1
2.. default-domain:: cpp
3
4.. cpp:namespace:: ceres
5
6.. _chapter-solving:
7
8=======
9Solving
10=======
11
12Introduction
13============
14
15Effective use of Ceres requires some familiarity with the basic
16components of a nonlinear least squares solver, so before we describe
17how to configure and use the solver, we will take a brief look at how
18some of the core optimization algorithms in Ceres work.
19
20Let :math:`x \in \mathbb{R}^n` be an :math:`n`-dimensional vector of
21variables, and
22:math:`F(x) = \left[f_1(x), ... ,  f_{m}(x) \right]^{\top}` be a
23:math:`m`-dimensional function of :math:`x`.  We are interested in
24solving the following optimization problem [#f1]_ .
25
26.. math:: \arg \min_x \frac{1}{2}\|F(x)\|^2\ . \\
27          L \le x \le U
28  :label: nonlinsq
29
30Where, :math:`L` and :math:`U` are lower and upper bounds on the
31parameter vector :math:`x`.
32
33Since the efficient global minimization of :eq:`nonlinsq` for
34general :math:`F(x)` is an intractable problem, we will have to settle
35for finding a local minimum.
36
37In the following, the Jacobian :math:`J(x)` of :math:`F(x)` is an
38:math:`m\times n` matrix, where :math:`J_{ij}(x) = \partial_j f_i(x)`
39and the gradient vector is :math:`g(x) = \nabla \frac{1}{2}\|F(x)\|^2
40= J(x)^\top F(x)`.
41
42The general strategy when solving non-linear optimization problems is
43to solve a sequence of approximations to the original problem
44[NocedalWright]_. At each iteration, the approximation is solved to
45determine a correction :math:`\Delta x` to the vector :math:`x`. For
46non-linear least squares, an approximation can be constructed by using
47the linearization :math:`F(x+\Delta x) \approx F(x) + J(x)\Delta x`,
48which leads to the following linear least squares problem:
49
50.. math:: \min_{\Delta x} \frac{1}{2}\|J(x)\Delta x + F(x)\|^2
51   :label: linearapprox
52
53Unfortunately, naively solving a sequence of these problems and
54updating :math:`x \leftarrow x+ \Delta x` leads to an algorithm that
55may not converge.  To get a convergent algorithm, we need to control
56the size of the step :math:`\Delta x`. Depending on how the size of
57the step :math:`\Delta x` is controlled, non-linear optimization
58algorithms can be divided into two major categories [NocedalWright]_.
59
601. **Trust Region** The trust region approach approximates the
61   objective function using using a model function (often a quadratic)
62   over a subset of the search space known as the trust region. If the
63   model function succeeds in minimizing the true objective function
64   the trust region is expanded; conversely, otherwise it is
65   contracted and the model optimization problem is solved again.
66
672. **Line Search** The line search approach first finds a descent
68   direction along which the objective function will be reduced and
69   then computes a step size that decides how far should move along
70   that direction. The descent direction can be computed by various
71   methods, such as gradient descent, Newton's method and Quasi-Newton
72   method. The step size can be determined either exactly or
73   inexactly.
74
75Trust region methods are in some sense dual to line search methods:
76trust region methods first choose a step size (the size of the trust
77region) and then a step direction while line search methods first
78choose a step direction and then a step size. Ceres implements
79multiple algorithms in both categories.
80
81.. _section-trust-region-methods:
82
83Trust Region Methods
84====================
85
86The basic trust region algorithm looks something like this.
87
88   1. Given an initial point :math:`x` and a trust region radius :math:`\mu`.
89   2. Solve
90
91      .. math::
92         \arg \min_{\Delta x}& \frac{1}{2}\|J(x)\Delta x + F(x)\|^2 \\
93         \text{such that} &\|D(x)\Delta x\|^2 \le \mu\\
94         &L \le x + \Delta x \le U.
95
96   3. :math:`\rho = \frac{\displaystyle \|F(x + \Delta x)\|^2 -
97      \|F(x)\|^2}{\displaystyle \|J(x)\Delta x + F(x)\|^2 -
98      \|F(x)\|^2}`
99   4. if :math:`\rho > \epsilon` then  :math:`x = x + \Delta x`.
100   5. if :math:`\rho > \eta_1` then :math:`\rho = 2  \rho`
101   6. else if :math:`\rho < \eta_2` then :math:`\rho = 0.5 * \rho`
102   7. Go to 2.
103
104Here, :math:`\mu` is the trust region radius, :math:`D(x)` is some
105matrix used to define a metric on the domain of :math:`F(x)` and
106:math:`\rho` measures the quality of the step :math:`\Delta x`, i.e.,
107how well did the linear model predict the decrease in the value of the
108non-linear objective. The idea is to increase or decrease the radius
109of the trust region depending on how well the linearization predicts
110the behavior of the non-linear objective, which in turn is reflected
111in the value of :math:`\rho`.
112
113The key computational step in a trust-region algorithm is the solution
114of the constrained optimization problem
115
116.. math::
117   \arg \min_{\Delta x}& \frac{1}{2}\|J(x)\Delta x + F(x)\|^2 \\
118   \text{such that} &\|D(x)\Delta x\|^2 \le \mu\\
119    &L \le x + \Delta x \le U.
120   :label: trp
121
122There are a number of different ways of solving this problem, each
123giving rise to a different concrete trust-region algorithm. Currently
124Ceres, implements two trust-region algorithms - Levenberg-Marquardt
125and Dogleg, each of which is augmented with a line search if bounds
126constraints are present [Kanzow]_. The user can choose between them by
127setting :member:`Solver::Options::trust_region_strategy_type`.
128
129.. rubric:: Footnotes
130
131.. [#f1] At the level of the non-linear solver, the block structure is
132         not relevant, therefore our discussion here is in terms of an
133         optimization problem defined over a state vector of size
134         :math:`n`. Similarly the presence of loss functions is also
135         ignored as the problem is internally converted into a pure
136         non-linear least squares problem.
137
138
139.. _section-levenberg-marquardt:
140
141Levenberg-Marquardt
142-------------------
143
144The Levenberg-Marquardt algorithm [Levenberg]_ [Marquardt]_ is the
145most popular algorithm for solving non-linear least squares problems.
146It was also the first trust region algorithm to be developed
147[Levenberg]_ [Marquardt]_. Ceres implements an exact step [Madsen]_
148and an inexact step variant of the Levenberg-Marquardt algorithm
149[WrightHolt]_ [NashSofer]_.
150
151It can be shown, that the solution to :eq:`trp` can be obtained by
152solving an unconstrained optimization of the form
153
154.. math:: \arg\min_{\Delta x}& \frac{1}{2}\|J(x)\Delta x + F(x)\|^2 +\lambda  \|D(x)\Delta x\|^2
155
156Where, :math:`\lambda` is a Lagrange multiplier that is inverse
157related to :math:`\mu`. In Ceres, we solve for
158
159.. math:: \arg\min_{\Delta x}& \frac{1}{2}\|J(x)\Delta x + F(x)\|^2 + \frac{1}{\mu} \|D(x)\Delta x\|^2
160   :label: lsqr
161
162The matrix :math:`D(x)` is a non-negative diagonal matrix, typically
163the square root of the diagonal of the matrix :math:`J(x)^\top J(x)`.
164
165Before going further, let us make some notational simplifications. We
166will assume that the matrix :math:`\sqrt{\mu} D` has been concatenated
167at the bottom of the matrix :math:`J` and similarly a vector of zeros
168has been added to the bottom of the vector :math:`f` and the rest of
169our discussion will be in terms of :math:`J` and :math:`f`, i.e, the
170linear least squares problem.
171
172.. math:: \min_{\Delta x} \frac{1}{2} \|J(x)\Delta x + f(x)\|^2 .
173   :label: simple
174
175For all but the smallest problems the solution of :eq:`simple` in
176each iteration of the Levenberg-Marquardt algorithm is the dominant
177computational cost in Ceres. Ceres provides a number of different
178options for solving :eq:`simple`. There are two major classes of
179methods - factorization and iterative.
180
181The factorization methods are based on computing an exact solution of
182:eq:`lsqr` using a Cholesky or a QR factorization and lead to an exact
183step Levenberg-Marquardt algorithm. But it is not clear if an exact
184solution of :eq:`lsqr` is necessary at each step of the LM algorithm
185to solve :eq:`nonlinsq`. In fact, we have already seen evidence
186that this may not be the case, as :eq:`lsqr` is itself a regularized
187version of :eq:`linearapprox`. Indeed, it is possible to
188construct non-linear optimization algorithms in which the linearized
189problem is solved approximately. These algorithms are known as inexact
190Newton or truncated Newton methods [NocedalWright]_.
191
192An inexact Newton method requires two ingredients. First, a cheap
193method for approximately solving systems of linear
194equations. Typically an iterative linear solver like the Conjugate
195Gradients method is used for this
196purpose [NocedalWright]_. Second, a termination rule for
197the iterative solver. A typical termination rule is of the form
198
199.. math:: \|H(x) \Delta x + g(x)\| \leq \eta_k \|g(x)\|.
200   :label: inexact
201
202Here, :math:`k` indicates the Levenberg-Marquardt iteration number and
203:math:`0 < \eta_k <1` is known as the forcing sequence.  [WrightHolt]_
204prove that a truncated Levenberg-Marquardt algorithm that uses an
205inexact Newton step based on :eq:`inexact` converges for any
206sequence :math:`\eta_k \leq \eta_0 < 1` and the rate of convergence
207depends on the choice of the forcing sequence :math:`\eta_k`.
208
209Ceres supports both exact and inexact step solution strategies. When
210the user chooses a factorization based linear solver, the exact step
211Levenberg-Marquardt algorithm is used. When the user chooses an
212iterative linear solver, the inexact step Levenberg-Marquardt
213algorithm is used.
214
215.. _section-dogleg:
216
217Dogleg
218------
219
220Another strategy for solving the trust region problem :eq:`trp` was
221introduced by M. J. D. Powell. The key idea there is to compute two
222vectors
223
224.. math::
225
226        \Delta x^{\text{Gauss-Newton}} &= \arg \min_{\Delta x}\frac{1}{2} \|J(x)\Delta x + f(x)\|^2.\\
227        \Delta x^{\text{Cauchy}} &= -\frac{\|g(x)\|^2}{\|J(x)g(x)\|^2}g(x).
228
229Note that the vector :math:`\Delta x^{\text{Gauss-Newton}}` is the
230solution to :eq:`linearapprox` and :math:`\Delta
231x^{\text{Cauchy}}` is the vector that minimizes the linear
232approximation if we restrict ourselves to moving along the direction
233of the gradient. Dogleg methods finds a vector :math:`\Delta x`
234defined by :math:`\Delta x^{\text{Gauss-Newton}}` and :math:`\Delta
235x^{\text{Cauchy}}` that solves the trust region problem. Ceres
236supports two variants that can be chose by setting
237:member:`Solver::Options::dogleg_type`.
238
239``TRADITIONAL_DOGLEG`` as described by Powell, constructs two line
240segments using the Gauss-Newton and Cauchy vectors and finds the point
241farthest along this line shaped like a dogleg (hence the name) that is
242contained in the trust-region. For more details on the exact reasoning
243and computations, please see Madsen et al [Madsen]_.
244
245``SUBSPACE_DOGLEG`` is a more sophisticated method that considers the
246entire two dimensional subspace spanned by these two vectors and finds
247the point that minimizes the trust region problem in this subspace
248[ByrdSchnabel]_.
249
250The key advantage of the Dogleg over Levenberg Marquardt is that if
251the step computation for a particular choice of :math:`\mu` does not
252result in sufficient decrease in the value of the objective function,
253Levenberg-Marquardt solves the linear approximation from scratch with
254a smaller value of :math:`\mu`. Dogleg on the other hand, only needs
255to compute the interpolation between the Gauss-Newton and the Cauchy
256vectors, as neither of them depend on the value of :math:`\mu`.
257
258The Dogleg method can only be used with the exact factorization based
259linear solvers.
260
261.. _section-inner-iterations:
262
263Inner Iterations
264----------------
265
266Some non-linear least squares problems have additional structure in
267the way the parameter blocks interact that it is beneficial to modify
268the way the trust region step is computed. e.g., consider the
269following regression problem
270
271.. math::   y = a_1 e^{b_1 x} + a_2 e^{b_3 x^2 + c_1}
272
273
274Given a set of pairs :math:`\{(x_i, y_i)\}`, the user wishes to estimate
275:math:`a_1, a_2, b_1, b_2`, and :math:`c_1`.
276
277Notice that the expression on the left is linear in :math:`a_1` and
278:math:`a_2`, and given any value for :math:`b_1, b_2` and :math:`c_1`,
279it is possible to use linear regression to estimate the optimal values
280of :math:`a_1` and :math:`a_2`. It's possible to analytically
281eliminate the variables :math:`a_1` and :math:`a_2` from the problem
282entirely. Problems like these are known as separable least squares
283problem and the most famous algorithm for solving them is the Variable
284Projection algorithm invented by Golub & Pereyra [GolubPereyra]_.
