大型稀疏矩阵特征值怎么求？-史上最强开源大型稀疏矩阵特征值求解系统Arpack

/* * This example demonstrates the use of C++ bindings to call arpack. * * Use arpack as you would have normally done, but, use [ae]upd instead * of *[ae]upd_. The main advantage is that compiler checks the argument types * and the correct function is called based on the type (float vs double vs * complex). Note: to debug arpack, call debug_c. This is a test program to * solve for the 9 eigenvalues of A*x = lambda*x where A is the diagonal * matrix with entries 1000, 999, ... , 2, 1 on the diagonal. */#include <cmath>#include <iostream>#include <vector>#include "arpack.hpp"#include "debug_c.hpp"  // debug arpack.#include "stat_c.hpp"   // arpack statistics.template <typename Real>void diagonal_matrix_vector_product(const Real* x, Real* y) {    for (int i = 0; i < 1000; ++i) {        y[i] = static_cast<Real>(i + 1) * x[i];    }}template <typename Real>void real_symmetric_runner(double const& tol_check, arpack::which const& ritz_option) {    const a_int N = 1000;    const a_int nev = 9;    const a_int ncv = 2 * nev + 1;    const a_int ldv = N;    const a_int ldz = N;    const a_int lworkl = ncv * (ncv + 8);    const a_int rvec = 1;      // need eigenvectors    const Real tol = 0.000001; // small tol => more stable checks after EV computation.    const Real sigma = 0.0;      // not referenced in this mode    std::vector<Real> resid(N);    std::vector<Real> V(ldv * ncv);    std::vector<Real> z(ldz * nev);    std::vector<Real> d(nev);    std::vector<Real> workd(3 * N);    std::vector<Real> workl(lworkl);    std::vector<a_int> select(ncv); // since HOWMNY = 'A', only used as workspace here    a_int iparam[11], ipntr[11];    iparam[0] = 1;      // ishift    iparam[2] = 10 * N; // on input: maxit; on output: actual iteration    iparam[3] = 1;      // NB, only 1 allowed    iparam[6] = 1;      // mode    a_int info = 0, ido = 0;    do {        arpack::saupd(ido, arpack::bmat::identity, N,            ritz_option, nev, tol, resid.data(), ncv,            V.data(), ldv, iparam, ipntr, workd.data(),            workl.data(), lworkl, info);        diagonal_matrix_vector_product(&(workd[ipntr[0] - 1]), &(workd[ipntr[1] - 1]));    } while (ido == 1 || ido == -1);    // check info and number of ev found by arpack.    if (info < 0 || iparam[4] < nev) { /*arpack may succeed to compute more EV than expected*/        std::cout << "ERROR in saupd: iparam[4] " << iparam[4] << ", nev " << nev            << ", info " << info << std::endl;        throw std::domain_error("Error inside ARPACK routines");    }    arpack::seupd(rvec, arpack::howmny::ritz_vectors, select.data(), d.data(),        z.data(), ldz, sigma, arpack::bmat::identity, N,        ritz_option, nev, tol, resid.data(), ncv,        V.data(), ldv, iparam, ipntr, workd.data(),        workl.data(), lworkl, info);    if (info < 0) throw std::runtime_error("Error in seupd, info " + std::to_string(info));    for (int i = 0; i < nev; ++i) {        Real val = d[i];        Real ref = static_cast<Real>(N - (nev - 1) + i);        Real eps = std::fabs(val - ref);        std::cout << val << " - " << ref << " - " << eps << std::endl;        /*eigen value order: smallest -> biggest*/        if (eps > tol_check) throw std::domain_error("Correct eigenvalues not computed");    }    std::cout << "------" << std::endl;}template <typename Real>void diagonal_matrix_vector_product(const std::complex<Real>* x, std::complex<Real>* y) {    for (int i = 0; i < 1000; ++i) {        // Use complex matrix (i, -i) instead of (i, i): this way "largest_magnitude"        // and "largest_imaginary" options produce different results that can be checked.        y[i] = x[i] * std::complex<Real>{Real(i + 1), -Real(i + 1)};    }}template <typename Real>void complex_nonsymmetric_runner(double const& tol_check, arpack::which const& ritz_option) {    const a_int N = 1000;    const a_int nev = 9;    const a_int ncv = 2 * nev + 1;    const a_int ldv = N;    const a_int ldz = N;    const a_int lworkl = ncv * (3 * ncv + 5);    const a_int rvec = 0;                   // eigenvectors omitted    const Real tol = 0.000001;                // small tol => more stable checks after EV computation.    