func updateBlas(t *testing.T, Y1, Y2, C1, C2, T, W *matrix.FloatMatrix) {
if W.Rows() != C1.Cols() {
panic("W.Rows != C1.Cols")
}
// W = C1.T
ScalePlus(W, C1, 0.0, 1.0, TRANSB)
//fmt.Printf("W = C1.T:\n%v\n", W)
// W = C1.T*Y1
blas.TrmmFloat(Y1, W, 1.0, linalg.OptLower, linalg.OptUnit, linalg.OptRight)
t.Logf("W = C1.T*Y1:\n%v\n", W)
// W = W + C2.T*Y2
blas.GemmFloat(C2, Y2, W, 1.0, 1.0, linalg.OptTransA)
t.Logf("W = W + C2.T*Y2:\n%v\n", W)
// --- here: W == C.T*Y ---
// W = W*T
blas.TrmmFloat(T, W, 1.0, linalg.OptUpper, linalg.OptRight)
t.Logf("W = C.T*Y*T:\n%v\n", W)
// --- here: W == C.T*Y*T ---
// C2 = C2 - Y2*W.T
blas.GemmFloat(Y2, W, C2, -1, 1.0, linalg.OptTransB)
t.Logf("C2 = C2 - Y2*W.T:\n%v\n", C2)
// W = Y1*W.T ==> W.T = W*Y1.T
blas.TrmmFloat(Y1, W, 1.0, linalg.OptLower,
linalg.OptUnit, linalg.OptRight, linalg.OptTrans)
t.Logf("W.T = W*Y1.T:\n%v\n", W)
// C1 = C1 - W.T
ScalePlus(C1, W, 1.0, -1.0, TRANSB)
//fmt.Printf("C1 = C1 - W.T:\n%v\n", C1)
// --- here: C = (I - Y*T*Y.T).T * C ---
}
func trmmTest(t *testing.T, A *matrix.FloatMatrix, flags Flags, nb int) bool {
var B0 *matrix.FloatMatrix
N := A.Cols()
S := 0
E := A.Cols()
side := linalg.OptLeft
if flags&RIGHT != 0 {
B0 = matrix.FloatWithValue(2, A.Rows(), 2.0)
side = linalg.OptRight
E = B0.Rows()
} else {
B0 = matrix.FloatWithValue(A.Rows(), 2, 2.0)
E = B0.Cols()
}
B1 := B0.Copy()
trans := linalg.OptNoTrans
if flags&TRANSA != 0 {
trans = linalg.OptTransA
}
diag := linalg.OptNonUnit
if flags&UNIT != 0 {
diag = linalg.OptUnit
}
uplo := linalg.OptUpper
if flags&LOWER != 0 {
uplo = linalg.OptLower
}
blas.TrmmFloat(A, B0, 1.0, uplo, diag, trans, side)
if A.Rows() < 8 {
//t.Logf("..A\n%v\n", A)
t.Logf(" BLAS B0:\n%v\n", B0)
}
Ar := A.FloatArray()
Br := B1.FloatArray()
if nb != 0 {
DTrmmBlk(Br, Ar, 1.0, flags, B1.LeadingIndex(), A.LeadingIndex(),
N, S, E, nb)
} else {
DTrmmUnblk(Br, Ar, 1.0, flags, B1.LeadingIndex(), A.LeadingIndex(),
N, S, E, 0)
}
result := B0.AllClose(B1)
t.Logf(" B0 == B1: %v\n", result)
if A.Rows() < 8 {
t.Logf(" DTrmm B1:\n%v\n", B1)
}
return result
}
func mcsdp(w *matrix.FloatMatrix) (*Solution, error) {
//
// Returns solution x, z to
//
// (primal) minimize sum(x)
// subject to w + diag(x) >= 0
//
// (dual) maximize -tr(w*z)
// subject to diag(z) = 1
// z >= 0.
//
n := w.Rows()
G := &matrixFs{n}
cngrnc := func(r, x *matrix.FloatMatrix, alpha float64) (err error) {
// Congruence transformation
//
// x := alpha * r'*x*r.
