LSST Applications  21.0.0+04719a4bac,21.0.0-1-ga51b5d4+f5e6047307,21.0.0-11-g2b59f77+a9c1acf22d,21.0.0-11-ga42c5b2+86977b0b17,21.0.0-12-gf4ce030+76814010d2,21.0.0-13-g1721dae+760e7a6536,21.0.0-13-g3a573fe+768d78a30a,21.0.0-15-g5a7caf0+f21cbc5713,21.0.0-16-g0fb55c1+b60e2d390c,21.0.0-19-g4cded4ca+71a93a33c0,21.0.0-2-g103fe59+bb20972958,21.0.0-2-g45278ab+04719a4bac,21.0.0-2-g5242d73+3ad5d60fb1,21.0.0-2-g7f82c8f+8babb168e8,21.0.0-2-g8f08a60+06509c8b61,21.0.0-2-g8faa9b5+616205b9df,21.0.0-2-ga326454+8babb168e8,21.0.0-2-gde069b7+5e4aea9c2f,21.0.0-2-gecfae73+1d3a86e577,21.0.0-2-gfc62afb+3ad5d60fb1,21.0.0-25-g1d57be3cd+e73869a214,21.0.0-3-g357aad2+ed88757d29,21.0.0-3-g4a4ce7f+3ad5d60fb1,21.0.0-3-g4be5c26+3ad5d60fb1,21.0.0-3-g65f322c+e0b24896a3,21.0.0-3-g7d9da8d+616205b9df,21.0.0-3-ge02ed75+a9c1acf22d,21.0.0-4-g591bb35+a9c1acf22d,21.0.0-4-g65b4814+b60e2d390c,21.0.0-4-gccdca77+0de219a2bc,21.0.0-4-ge8a399c+6c55c39e83,21.0.0-5-gd00fb1e+05fce91b99,21.0.0-6-gc675373+3ad5d60fb1,21.0.0-64-g1122c245+4fb2b8f86e,21.0.0-7-g04766d7+cd19d05db2,21.0.0-7-gdf92d54+04719a4bac,21.0.0-8-g5674e7b+d1bd76f71f,master-gac4afde19b+a9c1acf22d,w.2021.13
LSST Data Management Base Package
LeastSquares.cc
Go to the documentation of this file.
1 // -*- LSST-C++ -*-
2 
3 /*
4  * LSST Data Management System
5  * Copyright 2008, 2009, 2010, 2011 LSST Corporation.
6  *
7  * This product includes software developed by the
8  * LSST Project (http://www.lsst.org/).
9  *
10  * This program is free software: you can redistribute it and/or modify
11  * it under the terms of the GNU General Public License as published by
12  * the Free Software Foundation, either version 3 of the License, or
13  * (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18  * GNU General Public License for more details.
19  *
20  * You should have received a copy of the LSST License Statement and
21  * the GNU General Public License along with this program. If not,
22  * see <http://www.lsstcorp.org/LegalNotices/>.
23  */
24 
25 #include "Eigen/Eigenvalues"
26 #include "Eigen/SVD"
27 #include "Eigen/Cholesky"
28 #include "boost/format.hpp"
29 #include <memory>
30 
32 #include "lsst/pex/exceptions.h"
33 #include "lsst/log/Log.h"
34 
35 namespace {
36 LOG_LOGGER _log = LOG_GET("afw.math.LeastSquares");
37 }
38 
39 namespace lsst {
40 namespace afw {
41 namespace math {
42 
44 public:
45  enum StateFlags {
48  RHS_VECTOR = 0x004,
49  SOLUTION_ARRAY = 0x008,
52  DESIGN_AND_DATA = 0x040
53  };
54 
55  int state;
56  int dimension;
57  int rank;
60  double threshold;
61 
62  Eigen::MatrixXd design;
63  Eigen::VectorXd data;
64  Eigen::MatrixXd fisher;
65  Eigen::VectorXd rhs;
66 
67  ndarray::Array<double, 1, 1> solution;
68  ndarray::Array<double, 2, 2> covariance;
69  ndarray::Array<double, 1, 1> diagnostic;
70 
71  template <typename D>
72  void setRank(Eigen::MatrixBase<D> const& values) {
73  double cond = threshold * values[0];
74  if (cond <= 0.0) {
75  rank = 0;
76  } else {
77  for (rank = dimension; (rank > 1) && (values[rank - 1] < cond); --rank)
78  ;
79  }
80  }
81 
82  void ensure(int desired) {
84  if (desired & FULL_FISHER_MATRIX) desired |= LOWER_FISHER_MATRIX;
85  int toAdd = ~state & desired;
86  if (toAdd & LOWER_FISHER_MATRIX) {
87  assert(state & DESIGN_AND_DATA);
88  fisher = Eigen::MatrixXd::Zero(design.cols(), design.cols());
89  fisher.selfadjointView<Eigen::Lower>().rankUpdate(design.adjoint());
90  }
91  if (toAdd & FULL_FISHER_MATRIX) {
92  fisher.triangularView<Eigen::StrictlyUpper>() = fisher.adjoint();
93  }
94  if (toAdd & RHS_VECTOR) {
95  assert(state & DESIGN_AND_DATA);
