A BRIEF SURVEY OF METHODS FOR SOLVING NONLINEAR LEAST-SQUARES PROBLEMS

NUMERICAL ALGEBRA, doi:10.3934/naco.2019001
CONTROL AND OPTIMIZATION
Volume 9, Number 1, March 2019 pp. 1–13
A BRIEF SURVEY OF METHODS FOR SOLVING
NONLINEAR LEAST-SQUARES PROBLEMS
Hassan Mohammad1 ∗
, Mohammed Yusuf Waziri1
and
Sandra Augusta Santos2
1Department of Mathematical Sciences
Faculty of Physical Sciences
Bayero University, Kano, 700241, Nigeria
2Institute of Mathematics, Statistics and Scientific Computing
University of Campinas, Campinas, SP, 13083-970, Brazil
(Communicated by Hongchao Zhang)
Abstract. In this paper, we present a brief survey of methods for solving
nonlinear least-squares problems. We pay specific attention to methods that
take into account the special structure of the problems. Most of the methods
discussed belong to the quasi-Newton family (i.e. the structured quasi-Newton
methods (SQN)). Our survey comprises some of the traditional and modern de-
veloped methods for nonlinear least-squares problems. At the end, we suggest
a few topics for further research.
1. Introduction. This paper provides a survey of methods to find a minimizer for
the nonlinear least-squares problem (NLS). Specifically, we consider the problem of
determining a minimizer x∗
∈ Rn
to the problem of the form
min
x∈Rn
f(x), f(x) =
1
2
kF(x)k2
=
1
2
m
X
i=1
(Fi(x))2
. (1)
Here F : Rn
→ Rm
(m ≥ n), F(x) = (F1(x), F2(x), ..., Fm(x))T
, Fi : Rn
→ R,
and F is a twice continuously differentiable mapping. This type of problem arises
in many applications such as data fitting, optimal control, parameter estimation,
experimental design, data assimilation, and imaging problems (see, e.g. [12,17,22,
25,28,33,34,56] and references therein). These problems are solved either by general
unconstrained minimization algorithms, or by specialized algorithms that take into
account the particular structure of the objective function f [55,68].
The gradient and Hessian of the objective function (1) have a special form and
are given by
∇f(x) =
m
X
i=1
Fi(x)∇Fi(x) = J(x)T
F(x) (2)
2010 Mathematics Subject Classification. Primary: 93E24, 49M37; Secondary: 65K05, 90C53,
49J52.
Key words and phrases. Nonlinear least squares, structured quasi-Newton methods, sizing
techniques, hybridization methods, derivative-free methods.
∗ Corresponding author: Hassan Mohammad.
1

2 H. MOHAMMAD, M. Y. WAZIRI AND S. A. SANTOS
∇2
f(x) =
m
X
i=1
∇Fi(x)∇Fi(x)T
+
m
X
i=1
Fi(x)∇2
Fi(x) (3)
= J(x)T
J(x) + B(x), (4)
where Fi is the i-th component of F and ∇2
Fi(x) is its Hessian, J(x) = ∇F(x)T
is
the Jacobian of the so-called residual function F at x and B(x) is a square matrix
representing the second-order term of (3).
Basically, the iterative scheme for solving (1) has the general form: given an
initial approximation x0, a sequence of iterates {xk} is obtained via
xk+1 = xk + sk, k = 0, 1, ... (5)
where sk = αkdk, αk is the step length obtained by a suitable line search procedure,
and dk is the search direction. Along the text, we adopt the notation Fk = F(xk),
Jk = J(xk) and gk = ∇f(xk) = J(xk)T
F(xk).
It is vital to notice that the Hessian matrix (3) consists of two parts; the first-
order term contains only the first derivatives of the residual, and the second-order
term contains the second derivatives of the residuals.
There are many methods for solving (1), either first- or second-order based. Most
classical and modified algorithms require the computation of the gradient, and also
demand at least the first order-term of the Hessian matrix of the objective function
(1), which may be costly as dimension increases.
However, if the residue, i.e. the objective function value, is large at the minimizer,
then the computation of the exact second-order derivatives of F, although expensive,
might become relevant. Nonzero residues are present in many practical applications
(see [14,55]). Unfortunately, Newton-type methods that require the exact Hessian
matrix are not suitable for large-scale problems. As a consequence, alternative
approaches, which make use of the first derivative information, or just rest upon
function evaluations have been developed (see [42] for details).
The Gauss-Newton method (GN) [24] is a well-known method for solving (1),
and it requires the computation of the Jacobian Jk of the function Fk, and the
solution of a linear system at each iteration. Provided JT
k Jk is nonsingular, the
Gauss-Newton direction can be written as
dGN
k = −(JT
k Jk)−1
gk, (6)
where gk = JT
k Fk.
Another method for solving (1) is the Levenberg-Marquardt method (LM) [31,
36, 40]. This method adds some positive scalar correction to the term JT
k Jk in
the GN method. The correction is a regularization parameter added specifically to
remedy the possibility of the Jacobian matrix to become rank-deficient. The LM
direction is given by
dLM
k = −(JT
k Jk + µkI)−1
gk, (7)
where µk is a positive regularization or damping parameter.
Both GN and LM methods neglect the second order part of the Hessian matrix
of the objective function (1). As a result, they are expected to perform well when
solving small residual problems. Nevertheless, when solving large residual problems,
these methods may perform poorly [14, 55]. The need for storing the Jacobian
matrix of the GN and LM methods also contributes to their inefficiency in the case
of solving large-scale problems. Thus, derivative-free approaches are needed for
solving such cases (see, for example, [7,29,47,49–51,59–61,69,70]).

