Home > manopt > manifolds > ttfixedrank > TTeMPS_1.1 > operators > TTeMPS_op_NN.m

TTeMPS_op_NN

PURPOSE ^

SYNOPSIS ^

This is a script file.

DESCRIPTION ^





CROSS-REFERENCE INFORMATION ^

This function calls:

This function is called by:




SUBFUNCTIONS ^




SOURCE CODE ^


0001 classdef TTeMPS_op_NN < TTeMPS_op_laplace
0002     % Class for nearest-neighbor-type operators like the Henon-Heiles Potential
0003 
0004     %   TTeMPS Toolbox.
0005     %   Michael Steinlechner, 2013-2016
0006     %   Questions and contact: michael.steinlechner@epfl.ch
0007     %   BSD 2-clause license, see LICENSE.txt
0008 
0009     properties
0010         L
0011         L1
0012         L2
0013         L3
0014         B
0015         C
0016         % precomputed spectral decomp of 1D Laplace:
0017         E_L1
0018         V_L1
0019         E_L2
0020         V_L2
0021         E_L3
0022         V_L3
0023 
0024     end
0025 
0026     methods
0027 
0028         function A = update_properties( A );
0029 
0030             A.rank = [1,  3*ones(1, length(A.U)-1), 1];  % the TT rank is always three for such Laplace-like tensors
0031             size_col_ = cellfun( @(y) size(y,1), A.U);
0032             A.size_col = size_col_ ./ (A.rank(1:end-1).*A.rank(2:end));
0033             A.size_row = cellfun( @(y) size(y,2), A.U);
0034             A.order = length( A.size_row );
0035         end
0036     end
0037 
0038 
0039     methods( Access = public )
0040 
0041         function A = TTeMPS_op_NN( n, d )
0042             
0043             one = ones(n,1);
0044             q = linspace( -10, 2, n)';
0045             h = abs(q(2) - q(1));
0046             lambda = 0.11;
0047             L = -spdiags( [one, -2*one, one], [-1 0 1], n, n) / (h^2);
0048 
0049             D = spdiags(q, 0, n, n);
0050             % superclass constructor
0051             A = A@TTeMPS_op_laplace( L, d );
0052             % precompute eigenvalue information and exponential for use in local
0053             % preconditioner:
0054             A.L = L;
0055 
0056             A.L1 = L + 0.5*D.^2 - lambda/3*D.^3 + lambda^2/16*D.^4;;
0057             A.L2 = L + 0.5*D.^2 - lambda/3*D.^3 + lambda^2/8*D.^4;
0058             A.L3 = L + 0.5*D.^2 + lambda^2/16*D.^4;
0059 
0060             [A.V_L1, A.E_L1] = eig(full(A.L1));
0061             A.E_L1 = diag(A.E_L1);
0062             [A.V_L2, A.E_L2] = eig(full(A.L2));
0063             A.E_L2 = diag(A.E_L2);
0064             [A.V_L3, A.E_L3] = eig(full(A.L3));
0065             A.E_L3 = diag(A.E_L3);
0066 
0067             A.B = D + lambda/8*D.^2;
0068             A.C = lambda * D.^2;
0069             M = speye( n, n );
0070 
0071             e1 = sparse( 1, 1, 1, 3, 1 );
0072             e2 = sparse( 2, 1, 1, 3, 1 );
0073             e3 = sparse( 3, 1, 1, 3, 1 );
0074 
0075             l_mid = sparse( 3, 1, 1, 9, 1 );
0076             b_mid = sparse( 6, 1, 1, 9, 1 );
0077             m_mid = sparse( [1;9], [1;1], [1;1], 9, 1 );
0078             c_mid = sparse( 2, 1, 1, 9, 1 );
0079 
0080             A.U = cell( 1, d );
0081             A.U{1} = kron( A.L1, e1 ) + kron( A.B, e2 ) + kron( M, e3);
0082             A_mid = kron( A.L2, l_mid ) + kron( A.B, b_mid ) + kron( M, m_mid) + kron( A.C, c_mid );
0083             for i=2:d-1
0084                 A.U{i} = A_mid;
0085             end
0086             A.U{d} = kron( M, e1 ) + kron( A.C, e2 ) + kron( A.L3, e3);
0087 
0088             A = update_properties( A );
0089            
0090         end
0091 
0092         function expB = constr_precond_inner( A, X, mu )
0093 
0094             n = size(A.L2, 1);
0095             sz = [X.rank(mu), X.size(mu), X.rank(mu+1)]
0096 
0097             B1 = zeros( X.rank(mu) );
0098             % calculate B1 part:
0099             for i = 1:mu-1
0100                 % apply L to the i'th core
0101                 tmp = X;
0102                 Xi = matricize( tmp.U{i}, 2 );
0103                 if i == 1
0104                     Xi = A.L1*Xi;
0105                 else
0106                     Xi = A.L2*Xi;
0107                 end
0108                 tmp.U{i} = tensorize( Xi, 2, [X.rank(i), n, X.rank(i+1)] );
0109                 B1 = B1 + innerprod( X, tmp, 'LR', mu-1);
0110             end
0111 
0112             B3 = zeros( X.rank(mu+1) );
