Interna tional Jo urna l o f Applied Research 2024 ; 1 0(5 ): 184 -1 88

ISSN Print: 2394-7500

ISSN Online: 2394-5869
Impact Factor (RJIF): 8.4
Separable nonlinear least squares for estimating
IJAR 2024; 10(5): 184-188
www.allresearchjournal.com
nonlinear regression model
Received: 15-02-2024
Accepted: 20-03-2024
Mohamed Jaber, Mohamed Muftah, Yusriyah Hamad and Farag
Mohamed Jaber Hamad
Department of Statistics,
College of Science, University
of Misurata, Libya
DOI: https://doi.org/10.22271/allresearch.2024.v10.i5c.11750

Mohamed Muftah Abstract

Department of Mathematics, Regression analysis is a statistical technique used to examine the relationship between (dependent and
College Arts and Science, Al independent) variables. Regression analysis is typically used by academics to examine the impact of
Abayar, University of several independent factors, or explanatory variables, on a single variable, or response variable. The
Benghazi, Libya regression equation is used by the investigators to explain how the response and explanatory variables
relate to one another. We need to meet many assumptions to estimate the relationship (model). Several
Yusriyah Hamad techniques, including the ordinary least squares (OLS), and maximum likelihood approach (MLE) can
Department of Statistics, be used to estimate the parametric regression model. Moreover, the Spline or Kernel methods can be
College of Arts and Science, Al
used for estimating nonparametric regression. In this work, we attempt to demonstrate the significant
Maraj, University of Benghazi,
and practical method for estimating the nonlinear model. Separable nonlinear least squares (SNLS)
Libya
method is a special case of nonlinear least squares (NLS) method, for which the objective function is a
Farag Hamad mixture of linear and nonlinear functions. In this technique, the nonlinear function (model) can be
Department of Statistics, linearized by applying special transformation or by using expanded Taylor expansion to linearize
College of Arts and Science, Al functions. The separable nonlinear least squares (SNLS) are a very flexible technique that is used to
Maraj, University of Benghazi, linearize the nonlinear functions. The SNLS can be used after linearizing the nonlinear function
Libya through the transformation of the variable of interest. Moreover, the SNLS can be used to approximate
a wide variety of functional shapes. The results show that the SNLS performed very well in comparison
with the NLS. We can observe from the model goodness residuals standard error, AIC, and BIC, that
the SNLS method has provided an estimate equivalent to that NLS provided. Therefore, we can say that
it is useful to estimate nonlinear model separable. Furthermore, we plan to apply the SNLS to a more
complex model using different simulation studies to check the validity of the method.

Keywords: Linear, nonlinear, separable nonlinear least squares, Taylor expansion, gauss-newton
method

Introduction
The statistical tool used to investigate the relationship between (dependent and independent)
variables is called regression analysis. Usually, researchers use regression analysis to analyze
the effect of some independent variables (explanatory variables) on one variable (response
variable) [1, 2]. The investigators use the regression equation to describe the relationship
between the response and explanatory variables. The regression model includes one or more
hypothesized regression unknown parameters [3]. The regression model can be estimated
using several methods such as ordinary least squares (OLS) and maximum likelihood method
(MLE) for the parametric regression model [4]. Nonparametric regression can be estimated by
using the Kernel method or Spline method [5].
The most well-known and classic estimators for regression coefficients are the ordinary least
squares (OLS) estimators obtained by minimizing the sum of squared residuals [6]. The least
squares method needs the error to be assumed as independent and identically distributed with
mean zero and constant variance (the normality assumption). Under the Gauss-Markov
Corresponding Author: theorem, the estimated parameters are the best linear unbiased estimators (BLUE). In
Mohamed Jaber practice, there are many problems caused when the assumptions are violated, e.g. non-
Department of Statistics, normality, heteroscedasticity, and of particular interest is the nonlinear independence of
College of Science, University regresses (independent variables) [7].
of Misurata, Libya
~ 184 ~
International Journal of Applied Research https://www.allresearchjournal.com