285
286Similar structure can be found in the matrix factorization with
287missing data problem. There the corresponding algorithm is known as
288Wiberg's algorithm [Wiberg]_.
289
290Ruhe & Wedin present an analysis of various algorithms for solving
291separable non-linear least squares problems and refer to *Variable
292Projection* as Algorithm I in their paper [RuheWedin]_.
293
294Implementing Variable Projection is tedious and expensive. Ruhe &
295Wedin present a simpler algorithm with comparable convergence
296properties, which they call Algorithm II.  Algorithm II performs an
297additional optimization step to estimate :math:`a_1` and :math:`a_2`
298exactly after computing a successful Newton step.
299
300
301This idea can be generalized to cases where the residual is not
302linear in :math:`a_1` and :math:`a_2`, i.e.,
303
304.. math:: y = f_1(a_1, e^{b_1 x}) + f_2(a_2, e^{b_3 x^2 + c_1})
305
306In this case, we solve for the trust region step for the full problem,
307and then use it as the starting point to further optimize just `a_1`
308and `a_2`. For the linear case, this amounts to doing a single linear
309least squares solve. For non-linear problems, any method for solving
310the :math:`a_1` and :math:`a_2` optimization problems will do. The
311only constraint on :math:`a_1` and :math:`a_2` (if they are two
312different parameter block) is that they do not co-occur in a residual
313block.
314
315This idea can be further generalized, by not just optimizing
316:math:`(a_1, a_2)`, but decomposing the graph corresponding to the
317Hessian matrix's sparsity structure into a collection of
318non-overlapping independent sets and optimizing each of them.
319
320Setting :member:`Solver::Options::use_inner_iterations` to ``true``
321enables the use of this non-linear generalization of Ruhe & Wedin's
322Algorithm II.  This version of Ceres has a higher iteration
323complexity, but also displays better convergence behavior per
324iteration.
325
326Setting :member:`Solver::Options::num_threads` to the maximum number
327possible is highly recommended.
328
329.. _section-non-monotonic-steps:
330
331Non-monotonic Steps
332-------------------
333
334Note that the basic trust-region algorithm described in
335:ref:`section-trust-region-methods` is a descent algorithm in that it
336only accepts a point if it strictly reduces the value of the objective
337function.
338
339Relaxing this requirement allows the algorithm to be more efficient in
340the long term at the cost of some local increase in the value of the
341objective function.
342
343This is because allowing for non-decreasing objective function values
344in a principled manner allows the algorithm to *jump over boulders* as
345the method is not restricted to move into narrow valleys while
346preserving its convergence properties.
347
348Setting :member:`Solver::Options::use_nonmonotonic_steps` to ``true``
349enables the non-monotonic trust region algorithm as described by Conn,
350Gould & Toint in [Conn]_.
351
352Even though the value of the objective function may be larger
353than the minimum value encountered over the course of the
354optimization, the final parameters returned to the user are the
355ones corresponding to the minimum cost over all iterations.
356
357The option to take non-monotonic steps is available for all trust
358region strategies.
359
360
361.. _section-line-search-methods:
362
363Line Search Methods
364===================
365
366The line search method in Ceres Solver cannot handle bounds
367constraints right now, so it can only be used for solving
368unconstrained problems.
369
370Line search algorithms
371
372   1. Given an initial point :math:`x`
373   2. :math:`\Delta x = -H^{-1}(x) g(x)`
374   3. :math:`\arg \min_\mu \frac{1}{2} \| F(x + \mu \Delta x) \|^2`
375   4. :math:`x = x + \mu \Delta x`
376   5. Goto 2.
377
378Here :math:`H(x)` is some approximation to the Hessian of the
379objective function, and :math:`g(x)` is the gradient at
380:math:`x`. Depending on the choice of :math:`H(x)` we get a variety of
381different search directions :math:`\Delta x`.
382
383Step 4, which is a one dimensional optimization or `Line Search` along
384:math:`\Delta x` is what gives this class of methods its name.
385
386Different line search algorithms differ in their choice of the search
387direction :math:`\Delta x` and the method used for one dimensional
388optimization along :math:`\Delta x`. The choice of :math:`H(x)` is the
389primary source of computational complexity in these
390methods. Currently, Ceres Solver supports three choices of search
391directions, all aimed at large scale problems.
392
3931. ``STEEPEST_DESCENT`` This corresponds to choosing :math:`H(x)` to
394   be the identity matrix. This is not a good search direction for
395   anything but the simplest of the problems. It is only included here
396   for completeness.
397
3982. ``NONLINEAR_CONJUGATE_GRADIENT`` A generalization of the Conjugate
399   Gradient method to non-linear functions. The generalization can be
400   performed in a number of different ways, resulting in a variety of
401   search directions. Ceres Solver currently supports
402   ``FLETCHER_REEVES``, ``POLAK_RIBIERE`` and ``HESTENES_STIEFEL``
403   directions.
404
4053. ``BFGS`` A generalization of the Secant method to multiple
406   dimensions in which a full, dense approximation to the inverse
407   Hessian is maintained and used to compute a quasi-Newton step
408   [NocedalWright]_.  BFGS is currently the best known general
409   quasi-Newton algorithm.
410
4114. ``LBFGS`` A limited memory approximation to the full ``BFGS``
412   method in which the last `M` iterations are used to approximate the
413   inverse Hessian used to compute a quasi-Newton step [Nocedal]_,
414   [ByrdNocedal]_.
415
416Currently Ceres Solver supports both a backtracking and interpolation
417based Armijo line search algorithm, and a sectioning / zoom
418interpolation (strong) Wolfe condition line search algorithm.
419However, note that in order for the assumptions underlying the
420``BFGS`` and ``LBFGS`` methods to be guaranteed to be satisfied the
421Wolfe line search algorithm should be used.
422
423.. _section-linear-solver:
424
425LinearSolver
426============
427
428Recall that in both of the trust-region methods described above, the
429key computational cost is the solution of a linear least squares
430problem of the form
431
432.. math:: \min_{\Delta x} \frac{1}{2} \|J(x)\Delta x + f(x)\|^2 .
433   :label: simple2
434
435Let :math:`H(x)= J(x)^\top J(x)` and :math:`g(x) = -J(x)^\top
436f(x)`. For notational convenience let us also drop the dependence on
437:math:`x`. Then it is easy to see that solving :eq:`simple2` is
438equivalent to solving the *normal equations*.
439
440.. math:: H \Delta x = g
441   :label: normal
442
443Ceres provides a number of different options for solving :eq:`normal`.
444
445.. _section-qr:
446
447``DENSE_QR``
448------------
449
450For small problems (a couple of hundred parameters and a few thousand
451residuals) with relatively dense Jacobians, ``DENSE_QR`` is the method
452of choice [Bjorck]_. Let :math:`J = QR` be the QR-decomposition of
453:math:`J`, where :math:`Q` is an orthonormal matrix and :math:`R` is
454an upper triangular matrix [TrefethenBau]_. Then it can be shown that
455the solution to :eq:`normal` is given by
456
457.. math:: \Delta x^* = -R^{-1}Q^\top f
458
459
460Ceres uses ``Eigen`` 's dense QR factorization routines.
461
462.. _section-cholesky:
463
464``DENSE_NORMAL_CHOLESKY`` & ``SPARSE_NORMAL_CHOLESKY``
465------------------------------------------------------
466
467Large non-linear least square problems are usually sparse. In such
468cases, using a dense QR factorization is inefficient. Let :math:`H =
469R^\top R` be the Cholesky factorization of the normal equations, where
470:math:`R` is an upper triangular matrix, then the solution to
471:eq:`normal` is given by
472
473.. math::
474
475    \Delta x^* = R^{-1} R^{-\top} g.
476
477
478The observant reader will note that the :math:`R` in the Cholesky
479factorization of :math:`H` is the same upper triangular matrix
480:math:`R` in the QR factorization of :math:`J`. Since :math:`Q` is an
481orthonormal matrix, :math:`J=QR` implies that :math:`J^\top J = R^\top
482Q^\top Q R = R^\top R`. There are two variants of Cholesky
483factorization -- sparse and dense.
484
485``DENSE_NORMAL_CHOLESKY``  as the name implies performs a dense
486Cholesky factorization of the normal equations. Ceres uses
487``Eigen`` 's dense LDLT factorization routines.
488
489``SPARSE_NORMAL_CHOLESKY``, as the name implies performs a sparse
490Cholesky factorization of the normal equations. This leads to
491substantial savings in time and memory for large sparse
492problems. Ceres uses the sparse Cholesky factorization routines in
493Professor Tim Davis' ``SuiteSparse`` or ``CXSparse`` packages [Chen]_
494or the sparse Cholesky factorization algorithm in ``Eigen`` (which
495incidently is a port of the algorithm implemented inside ``CXSparse``)
496
497.. _section-schur:
498
499``DENSE_SCHUR`` & ``SPARSE_SCHUR``
500----------------------------------
501
502While it is possible to use ``SPARSE_NORMAL_CHOLESKY`` to solve bundle
503adjustment problems, bundle adjustment problem have a special
504structure, and a more efficient scheme for solving :eq:`normal`
505can be constructed.
506
507Suppose that the SfM problem consists of :math:`p` cameras and
508:math:`q` points and the variable vector :math:`x` has the block
509structure :math:`x = [y_{1}, ... ,y_{p},z_{1}, ... ,z_{q}]`. Where,
510:math:`y` and :math:`z` correspond to camera and point parameters,
511respectively.  Further, let the camera blocks be of size :math:`c` and
512the point blocks be of size :math:`s` (for most problems :math:`c` =
513:math:`6`--`9` and :math:`s = 3`). Ceres does not impose any constancy
514requirement on these block sizes, but choosing them to be constant
515simplifies the exposition.
516
517A key characteristic of the bundle adjustment problem is that there is
518no term :math:`f_{i}` that includes two or more point blocks.  This in
519turn implies that the matrix :math:`H` is of the form
520
521.. math:: H = \left[ \begin{matrix} B & E\\ E^\top & C \end{matrix} \right]\ ,
522   :label: hblock
523
524where, :math:`B \in \mathbb{R}^{pc\times pc}` is a block sparse matrix
525with :math:`p` blocks of size :math:`c\times c` and :math:`C \in
526\mathbb{R}^{qs\times qs}` is a block diagonal matrix with :math:`q` blocks
527of size :math:`s\times s`. :math:`E \in \mathbb{R}^{pc\times qs}` is a
528general block sparse matrix, with a block of size :math:`c\times s`
529for each observation. Let us now block partition :math:`\Delta x =
530[\Delta y,\Delta z]` and :math:`g=[v,w]` to restate :eq:`normal`
531as the block structured linear system
532
533.. math:: \left[ \begin{matrix} B & E\\ E^\top & C \end{matrix}
534                \right]\left[ \begin{matrix} \Delta y \\ \Delta z
535                    \end{matrix} \right] = \left[ \begin{matrix} v\\ w
536                    \end{matrix} \right]\ ,
537   :label: linear2
538
539and apply Gaussian elimination to it. As we noted above, :math:`C` is
540a block diagonal matrix, with small diagonal blocks of size
541:math:`s\times s`.  Thus, calculating the inverse of :math:`C` by
542inverting each of these blocks is cheap. This allows us to eliminate
543:math:`\Delta z` by observing that :math:`\Delta z = C^{-1}(w - E^\top
544\Delta y)`, giving us
545
546.. math:: \left[B - EC^{-1}E^\top\right] \Delta y = v - EC^{-1}w\ .
547   :label: schur
548
549The matrix
550
551.. math:: S = B - EC^{-1}E^\top
552
553is the Schur complement of :math:`C` in :math:`H`. It is also known as
554the *reduced camera matrix*, because the only variables
555participating in :eq:`schur` are the ones corresponding to the
556cameras. :math:`S \in \mathbb{R}^{pc\times pc}` is a block structured
557symmetric positive definite matrix, with blocks of size :math:`c\times
558c`. The block :math:`S_{ij}` corresponding to the pair of images
559:math:`i` and :math:`j` is non-zero if and only if the two images
560observe at least one common point.
561
562
563Now, eq-linear2 can be solved by first forming :math:`S`, solving for
564:math:`\Delta y`, and then back-substituting :math:`\Delta y` to
565obtain the value of :math:`\Delta z`.  Thus, the solution of what was
566an :math:`n\times n`, :math:`n=pc+qs` linear system is reduced to the
567inversion of the block diagonal matrix :math:`C`, a few matrix-matrix
568and matrix-vector multiplies, and the solution of block sparse
569:math:`pc\times pc` linear system :eq:`schur`.  For almost all
570problems, the number of cameras is much smaller than the number of
571points, :math:`p \ll q`, thus solving :eq:`schur` is
572significantly cheaper than solving :eq:`linear2`. This is the
573*Schur complement trick* [Brown]_.
