RMIT

Geospatial Science

NOTES ON
LEAST SQUARES
Copies of copyright material in this compilation have been made in
accordance with the provisions of Section V(b) of the Copyright Act for the
teaching purposes of the University.

RMIT
Geospatial Science
School of Mathematical and Geospatial Science
NOTES ON LEAST SQUARES
This edition printed in 2005
CONTENTS

CHAPTER PAGES

1. Introduction 3
2. Least Squares Adjustment of Indirect Observations 57
3. Propagation of Variances 37
4. Approximate Values 5
5. Propagation of Variances Applied to Least Squares 7
Adjustment of Indirect Observations
6. Least Squares Adjustment of Observations Only 39
7. Linearization Using Taylor's Theorem and the 15
Derivation of Some Common Surveying
Observation Equations
8. The Standard Error Ellipse 10
9. Least Squares Resection 14
10. Least Squares Bearing Intersection 12

APPENDICES

A Matrix Algebra 23

REFERENCES 2
RMIT University Geospatial Science

NOTES ON LEAST SQUARES

1. INTRODUCTION

The theory of least squares and its application to adjustment of survey measurements is well
known to every geodesist. The invention of the method is generally attributed to Karl
Freidrich Gauss (1777-1855) but could equally be credited to Adrien-Marie Legendre (1752-
1833).

Gauss used the method of least squares to compute the elements of the orbit of the minor
planet Ceres and predicted its position in October 1801 from a few observations made in the
previous year. He published the technique in 1809 in Theoria Motus Corporum Coelestium in
Sectionibus Conicis Solem Ambientium (Theory of the Motion of the Heavenly Bodies
Moving about the Sun in Conic Sections), mentioning that he had used it since 1795, and also
developed what we now know as the normal law of error, concluding that: "... the most
probable system of values of the quantities ... will be that in which the sum of the squares of
the differences between the actually observed and computed values multiplied by numbers
that measure the degree of precision, is a minimum." (Gauss 1809).

Legendre published an independent development of the technique in Nouvelles méthodes pour

la détermination des orbites des comètes (New methods for the determination of the orbits of
comets), Paris, 1806 and also as the "Méthod des moindres carriés" (Method of Least
Squares), published in the Mémoires de l'Institut national des sciences at arts, vol. 7, pt. 2,
Paris, 1810.

After these initial works, the topic was subjected to rigid analysis and by the beginning of the
20th century was the universal method for the treatment of observations. Merriman (1905)
compiled a list of 408 titles, including 72 books, written on the topic prior to 1877 and
publication has continued unabated since then. Leahy (1974) has an excellent summary of the

© 2005, R.E. Deakin Notes on Least Squares (2005) 1–1

RMIT University Geospatial Science

development of least squares and clearly identifies the historical connection with
mathematical statistics, which it pre-dates.
The current literature is extensive; the books Observations and Least Squares (Mikhail 1976)
and Analysis and Adjustment of Survey Measurements (Mikhail and Gracie 1981), and lecture
notes by Cross (1992), Krakiwsky (1975) and Wells and Krakiwsky (1971) stand out as the
simplest modern treatments of the topic.

Following Wells and Krakiwsky (1971, pp.8-9), it is interesting to analyse the following
quotation from Gauss' Theoria Motus (Gauss, 1809, p.249).

"If the astronomical observations and other quantities, on which the computation of
orbits is based, were absolutely correct, the elements also, whether deduced from
three or four observations, would be strictly accurate (so far indeed as the motion is
supposed to take place exactly according to the laws of KEPLER), and, therefore, if
other observations were used, they might be confirmed, but not corrected. But since
all our measurements and observations are nothing more than approximations to the
truth, the same must be true of all calculations resting upon them, and the highest
aim of all computations made concerning concrete phenomena must be to
approximate, as nearly as practicable, to the truth. But this can be accomplished in
no other way than by a suitable combination of more observations than the number
absolutely requisite for the determination of the unknown quantities. This problem
can only be properly undertaken when an approximate knowledge of the orbit has
been already attained, which is afterwards to be corrected so as to satisfy all the
observations in the most accurate manner possible."

This single paragraph, written almost 200 years ago, embodies the following concepts, which
are as relevant today as they were then.

(i) Mathematical models may be incomplete,

(ii) Physical measurements are inconsistent,

(iii) All that can be expected from computations based on inconsistent measurements are
estimates of the "truth",

(iv) Redundant measurements will reduce the effect of measurement inconsistencies,

© 2005, R.E. Deakin Notes on Least Squares (2005) 1–2

RMIT University Geospatial Science

(v) Initial approximations to the final estimates should be used, and finally,

(vi) Initial approximations should be corrected in such a way as to minimise the

inconsistencies between measurements (by which Gauss meant his method of least
squares).

