A Discussion on Least Squares Adjustment

with Worked Examples

A. Fotiou
Professor, Aristotle University of Thessaloniki, Dept. of Geodesy and Surveying,
Lab. of Geodetic Methods and Satellite Applications
[email protected]

Summary: Some remarks on least squares adjustment topics are presented, often bypassed
or given little attention. The minimum number of observable parameters, the relation be-
tween observables and other unknown parameters, the role of the functional model and the
linearization effect are fundamental to build a correct adjustment model. Discussing simple
worked examples from the fields of geodesy and surveying, with extensions to practical
applications, a closer look is gained. Especially, a better understanding of the least squares
principal concepts could be very helpful for students, young scientists and non-experts in-
volved in problems requiring adjustment of observations.

Key words: adjustment of observations, parametric degree, functional model, adjustment

methods, linearization effect, adjustment worked examples

1. Building the concept of the Least Squares adjustment model

A Least Squares (LS) adjustment problem includes observable parameters (true
/unknown values), observations (arithmetic values), observational errors (true
/unknown values) and often some other unknown parameters (true/unknown val-
ues). The relation between observables and/or unknown parameters form the func-
tional adjustment model, expressed by non-linear equations in the general case.
Moreover, the behavior of the errors and the observations is described by the same
stochastic model.
The goal of the adjustment is the estimation of unknown parameters, given a set of
observations or measurements, satisfying the minimization of squared errors. The
concept of observation is the same with that of measurement, so they are inter-
changeably used throughout the paper. Applying an adjustment algorithm/method,
the estimates of parameters and observables – adjusted/corrected values - satisfy
also the functional (compatibility). In addition, the estimated observational errors,
frequently called residuals, are the key to assess the adjustment results in a second
necessary step. Without error control, nothing can be stated for the quality of the
results. Since the number of observations is always greater than the minimum

Quod Erat Demonstrandum – In quest of the ultimate geodetic insight |

Special issue for Professor Emeritus Athanasios Dermanis |
School of Rural and Surveying Engineering, AUTh, 2018
needed for a solution, there can be more solutions by means of properly selected
subsets of observations. To derive a unique solution and take advantage of all
available observations simultaneously, a convenient criterion must be adopted.
This solution should be preferably best in terms of quality/accuracy (precision and
reliability) and possibly with a minimum cost.
Among various minimization/optimization criterions, the well-known LS criterion
is one that has been proved efficient for best estimates since the early years of
1800s, the era of Legendre, Gauss and other famous scientists afterwards. As the
magnitude of errors is the key for the quality control, its influence in the adjust-
ment results should be investigated (design/optimization problems).
The LS criterion can be applied either in a simple form, where all errors equally
effect the solution or in an extended form when the contribution of each error is
counted regarding its importance, something that is called weight. Large weights
lead generally to smaller error estimates or smaller weights tent to produce larger
errors. This fact can be realized in the functional expression of the LS criterion
with weights, where in the minimization process the weights act as zoom in or
zoom out scale factors. In the most general case it is possible to count also for the
interrelation of weights. The concept of a weight is proportionally related to the
precision of the corresponding measurement or inversely proportional to its uncer-
tainty; larger weights are associated with more precise measurements. Therefore, a
relation is established between weight and precision.
Considering observational errors as being just small quantities that tend to cancel
out as the number of observations increases more and more, applying the LS crite-
rion a solution is achieved. This is the case of a deterministic adjustment model.
Assigning proper weighs or generally a proper weight matrix demands for a sto-
chastic error model. Under this requirement, errors are considered as random or
stochastic variables, meaning that they are sample values of random variables that
take different values when the observations are repeated under similar conditions.
The same stochastic model holds also for the observations. In theory, an infinite
number of repeated observations may be described by a common probability den-
sity function, from where the expectation and the covariance matrix of random er-
rors can be determined. Even if it is not necessary the knowledge of the density
function, it is possible to express the stochastic model by the expectation of random
errors equal to zero and a covariance matrix that is determined and considered ab-
solutely or a priori known. So far, the LS mathematical adjustment model or sim-
ply the adjustment model (functional + stochastic) has been built, allowing for un-
biased estimations, i.e., those having mean values equal to their real ones for the
infinite sample. The application of the law of covariance propagation, also known
as the law of error propagation, provides the estimation of covariance matrices for
any estimated parameter.

A discussion on Least Squares Adjustment with Worked Examples 59

The elements of the error/observation covariance matrix are measures of precision.
A weight matrix should be then preferably determined so that the estimated pa-
rameter have maximum precision, i.e., minimum variances. A more general crite-
rion for unbiased estimations is expressed by the minimization of the variances of
the estimated parameters. For mathematical simplicity, a compromise is made and
linear estimations instead of non-linear are used. On the other hand, LS linear esti-
mates are identical to the minimum - variance linear estimates in case the weight
matrix is equal to the inverse of the error covariance matrix. This is the best possi-
ble choice for the weight matrix resulting in LS estimates characterized by maxi-
mum precision. Nonlinear LS estimates are identical to nonlinear estimates that
have minimum - variance only if random errors follow the Normal (Gauss) distri-
bution. Under the same stochastic behavior, LS linear estimates are also identical to
the ones derived by the method of Maximum Likelihood Estimation. In conclusion,
the LS adjustment of observations leads to Best Linear Unbiased Estimations
(BLUE).

The final step of the adjustment is the assessment of the results where estimated
covariance matrices are used together with error and other estimates. This is usu-
ally made through a statistical evaluation or reliability control. In this step, the dis-
tribution of errors must be known. Among various choices the Normal/Gauss dis-
tribution has been accepted for it is both a realistic and mathematically simple hy-
pothesis. First, the reliability of the results is examined by hypothesis testing, veri-
fying the validity of the a priori stated model hypotheses (null hypotheses). Usu-
ally, we try to detect and localize possible significant model errors, looking mainly
in the estimated size in relation to their intercorrelation as given by the estimated
covariance matrix. Ending with an acceptable set of observations and a realistic
covariance matrix, in association with an efficient functional model, measures of
marginal detectable errors or marginal effects on the estimated parameters can be
also derived (reliability measures). At this point, computed measures of precision
or uncertainty are also measures of accuracy/quality, like confidence intervals, ar-
eas, spaces. The assessment process could be carried out in a pre-analysis step, be-
fore the measurement campaign, in a trial and error optimization under certain
standards and specification criteria set a priori (design criteria, optimization prob-
lems).
The accuracy or quality of an estimated parameter express the degree of closeness
to its real (unknown) value; therefore, accuracy is always unknown. Moreover, the
precision of an observation or an estimated parameter express the degree of close-
ness between repeated observations or estimated parameters. Precision of a pa-
rameter is given usually by the variance or standard deviation, also called r.m.s.e
(root mean square error) for unbiased estimations. In case the observations are free
of systematic errors (biases) and outliers (gross errors), measures of precision are

