Mercurial > octave
diff liboctave/numeric/oct-norm.cc @ 31607:aac27ad79be6 stable
maint: Re-indent code after switch to using namespace macros.
* build-env.h, build-env.in.cc, Cell.h, __betainc__.cc, __eigs__.cc,
__ftp__.cc, __ichol__.cc, __ilu__.cc, __isprimelarge__.cc, __magick_read__.cc,
__pchip_deriv__.cc, amd.cc, base-text-renderer.cc, base-text-renderer.h,
besselj.cc, bitfcns.cc, bsxfun.cc, c-file-ptr-stream.h, call-stack.cc,
call-stack.h, ccolamd.cc, cellfun.cc, chol.cc, colamd.cc, dasrt.cc, data.cc,
debug.cc, defaults.cc, defaults.h, det.cc, display.cc, display.h, dlmread.cc,
dynamic-ld.cc, dynamic-ld.h, ellipj.cc, environment.cc, environment.h,
error.cc, error.h, errwarn.h, event-manager.cc, event-manager.h,
event-queue.cc, event-queue.h, fcn-info.cc, fcn-info.h, fft.cc, fft2.cc,
file-io.cc, filter.cc, find.cc, ft-text-renderer.cc, ft-text-renderer.h,
gcd.cc, gl-render.cc, gl-render.h, gl2ps-print.cc, gl2ps-print.h,
graphics-toolkit.cc, graphics-toolkit.h, graphics.cc, gsvd.cc, gtk-manager.cc,
gtk-manager.h, help.cc, help.h, hook-fcn.cc, hook-fcn.h, input.cc, input.h,
interpreter-private.cc, interpreter-private.h, interpreter.cc, interpreter.h,
inv.cc, jsondecode.cc, jsonencode.cc, latex-text-renderer.cc,
latex-text-renderer.h, load-path.cc, load-path.h, load-save.cc, load-save.h,
lookup.cc, ls-hdf5.cc, ls-mat4.cc, ls-mat5.cc, lsode.cc, lu.cc, mappers.cc,
matrix_type.cc, max.cc, mex.cc, mexproto.h, mxarray.h, mxtypes.in.h,
oct-errno.in.cc, oct-hdf5-types.cc, oct-hist.cc, oct-hist.h, oct-map.cc,
oct-map.h, oct-opengl.h, oct-prcstrm.h, oct-process.cc, oct-process.h,
oct-stdstrm.h, oct-stream.cc, oct-stream.h, oct-strstrm.h,
octave-default-image.h, ordqz.cc, ordschur.cc, pager.cc, pager.h, pinv.cc,
pow2.cc, pr-output.cc, psi.cc, qr.cc, quadcc.cc, rand.cc, regexp.cc,
settings.cc, settings.h, sighandlers.cc, sighandlers.h, sparse-xpow.cc,
sqrtm.cc, stack-frame.cc, stack-frame.h, stream-euler.cc, strfns.cc, svd.cc,
syminfo.cc, syminfo.h, symrcm.cc, symrec.cc, symrec.h, symscope.cc, symscope.h,
symtab.cc, symtab.h, sysdep.cc, sysdep.h, text-engine.cc, text-engine.h,
text-renderer.cc, text-renderer.h, time.cc, toplev.cc, typecast.cc,
url-handle-manager.cc, url-handle-manager.h, urlwrite.cc, utils.cc, utils.h,
variables.cc, variables.h, xdiv.cc, __delaunayn__.cc, __init_fltk__.cc,
__init_gnuplot__.cc, __ode15__.cc, __voronoi__.cc, audioread.cc, convhulln.cc,
gzip.cc, cdef-class.cc, cdef-class.h, cdef-fwd.h, cdef-manager.cc,
cdef-manager.h, cdef-method.cc, cdef-method.h, cdef-object.cc, cdef-object.h,
cdef-package.cc, cdef-package.h, cdef-property.cc, cdef-property.h,
cdef-utils.cc, cdef-utils.h, ov-base-diag.cc, ov-base-int.cc, ov-base-mat.cc,
ov-base-mat.h, ov-base-scalar.cc, ov-base.cc, ov-base.h, ov-bool-mat.cc,
ov-bool-mat.h, ov-bool-sparse.cc, ov-bool.cc, ov-builtin.h, ov-cell.cc,
ov-ch-mat.cc, ov-class.cc, ov-class.h, ov-classdef.cc, ov-classdef.h,
ov-complex.cc, ov-cx-diag.cc, ov-cx-mat.cc, ov-cx-sparse.cc, ov-dld-fcn.cc,
ov-dld-fcn.h, ov-fcn-handle.cc, ov-fcn-handle.h, ov-fcn.h, ov-float.cc,
ov-flt-complex.cc, ov-flt-cx-diag.cc, ov-flt-cx-mat.cc, ov-flt-re-diag.cc,
