
svthn = sv[0] * svth
n = 1
while sv[n] >= svthn and n < 10:
n = n + 1
# estimate A, C
Uhn = Uh[:, :n]
for ii in xrange(n):
blas.scal(sv[ii], Uhn, n=a, offset=ii * a)
syseC = Uhn[:p, :]
Als = Uhn[:-p, :]
Bls = Uhn[p:, :]
lapack.gels(Als, Bls)
syseA = Bls[:n, :]
Als[:, :] = Uhn[:-p, :]
Bls[:, :] = Uhn[p:, :]
blas.gemm(Als, syseA, Bls, beta=-1.0)
Aerr = blas.nrm2(Bls)
# stabilize A
Sc = matrix(0.0, (n, n), 'z')
w = matrix(0.0, (n, 1), 'z')
Vs = matrix(0.0, (n, n), 'z')
def F(w):
return (abs(w) < 1.0)
Sc[:, :] = syseA
ns = lapack.gees(Sc, w, Vs, select=F)
while ns < n:
#print "stabilize matrix A"
w[ns:] = w[ns:]**-1
Sc[::n + 1] = w
Sc = Vs * Sc * Vs.H
syseA[:, :] = Sc.real()
Sc[:, :] = syseA
ns = lapack.gees(Sc, w, Vs, select=F)
# estimate B,D,x0 stored in vector [x0; vec(D); vec(B)]
F1 = matrix(0.0, (p * N, n))
F1[:p, :] = syseC
for ii in xrange(1, N):
F1[ii * p:(ii + 1) * p, :] = F1[(ii - 1) * p:ii * p, :] * syseA
F2 = matrix(0.0, (p * N, p * m))
ut = u.T
for ii in xrange(p):
F2[ii::p, ii::p] = ut
F3 = matrix(0.0, (p * N, n * m))
F3t = matrix(0.0, (p * (N - 1), n * m))
for ii in xrange(1, N):
for jj in xrange(p):
for kk in xrange(n):
F3t[jj:jj + (N - ii) * p:p,
kk::n] = ut[:N - ii, :] * F1[(ii - 1) * p + jj, kk]
F3[ii * p:, :] = F3[ii * p:, :] + F3t[:(N - ii) * p, :]
F = matrix([[F1], [F2], [F3]])
yls = y[:]
Sls = matrix(0.0, (F.size[1], 1))
Uls = matrix(0.0, (F.size[0], F.size[1]))
Vtls = matrix(0.0, (F.size[1], F.size[1]))
lapack.gesvd(F, Sls, jobu='S', jobvt='S', U=Uls, Vt=Vtls)
Frank = len([ii for ii in xrange(Sls.size[0]) if Sls[ii] >= 1E-6])
#print 'Rank deficiency = ', F.size[1] - Frank
xx = matrix(0.0, (F.size[1], 1))
xx[:Frank] = Uls.T[:Frank, :] * yls
xx[:Frank] = base.mul(xx[:Frank], Sls[:Frank]**-1)
xx[:] = Vtls.T[:, :Frank] * xx[:Frank]
blas.gemv(F, xx, yls, beta=-1.0)
xxerr = blas.nrm2(yls)
x0 = xx[:n]
syseD = xx[n:n + p * m]
syseD.size = (p, m)
syseB = xx[n + p * m:]
syseB.size = (n, m)
return {'A': syseA, 'B': syseB, 'C': syseC, 'D': syseD, 'svN': svN, 'sv': \
sv, 'x0': x0, 'n': n, 'Aerr': Aerr, 'xxerr': xxerr}
def F(W):
"""
Create a solver for the linear equations
C * ux + G' * uzl - 2*A'(uzs21) = bx
-uzs11 = bX1
-uzs22 = bX2
G * ux - Dl^2 * uzl = bzl
[ -uX1 -A(ux)' ] [ uzs11 uzs21' ]
[ ] - r*r' * [ ] * r*r' = bzs
[ -A(ux) -uX2 ] [ uzs21 uzs22 ]
where Dl = diag(W['l']), r = W['r'][0].