574
575This still leaves open the question of solving :eq:`schur`. The
576method of choice for solving symmetric positive definite systems
577exactly is via the Cholesky factorization [TrefethenBau]_ and
578depending upon the structure of the matrix, there are, in general, two
579options. The first is direct factorization, where we store and factor
580:math:`S` as a dense matrix [TrefethenBau]_. This method has
581:math:`O(p^2)` space complexity and :math:`O(p^3)` time complexity and
582is only practical for problems with up to a few hundred cameras. Ceres
583implements this strategy as the ``DENSE_SCHUR`` solver.
584
585
586But, :math:`S` is typically a fairly sparse matrix, as most images
587only see a small fraction of the scene. This leads us to the second
588option: Sparse Direct Methods. These methods store :math:`S` as a
589sparse matrix, use row and column re-ordering algorithms to maximize
590the sparsity of the Cholesky decomposition, and focus their compute
591effort on the non-zero part of the factorization [Chen]_. Sparse
592direct methods, depending on the exact sparsity structure of the Schur
593complement, allow bundle adjustment algorithms to significantly scale
594up over those based on dense factorization. Ceres implements this
595strategy as the ``SPARSE_SCHUR`` solver.
596
597.. _section-cgnr:
598
599``CGNR``
600--------
601
602For general sparse problems, if the problem is too large for
603``CHOLMOD`` or a sparse linear algebra library is not linked into
604Ceres, another option is the ``CGNR`` solver. This solver uses the
605Conjugate Gradients solver on the *normal equations*, but without
606forming the normal equations explicitly. It exploits the relation
607
608.. math::
609    H x = J^\top J x = J^\top(J x)
610
611
612When the user chooses ``ITERATIVE_SCHUR`` as the linear solver, Ceres
613automatically switches from the exact step algorithm to an inexact
614step algorithm.
615
616.. _section-iterative_schur:
617
618``ITERATIVE_SCHUR``
619-------------------
620
621Another option for bundle adjustment problems is to apply PCG to the
622reduced camera matrix :math:`S` instead of :math:`H`. One reason to do
623this is that :math:`S` is a much smaller matrix than :math:`H`, but
624more importantly, it can be shown that :math:`\kappa(S)\leq
625\kappa(H)`.  Cseres implements PCG on :math:`S` as the
626``ITERATIVE_SCHUR`` solver. When the user chooses ``ITERATIVE_SCHUR``
627as the linear solver, Ceres automatically switches from the exact step
628algorithm to an inexact step algorithm.
629
630The cost of forming and storing the Schur complement :math:`S` can be
631prohibitive for large problems. Indeed, for an inexact Newton solver
632that computes :math:`S` and runs PCG on it, almost all of its time is
633spent in constructing :math:`S`; the time spent inside the PCG
634algorithm is negligible in comparison. Because PCG only needs access
635to :math:`S` via its product with a vector, one way to evaluate
636:math:`Sx` is to observe that
637
638.. math::  x_1 &= E^\top x
639.. math::  x_2 &= C^{-1} x_1
640.. math::  x_3 &= Ex_2\\
641.. math::  x_4 &= Bx\\
642.. math::   Sx &= x_4 - x_3
643   :label: schurtrick1
644
645Thus, we can run PCG on :math:`S` with the same computational effort
646per iteration as PCG on :math:`H`, while reaping the benefits of a
647more powerful preconditioner. In fact, we do not even need to compute
648:math:`H`, :eq:`schurtrick1` can be implemented using just the columns
649of :math:`J`.
650
651Equation :eq:`schurtrick1` is closely related to *Domain
652Decomposition methods* for solving large linear systems that arise in
653structural engineering and partial differential equations. In the
654language of Domain Decomposition, each point in a bundle adjustment
655problem is a domain, and the cameras form the interface between these
656domains. The iterative solution of the Schur complement then falls
657within the sub-category of techniques known as Iterative
658Sub-structuring [Saad]_ [Mathew]_.
659
660.. _section-preconditioner:
661
662Preconditioner
663--------------
664
665The convergence rate of Conjugate Gradients for
666solving :eq:`normal` depends on the distribution of eigenvalues
667of :math:`H` [Saad]_. A useful upper bound is
668:math:`\sqrt{\kappa(H)}`, where, :math:`\kappa(H)` is the condition
669number of the matrix :math:`H`. For most bundle adjustment problems,
670:math:`\kappa(H)` is high and a direct application of Conjugate
671Gradients to :eq:`normal` results in extremely poor performance.
672
673The solution to this problem is to replace :eq:`normal` with a
674*preconditioned* system.  Given a linear system, :math:`Ax =b` and a
675preconditioner :math:`M` the preconditioned system is given by
676:math:`M^{-1}Ax = M^{-1}b`. The resulting algorithm is known as
677Preconditioned Conjugate Gradients algorithm (PCG) and its worst case
678complexity now depends on the condition number of the *preconditioned*
679matrix :math:`\kappa(M^{-1}A)`.
680
681The computational cost of using a preconditioner :math:`M` is the cost
682of computing :math:`M` and evaluating the product :math:`M^{-1}y` for
683arbitrary vectors :math:`y`. Thus, there are two competing factors to
684consider: How much of :math:`H`'s structure is captured by :math:`M`
685so that the condition number :math:`\kappa(HM^{-1})` is low, and the
686computational cost of constructing and using :math:`M`.  The ideal
687preconditioner would be one for which :math:`\kappa(M^{-1}A)
688=1`. :math:`M=A` achieves this, but it is not a practical choice, as
689applying this preconditioner would require solving a linear system
690equivalent to the unpreconditioned problem.  It is usually the case
691that the more information :math:`M` has about :math:`H`, the more
692expensive it is use. For example, Incomplete Cholesky factorization
693based preconditioners have much better convergence behavior than the
694Jacobi preconditioner, but are also much more expensive.
695
696
697The simplest of all preconditioners is the diagonal or Jacobi
698preconditioner, i.e., :math:`M=\operatorname{diag}(A)`, which for
699block structured matrices like :math:`H` can be generalized to the
700block Jacobi preconditioner.
701
702For ``ITERATIVE_SCHUR`` there are two obvious choices for block
703diagonal preconditioners for :math:`S`. The block diagonal of the
704matrix :math:`B` [Mandel]_ and the block diagonal :math:`S`, i.e, the
705block Jacobi preconditioner for :math:`S`. Ceres's implements both of
706these preconditioners and refers to them as ``JACOBI`` and
707``SCHUR_JACOBI`` respectively.
708
709For bundle adjustment problems arising in reconstruction from
710community photo collections, more effective preconditioners can be
711constructed by analyzing and exploiting the camera-point visibility
712structure of the scene [KushalAgarwal]. Ceres implements the two
713visibility based preconditioners described by Kushal & Agarwal as
714``CLUSTER_JACOBI`` and ``CLUSTER_TRIDIAGONAL``. These are fairly new
715preconditioners and Ceres' implementation of them is in its early
716stages and is not as mature as the other preconditioners described
717above.
718
719.. _section-ordering:
720
721Ordering
722--------
723
724The order in which variables are eliminated in a linear solver can
725have a significant of impact on the efficiency and accuracy of the
726method. For example when doing sparse Cholesky factorization, there
727are matrices for which a good ordering will give a Cholesky factor
728with :math:`O(n)` storage, where as a bad ordering will result in an
729completely dense factor.
730
731Ceres allows the user to provide varying amounts of hints to the
732solver about the variable elimination ordering to use. This can range
733from no hints, where the solver is free to decide the best ordering
734based on the user's choices like the linear solver being used, to an
735exact order in which the variables should be eliminated, and a variety
736of possibilities in between.
737
738Instances of the :class:`ParameterBlockOrdering` class are used to
739communicate this information to Ceres.
740
741Formally an ordering is an ordered partitioning of the parameter
742blocks. Each parameter block belongs to exactly one group, and each
743group has a unique integer associated with it, that determines its
744order in the set of groups. We call these groups *Elimination Groups*
745
746Given such an ordering, Ceres ensures that the parameter blocks in the
747lowest numbered elimination group are eliminated first, and then the
748parameter blocks in the next lowest numbered elimination group and so
749on. Within each elimination group, Ceres is free to order the
750parameter blocks as it chooses. e.g. Consider the linear system
751
752.. math::
753  x + y &= 3\\
754  2x + 3y &= 7
755
756There are two ways in which it can be solved. First eliminating
757:math:`x` from the two equations, solving for y and then back
758substituting for :math:`x`, or first eliminating :math:`y`, solving
759for :math:`x` and back substituting for :math:`y`. The user can
760construct three orderings here.
761
7621. :math:`\{0: x\}, \{1: y\}` : Eliminate :math:`x` first.
7632. :math:`\{0: y\}, \{1: x\}` : Eliminate :math:`y` first.
7643. :math:`\{0: x, y\}`        : Solver gets to decide the elimination order.
765
766Thus, to have Ceres determine the ordering automatically using
767heuristics, put all the variables in the same elimination group. The
768identity of the group does not matter. This is the same as not
769specifying an ordering at all. To control the ordering for every
770variable, create an elimination group per variable, ordering them in
771the desired order.
772
773If the user is using one of the Schur solvers (``DENSE_SCHUR``,
774``SPARSE_SCHUR``, ``ITERATIVE_SCHUR``) and chooses to specify an
775ordering, it must have one important property. The lowest numbered
776elimination group must form an independent set in the graph
777corresponding to the Hessian, or in other words, no two parameter
778blocks in in the first elimination group should co-occur in the same
779residual block. For the best performance, this elimination group
780should be as large as possible. For standard bundle adjustment
781problems, this corresponds to the first elimination group containing
782all the 3d points, and the second containing the all the cameras
783parameter blocks.
784
785If the user leaves the choice to Ceres, then the solver uses an
786approximate maximum independent set algorithm to identify the first
787elimination group [LiSaad]_.
788
789.. _section-solver-options:
790
791:class:`Solver::Options`
792------------------------
793
794.. class:: Solver::Options
795
796   :class:`Solver::Options` controls the overall behavior of the
797   solver. We list the various settings and their default values below.
798
799.. function:: bool Solver::Options::IsValid(string* error) const
800
801   Validate the values in the options struct and returns true on
802   success. If there is a problem, the method returns false with
803   ``error`` containing a textual description of the cause.
804
805.. member:: MinimizerType Solver::Options::minimizer_type
806
807   Default: ``TRUST_REGION``
808
809   Choose between ``LINE_SEARCH`` and ``TRUST_REGION`` algorithms. See
810   :ref:`section-trust-region-methods` and
811   :ref:`section-line-search-methods` for more details.
812
813.. member:: LineSearchDirectionType Solver::Options::line_search_direction_type
814
815   Default: ``LBFGS``
816
817   Choices are ``STEEPEST_DESCENT``, ``NONLINEAR_CONJUGATE_GRADIENT``,
818   ``BFGS`` and ``LBFGS``.
819
820.. member:: LineSearchType Solver::Options::line_search_type
821
822   Default: ``WOLFE``
823
824   Choices are ``ARMIJO`` and ``WOLFE`` (strong Wolfe conditions).
825   Note that in order for the assumptions underlying the ``BFGS`` and
826   ``LBFGS`` line search direction algorithms to be guaranteed to be
827   satisifed, the ``WOLFE`` line search should be used.
828
829.. member:: NonlinearConjugateGradientType Solver::Options::nonlinear_conjugate_gradient_type
830
831   Default: ``FLETCHER_REEVES``
832
833   Choices are ``FLETCHER_REEVES``, ``POLAK_RIBIERE`` and
834   ``HESTENES_STIEFEL``.
835
836.. member:: int Solver::Options::max_lbfs_rank
837
838   Default: 20
839
840   The L-BFGS hessian approximation is a low rank approximation to the
841   inverse of the Hessian matrix. The rank of the approximation
842   determines (linearly) the space and time complexity of using the
843   approximation. Higher the rank, the better is the quality of the
844   approximation. The increase in quality is however is bounded for a
845   number of reasons.
846
847     1. The method only uses secant information and not actual
848        derivatives.
849
850     2. The Hessian approximation is constrained to be positive
851        definite.
852
853   So increasing this rank to a large number will cost time and space
854   complexity without the corresponding increase in solution
855   quality. There are no hard and fast rules for choosing the maximum
856   rank. The best choice usually requires some problem specific
857   experimentation.
