gTensor Utilities

Overview of Utilities

• Initialize a tensor
• Permute ndtensor/permute.hpp (Completed)
• Reshape ndtensor/reshape.hpp and generic/reshape.hpp (Completed)
• Diag generic/diag.hpp (Completed)
• Trace generic/trace.hpp (Completed)
• Rscale generic/rscale.hpp (Test function to be implemented)
• Contract gtensor_blas_lapack/contract.hpp (Completed)
• Mult ndtensor/mult.hpp (Completed)
• Deparallelise ndtensor/deparallelise.hpp (Test function to be implemented)
• Direct Sum /ndtensor/directsum.hpp (Completed)
• QR Decomposition /ndtensor/qrcompress.hpp (Completed)
• SVD Compression /ndtensor/svdcompress.hpp (Completed)
• Fusion ./ndtensor/fusion.hpp (Completed)
• Matrix exp /ndtensor/expm4.hpp (Completed)

Usage

Initialize a tensor

gc::TensorNd<data_type, rank, order> TensorName(TensorExtent)

• data_type could be double or complex<double>
• rank: rank of your tensor
• order: data storage order could be row or column major. Available input is CblasRowMajor and CblasColMajor
• TensorName: name of your tensor
• TensorExtent: dimension of your tensor with rank defined.
• Example
  • Initialize an rank 3 empty tensor EmptyT: gc::TensorNd<double, 3, CblasRowMajor> EmptyT;
  • Initialize an rank 3 empty tensor EmptyT with extent 234: gc::TensorNd<double, 3, CblasRowMajor> EmptyT(2,3,4);
  • Fill the tensor with 1.: EmptyT.fill(1.)
  • Fill the tensor with random number auto g = drand48; EmptyT.fill(g);
  • Fill the tensor with your own array by iterator:

/* test_construct.cpp 
 * Test function for tensor initialization 
 */

#include <iostream> 
#include <gtensor/ndtensor.hpp>
#include <gtensor/tensorio.hpp> 

using namespace gc;

// Initialize a tensor using vector
void t_construct() {
    gc::TensorNd<double, 3, CblasRowMajor> CTensor(2,2,2);
    std::vector<double> a = {1, 2, 3, 4, 5, 6, 7, 8};
    int i = 0; 
    for (auto it = CTensor.begin(); it != CTensor.end(); ++it, ++i) {
        *it = a[i]; 
    }
    std::cout << CTensor << std::endl;
}

// Initalize a tensor using fill
void t_fill() { 
    gc::TensorNd<double, 3, CblasRowMajor> CTensor(2,3,4); 
    CTensor.fill(1.0); 
    std::cout << CTensor << std::endl;
}

// Initalize a matrix using eye 
void t_eye() {
    auto mat_id = gc::eye(10);
    std::cout << mat_id << std::endl;
}


/* !!!!!!!!!!!!!Warning!!!!!!!!!!!!!!!!!!!
 * ones and zeros contains template arguments including
 * data_type (double/complex double), array major (row major/column major).
 * Default major for `ones` and `zeros` are column major.
 * Mixed usage of column major and row major would give errors.
 */ 
// Initaliaze a tensor fill with 1 with default major (column)  
void t_ones() {
    std::array<size_t, 3> shape = {2, 2, 2};
    auto t_one = gc::ones(shape); // Default 
    std::cout << t_one << std::endl;
}

// Initialize a tensor fill with 0 with default major (column)
void t_zeros() {
    std::array<size_t, 3> shape = {2, 3, 4}; 
    auto t_zero = gc::zeros(shape); 
    std::cout << t_zero << std::endl;
}

// Initialize a tensor fill with 1 with row major 
void t_ones_row() {
    auto t_one = gc::ones<double, 2, CblasRowMajor>({4,4}); 
    std::cout << t_ones_row << std::endl;
}

int main() 
{
    // t_construct();
    // t_fill();
    // t_eye();
    // t_ones();
    // t_zeros();
    t_ones_row();
    return 0;
}

Permute

We could permute the tensor by a similar fashion in Matlab permute(tensor,order)

• tensor is the target tensor to be permuted;
• order is an array that defines the new tensor extent.
• Output: tensor of the same rank with new extent.

