ngsolve 6.2.2506.post74.dev0__cp314-cp314-macosx_10_15_universal2.whl

ngsolve/eigenvalues.py

@@ -0,0 +1,364 @@
+from ngsolve.la import InnerProduct, MultiVector
+from math import sqrt
+from ngsolve import Projector, Norm, Matrix, Vector, IdentityMatrix
+
+def Orthogonalize (vecs, mat):
+    mv = []
+    for i in range(len(vecs)):
+        for j in range(i):
+            vecs[i] -= InnerProduct(vecs[i], mv[j]) * vecs[j]
+            
+        hv = mat.CreateRowVector()
+        hv.data = mat * vecs[i]
+        norm = sqrt(InnerProduct(vecs[i], hv))
+        vecs[i] *= 1/norm
+        hv *= 1/norm
+        mv.append (hv)
+
+
+def PINVIT1(mata, matm, pre, num=1, maxit=20, printrates=True, GramSchmidt=False):
+    """preconditioned inverse iteration"""
+    import scipy.linalg
+
+    r = mata.CreateRowVector()
+    Av = mata.CreateRowVector()
+    Mv = mata.CreateRowVector()
+
+    uvecs = []
+    for i in range(num):
+        uvecs.append (mata.CreateRowVector())
+    
+    vecs = []
+    for i in range(2*num):
+        vecs.append (mata.CreateRowVector())
+
+    for v in uvecs:
+        r.SetRandom()
+        v.data = pre * r
+
+    asmall = Matrix(2*num, 2*num)
+    msmall = Matrix(2*num, 2*num)
+    lams = num * [1]
+
+    for i in range(maxit):
+        
+        for j in range(num):
+            vecs[j].data = uvecs[j]
+            r.data = mata * vecs[j] - lams[j] * matm * vecs[j]
+            vecs[num+j].data = pre * r
+
+        if GramSchmidt:
+            Orthogonalize (vecs, matm)
+
+        for j in range(2*num):
+            Av.data = mata * vecs[j]
+            Mv.data = matm * vecs[j]
+            for k in range(2*num):
+                asmall[j,k] = InnerProduct(Av, vecs[k])
+                msmall[j,k] = InnerProduct(Mv, vecs[k])
+
+        ev,evec = scipy.linalg.eigh(a=asmall, b=msmall)
+        lams[:] = ev[0:num]
+        if printrates:
+            print (i, ":", lams)
+    
+        for j in range(num):
+            uvecs[j][:] = 0.0
+            for k in range(2*num):
+                uvecs[j].data += float(evec[k,j]) * vecs[k]
+
+    return lams, uvecs
+
+
+
+
+def PINVIT(mata, matm, pre, num=1, maxit=20, printrates=True, GramSchmidt=True):
+    """preconditioned inverse iteration"""
+    import scipy.linalg
+
+    r = mata.CreateRowVector()
+    
+    uvecs = MultiVector(r, num)
+    vecs = MultiVector(r, 2*num)
+    # hv = MultiVector(r, 2*num)
+
+    for v in vecs[0:num]:
+        v.SetRandom()
+    uvecs[:] = pre * vecs[0:num]
+    lams = Vector(num * [1])
+    
+    for i in range(maxit):
+        vecs[0:num] = mata * uvecs - (matm * uvecs).Scale (lams)
+        vecs[num:2*num] = pre * vecs[0:num]
+        vecs[0:num] = uvecs
+
+        vecs.Orthogonalize(matm)
+
+        # hv[:] = mata * vecs
+        # asmall = InnerProduct (vecs, hv)
+        # hv[:] = matm * vecs
+        # msmall = InnerProduct (vecs, hv)
+        asmall = InnerProduct (vecs, mata * vecs)
+        msmall = InnerProduct (vecs, matm * vecs)
+    
+        ev,evec = scipy.linalg.eigh(a=asmall, b=msmall)
+        lams = Vector(ev[0:num])
+        if printrates:
+            print (i, ":", list(lams))
+
+        uvecs[:] = vecs * Matrix(evec[:,0:num])
+    return lams, uvecs
+
+
+def LOBPCG(mata, matm, pre, num=1, maxit=20, initial=None, printrates=True, largest=False):
+    """Knyazev's cg-like extension of PINVIT"""
+    import scipy.linalg
+
+    r = mata.CreateRowVector()
+
+    if initial:
+        num=len(initial)
+        uvecs = initial
+    else:
+        uvecs = MultiVector(r, num)
+
+    vecs = MultiVector(r, 3*num)
+    for v in vecs:
+        r.SetRandom()
+        v.data = pre * r
+
+    if initial:
+         vecs[0:num] = uvecs       
+        
+    lams = Vector(num * [1])
+    
+    for i in range(maxit):
+        uvecs.data = mata * vecs[0:num] - (matm * vecs[0:num]).Scale (lams)
+        vecs[2*num:3*num] = pre * uvecs
+        
+        vecs.Orthogonalize(matm)
+
+        asmall = InnerProduct (vecs, mata * vecs)
+        msmall = InnerProduct (vecs, matm * vecs)
+    