858
859.. member:: bool Solver::Options::use_approximate_eigenvalue_bfgs_scaling
860
861   Default: ``false``
862
863   As part of the ``BFGS`` update step / ``LBFGS`` right-multiply
864   step, the initial inverse Hessian approximation is taken to be the
865   Identity.  However, [Oren]_ showed that using instead :math:`I *
866   \gamma`, where :math:`\gamma` is a scalar chosen to approximate an
867   eigenvalue of the true inverse Hessian can result in improved
868   convergence in a wide variety of cases.  Setting
869   ``use_approximate_eigenvalue_bfgs_scaling`` to true enables this
870   scaling in ``BFGS`` (before first iteration) and ``LBFGS`` (at each
871   iteration).
872
873   Precisely, approximate eigenvalue scaling equates to
874
875   .. math:: \gamma = \frac{y_k' s_k}{y_k' y_k}
876
877   With:
878
879  .. math:: y_k = \nabla f_{k+1} - \nabla f_k
880  .. math:: s_k = x_{k+1} - x_k
881
882  Where :math:`f()` is the line search objective and :math:`x` the
883  vector of parameter values [NocedalWright]_.
884
885  It is important to note that approximate eigenvalue scaling does
886  **not** *always* improve convergence, and that it can in fact
887  *significantly* degrade performance for certain classes of problem,
888  which is why it is disabled by default.  In particular it can
889  degrade performance when the sensitivity of the problem to different
890  parameters varies significantly, as in this case a single scalar
891  factor fails to capture this variation and detrimentally downscales
892  parts of the Jacobian approximation which correspond to
893  low-sensitivity parameters. It can also reduce the robustness of the
894  solution to errors in the Jacobians.
895
896.. member:: LineSearchIterpolationType Solver::Options::line_search_interpolation_type
897
898   Default: ``CUBIC``
899
900   Degree of the polynomial used to approximate the objective
901   function. Valid values are ``BISECTION``, ``QUADRATIC`` and
902   ``CUBIC``.
903
904.. member:: double Solver::Options::min_line_search_step_size
905
906   The line search terminates if:
907
908   .. math:: \|\Delta x_k\|_\infty < \text{min_line_search_step_size}
909
910   where :math:`\|\cdot\|_\infty` refers to the max norm, and
911   :math:`\Delta x_k` is the step change in the parameter values at
912   the :math:`k`-th iteration.
913
914.. member:: double Solver::Options::line_search_sufficient_function_decrease
915
916   Default: ``1e-4``
917
918   Solving the line search problem exactly is computationally
919   prohibitive. Fortunately, line search based optimization algorithms
920   can still guarantee convergence if instead of an exact solution,
921   the line search algorithm returns a solution which decreases the
922   value of the objective function sufficiently. More precisely, we
923   are looking for a step size s.t.
924
925   .. math:: f(\text{step_size}) \le f(0) + \text{sufficient_decrease} * [f'(0) * \text{step_size}]
926
927   This condition is known as the Armijo condition.
928
929.. member:: double Solver::Options::max_line_search_step_contraction
930
931   Default: ``1e-3``
932
933   In each iteration of the line search,
934
935   .. math:: \text{new_step_size} >= \text{max_line_search_step_contraction} * \text{step_size}
936
937   Note that by definition, for contraction:
938
939   .. math:: 0 < \text{max_step_contraction} < \text{min_step_contraction} < 1
940
941.. member:: double Solver::Options::min_line_search_step_contraction
942
943   Default: ``0.6``
944
945   In each iteration of the line search,
946
947   .. math:: \text{new_step_size} <= \text{min_line_search_step_contraction} * \text{step_size}
948
949   Note that by definition, for contraction:
950
951   .. math:: 0 < \text{max_step_contraction} < \text{min_step_contraction} < 1
952
953.. member:: int Solver::Options::max_num_line_search_step_size_iterations
954
955   Default: ``20``
956
957   Maximum number of trial step size iterations during each line
958   search, if a step size satisfying the search conditions cannot be
959   found within this number of trials, the line search will stop.
960
961   As this is an 'artificial' constraint (one imposed by the user, not
962   the underlying math), if ``WOLFE`` line search is being used, *and*
963   points satisfying the Armijo sufficient (function) decrease
964   condition have been found during the current search (in :math:`<=`
965   ``max_num_line_search_step_size_iterations``).  Then, the step size
966   with the lowest function value which satisfies the Armijo condition
967   will be returned as the new valid step, even though it does *not*
968   satisfy the strong Wolfe conditions.  This behaviour protects
969   against early termination of the optimizer at a sub-optimal point.
970
971.. member:: int Solver::Options::max_num_line_search_direction_restarts
972
973   Default: ``5``
974
975   Maximum number of restarts of the line search direction algorithm
976   before terminating the optimization. Restarts of the line search
977   direction algorithm occur when the current algorithm fails to
978   produce a new descent direction. This typically indicates a
979   numerical failure, or a breakdown in the validity of the
980   approximations used.
981
982.. member:: double Solver::Options::line_search_sufficient_curvature_decrease
983
984   Default: ``0.9``
985
986   The strong Wolfe conditions consist of the Armijo sufficient
987   decrease condition, and an additional requirement that the
988   step size be chosen s.t. the *magnitude* ('strong' Wolfe
989   conditions) of the gradient along the search direction
990   decreases sufficiently. Precisely, this second condition
991   is that we seek a step size s.t.
992
993   .. math:: \|f'(\text{step_size})\| <= \text{sufficient_curvature_decrease} * \|f'(0)\|
994
995   Where :math:`f()` is the line search objective and :math:`f'()` is the derivative
996   of :math:`f` with respect to the step size: :math:`\frac{d f}{d~\text{step size}}`.
997
998.. member:: double Solver::Options::max_line_search_step_expansion
999
1000   Default: ``10.0``
1001
1002   During the bracketing phase of a Wolfe line search, the step size
1003   is increased until either a point satisfying the Wolfe conditions
1004   is found, or an upper bound for a bracket containinqg a point
1005   satisfying the conditions is found.  Precisely, at each iteration
1006   of the expansion:
1007
1008   .. math:: \text{new_step_size} <= \text{max_step_expansion} * \text{step_size}
1009
1010   By definition for expansion
1011
1012   .. math:: \text{max_step_expansion} > 1.0
1013
1014.. member:: TrustRegionStrategyType Solver::Options::trust_region_strategy_type
1015
1016   Default: ``LEVENBERG_MARQUARDT``
1017
1018   The trust region step computation algorithm used by
1019   Ceres. Currently ``LEVENBERG_MARQUARDT`` and ``DOGLEG`` are the two
1020   valid choices. See :ref:`section-levenberg-marquardt` and
1021   :ref:`section-dogleg` for more details.
1022
1023.. member:: DoglegType Solver::Options::dogleg_type
1024
1025   Default: ``TRADITIONAL_DOGLEG``
1026
1027   Ceres supports two different dogleg strategies.
1028   ``TRADITIONAL_DOGLEG`` method by Powell and the ``SUBSPACE_DOGLEG``
1029   method described by [ByrdSchnabel]_ .  See :ref:`section-dogleg`
1030   for more details.
1031
1032.. member:: bool Solver::Options::use_nonmonotonic_steps
1033
1034   Default: ``false``
1035
1036   Relax the requirement that the trust-region algorithm take strictly
1037   decreasing steps. See :ref:`section-non-monotonic-steps` for more
1038   details.
1039
1040.. member:: int Solver::Options::max_consecutive_nonmonotonic_steps
1041
1042   Default: ``5``
1043
1044   The window size used by the step selection algorithm to accept
1045   non-monotonic steps.
1046
1047.. member:: int Solver::Options::max_num_iterations
1048
1049   Default: ``50``
1050
1051   Maximum number of iterations for which the solver should run.
1052
1053.. member:: double Solver::Options::max_solver_time_in_seconds
1054
1055   Default: ``1e6``
1056   Maximum amount of time for which the solver should run.
1057
1058.. member:: int Solver::Options::num_threads
1059
1060   Default: ``1``
1061
1062   Number of threads used by Ceres to evaluate the Jacobian.
1063
1064.. member::  double Solver::Options::initial_trust_region_radius
1065
1066   Default: ``1e4``
1067
1068   The size of the initial trust region. When the
1069   ``LEVENBERG_MARQUARDT`` strategy is used, the reciprocal of this
1070   number is the initial regularization parameter.
1071
1072.. member:: double Solver::Options::max_trust_region_radius
1073
1074   Default: ``1e16``
1075
1076   The trust region radius is not allowed to grow beyond this value.
1077
1078.. member:: double Solver::Options::min_trust_region_radius
1079
1080   Default: ``1e-32``
1081
1082   The solver terminates, when the trust region becomes smaller than
1083   this value.
1084
1085.. member:: double Solver::Options::min_relative_decrease
1086
1087   Default: ``1e-3``
1088
1089   Lower threshold for relative decrease before a trust-region step is
1090   accepted.
1091
1092.. member:: double Solver::Options::min_lm_diagonal
1093
1094   Default: ``1e6``
1095
1096   The ``LEVENBERG_MARQUARDT`` strategy, uses a diagonal matrix to
1097   regularize the the trust region step. This is the lower bound on
1098   the values of this diagonal matrix.
1099
1100.. member:: double Solver::Options::max_lm_diagonal
1101
1102   Default:  ``1e32``
1103
1104   The ``LEVENBERG_MARQUARDT`` strategy, uses a diagonal matrix to
1105   regularize the the trust region step. This is the upper bound on
1106   the values of this diagonal matrix.
1107
1108.. member:: int Solver::Options::max_num_consecutive_invalid_steps
1109
1110   Default: ``5``
1111
1112   The step returned by a trust region strategy can sometimes be
1113   numerically invalid, usually because of conditioning
1114   issues. Instead of crashing or stopping the optimization, the
1115   optimizer can go ahead and try solving with a smaller trust
1116   region/better conditioned problem. This parameter sets the number
1117   of consecutive retries before the minimizer gives up.
1118
1119.. member:: double Solver::Options::function_tolerance
1120
1121   Default: ``1e-6``
1122
1123   Solver terminates if
1124
1125   .. math:: \frac{|\Delta \text{cost}|}{\text{cost}} < \text{function_tolerance}
1126
1127   where, :math:`\Delta \text{cost}` is the change in objective
1128   function value (up or down) in the current iteration of
1129   Levenberg-Marquardt.
1130
1131.. member:: double Solver::Options::gradient_tolerance
1132
1133   Default: ``1e-10``
1134
1135   Solver terminates if
1136
1137   .. math:: \|x - \Pi \boxplus(x, -g(x))\|_\infty < \text{gradient_tolerance}
1138
1139   where :math:`\|\cdot\|_\infty` refers to the max norm, :math:`\Pi`
1140   is projection onto the bounds constraints and :math:`\boxplus` is
1141   Plus operation for the overall local parameterization associated
1142   with the parameter vector.
1143
1144.. member:: double Solver::Options::parameter_tolerance
1145
1146   Default: ``1e-8``
1147
1148   Solver terminates if
1149
1150   .. math:: \|\Delta x\| < (\|x\| + \text{parameter_tolerance}) * \text{parameter_tolerance}
1151
1152   where :math:`\Delta x` is the step computed by the linear solver in
1153   the current iteration of Levenberg-Marquardt.
1154
1155.. member:: LinearSolverType Solver::Options::linear_solver_type
1156
1157   Default: ``SPARSE_NORMAL_CHOLESKY`` / ``DENSE_QR``
1158
1159   Type of linear solver used to compute the solution to the linear
1160   least squares problem in each iteration of the Levenberg-Marquardt
1161   algorithm. If Ceres is built with support for ``SuiteSparse`` or
1162   ``CXSparse`` or ``Eigen``'s sparse Cholesky factorization, the
1163   default is ``SPARSE_NORMAL_CHOLESKY``, it is ``DENSE_QR``
1164   otherwise.
1165
1166.. member:: PreconditionerType Solver::Options::preconditioner_type
1167
1168   Default: ``JACOBI``
1169
1170   The preconditioner used by the iterative linear solver. The default
1171   is the block Jacobi preconditioner. Valid values are (in increasing
1172   order of complexity) ``IDENTITY``, ``JACOBI``, ``SCHUR_JACOBI``,
1173   ``CLUSTER_JACOBI`` and ``CLUSTER_TRIDIAGONAL``. See
1174   :ref:`section-preconditioner` for more details.
1175
1176.. member:: VisibilityClusteringType Solver::Options::visibility_clustering_type
1177
1178   Default: ``CANONICAL_VIEWS``
1179
1180   Type of clustering algorithm to use when constructing a visibility
1181   based preconditioner. The original visibility based preconditioning
1182   paper and implementation only used the canonical views algorithm.
1183
1184   This algorithm gives high quality results but for large dense
1185   graphs can be particularly expensive. As its worst case complexity
1186   is cubic in size of the graph.
1187
1188   Another option is to use ``SINGLE_LINKAGE`` which is a simple
1189   thresholded single linkage clustering algorithm that only pays
1190   attention to tightly coupled blocks in the Schur complement. This
1191   is a fast algorithm that works well.