60 A. Fotiou
also measures of accuracy as errors are influenced only by the (unavoidable) ran-
dom errors. Pre-processing and post-adjustment tests - usually statistical tests- refer
to the reliability control.
Some authors characterize outliers as mistakes and define errors as being only ran-
dom and systematic; obviously, mistakes have large values and normally could
easily be detected and excluded. We should note that random errors may have any
large value as illustrated by the normal distribution, though this is exceptionally not
likely for a correct model. A more appropriate term for all types of errors might be
the term model errors; the division to three or two categories supports mainly
teaching requirements and an ease understanding with respect to their sources.
Generally, it is difficult to distinguish the specific type of an estimated error, espe-
cially in models with many interrelated variables and parameters, for instance in
the adjustment of geodetic and surveying networks, where a weak configuration in
relation to the associated weights affects seriously error estimates, making thus dif-
ficult or even impossible a correct error localization.
Despite any statistical evaluation, there is not a hundred percent guaranty that all
possible existed errors have been removed. Based on a statistical decision, there is
always a probability (risk) to reject a correct null hypothesis - committing a type-I
error - or to accept a wrong one - committing a type-II error. In the present paper,
we will not deal with statistical assessment of the results. The goal is to discuss and
make some useful remarks on basic topics, sometimes bypassed or completely
omitted in literature. The number of the minimum observable parameters needed
for an adjustment, the relation between the number of observables and other un-
known parameters, the role of the adjustment model and the linearization process
are hot and essential topics for any adjustment problem.

2. Formulating the functional model

Using matrix notation, a list of symbols for parameters, variables and constants is
first given:
• the observable parameters or observables (unknown):
T
yα = ⎡⎣ y1α yα2 … yαn ⎤⎦ , i = 1, 2, ..., n
• the observations or measurements (arithmetic values):
T
y b = ⎡⎣ y1b yb2 … ynb ⎤⎦ , i = 1, 2, ..., n

• the errors (unknown):

T
v = [ v1 v2 … vn ] , i = 1, 2, ..., n
• the unknown parameters:

A discussion on Least Squares Adjustment with Worked Examples 61

 T
xα = ⎡⎣x1α xα2 … xαm ⎤⎦ , i = 1, 2, ..., m
• the approximate values of the unknown parameters:
T
xo = ⎡⎣x1o xo2 … xom ⎤⎦ , i = 1, 2, ..., m
• the corrections to the unknown parameters:
T
x = xα − xo = [ x1 x2 … xm ] , i = 1, 2, ..., m
• the approximate values of the observations:
T
yo = f (xo ) = ⎡⎣ y1o yo2 … yon ⎤⎦ , i = 1, 2, ..., n
The relation between observables, observations and errors is given by the conven-
tional with respect to the sign (+ or –) relation,
yb = yα + v (2.1)

With P, the weight matrix and C the variance covariance matrix of errors or
equally of observations, we usually have two types of stochastic models:
C = σ2Q, P = C−1 = (σ2Q) −1 ( σ2 known or a priori known and Q known) (2.2)

C = σ2Q, P = Q −1 ( σ2 unknown ( σ̂2 is used) and Q known) (2.3)

Model (2.2) reflects an absolutely known observation accuracy, that is a theoretical

case not so realistic as (2.3). On the other hand, the first choice offers much more
mathematical simplicity in estimating measures of accuracy and reliability and is
preferred in many software packages. The choice depends on the user’s opinion in
relation to the specific adjustment problem.
The LS-criterion, the minimization of the quantity φ, is expressed by

ϕ = ∑ vi2 = vT v = min. , P = I (simple form). (2.4)

ϕ = ∑ pi vi2 = vT P v = min., P = diag(p1,p2 ,...,pn ) (extended form) (2.5)

φ = vT P v = min., P full matrix (general form) (2.6)

In any LS adjustment problem, there must be a least number of observable parame-

ters that give a solution, i.e., making possible the determination of all unknown
parameters. We will call this minimum number, the parametric degree r.
The necessary and sufficient condition to characterize a problem as an adjustment
problem is that the number of observations must be greater than the parametric de-
gree (n > r) and at the same time a proper subset of observations no equal to r
(no = r) exists. Otherwise, the problem is not properly described, even if n > r, and

62 A. Fotiou
a solution does not exist. Always in an adjustment problem there are more than one
proper selections of no and therefore more than one solutions but not the LS-
solution.
The general form of the functional model is,
u(y α , xα ) = 0 (2.7)

The compatibility between estimates ( yˆ α , xˆ α ) and the functional model is ex-

pressed by u(yˆ α , xˆ α ) = 0 . Understanding that the number m of unknown parame-
ters in (2.7) is (n + m) while r is only required for a solution, there must be a num-
ber s of independent mathematical conditions/equations in the model,
s = (n + m) − r = (n − r) + m = f + m (2.8)

where f is the degrees of freedom representing the redundant observations.

In geodetic adjustment problems, as it is the case of trigonometric, levelling and
GPS/GNSS network adjustments, r is a constant regardless of the number n of ob-
servations. As the number n, can be as greater as possible, we can only select the
number m of the unknown parameters with respect to r. This choice determines the
specific form of the functional model (2.7) and thus the number s of equations,
something fundamental for the adjustment method that should be followed.

3. Basic observation adjustment methods

All the adjustment methods are equal, meaning that they lead to the same adjust-
ment results regardless the specific method. The choice of one or another method
depends on the type of the functional model. In the following, we will examine the
three basic adjustment methods, namely: the method of observation equations (the
method of parameters or the method of indirect observations), the method of condi-
tion equations (the method of direct observations) and the method of mixed or
compound equations (the general method), see e.g., see, e.g., Wells and Krakiwsky
1971, Mikhail and Ackermann 1976, Dermanis 1986/1987, Koch 1987, Dermanis
and Fotiou 1992, Rossikopoulos 1999, Fotiou 2007, Sneew et al. 2015

3.1 The method of observation equations

This is the most common adjustment method including unknown parameters xα

and having an obvious number of (independent) equations.
The choice m = r, and consequently s = n, makes possible (2.7) expressed by
y α = f (xα ) (3.1)

A discussion on Least Squares Adjustment with Worked Examples 63

Noting that each one of the observables is written as a function of only the un-
known parameters (constants may be included); no observable can be present on
the right side of (3.1). Because it is generally difficult to deal with non-linear sys-
tems and estimations, as a rule in all methods, linearization is adopted. Expanding
(3.1) in a Taylor series around the approximate point x o , computed by any suitable
means - often using a subset of observations - and keeping first order terms, we
have,
∂f
y α = f (xo ) + | (xα − xo ) = y o + Ax , y b − y o = Ax + v
α o
(3.2)
∂x
b = Ax + v (3.3)
where, the (n x m) matrix A is called the design matrix, having elements the partial
derivatives computed by means of x o and b is the (n x 1) vector of the reduced
observations noting that y o = f (xo ) is accurately computed.
Having thus (3.3), the linear system of observations, we note that this is also an
underdetermined system for there are s = n equations with (m + n) > n unknowns.
System (3.3) has generally an infinite number of solutions and the question arising
is which is the best one. Among all solutions, there exists one that satisfies the LS
criterion, and this is a best solution. Under the LS optimization constraint, and
without going into proofs, the (m x m) or (r x r) normal equations system is formed,
( A T PA)xˆ = A T Pb , or xˆ = u (3.4)

where  is the (symmetric) normal equation matrix and u the (m x 1) constant vec-
tor. From (3.4) we get the LS solution/estimates,
xˆ = (ATPA)−1(ATPb) = −1u [det(≠0)], xˆ α = xˆ ο + xˆ (3.5)

vˆ = b − Axˆ , yˆ α = y b − vˆ or yˆ α = f (xˆ α ) (3.6)

vˆ TPvˆ vˆ T Pvˆ vˆ T Pvˆ

σˆ 2 = = = (3.7)
n −r n −m f
The associated estimates of covariance matrices, are obtained by applying the co-
variance propagation law, i.e.,