SURVEY OF METHODS FOR NONLINEAR LEAST SQUARES 3
Our survey focuses mainly on two categories of methods for solving nonlinear
least-squares problems. The first category involves the use of structured quasi-
Newton methods (SQN), whereas the second one concerns derivative-free methods
for minimizing nonlinear least squares based on the Gauss-Newton method and
hybrid methods. These categories are detailed in Sections 2 and 3, respectively. In
Section 4 we describe additional works that do not fit into these two categories,
mainly concerning miscellaneous methods as well as convergence analysis results.
Final remarks and open research questions are raised in Section 5.
2. Structured Quasi-Newton Methods for Nonlinear Least Squares. Struc-
tured quasi-Newton methods (SQN) were devised to overcome the difficulties that
arise when solving large residual problems using the Gauss-Newton method [44,55].
These methods combine Gauss-Newton and quasi-Newton ideas in order to make
good use of the structure of the Hessian matrix of the objective function. The search
direction is obtained from
(Ak + Bk)dSQN
k = −gk, (8)
where Ak = JT
k Jk, and Bk is an approximation to the second-order part of the
Hessian matrix. The matrix Bk is updated using a suitable quasi-Newton update
such that
Bk+1sk = yk − Ak+1sk (9)
or
Bk+1sk = (Jk+1 − Jk)T
Fk+1, (10)
where sk = xk+1 − xk and yk = gk+1 − gk.
In [8], Brown and Dennis developed an algorithm for nonlinear least-squares curve
fitting. In [13] Dennis presented the structured quasi-Newton method as a technique
for solving large residual NLS problems. Although SQN methods are shown to
be very efficient (especially for large residual problems), there are some aspects
associated with such approaches that require special attention. Firstly, a suitable
quasi-Newton update has to be found, together with secant equations, such as (9)
or (10), listed above. Secondly, there is a demand for algorithms that automatically
switch to the Gauss-Newton method when solving small residual problems, or that
adopt a quasi-Newton update when solving large residual problems.
2.1. Sizing Techniques on SQN. Sizing techniques were introduced as a rem-
edy for the difficulties involved when solving small (or in particular zero) residual
problems using SQN methods. This is because, in general, quasi-Newton updates
do not approach zero matrices; therefore, SQN methods do not perform as well as
the GN and LM methods do for zero residual problems at the solution (i.e. when
F(x∗
) = 0).
In [3], Bartholomew-Biggs considered some SQN methods based on the observa-
tion of McKeown [37] about the incapability of GN-like algorithms on minimizing
some class of functions that are sums of squares. Bartholomew-Biggs was aware of
the shortcomings of the SQN methods when solving small residual problems, so he
proposed a sizing parameter
ζk =
FT
k+1Fk
FT
k Fk
(11)

on the matrix Bk, and updated it using the symmetric rank one (SR1) formula
Bk+1 = ζkBk +
(y∗
k − ζkBksk)(y∗
k − ζkBksk)T
(y∗
k − ζkBksk)T sk
, (12)
where
y∗
k = (Jk+1 − Jk)T
Fk+1. (13)
Dennis et al. [15] proposed another successful SQN algorithm called NL2SOL.
The NL2SOL maintains the secant approximation of the second-order part of the
Hessian and automatically switches to such an approximation when solving large
residual problems. The sizing parameter used by the NL2SOL algorithm is given
by
ζ̂ = min

sT
k y∗
k
sT
k Bksk
, 1

, (14)
where the scalar
sT
k y∗
k
sT
k Bksk
corresponds to the one used by Oren and Luenberger [46]
and Oren [45] in their popular self-sizing quasi-Newton algorithms. The approximat-
ed second-order part of the Hessian matrix in the NL2SOL algorithm was updated
using
Bk+1 = ζ̂kBk +
(y∗
k − ζ̂kBksk)yT
k + yk(y∗
k − ζ̂kBksk)T
sT
k yk
−
sT
k (y∗
k − ζ̂kBksk)
(sT
k yk)2
ykyT
k .
The SQN method with sizing parameter (14) yield an adjustable algorithm because,
if the residual turns to Fk+1 = 0, then y∗
k = 0 and ζ̂k = 0. As a result, the updated
matrix Bk+1 will be the zero matrix.
In [26], Huschens argued that there are no theoretical results backing the com-
monly used sizing techniques discussed above. He proposed a self-sizing strategy on
the SQN approach (and named it a totally structured secant method) that converges
locally and superlinearly for non-zero residual problems and converges locally and
quadratically for zero residual problems. Specifically, he dealt with the convex class
of the Broyden’s family.
Instead of obtaining the SQN direction using equation (8), Huschens proposed
an alternative way to compute the direction using
(Ak + kFkkBk)dSQN
k = −gk, (15)
where the matrix Bk =
m
X
i=1
Fi(x)
kFi(x)k
∇2
Fi(x) is the k-th approximation of the second-
order part of the Hessian. The matrix Bk is updated so that the secant equation
Bk+1sk = ŷk (16)
is satisfied, where ŷk = Ak+1sk + kFk+1k
kFkk y∗
k and y∗
k is given in Equation (13).
Another version of the convergence proof by Huschens was given by Yabe and
Ogasawara in [63], where they extend the result from a restricted convex class to a
wider class of the structured Broyden’s family.
In [73], Zhang et al. take advantage of the good convergence properties of the
totally structured secant method of Huschens and proposed a modified quasi-Newton
equation for nonlinear least-squares problems. Based on the numerical experiments
presented in the paper, the performance of the Zhang et al. modified secant equation
was shown to be better than the well-known structured secant equations (9) and
(10).