1192
1193   The optimal choice of the clustering algorithm depends on the
1194   sparsity structure of the problem, but generally speaking we
1195   recommend that you try ``CANONICAL_VIEWS`` first and if it is too
1196   expensive try ``SINGLE_LINKAGE``.
1197
1198.. member:: DenseLinearAlgebraLibrary Solver::Options::dense_linear_algebra_library_type
1199
1200   Default:``EIGEN``
1201
1202   Ceres supports using multiple dense linear algebra libraries for
1203   dense matrix factorizations. Currently ``EIGEN`` and ``LAPACK`` are
1204   the valid choices. ``EIGEN`` is always available, ``LAPACK`` refers
1205   to the system ``BLAS + LAPACK`` library which may or may not be
1206   available.
1207
1208   This setting affects the ``DENSE_QR``, ``DENSE_NORMAL_CHOLESKY``
1209   and ``DENSE_SCHUR`` solvers. For small to moderate sized probem
1210   ``EIGEN`` is a fine choice but for large problems, an optimized
1211   ``LAPACK + BLAS`` implementation can make a substantial difference
1212   in performance.
1213
1214.. member:: SparseLinearAlgebraLibrary Solver::Options::sparse_linear_algebra_library_type
1215
1216   Default:``SUITE_SPARSE``
1217
1218   Ceres supports the use of three sparse linear algebra libraries,
1219   ``SuiteSparse``, which is enabled by setting this parameter to
1220   ``SUITE_SPARSE``, ``CXSparse``, which can be selected by setting
1221   this parameter to ```CX_SPARSE`` and ``Eigen`` which is enabled by
1222   setting this parameter to ``EIGEN_SPARSE``.
1223
1224   ``SuiteSparse`` is a sophisticated and complex sparse linear
1225   algebra library and should be used in general.
1226
1227   If your needs/platforms prevent you from using ``SuiteSparse``,
1228   consider using ``CXSparse``, which is a much smaller, easier to
1229   build library. As can be expected, its performance on large
1230   problems is not comparable to that of ``SuiteSparse``.
1231
1232   Last but not the least you can use the sparse linear algebra
1233   routines in ``Eigen``. Currently the performance of this library is
1234   the poorest of the three. But this should change in the near
1235   future.
1236
1237   Another thing to consider here is that the sparse Cholesky
1238   factorization libraries in Eigen are licensed under ``LGPL`` and
1239   building Ceres with support for ``EIGEN_SPARSE`` will result in an
1240   LGPL licensed library (since the corresponding code from Eigen is
1241   compiled into the library).
1242
1243   The upside is that you do not need to build and link to an external
1244   library to use ``EIGEN_SPARSE``.
1245
1246.. member:: int Solver::Options::num_linear_solver_threads
1247
1248   Default: ``1``
1249
1250   Number of threads used by the linear solver.
1251
1252.. member:: shared_ptr<ParameterBlockOrdering> Solver::Options::linear_solver_ordering
1253
1254   Default: ``NULL``
1255
1256   An instance of the ordering object informs the solver about the
1257   desired order in which parameter blocks should be eliminated by the
1258   linear solvers. See section~\ref{sec:ordering`` for more details.
1259
1260   If ``NULL``, the solver is free to choose an ordering that it
1261   thinks is best.
1262
1263   See :ref:`section-ordering` for more details.
1264
1265.. member:: bool Solver::Options::use_post_ordering
1266
1267   Default: ``false``
1268
1269   Sparse Cholesky factorization algorithms use a fill-reducing
1270   ordering to permute the columns of the Jacobian matrix. There are
1271   two ways of doing this.
1272
1273   1. Compute the Jacobian matrix in some order and then have the
1274      factorization algorithm permute the columns of the Jacobian.
1275
1276   2. Compute the Jacobian with its columns already permuted.
1277
1278   The first option incurs a significant memory penalty. The
1279   factorization algorithm has to make a copy of the permuted Jacobian
1280   matrix, thus Ceres pre-permutes the columns of the Jacobian matrix
1281   and generally speaking, there is no performance penalty for doing
1282   so.
1283
1284   In some rare cases, it is worth using a more complicated reordering
1285   algorithm which has slightly better runtime performance at the
1286   expense of an extra copy of the Jacobian matrix. Setting
1287   ``use_postordering`` to ``true`` enables this tradeoff.
1288
1289.. member:: bool Solver::Options::dynamic_sparsity
1290
1291   Some non-linear least squares problems are symbolically dense but
1292   numerically sparse. i.e. at any given state only a small number of
1293   Jacobian entries are non-zero, but the position and number of
1294   non-zeros is different depending on the state. For these problems
1295   it can be useful to factorize the sparse jacobian at each solver
1296   iteration instead of including all of the zero entries in a single
1297   general factorization.
1298
1299   If your problem does not have this property (or you do not know),
1300   then it is probably best to keep this false, otherwise it will
1301   likely lead to worse performance.
1302
1303   This settings affects the `SPARSE_NORMAL_CHOLESKY` solver.
1304
1305.. member:: int Solver::Options::min_linear_solver_iterations
1306
1307   Default: ``1``
1308
1309   Minimum number of iterations used by the linear solver. This only
1310   makes sense when the linear solver is an iterative solver, e.g.,
1311   ``ITERATIVE_SCHUR`` or ``CGNR``.
1312
1313.. member:: int Solver::Options::max_linear_solver_iterations
1314
1315   Default: ``500``
1316
1317   Minimum number of iterations used by the linear solver. This only
1318   makes sense when the linear solver is an iterative solver, e.g.,
1319   ``ITERATIVE_SCHUR`` or ``CGNR``.
1320
1321.. member:: double Solver::Options::eta
1322
1323   Default: ``1e-1``
1324
1325   Forcing sequence parameter. The truncated Newton solver uses this
1326   number to control the relative accuracy with which the Newton step
1327   is computed. This constant is passed to
1328   ``ConjugateGradientsSolver`` which uses it to terminate the
1329   iterations when
1330
1331   .. math:: \frac{Q_i - Q_{i-1}}{Q_i} < \frac{\eta}{i}
1332
1333.. member:: bool Solver::Options::jacobi_scaling
1334
1335   Default: ``true``
1336
1337   ``true`` means that the Jacobian is scaled by the norm of its
1338   columns before being passed to the linear solver. This improves the
1339   numerical conditioning of the normal equations.
1340
1341.. member:: bool Solver::Options::use_inner_iterations
1342
1343   Default: ``false``
1344
1345   Use a non-linear version of a simplified variable projection
1346   algorithm. Essentially this amounts to doing a further optimization
1347   on each Newton/Trust region step using a coordinate descent
1348   algorithm.  For more details, see :ref:`section-inner-iterations`.
1349
1350.. member:: double Solver::Options::inner_itearation_tolerance
1351
1352   Default: ``1e-3``
1353
1354   Generally speaking, inner iterations make significant progress in
1355   the early stages of the solve and then their contribution drops
1356   down sharply, at which point the time spent doing inner iterations
1357   is not worth it.
1358
1359   Once the relative decrease in the objective function due to inner
1360   iterations drops below ``inner_iteration_tolerance``, the use of
1361   inner iterations in subsequent trust region minimizer iterations is
1362   disabled.
1363
1364.. member:: shared_ptr<ParameterBlockOrdering> Solver::Options::inner_iteration_ordering
1365
1366   Default: ``NULL``
1367
1368   If :member:`Solver::Options::use_inner_iterations` true, then the
1369   user has two choices.
1370
1371   1. Let the solver heuristically decide which parameter blocks to
1372      optimize in each inner iteration. To do this, set
1373      :member:`Solver::Options::inner_iteration_ordering` to ``NULL``.
1374
1375   2. Specify a collection of of ordered independent sets. The lower
1376      numbered groups are optimized before the higher number groups
1377      during the inner optimization phase. Each group must be an
1378      independent set. Not all parameter blocks need to be included in
1379      the ordering.
1380
1381   See :ref:`section-ordering` for more details.
1382
1383.. member:: LoggingType Solver::Options::logging_type
1384
1385   Default: ``PER_MINIMIZER_ITERATION``
1386
1387.. member:: bool Solver::Options::minimizer_progress_to_stdout
1388
1389   Default: ``false``
1390
1391   By default the :class:`Minimizer` progress is logged to ``STDERR``
1392   depending on the ``vlog`` level. If this flag is set to true, and
1393   :member:`Solver::Options::logging_type` is not ``SILENT``, the logging
1394   output is sent to ``STDOUT``.
1395
1396   For ``TRUST_REGION_MINIMIZER`` the progress display looks like
1397
1398   .. code-block:: bash
1399
1400      iter      cost      cost_change  |gradient|   |step|    tr_ratio  tr_radius  ls_iter  iter_time  total_time
1401         0  4.185660e+06    0.00e+00    1.09e+08   0.00e+00   0.00e+00  1.00e+04       0    7.59e-02    3.37e-01
1402         1  1.062590e+05    4.08e+06    8.99e+06   5.36e+02   9.82e-01  3.00e+04       1    1.65e-01    5.03e-01
1403         2  4.992817e+04    5.63e+04    8.32e+06   3.19e+02   6.52e-01  3.09e+04       1    1.45e-01    6.48e-01
1404
1405   Here
1406
1407   #. ``cost`` is the value of the objective function.
1408   #. ``cost_change`` is the change in the value of the objective
1409      function if the step computed in this iteration is accepted.
1410   #. ``|gradient|`` is the max norm of the gradient.
1411   #. ``|step|`` is the change in the parameter vector.
1412   #. ``tr_ratio`` is the ratio of the actual change in the objective
1413      function value to the change in the the value of the trust
1414      region model.
1415   #. ``tr_radius`` is the size of the trust region radius.
1416   #. ``ls_iter`` is the number of linear solver iterations used to
1417      compute the trust region step. For direct/factorization based
1418      solvers it is always 1, for iterative solvers like
1419      ``ITERATIVE_SCHUR`` it is the number of iterations of the
1420      Conjugate Gradients algorithm.
1421   #. ``iter_time`` is the time take by the current iteration.
1422   #. ``total_time`` is the the total time taken by the minimizer.
1423
1424   For ``LINE_SEARCH_MINIMIZER`` the progress display looks like
1425
1426   .. code-block:: bash
1427
1428      0: f: 2.317806e+05 d: 0.00e+00 g: 3.19e-01 h: 0.00e+00 s: 0.00e+00 e:  0 it: 2.98e-02 tt: 8.50e-02
1429      1: f: 2.312019e+05 d: 5.79e+02 g: 3.18e-01 h: 2.41e+01 s: 1.00e+00 e:  1 it: 4.54e-02 tt: 1.31e-01
1430      2: f: 2.300462e+05 d: 1.16e+03 g: 3.17e-01 h: 4.90e+01 s: 2.54e-03 e:  1 it: 4.96e-02 tt: 1.81e-01
1431
1432   Here
1433
1434   #. ``f`` is the value of the objective function.
1435   #. ``d`` is the change in the value of the objective function if
1436      the step computed in this iteration is accepted.
1437   #. ``g`` is the max norm of the gradient.
1438   #. ``h`` is the change in the parameter vector.
1439   #. ``s`` is the optimal step length computed by the line search.
1440   #. ``it`` is the time take by the current iteration.
1441   #. ``tt`` is the the total time taken by the minimizer.
1442
1443.. member:: vector<int> Solver::Options::trust_region_minimizer_iterations_to_dump
1444
1445   Default: ``empty``
1446
1447   List of iterations at which the trust region minimizer should dump
1448   the trust region problem. Useful for testing and benchmarking. If
1449   ``empty``, no problems are dumped.
1450
1451.. member:: string Solver::Options::trust_region_problem_dump_directory
1452
1453   Default: ``/tmp``
1454
1455    Directory to which the problems should be written to. Should be
1456    non-empty if
1457    :member:`Solver::Options::trust_region_minimizer_iterations_to_dump` is
1458    non-empty and
1459    :member:`Solver::Options::trust_region_problem_dump_format_type` is not
1460    ``CONSOLE``.
1461
1462.. member:: DumpFormatType Solver::Options::trust_region_problem_dump_format
1463
1464   Default: ``TEXTFILE``
1465
1466   The format in which trust region problems should be logged when
1467   :member:`Solver::Options::trust_region_minimizer_iterations_to_dump`
1468   is non-empty.  There are three options:
1469
1470   * ``CONSOLE`` prints the linear least squares problem in a human
1471      readable format to ``stderr``. The Jacobian is printed as a
1472      dense matrix. The vectors :math:`D`, :math:`x` and :math:`f` are
1473      printed as dense vectors. This should only be used for small
1474      problems.
1475
1476   * ``TEXTFILE`` Write out the linear least squares problem to the
1477     directory pointed to by
1478     :member:`Solver::Options::trust_region_problem_dump_directory` as
1479     text files which can be read into ``MATLAB/Octave``. The Jacobian
1480     is dumped as a text file containing :math:`(i,j,s)` triplets, the
1481     vectors :math:`D`, `x` and `f` are dumped as text files
1482     containing a list of their values.