Cxˆ = Cxˆ α = −1 ˆ =C

if P = C−1 , C ˆ α = σˆ 2−1 if P = Q −1 (3.8)
xˆ xˆ

Cvˆ = P−1 − A−1AT = C − Cyˆ α if P = C−1 , Cˆ vˆ = σˆ 2C vˆ if P = Q −1 (3.9)

Any other estimate based on the above best estimates is also a best estimate.
In case of initial linear equations, y α = Axα + t or b = Ax α + v , where a constant

64 A. Fotiou
vector t is possibly present. Comparing with (3.3), y b − t ≡ b and xα ≡ x meaning
that x o = 0 (no approximate values are needed), while matrices A,  remain the
same. The approximate values x o should be close enough to their correct esti-
mates, otherwise serious linearization errors will affect the solution. In any case,
the adjustment must be repeated with approximate values the previous estimates
until a negligible difference, between consecutive results, e.g. between the un-
known parameter estimates, is achieved.

3.2 The method of condition equations

In this method, m = 0 and s = n – r = f. As we have only observables in the func-
tional model, (2.7) becomes
g(y α ) = 0 (3.10)

having s independent condition equations. Following the linearization, and taken as

approximate values of observations the values of the observations, something that
is followed in the classical approach,
∂g
g(y α ) = g(y b ) + | b (y α − y b ) = 0 , or Bv = w (3.11)
∂y α
with obvious substitutions. B is the (s x s) matrix of partial derivatives computed
by means of y b and w = g (y b ) the (s x 1) vector of the so-called closing errors.
The linear system (3.10) is again an underdetermined system as there are s = (n –r)
equations with n > s unknowns. Imposing the LS criterion, the derived best LS so-
lution is given by the (s x s) normal equations system,
(BP −1B T )kˆ = w , Mkˆ = w , kˆ = M −1w , vˆ = P −1B T kˆ = P −1B T M −1w (3.12)

where M is the (symmetric) normal matrix (det(M ≠ 0)). The adjusted observa-
tions, and the posteriori variance factor when needed, are given by

vˆ TPvˆ vˆ TPvˆ vˆ T Pvˆ

yˆ α = y b − vˆ , σˆ 2 = = = (3.13)
n −r s f
Compatibility between adjusted observations and the functional model is expressed
by, g(yˆ α ) = 0 , Bvˆ = w . Also, the associated covariance matrices are given by

Cvˆ = P−1BTM−1BP−1 = P−1 − Cyˆ α ( P = C−1 ) or Cˆ vˆ = σˆ 2C vˆ ( P = Q −1 ) (3.14)

We remind that a complete elimination of the unknown parameters from the obser-
vation equations (3.1) results in condition equations (3.10).

A discussion on Least Squares Adjustment with Worked Examples 65

3.3 The method of mixed equations
This is the general adjustment method, realizing that it is a method of observation
equations with respect to the unknown parameters xα and simultaneously a
method of condition equations with respect to the observables.
Here, s = n + m – r with 0 < m ≤ r and thus n – r < s ≤ n. We see that there are m
unknown parameters, generally less than the number of r (m < r). In case m = r,
the model can be written in the form of the observation equations method but for
some reason, e.g. due to mathematical complication, the method of mixed equa-
tions may be preferred. Since m < r, some of the parameters xα have been elimi-
nated from (3.1); if all of them had been eliminated, then m = 0 and conditions
(3.10) would be derived.
The mixed model u(y α , xα ) = 0 is expanded in a Taylor series around the point
(xo , y b ) , as it is followed in the classical approach (discussed later). Keeping first
order terms, we will have,
∂u ∂u
u(y α , xα ) = u(y b , xo ) + | ( x α − x o ) + α | o,b ( y α − y b ) = 0 , or
α o,b
(3.15)
∂x ∂y
w + Ax − Bv = 0 (3.16)
where, the matrices (s x m) A, (s x n) B and the (s x 1) vector w = u(y b , xo ) are
evaluated at (xo , y b ) while x and v are estimated from the LS algorithm. As pre-
viously, (3.16) expresses an underdetermined system with s mixed equations and
(m + n) > s unknowns. The LS solution is given by the normal equation system,

( A T M −1A)xˆ = − A T M −1w , or xˆ = −u  xˆ = −−1u , [det(≠0)] (3.17)

vˆ T Pvˆ
ˆ2 =
xˆ α = xˆ ο + xˆ , vˆ = P−1BTM−1 (w + Axˆ ), yˆ α = y b − vˆ , σ (3.18)
s−m
The covariance matrices for the above estimates are then given by
ˆ =C
Cxˆ = Cxˆ α = −1 if P = C−1 , C ˆ α = σˆ 2−1 if P = Q −1 (3.19)
xˆ xˆ

ˆ = σˆ 2C (P = Q −1 ) (3.20)
Cvˆ = P−1BTM−1[I − A−1ATM−1 ]BP−1 (P = C−1 ) , C vˆ vˆ

3.4 Constraints on the unknown parameters

An extension to the above (2.7) and (3.1) models, that both contain unknown pa-
rameters xα , is to set constraints on all or some of the unknown parameters. This is
the case when k equations h ( x α ) = 0 must be satisfied. In consequence, the lin-
earized constraints,

66 A. Fotiou
 ∂h
h(xα ) = h(xo ) + | (x α − x o ) = −z + Hx = 0 ,
α o
or Hx = z (3.21)
∂x
must be included, as a new subset of equations, in the corresponding linear sys-
tems. For the related adjustment algorithms not given here, we underline some is-
sues. The number of k-constraints mean that one or more of the m parameters are
not independent and therefore not fundamental for the description of the problem.
The independent (m – k) parameters, (m – k) = r < m, if they are known, the rest k
parameters are determined; thus, we must include k equations in the functional
model.
In the method of observation equations with constraints, the new total number of
equations is given by sˊ= n + m – r = n + k where m – r = k or m = r + k. Like-
wise, in the method of mixed equations, the new total number of equations is,
sˊ= (n + m – r) + k = s + k, where m – r ≤ k or m ≤ r + k.
Sometimes it is better to eliminate a part or even all the constraints. A simple case
of constraints that are eliminated, refers to fix a number of point coordinates in a
geodetic network, in order to define the coordinate system or datum of the adjust-
ment.