On entry, x = (bx, bX1, bX2) and z = [ bzl; bzs[:] ].
On exit, x = (ux, uX1, uX2) and z = [ Dl*uzl; (r'*uzs*r)[:] ].
1. Compute matrices V1, V2 such that (with T = r*r')
[ V1 0 ] [ T11 T21' ] [ V1' 0 ] [ I S' ]
[ ] [ ] [ ] = [ ]
[ 0 V2' ] [ T21 T22 ] [ 0 V2 ] [ S I ]
and S = [ diag(s); 0 ], s a positive q-vector.
2. Factor the mapping X -> X + S * X' * S:
X + S * X' * S = L( L'( X )).
3. Compute scaled mappings: a matrix As with as its columns the
coefficients of the scaled mapping
L^-1( V2' * A() * V1' )
and the matrix Gs = Dl^-1 * G.
4. Cholesky factorization of H = C + Gs'*Gs + 2*As'*As.
"""
# 1. Compute V1, V2, s.
r = W['r'][0]
# LQ factorization R[:q, :] = L1 * Q1.
lapack.lacpy(r, Q1, m=q)
lapack.gelqf(Q1, tau1)
lapack.lacpy(Q1, L1, n=q, uplo='L')
lapack.orglq(Q1, tau1)
# LQ factorization R[q:, :] = L2 * Q2.
lapack.lacpy(r, Q2, m=p, offsetA=q)
lapack.gelqf(Q2, tau2)
lapack.lacpy(Q2, L2, n=p, uplo='L')
lapack.orglq(Q2, tau2)
# V2, V1, s are computed from an SVD: if
#
# Q2 * Q1' = U * diag(s) * V',
#
# then V1 = V' * L1^-1 and V2 = L2^-T * U.
# T21 = Q2 * Q1.T
blas.gemm(Q2, Q1, T21, transB='T')
# SVD T21 = U * diag(s) * V'. Store U in V2 and V' in V1.
lapack.gesvd(T21, s, jobu='A', jobvt='A', U=V2, Vt=V1)
# # Q2 := Q2 * Q1' without extracting Q1; store T21 in Q2
# this will requires lapack.ormlq or lapack.unmlq
# V2 = L2^-T * U
blas.trsm(L2, V2, transA='T')
# V1 = V' * L1^-1
blas.trsm(L1, V1, side='R')
# 2. Factorization X + S * X' * S = L( L'( X )).
#
# The factor L is stored as a diagonal matrix D and a sparse lower
# triangular matrix P, such that
#
# L(X)[:] = D**-1 * (I + P) * X[:]
# L^-1(X)[:] = D * (I - P) * X[:].
# SS is q x q with SS[i,j] = si*sj.
blas.scal(0.0, SS)
blas.syr(s, SS)
# For a p x q matrix X, P*X[:] is Y[:] where
#
# Yij = si * sj * Xji if i < j
# = 0 otherwise.
#
P.V = SS[Itril2]
# For a p x q matrix X, D*X[:] is Y[:] where
#
# Yij = Xij / sqrt( 1 - si^2 * sj^2 ) if i < j
# = Xii / sqrt( 1 + si^2 ) if i = j
# = Xij otherwise.
#
DV[Idiag] = sqrt(1.0 + SS[::q + 1])
DV[Itriu] = sqrt(1.0 - SS[Itril3]**2)
D.V = DV**-1
# 3. Scaled linear mappings
# Ask := V2' * Ask * V1'
blas.scal(0.0, As)
base.axpy(A, As)
for i in xrange(n):
# tmp := V2' * As[i, :]
blas.gemm(V2,
As,
tmp,
transA='T',
m=p,
n=q,
k=p,
ldB=p,
offsetB=i * p * q)
# As[:,i] := tmp * V1'
blas.gemm(tmp,
V1,
As,
transB='T',
m=p,
n=q,
k=q,
ldC=p,
offsetC=i * p * q)