1483
1484     A ``MATLAB/Octave`` script called
1485     ``ceres_solver_iteration_???.m`` is also output, which can be
1486     used to parse and load the problem into memory.
1487
1488.. member:: bool Solver::Options::check_gradients
1489
1490   Default: ``false``
1491
1492   Check all Jacobians computed by each residual block with finite
1493   differences. This is expensive since it involves computing the
1494   derivative by normal means (e.g. user specified, autodiff, etc),
1495   then also computing it using finite differences. The results are
1496   compared, and if they differ substantially, details are printed to
1497   the log.
1498
1499.. member:: double Solver::Options::gradient_check_relative_precision
1500
1501   Default: ``1e08``
1502
1503   Precision to check for in the gradient checker. If the relative
1504   difference between an element in a Jacobian exceeds this number,
1505   then the Jacobian for that cost term is dumped.
1506
1507.. member:: double Solver::Options::numeric_derivative_relative_step_size
1508
1509   Default: ``1e-6``
1510
1511   Relative shift used for taking numeric derivatives. For finite
1512   differencing, each dimension is evaluated at slightly shifted
1513   values, e.g., for forward differences, the numerical derivative is
1514
1515   .. math::
1516
1517     \delta &= numeric\_derivative\_relative\_step\_size\\
1518     \Delta f &= \frac{f((1 + \delta)  x) - f(x)}{\delta x}
1519
1520   The finite differencing is done along each dimension. The reason to
1521   use a relative (rather than absolute) step size is that this way,
1522   numeric differentiation works for functions where the arguments are
1523   typically large (e.g. :math:`10^9`) and when the values are small
1524   (e.g. :math:`10^{-5}`). It is possible to construct *torture cases*
1525   which break this finite difference heuristic, but they do not come
1526   up often in practice.
1527
1528.. member:: vector<IterationCallback> Solver::Options::callbacks
1529
1530   Callbacks that are executed at the end of each iteration of the
1531   :class:`Minimizer`. They are executed in the order that they are
1532   specified in this vector. By default, parameter blocks are updated
1533   only at the end of the optimization, i.e when the
1534   :class:`Minimizer` terminates. This behavior is controlled by
1535   :member:`Solver::Options::update_state_every_variable`. If the user
1536   wishes to have access to the update parameter blocks when his/her
1537   callbacks are executed, then set
1538   :member:`Solver::Options::update_state_every_iteration` to true.
1539
1540   The solver does NOT take ownership of these pointers.
1541
1542.. member:: bool Solver::Options::update_state_every_iteration
1543
1544   Default: ``false``
1545
1546   Normally the parameter blocks are only updated when the solver
1547   terminates. Setting this to true update them in every
1548   iteration. This setting is useful when building an interactive
1549   application using Ceres and using an :class:`IterationCallback`.
1550
1551:class:`ParameterBlockOrdering`
1552-------------------------------
1553
1554.. class:: ParameterBlockOrdering
1555
1556   ``ParameterBlockOrdering`` is a class for storing and manipulating
1557   an ordered collection of groups/sets with the following semantics:
1558
1559   Group IDs are non-negative integer values. Elements are any type
1560   that can serve as a key in a map or an element of a set.
1561
1562   An element can only belong to one group at a time. A group may
1563   contain an arbitrary number of elements.
1564
1565   Groups are ordered by their group id.
1566
1567.. function:: bool ParameterBlockOrdering::AddElementToGroup(const double* element, const int group)
1568
1569   Add an element to a group. If a group with this id does not exist,
1570   one is created. This method can be called any number of times for
1571   the same element. Group ids should be non-negative numbers.  Return
1572   value indicates if adding the element was a success.
1573
1574.. function:: void ParameterBlockOrdering::Clear()
1575
1576   Clear the ordering.
1577
1578.. function:: bool ParameterBlockOrdering::Remove(const double* element)
1579
1580   Remove the element, no matter what group it is in. If the element
1581   is not a member of any group, calling this method will result in a
1582   crash.  Return value indicates if the element was actually removed.
1583
1584.. function:: void ParameterBlockOrdering::Reverse()
1585
1586   Reverse the order of the groups in place.
1587
1588.. function:: int ParameterBlockOrdering::GroupId(const double* element) const
1589
1590   Return the group id for the element. If the element is not a member
1591   of any group, return -1.
1592
1593.. function:: bool ParameterBlockOrdering::IsMember(const double* element) const
1594
1595   True if there is a group containing the parameter block.
1596
1597.. function:: int ParameterBlockOrdering::GroupSize(const int group) const
1598
1599   This function always succeeds, i.e., implicitly there exists a
1600   group for every integer.
1601
1602.. function:: int ParameterBlockOrdering::NumElements() const
1603
1604   Number of elements in the ordering.
1605
1606.. function:: int ParameterBlockOrdering::NumGroups() const
1607
1608   Number of groups with one or more elements.
1609
1610
1611:class:`IterationCallback`
1612--------------------------
1613
1614.. class:: IterationSummary
1615
1616   :class:`IterationSummary` describes the state of the minimizer at
1617   the end of each iteration.
1618
1619.. member:: int32 IterationSummary::iteration
1620
1621   Current iteration number.
1622
1623.. member:: bool IterationSummary::step_is_valid
1624
1625   Step was numerically valid, i.e., all values are finite and the
1626   step reduces the value of the linearized model.
1627
1628    **Note**: :member:`IterationSummary::step_is_valid` is `false`
1629    when :member:`IterationSummary::iteration` = 0.
1630
1631.. member::  bool IterationSummary::step_is_nonmonotonic
1632
1633    Step did not reduce the value of the objective function
1634    sufficiently, but it was accepted because of the relaxed
1635    acceptance criterion used by the non-monotonic trust region
1636    algorithm.
1637
1638    **Note**: :member:`IterationSummary::step_is_nonmonotonic` is
1639    `false` when when :member:`IterationSummary::iteration` = 0.
1640
1641.. member:: bool IterationSummary::step_is_successful
1642
1643   Whether or not the minimizer accepted this step or not.
1644
1645   If the ordinary trust region algorithm is used, this means that the
1646   relative reduction in the objective function value was greater than
1647   :member:`Solver::Options::min_relative_decrease`. However, if the
1648   non-monotonic trust region algorithm is used
1649   (:member:`Solver::Options::use_nonmonotonic_steps` = `true`), then
1650   even if the relative decrease is not sufficient, the algorithm may
1651   accept the step and the step is declared successful.
1652
1653   **Note**: :member:`IterationSummary::step_is_successful` is `false`
1654   when when :member:`IterationSummary::iteration` = 0.
1655
1656.. member:: double IterationSummary::cost
1657
1658   Value of the objective function.
1659
1660.. member:: double IterationSummary::cost_change
1661
1662   Change in the value of the objective function in this
1663   iteration. This can be positive or negative.
1664
1665.. member:: double IterationSummary::gradient_max_norm
1666
1667   Infinity norm of the gradient vector.
1668
1669.. member:: double IterationSummary::gradient_norm
1670
1671   2-norm of the gradient vector.
1672
1673.. member:: double IterationSummary::step_norm
1674
1675   2-norm of the size of the step computed in this iteration.
1676
1677.. member:: double IterationSummary::relative_decrease
1678
1679   For trust region algorithms, the ratio of the actual change in cost
1680   and the change in the cost of the linearized approximation.
1681
1682   This field is not used when a linear search minimizer is used.
1683
1684.. member:: double IterationSummary::trust_region_radius
1685
1686   Size of the trust region at the end of the current iteration. For
1687   the Levenberg-Marquardt algorithm, the regularization parameter is
1688   1.0 / member::`IterationSummary::trust_region_radius`.
1689
1690.. member:: double IterationSummary::eta
1691
1692   For the inexact step Levenberg-Marquardt algorithm, this is the
1693   relative accuracy with which the step is solved. This number is
1694   only applicable to the iterative solvers capable of solving linear
1695   systems inexactly. Factorization-based exact solvers always have an
1696   eta of 0.0.
1697
1698.. member:: double IterationSummary::step_size
1699
1700   Step sized computed by the line search algorithm.
1701
1702   This field is not used when a trust region minimizer is used.
1703
1704.. member:: int IterationSummary::line_search_function_evaluations
1705
1706   Number of function evaluations used by the line search algorithm.
1707
1708   This field is not used when a trust region minimizer is used.
1709
1710.. member:: int IterationSummary::linear_solver_iterations
1711
1712   Number of iterations taken by the linear solver to solve for the
1713   trust region step.
1714
1715   Currently this field is not used when a line search minimizer is
1716   used.
1717
1718.. member:: double IterationSummary::iteration_time_in_seconds
1719
1720   Time (in seconds) spent inside the minimizer loop in the current
1721   iteration.
1722
1723.. member:: double IterationSummary::step_solver_time_in_seconds
1724
1725   Time (in seconds) spent inside the trust region step solver.
1726
1727.. member:: double IterationSummary::cumulative_time_in_seconds
1728
1729   Time (in seconds) since the user called Solve().
1730
1731
1732.. class:: IterationCallback
1733
1734   Interface for specifying callbacks that are executed at the end of
1735   each iteration of the minimizer.
1736
1737   .. code-block:: c++
1738
1739      class IterationCallback {
1740       public:
1741        virtual ~IterationCallback() {}
1742        virtual CallbackReturnType operator()(const IterationSummary& summary) = 0;
1743      };
1744
1745
1746  The solver uses the return value of ``operator()`` to decide whether
1747  to continue solving or to terminate. The user can return three
1748  values.
1749
1750  #. ``SOLVER_ABORT`` indicates that the callback detected an abnormal
1751     situation. The solver returns without updating the parameter
1752     blocks (unless ``Solver::Options::update_state_every_iteration`` is
1753     set true). Solver returns with ``Solver::Summary::termination_type``
1754     set to ``USER_FAILURE``.
1755
1756  #. ``SOLVER_TERMINATE_SUCCESSFULLY`` indicates that there is no need
1757     to optimize anymore (some user specified termination criterion
1758     has been met). Solver returns with
1759     ``Solver::Summary::termination_type``` set to ``USER_SUCCESS``.
1760
1761  #. ``SOLVER_CONTINUE`` indicates that the solver should continue
1762     optimizing.
1763
1764  For example, the following :class:`IterationCallback` is used
1765  internally by Ceres to log the progress of the optimization.
1766
1767  .. code-block:: c++
1768
1769    class LoggingCallback : public IterationCallback {
1770     public:
1771      explicit LoggingCallback(bool log_to_stdout)
1772          : log_to_stdout_(log_to_stdout) {}
1773
1774      ~LoggingCallback() {}
1775
1776      CallbackReturnType operator()(const IterationSummary& summary) {
1777        const char* kReportRowFormat =
1778            "% 4d: f:% 8e d:% 3.2e g:% 3.2e h:% 3.2e "
1779            "rho:% 3.2e mu:% 3.2e eta:% 3.2e li:% 3d";
1780        string output = StringPrintf(kReportRowFormat,
1781                                     summary.iteration,
1782                                     summary.cost,
1783                                     summary.cost_change,
1784                                     summary.gradient_max_norm,
1785                                     summary.step_norm,
1786                                     summary.relative_decrease,
1787                                     summary.trust_region_radius,
1788                                     summary.eta,
1789                                     summary.linear_solver_iterations);
1790        if (log_to_stdout_) {
1791          cout << output << endl;
1792        } else {
1793          VLOG(1) << output;
1794        }
1795        return SOLVER_CONTINUE;
1796      }
1797
1798     private:
1799      const bool log_to_stdout_;
1800    };
1801
1802
1803
1804:class:`CRSMatrix`
1805------------------
1806
1807.. class:: CRSMatrix
1808
1809   A compressed row sparse matrix used primarily for communicating the
1810   Jacobian matrix to the user.
1811
1812.. member:: int CRSMatrix::num_rows
1813
1814   Number of rows.
1815
1816.. member:: int CRSMatrix::num_cols
1817
1818   Number of columns.
1819
1820.. member:: vector<int> CRSMatrix::rows
1821
1822   :member:`CRSMatrix::rows` is a :member:`CRSMatrix::num_rows` + 1
1823   sized array that points into the :member:`CRSMatrix::cols` and
1824   :member:`CRSMatrix::values` array.
1825
1826.. member:: vector<int> CRSMatrix::cols
1827
1828   :member:`CRSMatrix::cols` contain as many entries as there are
1829   non-zeros in the matrix.
1830
1831   For each row ``i``, ``cols[rows[i]]`` ... ``cols[rows[i + 1] - 1]``
1832   are the indices of the non-zero columns of row ``i``.
1833
1834.. member:: vector<int> CRSMatrix::values
1835
1836   :member:`CRSMatrix::values` contain as many entries as there are
1837   non-zeros in the matrix.