Example:

/* test_permute.cpp 
 * Test function for tensor permutation
 */

#include <iostream> 
#include <gtensor_operation/ndtensor_ext.hpp> 
#include <gtensor_operation/ndtensor_operation.hpp>
using namespace gc;

// Construct a tensor 
gc::TensorNd<double, 3, CblasRowMajor> callTensor() {
    gc::TensorNd<double, 3, CblasRowMajor> tensor(2, 3, 4);
    int i = 0; 
    for (auto it = tensor.begin(); it != tensor.end(); ++it, ++i) {
        *it = i;
    }
    return tensor;
}

// Permute a tensor 
void t_permute() {
    auto PTensor = callTensor(); 
    //const std::array<size_t, 3> order = {2, 1, 0};
    //std::cout << gc::permute(PTensor, order) << std::endl;
    std::cout << gc::permute(PTensor, {2, 1, 0}) << std::endl;
}

int main() 
{
    t_permute();
    return 0;
}

Reshape

Reshape is slightly different as permute function. permute is a standalone function whereas reshape is a member function of the tensor class. tensor.reshape(shape)

• tensor is the target tensor
• shape is the new shape defined by an array

Example:

/* test_reshape.cpp 
 * Test function for tensor reshape
 */

#include <iostream> 
#include <gtensor_operation/ndtensor_ext.hpp> 
#include <gtensor_operation/ndtensor_operation.hpp>
using namespace gc;

// Construct a tensor 
gc::TensorNd<double, 3, CblasRowMajor> callTensor() {
    gc::TensorNd<double, 3, CblasRowMajor> tensor(2, 3, 4);
    int i = 0; 
    for (auto it = tensor.begin(); it != tensor.end(); ++it, ++i) {
        *it = i;
    }
    return tensor;
}

// Reshape a tensor 
void t_reshape() {
    auto RTensor = callTensor(); 
    //std::array<size_t, 4> shape = {2, 2, 2, 3};
    std::cout << "Before reshaping" << std::endl
        << RTensor << std::endl 
        << "After reshaping" << std::endl 
        << RTensor.reshape<4>({2, 2, 2, 3}) << std::endl;
    // Note 4 is a template argument to specify the new rank.
}

int main() 
{
    t_reshape();
    return 0;
}

Diag

Similar to Matlab, diag takes the diagonal elements for rank 2 tensors (matrices) diag(tensor)

• tensor is the target tensor

Example:

/* test_diag.cpp 
 * Test function for diag (taking diagonal elements)
 */

#include <iostream> 
#include <gtensor_operation/ndtensor_ext.hpp> 
#include <gtensor_operation/ndtensor_operation.hpp>
using namespace gc;

// Construct a matrix
gc::TensorNd<double, 2, CblasRowMajor> callMat() {
    gc::TensorNd<double, 2, CblasRowMajor> tensor(4, 4);
    int i = 0; 
    for (auto it = tensor.begin(); it != tensor.end(); ++it, ++i) {
        *it = i;
    }
    return tensor;
}

// Take the diagonal elements 
void t_diag() {
    auto mat = callMat(); 
    std::cout << "The matrix" << std::endl
        << mat << std::endl 
        << "Diagonal elements are" << std::endl 
        << diag(mat) << std::endl;
}

int main() 
{
    t_diag();
    return 0;
}

Trace

Trace applies to rank 2 tensor (matrix) trace(tensor)

• tensor: target tensor
• return a double or double complex number

Example:

/* test_trace.cpp 
 * Test function for trace (taking the sum of diagonal elements)
 */

#include <iostream> 
#include <gtensor_operation/ndtensor_ext.hpp> 
#include <gtensor_operation/ndtensor_operation.hpp>
using namespace gc;

// Construct a matrix
gc::TensorNd<double, 2, CblasRowMajor> callMat() {
    gc::TensorNd<double, 2, CblasRowMajor> tensor(4,4);
    int i = 0; 
    for (auto it = tensor.begin(); it != tensor.end(); ++it, ++i) {
        *it = i;
    }
    return tensor;
}