96  rhs = design.adjoint() * data;
97  }
98  if (toAdd & SOLUTION_ARRAY) {
99  if (solution.isEmpty()) solution = ndarray::allocate(dimension);
100  updateSolution();
101  }
102  if (toAdd & COVARIANCE_ARRAY) {
103  if (covariance.isEmpty()) covariance = ndarray::allocate(dimension, dimension);
105  }
106  if (toAdd & DIAGNOSTIC_ARRAY) {
107  if (diagnostic.isEmpty()) diagnostic = ndarray::allocate(dimension);
109  }
110  state |= toAdd;
111  }
112 
113  virtual void factor() = 0;
114 
115  virtual void updateRank() = 0;
116 
117  virtual void updateSolution() = 0;
118  virtual void updateCovariance() = 0;
119 
120  virtual void updateDiagnostic() = 0;
121 
122  Impl(int dimension_, double threshold_ = std::numeric_limits<double>::epsilon())
123  : state(0), dimension(dimension_), rank(dimension_), threshold(threshold_) {}
124 
125  virtual ~Impl() {}
126 };
127 
128 namespace {
129 
130 class EigensystemSolver : public LeastSquares::Impl {
131 public:
132  explicit EigensystemSolver(int dimension) : Impl(dimension), _eig(dimension), _svd(), _tmp(dimension) {}
133 
134  void factor() override {
135  ensure(LOWER_FISHER_MATRIX | RHS_VECTOR);
136  _eig.compute(fisher);
137  if (_eig.info() == Eigen::Success) {
138  setRank(_eig.eigenvalues().reverse());
139  LOGL_DEBUG(_log, "SelfAdjointEigenSolver succeeded: dimension=%d, rank=%d", dimension, rank);
140  } else {
141  // Note that the fallback is using SVD of the Fisher to compute the Eigensystem, because those
142  // are the same for a symmetric matrix; this is very different from doing a direct SVD of
143  // the design matrix.
144  ensure(FULL_FISHER_MATRIX);
145  _svd.compute(fisher, Eigen::ComputeFullU); // Matrix is symmetric, so V == U == eigenvectors
146  setRank(_svd.singularValues());
147  LOGL_DEBUG(_log,
148  "SelfAdjointEigenSolver failed; falling back to equivalent SVD: dimension=%d, rank=%d",
149  dimension, rank);
150  }
151  }
152 
153  void updateRank() override {
154  if (_eig.info() == Eigen::Success) {
155  setRank(_eig.eigenvalues().reverse());
156  } else {
157  setRank(_svd.singularValues());
158  }
159  }
160 
161  void updateDiagnostic() override {
162  if (whichDiagnostic == LeastSquares::NORMAL_CHOLESKY) {
163  throw LSST_EXCEPT(
164  pex::exceptions::LogicError,
165  "Cannot compute NORMAL_CHOLESKY diagnostic from NORMAL_EIGENSYSTEM factorization.");
166  }
167  if (_eig.info() == Eigen::Success) {
168  ndarray::asEigenMatrix(diagnostic) = _eig.eigenvalues().reverse();
169  } else {
170  ndarray::asEigenMatrix(diagnostic) = _svd.singularValues();
171  }
172  if (whichDiagnostic == LeastSquares::DIRECT_SVD) {
173  ndarray::asEigenArray(diagnostic) = ndarray::asEigenArray(diagnostic).sqrt();
174  }
175  }
176 
177  void updateSolution() override {
178  if (rank == 0) {
179  ndarray::asEigenMatrix(solution).setZero();
180  return;
181  }
182  if (_eig.info() == Eigen::Success) {
183  _tmp.head(rank) = _eig.eigenvectors().rightCols(rank).adjoint() * rhs;
184  _tmp.head(rank).array() /= _eig.eigenvalues().tail(rank).array();
185  ndarray::asEigenMatrix(solution) = _eig.eigenvectors().rightCols(rank) * _tmp.head(rank);
186  } else {
187  _tmp.head(rank) = _svd.matrixU().leftCols(rank).adjoint() * rhs;
188  _tmp.head(rank).array() /= _svd.singularValues().head(rank).array();
189  ndarray::asEigenMatrix(solution) = _svd.matrixU().leftCols(rank) * _tmp.head(rank);
190  }
191  }
192 
193  void updateCovariance() override {
194  if (rank == 0) {
195  ndarray::asEigenMatrix(covariance).setZero();
196  return;
197  }
198  if (_eig.info() == Eigen::Success) {