# As := D * (I - P) * As
# = L^-1 * As.
blas.copy(As, As2)
base.gemm(P, As, As2, alpha=-1.0, beta=1.0)
base.gemm(D, As2, As)
# Gs := Dl^-1 * G
blas.scal(0.0, Gs)
base.axpy(G, Gs)
for k in xrange(n):
blas.tbmv(W['di'], Gs, n=m, k=0, ldA=1, offsetx=k * m)
# 4. Cholesky factorization of H = C + Gs' * Gs + 2 * As' * As.
blas.syrk(As, H, trans='T', alpha=2.0)
blas.syrk(Gs, H, trans='T', beta=1.0)
base.axpy(C, H)
lapack.potrf(H)
def f(x, y, z):
"""
Solve
C * ux + G' * uzl - 2*A'(uzs21) = bx
-uzs11 = bX1
-uzs22 = bX2
G * ux - D^2 * uzl = bzl
[ -uX1 -A(ux)' ] [ uzs11 uzs21' ]
[ ] - T * [ ] * T = bzs.
[ -A(ux) -uX2 ] [ uzs21 uzs22 ]
On entry, x = (bx, bX1, bX2) and z = [ bzl; bzs[:] ].
On exit, x = (ux, uX1, uX2) and z = [ D*uzl; (r'*uzs*r)[:] ].
Define X = uzs21, Z = T * uzs * T:
C * ux + G' * uzl - 2*A'(X) = bx
[ 0 X' ] [ bX1 0 ]
T * [ ] * T - Z = T * [ ] * T
[ X 0 ] [ 0 bX2 ]
G * ux - D^2 * uzl = bzl
[ -uX1 -A(ux)' ] [ Z11 Z21' ]
[ ] - [ ] = bzs
[ -A(ux) -uX2 ] [ Z21 Z22 ]
Return x = (ux, uX1, uX2), z = [ D*uzl; (rti'*Z*rti)[:] ].
We use the congruence transformation
[ V1 0 ] [ T11 T21' ] [ V1' 0 ] [ I S' ]
[ ] [ ] [ ] = [ ]
[ 0 V2' ] [ T21 T22 ] [ 0 V2 ] [ S I ]
and the factorization
X + S * X' * S = L( L'(X) )
to write this as
C * ux + G' * uzl - 2*A'(X) = bx
L'(V2^-1 * X * V1^-1) - L^-1(V2' * Z21 * V1') = bX
G * ux - D^2 * uzl = bzl
[ -uX1 -A(ux)' ] [ Z11 Z21' ]
[ ] - [ ] = bzs,
[ -A(ux) -uX2 ] [ Z21 Z22 ]
or
C * ux + Gs' * uuzl - 2*As'(XX) = bx
XX - ZZ21 = bX
Gs * ux - uuzl = D^-1 * bzl
-As(ux) - ZZ21 = bbzs_21
-uX1 - Z11 = bzs_11
-uX2 - Z22 = bzs_22
if we introduce scaled variables
uuzl = D * uzl
XX = L'(V2^-1 * X * V1^-1)
= L'(V2^-1 * uzs21 * V1^-1)
ZZ21 = L^-1(V2' * Z21 * V1')
and define
bbzs_21 = L^-1(V2' * bzs_21 * V1')
[ bX1 0 ]
bX = L^-1( V2' * (T * [ ] * T)_21 * V1').
[ 0 bX2 ]
Eliminating Z21 gives
C * ux + Gs' * uuzl - 2*As'(XX) = bx
Gs * ux - uuzl = D^-1 * bzl
-As(ux) - XX = bbzs_21 - bX
-uX1 - Z11 = bzs_11
-uX2 - Z22 = bzs_22
and eliminating uuzl and XX gives
H * ux = bx + Gs' * D^-1 * bzl + 2*As'(bX - bbzs_21)
Gs * ux - uuzl = D^-1 * bzl
-As(ux) - XX = bbzs_21 - bX
-uX1 - Z11 = bzs_11
-uX2 - Z22 = bzs_22.
In summary, we can use the following algorithm:
1. bXX := bX - bbzs21
[ bX1 0 ]
= L^-1( V2' * ((T * [ ] * T)_21 - bzs_21) * V1')
[ 0 bX2 ]
2. Solve H * ux = bx + Gs' * D^-1 * bzl + 2*As'(bXX).
3. From ux, compute
uuzl = Gs*ux - D^-1 * bzl and
X = V2 * L^-T(-As(ux) + bXX) * V1.