// Take the sum of the diagonal elements 
void t_trace() {
    auto mat = callMat(); 
    std::cout << "The matrix" << std::endl
        << mat << std::endl 
        << "The trace is" << std::endl 
        << trace(mat) << std::endl;
}

int main() 
{
    t_trace();
    return 0;
}

Rscale

Rscale is implemented to reduce the computational cost for multiplication between diagonal matrix and a dense matrix. It is an auxiliary function for SVD compress. $B=AB$ Rscale (Double Check) and $B=BA$ Cscale (Double Check)

(standalone test function to be updated).

Contract

Tensor contraction. The usage is similar as that in Verstrate's demo code.

$$T_{ijlm} = \sum_k A_{ijk} B_{klm}$$

Example:

/* test_contract.cpp 
 * Test function for tensor contraction
 */

#include <iostream> 
#include <gtensor_operation/ndtensor_ext.hpp> 
#include <gtensor_operation/ndtensor_operation.hpp>
using namespace gc;

// Construct a tensor 
gc::TensorNd<double, 3, CblasRowMajor> callTensor(size_t _i,
        size_t _j, size_t _k) {
    gc::TensorNd<double, 3, CblasRowMajor> tensor(_i,_j,_k);
    int i = 0; 
    for (auto it = tensor.begin(); it != tensor.end(); ++it, ++i) {
        *it = i;
    }
    return tensor;
}

// Tensor contraction test
void t_contract() {
    auto FTensor1 = callTensor(2, 3, 4);
    auto FTensor2= callTensor(2, 2, 3);
    std::array<size_t, 1> ax1 = {1}, ax2 = {2};
    auto result = gc::contract(FTensor1, ax1, FTensor2, ax2);
    std::cout << "Tensor 1" << std::endl
        << FTensor1 << std::endl 
        << "Tensor 2" << std::endl 
        << FTensor2 << std::endl
        << "Contracted on indices 1 and 2 (0, 1, 2)" << std::endl
        << result << std::endl;
} 

int main() 
{
    t_contract();
    return 0;
}

Multiplication

Multiplication is defined for matrix (rank 2 tensor). * is overloaded to realize it.

Note that the major convention cannot be mixed.

Example:

/* test_mult.cpp 
 * Test function for mult (matrix)
 */

#include <iostream> 
#include <gtensor_operation/ndtensor_ext.hpp> 
#include <gtensor_operation/ndtensor_operation.hpp>
using namespace gc;

// Construct a matrix
gc::TensorNd<double, 2, CblasRowMajor> callMat() {
    gc::TensorNd<double, 2, CblasRowMajor> tensor(4,4);
    int i = 0; 
    for (auto it = tensor.begin(); it != tensor.end(); ++it, ++i) {
        *it = i;
    }
    return tensor;
}

// Take the diagonal elements 
void t_mult() {
    auto mat = callMat(); 
    std::cout << "The matrix" << std::endl
        << mat << std::endl 
        << "The product is for mat*mat is" << std::endl 
        << mat*mat << std::endl;
}

int main() 
{
    t_mult();
    return 0;
}

Deparallelisation

Deparallelisation is used for mps/mpo compress. Notes, here compress does not refer to SVD compress. By deparallelisation, the compression does not introduce any error. For example, we have matrix $$\begin{pmatrix} 1 & 2 & 1 \\ 3 & 6 & 5 \\ 2 & 4 & 2 \end{pmatrix}$$ where the first two two column are differ by a factor of 2 (i.e., they are linear dependent). Such matrix can be decompose into two matrix with dimension (3,2) and (2,3). As such, the dimension could be decreased.

(standalone test function to be updated)

Direct Sum

Direct sum for tensor is an important function. It is necessary for MPS/MPO addition.