These notes contain a development of Least Squares processes applicable to surveying and
geodesy. Examples and exercises of least squares processes are given using MATLAB, an
interactive, matrix-based system for scientific and engineering computation and visualization.
The name MATLAB is derived from MATrix LABoratory and is licensed by The
MathWorks, Inc.

© 2005, R.E. Deakin Notes on Least Squares (2005) 1–3

RMIT University Geospatial Science

2. LEAST SQUARES ADJUSTMENT OF INDIRECT OBSERVATIONS

2.1. Introduction

The modern professional surveyor must be competent in all aspects of surveying

measurements such as height differences, linear distances, horizontal and vertical angle
measurements and combinations thereof which form the fundamental observations used to
determine position in space. To obtain these measurements with any degree of confidence the
surveyor must be aware of the principles and operation of various pieces of surveying
equipment as well as the nature of measurements and the possible effects of errors on these
measurements and any derived quantities. The nature of errors in measurements, studied by
Gauss and leading to his theory of errors (the normal law of error) is the basis of statistical
rules and tests that the surveyor employs to assess the quality of measurements; these rules
and tests are covered in basic statistics courses during the undergraduate degree program. In
the simple least squares processes and applications which follow it is sufficient to assume that
the measurements are affected by small accidental or random errors and the least squares
"solutions" provide a means of determining the best estimate of a measured quantity. Least
squares solutions also imply that the quantity of interest has been determined from a
redundant system of measurements, i.e., there are more measurements than the minimum
number required to calculate the quantity.

2.1.1. Definition and classification of measurements

Crandall and Seabloom (1970, pp. 4-5) give a definition of a measurement as:

A measurement is a comparison between an unknown quantity and a

predefined standard, determined by some measuring device and hence,
any measured value is an approximation of the exact or true value, not
the true value itself. Since the true value of a quantity cannot be
measured, any measurement contains by definition, an error.

Direct measurements (or observations) are those that are made directly upon the quantity to be
determined. Measurements of a line by direct chaining, or Electronic Distance Measurement
(EDM), or measurement of an angle by theodolite or Total Station are examples of direct
measurements.

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–1

RMIT University Geospatial Science

Indirect measurements (or observations) are not made upon the quantity itself but are made on
some other quantity or quantities related to it. For example, the coordinates of a point P are
indirectly determined by measuring bearings and distances to P from other points; the latitude
of P may be determined from altitudes to certain stars; and the height of P may be determined
by measured height differences from a known point.

2.1.2. Classification of errors of measurement

Since, by definition, every measurement contains an error it is necessary to consider the

various kinds of errors that occur in practice. Rainsford (1968, p. 1) provides a derivation of
the word error as:

coming from the Latin errare which means to wander and not to sin.

Rainsford divides errors into four classes

(a) blunders or mistakes

(b) constant errors
(c) systematic errors
(d) accidental or random errors

Blunders or mistakes are definite mis-readings, booking errors or other like occurrences.
They are usually caused by poor measurement technique and/or a lack of attention to detail by
the person making the measurement. They may be eliminated or minimized by correct and
careful measurement techniques, and a thorough understanding of the operation of the
equipment used for the measurement.

Constant errors are those that do not vary throughout the particular measurement period.
They are always of the same sign. Neglecting to standardize a measuring tape introduces a
constant error; failure to use the correct prism-offset value introduces constant errors in EDM
measurements. A faulty joint between sections of a levelling staff will introduce a constant
error into height differences from spirit levelling. Constant errors can be eliminated from
measurements by a thorough understanding of the measurement process and the equipment
used.

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–2

RMIT University Geospatial Science

Systematic errors are those errors that follow some fixed law (possibly unknown) dependent
on local conditions and/or the equipment being used. For example, if the temperature and
pressure (which are indicators of atmospheric conditions) are not measured when using EDM
equipment then a systematic error may be introduced, since the modulated electromagnetic
beam of the EDM passes through the atmosphere and its time of travel (indirectly measured
by phase comparison of emitted and reflected beams) is affected by atmospheric conditions.
All EDM measurements must be corrected for atmospheric conditions that depart from
"standard conditions".

Accidental or Random errors are the small errors remaining in a measurement after mistakes,
constant errors and systematic errors have been eliminated. They are due to the imperfection
of the instruments used, the fallibility of the observer and the changing environmental
conditions in which the measurements are made, all of which affect the measurement to a
lesser or greater degree.

Bearing in mind the aforementioned, it could be said that all careful measurements (where
mistakes, constant errors and systematic errors have been eliminated) contain small random
errors and from experience, three axioms relating to random errors can be stated.

1. Small errors occur more frequently, or are more probable then

large errors.
2. Positive and negative errors of the same magnitude are equally
probable
3. Very large errors do not occur.

These axioms are the basic premises on which the theory of errors (the normal law of error) is
founded.