ov-flt-re-mat.cc, ov-flt-re-mat.h, ov-intx.h, ov-java.cc, ov-lazy-idx.cc,
ov-legacy-range.cc, ov-magic-int.cc, ov-mex-fcn.cc, ov-mex-fcn.h,
ov-null-mat.cc, ov-perm.cc, ov-range.cc, ov-re-diag.cc, ov-re-mat.cc,
ov-re-mat.h, ov-re-sparse.cc, ov-scalar.cc, ov-str-mat.cc, ov-struct.cc,
ov-typeinfo.cc, ov-typeinfo.h, ov-usr-fcn.cc, ov-usr-fcn.h, ov.cc, ov.h, ovl.h,
octave.cc, octave.h, op-b-sbm.cc, op-bm-sbm.cc, op-cs-scm.cc, op-fm-fcm.cc,
op-fs-fcm.cc, op-s-scm.cc, op-scm-cs.cc, op-scm-s.cc, op-sm-cs.cc, ops.h,
anon-fcn-validator.cc, anon-fcn-validator.h, bp-table.cc, bp-table.h,
comment-list.cc, comment-list.h, filepos.h, lex.h, oct-lvalue.cc, oct-lvalue.h,
parse.h, profiler.cc, profiler.h, pt-anon-scopes.cc, pt-anon-scopes.h,
pt-arg-list.cc, pt-arg-list.h, pt-args-block.cc, pt-args-block.h,
pt-array-list.cc, pt-array-list.h, pt-assign.cc, pt-assign.h, pt-binop.cc,
pt-binop.h, pt-bp.cc, pt-bp.h, pt-cbinop.cc, pt-cbinop.h, pt-cell.cc,
pt-cell.h, pt-check.cc, pt-check.h, pt-classdef.cc, pt-classdef.h, pt-cmd.h,
pt-colon.cc, pt-colon.h, pt-const.cc, pt-const.h, pt-decl.cc, pt-decl.h,
pt-eval.cc, pt-eval.h, pt-except.cc, pt-except.h, pt-exp.cc, pt-exp.h,
pt-fcn-handle.cc, pt-fcn-handle.h, pt-id.cc, pt-id.h, pt-idx.cc, pt-idx.h,
pt-jump.h, pt-loop.cc, pt-loop.h, pt-mat.cc, pt-mat.h, pt-misc.cc, pt-misc.h,
pt-pr-code.cc, pt-pr-code.h, pt-select.cc, pt-select.h, pt-spmd.cc, pt-spmd.h,
pt-stmt.cc, pt-stmt.h, pt-tm-const.cc, pt-tm-const.h, pt-unop.cc, pt-unop.h,
pt-walk.cc, pt-walk.h, pt.cc, pt.h, token.cc, token.h, Range.cc, Range.h,
idx-vector.cc, idx-vector.h, range-fwd.h, CollocWt.cc, CollocWt.h,
aepbalance.cc, aepbalance.h, chol.cc, chol.h, gepbalance.cc, gepbalance.h,
gsvd.cc, gsvd.h, hess.cc, hess.h, lo-mappers.cc, lo-mappers.h, lo-specfun.cc,
lo-specfun.h, lu.cc, lu.h, oct-convn.cc, oct-convn.h, oct-fftw.cc, oct-fftw.h,
oct-norm.cc, oct-norm.h, oct-rand.cc, oct-rand.h, oct-spparms.cc,
oct-spparms.h, qr.cc, qr.h, qrp.cc, qrp.h, randgamma.cc, randgamma.h,
randmtzig.cc, randmtzig.h, randpoisson.cc, randpoisson.h, schur.cc, schur.h,
sparse-chol.cc, sparse-chol.h, sparse-lu.cc, sparse-lu.h, sparse-qr.cc,
sparse-qr.h, svd.cc, svd.h, child-list.cc, child-list.h, dir-ops.cc, dir-ops.h,
file-ops.cc, file-ops.h, file-stat.cc, file-stat.h, lo-sysdep.cc, lo-sysdep.h,
lo-sysinfo.cc, lo-sysinfo.h, mach-info.cc, mach-info.h, oct-env.cc, oct-env.h,
oct-group.cc, oct-group.h, oct-password.cc, oct-password.h, oct-syscalls.cc,
oct-syscalls.h, oct-time.cc, oct-time.h, oct-uname.cc, oct-uname.h,
action-container.cc, action-container.h, base-list.h, cmd-edit.cc, cmd-edit.h,
cmd-hist.cc, cmd-hist.h, f77-fcn.h, file-info.cc, file-info.h,
lo-array-errwarn.cc, lo-array-errwarn.h, lo-hash.cc, lo-hash.h, lo-ieee.h,
lo-regexp.cc, lo-regexp.h, lo-utils.cc, lo-utils.h, oct-base64.cc,
oct-base64.h, oct-glob.cc, oct-glob.h, oct-inttypes.h, oct-mutex.cc,
oct-mutex.h, oct-refcount.h, oct-shlib.cc, oct-shlib.h, oct-sparse.cc,
oct-sparse.h, oct-string.h, octave-preserve-stream-state.h, pathsearch.cc,
pathsearch.h, quit.cc, quit.h, unwind-prot.cc, unwind-prot.h, url-transfer.cc,
url-transfer.h:
Re-indent code after switch to using namespace macros.