In summary, the introduction of SQN methods reduced some of the drawback-
s concerning the minimization of large residual problems. These approaches con-
tributed to choosing the best secant equation from (9) and (10) for the quasi-Newton
update (see [15]). Sizing techniques on SQN provide an environment for developing
switching algorithms that make use of the good performance of the Gauss-Newton
method when dealing with zero residual problems.
2.2. Hybridization Methods for Solving Nonlinear Least Squares. Hybrid
methods for minimizing nonlinear least-squares problems take advantage of two
approaches in just one algorithm. In [37], McKeown illustrates several examples of
the quasi-Newton methods that do not take into consideration the special form of
the Hessian matrix (3), but even though, perform much better than GN and LM
methods.
Betts in [5] developed a hybrid algorithm for solving nonlinear least-squares prob-
lems. Betts algorithm considered the estimate for the Hessian matrix as the sum of
the first and second order term, where the first order term is the usual Gauss-Newton
method and the second-order term is obtained using a rank one quasi-Newton up-
dating formula.
Based on the arguments of McKeown, Nazareth [42] reviewed some of the meth-
ods that implicitly or explicitly considered approximating the second order term of
the Hessian matrix. In [41], Nazareth presented his contribution for minimizing a
sum of squares of nonlinear functions. He considered a convex combination of the
first order term and a matrix obtained by adopting a quasi-Newton approximation
on the second-order term of the Hessian. The Hessian approximation proposed by
Nazareth is given by
Hk = αkAk + (1 − αk)Bk + λkI, (17)
where αk ∈ [0, 1], λk 0 is a regularization parameter, and I is the identity matrix
in Rn×n
. The search direction is obtained by solving
Hkdk = −gk. (18)
It can be observed that if αk = 1, the resulting direction corresponds to the
Levenberg-Marquardt direction, while if αk = 0, we have the regularized quasi-
Newton direction. Thus, the direction obtained by solving (18) contains the clas-
sical Levenberg-Marquardt, and the regularized quasi-Newton directions as special
cases.
Al-Baali and Fletcher [2] considered various possibilities for hybridizing the
Gauss-Newton method with quasi-Newton updates for nonlinear least squares. They
used the Broyden-Fletcher-Goldfarb-Shanno (BFGS) update in approximating the
second-order term of the Hessian matrix. The authors have also used a switching
criterion that measures the difference between the current Hessian and any positive
definite matrix. Depending on the size of this measure, their algorithms switch to
either the GN or the BFGS direction. The matrix updating is defined by
Bk+1 =
(
Bk +
y∗
ky∗T
k
sT
k y∗
k
−
BksksT
k Bk
sT
k Bksk
, if∆(Bk) ∆(Ak)
Ak+1, otherwise,
where ∆(.) is the Hessian error measure.
In [16], Fletcher and Xu pointed out that the rate of convergence of the algorithm
proposed in [2] is not superlinear as claimed. They argued that the measure ∆(.)
does not, in general, distinguish between zero and non-zero residual problems, they