const std::complex<Real> sigma(0.0, 0.0); // not referenced in this mode    std::vector<std::complex<Real>> resid(N);    std::vector<std::complex<Real>> V(ldv * ncv);    std::vector<std::complex<Real>> z(ldz * nev);    std::vector<std::complex<Real>> d(nev);    std::vector<std::complex<Real>> workd(3 * N);    std::vector<std::complex<Real>> workl(lworkl);    std::vector<std::complex<Real>> workev(2 * ncv);    std::vector<Real> rwork(ncv);    std::vector<a_int> select(ncv); // since HOWMNY = 'A', only used as workspace here    a_int iparam[11], ipntr[14];    iparam[0] = 1;       // ishift    iparam[2] = 10 * N;  // on input: maxit; on output: actual iteration    iparam[3] = 1;       // NB, only 1 allowed    iparam[6] = 1;       // mode    a_int info = 0, ido = 0;    do {        arpack::naupd(ido, arpack::bmat::identity, N,            ritz_option, nev, tol, resid.data(), ncv,            V.data(), ldv, iparam, ipntr, workd.data(),            workl.data(), lworkl, rwork.data(), info);        diagonal_matrix_vector_product(&(workd[ipntr[0] - 1]), &(workd[ipntr[1] - 1]));    } while (ido == 1 || ido == -1);    // check info and number of ev found by arpack.    if (info < 0 || iparam[4] < nev) { /*arpack may succeed to compute more EV than expected*/        std::cout << "ERROR in naupd: iparam[4] " << iparam[4] << ", nev " << nev            << ", info " << info << std::endl;        throw std::domain_error("Error inside ARPACK routines");    }    arpack::neupd(rvec, arpack::howmny::ritz_vectors, select.data(), d.data(),        z.data(), ldz, sigma, workev.data(), arpack::bmat::identity, N,        ritz_option, nev, tol, resid.data(), ncv,        V.data(), ldv, iparam, ipntr, workd.data(),        workl.data(), lworkl, rwork.data(), info);    if (info < 0) throw std::runtime_error("Error in neupd, info " + std::to_string(info));    if (ritz_option == arpack::which::largest_magnitude) {        for (int i = 0; i < nev; ++i) {            Real rval = std::real(d[i]);            Real rref = static_cast<Real>(N - (nev - 1) + i);            Real reps = std::fabs(rval - rref);            Real ival = std::imag(d[i]);            Real iref = -static_cast<Real>(N - (nev - 1) + i);            Real ieps = std::fabs(ival - iref);            std::cout << rval << " " << ival << " - " << rref << " " << iref << " - " << reps << " " << ieps << std::endl;            if (reps > tol_check || ieps > tol_check) throw std::domain_error("Correct eigenvalues not computed");        }    }    else if (ritz_option == arpack::which::largest_imaginary) {        for (int i = 0; i < nev; ++i) {            Real rval = std::real(d[i]);            Real rref = static_cast<Real>(nev - i);            Real reps = std::fabs(rval - rref);            Real ival = std::imag(d[i]);            Real iref = -static_cast<Real>(nev - i);            Real ieps = std::fabs(ival - iref);            std::cout << rval << " " << ival << " - " << rref << " " << iref << " - " << reps << " " << ieps << std::endl;            if (reps > tol_check || ieps > tol_check) throw std::domain_error("Correct eigenvalues not computed");        }    }    else {        throw std::domain_error("The input Ritz option is not allowed in this test file.");    }    std::cout << "------" << std::endl;}int main() {    sstats_c();    // arpack without debug    real_symmetric_runner<float>(1., arpack::which::largest_magnitude);    real_symmetric_runner<float>(1., arpack::which::largest_algebraic);    real_symmetric_runner<double>(1.e-05, arpack::which::largest_magnitude);    real_symmetric_runner<double>(1.e-05, arpack::which::largest_algebraic);    a_int nopx_c, nbx_c, nrorth_c, nitref_c, nrstrt_c;    float tsaupd_c, tsaup2_c, tsaitr_c, tseigt_c, tsgets_c, tsapps_c, tsconv_c;    float tnaupd_c, tnaup2_c, tnaitr_c, tneigt_c, tngets_c, tnapps_c, tnconv_c;    float tcaupd_c, tcaup2_c, tcaitr_c, tceigt_c, tcgets_c, tcapps_c, tcconv_c;    float tmvopx_c, tmvbx_c, tgetv0_c, titref_c, trvec_c;    stat_c(nopx_c, nbx_c, nrorth_c, nitref_c, nrstrt_c, tsaupd_c, tsaup2_c,        tsaitr_c, tseigt_c, tsgets_c, tsapps_c, tsconv_c, tnaupd_c, tnaup2_c,        tnaitr_c, tneigt_c, tngets_c, tnapps_c, tnconv_c, tcaupd_c, tcaup2_c,        tcaitr_c, tceigt_c, tcgets_c, tcapps_c, tcconv_c, tmvopx_c, tmvbx_c,        tgetv0_c, titref_c, trvec_c);    std::cout << "Timers : nopx " << nopx_c << ", tmvopx " << tmvopx_c;    std::cout << " - nbx " << nbx_c << ", tmvbx " << tmvbx_c << std::endl;    std::cout << "------" << std::endl;    // set debug flags    debug_c(6, -6, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,        1);    // arpack with debug    complex_nonsymmetric_runner<float>(1., arpack::which::largest_magnitude);    complex_nonsymmetric_runner<float>(1., arpack::which::largest_imaginary);    complex_nonsymmetric_runner<double>(1.e-05, arpack::which::largest_magnitude);    complex_nonsymmetric_runner<double>(1.e-05, arpack::which::largest_imaginary);    return 0;}