Unlike the linear regression model, the nonlinear regression nonlinear least squares, like in linear least squares case. The
model is not restricted to belonging to a specific relation. To nonlinear regression model, given by,
apply the linear regression model, we need to satisfy several
assumptions such as linearity, no multicollinearity, and Y = f(X; θ) + ε
normality. While to apply the nonlinear regression, we need
a model to fit with data and an initial guess to start Where
estimating the model parameters. The Gauss-Newton Y = (y1 , y2 , … , yn )T , f(X; θ) =
method (GNM) is the most popular method that is use to fit (f(X1 , θ), f( X2 , θ), … , f(Xn , θ))T , X = (X1 , X2 , … , Xp )T , θ =
the nonlinear regression model [8].
(θ1 , θ2 , … , θp )T , and ε = (ε1 , ε2 , … , εn )T ; ε~NID(0, σ2 I)
When the relationships in data are not linear, additional
flexibility is needed to apply the traditional approach of the
Nonlinear least squares. However, recent advances in The least squares estimate of θ, labeled by θ̂, is the choice
statistical techniques help to analyze data where questions of of parameters that minimizes the sum of squared residuals
nonlinearity arise [9]. Smoothing splines and semi-parametric n
regression, which allow more flexibility than the nonlinear
regression models, are examples of these techniques. This ss(θ) = ∑[Yi − f(Xi ; θ)]2 , i = 1,2, … n
work will provide a review of nonlinear least squares and i=1

separable nonlinear least squares with applications of these

techniques. Separable nonlinear least squares (SNLS) Or, it can be written as:
method is a special case of nonlinear least squares (NLS)
method, for which the objective function is a mixture of ss(θ) = εT ε = [Y − f(X; θ)]T [Y − f(X; θ)]
linear and nonlinear functions [10].
Moreover, in this work, we will introduce basic concepts of The partial derivatives of ss(θ), with respect to each θj in
the separable nonlinear least square technique, where the turn, set equal to zero to obtain the p normal equations [16].
basic idea will be clarified by an example. Whereas, the Each normal equation has the general form
separable least squares regression is concerned with the
flexible incorporation of nonlinear functional relationships ∂
[ss(θ)]θ=θ̂ = 0
in regression analyses. It has many applications in many ∂θj
different areas, especially in operations research, and
industry engineering [11]. Furthermore, the separable n
∂
nonlinear least square can be used in many practical cases ∑[Yi − f(Xi ; θ)][ f(Xi ; θ)]θ=θ̂ = 0, j = 1,2, … , p
such as fuzz regression model [12, 13]. In details, the ∂θj
i=1
advantages and disadvantages of these advanced regression
techniques will be evaluated and discussed for a partly ∂
Where, f(Xi ; θ) the partial derivatives of a nonlinear
linear regression model and partly nonlinear regression ∂θj
model. model are functions of the parameters. A major difficulty
with nonlinear least squares arises in trying to solve the
Methodology normal equations for θ̂, since clear solutions cannot be
The nonlinear models are used to describe a more obtained, iterative numerical methods are used. These
complicated relationship between the response and methods require initial guesses, or starting values. For the
explanatory variable. Moreover, when the relationships starting value parameters are labeled as θ0 and find θ1 , θ2 , …
between the response and explanatory are not linear relation until we obtain a sufficiently small adjustment being made
more flexibility is needed to apply the traditional approach at each step, when this happens, the process is said to have
of the Nonlinear least squares. Unlike the linear regression converged to a solution [17].
model, the parameters may not linear function in the The Gauss-Newton method (GNM) uses a linearization
predictors. Therefore, the assumptions of applying ordinary based on a Taylor expansion in the parameter space to
least squares are violated [14]. The basic form for a nonlinear estimate parameter values. For the Taylor expansion of
model between the response y and a predictor x is given as, f(Xi ; θ) around the starting value θ0 , to obtain a linear
approximation of the model in the region nears the starting
Y = f(X; θ) + ε values. If θ is close to θ0 , the following approximation
holds:
where, f is a nonlinear function involving the predictor and
the parameter vector θ, relating E(Y), and θ are the vector of p
P parameters [15]. Also, the error term is assumed to have the ∂
f(Xi ; θ) ≈ f(Xi ; θ 0)
+ ∑[ f(Xi ; θj )] (θj − θ0j )
same properties as in the linear regression models. In the ∂θj
j=0 θ=θ0
nonlinear regression models, at least one of the derivatives
of the expectation function f with respect to the parameters T T
will have at least one of the parameters [2]. Where, θ = (θ1 , θ2 , … , θp ) , θ0 = (θ00 , θ10 , … , θ0p ) , fi0 =
∂
f(XI ; θ0 ), β0j = (θj − θ0j ), and Fij0 = [ f(Xi ; θ)]
Estimation of Nonlinear Regression model ∂θj
θ=θ0
Nonlinear Least Squares Estimates (NLS)
The least squares method is used to estimate the parameters Therefore, the following equation can be written as
of the nonlinear models. To estimate the parameters using