also constructed a counter-example to support their observations (details can be
found in [16]). In view of this, Fletcher and Xu proposed a general and most
efficient algorithm with a switching strategy that uses the condition
(fk − fk+1) ≤ ǫfk, (19)
where 0 ǫ 1. If (19) is satisfied then the algorithm switches to SQN direction,
otherwise the GN direction is used by the algorithm. Therefore, the algorithm
automatically adopts the GN direction for zero residual problems, and the SQN
direction for non-zero residual problems.
Many studies have been made to the development of hybrid SQN methods (see,
for example, [1,35,38,57]). Spedicato and Vespucci [54] presented an extensive set
of numerical experiments concerning 24 hybrid algorithms using the test problems
introduced by McKeown [37]. The results show that most of the hybrid SQN meth-
ods performed well especially for large residual problems, with better results than
those of the specialized GN method.
2.3. Factorized Versions of Structured Quasi-Newton Methods. One of the
issues associated with SQN methods is the need of the approximated Hessian matrix
to maintain its positive definiteness so that a line search approach can be used to
achieve global convergence. In an attempt to overcome this difficulty, Yabe and
Takahashi [64–66], and Sheng and Zou [53], independently, proposed a factorized
version of SQN methods which maintains the positive definiteness of the Hessian
matrix in the system Hkdk = −gk. In the factorized version of SQN, the search
direction is given by
dk = −[(Lk + Jk)T
(Lk + Jk)]−1
gk, (20)
where the matrix Lk (called L-update) is the same size as the Jacobian matrix Jk,
acting as its correction and expecting to generate a descent direction.
In [65], Yabe and Takahashi studied the modified version of Sheng and Zou factor-
ized SQN method (called SZ-update), creating a generalized update which encom-
passes the SZ-update as a special case. By applying sizing techniques, Yabe [62], fur-
ther proposed the factorized version of the structured Broyden-like family. In [67],
Yabe and Yamaki established and proved the local and superlinear convergence of
the proposed methods.
In [72], Zhang et al. considered sized factorized secant methods. Interestingly,
their methods ensure quadratic convergence for zero residual problems, and super-
linear convergence for nonzero residual problems. Instead of computing the search
direction using (20), they consider
dk = −[(kFkkLk + Jk)T
(kFkkLk + Jk)]−1
gk, (21)
where the factor kFkk acts as the sizing factor of the matrix Lk, in which kFkkLk
approaches zero for zero residual problems, even if the Lk update does not approach
the zero matrix.
Zhou and Chen [75] proposed a global hybrid Gauss-Newton structured BFGS
method that automatically reduces to Gauss-Newton step for zero residual problem-
s, and to the structured BFGS step for nonzero residual problems. Unlike Fletcher
and Xu, Zhou and Chen proposed a switching strategy for which
zT
k sk
kskk2
≥ ǫ, (22)

where 0 ǫ ≤ 1
2 λmin(H(x∗
)) is a constant and λmin(H(x∗
)) is the smallest eigen-
value of the Hessian H at the minimizer x∗
. zk is given by
zk = y∗
k
kFk+1k
kFkk
.
Condition (22) plays a role similar to (19); in addition, it can be used to ensure
that the update matrix in the BFGS step is safely positive definite.
Wang et al. [58] exploited the idea of the modified BFGS method for non-convex
minimization proposed by Li and Fukushima [32], and proposed a hybrid Gauss-
Newton structured modified BFGS method for solving nonlinear least-squares prob-
lems.
The hybrid Gauss-Newton structured quasi-Newton algorithms show compelling
and improved numerical performance compared to the classical algorithm for solv-
ing nonlinear least squares problems. However, their performance on large-scale
problems is very poor because they require a considerable large amount of storage.
Based on this drawback, matrix-free approaches are preferable for solving large-scale
nonlinear least-squares problems (cf. [39]).
3. Derivative-free Methods for Minimizing Sum of Squares of Functions.
Besides Gauss-Newton, Levenberg-Marquardt and SQN methods, other strategies
based on the derivative-free approach for solving nonlinear least squares problems
exist. These methods are intentionally developed to avoid computing the first de-
rivative of the function when solving (1). To the best of our knowledge, the first
derivative-free algorithm for nonlinear least-squares problems was devised in 1965
by Powell [49]. Powell’s algorithm is similar to the Gauss-Newton step, except for
the fact that the former approximates the first derivative along a set of n linearly
independent directions rather than along the coordinate axis.
Peckham [47] also proposed a method for minimizing a function, given by the
sum of squares, without computing the gradient. Peckham considered the different
approach of approximating the first and second order terms of (3). In [7], Brown and
Dennis proposed two derivative-free algorithms for minimizing sums of the squares
of nonlinear functions. These methods are simply the analogs of the Gauss-Newton
and Levenberg-Marquardt methods where the Jacobian matrix is approximated
using finite difference techniques. The direction of these methods is given by
dk = −(DT
k Dk + µkI)−1
DkFk, (23)
where the matrix Dk is componentwise computed by finite differences as
Dij
k =
Fi(xk + hj
kuj) − Fi(xk)
hj
k
with uj the j-th unit column vector and µk and hj
k given by
µk = ckFkk,
with c =





10, if 10 ≤ kFkk
1, if 1 kFkk 10
0.01, if kFkk ≤ 1,
and hj
k = min{kFkk, δj
k}.