2.1.3. Errors, corrections and residuals

A measured quantity has a true value and a most probable value. The most probable value is
often called the best estimate and the two terms can be taken as synonymous.

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–3

RMIT University Geospatial Science

No matter how many times a quantity is measured, its true value will remain unknown and
only a best estimate can be obtained from the measurements. In the case of a single measured
quantity, the best estimate is the arithmetic mean (or average) of the measurements.

If a quantity has been measured a number of times, the difference between the true (but
unknown) value and any measurement is the true error and the difference between the best
estimate and any measurement is the apparent error.

These relationships can be established by defining a correction to have the same magnitude as
an error but the opposite sign. In surveying, the terms correction and residual are regarded as
synonymous, and are universally denoted by the letter v.

Suppose an unknown quantity x is measured n times giving values x1 , x2 , x3 , … , xn . The

true value (unknown) of the measured quantity is μ (mu) and is estimated by the arithmetic
mean x where
n

x +x + + xn ∑x k
x= 1 2 = k =1
(2.1)
n n

The arithmetic mean is regarded as the best estimate or most probable value. A correction v
having the same magnitude as an error but the opposite sign is defined as

vk = x − xk

Since these corrections relate to the measurements and arithmetic mean, they could be called
apparent corrections and hence according to our definition of corrections and errors, apparent
errors −v are defined as

− vk = xk − x

In a similar manner, we may define true errors ε (epsilon) as

ε k = xk − μ

These relationships may be expressed as

measurement + residual = best estimate

measurement − best estimate = apparent error
measurement − true value = true error

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–4

RMIT University Geospatial Science

True errors are unknown and are approximated by apparent errors. The closer the best
estimate (or most probable value) approaches the true value, the closer the apparent error
approaches the true error. The laws defining the nature and behaviour of true errors were
derived from practical axioms deduced from the nature of apparent errors and hence any
theory of errors may also be regarded as a theory of corrections (or residuals) and the
distinction between true errors and apparent errors is ignored for all practical purposes.

The following sections contain simple examples of least squares processes, the mean, the
weighted mean, line of best fit (linear regression) and polynomial curve fitting. In each case,
Gauss' least squares principle: "... the most probable system of values of the quantities ... will
be that in which the sum of the squares of the differences between the actually observed and
computed values multiplied by numbers that measure the degree of precision, is a minimum."
will be employed to determine equations or systems of equations that may be regarded as least
squares solutions to the problems. Furthermore, it is assumed that all measurements are free
of mistakes, constant errors and systematic errors and "contain" only random errors and that
the precision of the measurements is known a priori (Latin a priori from what is before).
Solutions to some of the examples are provided as MATLAB script files (.m files).

2.2. The Mean

It is well known practice that when a single quantity is measured a number of times the
arithmetic mean is taken as the best estimate of the measured quantity. Few people realise
that when they adopt this practice they are employing Gauss' least squares principle.

Consider a series of measurements x1 , x2 , x3 , … , xn of a quantity and denote the best

estimate of this quantity as p. According to our general definition of measurements and
corrections we may write: measurement + correction (or residual) = best estimate or

x1 + v1 = p, x2 + v2 = p, x3 + v3 = p, , xn + vn = p

These equations can be rearranged as

v1 = p − x1 , v2 = p − x2 , v3 = p − x3 , , vn = p − xn

Now if all the measurements can be regarded as having equal precision we may state the least
squares principle as

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–5

RMIT University Geospatial Science

The best estimate p is that value which makes the sum of the squares
of the residuals a minimum.

We may define a least squares function ϕ (phi) as

n
ϕ = the sum of the squares of the residuals = ∑ vk2 (2.2)
k =1

n
ϕ = ∑ vk2 = ( p − x1 ) + ( p − x2 ) + + ( p − xn )
2 2 2
or
k =1

We say that ϕ is a function of p, the single parameter or variable in this equation. The
minimum value of the function (i.e. making the sum of squares of residuals a minimum) can
dϕ
be found by equating the derivative to zero, i.e.,
dp

dϕ
ϕ is a minimum when =0
dp

dϕ
and = 2 ( p − x1 ) + 2 ( p − x2 ) + + 2 ( p − xn ) = 0
dp

Cancelling the 2's and rearranging gives the best estimate p as the arithmetic mean.
n

x + x + x3 + + xn ∑x k
p= 1 2 = k =1
(2.3)
n n

Hence, the arithmetic mean of a series of measurements is the best estimate according to
Gauss' least squares principle.

2.3. The Weighted Mean

Before demonstrating that the weighted mean of a set of observations is the result of a least
squares process, some discussion of the term weight and its connection with precision is
required.