author | Rik <rik@octave.org> |
---|---|
date | Thu, 01 Dec 2022 18:02:15 -0800 |
parents | e88a07dec498 |
children | 23664317f0d3 597f3ee61a48 |
line wrap: on
line diff
--- a/liboctave/numeric/oct-norm.cc Thu Dec 01 14:23:45 2022 -0800 +++ b/liboctave/numeric/oct-norm.cc Thu Dec 01 18:02:15 2022 -0800 @@ -65,261 +65,261 @@ OCTAVE_BEGIN_NAMESPACE(octave) - // Theory: norm accumulator is an object that has an accum method able - // to handle both real and complex element, and a cast operator - // returning the intermediate norm. Reference: Higham, N. "Estimating - // the Matrix p-Norm." Numer. Math. 62, 539-555, 1992. +// Theory: norm accumulator is an object that has an accum method able +// to handle both real and complex element, and a cast operator +// returning the intermediate norm. Reference: Higham, N. "Estimating +// the Matrix p-Norm." Numer. Math. 62, 539-555, 1992. - // norm accumulator for the p-norm - template <typename R> - class norm_accumulator_p +// norm accumulator for the p-norm +template <typename R> +class norm_accumulator_p +{ +public: + norm_accumulator_p () { } // we need this one for Array + norm_accumulator_p (R pp) : m_p(pp), m_scl(0), m_sum(1) { } + + template <typename U> + void accum (U val) { - public: - norm_accumulator_p () { } // we need this one for Array - norm_accumulator_p (R pp) : m_p(pp), m_scl(0), m_sum(1) { } - - template <typename U> - void accum (U val) - { - octave_quit (); - R t = std::abs (val); - if (m_scl == t) // we need this to handle Infs properly + octave_quit (); + R t = std::abs (val); + if (m_scl == t) // we need this to handle Infs properly + m_sum += 1; + else if (m_scl < t) + { + m_sum *= std::pow (m_scl/t, m_p); m_sum += 1; - else if (m_scl < t) - { - m_sum *= std::pow (m_scl/t, m_p); - m_sum += 1; - m_scl = t; - } - else if (t != 0) - m_sum += std::pow (t/m_scl, m_p); - } + m_scl = t; + } + else if (t != 0) + m_sum += std::pow (t/m_scl, m_p); + } + + operator R () { return m_scl * std::pow (m_sum, 1/m_p); } - operator R () { return m_scl * std::pow (m_sum, 1/m_p); } +private: + R m_p, m_scl, m_sum; +}; + +// norm accumulator for the minus p-pseudonorm +template <typename R> +class norm_accumulator_mp +{ +public: + norm_accumulator_mp () { } // we need this one for Array + norm_accumulator_mp (R pp) : m_p(pp), m_scl(0), m_sum(1) { } - private: - R m_p, m_scl, m_sum; - }; - - // norm accumulator for the minus p-pseudonorm - template <typename R> - class norm_accumulator_mp + template <typename U> + void accum (U val) { - public: - norm_accumulator_mp () { } // we need this one for Array - norm_accumulator_mp (R pp) : m_p(pp), m_scl(0), m_sum(1) { } + octave_quit (); + R t = 1 / std::abs (val); + if (m_scl == t) + m_sum += 1; + else if (m_scl < t) + { + m_sum *= std::pow (m_scl/t, m_p); + m_sum += 1; + m_scl = t; + } + else if (t != 0) + m_sum += std::pow (t/m_scl, m_p); + } + + operator R () { return m_scl * std::pow (m_sum, -1/m_p); } - template <typename U> - void accum (U val) - { - octave_quit (); - R t = 1 / std::abs (val); - if (m_scl == t) +private: + R m_p, m_scl, m_sum; +}; + +// norm accumulator for the 2-norm (euclidean) +template <typename R> +class norm_accumulator_2 +{ +public: + norm_accumulator_2 () : m_scl(0), m_sum(1) { } + + void accum (R val) + { + R t = std::abs (val); + if (m_scl == t) + m_sum += 1; + else if (m_scl < t) + { + m_sum *= pow2 (m_scl/t); m_sum += 1; - else if (m_scl < t) - { - m_sum *= std::pow (m_scl/t, m_p); - m_sum += 1; - m_scl = t; - } - else if (t != 0) - m_sum += std::pow (t/m_scl, m_p); - } + m_scl = t; + } + else if (t != 0) + m_sum += pow2 (t/m_scl); + } + + void accum (std::complex<R> val) + { + accum (val.real ()); + accum (val.imag ()); + } + + operator R () { return m_scl * std::sqrt (m_sum); } + +private: + static inline R pow2 (R x) { return x*x; } + + //-------- + + R m_scl, m_sum; +}; - operator R () { return m_scl * std::pow (m_sum, -1/m_p); } +// norm accumulator for the 1-norm (city metric) +template <typename R> +class norm_accumulator_1 +{ +public: + norm_accumulator_1 () : m_sum (0) { } + template <typename U> + void accum (U val) + { + m_sum += std::abs (val); + } + + operator R () { return m_sum; } + +private: + R m_sum; +}; + +// norm accumulator for the