4. An introductory example
The role of the parametric degree r is fundamental in any adjustment model. More
over the selection of the adjustment method and the linearization are important is-
sues. With the help of worked examples we will try to clarify some related issues.
Consider a problem where the shape of a triangle on a horizontal plane must be
determined. For this purpose, we measured just two angles of the triangle. Obvi-
ously, the third angle is the difference of their sum from 180º. Geometrically, the
triangle’s shape is determined by taking an arbitrary length of the side with the two
angles at each end and intersecting the two lines formed by each angle (the third
vertex is defined). Note that there are two solutions, two intersections at each half-
plane; we choose one of them if there is some knowledge of the relative orienta-
tion. However, it is not possible with two measured angles to have error control;
any serious error in the observations affects directly the position of the third vertex
resulting in an erroneous shape. Realizing that the shape is determined by a mini-
mum number of two angles (observables), the parametric degree r = 2. Also, n = 2
and therefore, the condition n > r for an LS adjustment is not fulfilled.
Let’s modify the design and measure also the third angle, in total the three angles
(ωΑ , ωB , ωC ) . Suppose that measurements have known precision given by their
standard deviations σωΑ , σωΒ , σωC . While r remains constant (r = 2), regardless the
increase of observations, we will have, n = 3 > 2. In addition, there is a minimum

A discussion on Least Squares Adjustment with Worked Examples 67

suitable number of observables, no = r =2 that gives a solution, e.g. any pair of an-
gles determines the shape. Since the necessary and sufficient condition is fulfilled,
the LS-best shape of the triangle can be estimated. Of course, any empirical solu-
tion can be obtained but this is not a best LS-solution.
In the following, we will present the formulation of the functional LS model in
each one of the three basic adjustment methods, underling that the ad-
justed/estimated results from any method are the same.

4.1 Using the method of observation equation

First, we must select m = r = 2 unknown parameters that fully describe the shape of
the triangle and then formulate the three observation equations. One choice is to
take any pair of the observables as being also unknown parameters. According to
this choice, we’ll have,
T T T
yα = [ ωΑ ωΒ ωC ] , yb = [ ω'Α ω'Β ω'C ] , v = ⎡⎣ vω A vω Β vω C ⎤⎦

T
and the unknown parameters, x = [ zA zB ] .
α

Covariance and the weight matrices of the errors/observations are

C = diag(σω2 Α , σω2 Β , σω2 C ) , P = C−1 = diag(1 / σω2 Α ,1 / σω2 Β ,1 / σω2 C )

The observation equations y α = f (xα ) are then given by

ωΑ = z A ωB = z B ωC = 180° − (z A + z B )

Although, these equations are directly linear and one should take advantage of it
(see below), we will treat the model as if it was nonlinear.
T Τ
With x = ⎡⎣zA zB ⎤⎦ , e.g., z oA = ω 'A , z oB = ω 'B , x = x − x = [ δzΑ δzΒ ]
o o o α o

T T
and yo = f (xo ) = ⎡⎣ωoΑ ωoΒ ωoC ⎤⎦ = ⎡⎣zoΑ zoΒ 180° − (zoΑ + zoΒ )⎤⎦

the algorithm is applied as given above, noting that correct units should be used
throughout the computational steps. Applying the algorithm, we begin with the
evaluation of matrices A, b of the linear system b = Ax + v :

⎡ 1 0 ⎤ ⎡ω 'Α − zoA ⎤ ⎡ 0 ⎤
⎢ ⎥ ⎢
A = ⎢⎢ 0 1 ⎥⎥ , b 0 o
b = y − y = ⎢ ω 'Β − zΒ ⎥ = ⎢ 0 ⎥
⎥
⎢⎣−1 −1⎥⎦ ⎢ o ⎥ ⎢ω ' − (180° − (ω ' + ω ' )) ⎥
⎣⎢ω 'C − ωC ⎦⎥ ⎣ C Α Β ⎦

The easily formed (2x2) normal matrix  must be inverted and all the relevant es-

68 A. Fotiou
 ˆ Α,ω
timates {δzˆ Α , δzˆ Β , zˆ Α , zˆ Β , vˆ ωA , vˆ ωB , vˆ ωc , ω ˆ B, ω
ˆ C } are then computed. In this ex-
ample, having initially linear equations, we could alternatively proceed on with a
direct estimation of ( zˆ Α , zˆ Β ) instead of their corrections. In this case matrix A is
the same while b is different. The same holds for  and u.

4.1.1 Using coordinates as model parameters

An alternative approach to formulate the functional model is to use (2-d) coordi-
nates as model parameters. Due to the linearization, approximate coordinates for
points under estimation are computed by any suitable way, usually by a subset of
observations and some known point coordinates, given or arbitrary defined. Coor-
dinates are preferably used in geodetic and surveying adjustment problems as they
offer ease to express equations and compute any quantity. Coordinates cannot be
derived from classical measurements like angles and distances. Therefore, in our
measurement scheme, we must include or assign known coordinates to some
points, either with respect to an official or arbitrary defined coordinate system.
Assume a geodetic network with many vertices, the triangle being a trivial case,
with a number of measured angles, more than the parametric degree r. Assigning
coordinates to points-vertices, the shape, size, orientation, and the position of the
network are fully defined. On the other hand, the observables (here angles) define
only the shape of the network. Therefore, to describe fully the problem, we must
formulate, by some means, the rest information carried by the coordinates, i.e. we
must define the scale/size, orientation and position with respect to the coordinate
system. This is possible by means of imposing constraints on point coordinates. A
simple, usually followed, type of constraints is to keep fixed some coordinates,
such that they do not influence what is defined by the observables (minimal con-
straints) - here the shape of the network/triangle. In a 2-d network with four de-
grees of freedom in general, the position is defined by fixing the coordinates of one
point. The orientation and the size are also defined by fixing the coordinates of an-
other point. Generally, in a 2-d angular network any four coordinates, often two per
point, eliminate the four degrees of freedom in plane; otherwise there is a so-called
datum deficiency.
In the triangle with the three measured angles, the position, size and orientation are
defined by keeping fixed the coordinates for two vertices, e.g., A and B. By con-
sidering arbitrary values, e.g., (xA = 0 m, yA = 0 m) and arbitrary orientation, e.g.,
plane azimuth αAB = 90º and scale, e.g., length dAB = 100 m, then (xB=100 m,
yB=0 m) are obviously obtained. Keeping A and B fixed their coordinates are ab-
solutely known, i.e., constants. Applying the LS-algorithm, with unknown parame-
ters and approximate values
T T
xα = [ xC yC ] , xo = ⎡⎣xoC yoC ⎤⎦ (computed from simple trigonometry)

A discussion on Least Squares Adjustment with Worked Examples 69

the three non-linear equations y α = f (xα ) are given by,

yC yC xC x −x
ωA = arctan , ωB = arctan , ωC = arctan + arctan B C
xC x B − xC yC yC
Also, the analytical structure of A and b is,

⎡ ∂ωA ∂ωA ⎤ ⎢⎡ yoC x oC ⎤

− ⎥
⎢ |o |o ⎥
(doAC )2 (doAC )2
⎢ ∂x C ∂yC ⎥ ⎢ ⎥
⎢ ∂ω ⎢ ⎥
∂ωB ⎥ ⎢ yoC o
xC ⎥
A = ⎢ B |o |o ⎥ = − o 2
⎢ ∂x C ∂yC ⎥ ⎢ (doBC )2 (d BC ) ⎥
⎢ ∂ω ⎢ ⎥
∂ωC ⎥ ⎢ o o
x oC x oC ⎥
⎢ C |o |o ⎥ − yC + yC −
⎣⎢ ∂x C ∂yC ⎦⎥ ⎢ (do )2 (do )2 (doBC )2 (doAC )2 ⎦
⎥
⎣ BC AC