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–6

RMIT University Geospatial Science

2.3.1. Measures of Precision of a Finite Population

In every least squares process it is assumed that the precision of measurements is known. The
precision is a measure of the dispersion (or spread) of a number of measurements from their
mean (or average) value. A common statistical measure of precision is the variance σ 2 and
the positive square root of the variance is the standard deviation σ . Equations for the
variance and standard deviation differ depending on whether the population of measurements
is finite or infinite and a population is a potential set of quantities that we want to make
inference about based on a sample from that population.

Following Deakin and Kildea (1999), consider a finite population, such as the examination
marks mk of a group of N students in a single subject. Since we have complete information

about the population, i.e., its size is known, the mean μ , the variance σ 2 and the standard
deviation σ of the finite population are
N

∑m k
μ= k =1
(2.4)
N

∑ (m − μ)
2
k
σ2 = k =1
(2.5)
N

∑(m − μ)
2
k
σ= k =1
(2.6)
N

Note that the variance σ 2 is the average squared difference of a member of the population mk

from the population mean μ . The mean, variance and standard deviation are known as
population parameters.

2.3.2. Estimates of Precision of Samples of an Infinite Population

Consider surveying measurements, drawn from infinite populations with the attendant
difficulties of estimation since population averages can never be known. In such cases we are
usually dealing with small samples of measurements of size n and we can only obtain
estimates of the unknown population parameters μ , σ 2 and σ . For a sample of n

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–7

RMIT University Geospatial Science

measurements x1 , x2 , x3 , … , xn from an infinite population, estimates of the mean, variance

and standard deviation, denoted by x , s x2 and s x are

1 n
x= ∑ xk
n k =1
(2.7)

1 n
s x2 = ∑ ( xk − x )
2
(2.8)
n − 1 k =1

1 n
sx = ∑ ( xk − x )
2
(2.9)
n − 1 k =1

Note the divisor n − 1 (known as the degrees of freedom) in equations for the estimates of
variance and standard deviation. This ensures that s x2 is an unbiased estimate of the

population variance σ 2 , but does not ensure that s x is an unbiased estimate of the population

standard deviation σ ; the action of "taking a square-root" changes the property of

unbiasedness. This is more an accident of mathematics rather than a cause of faulty
estimation but it is not well appreciated in general. Deakin and Kildea (1999, p. 76) show that
an unbiased estimator s ∗x of the population standard deviation σ is given by

n
1
s ∗x = ∑( x − x)
2
k (2.10)
cn k =1

Values of cn are given in Table 2.1

n 2 3 4 5 10 15 20 30 90
n-1 1 2 3 4 9 14 19 29 89
cn 0.64 1.57 2.55 3.53 8.51 13.51 18.51 28.50 88.50

Table 2.1 Values of divisor cn for unbiased estimation of σ

In these notes, it is always assumed that the terms mean, variance and standard deviation refer
to estimates of population values.

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–8

RMIT University Geospatial Science

2.3.3. Relationship between Weights and Estimates of Variance

Another measure of precision, often used in least squares applications is weight w and the
weight of an observation (or measurement) is defined as being inversely proportional to the
variance

1
wk ∝ (2.11)
sk2

σ 02
or wk = (2.12)
sk2

σ 02 is a constant of proportionality known as the reference variance or variance factor. This is

the classical definition of weight and if an observation has unit weight ( wk = 1) its variance

equals σ 02 , hence the reference variance is sometimes called the variance of an observation of
unit weight; a term often encountered in older surveying texts. In this definition of weight,
there is an assumption that the measurements are uncorrelated (a statistical term relating to the
dependence between measurements, see section 2.5). In cases where measurements are
correlated, weights are not an adequate means of describing relative precisions.

As an example of the connection between weights and standard deviations consider three
uncorrelated (i.e., independent) observations of a particular distance, where each observation
is the mean of several measurements and standard deviations of each observation have been
determined from the measurements

observation 1 136.225 m (st. dev. 0.010 m)

observation 2 136.233 m (st. dev. 0.032 m)
observation 3 136.218 m (st. dev. 0.024 m)

Since the weight is inversely proportional to the variance, the observation with the smallest
weight will have the largest variance (standard deviation squared). For convenience, this
observation is given unit weight i.e., w2 = 1 and the other observations (with smaller
variances) will have higher weight. Hence from (2.12)

σ 02
w2 = 1 = and σ 02 = ( 0.032 )
2

( 0.032 )
2

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–9

RMIT University Geospatial Science

The weights of the three observations are then

( 0.032 ) = 10.24
2

w1 =
( 0.010 )
2

( 0.032 )
2

w2 = 2 =
1
( 0.032 )
( 0.032 )
2

w3 = = 1.78
( 0.024 )
2

Weights are often assigned to observations using "other information". Say for example, a
distance is measured three times and a mean value determined. If two other determinations of
the distance are from the means of six and four measurements respectively, the weights of the
three observations may simply be assigned the values 3, 6 and 4. This assignment of weights
is a very crude reflection of the (likely) relative precisions of the observations since it is
known that to double the precision of a mean of a set of measurements, we must quadruple
the number of measurements taken (Deakin and Kildea, 1999, p. 76).