inf-norm (max metric) +template <typename R> +class norm_accumulator_inf +{ +public: + norm_accumulator_inf () : m_max (0) { } + template <typename U> + void accum (U val) + { + if (math::isnan (val)) + m_max = numeric_limits<R>::NaN (); + else + m_max = std::max (m_max, std::abs (val)); + } + + operator R () { return m_max; } - private: - R m_p, m_scl, m_sum; - }; +private: + R m_max; +}; + +// norm accumulator for the -inf pseudonorm (min abs value) +template <typename R> +class norm_accumulator_minf +{ +public: + norm_accumulator_minf () : m_min (numeric_limits<R>::Inf ()) { } + template <typename U> + void accum (U val) + { + if (math::isnan (val)) + m_min = numeric_limits<R>::NaN (); + else + m_min = std::min (m_min, std::abs (val)); + } + + operator R () { return m_min; } + +private: + R m_min; +}; - // norm accumulator for the 2-norm (euclidean) - template <typename R> - class norm_accumulator_2 +// norm accumulator for the 0-pseudonorm (hamming distance) +template <typename R> +class norm_accumulator_0 +{ +public: + norm_accumulator_0 () : m_num (0) { } + template <typename U> + void accum (U val) { - public: - norm_accumulator_2 () : m_scl(0), m_sum(1) { } + if (val != static_cast<U> (0)) ++m_num; + } + + operator R () { return m_num; } + +private: + unsigned int m_num; +}; + +// OK, we're armed :) Now let's go for the fun - void accum (R val) +template <typename T, typename R, typename ACC> +inline void vector_norm (const Array<T>& v, R& res, ACC acc) +{ + for (octave_idx_type i = 0; i < v.numel (); i++) + acc.accum (v(i)); + + res = acc; +} + +// dense versions +template <typename T, typename R, typename ACC> +void column_norms (const MArray<T>& m, MArray<R>& res, ACC acc) +{ + res = MArray<R> (dim_vector (1, m.columns ())); + for (octave_idx_type j = 0; j < m.columns (); j++) { - R t = std::abs (val); - if (m_scl == t) - m_sum += 1; - else if (m_scl < t) - { - m_sum *= pow2 (m_scl/t); - m_sum += 1; - m_scl = t; - } - else if (t != 0) - m_sum += pow2 (t/m_scl); + ACC accj = acc; + for (octave_idx_type i = 0; i < m.rows (); i++) + accj.accum (m(i, j)); + + res.xelem (j) = accj; } - - void accum (std::complex<R> val) - { - accum (val.real ()); - accum (val.imag ()); - } +} - operator R () { return m_scl * std::sqrt (m_sum); } - - private: - static inline R pow2 (R x) { return x*x; } - - //-------- - - R m_scl, m_sum; - }; - - // norm accumulator for the 1-norm (city metric) - template <typename R> - class norm_accumulator_1 - { - public: - norm_accumulator_1 () : m_sum (0) { } - template <typename U> - void accum (U val) +template <typename T, typename R, typename ACC> +void row_norms (const MArray<T>& m, MArray<R>& res, ACC acc) +{ + res = MArray<R> (dim_vector (m.rows (), 1)); + std::vector<ACC> acci (m.rows (), acc); + for (octave_idx_type j = 0; j < m.columns (); j++) { - m_sum += std::abs (val); + for (octave_idx_type i = 0; i < m.rows (); i++) + acci[i].accum (m(i, j)); } - operator R () { return m_sum; } - - private: - R m_sum; - }; - - // norm accumulator for the inf-norm (max metric) - template <typename R> - class norm_accumulator_inf - { - public: - norm_accumulator_inf () : m_max (0) { } - template <typename U> - void accum (U val) - { - if (math::isnan (val)) - m_max = numeric_limits<R>::NaN (); - else - m_max = std::max (m_max, std::abs (val)); - } - - operator R () { return m_max; } - - private: - R m_max; - }; + for (octave_idx_type i = 0; i < m.rows (); i++) + res.xelem (i) = acci[i]; +} - // norm accumulator for the -inf pseudonorm (min abs value) - template <typename R> - class norm_accumulator_minf - { - public: - norm_accumulator_minf () : m_min (numeric_limits<R>::Inf ()) { } - template <typename U> - void accum (U val) +// sparse versions +template <typename T, typename R, typename ACC> +void column_norms (const MSparse<T>& m, MArray<R>& res, ACC acc) +{ + res = MArray<R> (dim_vector (1, m.columns ())); + for (octave_idx_type j = 0; j < m.columns (); j++) { - if (math::isnan (val)) - m_min = numeric_limits<R>::NaN (); - else - m_min = std::min (m_min, std::abs (val)); - } + ACC accj = acc; + for (octave_idx_type k = m.cidx (j); k < m.cidx (j+1); k++) + accj.accum (m.data (k)); - operator R () { return m_min; } - - private: - R m_min; - }; + res.xelem (j) = accj; + } +} - // norm accumulator for the 0-pseudonorm (hamming distance) - template <typename R> - class norm_accumulator_0 - { - public: - norm_accumulator_0 () : m_num (0) { } - template <typename U> - void accum (U val) +template <typename T, typename R, typename ACC> +void row_norms (const MSparse<T>& m, MArray<R>& res, ACC acc) +{ + res = MArray<R> (dim_vector (m.rows (), 1)); + std::vector<ACC> acci (m.rows (), acc); + for (octave_idx_type j = 0; j < m.columns (); j++) { - if (val != static_cast<U> (0)) ++m_num; + for (octave_idx_type k = m.cidx (j); k < m.cidx (j+1); k++) + acci[m.ridx (k)].accum (m.data (k)); } - operator R () { return m_num; } - - private: - unsigned int m_num; - }; - - // OK, we're armed :) Now let's go for the fun - - template <typename T, typename R, typename ACC> - inline void vector_norm (const Array<T>& v, R& res, ACC acc) - { - for (octave_idx_type i = 0; i < v.numel (); i++) - acc.accum (v(i)); - - res = acc; - } - - // dense versions - template <typename T, typename R, typename ACC> - void column_norms (const MArray<T>& m, MArray<R>& res, ACC acc) - { - res = MArray<R> (dim_vector (1, m.columns ())); - for (octave_idx_type j = 0; j < m.columns (); j++) - { - ACC accj = acc; - for (octave_idx_type i = 0; i < m.rows (); i++) - accj.accum (m(i, j)); - - res.xelem (j) = accj; - } - } + for (octave_idx_type i = 0; i < m.rows (); i++) + res.xelem (i) = acci[i]; +} - template <typename T, typename R, typename ACC> - void row_norms (const MArray<T>& m, MArray<R>& res, ACC acc) - { - res = MArray<R> (dim_vector (m.rows (), 1)); - std::vector<ACC> acci (m.rows (), acc); - for (octave_idx_type j = 0; j < m.columns (); j++) - { - for (octave_idx_type i = 0; i < m.rows (); i++) - acci[i].accum (m(i, j)); - } - - for (octave_idx_type i = 0; i < m.rows (); i++) - res.xelem (i) = acci[i]; - } - - // sparse versions - template <typename T, typename R, typename ACC> - void column_norms (const MSparse<T>& m, MArray<R>& res, ACC acc) - { - res = MArray<R> (dim_vector (1, m.columns ())); - for (octave_idx_type j = 0; j < m.columns (); j++) - { - ACC accj = acc; - for (octave_idx_type k = m.cidx (j); k < m.cidx (j+1); k++) - accj.accum (m.data (k)); - - res.xelem (j) = accj; - } - } - - template <typename T, typename R, typename ACC> - void row_norms (const MSparse<T>& m, MArray<R>& res, ACC acc) - { - res = MArray<R> (dim_vector (m.rows (), 1)); - std::vector<ACC> acci (m.rows (), acc); - for (octave_idx_type j = 0; j < m.columns (); j++) - { - for (octave_idx_type k = m.cidx (j); k < m.cidx (j+1); k++) - acci[m.ridx (k)].accum (m.data (k)); - } - - for (octave_idx_type i = 0; i < m.rows (); i++) - res.xelem (i) = acci[i]; - } - - // now the dispatchers +// now the dispatchers #define DEFINE_DISPATCHER(FCN_NAME, ARG_TYPE, RES_TYPE) \ template <typename T, typename R> \ RES_TYPE FCN_NAME (const ARG_TYPE& v, R p) \ @@ -345,224 +345,224 @@ return res; \ } - DEFINE_DISPATCHER (vector_norm, MArray<T>, R) - DEFINE_DISPATCHER (column_norms, MArray<T>, MArray<R>) - DEFINE_DISPATCHER (row_norms, MArray<T>, MArray<R>) - DEFINE_DISPATCHER (column_norms, MSparse<T>, MArray<R>) - DEFINE_DISPATCHER (row_norms, MSparse<T>, MArray<R>) +DEFINE_DISPATCHER (vector_norm, MArray<T>, R) +DEFINE_DISPATCHER (column_norms, MArray<T>, MArray<R>) +DEFINE_DISPATCHER (row_norms, MArray<T>, MArray<R>) +DEFINE_DISPATCHER (column_norms, MSparse<T>, MArray<R>) +DEFINE_DISPATCHER (row_norms, MSparse<T>, MArray<R>) - // The approximate subproblem in Higham's method. Find lambda and mu such - // that norm ([lambda, mu], p) == 1 and norm (y*lambda + col*mu, p) is - // maximized. - // Real version. As in Higham's paper. - template <typename ColVectorT, typename R> - static void - higham_subp (const ColVectorT& y, const ColVectorT& col, - octave_idx_type nsamp, R p, R& lambda, R& mu) - { - R nrm = 0; - for (octave_idx_type i = 0; i < nsamp; i++) - { - octave_quit (); - R fi = i * static_cast<R> (M_PI) / nsamp; - R lambda1 = cos (fi); - R mu1 = sin (fi); - R lmnr = std::pow (std::pow (std::abs (lambda1), p) + - std::pow (std::abs (mu1), p), 1/p); - lambda1 /= lmnr; mu1 /= lmnr; - R nrm1 = vector_norm (lambda1 * y + mu1 * col, p); - if (nrm1 > nrm) - { - lambda = lambda1; - mu = mu1; - nrm = nrm1; - } - } - } +// The approximate subproblem in Higham's method. Find lambda and mu such +// that norm ([lambda, mu], p) == 1 and norm (y*lambda + col*mu, p) is +// maximized. +// Real version. As in Higham's paper. +template <typename ColVectorT, typename R> +static void +higham_subp (const ColVectorT& y, const ColVectorT& col, + octave_idx_type nsamp, R p, R& lambda, R& mu) +{ + R nrm = 0; + for (octave_idx_type i = 0; i < nsamp; i++) + { + octave_quit (); + R fi = i * static_cast<R> (M_PI) / nsamp; + R lambda1 = cos (fi); + R mu1 = sin (fi); + R lmnr = std::pow (std::pow (std::abs (lambda1), p) + + std::pow (std::abs (mu1), p), 1/p); + lambda1 /= lmnr; mu1 /= lmnr; + R nrm1 = vector_norm (lambda1 * y + mu1 * col, p); + if (nrm1 > nrm) + { + lambda = lambda1; + mu = mu1; + nrm = nrm1; + } + } +} - // Complex version. Higham's paper does not deal with complex case, so we - // use a simple extension. First, guess the magnitudes as in real version, - // then try to rotate lambda to improve further. - template <typename ColVectorT, typename R> - static void - higham_subp (const ColVectorT& y, const ColVectorT& col, - octave_idx_type nsamp, R p, - std::complex<R>& lambda, std::complex<R>& mu) - { - typedef std::complex<R> CR; - R nrm = 0; - lambda = 1.0; - CR lamcu = lambda / std::abs (lambda); - // Probe magnitudes - for (octave_idx_type i = 0; i < nsamp; i++) - { - octave_quit (); - R fi = i * static_cast<R> (M_PI) / nsamp; - R lambda1 = cos (fi); - R mu1 = sin (fi); - R lmnr = std::pow (std::pow (std::abs (lambda1), p) + - std::pow (std::abs (mu1), p), 1/p); - lambda1 /= lmnr; mu1 /= lmnr; - R nrm1 = vector_norm (lambda1 * lamcu * y + mu1 * col, p); - if (nrm1 > nrm) - { - lambda = lambda1 * lamcu; - mu = mu1; - nrm = nrm1; - } - } - R lama = std::abs (lambda); - // Probe orientation - for (octave_idx_type i = 0; i < nsamp; i++) - { - octave_quit (); - R fi = i * static_cast<R> (M_PI) / nsamp; - lamcu = CR (cos (fi), sin (fi)); - R nrm1 = vector_norm (lama * lamcu * y + mu * col, p); - if (nrm1 > nrm) - { - lambda = lama * lamcu; - nrm = nrm1; - } - } - } +// Complex version. Higham's paper does not deal with complex case, so we +// use a simple extension. First, guess the magnitudes as in real version, +// then try to rotate lambda to improve further. +template <typename ColVectorT, typename R> +static void +higham_subp (const ColVectorT& y, const ColVectorT& col, + octave_idx_type nsamp, R p, + std::complex<R>& lambda, std::complex<R>& mu) +{ + typedef std::complex<R> CR; + R nrm = 0; + lambda = 1.0; + CR lamcu = lambda / std::abs (lambda); + // Probe magnitudes + for (octave_idx_type i = 0; i < nsamp; i++) + { + octave_quit (); + R fi = i * static_cast<R> (M_PI) / nsamp; + R lambda1 = cos (fi); + R mu1 = sin (fi); + R lmnr = std::pow (std::pow (std::abs (lambda1), p) + + std::pow (std::abs (mu1), p), 1/p); + lambda1 /= lmnr; mu1 /= lmnr; + R nrm1 = vector_norm (lambda1 * lamcu * y + mu1 * col, p); + if (nrm1 > nrm) + { + lambda = lambda1 * lamcu; + mu = mu1; + nrm = nrm1; + } + } + R lama = std::abs (lambda); + // Probe orientation + for (octave_idx_type i = 0; i < nsamp; i++) + { + octave_quit (); + R fi = i * static_cast<R> (M_PI) / nsamp; + lamcu = CR (cos (fi), sin (fi)); + R nrm1 = vector_norm (lama * lamcu * y + mu * col, p); + if (nrm1 > nrm) + { + lambda = lama * lamcu; + nrm = nrm1; + } + } +} - // the p-dual element (should work for both real and complex) - template <typename T, typename R> - inline T elem_dual_p (T x, R p) - { - return math::signum (x) * std::pow (std::abs (x), p-1); - } +// the p-dual element (should work for both real and complex) +template <typename T, typename R> +inline T elem_dual_p (T x, R p) +{ + return math::signum (x) * std::pow (std::abs (x), p-1); +} - // the VectorT is used for vectors, but actually it has to be - // a Matrix type to allow all the operations. For instance SparseMatrix - // does not support multiplication with column/row vectors. - // the dual vector - template <typename VectorT, typename R> - VectorT dual_p (const VectorT& x, R p, R q) - { - VectorT res (x.dims ()); - for (octave_idx_type i = 0; i < x.numel (); i++) - res.xelem (i) = elem_dual_p (x(i), p); - return res / vector_norm (res, q); - } +// the VectorT is used