~ 185 ~
International Journal of Applied Research https://www.allresearchjournal.com

p
Gauss-Newton Algorithm for (NLS)
f(Xi ; θ) = fi0 + ∑ Fij0 β0j Taylor expansion for f(X𝑖 ; θ) is
j=0
3
∂
f(Xi ; θ) ≈ f(Xi ; θ0 ) + ∑ [ f(Xi ; θj )] (θj − θ0j )
∂θj
Also, the nonlinear model can be written as j=0 θ=θ0

p ∂
f(Xi ; θ) ≈ f(Xi ; θ0 ) + [θ (1 − θ2 e−θ3X )](θ1 − θ10 )
Yi0 = ∑ Fij0 β0j + εi , where Yi0 = Yi − fi0 ∂θ1 1
j=0
∂ ∂
+ [θ (1 − θ2 e−θ3X )](θ2 − θ02 ) + [θ (1 − θ2 e−θ3X )](θ3 − θ03 )
∂θ2 1 ∂θ3 1
Which is of the similar form of the multiple linear f(X i ; θ) ≈ f(X i ; θ0 ) + (1 − θ2 e−θ3X )(θ1 − θ10 ) − θ1 e−θ3X (θ2 − θ02 )
regression model. Using the matrix notation can be written + θ1 θ2 Xe−θ3X (θ3 − θ03 )
as
Initial guesses or starting values are required for estimating
Y 0 = F 0 β0 + ε the model parameters using Gauss-Newton algorithm.
Moreover, we can inference about the model parameters by
Where, applying several assumptions around the estimated
parameters θ̂ [20].
Y10 0 0
. . . F1p 0 β00 ε1
F10 F11
ε2 1. The estimated parameters θ̂ has approximate normal
Y20 0
F20 0
F21 . . . F2p 0 β10
Y 0 = .. , F 0 = . .. . . . . , β0 = .. , and ε = ...
distribution with approximate mean 𝜃 and approximate
.. .. . .. .. .. .. .. . covariance matrix σ2 (F ′ F)−1 .
0 0 0
0
[Yn ] [ Fn0 Fn1 . . . Fnp ]
0
[ βp ] [εn ] 2. An approximate (1 − 𝛼)100% joint confidence region
for θ, which is an ellipsoid is given by:
The least squares estimate of the parameters β0 are obtained 3. (θ − θ̂)′ F̂ ′ F̂(θ − θ̂) ≤ ps 2 F(p,n−p,α)
as 4. An approximate (1 − 𝛼)100% marginal confidence
interval for θi is
−1
δ0 = (F 0 F 0 )
T
F0Y0 5. θ̂i ± t (n−p,α) se(θ̂)
2
6. An approximate (1 − 𝛼)100% confidence interval for
New values of the parameters are obtained by adding the the expected response variable at X0 is
estimated shift to the initial values using Gauss-Newton
increment by 7. f(θ̂, X0 ) ± t (n−p,α) s√V0′ (F̂ ′ F̂)−1 V0
2
8. An approximate (1 − 𝛼)100% confidence interval for
θ1 = θ0 + δ0 the predicted mean of the response variable at X0 is:
θ2 = θ.1 + δ1
f(θ̂, X0 ) ± t (n−p,α) s√1 + V0′ (F̂ ′ F̂)−1 V0 ; V0
.. 2
a
θ =θ a−1
+δ a−1 ∂f(θ̂, X0 )
=[ ]
∂θ θ=θ0
Where, δa is called the Gauss-Newton increment. The
model is then linearized about the new values of the Separable of Nonlinear Least Squares (SNLS)
parameters and linear least squares are again applied to find In the separable least squares, the objective function is a
the second set of adjustments, and so forth until the desired mixture of two components (linear and nonlinear functions)
degree of convergence is attained [18]. The adjustments [21]
. The separable of nonlinear least squares is a special case
obtained from the Gauss-Newton method can be too large of nonlinear least squares in which the function can be
and bypass the solution, in which case the residual sum of derived into two parts [22]. The method can be used in many
squares may increase at that step rather than decrease. applications such as numerical analysis, neural networks,
Moreover, the Gauss-Newton algorithm will converge only and Environmental Sciences. However, the SNLS is an
with a good initial guess when the matrix F 0 is a full rank invalid method when there are some constrains on the linear
matrix in a neighborhood of the least square’s solution [19]. part of variables [23, 24]. Here, we proposed SNLS to solve a
Otherwise, there is no guarantee that the Gauss-Newton function that was solved by NSL.
algorithm will converge.
In practice, the previous technique can be used only when ̃ + εi ,
f(X i , θ) = θ1 (1 − θ2 e−θ3 X ) + εi = θ1 − θ1 θ2 e−θ3X = β0 + β1 X
the function f(Xi ; θ) is continuously and differentiable
respect to the parameter θ. If the function f(Xi ; θ) is not ̃ = e−θ3X
β0 = θ1 , β1 = −θ1 θ2 , and X
continuous and differentiable with respect to the parameter
θ, it is usually necessary to modify the model or use another ̃ 𝑖 ; θ3 ) is
Taylor expansion for f(X
technique to estimate the nonlinear model. To apply the
Gauss-Newton Algorithm for the nonlinear model, we need ∞
to find the Taylor expansion for the model. (−θ3 X)j
̃ i ; θ3 ) ≈ f(X
f(X ̃ i ; θ03 ) + ∑
j!
j=1
f(Xi , θ) = θ1 (1 − θ2 e−θ3X ) + εi