//
// r and x are square matrices.
//
err = nil
// tx = matrix(x, (n,n)) is copying and reshaping
// scale diagonal of x by 1/2, (x is (n,n))
tx := x.Copy()
matrix.Reshape(tx, n, n)
tx.Diag().Scale(0.5)
// a := tril(x)*r
// (python: a = +r is really making a copy of r)
a := r.Copy()
err = blas.TrmmFloat(tx, a, 1.0, linalg.OptLeft)
// x := alpha*(a*r' + r*a')
err = blas.Syr2kFloat(r, a, tx, alpha, 0.0, linalg.OptTrans)
// x[:] = tx[:]
tx.CopyTo(x)
return
}
Fkkt := func(W *sets.FloatMatrixSet) (KKTFunc, error) {
// Solve
// -diag(z) = bx
// -diag(x) - inv(rti*rti') * z * inv(rti*rti') = bs
//
// On entry, x and z contain bx and bs.
// On exit, they contain the solution, with z scaled
// (inv(rti)'*z*inv(rti) is returned instead of z).
//
// We first solve
//
// ((rti*rti') .* (rti*rti')) * x = bx - diag(t*bs*t)
//
// and take z = -rti' * (diag(x) + bs) * rti.
var err error = nil
rti := W.At("rti")[0]
// t = rti*rti' as a nonsymmetric matrix.
t := matrix.FloatZeros(n, n)
err = blas.GemmFloat(rti, rti, t, 1.0, 0.0, linalg.OptTransB)
if err != nil {
return nil, err
}
// Cholesky factorization of tsq = t.*t.
tsq := matrix.Mul(t, t)
err = lapack.Potrf(tsq)
if err != nil {
return nil, err
}
f := func(x, y, z *matrix.FloatMatrix) (err error) {
// tbst := t * zs * t = t * bs * t
tbst := z.Copy()
matrix.Reshape(tbst, n, n)
cngrnc(t, tbst, 1.0)
// x := x - diag(tbst) = bx - diag(rti*rti' * bs * rti*rti')
diag := tbst.Diag().Transpose()
x.Minus(diag)
// x := (t.*t)^{-1} * x = (t.*t)^{-1} * (bx - diag(t*bs*t))
err = lapack.Potrs(tsq, x)
// z := z + diag(x) = bs + diag(x)
// z, x are really column vectors here
z.AddIndexes(matrix.MakeIndexSet(0, n*n, n+1), x.FloatArray())
// z := -rti' * z * rti = -rti' * (diag(x) + bs) * rti
cngrnc(rti, z, -1.0)
return nil
}
return f, nil
}
c := matrix.FloatWithValue(n, 1, 1.0)
// initial feasible x: x = 1.0 - min(lmbda(w))
lmbda := matrix.FloatZeros(n, 1)
wp := w.Copy()
lapack.Syevx(wp, lmbda, nil, 0.0, nil, []int{1, 1}, linalg.OptRangeInt)
x0 := matrix.FloatZeros(n, 1).Add(-lmbda.GetAt(0, 0) + 1.0)
s0 := w.Copy()
s0.Diag().Plus(x0.Transpose())
matrix.Reshape(s0, n*n, 1)
// initial feasible z is identity
z0 := matrix.FloatIdentity(n)
matrix.Reshape(z0, n*n, 1)
dims := sets.DSetNew("l", "q", "s")
dims.Set("s", []int{n})
primalstart := sets.FloatSetNew("x", "s")
dualstart := sets.FloatSetNew("z")
primalstart.Set("x", x0)
primalstart.Set("s", s0)
dualstart.Set("z", z0)
var solopts SolverOptions
solopts.MaxIter = 30
solopts.ShowProgress = false
h := w.Copy()
matrix.Reshape(h, h.NumElements(), 1)
return ConeLpCustomMatrix(c, G, h, nil, nil, dims, Fkkt, &solopts, primalstart, dualstart)
}
/*
Applies Nesterov-Todd scaling or its inverse.