T
b = y b − y 0 = ⎡⎣ω 'Α − ω0A ω 'Β − ω0Β ω 'C − ω0C ⎤⎦

4.2 Using the method of condition equations

From the observation equations, eliminating the two unknown parameters, we get
one independent condition equation (s = n – r = 3 – 2 = 1), i.e.,

g(yα ) = ωΑ + ωΒ + ωC −180° = 0
which is also linear. This equation could easily be written directly realizing the
geometric angle condition. The matrices of the linear system are then formed,
B = [1 1 1] , w = [ ω 'Α + ω 'Β + ω 'C − 180°]

Generally, the number of independent conditions is easily determined if the para-

metric degree r had been previously determined (s = n – r), e.g. thinking in terms of
the observation equations method. To find the correct r and mainly to formulate the
condition equations, it may be a simple matter for some problems having a simple
geometric configuration/design; otherwise it is rather difficult with a possible risk
for an incorrect model.
In this example, matrix M to be inverted is one (1x1)-one element. Generally, in
view of the computational ‘difficulty’ compared to the method of observation
equation where  is (mxm) or (rxr), we note that in cases that s = n – r < r or n < 2r
the condition method is preferable. This is mainly the reason for using the condi-
tion method in the past when computing means were limited or even absent. On the
contrary, formulating conditions is rather a difficult task in a complicated configu-
ration like a classical geodetic trigonometric network. Understanding that in geo-

70 A. Fotiou
detic networks coordinates estimation is of primary importance, the ease of ex-
pressing the observation equations and the direct estimation of covariance matrices
of unknown parameters, in relation to the plenty of computing power, makes the
observation equation method almost the standard choice since some decades.

4.3 Using the method of mixed equations

In this method, m≤ r. Because m = r = 2 and s = n + m – r = 3, the functional mixed
model is obtained directly by simply writing the observation equations as
u1 = ωΑ − z A = 0 , u 2 = ωB − z B = 0 , u 3 = ωC + z A + z B − 180° = 0
from where,

⎡−1 0⎤ ⎡1 0 0 ⎤ ⎡ ω 'Α − z0A ⎤

⎢ ⎥ ⎢ ⎥ ⎢ ⎥
A = ⎢0 −1⎥ , B = ⎢0 1 0⎥ = I , w = ⎢ ω 'Β − z0Β ⎥
⎢⎣ 1 1 ⎥⎦ ⎢⎣0 0 1⎥⎦ ⎢ 0 0 ⎥
⎢⎣ω 'C + zA + zΒ − 180°⎥⎦

Applying the LS algorithm best estimates are then derived.

In the common case where m<r, we would have had a different functional model.
For instance, by eliminating the parameter z B : from equation u2, z B = ωB and sub-
stituting in u3 results in u 3 = ωC + z A + ωC − 180° = 0 . Finally, in the remaining two
equations u1 and u3, we’ll have only one unknown parameter ( z A ), noting that u3 is
of a mixed type and cannot be expressed as y α = f (xα ) , i.e., we have,
u1 = ωΑ − z A = 0 , u 3 = z A + ωB + ωC − 180° = 0

The application of the algorithm results in ẑ A . After the end of the adjustment al-
ˆ B is also obtained.
gorithm the estimate ẑ B = ω
We can also apply the mixed equations method using coordinates as previously
described. As far as the algorithmic and computing ease, we will see that in some
problems such as the best fit of a function to data points (see below), the mixed
equation method takes precedence.

4.4 Including observations of different type

Apart from the three measured (horizontal) angles, let’s also measure the horizontal
distance dAB, having in total n = 4 observations. In this case the observations de-
fine the shape and the scale/size of the triangle. The minimum number of observ-
ables for a solution (shape + scale) is now three (r = 3), e.g., two angles to define
the shape and the measured distance to define the scale of the triangle.
In the method of observation equations with s = 4 and m = r = 3, we can proceed

A discussion on Least Squares Adjustment with Worked Examples 71

on as previously having now one more unknown parameter ( z AB ) and thus one
more equation, e.g., d AB = z AB . Using coordinates as parameters, we must realize
that the minimal constraints are three; two for the position of the triangle, e.g., fix-
ing (x A , y A ) and one for the orientation, e.g., by fixing y B or x B . Considering
(x B , x C , y C ) as the three unknown parameters (m=3), the additional observation
equation is d AB = (x B − x A ) 2 + (y B − y A ) 2 . The functional model is easily formed
when more distance observations are included, noting that m = r = 3 in any case.
In the method of condition equations, s = n – r = 4 – 3 = 1 and the obvious condi-
α
tion is g(y ) = g1 = ωΑ + ωΒ + ωC −180° = 0 , leaving ´out of the model´ the dis-
tance d AB , meaning that the scale/size of the triangle is not adjusted and simply is
defined from the distance observation without error control on it. The same holds
also for the method of observation equations, not directly shown as in the condition
equation. Should this be avoided, one more or two distance observations should be
observed. With one more distance observation, n=5, s = n – r = 5 – 3 = 2, and the
second condition equation comes from the law of sines where two measured dis-
tances are included. With three distance observations, n = 3, s = n – r = 6 – 3 = 3,
and the third condition equation comes again from the law of sines or even from
the law of cosines.
If we would prefer to use the method of mixed equations with m<r, one at least of
the unknown parameters should be eliminated from the equations of the method of
observation equations.

5. Introducing redundant constraints on unknown parameters

In geodetic and surveying networks coordinates are used as unknown parameters.
In the previous example, we discussed the definition of the coordinate system
through minimal constraints on a number of point coordinates that are kept fixed.
Minimal constraints do not affect the estimated errors and observable parameters,
regardless of their choice; introducing them in the adjustment model, statistical
tests on the quality of observations are meaningful. On the other hand, they affect
the estimates of the unknown parameters/coordinates and their covariance matrices.
Very often, compliance with technical specifications or operational requirements
demand for redundant constraints based on the supposed higher accuracy of the
control points included in the network and kept fixed, to increase the overall accu-
racy. This is normally true but sometimes, a significant problem may have de-
stroyed the (assumed) superior quality of the control points. Redundant constraints
do affect everything in the adjustment results. Normally, their quality must be
tested, usually statistically; otherwise the reason for unreliable results cannot be

72 A. Fotiou
safely detected, i.e. if the reason is the quality of the observations or the quality of
the redundant constraints or both.

5.1 A worked example with and without constraints

In a levelling network with four points (1, 2, 3, 4), five height differences were
measured (n=5),

yα = [ δh12 δh13 δh23 δh24 δh34 ]T , yb = [ δh12

′ δh13
′ δh′23 δh′24 δh′34 ]T
The goal is to determine the best estimates of height differences between any pair
of points under the condition δh14 = q14 = constant .
The observations, obviously, define height differences. Ignoring, for the moment,
the condition that should be satisfied, we need only three (r = 3) observables for a
solution, e.g., (δh12 δh13 δh 24 ) : from (δh12 δh13 ) and the loop-condition on
(1-2-3-1), δh12 + δh 23 − δh13 = 0 , δh 23 is defined, while from the loop-condition on
(2-4-3-2), δh 23 + δh 34 − δh 24 = 0 , δh 34 is also defined. We see that n (=4) > r (=3),
and a suitable number no = r = 3 of observables exist; therefore, we have an ad-
justment problem.