~ 186 ~
International Journal of Applied Research https://www.allresearchjournal.com

̃)2 (θ3 X
(θ3 X ̃)3 the estimated parameters of the
̃ i ; θ3 ) ≈ f(X
f(X ̃ i ; θ03 )−θ3 ̃
X+ − +⋯
2! 3! f(Xi , θ) = θ1 (1 − θ2 e−θ3X ) + εi using data science about
chloride ion transport through blood cell walls the data set
For any given value for θ3 , the θ1 and θ2 can be estimated includes two factors (y donates the chloride concentration
by applying linear least squares method as: (in percent) and x donates to the time (in minutes)). For
more details, see [25]. The review study was performed to
Estimating Nonlinear Model compare between the nonlinear least squares and separable
The study implemented to comparison between the nonlinear least square. In this study, we give a short
estimated model using nonlinear least squares (Gauss- application on the separable nonlinear least square method
Newton algorithm) and separable nonlinear least squares. unseparated scheme for NLS. The results, of estimating
The study was carried out to estimate the nonlinear model nonlinear model using NLS and SNLS are demonstrated in
parameters (θ1 , θ2 , and θ3 ). We used RStudio to generate Table 1:

Table 1: Estimated model parameters using nonlinear least squares and separable nonlinear least squares
NLS SNLS
Parameters
Estimate St. Error T P-value Estimate St. Error T P-value
𝜃1 39.09 0.974 40.12 <2e-16*** 28.835 1.0952 26.328 <2e-16***
𝜃2 0.828 0.008 99.80 <2e-16*** 0.638 0.1132 5.634 7.64e-07***
𝜃3 0.158 0.010 15.18 <2e-16*** 0.227 1.5085 -15.113 < 2e-16***
Residuals standard error=1.92 Residuals standard error=1.95
AIC=-20.09 & BIC=-12.13 AIC=-18.12 & BIC=-10.17

From the results, we can see that the estimated model using though the estimated value of the parameter θ1 (28.835) was
SNLS is comparable with the estimated model using NLS. slightly different from the estimated value of θ1 using NLS
Moreover, based on the model goodness of fits both (39.09).
methods NLS and SNLS performed well with the data. The
estimated model using SNLS still performed well even

Fig 1: The typical residuals of nonlinear regression model with estimated model using NLS and SNLS

The above plots, show the mathematical function that model can be estimated using several methods such as
explains the relationship between the dependent variable y ordinary least squares (OLS) and maximum likelihood
and the response variable x throughout the nonlinear method (MLE). While the nonparametric regression can be
relationship. It can be observed from the above figures, that estimated by using the Kernel method or Spline method.
the estimated model using SNLS is close to the estimated In this work, we try to show the important and useful
model using NLS. However, the relationship looks linear technique for estimating the nonlinear model. In this
which can be easily estimated by OLS but the linearity technique, the nonlinear function (model) can be linearized
assumption for the model parameters is violated. by applying special transformation or by expanding using
Taylor expansion to linearize functions. The separable
Conclusion nonlinear least squares are a very flexible technique that
The statistical tool used to investigate the relationship used to linearize the nonlinear functions. The SNLS can be
between (dependent and independent) variables is called used after linearizing the nonlinear function through the
regression analysis. Usually, researchers use regression transformation of the variable of interest and the explanatory
analysis to analyze the effect of some independent variables variables. Moreover, the SNLS can be used to approximate
(explanatory variables) on one variable (response variable. a wide variety of functional shapes. The results show that
The investigators use the regression equation to describe the the SNLS performed very well in comparison with the NLS.
relationship between the response and explanatory variables. We can observe from the model goodness residuals standard
The relationship might be linear and might be a nonlinear error, AIC, and BIC that the SNLS method has provided an
relationship. To estimate the relationship (model), we need estimate equivalent to that NLS provided. Therefore, we can
to satisfy several assumptions. The parametric regression say that it is useful to estimate nonlinear model separable.
~ 187 ~
International Journal of Applied Research https://www.allresearchjournal.com