4. Return ux, uuzl,
rti' * Z * rti = r' * [ -bX1, X'; X, -bX2 ] * r
and uX1 = -Z11 - bzs_11, uX2 = -Z22 - bzs_22.
"""
# Save bzs_11, bzs_22, bzs_21.
lapack.lacpy(z, bz11, uplo='L', m=q, n=q, ldA=p + q, offsetA=m)
lapack.lacpy(z, bz21, m=p, n=q, ldA=p + q, offsetA=m + q)
lapack.lacpy(z,
bz22,
uplo='L',
m=p,
n=p,
ldA=p + q,
offsetA=m + (p + q + 1) * q)
# zl := D^-1 * zl
# = D^-1 * bzl
blas.tbmv(W['di'], z, n=m, k=0, ldA=1)
# zs := r' * [ bX1, 0; 0, bX2 ] * r.
# zs := [ bX1, 0; 0, bX2 ]
blas.scal(0.0, z, offset=m)
lapack.lacpy(x[1], z, uplo='L', m=q, n=q, ldB=p + q, offsetB=m)
lapack.lacpy(x[2],
z,
uplo='L',
m=p,
n=p,
ldB=p + q,
offsetB=m + (p + q + 1) * q)
# scale diagonal of zs by 1/2
blas.scal(0.5, z, inc=p + q + 1, offset=m)
# a := tril(zs)*r
blas.copy(r, a)
blas.trmm(z,
a,
side='L',
m=p + q,
n=p + q,
ldA=p + q,
ldB=p + q,
offsetA=m)
# zs := a'*r + r'*a
blas.syr2k(r,
a,
z,
trans='T',
n=p + q,
k=p + q,
ldB=p + q,
ldC=p + q,
offsetC=m)
# bz21 := L^-1( V2' * ((r * zs * r')_21 - bz21) * V1')
#
# [ bX1 0 ]
# = L^-1( V2' * ((T * [ ] * T)_21 - bz21) * V1').
# [ 0 bX2 ]
# a = [ r21 r22 ] * z
# = [ r21 r22 ] * r' * [ bX1, 0; 0, bX2 ] * r
# = [ T21 T22 ] * [ bX1, 0; 0, bX2 ] * r
blas.symm(z,
r,
a,
side='R',
m=p,
n=p + q,
ldA=p + q,
ldC=p + q,
offsetB=q)
# bz21 := -bz21 + a * [ r11, r12 ]'
# = -bz21 + (T * [ bX1, 0; 0, bX2 ] * T)_21
blas.gemm(a,
r,
bz21,
transB='T',
m=p,
n=q,
k=p + q,
beta=-1.0,
ldA=p + q,
ldC=p)
# bz21 := V2' * bz21 * V1'
# = V2' * (-bz21 + (T*[bX1, 0; 0, bX2]*T)_21) * V1'
blas.gemm(V2, bz21, tmp, transA='T', m=p, n=q, k=p, ldB=p)
blas.gemm(tmp, V1, bz21, transB='T', m=p, n=q, k=q, ldC=p)
# bz21[:] := D * (I-P) * bz21[:]
# = L^-1 * bz21[:]
# = bXX[:]
blas.copy(bz21, tmp)
base.gemv(P, bz21, tmp, alpha=-1.0, beta=1.0)
base.gemv(D, tmp, bz21)