for vectors, but actually it has to be +// a Matrix type to allow all the operations. For instance SparseMatrix +// does not support multiplication with column/row vectors. +// the dual vector +template <typename VectorT, typename R> +VectorT dual_p (const VectorT& x, R p, R q) +{ + VectorT res (x.dims ()); + for (octave_idx_type i = 0; i < x.numel (); i++) + res.xelem (i) = elem_dual_p (x(i), p); + return res / vector_norm (res, q); +} - // Higham's hybrid method - template <typename MatrixT, typename VectorT, typename R> - R higham (const MatrixT& m, R p, R tol, int maxiter, - VectorT& x) - { - x.resize (m.columns (), 1); - // the OSE part - VectorT y(m.rows (), 1, 0), z(m.rows (), 1); - typedef typename VectorT::element_type RR; - RR lambda = 0; - RR mu = 1; - for (octave_idx_type k = 0; k < m.columns (); k++) - { - octave_quit (); - VectorT col (m.column (k)); - if (k > 0) - higham_subp (y, col, 4*k, p, lambda, mu); - for (octave_idx_type i = 0; i < k; i++) - x(i) *= lambda; - x(k) = mu; - y = lambda * y + mu * col; - } +// Higham's hybrid method +template <typename MatrixT, typename VectorT, typename R> +R higham (const MatrixT& m, R p, R tol, int maxiter, + VectorT& x) +{ + x.resize (m.columns (), 1); + // the OSE part + VectorT y(m.rows (), 1, 0), z(m.rows (), 1); + typedef typename VectorT::element_type RR; + RR lambda = 0; + RR mu = 1; + for (octave_idx_type k = 0; k < m.columns (); k++) + { + octave_quit (); + VectorT col (m.column (k)); + if (k > 0) + higham_subp (y, col, 4*k, p, lambda, mu); + for (octave_idx_type i = 0; i < k; i++) + x(i) *= lambda; + x(k) = mu; + y = lambda * y + mu * col; + } - // the PM part - x = x / vector_norm (x, p); - R q = p/(p-1); + // the PM part + x = x / vector_norm (x, p); + R q = p/(p-1); - R gamma = 0, gamma1; - int iter = 0; - while (iter < maxiter) - { - octave_quit (); - y = m*x; - gamma1 = gamma; - gamma = vector_norm (y, p); - z = dual_p (y, p, q); - z = z.hermitian (); - z = z * m; + R gamma = 0, gamma1; + int iter = 0; + while (iter < maxiter) + { + octave_quit (); + y = m*x; + gamma1 = gamma; + gamma = vector_norm (y, p); + z = dual_p (y, p, q); + z = z.hermitian (); + z = z * m; - if (iter > 0 && (vector_norm (z, q) <= gamma - || (gamma - gamma1) <= tol*gamma)) - break; + if (iter > 0 && (vector_norm (z, q) <= gamma + || (gamma - gamma1) <= tol*gamma)) + break; - z = z.hermitian (); - x = dual_p (z, q, p); - iter++; - } + z = z.hermitian (); + x = dual_p (z, q, p); + iter++; + } - return gamma; - } + return gamma; +} - // derive column vector and SVD types +// derive column vector and SVD types - static const char *p_less1_gripe = "xnorm: p must be >= 1"; +static const char *p_less1_gripe = "xnorm: p must be >= 1"; - // Static constant to control the maximum number of iterations. 100 seems to - // be a good value. Eventually, we can provide a means to change this - // constant from Octave. - static int max_norm_iter = 100; +// Static constant to control the maximum number of iterations. 100 seems to +// be a good value. Eventually, we can provide a means to change this +// constant from Octave. +static int max_norm_iter = 100; - // version with SVD for dense matrices - template <typename MatrixT, typename VectorT, typename R> - R svd_matrix_norm (const MatrixT& m, R p, VectorT) - { - // NOTE: The octave:: namespace tags are needed for the following - // function calls until the deprecated inline functions are removed - // from oct-norm.h. +// version with SVD for dense matrices +template <typename MatrixT, typename VectorT, typename R> +R svd_matrix_norm (const MatrixT& m, R p, VectorT) +{ + // NOTE: The octave:: namespace tags are needed for the following + // function calls until the deprecated inline functions are removed + // from oct-norm.h. - R res = 0; - if (p == 2) - { - math::svd<MatrixT> fact (m, math::svd<MatrixT>::Type::sigma_only); - res = fact.singular_values () (0, 0); - } - else if (p == 1) - res = octave::xcolnorms (m, static_cast<R> (1)).max (); - else if (lo_ieee_isinf (p) && p > 1) - res = octave::xrownorms (m, static_cast<R> (1)).max (); - else if (p > 1) - { - VectorT x; - const R sqrteps = std::sqrt (std::numeric_limits<R>::epsilon ()); - res = higham (m, p, sqrteps, max_norm_iter, x); - } - else - (*current_liboctave_error_handler) ("%s", p_less1_gripe); + R res = 0; + if (p == 2) + { + math::svd<MatrixT> fact (m, math::svd<MatrixT>::Type::sigma_only); + res = fact.singular_values () (0, 0); + } + else if (p == 1) + res = octave::xcolnorms (m, static_cast<R> (1)).max (); + else if (lo_ieee_isinf (p) && p > 1) + res = octave::xrownorms (m, static_cast<R> (1)).max (); + else if (p > 1) + { + VectorT x; + const R sqrteps = std::sqrt (std::numeric_limits<R>::epsilon ()); + res = higham (m, p, sqrteps, max_norm_iter, x); + } + else + (*current_liboctave_error_handler) ("%s", p_less1_gripe); - return res; - } + return res; +} - // SVD-free version for sparse matrices - template <typename MatrixT, typename VectorT, typename R> - R matrix_norm (const MatrixT& m, R p, VectorT) - { - // NOTE: The octave:: namespace tags are needed for the following - // function calls until the deprecated inline functions are removed - // from oct-norm.h. +// SVD-free version for sparse matrices +template <typename MatrixT, typename VectorT, typename R> +R matrix_norm (const MatrixT& m, R p, VectorT) +{ + // NOTE: The octave:: namespace tags are needed for the following + // function calls until the deprecated inline functions are removed + // from oct-norm.h. - R res = 0; - if (p == 1) - res = octave::xcolnorms (m, static_cast<R> (1)).max (); - else if (lo_ieee_isinf (p) && p > 1) - res = octave::xrownorms (m, static_cast<R> (1)).max (); - else if (p > 1) - { - VectorT x; - const R sqrteps = std::sqrt (std::numeric_limits<R>::epsilon ()); - res = higham (m, p, sqrteps, max_norm_iter, x); - } - else - (*current_liboctave_error_handler) ("%s", p_less1_gripe); + R res = 0; + if (p == 1) + res = octave::xcolnorms (m, static_cast<R> (1)).max (); + else if (lo_ieee_isinf (p) && p > 1) + res = octave::xrownorms (m, static_cast<R> (1)).max (); + else if (p > 1) + { + VectorT x; + const R sqrteps = std::sqrt (std::numeric_limits<R>::epsilon ()); + res = higham (m, p, sqrteps, max_norm_iter, x); + } + else + (*current_liboctave_error_handler) ("%s", p_less1_gripe); - return res; - } + return res; +} - // and finally, here's what we've promised in the header file +// and finally, here's what we've promised in the header file #define DEFINE_XNORM_FCNS(PREFIX, RTYPE) \ RTYPE xnorm (const PREFIX##ColumnVector& x, RTYPE p) \ @@ -582,21 +582,21 @@ return vector_norm (x, static_cast<RTYPE> (2)); \ } - DEFINE_XNORM_FCNS(, double) - DEFINE_XNORM_FCNS(Complex, double) - DEFINE_XNORM_FCNS(Float, float) - DEFINE_XNORM_FCNS(FloatComplex, float) +DEFINE_XNORM_FCNS(, double) +DEFINE_XNORM_FCNS(Complex, double) +DEFINE_XNORM_FCNS(Float, float) +DEFINE_XNORM_FCNS(FloatComplex, float) - // this is needed to avoid copying the sparse matrix for xfrobnorm - template <typename T, typename R> - inline void array_norm_2 (const T *v, octave_idx_type n, R& res) - { - norm_accumulator_2<R> acc; - for (octave_idx_type i = 0; i < n; i++) - acc.accum (v[i]); +// this is needed to avoid copying the sparse matrix for xfrobnorm +template <typename T, typename R> +inline void array_norm_2 (const T *v, octave_idx_type n, R& res) +{ + norm_accumulator_2<R> acc; + for (octave_idx_type i = 0; i < n; i++) + acc.accum (v[i]); - res = acc; - } + res = acc; +} #define DEFINE_XNORM_SPARSE_FCNS(PREFIX, RTYPE) \ RTYPE xnorm (const Sparse##PREFIX##Matrix& x, RTYPE p) \ @@ -610,8 +610,8 @@ return res; \ } - DEFINE_XNORM_SPARSE_FCNS(, double) - DEFINE_XNORM_SPARSE_FCNS(Complex, double) +DEFINE_XNORM_SPARSE_FCNS(, double) +DEFINE_XNORM_SPARSE_FCNS(Complex, double) #define DEFINE_COLROW_NORM_FCNS(PREFIX, RPREFIX, RTYPE) \ RPREFIX##RowVector \ @@ -625,12 +625,12 @@ return row_norms (m, p); \ } \ - DEFINE_COLROW_NORM_FCNS(, , double) - DEFINE_COLROW_NORM_FCNS(Complex, , double) - DEFINE_COLROW_NORM_FCNS(Float, Float, float) - DEFINE_COLROW_NORM_FCNS(FloatComplex, Float, float) +DEFINE_COLROW_NORM_FCNS(,, double) +DEFINE_COLROW_NORM_FCNS(Complex,, double) +DEFINE_COLROW_NORM_FCNS(Float, Float, float) +DEFINE_COLROW_NORM_FCNS(FloatComplex, Float, float) - DEFINE_COLROW_NORM_FCNS(Sparse, , double) - DEFINE_COLROW_NORM_FCNS(SparseComplex, , double) +DEFINE_COLROW_NORM_FCNS(Sparse,, double) +DEFINE_COLROW_NORM_FCNS(SparseComplex,, double) OCTAVE_END_NAMESPACE(octave)