1
4

Using the observation equations method, the problem can be solved by choosing
three of the observables (m = r = 3) as also being unknown parameters and subse-
quently formulate the respective equations, as we did in the previous worked ex-
ample. Alternatively, we can choose heights to describe the problem and the three
T
unknown parameters, e.g., the heights of points (2, 3, 4), x = [ h2 h3 h4 ] , while a
α

fixed height will be given to the fourth point, i.e., h1 = q1 = constant , in order to
define the height system/datum by minimal constraints (k=1). From three heights
among five points and one known height any other height difference is well de-
fined.

A discussion on Least Squares Adjustment with Worked Examples 73

Including the constraint, δh14 = q14 , we easily notice that the parametric degree
changes and becomes r = 2, as we need only two observables to obtain a solution,
e.g., (δh12 δh13 ) . In the case of coordinates, we have only two known heights
(m=2), because, from δh14 = h 4 − h1 = q14 = constant we get one more known
height, h 4 = h1 + δh14 = h1 + q14 = constant , eliminating thus the additional con-
straint equation by fixing two heights. The adjustment algorithm, i.e. observation
equation without constraints, can be normally applied keeping in mind that the ap-
plied two constraints are redundant and should be tested.
For the implementation of the condition method, we notice that without constraints,
s = 5 – 3 = 2, and the two (independent) conditions are directly formed by any pair
of equations among the loops (1-2-3-1), (2-3-4-2) and (1-2-4-3-1). Setting the con-
straint, we have, s = 5 – 2 = 3. The three (independent) conditions are directly
formed by any pair of equations among the loops (1-2-3-1), (2-3-4-2) and (1-2-4-3-
1) plus one condition reflecting the need that one given height difference is con-
stant. The last condition could be formed by a loop-equation that includes the con-
stant height difference, e.g., δh12 + δh 24 − q14 = 0 .
The mixed equation method can also be applied considering the above remarks.

6. Derived observations
Sometimes, instead of the original observations, derived observations are used. The
advantage could be a simpler functional model but with a probably more compli-
cated stochastic model and a demanding software. A simple example will clarify
these concepts. Consider a measured plane (convex) quadrilateral (1-2-3-4) with
observables the 12 horizontal directions, three from each vertex to the other ones.

3
4

2
1

A horizontal direction βij is an angular quantity measured from a point (i) to an-
other point (j) with respect to a reference direction θio , e.g. the zero direction of
the horizontal circle of the theodolite. The reference direction is an orientation pa-
rameter that must be defined, among other parameters, in the method of observa-
tion equations (or even in the mixed model); otherwise the concept of a direction is
meaningless and the problem is not correctly described. A series of measured di-

74 A. Fotiou
rections refers to the directions from one point to other points, having all of them
the same orientation direction. On the other hand, the difference between any two
directions form the respective horizontal angle, ωijk = βik − βij , where i  k is the
right sighting and i  j the left sighting. From a horizontal angle and one of its di-
rections the other direction is directly derived.
Directions as observable parameters define not only the shape of the quadrilateral
but also the four orientation parameters, one at each point. The parametric degree r
= 8 and this should be explained, through a construction scheme, as it is not so
clear: Let’s take an arbitrary length e.g., for the distance d12 with an arbitrary ori-
entation. From point 1 we use β14 as an orientation direction and draw β13 , β12 .
From point 2 take as reference β 21 and draw β 24 , β 23 so that point 3 is determined
by the intersection with β13 . The intersection of the respective directions defines
points 3 and 4. Up to now we have used six directions so that the shape of the
quadrilateral is well defined; any other angle can be derived using simple geome-
try. However, there is something missing, and this is the possibility to describe di-
rections from points 3 and 4. This can be feasible by using a reference direction for
point 3, e.g., β32 and one for point 4, e.g., β 43 . At this point the problem has been
fully described by a least number of eight observables, hence the parametric degree
r = 8 and the degrees of freedom f = 12 – 8 = 4. Other equivalent subsets of eight
observables are possible.
Having n = 12 and r = 8, we must choose m = r = 8 unknown parameters in case
the method of observation equations is used. One option is to take a suitable set of
eight unknown parameters among the set of the twelve observables and form the
function model accordingly (some geometric manipulations needed). The other
option, almost always preferred, is to use point coordinates to describe the prob-
lem. In this case the reference system/datum should be also defined, e.g., by fixing
coordinates (x1=0, y1=0) and (x2=q, y2=0), q being a constant (minimal constraints).
Therefore, with four unknown parameters, being (x3, y3), (x4, y4) and with four un-
known orientation parameters (θ1, θ2, θ3, θ4), eight unknown parameters are used in
total. We remind that the observation equation of a horizontal direction is given by,
⎛ x j − xi ⎞
βij = αij − θi = arctan ⎜ ⎟ − θi
⎝ y − yi ⎠
where, αij is the plane azimuth (angle from the grid north to the direction ij).
Usually, in geodetic networks, θi coincides with the most left measured direction
in a series from a station point; its approximate value is given by the azimuth com-
puted from approximate coordinates.
In case of condition equations, the number of the independent equations should be

A discussion on Least Squares Adjustment with Worked Examples 75

s = n – r = 12 – 8 = 4; three simple angular equations – one for each of three out of
four formed triangles, and one ‘side equation’ which is a bit complicated and will
not be given here. Because the condition equations use the directions as pairs of
differences between them (angles), no orientation parameter is needed.

6.1 Using derived angles instead of the initial directions

A different functional model makes use of the eight horizontal angles derived from
the above twelve measured directions. In terms angles as observables, we need at
least four angles (r = 4) to define the shape, e.g. the four angles at the end points of
side 1–2 (any other angle is well defined). This is the case of the so-called derived
observations (here derived angles) instead of the original measured observations
(here directions). No orientation parameters exist in the respective eight angle ob-
servation equations, as expressed with respect to the four unknown parameters
(m = r = 4), in the method of observation equations. In the method of condition
equations, the number of conditions s = n – r = 8 – 4 = 4 should be formed, three
angle equations and one side equation (like the directions scheme). The results be-
tween the adjustments, directions against angles, will be the same only if we will
use the correct stochastic model for the derived angles; correlation exist between
consecutive angles as their derivation uses a common direction. For instance, sup-
posing uncorrelated directions, and taking σβ212 = σβ213 = σβ214 = σ12 , the derived con-
secutive angles
ω132 ( = β12 − β13 ) and ω143 ( = β13 − β14 )
will have,
σω2 132 = σω2 143 = 2σ12 and σω132ω143 = −σ12 ,

as obtained by the error propagation law.