# Solve H * ux = bx + Gs' * D^-1 * bzl + 2*As'(bXX).
# x[0] := x[0] + Gs'*zl + 2*As'(bz21)
# = bx + G' * D^-1 * bzl + 2 * As'(bXX)
blas.gemv(Gs, z, x[0], trans='T', alpha=1.0, beta=1.0)
blas.gemv(As, bz21, x[0], trans='T', alpha=2.0, beta=1.0)
# x[0] := H \ x[0]
# = ux
lapack.potrs(H, x[0])
# uuzl = Gs*ux - D^-1 * bzl
blas.gemv(Gs, x[0], z, alpha=1.0, beta=-1.0)
# bz21 := V2 * L^-T(-As(ux) + bz21) * V1
# = X
blas.gemv(As, x[0], bz21, alpha=-1.0, beta=1.0)
blas.tbsv(DV, bz21, n=p * q, k=0, ldA=1)
blas.copy(bz21, tmp)
base.gemv(P, tmp, bz21, alpha=-1.0, beta=1.0, trans='T')
blas.gemm(V2, bz21, tmp)
blas.gemm(tmp, V1, bz21)
# zs := -zs + r' * [ 0, X'; X, 0 ] * r
# = r' * [ -bX1, X'; X, -bX2 ] * r.
# a := bz21 * [ r11, r12 ]
# = X * [ r11, r12 ]
blas.gemm(bz21, r, a, m=p, n=p + q, k=q, ldA=p, ldC=p + q)
# z := -z + [ r21, r22 ]' * a + a' * [ r21, r22 ]
# = rti' * uzs * rti
blas.syr2k(r,
a,
z,
trans='T',
beta=-1.0,
n=p + q,
k=p,
offsetA=q,
offsetC=m,
ldB=p + q,
ldC=p + q)
# uX1 = -Z11 - bzs_11
# = -(r*zs*r')_11 - bzs_11
# uX2 = -Z22 - bzs_22
# = -(r*zs*r')_22 - bzs_22
blas.copy(bz11, x[1])
blas.copy(bz22, x[2])
# scale diagonal of zs by 1/2
blas.scal(0.5, z, inc=p + q + 1, offset=m)
# a := r*tril(zs)
blas.copy(r, a)
blas.trmm(z,
a,
side='R',
m=p + q,
n=p + q,
ldA=p + q,
ldB=p + q,
offsetA=m)
# x[1] := -x[1] - a[:q,:] * r[:q, :]' - r[:q,:] * a[:q,:]'
# = -bzs_11 - (r*zs*r')_11
blas.syr2k(a, r, x[1], n=q, alpha=-1.0, beta=-1.0)
# x[2] := -x[2] - a[q:,:] * r[q:, :]' - r[q:,:] * a[q:,:]'
# = -bzs_22 - (r*zs*r')_22
blas.syr2k(a,
r,
x[2],
n=p,
alpha=-1.0,
beta=-1.0,
offsetA=q,
offsetB=q)
# scale diagonal of zs by 1/2
blas.scal(2.0, z, inc=p + q + 1, offset=m)
return f
def solve(x, y, z):
"""
1. Solve for usx[0]:
Asc'(Asc(usx[0]))
= bx0 + Asc'( ( bsz0 - bsz1 + S * bsx[1] * S ) ./ sqrtG)
= bx0 + Asc'( ( bsz0 + S * ( bsx[1] - bssz1) S )
./ sqrtG)
where bsx[1] = U^-1 * bx[1] * U^-T, bsz0 = U' * bz0 * U,
bsz1 = U' * bz1 * U, bssz1 = S^-1 * bsz1 * S^-1
2. Solve for usx[1]:
usx[1] + S * usx[1] * S
= S * ( As(usx[0]) + bsx[1] - bsz0 ) * S - bsz1
usx[1]
= ( S * (As(usx[0]) + bsx[1] - bsz0) * S - bsz1) ./ Gamma
= -bsz0 + (S * As(usx[0]) * S) ./ Gamma
+ (bsz0 - bsz1 + S * bsx[1] * S ) . / Gamma
= -bsz0 + (S * As(usx[0]) * S) ./ Gamma
+ (bsz0 + S * ( bsx[1] - bssz1 ) * S ) . / Gamma
Unscale ux[1] = Uti * usx[1] * Uti'
3. Compute usz0, usz1
r0' * uz0 * r0 = r0^-1 * ( A(ux[0]) - ux[1] - bz0 ) * r0^-T