1838
1839   For each row ``i``,
1840   ``values[rows[i]]`` ... ``values[rows[i + 1] - 1]`` are the values
1841   of the non-zero columns of row ``i``.
1842
1843e.g, consider the 3x4 sparse matrix
1844
1845.. code-block:: c++
1846
1847   0 10  0  4
1848   0  2 -3  2
1849   1  2  0  0
1850
1851The three arrays will be:
1852
1853.. code-block:: c++
1854
1855            -row0-  ---row1---  -row2-
1856   rows   = [ 0,      2,          5,     7]
1857   cols   = [ 1,  3,  1,  2,  3,  0,  1]
1858   values = [10,  4,  2, -3,  2,  1,  2]
1859
1860
1861:class:`Solver::Summary`
1862------------------------
1863
1864.. class:: Solver::Summary
1865
1866   Summary of the various stages of the solver after termination.
1867
1868
1869.. function:: string Solver::Summary::BriefReport() const
1870
1871   A brief one line description of the state of the solver after
1872   termination.
1873
1874.. function:: string Solver::Summary::FullReport() const
1875
1876   A full multiline description of the state of the solver after
1877   termination.
1878
1879.. function:: bool Solver::Summary::IsSolutionUsable() const
1880
1881   Whether the solution returned by the optimization algorithm can be
1882   relied on to be numerically sane. This will be the case if
1883   `Solver::Summary:termination_type` is set to `CONVERGENCE`,
1884   `USER_SUCCESS` or `NO_CONVERGENCE`, i.e., either the solver
1885   converged by meeting one of the convergence tolerances or because
1886   the user indicated that it had converged or it ran to the maximum
1887   number of iterations or time.
1888
1889.. member:: MinimizerType Solver::Summary::minimizer_type
1890
1891   Type of minimization algorithm used.
1892
1893.. member:: TerminationType Solver::Summary::termination_type
1894
1895   The cause of the minimizer terminating.
1896
1897.. member:: string Solver::Summary::message
1898
1899   Reason why the solver terminated.
1900
1901.. member:: double Solver::Summary::initial_cost
1902
1903   Cost of the problem (value of the objective function) before the
1904   optimization.
1905
1906.. member:: double Solver::Summary::final_cost
1907
1908   Cost of the problem (value of the objective function) after the
1909   optimization.
1910
1911.. member:: double Solver::Summary::fixed_cost
1912
1913   The part of the total cost that comes from residual blocks that
1914   were held fixed by the preprocessor because all the parameter
1915   blocks that they depend on were fixed.
1916
1917.. member:: vector<IterationSummary> Solver::Summary::iterations
1918
1919   :class:`IterationSummary` for each minimizer iteration in order.
1920
1921.. member:: int Solver::Summary::num_successful_steps
1922
1923   Number of minimizer iterations in which the step was
1924   accepted. Unless :member:`Solver::Options::use_non_monotonic_steps`
1925   is `true` this is also the number of steps in which the objective
1926   function value/cost went down.
1927
1928.. member:: int Solver::Summary::num_unsuccessful_steps
1929
1930   Number of minimizer iterations in which the step was rejected
1931   either because it did not reduce the cost enough or the step was
1932   not numerically valid.
1933
1934.. member:: int Solver::Summary::num_inner_iteration_steps
1935
1936   Number of times inner iterations were performed.
1937
1938.. member:: double Solver::Summary::preprocessor_time_in_seconds
1939
1940   Time (in seconds) spent in the preprocessor.
1941
1942.. member:: double Solver::Summary::minimizer_time_in_seconds
1943
1944   Time (in seconds) spent in the Minimizer.
1945
1946.. member:: double Solver::Summary::postprocessor_time_in_seconds
1947
1948   Time (in seconds) spent in the post processor.
1949
1950.. member:: double Solver::Summary::total_time_in_seconds
1951
1952   Time (in seconds) spent in the solver.
1953
1954.. member:: double Solver::Summary::linear_solver_time_in_seconds
1955
1956   Time (in seconds) spent in the linear solver computing the trust
1957   region step.
1958
1959.. member:: double Solver::Summary::residual_evaluation_time_in_seconds
1960
1961   Time (in seconds) spent evaluating the residual vector.
1962
1963.. member:: double Solver::Summary::jacobian_evaluation_time_in_seconds
1964
1965   Time (in seconds) spent evaluating the Jacobian matrix.
1966
1967.. member:: double Solver::Summary::inner_iteration_time_in_seconds
1968
1969   Time (in seconds) spent doing inner iterations.
1970
1971.. member:: int Solver::Summary::num_parameter_blocks
1972
1973   Number of parameter blocks in the problem.
1974
1975.. member:: int Solver::Summary::num_parameters
1976
1977   Number of parameters in the problem.
1978
1979.. member:: int Solver::Summary::num_effective_parameters
1980
1981   Dimension of the tangent space of the problem (or the number of
1982   columns in the Jacobian for the problem). This is different from
1983   :member:`Solver::Summary::num_parameters` if a parameter block is
1984   associated with a :class:`LocalParameterization`.
1985
1986.. member:: int Solver::Summary::num_residual_blocks
1987
1988   Number of residual blocks in the problem.
1989
1990.. member:: int Solver::Summary::num_residuals
1991
1992   Number of residuals in the problem.
1993
1994.. member:: int Solver::Summary::num_parameter_blocks_reduced
1995
1996   Number of parameter blocks in the problem after the inactive and
1997   constant parameter blocks have been removed. A parameter block is
1998   inactive if no residual block refers to it.
1999
2000.. member:: int Solver::Summary::num_parameters_reduced
2001
2002   Number of parameters in the reduced problem.
2003
2004.. member:: int Solver::Summary::num_effective_parameters_reduced
2005
2006   Dimension of the tangent space of the reduced problem (or the
2007   number of columns in the Jacobian for the reduced problem). This is
2008   different from :member:`Solver::Summary::num_parameters_reduced` if
2009   a parameter block in the reduced problem is associated with a
2010   :class:`LocalParameterization`.
2011
2012.. member:: int Solver::Summary::num_residual_blocks_reduced
2013
2014   Number of residual blocks in the reduced problem.
2015
2016.. member:: int Solver::Summary::num_residuals_reduced
2017
2018   Number of residuals in the reduced problem.
2019
2020.. member:: int Solver::Summary::num_threads_given
2021
2022   Number of threads specified by the user for Jacobian and residual
2023   evaluation.
2024
2025.. member:: int Solver::Summary::num_threads_used
2026
2027   Number of threads actually used by the solver for Jacobian and
2028   residual evaluation. This number is not equal to
2029   :member:`Solver::Summary::num_threads_given` if `OpenMP` is not
2030   available.
2031
2032.. member:: int Solver::Summary::num_linear_solver_threads_given
2033
2034   Number of threads specified by the user for solving the trust
2035   region problem.
2036
2037.. member:: int Solver::Summary::num_linear_solver_threads_used
2038
2039   Number of threads actually used by the solver for solving the trust
2040   region problem. This number is not equal to
2041   :member:`Solver::Summary::num_linear_solver_threads_given` if
2042   `OpenMP` is not available.
2043
2044.. member:: LinearSolverType Solver::Summary::linear_solver_type_given
2045
2046   Type of the linear solver requested by the user.
2047
2048.. member:: LinearSolverType Solver::Summary::linear_solver_type_used
2049
2050   Type of the linear solver actually used. This may be different from
2051   :member:`Solver::Summary::linear_solver_type_given` if Ceres
2052   determines that the problem structure is not compatible with the
2053   linear solver requested or if the linear solver requested by the
2054   user is not available, e.g. The user requested
2055   `SPARSE_NORMAL_CHOLESKY` but no sparse linear algebra library was
2056   available.
2057
2058.. member:: vector<int> Solver::Summary::linear_solver_ordering_given
2059
2060   Size of the elimination groups given by the user as hints to the
2061   linear solver.
2062
2063.. member:: vector<int> Solver::Summary::linear_solver_ordering_used
2064
2065   Size of the parameter groups used by the solver when ordering the
2066   columns of the Jacobian.  This maybe different from
2067   :member:`Solver::Summary::linear_solver_ordering_given` if the user
2068   left :member:`Solver::Summary::linear_solver_ordering_given` blank
2069   and asked for an automatic ordering, or if the problem contains
2070   some constant or inactive parameter blocks.
2071
2072.. member:: bool Solver::Summary::inner_iterations_given
2073
2074   `True` if the user asked for inner iterations to be used as part of
2075   the optimization.
2076
2077.. member:: bool Solver::Summary::inner_iterations_used
2078
2079   `True` if the user asked for inner iterations to be used as part of
2080   the optimization and the problem structure was such that they were
2081   actually performed. e.g., in a problem with just one parameter
2082   block, inner iterations are not performed.
2083
2084.. member:: vector<int> inner_iteration_ordering_given
2085
2086   Size of the parameter groups given by the user for performing inner
2087   iterations.
2088
2089.. member:: vector<int> inner_iteration_ordering_used
2090
2091   Size of the parameter groups given used by the solver for
2092   performing inner iterations. This maybe different from
2093   :member:`Solver::Summary::inner_iteration_ordering_given` if the
2094   user left :member:`Solver::Summary::inner_iteration_ordering_given`
2095   blank and asked for an automatic ordering, or if the problem
2096   contains some constant or inactive parameter blocks.
2097
2098.. member:: PreconditionerType Solver::Summary::preconditioner_type
2099
2100   Type of preconditioner used for solving the trust region step. Only
2101   meaningful when an iterative linear solver is used.
2102
2103.. member:: VisibilityClusteringType Solver::Summary::visibility_clustering_type
2104
2105   Type of clustering algorithm used for visibility based
2106   preconditioning. Only meaningful when the
2107   :member:`Solver::Summary::preconditioner_type` is
2108   ``CLUSTER_JACOBI`` or ``CLUSTER_TRIDIAGONAL``.
2109
2110.. member:: TrustRegionStrategyType Solver::Summary::trust_region_strategy_type
2111
2112   Type of trust region strategy.
2113
2114.. member:: DoglegType Solver::Summary::dogleg_type
2115
2116   Type of dogleg strategy used for solving the trust region problem.
2117
2118.. member:: DenseLinearAlgebraLibraryType Solver::Summary::dense_linear_algebra_library_type
2119
2120   Type of the dense linear algebra library used.
2121
2122.. member:: SparseLinearAlgebraLibraryType Solver::Summary::sparse_linear_algebra_library_type
2123
2124   Type of the sparse linear algebra library used.
2125
2126.. member:: LineSearchDirectionType Solver::Summary::line_search_direction_type
2127
2128   Type of line search direction used.
2129
2130.. member:: LineSearchType Solver::Summary::line_search_type
2131
2132   Type of the line search algorithm used.
2133
2134.. member:: LineSearchInterpolationType Solver::Summary::line_search_interpolation_type
2135
2136   When performing line search, the degree of the polynomial used to
2137   approximate the objective function.
2138
2139.. member:: NonlinearConjugateGradientType Solver::Summary::nonlinear_conjugate_gradient_type
2140
2141   If the line search direction is `NONLINEAR_CONJUGATE_GRADIENT`,
2142   then this indicates the particular variant of non-linear conjugate
2143   gradient used.
2144
2145.. member:: int Solver::Summary::max_lbfgs_rank
2146
2147   If the type of the line search direction is `LBFGS`, then this
2148   indicates the rank of the Hessian approximation.
2149
2150Covariance Estimation
2151=====================
2152
2153Background
2154----------
2155
2156One way to assess the quality of the solution returned by a
2157non-linear least squares solve is to analyze the covariance of the
2158solution.
2159
2160Let us consider the non-linear regression problem
2161
2162.. math::  y = f(x) + N(0, I)
2163
2164i.e., the observation :math:`y` is a random non-linear function of the
2165independent variable :math:`x` with mean :math:`f(x)` and identity
2166covariance. Then the maximum likelihood estimate of :math:`x` given
2167observations :math:`y` is the solution to the non-linear least squares
2168problem:
2169
2170.. math:: x^* = \arg \min_x \|f(x)\|^2
2171
2172And the covariance of :math:`x^*` is given by
2173
2174.. math:: C(x^*) = \left(J'(x^*)J(x^*)\right)^{-1}
2175
2176Here :math:`J(x^*)` is the Jacobian of :math:`f` at :math:`x^*`. The
2177above formula assumes that :math:`J(x^*)` has full column rank.
2178
2179If :math:`J(x^*)` is rank deficient, then the covariance matrix :math:`C(x^*)`
2180is also rank deficient and is given by the Moore-Penrose pseudo inverse.