7. Best fitting problems with an example for a best fit circle

In problems, such as the best fit of a function to data points expressed by measured
coordinates, e.g., line, plane curve, plane, circle, sphere, ellipsoid, some difficulties
arise. Among them the selection of the functional model and method, the lineariza-
tion process and the implementation of the algorithm are of primary importance.
Next, we will try to point out some of these issues through a worked example of a
best fit circle.
Given N data points (x ′i , y′i ) , i=1, 2, …N, measured along the circumference of a
circle (x c , y c , R) , with (x c , y c ) the coordinates of the circle center and R its ra-
dius. The familiar circle equation is given by,

76 A. Fotiou
 (x i − x c ) 2 + (yi − yc ) 2 − R 2 = 0 (7.1)

Due to errors, all points do not satisfy the same equation; any triplet of data points
determine a different circle. The goal is to find the best solution, i.e. the best circle
under some fitting criterion. Generally, in fitting problems, data points should have
a suitable distribution, here along a large arc and preferably along the whole circle,
so that the circle is correctly defined avoiding probable inconsistencies and bad
approximate values. However, small arcs should be avoided as considerable errors
may grossly affect estimates of circle parameters.
Noting with (x i , yi ) the observables and (v xi , v yi ) the errors, an efficient fitting
criterion, is based on the minimization of squares of both errors. This can be ex-
pressed by the minimization of the geometric distances of the observed points from
the best fit circle,

( )
2
∑ di2 = ∑ (x ′i − x c ) 2 + (y′i − y c ) 2 − R (7.2)

Such distances are reckoned along the directions from any data point to the center
of the circle in case the errors/observations have the same uncertainty, a common
situation in practice (geometric fit, orthogonal fit).
A criterion not so strict but competitive and often preferable to the previous one for
its mathematical simplicity, is based on the minimization of the squares of the so-
called algebraic distances, i.e.,

∑ di2 = ∑ ( (x ′i − x c )2 + (y′i − yc )2 − R 2 )
2
(7.3)

In literature (see, e.g., Chernov 2010), there are many approaches and model modi-
fications to both above model implementations, e.g., either using the model
A(x i2 + yi2 ) + Bx i + Cyi + D = 0 (under constraint on A, B, C, D)
or z i + Bx i + Cyi + D = 0 , from where, with the substitutions,

B = −2x c , C = −2yc , D = x c2 + yc2 , zi = x i2 + yi2

we derive a linear model with respect to the unknown parameters with simple and
fast computational algorithms. We do not intend here to go into details; instead we
will point out how this problem could be faced with each one of the three basic ad-
justment methods and how one could face the effect of the linearization to obtain a
best or at least a sufficient solution.

7.1 Using the method of observation equations

The model y α = f (xα ) for a circle, can be described by

A discussion on Least Squares Adjustment with Worked Examples 77

 xi = xc + R cos θi , yi = yc + Rsin θi (7.4)
where, θi the angle from the x-axis anticlockwise to the point on the circumfer-
ence. In total, we have m = r = 3 + N unknown parameters with n = 2N observa-
tions. For N > 3 so that n = 2N > r, an adjustment problem of observations is evi-
dent. The adjustment can be normally applied, noting that the normal equation ma-
trix is (mxm) and that good starting values for the approximate coordinates are eas-
ily computed by suitable triplet/triplets of data points, for the center and the radius.
In addition, from each data point an approximate angle θ is also obtained. Con-
straints on the unknown parameters may be included, such as the circle must have a
known radius, the circle is passing from a known point or the circle is tangential to
a known line.
A worthwhile notice for the parametric degree is that r (= N + 3) depends on the
number of observations, something that do not happen in geodetic and surveying
network adjustment problems. The same dependence of the parametric degree on
data points is true in all best fitting problems, where given/measured coordinates
are treated as observables. The same situation exists in the coordinate transforma-
tion problems between different systems, as it is the well-known similarity trans-
formation, where coordinates of common points in both systems are affected by
errors.

7.2 Using the method of mixed equations

Eliminating all the θi parameters from (7.4), by adding the squares of coordinate
differences, we will obtain a different functional model,
u i = (x i − x c ) 2 + (yi − yc ) 2 − R 2 = 0 (7.5)

that describe mixed equations u(y α , xα ) = 0 . In this model, we have m = 3 un-

known parameters (m < r (= 3 + N)) and s = N mixed equations (one per point).
For uncorrelated observations, a common practice, the advantage with this method
is that the normal equation matrix 3x3 instead of (3+N)x(3+N) in the method of
observation equations; certainly, there are some matrix-partitioned techniques re-
ducing the computation effort for many data points in the method of observation
equations. Overall, the computational ease makes the mixed equation method the
preferred one. Constraints can also be included.

7.3 Using the method of condition equations

Even though the condition method does not offer any serious advantage, for the
shake of completeness, we note that the functional model, g(y α ) = 0 , can be de-
rived by complete elimination of the unknown parameters. From the mixed equa-

78 A. Fotiou
tions (7.5), we take three of them and analytically express the coordinates of the
circle and its radius in terms of the coordinates of the three used points. Next, we
substitute these coordinates to the remaining (N–3) equations and in this form no
unknown parameter except the observables is included. Matrix M to be inversed is
(N–3)x(N–3) and nothing notable is gained. Moreover, equation complexity is ob-
vious. Therefore, the mixed equation model is again the preferred one among the
three basic adjustment methods.

7.4 The linearization effect

LS adjustment follows normally the expansion of the initial functional model to a
Taylor series to derive linear equations. In some problems, if an exact non-linear
solution is possible, it should be preferred or at least compared to the obtained one
through linearization, avoiding possible divergence of a correct solution.
Partial derivatives in the method of observation equations (matrix A), depend only
on the approximate values of unknown parameters xα . Starting with good initial
values x o , some iterations are generally needed to achieve convergence and a cor-
rect solution.
In the method of condition equations where there are no (unknown) parameters the
point of expansion in most of the classical treatments is around y b and not
y o (= f (xo )) . This fact imposes some questions for the correct solution. The algo-
rithmic iterations in the adjustment process start with initial values –the classical
approach- and then repeated accordingly, in a similar treatment as described in the
following more general case of the mixed model. However, the linearization error
(neglecting higher order terms) may affect the solution to a false point despite an
achieved convergence. Moreover, matrix B acquires variability/stochastic proper-
ties under the theoretical infinite repetitions of the observations, as it depends on
the observations. To avoid stochastic problems, since the experiment is usually car-
ried out once, we suppose that y o ‘happens’ to be identical with y b .
Similarly, in the mixed equations model, with
T T
xα = [ x c yc R] , x = [ δx c δy c δR ]
T T
y α = [ x1 y1 x 2 y 2  x N y N ] , y b = [ x1′ y1′ x ′2 y′2  x ′N y′N ]

the expansion is traditionally developed around the point (xo , y b ) instead of

(xo , y o ) . Again, we have a similar risk as mentioned above, as matrices B and A
depend also on the observations.
In case the initial equations of the functional model are linear or partly linear the
algorithm allows for simplifications and similarities with the classical approaches.