+        ev,evec = scipy.linalg.eigh(a=asmall, b=msmall)
+
+        if not largest:
+            lams = Vector(ev[0:num])
+            if printrates:
+                print (i, ":", list(lams), flush=True)
+
+            uvecs[:] = vecs * Matrix(evec[:,0:num])
+            vecs[num:2*num] = vecs[0:num]
+            vecs[0:num] = uvecs
+        else:
+            lams = Vector(ev[2*num:3*num])
+            if printrates:
+                print (i, ":", list(lams), flush=True)
+
+            uvecs[:] = vecs * Matrix(evec[:,2*num:3*num])
+            vecs[num:2*num] = vecs[0:num]
+            vecs[0:num] = uvecs
+        
+    return lams, uvecs
+
+
+
+
+
+def Arnoldi (mat, tol=1e-10, maxiter=200):
+    import scipy.linalg
+    H = Matrix(maxiter,maxiter, complex=mat.is_complex)
+    H[:,:] = 0
+    v = mat.CreateVector(colvector=False)
+    abv = MultiVector(v, 0)
+    v.SetRandom()
+    v /= Norm(v)
+
+    for i in range(maxiter):
+        abv.Append(v)
+        v = (mat*v).Evaluate()
+        for j in range(i+1):
+            H[j,i] = InnerProduct(v, abv[j])
+            v -= H[j,i]*abv[j]
+        if i+1 < maxiter:
+            H[i+1,i] = Norm(v)
+        v = 1/Norm(v)*v
+
+    lam,ev = scipy.linalg.eig(H)
+    return Vector(lam), (abv*Matrix(ev)).Evaluate()
+    
+    
+
+
+
+
+
+# SOAR: A SECOND-ORDER ARNOLDI METHOD FOR THE SOLUTION OF THE QUADRATIC EIGENVALUE PROBLEM
+# Z. Bai and Y. Su, SIAM J. Matrix Anal. Appl 26, pp 640-659 (2005)
+# author: A. Schoefl
+
+def SOAR (A, B, maxiter=200):
+# first version without memory saving and breakdown 
+
+    q = A.CreateVector()
+    p = A.CreateVector()
+    s = A.CreateVector()
+    r = A.CreateVector()
+
+    Q = MultiVector(q,0)
+    P = MultiVector(p,0)
+    p[:] = 0
+    q.SetRandom()
+    q.FV().imag = 0      
+    q /= Norm(q)
+    T = Matrix(maxiter, complex=A.is_complex)
+    T[:,:] = 0
+    
+    for j in range(maxiter):
+        Q.Append(q)
+        P.Append(p)
+        r.data = A*q + B*p
+        s.data = q
+
+        for i in range(j+1):
+
+            T[i,j] = InnerProduct(r, Q[i])
+            r -= T[i,j]*Q[i]
+            s -= T[i,j]*P[i]
+        
+        if j+1 < maxiter:
+            T[j+1,j] = Norm(r)
+            if T[j+1,j] == 0:
+                print("SOAR stopped at iteration j = ", j)
+                break
+
+            q.data = 1/T[j+1,j]*r
+            p.data = 1/T[j+1,j]*s
+
+    return Q
+
+
+
+
+# author: A. Schoefl
+def TOAR (A, B, maxiter=200):
+
+    r = A.CreateVector()
+
+    tmp_np = np.zeros((len(r), 2))#, dtype=complex)
+    r.SetRandom()
+    r.FV().imag = 0      
+    r /= Norm(r)
+    tmp_np[:,0] = r.FV().real
+    r.SetRandom()
+    r.FV().imag = 0      
+    r /= Norm(r)
+    tmp_np[:,1] = r.FV().real
+
+    # tmp_np[:,1] = tmp_np[:,0] # just for testing
+
+
+    Q_np, X_np, perm = la.qr(tmp_np, pivoting=True, mode="economic")
+    # print(Q_np)
+    # print(Q_np.shape, X_np.shape, X_np)
+
+    Q = MultiVector(r,0)
+    # r.FV().NumPy()[:] = Q_np[:,0]
+    
+    r.FV().real[:] = Q_np[:,0]
+    r.FV().imag = 0
+    Q.Append(r)
+    # assign rank
+    if np.isclose(X_np[1,1], 0):
+        eta = 1
+    else:
+        eta = 2
+
+        # r.FV().NumPy()[:] = Q_np[:,1]
+        r.FV().real[:] = Q_np[:,1]
+        r.FV().imag = 0
+        Q.Append(r)
+
+    # print (Q_np)
+    # print (Q[0], Q[1])
+    print (InnerProduct(Q,Q))
+        
+    gamma = np.linalg.norm(tmp_np, ord='fro')
+    
+    U1 = Matrix(eta,1, True) 
+    U1.NumPy()[:,0] = X_np[:eta,1]/gamma
+
+    U2 = Matrix(eta,1, True) 
+    U2.NumPy()[:,0] = X_np[:eta,0]/gamma
+    print ("X_nb = ", X_np[:eta,:])
+    print ("U1 = ", U1)
+    print ("U2 = ", U2)
+
+    H = Matrix(maxiter, maxiter-1, True)
+
+    # TODO: would be more efficient in C++
+    for j in range(maxiter-1):
+
+        # print(U1, "\n", U2)
+
+        r.data = A*(Q*U1[:,j])+B*(Q*U2[:,j])
+        s = Vector(eta, True)
+        # MGS: orthogonalize r against Q
+        for i in range(eta):
+            s[i] = InnerProduct(r, Q[i], conjugate=True) # TODO: order of Q[i], r correct? 