2.3.4. Derivation of Equation for the Weighted Mean

Consider a set of measurements of a quantity as x1 , x2 , x3 , … , xn each having weight

w1 , w2 , w3 , … , wn and denote the best estimate of this quantity as q. According to our general

definition of measurements and corrections we may write:

measurement + correction (or residual) = best estimate

or x1 + v1 = q, x2 + v2 = q, x3 + v3 = q, , xn + v n = q

These equations can be rearranged as

v1 = q − x1 , v2 = q − x2 , v3 = q − x3 , , v n = q − xn

Now each measurement has a weight reflecting relative precision and we may state the least
squares principle as
The best estimate q is that value which makes the sum of the squares
of the residuals, multiplied by their weights, a minimum.
We may define a least squares function ϕ (phi) as

n
ϕ = the sum of the weighted squared residuals = ∑ wk vk2 (2.13)
k =1

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–10

RMIT University Geospatial Science

n
ϕ = ∑ wk vk2 = w1 ( q − x1 ) + w2 ( q − x2 ) + + wn ( q − xn )
2 2 2
or
k =1

We say that ϕ is a function of q, the single parameter or variable in this equation. The
minimum value of the function (i.e., making the sum of the weighted squared residuals a
dϕ
minimum) can be found by equating the derivative to zero, i.e.,
dq

dϕ
ϕ is a minimum when =0
dq

dϕ
and = 2 w1 ( q − x1 ) + 2 w2 ( q − x2 ) + + 2 wn ( q − xn ) = 0
dq

Cancelling the 2's and expanding gives

w1q − w1 x1 + w2 q − w2 x2 + + wn q − wn xn = 0

Rearranging gives the weighted arithmetic mean q

n

w x + w2 x2 + + wn xn ∑w x k k
q= 1 1 = k =1
(2.14)
w1 + w2 + + wn n

∑w
k =1
k

Hence, the weighted arithmetic mean of a series of measurements xk each having weight wk
is the best estimate according to Gauss' least squares principle.

It should be noted that the equation for the weighted mean (2.14) is valid only for
measurements that are statistically independent. If observations are dependent, then a
measure of the dependence between the measurements, known as covariance, must be taken
into account. A more detailed discussion of weights, variances and covariances is given in
later sections of these notes.

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–11

RMIT University Geospatial Science

2.4. Line of Best Fit

5 •

c
•4 x+
y =m

2• C • 3

1•

Figure 2.1 Line of Best Fit through data points 1 to 5

The line of best fit shown in the Figure 2.1 has the equation y = m x + c where m is the slope

⎛ y − y1 ⎞
of the line ⎜ m = tan θ = 2 ⎟ and c is the intercept of the line on the y axis.
⎝ x2 − x1 ⎠

m and c are the parameters and the data points are assumed to accord with the mathematical
model y = m x + c . Obviously, only two points are required to define a straight line and so
three of the five points in Figure 2.1 are redundant measurements (or observations). In this
example the x,y coordinate pairs of each data point are considered as indirect measurements of
the parameters m and c of the mathematical model.

To estimate (or compute) values for m and c, pairs of points in all combinations (ten in all)
could be used to obtain average values of the parameters; or perhaps just two points selected
as representative could be used to determine m and c.

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–12

RMIT University Geospatial Science

A better way is to determine a line such that it passes as close as possible to all the data
points. Such a line is known as a Line of Best Fit and is obtained (visually) by minimising the
differences between the line and the data points. No account is made of the "sign" of these
differences, which can be considered as corrections to the measurements or residuals. The
Line of Best Fit could also be defined as the result of a least squares process that determines
estimates of the parameters m and c such that those values will make the sum of the squares of
the residuals, multiplied by their weights, a minimum. Two examples will be considered, the
first with all measurements considered as having equal precisions, i.e., all weights of equal
value, and the second, measurements having different precisions, i.e., unequal weights.

2.4.1. Line of Best Fit (equal weights)

In Figure 2.1 there are five data points whose x,y coordinates (scaled from the diagram in
mm's) are

Point x y
1 −40.0 −24.0
2 −15.0 −24.0
3 10.0 −12.0
4 38.0 15.0
5 67.0 30.0

Table 2.2 Coordinates of data points (mm's) shown in Figure 2.1

Assume that the data points accord with the mathematical model y = m x + c and each
measurement has equal precision. Furthermore, assume that the residuals are associated with
the y values only, which leads to an observation equation of the form

y k + v k = m xk + c (2.15)

By adopting this observation equation we are actually saying that the measurements (the x,y
coordinates) don't exactly fit the mathematical model, i.e., there are inconsistencies between
the model and the actual measurements, and these inconsistencies (in both x and y
measurements) are grouped together as residuals vk and simply added to the left-hand-side of
the mathematical model. This is simply a convenience. We could write an observation
equation of the form

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–13

RMIT University Geospatial Science

(
yk + v yk = m xk + v xk + c )
v xk , v yk are residuals associated with the x and y coordinates of the kth point. Observation

equations of this form require more complicated least squares solutions and are not considered
in this elementary section.