199  ndarray::asEigenMatrix(covariance) =
200  _eig.eigenvectors().rightCols(rank) *
201  _eig.eigenvalues().tail(rank).array().inverse().matrix().asDiagonal() *
202  _eig.eigenvectors().rightCols(rank).adjoint();
203  } else {
204  ndarray::asEigenMatrix(covariance) =
205  _svd.matrixU().leftCols(rank) *
206  _svd.singularValues().head(rank).array().inverse().matrix().asDiagonal() *
207  _svd.matrixU().leftCols(rank).adjoint();
208  }
209  }
210 
211 private:
212  Eigen::SelfAdjointEigenSolver<Eigen::MatrixXd> _eig;
213  Eigen::JacobiSVD<Eigen::MatrixXd> _svd; // only used if Eigendecomposition fails, should be very rare
214  Eigen::VectorXd _tmp;
215 };
216 
217 class CholeskySolver : public LeastSquares::Impl {
218 public:
219  explicit CholeskySolver(int dimension) : Impl(dimension, 0.0), _ldlt(dimension) {}
220 
221  void factor() override {
222  ensure(LOWER_FISHER_MATRIX | RHS_VECTOR);
223  _ldlt.compute(fisher);
224  }
225 
226  void updateRank() override {}
227 
228  void updateDiagnostic() override {
229  if (whichDiagnostic != LeastSquares::NORMAL_CHOLESKY) {
230  throw LSST_EXCEPT(
231  pex::exceptions::LogicError,
232  "Can only compute NORMAL_CHOLESKY diagnostic from NORMAL_CHOLESKY factorization.");
233  }
234  ndarray::asEigenMatrix(diagnostic) = _ldlt.vectorD();
235  }
236 
237  void updateSolution() override { ndarray::asEigenMatrix(solution) = _ldlt.solve(rhs); }
238 
239  void updateCovariance() override {
240  auto cov = ndarray::asEigenMatrix(covariance);
241  cov.setIdentity();
242  cov = _ldlt.solve(cov);
243  }
244 
245 private:
246  Eigen::LDLT<Eigen::MatrixXd> _ldlt;
247 };
248 
249 class SvdSolver : public LeastSquares::Impl {
250 public:
251  explicit SvdSolver(int dimension) : Impl(dimension), _svd(), _tmp(dimension) {}
252 
253  void factor() override {
254  if (!(state & DESIGN_AND_DATA)) {
255  throw LSST_EXCEPT(pex::exceptions::InvalidParameterError,
256  "Cannot initialize DIRECT_SVD solver with normal equations.");
257  }
258  _svd.compute(design, Eigen::ComputeThinU | Eigen::ComputeThinV);
259  setRank(_svd.singularValues());
260  LOGL_DEBUG(_log, "Using direct SVD method; dimension=%d, rank=%d", dimension, rank);
261  }
262 
263  void updateRank() override { setRank(_svd.singularValues()); }
264 
265  void updateDiagnostic() override {
266  switch (whichDiagnostic) {
268  ndarray::asEigenArray(diagnostic) = _svd.singularValues().array().square();
269  break;
271  throw LSST_EXCEPT(
272  pex::exceptions::LogicError,
273  "Can only compute NORMAL_CHOLESKY diagnostic from DIRECT_SVD factorization.");
275  ndarray::asEigenMatrix(diagnostic) = _svd.singularValues();
276  break;
277  }
278  }
279 
280  void updateSolution() override {
281  if (rank == 0) {
282  ndarray::asEigenMatrix(solution).setZero();
283  return;
284  }
285  _tmp.head(rank) = _svd.matrixU().leftCols(rank).adjoint() * data;
286  _tmp.head(rank).array() /= _svd.singularValues().head(rank).array();
287  ndarray::asEigenMatrix(solution) = _svd.matrixV().leftCols(rank) * _tmp.head(rank);
288  }
289 
290  void updateCovariance() override {
291  if (rank == 0) {
292  ndarray::asEigenMatrix(covariance).setZero();
293  return;
294  }
295  ndarray::asEigenMatrix(covariance) =
296  _svd.matrixV().leftCols(rank) *
297  _svd.singularValues().head(rank).array().inverse().square().matrix().asDiagonal() *
298  _svd.matrixV().leftCols(rank).adjoint();
299  }
300 
301 private:
302  Eigen::JacobiSVD<Eigen::MatrixXd> _svd;
303  Eigen::VectorXd _tmp;
304 };
305 
306 } // namespace
307 
308 void LeastSquares::setThreshold(double threshold) {
309  _impl->threshold = threshold;