+            r-=s[i]*Q[i]
+        alpha = InnerProduct(r,r, conjugate=True).real
+        
+        # MGS 
+        for i in range(j):
+            # TODO: I have to re-read in proof if should be conjugation makes sense
+            # (currently everything is real valued)
+            H[i,j] = InnerProduct(s, U1[:,i], conjugate=False) + \
+                        InnerProduct( U1[:,j], U2[:,i], conjugate=True)
+            s -= H[i,j]*U1[:,i]
+            U1[:,j] -= H[i,j]*U2[:,i]
+
+        H[j+1,j] = sqrt(alpha+InnerProduct(s,s,conjugate=True).real+
+                    InnerProduct(U1[:,j],U1[:,j], conjugate=True).real)
+
+        alpha = sqrt(alpha)
+
+        # breakdown
+        if H[j+1,j] == 0: 
+            print("breakdown in iteration ", j)
+            return Q
+
+        # deflation
+        if alpha == 0:
+            print("deflation in iteration ", j)
+            tmp = U1
+            U1 = Matrix(eta, j+2, True) 
+            U1[:,:j+1] = tmp
+            U1[:,j+1] = 1/H[j+1,j]*s
+
+            tmp = U2
+            U2 = Matrix(eta, j+2, True) 
+            U2[:,:j+1] = tmp
+            U2[:,j+1] = 1/H[j+1,j]*U1[:,j]
+        else:
+            # update rank
+            eta+=1
+
+            tmp = U2
+            U2 = Matrix(eta, j+2, True) 
+            U2[:-1,:j+1] = tmp
+            U2[:-1,j+1] = 1/H[j+1,j]*U1[:,j]
+            U2[eta-1, :] = 0
+
+            tmp = U1
+            U1 = Matrix(eta, j+2, True) 
+            U1[:-1,:j+1] = tmp[:,:]
+            U1[:-1,j+1] = 1/H[j+1,j]*s
+            U1[eta-1, j+1] = alpha/H[j+1,j]
+
+            Q.Append(1/alpha*r)                        
+
+
+    return Q
+

ngsolve/eigenvalues.pyi

@@ -0,0 +1,30 @@
+from __future__ import annotations
+from math import sqrt
+from ngsolve.bla import Matrix
+from ngsolve.bla import Norm
+from ngsolve.bla import Vector
+from ngsolve.la import IdentityMatrix
+from ngsolve.la import InnerProduct
+from ngsolve.la import MultiVector
+from ngsolve.la import Projector
+__all__: list[str] = ['Arnoldi', 'IdentityMatrix', 'InnerProduct', 'LOBPCG', 'Matrix', 'MultiVector', 'Norm', 'Orthogonalize', 'PINVIT', 'PINVIT1', 'Projector', 'SOAR', 'TOAR', 'Vector', 'sqrt']
+def Arnoldi(mat, tol = 1e-10, maxiter = 200):
+    ...
+def LOBPCG(mata, matm, pre, num = 1, maxit = 20, initial = None, printrates = True, largest = False):
+    """
+    Knyazev's cg-like extension of PINVIT
+    """
+def Orthogonalize(vecs, mat):
+    ...
+def PINVIT(mata, matm, pre, num = 1, maxit = 20, printrates = True, GramSchmidt = True):
+    """
+    preconditioned inverse iteration
+    """
+def PINVIT1(mata, matm, pre, num = 1, maxit = 20, printrates = True, GramSchmidt = False):
+    """
+    preconditioned inverse iteration
+    """
+def SOAR(A, B, maxiter = 200):
+    ...
+def TOAR(A, B, maxiter = 200):
+    ...