Equations (2.15) can be re-written as residual equations of the form

vk = m xk + c − y k (2.16)

The distinction here between observation equations and residual equations is simply that
residual equations have only residuals on the left of the equals sign. Rearranging observation
equations into residual equations is an interim step to simplify the function ϕ = sum of
squares of residuals.

Since all observations are of equal precision (equal weights), the least squares function to be
minimised is
n
ϕ = the sum of the squares of the residuals = ∑ vk2
k =1

5
or ϕ = ∑ vk2 = ( m x1 + c − y1 ) 2 + ( m x2 + c − y2 )2 + .... + (m x5 + c − y5 )2
k =1

ϕ is a function of the u = 2 "unknown" parameters m and c and so to minimise the sum of

∂ϕ ∂ϕ
squares of residuals, the partial derivatives and are equated to zero.
∂m ∂c

∂φ
= 2( m x1 + c − y1 )( x1 ) + 2( m x2 + c − y2 )( x2 ) + ... + 2( m x5 + c − y5 )( x5 ) = 0
∂m
∂φ
= 2( m x1 + c − y1 )(1) + 2( m x2 + c − y2 )(1) + ... + 2( m x5 + c − y5 )(1) =0
∂c

Cancelling the 2's, simplifying and re-arranging gives two normal equations of the form
n n n
m ∑ xk2 + c ∑ xk = ∑x k yk
k =1 k =1 k =1
n n
(2.17)
m ∑ xk + c n = ∑y k
k =1 k =1

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–14

RMIT University Geospatial Science

The normal equations can be expressed in matrix form as

⎡ ∑ xk2 ∑ x ⎥⎤ ⎡m ⎤ ⎡ ∑ xk y k ⎤
⎢
k
= ⎢ ⎥ (2.18)
⎣ ∑ xk n ⎦ ⎢⎣ c ⎥⎦ ⎣ ∑ yk ⎦

or Nx = t (2.19)

Matrix algebra is a powerful mathematical tool that simplifies the theory associated with least
squares. The student should become familiar with the terminology and proficient with the
algebra. Appendix A contains useful information relating to matrix algebra.

⎡n n12 ⎤
N = ⎢ 11 is the ( u, u ) normal equation coefficient matrix
⎣ n21 n22 ⎥⎦

⎡x ⎤
x = ⎢ 1⎥ is the ( u,1) vector of parameters (or "unknowns"), and
⎣ x2 ⎦

⎡t ⎤
t = ⎢ 1⎥ is the ( u,1) vector of numeric terms.
⎣ t2 ⎦

The solution of the normal equations for the vector of parameters is

x = N −1 t (2.20)

In this example (two equations in two unknowns) the matrix inverse N −1 is easily obtained
(see Appendix A 4.8) and the solution of (2.20) is given as

⎡x ⎤ 1 ⎡ n22 −n12 ⎤ ⎡ t1 ⎤
x = ⎢ 1⎥ = (2.21)
⎣ x2 ⎦ ( n11 n22 − n12 n21 ) ⎢⎣ −n21 n11 ⎥⎦ ⎢⎣ t2 ⎥⎦

From the data given in Table 2.2, the normal equations are

⎡ 7858.00 60.00⎤ ⎡ m ⎤ ⎡3780.00⎤

⎢ 60.00 5.00⎥ ⎢ c ⎥ = ⎢ −15.00⎥
⎣ ⎦⎣ ⎦ ⎣ ⎦

and the solutions for the best estimates of the parameters m and c are

⎡m⎤ 1 ⎡ 5.00 −60.00 ⎤ ⎡ 0.554777 ⎤

⎢ c ⎥ = 7858.00 5.00 − 60.00 60.00 ⎢ −60.00 7858.00⎥ = ⎢ −9.657327 ⎥
⎣ ⎦ ( )( ) ( )( )⎣ ⎦ ⎣ ⎦

Substitution of the best estimates of the parameters m and c into the residual equations
vk = m xk + c − yk gives the residuals (mm's) as

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–15

RMIT University Geospatial Science

v1 = −7.8
v2 = 6.0
v3 = 7.9
v4 = −3.6
v5 = −2.5

2.4.2. Line of Best Fit (unequal weights)

Consider again the Line of best Fit shown in Figure 2.1 but this time the x,y coordinate pairs
are weighted, i.e., some of the data points are considered to have more precise coordinates
than others. Table 2.3 shows the x,y coordinates (scaled from the diagram in mm's) with
weights.