r1' * uz1 * r1 = r1^-1 * ( -ux[1] - bz1 ) * r1^-T
"""
# z0 := U' * z0 * U
# = bsz0
__cngrnc(U, z, trans = 'T')
# z1 := Us' * bz1 * Us
# = S^-1 * U' * bz1 * U * S^-1
# = S^-1 * bsz1 * S^-1
__cngrnc(Us, z, trans = 'T', offsetx = msq)
# x[1] := Uti' * x[1] * Uti
# = bsx[1]
__cngrnc(Uti, x[1], trans = 'T')
# x[1] := x[1] - z[msq:]
# = bsx[1] - S^-1 * bsz1 * S^-1
blas.axpy(z, x[1], alpha = -1.0, offsetx = msq)
# x1 = (S * x[1] * S + z[:msq] ) ./ sqrtG
# = (S * ( bsx[1] - S^-1 * bsz1 * S^-1) * S + bsz0 ) ./ sqrtG
# = (S * bsx[1] * S - bsz1 + bsz0 ) ./ sqrtG
# in packed storage
blas.copy(x[1], x1)
blas.tbmv(S, x1, n = msq, k = 0, ldA = 1)
blas.axpy(z, x1, n = msq)
blas.tbsv(sqrtG, x1, n = msq, k = 0, ldA = 1)
misc.pack2(x1, {'l': 0, 'q': [], 's': [m]})
# x[0] := x[0] + Asc'*x1
# = bx0 + Asc'( ( bsz0 - bsz1 + S * bsx[1] * S ) ./ sqrtG)
# = bx0 + As'( ( bz0 - bz1 + S * bx[1] * S ) ./ Gamma )
blas.gemv(Asc, x1, x[0], m = mpckd, trans = 'T', beta = 1.0)
# x[0] := H^-1 * x[0]
# = ux[0]
lapack.potrs(H, x[0])
# x1 = Asc(x[0]) .* sqrtG (unpacked)
# = As(x[0])
blas.gemv(Asc, x[0], tmp, m = mpckd)
misc.unpack(tmp, x1, {'l': 0, 'q': [], 's': [m]})
blas.tbmv(sqrtG, x1, n = msq, k = 0, ldA = 1)
# usx[1] = (x1 + (x[1] - z[:msq])) ./ sqrtG**2
# = (As(ux[0]) + bsx[1] - bsz0 - S^-1 * bsz1 * S^-1)
# ./ Gamma
# x[1] := x[1] - z[:msq]
# = bsx[1] - bsz0 - S^-1 * bsz1 * S^-1
blas.axpy(z, x[1], -1.0, n = msq)
# x[1] := x[1] + x1
# = As(ux) + bsx[1] - bsz0 - S^-1 * bsz1 * S^-1
blas.axpy(x1, x[1])
# x[1] := x[1] / Gammma
# = (As(ux) + bsx[1] - bsz0 + S^-1 * bsz1 * S^-1 ) / Gamma
# = S^-1 * usx[1] * S^-1
blas.tbsv(Gamma, x[1], n = msq, k = 0, ldA = 1)
# z[msq:] := r1' * U * (-z[msq:] - x[1]) * U * r1
# := -r1' * U * S^-1 * (bsz1 + ux[1]) * S^-1 * U * r1
# := -r1' * uz1 * r1
blas.axpy(x[1], z, n = msq, offsety = msq)
blas.scal(-1.0, z, offset = msq)
__cngrnc(U, z, offsetx = msq)
__cngrnc(W['r'][1], z, trans = 'T', offsetx = msq)
# x[1] := S * x[1] * S
# = usx1
blas.tbmv(S, x[1], n = msq, k = 0, ldA = 1)
# z[:msq] = r0' * U' * ( x1 - x[1] - z[:msq] ) * U * r0
# = r0' * U' * ( As(ux) - usx1 - bsz0 ) * U * r0
# = r0' * U' * usz0 * U * r0
# = r0' * uz0 * r0
blas.axpy(x1, z, -1.0, n = msq)
blas.scal(-1.0, z, n = msq)
blas.axpy(x[1], z, -1.0, n = msq)
__cngrnc(U, z)
__cngrnc(W['r'][0], z, trans = 'T')
# x[1] := Uti * x[1] * Uti'
# = ux[1]
__cngrnc(Uti, x[1])