Computes
x := W*x (trans is false 'N', inverse = false 'N')
x := W^T*x (trans is true 'T', inverse = false 'N')
x := W^{-1}*x (trans is false 'N', inverse = true 'T')
x := W^{-T}*x (trans is true 'T', inverse = true 'T').
x is a dense float matrix.
W is a MatrixSet with entries:
- W['dnl']: positive vector
- W['dnli']: componentwise inverse of W['dnl']
- W['d']: positive vector
- W['di']: componentwise inverse of W['d']
- W['v']: lists of 2nd order cone vectors with unit hyperbolic norms
- W['beta']: list of positive numbers
- W['r']: list of square matrices
- W['rti']: list of square matrices. rti[k] is the inverse transpose
of r[k].
The 'dnl' and 'dnli' entries are optional, and only present when the
function is called from the nonlinear solver.
*/
func scale(x *matrix.FloatMatrix, W *sets.FloatMatrixSet, trans, inverse bool) (err error) {
/*DEBUGGED*/
var wl []*matrix.FloatMatrix
var w *matrix.FloatMatrix
ind := 0
err = nil
// var minor int = 0
//if ! checkpnt.MinorEmpty() {
// minor = checkpnt.MinorTop()
//}
//fmt.Printf("\n%d.%04d scaling x=\n%v\n", checkpnt.Major(), minor, x.ToString("%.17f"))
// Scaling for nonlinear component xk is xk := dnl .* xk; inverse
// scaling is xk ./ dnl = dnli .* xk, where dnl = W['dnl'],
// dnli = W['dnli'].
if wl = W.At("dnl"); wl != nil {
if inverse {
w = W.At("dnli")[0]
} else {
w = W.At("dnl")[0]
}
for k := 0; k < x.Cols(); k++ {
err = blas.TbmvFloat(w, x, &la_.IOpt{"n", w.Rows()}, &la_.IOpt{"k", 0},
&la_.IOpt{"lda", 1}, &la_.IOpt{"offsetx", k * x.Rows()})
if err != nil {
//fmt.Printf("1. TbmvFloat: %v\n", err)
return
}
}
ind += w.Rows()
}
//if ! checkpnt.MinorEmpty() {
// checkpnt.Check("000scale", minor)
//}
// Scaling for linear 'l' component xk is xk := d .* xk; inverse
// scaling is xk ./ d = di .* xk, where d = W['d'], di = W['di'].
if inverse {
w = W.At("di")[0]
} else {
w = W.At("d")[0]
}
for k := 0; k < x.Cols(); k++ {
err = blas.TbmvFloat(w, x, &la_.IOpt{"n", w.Rows()}, &la_.IOpt{"k", 0},
&la_.IOpt{"lda", 1}, &la_.IOpt{"offsetx", k*x.Rows() + ind})
if err != nil {
//fmt.Printf("2. TbmvFloat: %v\n", err)
return
}
}
ind += w.Rows()
//if ! checkpnt.MinorEmpty() {
// checkpnt.Check("010scale", minor)
//}
// Scaling for 'q' component is
//
// xk := beta * (2*v*v' - J) * xk
// = beta * (2*v*(xk'*v)' - J*xk)
//
// where beta = W['beta'][k], v = W['v'][k], J = [1, 0; 0, -I].
//
//Inverse scaling is
//
// xk := 1/beta * (2*J*v*v'*J - J) * xk
// = 1/beta * (-J) * (2*v*((-J*xk)'*v)' + xk).