Point x y weight w
1 −40.0 −24.0 2
2 −15.0 −24.0 5
3 10.0 −12.0 7
4 38.0 15.0 3
5 67.0 30.0 3

Table 2.3 Coordinates (mm) and weights of data points shown in Figure 2.1

Similarly to before, a residual equation of the form given by (2.16) can be written for each
observation but this time a weight wk is associated with each equation and the least squares
function to be minimised is
n
ϕ = the sum of the weighted squared residuals = ∑ wk vk2
k =1

5
or ϕ = ∑ wk vk2 = w1 ( m x1 + c − y1 ) 2 + w2 ( m x2 + c − y2 )2 + .... + w5 ( m x5 + c − y5 ) 2
k =1

ϕ is a function of the u = 2 "unknown" parameters m and c and so to minimise ϕ the partial

∂ϕ ∂ϕ
derivatives and are equated to zero.
∂m ∂c

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–16

RMIT University Geospatial Science

∂φ
= 2 w1 ( m x1 + c − y1 )( x1 ) + 2 w2 ( m x2 + c − y2 )( x2 ) + ... + 2 w5 ( m x5 + c − y5 )( x5 ) = 0
∂m
∂φ
= 2 w1 ( m x1 + c − y1 )(1) + 2 w2 ( m x2 + c − y2 )(1) + ... + 2 w5 ( m x5 + c − y5 )(1) =0
∂c

Cancelling the 2's simplifying and re-arranging gives two normal equations of the form
n n n
m ∑ wk xk2 + c ∑ wk xk = ∑w x k k yk
k =1 k =1 k =1
n n n
m ∑ wk xk + c ∑ wk = ∑w y k k
k =1 k =1 k =1

The normal equations expressed in matrix form Nx = t are

⎡ ∑ wk xk2 ∑w x ⎤ ⎡m⎤ ⎡ ∑ wk xk yk ⎤
⎥⎢ ⎥ = ⎢
k k
⎢ ⎥
⎣ ∑ wk xk ∑w k ⎦⎣c⎦ ⎣ ∑ wk yk ⎦

Substituting the data in Table 2.3, the normal equations are

⎡ 22824.00 230.00⎤ ⎡ m ⎤ ⎡10620.00⎤

⎢ 230.00 20.00⎥ ⎢ c ⎥ = ⎢ −117.00⎥
⎣ ⎦⎣ ⎦ ⎣ ⎦

The solution for the best estimates of the parameters m and c is found in exactly the same
manner as before (see section 2.4.1)

m = 0.592968
c = −12.669131

Substitution of m and c into the residual equations vk = m xk + c − yk gives the residuals

(mm's) as

v1 = −12.4
v2 = 2.4
v3 = 5.3
v4 = −5.1
v5 = −2.9

Comparing these residuals with those from the Line of Best Fit (equal weights), shows that
the line has been pulled closer to points 2 and 3, i.e.; the points having largest weight.

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–17

RMIT University Geospatial Science

2.5. Variances, Covariances, Cofactors and Weights

Some of the information in this section has been introduced in previously in section 2.3 The
Weighted Mean and is re-stated here in the context of developing general matrix expressions
for variances, covariances, cofactors and weights of sets or arrays of measurements.

In surveying applications, we may regard a measurement x as a possible value of a continuous

random variable drawn from an infinite population. To model these populations, and thus
estimate the quality of the measurements, probability density functions have been introduced.
In surveying, Normal (Gaussian) probability density functions are the usual model. A
probability density function is a non-negative function where the area under the curve is one.
+∞
For f ( x ) ≥ 0 and ∫ −∞
f ( x ) dx = 1 the values of f ( x ) are not probabilities. The probability a
b
member of the population lies in the interval a to b is ∫ a
f ( x ) dx . The population mean μ ,

population variance σ x2 and the family of Normal probability density functions are given by
Kreyszig (1970) as
+∞
μx = ∫ x f ( x ) dx (2.22)
−∞

+∞
σ x2 = ∫ ( x − μx )
2
f ( x ) dx (2.23)
−∞

2
1 ⎛ x−μx ⎞
− ⎜ ⎟
1 2⎝ σ x ⎠
f ( x; μ x , σ x ) = e (2.24)
σ x 2π

Since the population is infinite, means and variances are never known, but may be estimated
from a sample of size n. The sample mean x and sample variance s x2 , are unbiased estimates

of the population mean μ x and population variance σ x2

1 n
x= ∑ xk
n k =1
(2.25)

1 n
s x2 = ∑ ( xk − x )
2
(2.26)
n − 1 k =1

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–18

RMIT University Geospatial Science

The sample standard deviation s x is the positive square root of the sample variance and is a

measure of the precision (or dispersion) of the measurements about the mean x .