Notice that hk and δk are n dimensional vectors, the index j denotes the j-th
component of these vectors, and
δj
k =
(
10−9
, |xj
k| 10−6
0.001|xj
k|, otherwise.
(24)
Ralston and Jennrich [50] developed another derivative-free algorithm for mini-
mizing sums of squares of a nonlinear function called DUD, for “don’t use deriva-
tives”. The algorithm is designed specifically for data fitting applications in which
the computations of the derivatives are too expensive. Ralston and Jennrich showed
that their approach is competitive with the best existing derivative-free algorithms
at that time, like Powell’s [49] and Peckham’s [47].
A trial to develop a derivative-free hybrid SQN method for nonlinear least-squares
problems was given by Xu in [59]. Xu proposed a modification to the hybrid Gauss-
Newton BFGS method in such a way that the Broyden’s rank-one update is im-
plemented for the algorithm to perform the Gauss-Newton step, whereas for the
algorithm to take the BFGS step, the Jacobian matrix associated with the gradient
is estimated using the finite difference operator. A good numerical behavior was
observed from the algorithm, possibly because it combines the best features of the
GN-Broyden and the finite-difference BFGS methods.
Working from a different perspective, but also as an attempt to develop a
derivative-free algorithm for solving nonlinear least-squares problems, Shakhno and
Gnatyshyn [52] approximated the Jacobian matrix in the Gauss-Newton step using
a divided differences operator. Following the ideas of Shakhno and Gnatyshyn, Ren
et al. [51] improved the local convergence result determining its radius of conver-
gence.
Zhang et al. [70] and Zhang and Conn [69] developed a class of derivative-free
algorithms for least-squares minimization. By taking advantage of the structure of
the problem, they built polynomial interpolation models for each of the residual
functions. The global convergence result of the proposed technique was obtained
using the trust-region framework. It is worth mentioning that this new framework
of derivative-free algorithms for solving nonlinear least-squares problems shows an
excellent numerical performance compared with the algorithms that approximate
the gradients using finite differences (cf. [69,70]).
Recently, by taking advantage of the development of Automatic Differentiation
(AD) technology [23], Xu et al. [60] proposed an efficient quasi-Newton approach for
solving nonlinear least-squares problems. The method uses the application of AD
to directly and efficiently compute the gradient of the objective function without
computing the explicit Jacobian matrix, which is sometimes expensive. The Jaco-
bian matrix appearing in the first order term of the Hessian matrix is approximated
using the Broyden matrix Bk [9] updated in each iteration by the Broyden rank-one
formula
Bk+1 = Bk +
(vk − Bksk)sT
k
sT
k sk
, (25)
such that Bk+1sk = vk, where vk = Fk+1 − Fk.
A fully Jacobian-free method for minimizing nonlinear least squares was recently
proposed by Xu et al. [61]. This new approach was realized by the novel combination
of preconditioning and AD techniques. In comparison to the previously discussed
approaches, this new approach does not depend on the structure of the gradient

and the Hessian matrix of the objective function, making it suitable for general
large-scale unconstrained optimization problems.
4. Some Miscellaneous Methods for Nonlinear Least Squares. Despite the
fact that our survey is mainly on SQN methods, there exist some interesting methods
in the literature that are worth mentioning. We list some of them here.
Gould et al. [20] introduced an algorithm for nonlinear least-squares problems us-
ing the combination of filter and trust region techniques. Bellavia et al. [4] presented
a regularized Euclidean residual algorithm for nonlinear least squares and nonlin-
ear equations. They proved the global and quadratic convergent of the algorithm
without the non-singularity assumption of the Jacobian.
Although the structured nonlinear conjugate gradient methods are hardly stud-
ied, Kobayashi et al. [30] exploits the special structure of the Hessian of the nonlinear
least-squares objective function and introduced some conjugate-gradient method-
s with structured secant condition for nonlinear least-squares problems. For box-
constrained nonlinear least-squares problems, an inexact Gauss-Newton trust-region
method was proposed by Porcelli in [48].
In another attempt, Gonçalves and Santos [19] proposed a simple spectral correc-
tion for the GN method, by approximating the residual Hessian in the second-order
part of (3) using a sign-free spectral parameter. They showed that their approach
is locally convergent when applied to the class of quadratic residual problems. In
addition, convergence results were given for problems without ensured convergence
of the GN method. Subsequently, Gonçalves and Santos [18] extended the idea, pro-
viding global convergence results to the case of non-quadratic residuals, by adopting
the nonmonotone line search strategy of Zhang and Hager [71] as the globalization
tool. Specifically, the direction of the approach of Gonçalves and Santos is obtained
by solving
(JT
k Jk + µkI)dk = −gk, (26)
where the parameter µk is the spectral coefficient given by
µk =
sT
k (Jk+1 − Jk)T
Fk+1
sT
k sk
. (27)
In [27], Karas et al. have devised explicit algebraic rules for computing the
regularization parameter of the LM method, with global and local convergence
analysis. Resting upon the regularizing aspect of the LM method, further analysis
of the performance of methods with second- and higher-order regularization terms
have been developed (see, e.g. [6,10,11,21,43,74]).
5. Conclusions. We have briefly reviewed methods for solving nonlinear least-
squares problems, paying special attention to the structured quasi-Newton methods.
The interested reader may refer to Ya-Xiang Yuan [68] review paper for some recent
numerical methods based on Levenberg-Marquardt type, quasi-Newton’s type, and
trust-region algorithm.
Due to the nature of these problems, little attention was given to developing
derivative-free methods for finding their solution. Based on this observation, we
recommend further research on derivative-free methods for large-scale nonlinear
least-squares problems.
Finally, we hope this small review will help students and researchers familiarize
themselves with the required literature needed to understand classical and modern
methods for solving nonlinear least-squares problems.