//wf := matrix.FloatZeros(x.Cols(), 1)
w = matrix.FloatZeros(x.Cols(), 1)
for k, v := range W.At("v") {
m := v.Rows()
if inverse {
blas.ScalFloat(x, -1.0, &la_.IOpt{"offset", ind}, &la_.IOpt{"inc", x.Rows()})
}
err = blas.GemvFloat(x, v, w, 1.0, 0.0, la_.OptTrans, &la_.IOpt{"m", m},
&la_.IOpt{"n", x.Cols()}, &la_.IOpt{"offsetA", ind},
&la_.IOpt{"lda", x.Rows()})
if err != nil {
//fmt.Printf("3. GemvFloat: %v\n", err)
return
}
err = blas.ScalFloat(x, -1.0, &la_.IOpt{"offset", ind}, &la_.IOpt{"inc", x.Rows()})
if err != nil {
return
}
err = blas.GerFloat(v, w, x, 2.0, &la_.IOpt{"m", m},
&la_.IOpt{"n", x.Cols()}, &la_.IOpt{"lda", x.Rows()},
&la_.IOpt{"offsetA", ind})
if err != nil {
//fmt.Printf("4. GerFloat: %v\n", err)
return
}
var a float64
if inverse {
blas.ScalFloat(x, -1.0,
&la_.IOpt{"offset", ind}, &la_.IOpt{"inc", x.Rows()})
// a[i,j] := 1.0/W[i,j]
a = 1.0 / W.At("beta")[0].GetIndex(k)
} else {
a = W.At("beta")[0].GetIndex(k)
}
for i := 0; i < x.Cols(); i++ {
blas.ScalFloat(x, a, &la_.IOpt{"n", m}, &la_.IOpt{"offset", ind + i*x.Rows()})
}
ind += m
}
//if ! checkpnt.MinorEmpty() {
// checkpnt.Check("020scale", minor)
//}
// Scaling for 's' component xk is
//
// xk := vec( r' * mat(xk) * r ) if trans = 'N'
// xk := vec( r * mat(xk) * r' ) if trans = 'T'.
//
// r is kth element of W['r'].
//
// Inverse scaling is
//
// xk := vec( rti * mat(xk) * rti' ) if trans = 'N'
// xk := vec( rti' * mat(xk) * rti ) if trans = 'T'.
//
// rti is kth element of W['rti'].
maxn := 0
for _, r := range W.At("r") {
if r.Rows() > maxn {
maxn = r.Rows()
}
}
a := matrix.FloatZeros(maxn, maxn)
for k, v := range W.At("r") {
t := trans
var r *matrix.FloatMatrix
if !inverse {
r = v
t = !trans
} else {
r = W.At("rti")[k]
}
n := r.Rows()
for i := 0; i < x.Cols(); i++ {
// scale diagonal of xk by 0.5
blas.ScalFloat(x, 0.5, &la_.IOpt{"offset", ind + i*x.Rows()},
&la_.IOpt{"inc", n + 1}, &la_.IOpt{"n", n})
// a = r*tril(x) (t is 'N') or a = tril(x)*r (t is 'T')
blas.Copy(r, a)
if !t {
err = blas.TrmmFloat(x, a, 1.0, la_.OptRight, &la_.IOpt{"m", n},
&la_.IOpt{"n", n}, &la_.IOpt{"lda", n}, &la_.IOpt{"ldb", n},
&la_.IOpt{"offsetA", ind + i*x.Rows()})
if err != nil {
//fmt.Printf("5. TrmmFloat: %v\n", err)
return
}
// x := (r*a' + a*r') if t is 'N'
err = blas.Syr2kFloat(r, a, x, 1.0, 0.0, la_.OptNoTrans, &la_.IOpt{"n", n},
&la_.IOpt{"k", n}, &la_.IOpt{"ldb", n}, &la_.IOpt{"ldc", n},
&la_.IOpt{"offsetC", ind + i*x.Rows()})
if err != nil {
//fmt.Printf("6. Syr2kFloat: %v\n", err)
return
}
} else {
err = blas.TrmmFloat(x, a, 1.0, la_.OptLeft, &la_.IOpt{"m", n},
&la_.IOpt{"n", n}, &la_.IOpt{"lda", n}, &la_.IOpt{"ldb", n},
&la_.IOpt{"offsetA", ind + i*x.Rows()})
if err != nil {
//fmt.Printf("7. TrmmFloat: %v\n", err)
return
}
// x := (r'*a + a'*r) if t is 'T'
err = blas.Syr2kFloat(r, a, x, 1.0, 0.0, la_.OptTrans, &la_.IOpt{"n", n},
&la_.IOpt{"k", n}, &la_.IOpt{"ldb", n}, &la_.IOpt{"ldc", n},
&la_.IOpt{"offsetC", ind + i*x.Rows()})
if err != nil {
//fmt.Printf("8. Syr2kFloat: %v\n", err)
return
}
}
}
ind += n * n
}
//if ! checkpnt.MinorEmpty() {
// checkpnt.Check("030scale", minor)
//}
return
}
/*
Returns the Nesterov-Todd scaling W at points s and z, and stores the
scaled variable in lmbda.