0113             % calculate B3 part:
0114             for i = mu+1:A.order
0115                 tmp = X;
0116                 Xi = matricize( tmp.U{i}, 2 );
0117                 if i == A.order
0118                     Xi = A.L2*Xi;
0119                 else
0120                     Xi = A.L3*Xi;
0121                 end
0122                 tmp.U{i} = tensorize( Xi, 2, [X.rank(i), n, X.rank(i+1)] );
0123                 B3 = B3 + innerprod( X, tmp, 'RL', mu+1);
0124             end
0125             
0126             [V1,e1] = eig(B1);
0127             e1 = diag(e1);
0128             [V3,e3] = eig(B3);
0129             e3 = diag(e3);
0130 
0131             if mu == 1
0132                 lmin = min(e1) + min(A.E_L1) + min(e3);
0133                 lmax = max(e1) + max(A.E_L1) + max(e3);
0134             elseif mu == A.order
0135                 lmin = min(e1) + min(A.E_L3) + min(e3);
0136                 lmax = max(e1) + max(A.E_L3) + max(e3);
0137             else
0138                 lmin = min(e1) + min(A.E_L2) + min(e3);
0139                 lmax = max(e1) + max(A.E_L2) + max(e3);
0140             end
0141 
0142             R = lmax/lmin
0143             
0144             [omega, alpha] = load_coefficients( R );
0145 
0146             k = 3;
0147             omega = omega/lmin;
0148             alpha = alpha/lmin;
0149 
0150             expB = cell(3,k);
0151             
0152             for i = 1:k
0153                 expB{1,i} = omega(i) * V1*diag( exp( -alpha(i)*e1 ))*V1';    % include omega in first part
0154                 if mu == 1
0155                     expB{2,i} = A.V_L1*diag( exp( -alpha(i)*A.E_L1 ))*A.V_L1';
0156                 elseif mu == A.order
0157                     expB{2,i} = A.V_L3*diag( exp( -alpha(i)*A.E_L3 ))*A.V_L3';
0158                 else
0159                     expB{2,i} = A.V_L2*diag( exp( -alpha(i)*A.E_L2 ))*A.V_L2';
0160                 end
0161                 expB{3,i} = V3*diag( exp( -alpha(i)*e3 ))*V3';
0162             end
0163         end
0164 
0165         function expB = constr_precond_outer( A, X, mu1, mu2 )
0166             
0167             n = size(A.L2, 1);
0168 
0169             B1 = zeros( X.rank(mu1) );
0170             % calculate B1 part:
0171             for i = 1:mu1-1
0172                 % apply L to the i'th core
0173                 tmp = X;
0174                 Xi = matricize( tmp.U{i}, 2 );
0175                 if i == 1
0176                     Xi = A.L1*Xi;
0177                 else
0178                     Xi = A.L2*Xi;
0179                 end
0180                 tmp.U{i} = tensorize( Xi, 2, [X.rank(i), n, X.rank(i+1)] );
0181                 B1 = B1 + innerprod( X, tmp, 'LR', mu1-1);
0182             end
0183 
0184             B3 = zeros( X.rank(mu2+1) );
0185             % calculate B3 part:
0186             for i = mu2+1:A.order
0187                 tmp = X;
0188                 Xi = matricize( tmp.U{i}, 2 );
0189                 if i == A.order
0190                     Xi = A.L2*Xi;
0191                 else
0192                     Xi = A.L3*Xi;
0193                 end
0194                 tmp.U{i} = tensorize( Xi, 2, [X.rank(i), n, X.rank(i+1)] );
0195                 B3 = B3 + innerprod( X, tmp, 'RL', mu2+1);
0196             end
0197             
0198             [V1,e1] = eig(B1);
0199             e1 = diag(e1);
0200             [V3,e3] = eig(B3);
0201             e3 = diag(e3);
0202 
0203             if mu1 == 1
0204                 lmin = min(e1) + min(A.E_L1) + min(A.E_L2) + min(e3);
0205                 lmax = max(e1) + max(A.E_L1) + max(A.E_L2) + max(e3);
0206             elseif mu2 == A.order
0207                 lmin = min(e1) + min(A.E_L2) + min(A.E_L3) + min(e3);
0208                 lmax = max(e1) + max(A.E_L2) + max(A.E_L3) + max(e3);
0209             else
0210                 lmin = min(e1) + 2*min(A.E_L2) + min(e3);
0211                 lmax = max(e1) + 2*max(A.E_L2) + max(e3);
0212             end
0213 
0214             R = lmax/lmin
0215             
0216             [omega, alpha] = load_coefficients( R );
0217 
0218             k = 3;
0219             omega = omega/lmin;
0220             alpha = alpha/lmin;
0221 
0222             expB = cell(4,k);
0223             
0224             for i = 1:k