A discussion on Least Squares Adjustment with Worked Examples 79

We must face the linearization effects by either a non-linear solution, if possible, or
by iterating properly the linear one, introducing some modification on the general
adjustment model of Gauss-Helmert.
To show the effect of the model modification, we’ll take the general/mixed model
for the circle fit. The linearization, taking only the first order terms, has as follows:
∂u i ∂u
u i (x c , yc ,R, x i , yi ) = u i (x co , yoc ,R o , x io , yio ) + (x c − x oc ) + i (yc − yoc ) +
∂x c o ∂yc o

∂u i ∂u ∂u
+ (R − R o ) + i (x i − x io ) + i (yi − yio ) + ... = 0 . (7.6)
∂R o ∂x i o ∂yi o

Noting that
x c = x oc + δx c , y c = y oc + δy c , R = R o + δR ,
and
xi = x′i − vxi , yi = y′i − vyi , xio = x′i − voxi , yio = y′i − voyi ,
from where,
x i − x io = (x ′i − v xi ) − (x′i − voxi ) = − v xi + voxi , yi − yio = − v yi + voyi ,
we have,
u i (x oc , yco , R o , x io , yio ) = (x io − x oc ) 2 + (yio − yoc ) 2 − (R o )2 =
= (x′i − voxi − x co )2 + (y'i − voyi − yco )2 − (R o )2

∂u i ∂u ∂u
(x c − x oc ) + i (yc − yoc ) + i (R − R o ) =
∂x c o ∂yc o ∂R o

= −2(x io − x co ) δx c − 2(yio − yoc ) δyc − 2R o δR

⎡δx c ⎤
= −2 ⎡⎣(x ′i − voxi − x oc ) (y'i − voyi − yoc ) R ⎦ ⎢ δyc ⎥⎥ = A i xi
o⎤⎢

⎢⎣ δR ⎥⎦

∂u i ∂u
(x i − x io ) + i (yi − yio ) = 2(x io − x co ) (x i − x io ) + 2(yio − yoc )(yi − yio ) =
∂x i o ∂yi o

= 2(x ′i − voxi − x co )(x ′i − v x i − voxi ) + 2(y 'i − voyi − yoc )(y 'i − v yi − voyi ) =

⎡ v xi ⎤
= −2 ⎡⎣(x ′i − vox i − x oc ) (y 'i − voyi − yoc ) ⎤⎦ ⎢ ⎥ +
⎣⎢ v yi ⎦⎥

80 A. Fotiou
 ⎡ x ′ − v ox ⎤ ⎡ x ′i − vox ⎤
o ⎤⎢ i
+ 2 ⎡⎣ (x ′i − v ox i x co ) v oyi ⎥ = −B i v i + B i ⎢ ⎥.
i i
− (y 'i − − yc ) ⎦
⎢ y 'i − v oy ⎥ ⎢ y 'i − voy ⎥
⎣ i ⎦ ⎣ i ⎦

With
⎡ x ′i − v ox ⎤
w i = (x ′i − v ox i x co ) 2 v oyi y oc ) 2 o 2
− (R ) + Bi ⎢ ⎥
i
− + (y 'i − −
⎢ y 'i − v oy ⎥
⎣ i ⎦

we obtain the linear system,

w + Ax − Bv = 0 (7.7)
A clarification on the iterating LS adjustment algorithm until convergence is
achieved. We begin with the zeroth iteration or the first solution, by choosing
y o = y b from where v o is zeroed and matrices A, B and w are functions of
(y b , xo ) as happens in the classical treatment. Therefore, the LS solution gives
( vˆ o , xˆ o ) from where (yˆ o = y b − vˆ o , xˆ α (o) = xo + xˆ o ) . With the new approximate
values (yˆ o , xˆ α (o) ) , the first iteration or 2nd solution gives ( vˆ o(1) , xˆ o(1) ) and
(yˆ o(1) = y b − vˆ o(1) , xˆ α (1) = xˆ α (o) + xˆ o(1) ) . Similarly, the nth iteration (n+1 solution)
results in ( vˆ o(n) , xˆ o(n) ) and (yˆ o(n) = y b − vˆ o(n) , xˆ α (n) = xˆ α (n −1) + xˆ o(n) ) . Setting a
converge limit on the (absolute) differences between two successive estimations,
i.e., (| vˆ o(n) − vˆ o(n −1) | ≤ e v , | xˆ o(n) − xˆ o(n −1) | ≤ e x ) , the final LS-estimates are ob-
tained. To prevent many iterations and probably a diverse solution to a local mini-
mum, starting values x o should be computed as close as possible to the correct
estimates, usually by a subset of observations.
Details for the previous scheme of iterations are extensively discussed by Pope
(1972) where some pitfalls are also underlined. Some other examples and ap-
proaches are given, e.g., in Schaffrin and Wieser (2008), Pan et al. (2015). A sim-
plified mathematical and algorithmic implementation of the above Modified Mixed
Model (MMM) in practical fitting problems is presented in Fotiou (2017a, 2017b).

8. Concluding remarks
A means to study the real world and to obtain quantitative determination for vari-
ous physical problems is provided by the analysis of observations in relation to a
mathematical model that reflects the unknown reality. For that purpose, the ad-
justment of observations consists an indispensable part in geodetic sciences, geo-
matics and other related fields.
Central LS adjustment topics, about the correct choice of the adjustment
model/method and the efficient algorithmic implementation for parameter esti-

A discussion on Least Squares Adjustment with Worked Examples 81

mates and error control, need a closer look as the results may be crucial in many
practical applications. A better understanding of the involved concepts, models and
methods can be gained through simple worked examples presented in this paper,
building step by step acquaintance and developing a good and effective theoretical
framework.

References
Chernov, N. (2010). Circular and Linear Regression: Fitting Circles and Lines by Least
Squares. CRC Press, A Chapman &Hall Book.
Dermanis, Α. (1986). Adjustment of Observations and Estimation Theory, Vol. I (in greek).
Ziti editions, Thessaloniki.
Dermanis, Α. (1987). Adjustment of Observations and Estimation Theory, Vol. II (in
greek). Ziti editions, Thessaloniki.
Dermanis, Α. and A. Fotiou (1992). Methods and Application of Observation Adjustment
(in greek). Ziti Publisher, Thessaloniki.
Fotiou, A. (2007). Geometric Geodesy-Theory and Practice. Ziti editions, Thessaloniki (in
greek).
Fotiou, A. (2017a). Line of Best Fit by a Least Squares Modified Mixed Model. Volume
‘Living with GIS’ in honour of the memory of Prof. I. Paraschakis, School of Rural and
Surveying Engineering, Aristotle University of Thessaloniki, 57-80.
Fotiou, A. (2017b). Computationally Efficient Methods and Solutions with Least Squares
Similarity Transformation Models. Volume ‘Living with GIS’ in honour of the memory
of Prof. I. Paraschakis, School of Rural and Surveying Engineering, Aristotle University
of Thessaloniki, 107-128.
Koch, K.R. (1987). Parameter Estimation and Hypothesis Testing in Linear Models.
Springer-Verlang.
Mikhail, E.M. and F. Ackermann (1976). Observations and Least Squares. IEP-A Dun-
Donnelley Publisher.
Pan, G., Y. Zhou, H. Sun, and W. Guo (2015). Linear observation based total least squares.
Survey Review, 47:340, 18-27.
Pope, A. J. (1972). Some pitfalls to be avoided in the iterative adjustment of nonlinear
problems. In Proc.: 38th Ann. Meeting, American Society of Photogrammetry, Washing-
ton, D.C., 449-473.
Rossikopoulos, D. (1999). Surveying Networks and Computations (in greek). Ziti editions.
Schaffrin, B. and A. Wieser (2008). On weighted total least-squares adjustment for linear
regression. J Geod., 82, 415– 421.
Sneew, N., F. Krumm, M. Roth (2015). Adjustment Theory. Lecture Notes, Rev. 4.43,
Geod. Inst. Univ. Stuttgart.
Wells, D.E. and E.J. Krakiwsky (1971). The Method of Least Squares. Lecture Notes, De-
partment of Geodesy and Geomatics Engineering, U.N.B.

82 A. Fotiou