For example,

• Case 1: the direct sum of a matrix A(2,3) and matrix B(3,3). We add up its axis(dimension) 0.

$$\begin{pmatrix}1 & 2 & 3 \\ 4 & 5 & 6\end{pmatrix} \oplus_{(0)} \begin{pmatrix}7 & 8 & 9 \\ 10 & 11 & 12 \\ 13 & 14 & 15 \end{pmatrix} = \begin{pmatrix}1 & 2 & 3 \\ 4 & 5 & 6 \\ 7 & 8 & 9 \\ 10 & 11 & 12 \\ 13 & 14 & 15 \end{pmatrix} $$

• Case 2: the direct sum of a matrix A(2,3) and matrix B(3,3). We add up its axis(dimension) 1. However, here the fixed axis does not have same size and the direct sum is not defined
• Case 3: the direct sum of a matrix A(2,2) and B(2,2) for both axes. The sum matrix is C(4,4). $$\begin{pmatrix}a & b \\ c & d\end{pmatrix} \oplus_{(0,1)} \begin{pmatrix}e & f \\ g & h\end{pmatrix} = \begin{pmatrix}a & b & 0 & 0 \\ c & d & 0 & 0 \\ 0 & 0 & e & f \\ 0 & 0 & g & h \end{pmatrix} $$
• Case 4: the direct sum of a matrix A(2,3) and B(4,5) for both axes. The sum matrix is C(6,8). $$ \mathbf{A}+\mathbf{B} = \begin{pmatrix} \mathbf{A} & 0 \\ 0 & \mathbf{B}\end{pmatrix} $$

Example:

/* test_directSum.cpp 
 * Test function for directSum (taking directSum)
 */

#include <iostream> 
#include <gtensor_operation/ndtensor_ext.hpp> 
#include <gtensor_operation/ndtensor_operation.hpp>
using namespace gc;

// Construct a matrix
gc::TensorNd<double, 2, CblasRowMajor> callMat(size_t n, size_t m) {
    gc::TensorNd<double, 2, CblasRowMajor> tensor(n,m);
    int i = 0; 
    for (auto it = tensor.begin(); it != tensor.end(); ++it, ++i) {
        *it = i;
    }
    return tensor;
}

// Case 1
// Take the direct sum (Summation one index)   
void t_directSum1() {
    std::cout << "mat1(2,3) +(0) mat2(3,3)" << std::endl;
    auto mat1 = callMat(2,3);
    auto mat2 = callMat(3,3); 
    std::array<size_t,1> ax = {0};
    std::cout << "Matrix 1" << std::endl
        << mat1 << std::endl
        << "Matrix 2" << std::endl 
        << mat2 << std::endl
        << "The direct sum is" << std::endl 
        << directSum(mat1, mat2, ax) << std::endl;
}

// Case 2 
// Take the direct sum (Summation one index)   
// This function would not be able to run as the dimension of 
// the fixed index should be the same
void t_directSum2() {
    std::cout << "mat1(2,3) +(1) mat2(3,3)" << std::endl;
    auto mat1 = callMat(2,3);
    auto mat2 = callMat(3,3); 
    std::array<size_t,1> ax = {1};
    std::cout << "Matrix 1" << std::endl
        << mat1 << std::endl
        << "Matrix 2" << std::endl 
        << mat2 << std::endl
        << "The direct sum is" << std::endl 
        << directSum(mat1, mat2, ax) << std::endl;
}

// Case 3
// Take the diret sum (Summation both index)
void t_directSum3() { 
    std::cout << "mat1(2,2) +(0,1) mat2(2,2)" << std::endl;
    auto mat1 = callMat(2,2);
    auto mat2 = callMat(2,2);
    std::array<size_t,2> ax = {0, 1};
    std::cout << "Matrix 1" << std::endl
        << mat1 << std::endl
        << "Matrix 2" << std::endl 
        << mat2 << std::endl
        << "The direct sum is" << std::endl
        << directSum(mat1, mat2, ax) << std::endl;
} 