W * z = W^{-T} * s = lmbda.
W is a MatrixSet with entries:
- W['dnl']: positive vector
- W['dnli']: componentwise inverse of W['dnl']
- W['d']: positive vector
- W['di']: componentwise inverse of W['d']
- W['v']: lists of 2nd order cone vectors with unit hyperbolic norms
- W['beta']: list of positive numbers
- W['r']: list of square matrices
- W['rti']: list of square matrices. rti[k] is the inverse transpose
of r[k].
*/
func computeScaling(s, z, lmbda *matrix.FloatMatrix, dims *sets.DimensionSet, mnl int) (W *sets.FloatMatrixSet, err error) {
/*DEBUGGED*/
err = nil
W = sets.NewFloatSet("dnl", "dnli", "d", "di", "v", "beta", "r", "rti")
// For the nonlinear block:
//
// W['dnl'] = sqrt( s[:mnl] ./ z[:mnl] )
// W['dnli'] = sqrt( z[:mnl] ./ s[:mnl] )
// lambda[:mnl] = sqrt( s[:mnl] .* z[:mnl] )
var stmp, ztmp, lmd *matrix.FloatMatrix
if mnl > 0 {
stmp = matrix.FloatVector(s.FloatArray()[:mnl])
ztmp = matrix.FloatVector(z.FloatArray()[:mnl])
//dnl := stmp.Div(ztmp)
//dnl.Apply(dnl, math.Sqrt)
dnl := matrix.Sqrt(matrix.Div(stmp, ztmp))
//dnli := dnl.Copy()
//dnli.Apply(dnli, func(a float64)float64 { return 1.0/a })
dnli := matrix.Inv(dnl)
W.Set("dnl", dnl)
W.Set("dnli", dnli)
//lmd = stmp.Mul(ztmp)
//lmd.Apply(lmd, math.Sqrt)
lmd = matrix.Sqrt(matrix.Mul(stmp, ztmp))
lmbda.SetIndexesFromArray(lmd.FloatArray(), matrix.MakeIndexSet(0, mnl, 1)...)
} else {
// set for empty matrices
//W.Set("dnl", matrix.FloatZeros(0, 1))
//W.Set("dnli", matrix.FloatZeros(0, 1))
mnl = 0
}
// For the 'l' block:
//
// W['d'] = sqrt( sk ./ zk )
// W['di'] = sqrt( zk ./ sk )
// lambdak = sqrt( sk .* zk )
//
// where sk and zk are the first dims['l'] entries of s and z.
// lambda_k is stored in the first dims['l'] positions of lmbda.
m := dims.At("l")[0]
//td := s.FloatArray()
stmp = matrix.FloatVector(s.FloatArray()[mnl : mnl+m])
//zd := z.FloatArray()
ztmp = matrix.FloatVector(z.FloatArray()[mnl : mnl+m])
//fmt.Printf(".Sqrt()=\n%v\n", matrix.Div(stmp, ztmp).Sqrt().ToString("%.17f"))
//d := stmp.Div(ztmp)
//d.Apply(d, math.Sqrt)
d := matrix.Div(stmp, ztmp).Sqrt()
//di := d.Copy()
//di.Apply(di, func(a float64)float64 { return 1.0/a })
di := matrix.Inv(d)
//fmt.Printf("d:\n%v\n", d)
//fmt.Printf("di:\n%v\n", di)
W.Set("d", d)
W.Set("di", di)
//lmd = stmp.Mul(ztmp)
//lmd.Apply(lmd, math.Sqrt)
lmd = matrix.Mul(stmp, ztmp).Sqrt()
// lmd has indexes mnl:mnl+m and length of m
lmbda.SetIndexesFromArray(lmd.FloatArray(), matrix.MakeIndexSet(mnl, mnl+m, 1)...)