2181
2182.. math:: C(x^*) =  \left(J'(x^*)J(x^*)\right)^{\dagger}
2183
2184Note that in the above, we assumed that the covariance matrix for
2185:math:`y` was identity. This is an important assumption. If this is
2186not the case and we have
2187
2188.. math:: y = f(x) + N(0, S)
2189
2190Where :math:`S` is a positive semi-definite matrix denoting the
2191covariance of :math:`y`, then the maximum likelihood problem to be
2192solved is
2193
2194.. math:: x^* = \arg \min_x f'(x) S^{-1} f(x)
2195
2196and the corresponding covariance estimate of :math:`x^*` is given by
2197
2198.. math:: C(x^*) = \left(J'(x^*) S^{-1} J(x^*)\right)^{-1}
2199
2200So, if it is the case that the observations being fitted to have a
2201covariance matrix not equal to identity, then it is the user's
2202responsibility that the corresponding cost functions are correctly
2203scaled, e.g. in the above case the cost function for this problem
2204should evaluate :math:`S^{-1/2} f(x)` instead of just :math:`f(x)`,
2205where :math:`S^{-1/2}` is the inverse square root of the covariance
2206matrix :math:`S`.
2207
2208Gauge Invariance
2209----------------
2210
2211In structure from motion (3D reconstruction) problems, the
2212reconstruction is ambiguous upto a similarity transform. This is
2213known as a *Gauge Ambiguity*. Handling Gauges correctly requires the
2214use of SVD or custom inversion algorithms. For small problems the
2215user can use the dense algorithm. For more details see the work of
2216Kanatani & Morris [KanataniMorris]_.
2217
2218
2219:class:`Covariance`
2220-------------------
2221
2222:class:`Covariance` allows the user to evaluate the covariance for a
2223non-linear least squares problem and provides random access to its
2224blocks. The computation assumes that the cost functions compute
2225residuals such that their covariance is identity.
2226
2227Since the computation of the covariance matrix requires computing the
2228inverse of a potentially large matrix, this can involve a rather large
2229amount of time and memory. However, it is usually the case that the
2230user is only interested in a small part of the covariance
2231matrix. Quite often just the block diagonal. :class:`Covariance`
2232allows the user to specify the parts of the covariance matrix that she
2233is interested in and then uses this information to only compute and
2234store those parts of the covariance matrix.
2235
2236Rank of the Jacobian
2237--------------------
2238
2239As we noted above, if the Jacobian is rank deficient, then the inverse
2240of :math:`J'J` is not defined and instead a pseudo inverse needs to be
2241computed.
2242
2243The rank deficiency in :math:`J` can be *structural* -- columns
2244which are always known to be zero or *numerical* -- depending on the
2245exact values in the Jacobian.
2246
2247Structural rank deficiency occurs when the problem contains parameter
2248blocks that are constant. This class correctly handles structural rank
2249deficiency like that.
2250
2251Numerical rank deficiency, where the rank of the matrix cannot be
2252predicted by its sparsity structure and requires looking at its
2253numerical values is more complicated. Here again there are two
2254cases.
2255
2256  a. The rank deficiency arises from overparameterization. e.g., a
2257     four dimensional quaternion used to parameterize :math:`SO(3)`,
2258     which is a three dimensional manifold. In cases like this, the
2259     user should use an appropriate
2260     :class:`LocalParameterization`. Not only will this lead to better
2261     numerical behaviour of the Solver, it will also expose the rank
2262     deficiency to the :class:`Covariance` object so that it can
2263     handle it correctly.
2264
2265  b. More general numerical rank deficiency in the Jacobian requires
2266     the computation of the so called Singular Value Decomposition
2267     (SVD) of :math:`J'J`. We do not know how to do this for large
2268     sparse matrices efficiently. For small and moderate sized
2269     problems this is done using dense linear algebra.
2270
2271
2272:class:`Covariance::Options`
2273
2274.. class:: Covariance::Options
2275
2276.. member:: int Covariance::Options::num_threads
2277
2278   Default: ``1``
2279
2280   Number of threads to be used for evaluating the Jacobian and
2281   estimation of covariance.
2282
2283.. member:: CovarianceAlgorithmType Covariance::Options::algorithm_type
2284
2285   Default: ``SUITE_SPARSE_QR`` if ``SuiteSparseQR`` is installed and
2286   ``EIGEN_SPARSE_QR`` otherwise.
2287
2288   Ceres supports three different algorithms for covariance
2289   estimation, which represent different tradeoffs in speed, accuracy
2290   and reliability.
2291
2292   1. ``DENSE_SVD`` uses ``Eigen``'s ``JacobiSVD`` to perform the
2293      computations. It computes the singular value decomposition
2294
2295      .. math::   U S V^\top = J
2296
2297      and then uses it to compute the pseudo inverse of J'J as
2298
2299      .. math::   (J'J)^{\dagger} = V  S^{\dagger}  V^\top
2300
2301      It is an accurate but slow method and should only be used for
2302      small to moderate sized problems. It can handle full-rank as
2303      well as rank deficient Jacobians.
2304
2305   2. ``EIGEN_SPARSE_QR`` uses the sparse QR factorization algorithm
2306      in ``Eigen`` to compute the decomposition
2307
2308       .. math::
2309
2310          QR &= J\\
2311          \left(J^\top J\right)^{-1} &= \left(R^\top R\right)^{-1}
2312
2313      It is a moderately fast algorithm for sparse matrices.
2314
2315   3. ``SUITE_SPARSE_QR`` uses the sparse QR factorization algorithm
2316      in ``SuiteSparse``. It uses dense linear algebra and is multi
2317      threaded, so for large sparse sparse matrices it is
2318      significantly faster than ``EIGEN_SPARSE_QR``.
2319
2320   Neither ``EIGEN_SPARSE_QR`` nor ``SUITE_SPARSE_QR`` are capable of
2321   computing the covariance if the Jacobian is rank deficient.
2322
2323.. member:: int Covariance::Options::min_reciprocal_condition_number
2324
2325   Default: :math:`10^{-14}`
2326
2327   If the Jacobian matrix is near singular, then inverting :math:`J'J`
2328   will result in unreliable results, e.g, if
2329
2330   .. math::
2331
2332     J = \begin{bmatrix}
2333         1.0& 1.0 \\
2334         1.0& 1.0000001
2335         \end{bmatrix}
2336
2337   which is essentially a rank deficient matrix, we have
2338
2339   .. math::
2340
2341     (J'J)^{-1} = \begin{bmatrix}
2342                  2.0471e+14&  -2.0471e+14 \\
2343                  -2.0471e+14   2.0471e+14
2344                  \end{bmatrix}
2345
2346
2347   This is not a useful result. Therefore, by default
2348   :func:`Covariance::Compute` will return ``false`` if a rank
2349   deficient Jacobian is encountered. How rank deficiency is detected
2350   depends on the algorithm being used.
2351
2352   1. ``DENSE_SVD``
2353
2354      .. math:: \frac{\sigma_{\text{min}}}{\sigma_{\text{max}}}  < \sqrt{\text{min_reciprocal_condition_number}}
2355
2356      where :math:`\sigma_{\text{min}}` and
2357      :math:`\sigma_{\text{max}}` are the minimum and maxiumum
2358      singular values of :math:`J` respectively.
2359
2360   2. ``EIGEN_SPARSE_QR`` and ``SUITE_SPARSE_QR``
2361
2362       .. math:: \operatorname{rank}(J) < \operatorname{num\_col}(J)
2363
2364       Here :\math:`\operatorname{rank}(J)` is the estimate of the
2365       rank of `J` returned by the sparse QR factorization
2366       algorithm. It is a fairly reliable indication of rank
2367       deficiency.
2368
2369.. member:: int Covariance::Options::null_space_rank
2370
2371    When using ``DENSE_SVD``, the user has more control in dealing
2372    with singular and near singular covariance matrices.
2373
2374    As mentioned above, when the covariance matrix is near singular,
2375    instead of computing the inverse of :math:`J'J`, the Moore-Penrose
2376    pseudoinverse of :math:`J'J` should be computed.
2377
2378    If :math:`J'J` has the eigen decomposition :math:`(\lambda_i,
2379    e_i)`, where :math:`lambda_i` is the :math:`i^\textrm{th}`
2380    eigenvalue and :math:`e_i` is the corresponding eigenvector, then
2381    the inverse of :math:`J'J` is
2382
2383    .. math:: (J'J)^{-1} = \sum_i \frac{1}{\lambda_i} e_i e_i'
2384
2385    and computing the pseudo inverse involves dropping terms from this
2386    sum that correspond to small eigenvalues.
2387
2388    How terms are dropped is controlled by
2389    `min_reciprocal_condition_number` and `null_space_rank`.
2390
2391    If `null_space_rank` is non-negative, then the smallest
2392    `null_space_rank` eigenvalue/eigenvectors are dropped irrespective
2393    of the magnitude of :math:`\lambda_i`. If the ratio of the
2394    smallest non-zero eigenvalue to the largest eigenvalue in the
2395    truncated matrix is still below min_reciprocal_condition_number,
2396    then the `Covariance::Compute()` will fail and return `false`.
2397
2398    Setting `null_space_rank = -1` drops all terms for which
2399
2400    .. math::  \frac{\lambda_i}{\lambda_{\textrm{max}}} < \textrm{min_reciprocal_condition_number}
2401
2402    This option has no effect on ``EIGEN_SPARSE_QR`` and
2403    ``SUITE_SPARSE_QR``.
2404
2405.. member:: bool Covariance::Options::apply_loss_function
2406
2407   Default: `true`
2408
2409   Even though the residual blocks in the problem may contain loss
2410   functions, setting ``apply_loss_function`` to false will turn off
2411   the application of the loss function to the output of the cost
2412   function and in turn its effect on the covariance.
2413
2414.. class:: Covariance
2415
2416   :class:`Covariance::Options` as the name implies is used to control
2417   the covariance estimation algorithm. Covariance estimation is a
2418   complicated and numerically sensitive procedure. Please read the
2419   entire documentation for :class:`Covariance::Options` before using
2420   :class:`Covariance`.
2421
2422.. function:: bool Covariance::Compute(const vector<pair<const double*, const double*> >& covariance_blocks, Problem* problem)
2423
2424   Compute a part of the covariance matrix.
2425
2426   The vector ``covariance_blocks``, indexes into the covariance
2427   matrix block-wise using pairs of parameter blocks. This allows the
2428   covariance estimation algorithm to only compute and store these
2429   blocks.
2430
2431   Since the covariance matrix is symmetric, if the user passes
2432   ``<block1, block2>``, then ``GetCovarianceBlock`` can be called with
2433   ``block1``, ``block2`` as well as ``block2``, ``block1``.
2434
2435   ``covariance_blocks`` cannot contain duplicates. Bad things will
2436   happen if they do.
2437
2438   Note that the list of ``covariance_blocks`` is only used to
2439   determine what parts of the covariance matrix are computed. The
2440   full Jacobian is used to do the computation, i.e. they do not have
2441   an impact on what part of the Jacobian is used for computation.
2442
2443   The return value indicates the success or failure of the covariance
2444   computation. Please see the documentation for
2445   :class:`Covariance::Options` for more on the conditions under which
2446   this function returns ``false``.
2447
2448.. function:: bool GetCovarianceBlock(const double* parameter_block1, const double* parameter_block2, double* covariance_block) const
2449
2450   Return the block of the covariance matrix corresponding to
2451   ``parameter_block1`` and ``parameter_block2``.
2452
2453   Compute must be called before the first call to ``GetCovarianceBlock``
2454   and the pair ``<parameter_block1, parameter_block2>`` OR the pair
2455   ``<parameter_block2, parameter_block1>`` must have been present in the
2456   vector covariance_blocks when ``Compute`` was called. Otherwise
2457   ``GetCovarianceBlock`` will return false.
2458
2459   ``covariance_block`` must point to a memory location that can store
2460   a ``parameter_block1_size x parameter_block2_size`` matrix. The
2461   returned covariance will be a row-major matrix.
2462
2463Example Usage
2464-------------
2465
2466.. code-block:: c++
2467
2468 double x[3];
2469 double y[2];
2470
2471 Problem problem;
2472 problem.AddParameterBlock(x, 3);
2473 problem.AddParameterBlock(y, 2);
2474 <Build Problem>
2475 <Solve Problem>
2476
2477 Covariance::Options options;
2478 Covariance covariance(options);
2479
2480 vector<pair<const double*, const double*> > covariance_blocks;
2481 covariance_blocks.push_back(make_pair(x, x));
2482 covariance_blocks.push_back(make_pair(y, y));
2483 covariance_blocks.push_back(make_pair(x, y));
2484
2485 CHECK(covariance.Compute(covariance_blocks, &problem));
2486
2487 double covariance_xx[3 * 3];
2488 double covariance_yy[2 * 2];
2489 double covariance_xy[3 * 2];
2490 covariance.GetCovarianceBlock(x, x, covariance_xx)
2491 covariance.GetCovarianceBlock(y, y, covariance_yy)
2492 covariance.GetCovarianceBlock(x, y, covariance_xy)
2493