We conclude this brief survey stating some open (based on this survey and to the
best of our knowledge) research topics:
1. It seems that hybrid GN-SQN algorithms perform better than the individual
GN or SQN algorithm in practice. Is it possible to develop a single algorithm
that solves nonlinear least-squares problems with better efficiency than the
hybrid algorithm?
2. The hybrid algorithm proposed by Nazareth [41], adaptively solves nonlinear
least-squares problems. We recommend further research on choosing a best
convex combination parameter and, at the same time developing the conver-
gence analysis of the algorithm.
3. Considering the good convergence property of the class of SQN algorithms,
more research is needed to develop a simplified version of factorized SQN
algorithms.
4. Due to the advancement of automatic differentiation (AD), the efficiency of
the derivative-free hybrid method presented by Xu in [59] may likely improve
by incorporating AD techniques.
5. Huschens-like methods for nonlinear least squares exhibit good local conver-
gence properties. Does this property hold globally?
6. For large-scale problems, assembling a complete Hessian matrix may cause
insufficient memory for storage. Keeping this in mind, is it possible to develop
an algorithm that does not require the space for storing a Hessian matrix for
minimizing the sums of squares of functions, considering its structure?
Acknowledgements. We would like to thank three anonymous referees for their
valuable comments and for recommending additional references that enhance the
quality of this review paper. We also wish to thank Hasan Shahid, of Brown Uni-
versity, USA, for his careful reading and helpful comments.
REFERENCES
[1] M. Al-Baali, Quasi-newton algorithms for large-scale nonlinear least-squares, in High Perfor-
mance Algorithms and Software for Nonlinear Optimization, Springer, (2003), 1–21.
[2] M. Al-Baali and R. Fletcher, Variational methods for non-linear least-squares, J. Oper. Res.
Soc., 36 (1985), 405–421.
[3] M. C. Bartholomew-Biggs, The estimation of the Hessian matrix in nonlinear least squares
problems with non-zero residuals, Math. Program., 12 (1977), 67–80.
[4] S. Bellavia, C. Cartis, N. I. M. Gould, B. Morini and P. L. Toint, Convergence of a regularized
Euclidean residual algorithm for nonlinear least-squares, SIAM J. Numer. Anal., 48 (2010),
1–29.
[5] J. T. Betts, Solving the nonlinear least square problem: Application of a general method, J.
Optim. Theory Appl., 18 (1976), 469–483.
[6] E. G. Birgin, J. L. Gardenghi, J. M. Martı́nez, S. A. Santos and P. L. Toint, Worst-case
evaluation complexity for unconstrained nonlinear optimization using high-order regularized
models, Math. Program., 163 (2017), 359–368.
[7] K. M. Brown and J. E. Dennis, Derivative-free analogues of the Levenberg-Marquardt and
Gauss algorithms for nonlinear least squares approximation, Numer. Math., 18 (1971), 289–
297.
[8] K. M. Brown and J. E. Dennis, A new algorithm for nonlinear least-squares curve fitting, in
Mathematical Software (ed. J. R. Rice), Academic Press, New York, (1971), 391–396.
[9] C. G. Broyden, A class of methods for solving nonlinear simultaneous equations, Math. Com-
put., 19 (1965), 577–593.
[10] C. Cartis, N. I. M. Gould and P. L. Toint, On the evaluation complexity of cubic regularization
methods for potentially rank-deficient nonlinear least-squares problems and its relevance to
constrained nonlinear optimization, SIAM J. Optim., 23 (2013), 1553–1574.