//fmt.Printf("after l:\n%v\n", lmbda)
/*
For the 'q' blocks, compute lists 'v', 'beta'.
The vector v[k] has unit hyperbolic norm:
(sqrt( v[k]' * J * v[k] ) = 1 with J = [1, 0; 0, -I]).
beta[k] is a positive scalar.
The hyperbolic Householder matrix H = 2*v[k]*v[k]' - J
defined by v[k] satisfies
(beta[k] * H) * zk = (beta[k] * H) \ sk = lambda_k
where sk = s[indq[k]:indq[k+1]], zk = z[indq[k]:indq[k+1]].
lambda_k is stored in lmbda[indq[k]:indq[k+1]].
*/
ind := mnl + dims.At("l")[0]
var beta *matrix.FloatMatrix
for _, k := range dims.At("q") {
W.Append("v", matrix.FloatZeros(k, 1))
}
beta = matrix.FloatZeros(len(dims.At("q")), 1)
W.Set("beta", beta)
vset := W.At("v")
for k, m := range dims.At("q") {
v := vset[k]
// a = sqrt( sk' * J * sk ) where J = [1, 0; 0, -I]
aa := jnrm2(s, m, ind)
// b = sqrt( zk' * J * zk )
bb := jnrm2(z, m, ind)
// beta[k] = ( a / b )**1/2
beta.SetIndex(k, math.Sqrt(aa/bb))
// c = sqrt( (sk/a)' * (zk/b) + 1 ) / sqrt(2)
c0 := blas.DotFloat(s, z, &la_.IOpt{"n", m},
&la_.IOpt{"offsetx", ind}, &la_.IOpt{"offsety", ind})
cc := math.Sqrt((c0/aa/bb + 1.0) / 2.0)
// vk = 1/(2*c) * ( (sk/a) + J * (zk/b) )
blas.CopyFloat(z, v, &la_.IOpt{"offsetx", ind}, &la_.IOpt{"n", m})
blas.ScalFloat(v, -1.0/bb)
v.SetIndex(0, -1.0*v.GetIndex(0))
blas.AxpyFloat(s, v, 1.0/aa, &la_.IOpt{"offsetx", ind}, &la_.IOpt{"n", m})
blas.ScalFloat(v, 1.0/2.0/cc)
// v[k] = 1/sqrt(2*(vk0 + 1)) * ( vk + e ), e = [1; 0]
v.SetIndex(0, v.GetIndex(0)+1.0)
blas.ScalFloat(v, (1.0 / math.Sqrt(2.0*v.GetIndex(0))))
/*
To get the scaled variable lambda_k
d = sk0/a + zk0/b + 2*c
lambda_k = [ c;
(c + zk0/b)/d * sk1/a + (c + sk0/a)/d * zk1/b ]
lambda_k *= sqrt(a * b)
*/
lmbda.SetIndex(ind, cc)
dd := 2*cc + s.GetIndex(ind)/aa + z.GetIndex(ind)/bb
blas.CopyFloat(s, lmbda, &la_.IOpt{"offsetx", ind + 1}, &la_.IOpt{"offsety", ind + 1},
&la_.IOpt{"n", m - 1})
zz := (cc + z.GetIndex(ind)/bb) / dd / aa
ss := (cc + s.GetIndex(ind)/aa) / dd / bb
blas.ScalFloat(lmbda, zz, &la_.IOpt{"offset", ind + 1}, &la_.IOpt{"n", m - 1})
blas.AxpyFloat(z, lmbda, ss, &la_.IOpt{"offsetx", ind + 1},
&la_.IOpt{"offsety", ind + 1}, &la_.IOpt{"n", m - 1})
blas.ScalFloat(lmbda, math.Sqrt(aa*bb), &la_.IOpt{"offset", ind}, &la_.IOpt{"n", m})
ind += m
//fmt.Printf("after q[%d]:\n%v\n", k, lmbda)
}
/*
For the 's' blocks: compute two lists 'r' and 'rti'.