310  _impl->state &= ~Impl::SOLUTION_ARRAY;
311  _impl->state &= ~Impl::COVARIANCE_ARRAY;
312  _impl->updateRank();
313 }
314 
315 double LeastSquares::getThreshold() const { return _impl->threshold; }
316 
317 ndarray::Array<double const, 1, 1> LeastSquares::getSolution() {
318  _impl->ensure(Impl::SOLUTION_ARRAY);
319  return _impl->solution;
320 }
321 
322 ndarray::Array<double const, 2, 2> LeastSquares::getCovariance() {
323  _impl->ensure(Impl::COVARIANCE_ARRAY);
324  return _impl->covariance;
325 }
326 
327 ndarray::Array<double const, 2, 2> LeastSquares::getFisherMatrix() {
328  _impl->ensure(Impl::FULL_FISHER_MATRIX);
329  // Wrap the Eigen::MatrixXd in an ndarray::Array, using _impl as the reference-counted owner.
330  // Doesn't matter if we swap strides, because it's symmetric.
331  return ndarray::external(_impl->fisher.data(), ndarray::makeVector(_impl->dimension, _impl->dimension),
332  ndarray::makeVector(_impl->dimension, 1), _impl);
333 }
334 
335 ndarray::Array<double const, 1, 1> LeastSquares::getDiagnostic(Factorization factorization) {
336  if (_impl->whichDiagnostic != factorization) {
337  _impl->state &= ~Impl::DIAGNOSTIC_ARRAY;
338  _impl->whichDiagnostic = factorization;
339  }
340  _impl->ensure(Impl::DIAGNOSTIC_ARRAY);
341  return _impl->diagnostic;
342 }
343 
344 int LeastSquares::getDimension() const { return _impl->dimension; }
345 
346 int LeastSquares::getRank() const { return _impl->rank; }
347 
348 LeastSquares::Factorization LeastSquares::getFactorization() const { return _impl->factorization; }
349 
350 LeastSquares::LeastSquares(Factorization factorization, int dimension) {
351  switch (factorization) {
352  case NORMAL_EIGENSYSTEM:
353  _impl = std::make_shared<EigensystemSolver>(dimension);
354  break;
355  case NORMAL_CHOLESKY:
356  _impl = std::make_shared<CholeskySolver>(dimension);
357  break;
358  case DIRECT_SVD:
359  _impl = std::make_shared<SvdSolver>(dimension);
360  break;
361  }
362  _impl->factorization = factorization;
363 }
364 
365 LeastSquares::LeastSquares(LeastSquares const&) = default;
369 
370 LeastSquares::~LeastSquares() = default;
371 
372 Eigen::MatrixXd& LeastSquares::_getDesignMatrix() { return _impl->design; }
373 Eigen::VectorXd& LeastSquares::_getDataVector() { return _impl->data; }
374 
375 Eigen::MatrixXd& LeastSquares::_getFisherMatrix() { return _impl->fisher; }
376 Eigen::VectorXd& LeastSquares::_getRhsVector() { return _impl->rhs; }
377 
378 void LeastSquares::_factor(bool haveNormalEquations) {
379  if (haveNormalEquations) {
380  if (_getFisherMatrix().rows() != _impl->dimension) {
381  throw LSST_EXCEPT(pex::exceptions::InvalidParameterError,
382  (boost::format("Number of rows of Fisher matrix (%d) does not match"
383  " dimension of LeastSquares solver.") %
384  _getFisherMatrix().rows() % _impl->dimension)
385  .str());
386  }
387  if (_getFisherMatrix().cols() != _impl->dimension) {
388  throw LSST_EXCEPT(pex::exceptions::InvalidParameterError,
389  (boost::format("Number of columns of Fisher matrix (%d) does not match"
390  " dimension of LeastSquares solver.") %
391  _getFisherMatrix().cols() % _impl->dimension)
392  .str());
393  }
394  if (_getRhsVector().size() != _impl->dimension) {
395  throw LSST_EXCEPT(pex::exceptions::InvalidParameterError,
396  (boost::format("Number of elements in RHS vector (%d) does not match"
397  " dimension of LeastSquares solver.") %
398  _getRhsVector().size() % _impl->dimension)
399  .str());
400  }
402  } else {
403  if (_getDesignMatrix().cols() != _impl->dimension) {
404  throw LSST_EXCEPT(