[11] C. Cartis, N. I. M. Gould and P. L. Toint, On the evaluation complexity of constrained
nonlinear least-squares and general constrained nonlinear optimization using second-order
methods, SIAM J. Numer. Anal., 53 (2015), 836–851.
[12] A. Cornelio, Regularized nonlinear least squares methods for hit position reconstruction in
small gamma cameras, Appl. Math. Comput., 217 (2011), 5589–5595.
[13] J. E. Dennis, Some computational techniques for the nonlinear least squares problem, in
Numerical Solution of Systems of Nonlinear Algebraic Equations (eds. G. D. Byrne and
C. A. Hall), Academic Press, New York, (1973), 157–183.
[14] J. E. Dennis and R. B. Schnabel Jr, Numerical Methods for Unconstrained Optimization and
Nonlinear Equations, Prentice-Hall, 1983.
[15] J. E. Dennis Jr, D. M. Gay and R. E. Walsh, An adaptive nonlinear least-squares algorithm,
ACM Trans. Math. Software (TOMS), 7 (1981), 348–368.
[16] R. Fletcher and C. Xu, Hybrid methods for nonlinear least squares, IMA J. Numer. Anal., 7
(1987), 371–389.
[17] G. Golub and V. Pereyra, Separable nonlinear least squares: the variable projection method
and its applications, Inverse Problems, 19 (2003), R1–R26.
[18] D. S. Gonçalves and S. A. Santos, A globally convergent method for nonlinear least-squares
problems based on the Gauss-Newton model with spectral correction, Bull. Comput. Appl.
Math., 4 (2016), 7–26.
[19] D. S. Gonçalves and S. A. Santos, Local analysis of a spectral correction for the Gauss-Newton
model applied to quadratic residual problems, Numer. Algorithms, 73 (2016), 407–431.
[20] N. Gould, S. Leyffer and P. L. Toint, A multidimensional filter algorithm for nonlinear equa-
tions and nonlinear least-squares, SIAM J. Optim., 15 (2004), 17–38.
[21] G. N. Grapiglia, J. Yuan and Y. X. Yuan, On the convergence and worst-case complexity
of trust-region and regularization methods for unconstrained optimization, Math. Program.,
152 (2015), 491–520.
[22] S. Gratton, A. S. Lawless and N. K. Nichols, Approximate Gauss-Newton methods for non-
linear least squares problems, SIAM J. Optim., 18 (2007), 106–132.
[23] A. Griewank and A. Walther, Evaluating Derivatives: Principles and Techniques of Algorith-
mic Differentiation, SIAM, Vol. 105, 2008.
[24] H. O. Hartley, The modified Gauss-Newton method for the fitting of non-linear regression
functions by least squares, Technometrics, 3 (1961), 269–280.
[25] S. Henn, A Levenberg-Marquardt scheme for nonlinear image registration, BIT, 43 (2003),
743–759.
[26] J. Huschens, On the use of product structure in secant methods for nonlinear least squares
problems, SIAM J. Optim., 4 (1994), 108–129.
[27] E. W. Karas, S. A. Santos and B. F. Svaiter, Algebraic rules for computing the regularization
parameter of the Levenberg-Marquardt method, Comput. Optim. Appl., 65 (2016), 723–751.
[28] S. Kim, K. Koh, M. Lustig, S. Boyd and D. Gorinevsky, An interior-point method for large-
scale ℓ1-regularized least squares, IEEE. J. Sel. Top. Signa., 1 (2007), 606–617.
[29] D. A. Knoll and D. E. Keyes, Jacobian-free Newton–Krylov methods: a survey of approaches
and applications, J. Comput. Phys., 193 (2004), 357–397.
[30] M. Kobayashi, Y. Narushima and H. Yabe, Nonlinear conjugate gradient methods with struc-
tured secant condition for nonlinear least squares problems, J. Comput. Appl. Math., 234
(2010), 375–397.
[31] K. Levenberg, A method for the solution of certain non-linear problems in least squares,
Quarter. Appl. Math., 2 (1944), 164–168.
[32] D.-H. Li and M. Fukushima, A modified BFGS method and its global convergence in non-
convex minimization, J. Comput. Appl. Math., 129 (2001), 15–35.
[33] J. Li, F. Ding and G. Yang, Maximum likelihood least squares identification method for input
nonlinear finite impulse response moving average systems, Math. Comput. Model., 55 (2012),
442–450.
[34] D. C. López, T. Barz, S. Korkel and G. Wozny, Nonlinear ill-posed problem analysis in model-
based parameter estimation and experimental design, Comput. Chem. Eng., 77 (2015), 24–42.
[35] L. Lukšan, Hybrid methods for large sparse nonlinear least squares, J. Optim. Theory Appl.,
89 (1996), 575–595.
[36] D. W. Marquardt, An algorithm for least-squares estimation of nonlinear parameters, J. Soc.
Ind. Appl. Math., 11 (1963), 431–441.