// Case 4
// Take the diret sum (Summation both index)
void t_directSum4() { 
    std::cout << "mat1(2,3) +(0,1) mat2(4,5)" << std::endl;
    auto mat1 = callMat(2,3);
    auto mat2 = callMat(4,5);
    std::array<size_t,2> ax = {0, 1};
    std::cout << "Matrix 1" << std::endl
        << mat1 << std::endl
        << "Matrix 2" << std::endl 
        << mat2 << std::endl
        << "The direct sum is" << std::endl
        << directSum(mat1, mat2, ax) << std::endl;
}

int main() 
{
    t_directSum1();
 // t_directSum2();
    t_directSum3();
    t_directSum4();
    return 0;
}

Matrix Exponential

gc:expm(matrix)

• Pade approximation to 6/7th order

Example:

/* test_expm.cpp 
 * Test function for matrix exponential
 */

#include <iostream> 
#include <gtensor_operation/ndtensor_ext.hpp> 
#include <gtensor_operation/ndtensor_operation.hpp>
using namespace gc;

// Construct a matrix
gc::TensorNd<double, 2, CblasRowMajor> callMat() {
    gc::TensorNd<double, 2, CblasRowMajor> tensor(4,4);
    int i = 0; 
    for (auto it = tensor.begin(); it != tensor.end(); ++it, ++i) {
        *it = i;
    }
    return tensor;
}

// Take matrix exp
void t_expm() {
    auto mat = callMat();
    std::cout << "The matrix" << std::endl
        << mat << std::endl;
    gc::expm(mat); 
    std::cout << "The matrix exp is" << std::endl 
        << mat << std::endl;
}

int main() 
{
    t_expm();
    return 0;
}

QR Decomposition

The index convention for QR decomposition is defined in the following way:

• Giving a rank 4 tensor qrtensor(a,b,c,d), with a,b,c,d denoting the dimensions of respective indices.
• If the axis is defined as 3, i.e., the 2nd index with dimension b, qrtensor would be reshape as a matrix with dimension (abc,d)
• The outcome matrix would be Q=(abc,d), R=(d,d)
• If the axes are defined as 1,2, i.e, the 2nd and 3rd indices with dimension b, c respectively, qrtensor would be reshape as matrix with dimension (ad,bc)
• The outcome matrix would be Q=(ad,bc), R=(bc,bc) if ad > bc or Q=(ad,ad) R=(ad,bc) if ad < bc.

/* test_qrcompress.cpp 
 * Test function for qr decomposition
 */

#include <iostream> 
#include <gtensor_operation/ndtensor_ext.hpp> 
#include <gtensor_operation/ndtensor_operation.hpp>
using namespace gc;


// Construct a tensor 
gc::TensorNd<double, 4, CblasRowMajor> callTensor() {
    gc::TensorNd<double, 4, CblasRowMajor> tensor(2,3,4,2);
    int i = 0; 
    for (auto it = tensor.begin(); it != tensor.end(); ++it, ++i) {
        *it = i;
    }
    return tensor;
}

// QR Decomposition to a Tensor
void t_qr_oneindex() {
    auto QRTensor = callTensor(); 
    std::array<size_t, 1> ax = {3};
    auto result = gc::qrCompress(QRTensor,ax);
    std::cout << "The tensor" << std::endl
        << QRTensor << std::endl 
        << "Q Tensor" << std::endl 
        << result.first << std::endl
        << "R Tensor" << std::endl
        << result.second << std::endl;
        
} 


void t_qr_twoindices() {
    auto QRTensor = callTensor(); 
    std::array<size_t, 2> ax = {1, 2};
    auto result = gc::qrCompress(QRTensor,ax);
    std::cout << "The tensor" << std::endl
        << QRTensor << std::endl 
        << "Q Tensor" << std::endl 
        << result.first << std::endl
        << "R Tensor" << std::endl
        << result.second << std::endl;
} 


int main() 
{
    t_qr_oneindex();
    t_qr_twoindices();
    return 0;
}

SVD Decomposition

SVD Compress a single tensor depending on the max bond dimension and svdcutoff. The max bond dimension is first checked. No truncation would be done if the bond dimension is within the max bond dimension.

gc::svdCompress(tensor,ax,maxbonddimension,svdcutoff,verbose)

• tensor: tensor to be compressed
• ax: the axes on the tensor to be decomposed
• maxbonddimension: the upper bound for bond dimension (elements to be kept)
• svdcutoff: the cutoff error
• verbose: 0, 1, 2 for runtime output

The index convention follows the same spirit as QR compression/decomposition.