Moreover, we plan to apply the SNLS to a more complex 23. Ng MP, Grunwald GK. Nonlinear regression analysis of
model using different simulation studies to check the the joint-regression model. Biometrics; c1997. p. 1366-
validity of the method. 1372.
24. Baty F, Ritz C, Charles S, Brutsche M, Flandrois JP,
References Delignette-Muller ML. A toolbox for nonlinear
1. Sykes AO. An introduction to regression analysis; regression in R: The package nlstools. J Stat Softw.
c1993. 2015;66:11-21.
2. Seber GA, Lee AJ. Polynomial regression. Linear 25. Ruckstuhl A. Introduction to nonlinear regression. IDP
Regres Anal; c2003. p. 165-185. Inst Datenanalyse Prozessdesign Zurcher Hochsch
3. Weisberg S. Applied linear regression. Vol. 528. John Angew Wiss; c2010, 365.
Wiley & Sons; c2005.
4. Bro R, Sidiropoulos ND, Smilde AK. Maximum
likelihood fitting using ordinary least squares
algorithms. J Chemom. 2002;16(8‐10):387-400.
5. Opsomer JD, Breidt FJ. Nonparametric Regression
Using Kernel and Spline Methods; c2011.
6. Jaber M. Comparisons between Nonlinear Regression
Models and Semi-parametric Regression Models; c2009
Jul 20.
7. Thinh R. A Comparison of Linear Regression Models
for Heteroscedastic and Non-Normal Data; c2018.
8. Turner H, Firth D. Generalized nonlinear models in R:
An overview of the GNM package; c2007.
9. Hamad F, Kachouie NN. A hybrid method to estimate
the full parametric hazard model. Commun Stat-Theor
Methods. 2019;48(22):5477-5491.
10. Sherrod PH. Nonlinear regression analysis program.
Nashv. TN USA; c2005.
11. Eissa TEA. Effect of Industry Clusters and Commodity
Flow Characteristics on the Selection of Freight
Transportation Modes in the United States with
Advanced Econometric Approaches; c2019.
12. Omar E, Jaber M, Hamad F, Kyamo M, Eissa T. Fuzzy
Based Clustering Algorithm For Manets. Data Acquis
Process. 2023;38:5416.
13. Omara E, et al. An Improved Fuzzy Based Clustering
Algorithm for MANETs. Scand J Inf. Syst.
2022;34(2):33-38.
14. Hamad F, Younus N, Muftah MM, Jaber M. Specify
underlining distribution for clustering linearly separable
data: Normal and uniform distribution case. J Data
Acquis Process. 2023;38:4675.
15. Walkowiak R, Kala R. Two-phase nonlinear regression
with smooth transition. Commun Stat-Simul Comput.
2000;29(2):385-397.
16. Argyros IK, Hilout S. On the Gauss-Newton method. J
Appl Math Comput. 2011;35:537-550.
17. Jaber M, Hamad F, Breininger RD, Kachouie NN. An
Enhanced Spatial Capture Model for Population
Analysis Using Unidentified Counts through Camera
Encounters. Axioms. 2023;12(12):1094.
18. Milliken GA. Nonlinear regression analysis and its
applications. 1990.
19. Jaber M. A Spatiotemporal Bayesian Model for
Population Analysis; c2022.
20. Bates DM, Watts DG. Nonlinear regression analysis
and its applications. Wiley Ser. Probab. Stat; c1988.
21. Smyth GK. Nonlinear regression. Encycl
Environmetrics. 2002;3:1405-1411.
22. Gharibi W, Al-Mushayt OS. A note on separable
nonlinear least squares problem. Presented at: 2011
International Conference on Future Computer Sciences
and Application, IEEE; c2011. p. 54-56.

~ 188 ~