405  pex::exceptions::InvalidParameterError,
406  "Number of columns of design matrix does not match dimension of LeastSquares solver.");
407  }
408  if (_getDesignMatrix().rows() != _getDataVector().size()) {
409  throw LSST_EXCEPT(pex::exceptions::InvalidParameterError,
410  (boost::format("Number of rows of design matrix (%d) does not match number of "
411  "data points (%d)") %
412  _getDesignMatrix().rows() % _getDataVector().size())
413  .str());
414  }
415  if (_getDesignMatrix().cols() > _getDataVector().size()) {
416  throw LSST_EXCEPT(
417  pex::exceptions::InvalidParameterError,
418  (boost::format("Number of columns of design matrix (%d) must be smaller than number of "
419  "data points (%d)") %
420  _getDesignMatrix().cols() % _getDataVector().size())
421  .str());
422  }
423  _impl->state = Impl::DESIGN_AND_DATA;
424  }
425  _impl->factor();
426 }
427 } // namespace math
428 } // namespace afw
429 } // namespace lsst
char * data
Definition: BaseRecord.cc:62
#define LSST_EXCEPT(type,...)
Create an exception with a given type.
Definition: Exception.h:48
LSST DM logging module built on log4cxx.
#define LOG_GET(logger)
Returns a Log object associated with logger.
Definition: Log.h:75
#define LOG_LOGGER
Definition: Log.h:703
#define LOGL_DEBUG(logger, message...)
Log a debug-level message using a varargs/printf style interface.
Definition: Log.h:504
ndarray::Array< double, 1, 1 > solution
Definition: LeastSquares.cc:67
ndarray::Array< double, 2, 2 > covariance
Definition: LeastSquares.cc:68
ndarray::Array< double, 1, 1 > diagnostic
Definition: LeastSquares.cc:69
void setRank(Eigen::MatrixBase< D > const &values)
Definition: LeastSquares.cc:72
Impl(int dimension_, double threshold_=std::numeric_limits< double >::epsilon())
Solver for linear least-squares problems.
Definition: LeastSquares.h:67
ndarray::Array< double const, 1, 1 > getSolution()
Return the vector solution to the least squares problem.
ndarray::Array< double const, 2, 2 > getFisherMatrix()
Return the Fisher matrix (inverse of the covariance) of the parameters.
int getRank() const
Return the rank of the problem (number of nonzero Eigenvalues).
ndarray::Array< double const, 2, 2 > getCovariance()
Return the covariance matrix of the least squares problem.
LeastSquares & operator=(LeastSquares const &)
Factorization
Private implementation; forward-declared publicly so we can inherit from it in .cc.
Definition: LeastSquares.h:71
@ NORMAL_EIGENSYSTEM
Use the normal equations with a symmetric Eigensystem decomposition.
Definition: LeastSquares.h:72
@ NORMAL_CHOLESKY
Use the normal equations with a Cholesky decomposition.
Definition: LeastSquares.h:80
@ DIRECT_SVD
Use a thin singular value decomposition of the design matrix.
Definition: LeastSquares.h:88
LeastSquares(Factorization factorization, int dimension)
Construct a least-squares object for the given factorization and dimensionality.
int getDimension() const
Return the number of parameters.
Factorization getFactorization() const
Retun the type of factorization used by the solver.
ndarray::Array< double const, 1, 1 > getDiagnostic(Factorization factorization)
Return a factorization-dependent vector that can be used to characterize the stability of the solutio...
double getThreshold() const
Get the threshold used to determine when to truncate Eigenvalues.
void setThreshold(double threshold)
Set the threshold used to determine when to truncate Eigenvalues.
def format(config, name=None, writeSourceLine=True, prefix="", verbose=False)
Definition: history.py:174
A base class for image defects.
MatrixQ covariance
Definition: simpleShape.cc:152