[37] J. J. McKeown, Specialised versus general-purpose algorithms for minimising functions that
are sums of squared terms, Math. Program., 9 (1975), 57–68.
[38] F. Modarres, M. A. Hassan and W. J. Leong, Structured symmetric rank-one method for
unconstrained optimization, Int. J. Comput. Math., 88 (2011), 2608–2617.
[39] H. Mohammad and S. A. Santos, A structured diagonal Hessian approximation method with
evaluation complexity analysis for nonlinear least squares, Computational and Applied Math-
ematics, online.
[40] D. D. Morrison, Methods for nonlinear least squares problems and convergence proofs, in
Proceedings of the Seminar on Tracking Programs and Orbit Determination (eds. J. Lorell
and F. Yagi), Jet Propulsion Laboratory, Pasadena, (1960), 1–9.
[41] L. Nazareth, An Adaptive Method for Minimizing a Sum of Squares of Nonlinear Functions,
NASA Report WP-83-99, 1983.
[42] L. Nazareth, Some recent approaches to solving large residual nonlinear least squares prob-
lems, SIAM Review, 22 (1980), 1–11.
[43] Y. Nesterov, Modified Gauss-Newton scheme with worst case guarantees for global perfor-
mance, Optim. Methods Softw., 22 (2007), 469–483.
[44] J. Nocedal and S. J. Wright, Numerical Optimization, Springer Science, 2006.
[45] S. S. Oren, On the selection of parameters in self scaling variable metric algorithms, Math.
Program., 7 (1974), 351–367.
[46] S. S. Oren and D. G. Luenberger, Self-scaling variable metric (SSVM) algorithms: Part I:
Criteria and sufficient conditions for scaling a class of algorithms, Management Science, 20
(1974), 845–862.
[47] G. Peckham, A new method for minimising a sum of squares without calculating gradients,
The Comput. J., 13 (1970), 418–420.
[48] M. Porcelli, On the convergence of an inexact Gauss–Newton trust-region method for nonlin-
ear least-squares problems with simple bounds, Optim. Lett., 7 (2013), 447–465.
[49] M. J. D. Powell, A method for minimizing a sum of squares of non-linear functions without
calculating derivatives, The Comput. J., 7 (1965), 303–307.
[50] M. L. Ralston and R. I. Jennrich, DUD, A derivative-free algorithm for nonlinear least squares,
Technometrics, 20 (1978), 7–14.
[51] H. Ren, I. K. Argyros and S. Hilout, A derivative-free iterative method for solving least
squares problems, Numer. Algorithms, 58 (2011), 555–571.
[52] S. M. Shakhno and O. P. Gnatyshyn, On an iterative algorithm of order 1.839 . . . for solving
the nonlinear least squares problems, Appl. Math. Comput., 161 (2005), 253–264.
[53] S. Sheng and Z. Zou, A New Secant Method for Nonlinear Least Squares Problems, Report
NANOG 1988-03, Nanjing University, China, 1988.
[54] E. Spedicato and M. T. Vespucci, Numerical experiments with variations of the Gauss-Newton
algorithm for nonlinear least squares, J. Optim. Theory Appl., 57 (1988), 323–339.
[55] W. Sun and Y. X. Yuan, Optimization Theory and Methods: Nonlinear Programming,
Springer Science, 2006.
[56] L. Tang, A regularization homotopy iterative method for ill-posed nonlinear least squares
problem and its application, in Advances in Civil Engineering, ICCET 2011, vol. 90 of
Applied Mechanics and Materials, Trans Tech Publications, 2011, 3268–3273.
[57] P. L. Toint, On large scale nonlinear least squares calculations, SIAM J. Sci. Stat. Comput.,
8 (1987), 416–435.
[58] F. Wang, D. H. Li and L. Qi, Global convergence of Gauss-Newton-MBFGS method for
solving the nonlinear least squares problem, Advan. Model. Optim., 12 (2010), 1–19.
[59] C. X. Xu, Hybrid method for nonlinear least-square problems without calculating derivatives,
J. Optim. Theory Appl., 65 (1990), 555–574.
[60] W. Xu, T. F. Coleman and G. Liu, A secant method for nonlinear least-squares minimization,
Comput. Optim. Appl., 51 (2012), 159–173.
[61] W. Xu, N. Zheng and K. Hayami, Jacobian-free implicit inner-iteration preconditioner for
nonlinear least squares problems, J. Sci. Comput., 68 (2016), 1055–1081.
[62] H. Yabe, Variations of structured Broyden families for nonlinear least squares problems,
Optim. Methods Softw., 2 (1993), 107–144.
[63] H. Yabe and H. Ogasawara, Quadratic and superlinear convergence of the Huschen’s method
for nonlinear least squares problems, Comput. Optim. Appl., 10 (1998), 79–103.
[64] H. Yabe and T. Takahashi, Structured quasi-newton methods for nonlinear least squares
problems, TRU Math., 24 (1988), 195–209.

[65] H. Yabe and T. Takahashi, Factorized quasi-newton methods for nonlinear least squares
problems, Math. Program., 51 (1991), 75–100.
[66] H. Yabe and T. Takahashi, Numerical Comparison among structured quasi-Newton methods
for nonlinear least squares problems, J. Oper. Res. Soc. Japan, 34 (1991), 287–305.
[67] H. Yabe and N. Yamaki, Convergence of a factorized Broyden-like family for nonlinear least
squares problems, SIAM J. Optim., 5 (1995), 770–791.
[68] Y. X. Yuan, Recent advances in numerical methods for nonlinear equations and nonlinear
least squares, Numer. Algebr., Contr. Optim., 1 (2011), 15–34.
[69] H. Zhang and A. R. Conn, On the local convergence of a derivative-free algorithm for least-
squares minimization, Comput. Optim. Appl., 51 (2012), 481–507.
[70] H. Zhang, A. R. Conn and K. Scheinberg, A derivative-free algorithm for least-squares mini-
mization, SIAM J. Optim., 20 (2010), 3555–3576.
[71] H. Zhang and W. W. Hager, A nonmonotone line search technique and its application to
unconstrained optimization, SIAM J. Optim., 14 (2004), 1043–1056.
[72] J. Z. Zhang, L. H. Chen and N. Y. Deng, A family of scaled factorized Broyden-like methods
for nonlinear least squares problems, SIAM J. Optim., 10 (2000), 1163–1179.
[73] J. Zhang, Y. Xue and K. Zhang, A structured secant method based on a new quasi-newton
equation for nonlinear least squares problems, BIT, 43 (2003), 217–229.
[74] R. Zhao and J. Fan, Global complexity bound of the Levenberg-Marquardt method, Optim.
Methods Softw., 31 (2016), 805–814.
[75] W. Zhou and X. Chen, Global convergence of new hybrid Gauss-Newton structured BFGS
method for nonlinear least squares problems, SIAM J. Optim., 20 (2010), 2422–2441.
Received September 2017; 1st
revision April 2018; final revision May 2018.
E-mail address: hmuhd.mth@buk.edu.ng
E-mail address: mywaziri.mth@buk.edu.ng
E-mail address: sandra@ime.unicamp.br

A BRIEF SURVEY OF METHODS FOR SOLVING NONLINEAR LEAST-SQUARES PROBLEMS

More Related Content

Similar to A BRIEF SURVEY OF METHODS FOR SOLVING NONLINEAR LEAST-SQUARES PROBLEMS (20)

More from Shannon Green (20)

Recently uploaded (20)

A BRIEF SURVEY OF METHODS FOR SOLVING NONLINEAR LEAST-SQUARES PROBLEMS