r[k]' * sk^{-1} * r[k] = diag(lambda_k)^{-1}
r[k]' * zk * r[k] = diag(lambda_k)
where sk and zk are the entries inds[k] : inds[k+1] of
s and z, reshaped into symmetric matrices.
rti[k] is the inverse of r[k]', so
rti[k]' * sk * rti[k] = diag(lambda_k)^{-1}
rti[k]' * zk^{-1} * rti[k] = diag(lambda_k).
The vectors lambda_k are stored in
lmbda[ dims['l'] + sum(dims['q']) : -1 ]
*/
for _, k := range dims.At("s") {
W.Append("r", matrix.FloatZeros(k, k))
W.Append("rti", matrix.FloatZeros(k, k))
}
maxs := maxdim(dims.At("s"))
work := matrix.FloatZeros(maxs*maxs, 1)
Ls := matrix.FloatZeros(maxs*maxs, 1)
Lz := matrix.FloatZeros(maxs*maxs, 1)
ind2 := ind
for k, m := range dims.At("s") {
r := W.At("r")[k]
rti := W.At("rti")[k]
// Factor sk = Ls*Ls'; store Ls in ds[inds[k]:inds[k+1]].
blas.CopyFloat(s, Ls, &la_.IOpt{"offsetx", ind2}, &la_.IOpt{"n", m * m})
lapack.PotrfFloat(Ls, &la_.IOpt{"n", m}, &la_.IOpt{"lda", m})
// Factor zs[k] = Lz*Lz'; store Lz in dz[inds[k]:inds[k+1]].
blas.CopyFloat(z, Lz, &la_.IOpt{"offsetx", ind2}, &la_.IOpt{"n", m * m})
lapack.PotrfFloat(Lz, &la_.IOpt{"n", m}, &la_.IOpt{"lda", m})
// SVD Lz'*Ls = U*diag(lambda_k)*V'. Keep U in work.
for i := 0; i < m; i++ {
blas.ScalFloat(Ls, 0.0, &la_.IOpt{"offset", i * m}, &la_.IOpt{"n", i})
}
blas.CopyFloat(Ls, work, &la_.IOpt{"n", m * m})
blas.TrmmFloat(Lz, work, 1.0, la_.OptTransA, &la_.IOpt{"lda", m}, &la_.IOpt{"ldb", m},
&la_.IOpt{"n", m}, &la_.IOpt{"m", m})
lapack.GesvdFloat(work, lmbda, nil, nil,
la_.OptJobuO, &la_.IOpt{"lda", m}, &la_.IOpt{"offsetS", ind},
&la_.IOpt{"n", m}, &la_.IOpt{"m", m})
// r = Lz^{-T} * U
blas.CopyFloat(work, r, &la_.IOpt{"n", m * m})
blas.TrsmFloat(Lz, r, 1.0, la_.OptTransA,
&la_.IOpt{"lda", m}, &la_.IOpt{"n", m}, &la_.IOpt{"m", m})
// rti = Lz * U
blas.CopyFloat(work, rti, &la_.IOpt{"n", m * m})
blas.TrmmFloat(Lz, rti, 1.0,
&la_.IOpt{"lda", m}, &la_.IOpt{"n", m}, &la_.IOpt{"m", m})
// r := r * diag(sqrt(lambda_k))
// rti := rti * diag(1 ./ sqrt(lambda_k))
for i := 0; i < m; i++ {
a := math.Sqrt(lmbda.GetIndex(ind + i))
blas.ScalFloat(r, a, &la_.IOpt{"offset", m * i}, &la_.IOpt{"n", m})
blas.ScalFloat(rti, 1.0/a, &la_.IOpt{"offset", m * i}, &la_.IOpt{"n", m})
}
ind += m
ind2 += m * m
}
return
}