0225                 expB{1,i} = omega(i) * V1*diag( exp( -alpha(i)*e1 ))*V1';    % include omega in first part
0226                 if mu1 == 1
0227                     expB{2,i} = A.V_L1*diag( exp( -alpha(i)*A.E_L1 ))*A.V_L1';
0228                 else
0229                     expB{2,i} = A.V_L2*diag( exp( -alpha(i)*A.E_L2 ))*A.V_L2';
0230                 end
0231 
0232                 if mu2 == A.order
0233                     expB{3,i} = A.V_L3*diag( exp( -alpha(i)*A.E_L3 ))*A.V_L3';
0234                 else
0235                     expB{3,i} = A.V_L2*diag( exp( -alpha(i)*A.E_L2 ))*A.V_L2';
0236                 end
0237 
0238                 expB{4,i} = V3*diag( exp( -alpha(i)*e3 ))*V3';
0239             end
0240         end
0241 
0242         function P = constr_precond( A, k )
0243 
0244             d = A.order;
0245             size(A.L1)
0246             ev1 = eig(A.L1);
0247             ev2 = eig(A.L2);
0248             ev3 = eig(A.L3);
0249 
0250             lmin = min(ev1) + (d-2)*min(ev2) + min(ev3);
0251             lmax = max(ev1) + (d-2)*max(ev2) + max(ev3);
0252 
0253             R = lmax/lmin
0254 
0255             %  http://www.mis.mpg.de/scicomp/EXP_SUM/1_x/1_xk07_2E2
0256             %  0.0133615547183825570028305575534521842940   {omega[1]}
0257             %  0.0429728469424360175410925952177443321034   {omega[2]}
0258             %  0.1143029399081515586560726591147663100401   {omega[3]}
0259             %  0.2838881266934189482611071431161775535656   {omega[4]}
0260             %  0.6622322841999484042811198458711174907876   {omega[5]}
0261             %  1.4847175320092703810050463464342840325116   {omega[6]}
0262             %  3.4859753729916252771962870138366952232900   {omega[7]}
0263             %  0.0050213411684266507485648978019454613531   {alpha[1]}
0264             %  0.0312546410994290844202411500801774835168   {alpha[2]}
0265             %  0.1045970270084145620410366606112262388706   {alpha[3]}
0266             %  0.2920522758702768403556507270657505159761   {alpha[4]}
0267             %  0.7407504784499061527671195936939341208927   {alpha[5]}
0268             %  1.7609744335543204401530945069076494746696   {alpha[6]}
0269             %  4.0759036969145123916954953635638503328664   {alpha[7]}
0270             
0271             if k == 3
0272                 [omega, alpha] = load_coefficients( R );
0273 
0274             elseif k == 7
0275                 omega = [0.0133615547183825570028305575534521842940 0.0429728469424360175410925952177443321034 0.1143029399081515586560726591147663100401,...
0276                          0.2838881266934189482611071431161775535656 0.6622322841999484042811198458711174907876 1.4847175320092703810050463464342840325116,...
0277                          3.4859753729916252771962870138366952232900];
0278                 alpha = [0.0050213411684266507485648978019454613531 0.0312546410994290844202411500801774835168 0.1045970270084145620410366606112262388706,...
0279                          0.2920522758702768403556507270657505159761 0.7407504784499061527671195936939341208927 1.7609744335543204401530945069076494746696,...
0280                          4.0759036969145123916954953635638503328664];
0281             else
0282                 error('Unknown rank specified. Choose either k=3 or k=7');
0283             end
0284 
0285             omega = omega/lmin;
0286             alpha = alpha/lmin;
0287 
0288             E1 = reshape( expm( -alpha(1) * A.L1), [1, A.size_row(1), A.size_col(1), 1]);
0289             E2 = reshape( expm( -alpha(1) * A.L2), [1, A.size_row(2), A.size_col(2), 1]);
0290             E3 = reshape( expm( -alpha(1) * A.L3), [1, A.size_row(d), A.size_col(d), 1]);
0291             P = omega(1)*TTeMPS_op( [E1, repmat({E2},1,d-2), E3] );
0292             for i = 2:k
0293                 E1 = reshape( expm( -alpha(i) * A.L1), [1, A.size_row(1), A.size_col(1), 1]);
0294                 E2 = reshape( expm( -alpha(i) * A.L2), [1, A.size_row(2), A.size_col(2), 1]);
0295                 E3 = reshape( expm( -alpha(i) * A.L3), [1, A.size_row(d), A.size_col(d), 1]);
0296                 P = P + omega(i)*TTeMPS_op( [E1, repmat({E2},1,d-1), E3] );
0297             end
0298 
0299         end
0300 
0301     end
0302         
0303 end


Generated on Fri 30-Sep-2022 13:18:25 by m2html © 2005