/* test_svdcompress.cpp 
 * Test function for SVD decomposition
 */

#include <iostream> 
#include <gtensor_operation/ndtensor_ext.hpp> 
#include <gtensor_operation/ndtensor_operation.hpp>
using namespace gc;

// Construct a tensor 
gc::TensorNd<double, 4, CblasRowMajor> callTensor() {
    gc::TensorNd<double, 4, CblasRowMajor> tensor(2,3,4,2);
    int i = 0; 
    for (auto it = tensor.begin(); it != tensor.end(); ++it, ++i) {
        *it = i;
    }
    return tensor;
}

// SVD Decomposition to a Tensor
void t_svd_oneindex() {
    auto SVDTensor = callTensor(); 
    std::array<size_t, 1> ax = {3};
    auto result = gc::svdCompress(SVDTensor, ax, -1, 0, 1);
    std::cout << "The tensor" << std::endl
        << SVDTensor << std::endl 
        << "S Tensor" << std::endl 
        << std::get<0>(result) << std::endl
        << "V Tensor" << std::endl
        << std::get<1>(result) << std::endl 
        << "D Tensor" << std::endl
        << std::get<2>(result) << std::endl;
} 


void t_svd_twoindices() {
    auto SVDTensor = callTensor(); 
    std::array<size_t, 2> ax = {2,3};
    auto result = gc::svdCompress(SVDTensor, ax, 10, 0, 1);
    std::cout << "The tensor" << std::endl
        << SVDTensor << std::endl 
        << "S Tensor" << std::endl 
        << std::get<0>(result) << std::endl
        << "V Tensor" << std::endl
        << std::get<1>(result) << std::endl 
        << "D Tensor" << std::endl
        << std::get<2>(result) << std::endl;
}

int main() 
{
    t_svd_oneindex();
    t_svd_twoindices();
    return 0;
}

Index fusion

Index fusion is a procedure for tensor contraction. Hence it is implemented for a pair of tensor instead of one. To use it for a single tensor, the second tensor input could be left as empty.

Alternatively, the contraction on the index could be replaced by using permute and reshape functions.

/* test_fusion.cpp 
 * Test function for tensor index fusion
 */

#include <iostream> 
#include <gtensor_operation/ndtensor_ext.hpp> 
#include <gtensor_operation/ndtensor_operation.hpp>
using namespace gc;

// Construct a tensor 
gc::TensorNd<double, 4, CblasRowMajor> callTensor() {
    gc::TensorNd<double, 4, CblasRowMajor> tensor(2,3,4,2);
    int i = 0; 
    for (auto it = tensor.begin(); it != tensor.end(); ++it, ++i) {
        *it = i;
    }
    return tensor;
}

// Tensor index fusion
void t_fusion() {
    auto FTensor = callTensor(); 
    auto FTensor1 = callTensor();
    auto FTensor2= callTensor();
    std::array<size_t, 2> ax = {1,2};
    auto result = gc::fusion(FTensor, ax, FTensor, ax);
    std::cout << "The tensor" << std::endl
        << FTensor << std::endl 
        << "Fusion 1" << std::endl 
        << result.first << std::endl
        << "Fusion 2" << std::endl
        << result.second << std::endl
        << "What happend to FTensor" << std::endl
        << FTensor << std::endl;

    std::array<size_t, 2> ax1 = {0, 1}, ax2= {1, 2}, ax3 = {2, 3};
    std::cout << "The tensor" << std::endl;
//  auto result1 = gc::fusion(FTensor, ax1, FTensor2, ax2); Error
//  After fusion, the left tail dimension and right leading dimension has to be the same.
    auto result1 = gc::fusion(FTensor1,ax1,FTensor2, ax1); // Passed
    std::cout << "Fusion 1" << std::endl
        << result1.first << std::endl
        << "Fusion 2"  << std::endl
        << result1.second << std::endl; 
} 

int main() 
{
    t_fusion();
    return 0;
}