In least squares applications, an observation may be the mean of a number measurements or a

single measurement. In either case, it is assumed to be from an infinite population of
measurements having a certain (population) standard deviation and that an estimate this
standard deviation is known.

When two or more observations are jointly used in a least squares solution then the
interdependence of these observations must be considered. Two measures of this
interdependence are covariance and correlation. For two random variables x and y with a
joint probability density function f ( x, y ) the covariance σ x y is

∫ ( x − μ ) ( y − μ ) f ( x, y ) dx dy
+∞ +∞
σxy = ∫ x y (2.27)
−∞ −∞

and the correlation coefficient ρ is given by

σxy
ρx y = (2.28)
σ xσ y

The correlation coefficient ρ will vary between ±1 . If ρ x y = 0 random variables x and y

are said to be uncorrelated and, if ρ x y = ± 1 , x and y are linked by a linear relationship

(Kreyszig 1970, pp.335-9). Correlation and statistical dependence are not the same, although
both concepts are used synonymously. It can be shown that the covariance σ x y is always

zero when the random variables are statistically independent (Kreyszig 1970, p.137-9).
Unfortunately, the reverse is not true in general. Zero covariance does not necessarily imply
statistical independence. Nevertheless, for multivariate Normal probability density functions,
zero covariance (no correlation) is a sufficient condition for statistical independence (Mikhail
1976, p.19).

The sample covariance s x y between n pairs of values ( x1 , y1 ) , ( x2 , y2 ) , ..., ( xn , yn ) with

means x and y is (Mikhail 1976, p.43)

1 n
sx y = ∑ ( xk − x )( yk − y )
n − 1 k =1
(2.29)

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–19

RMIT University Geospatial Science

Variances and covariances of observations can be conveniently represented using matrices.

For n observations x1 , x2 , x3 , ..., xn with variances σ 12 , σ 22 , σ 32 ,..., σ n2 and covariances

σ 12 , σ 13 ,... the variance-covariance matrix Σ is defined as

⎡ σ 12 σ 12 σ 13 ... σ 1n ⎤
⎢ ⎥
⎢ σ 21 σ 22 σ 23 ... σ 2 n ⎥
Σ = (2.30)
⎢ .... .... .... .... ⎥
⎢ 2 ⎥
⎣⎢σ n1 σ n 2 σ n 3 ... σ n n ⎥⎦

Note that the variance-covariance matrix Σ is symmetric since in general σ k j = σ j k .

In practical applications of least squares, population variances and covariances are unknown
and are replaced by estimates s12 , s22 , … , sn2 and s12 , s13 , … or by other numbers

representing relative variances and covariances. These are known as cofactors and the
cofactor matrix Q, which is symmetric, is defined as

⎡ q11 q12 q13 ... q1n ⎤

⎢q q22 q23 ... q2 n ⎥
Q = ⎢ ⎥
21
(2.31)
⎢ .... .... .... .... ⎥
⎢ ⎥
⎣⎢ qn1 qn 2 qn 3 ... qn n ⎦⎥

The relationship between variance-covariance matrices and cofactor matrices is

Σ = σ 02 Q (2.32)

σ 02 is a scalar quantity known as the variance factor. The variance factor is also known as the
reference variance and the variance of an observation of unit weight (see section 2.3 for
further discussion on this subject).

The inverse of the cofactor matrix Q is the weight matrix W.

W = Q −1 (2.33)

Note that since Q is symmetric, its inverse W is also symmetric. In the case of uncorrelated
observations, the variance-covariance matrix Σ and the cofactor matrix Q are both diagonal
matrices (see Appendix A) and the weight of an observation w is a value that is inversely
proportional to the estimate of the variance i.e.,

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–20

RMIT University Geospatial Science

wkk = σ 02 qkk or wkk = σ 02 skk2 (2.34)

For uncorrelated observations, the off-diagonal terms will be zero and the double subscripts
may be replaced by single subscripts; equation (2.34) becomes

wk = σ 02 sk2 (2.35)

This is the classical definition of a weight where σ 02 is a constant of proportionality.

Note: The concept of weights has been extensively used in classical least squares theory but
is limited in its definition to the case of independent (or uncorrelated) observations.
(Mikhail 1976, pp.64-65 and Mikhail and Gracie 1981, pp.66-68).

2.6. Matrices and Simple Least Squares Problems

Matrix algebra is a powerful mathematical tool that can be employed to develop standard
solutions to least squares problems. The previous examples of the Line of Best Fit will be
used to show the development of standard matrix equations that can be used for any least
squares solution.

In previous developments, we have used a least squares function ϕ as meaning either the sum
of squares of residuals or the sum of squares of residuals multiplied by weights.

In the Line of Best Fit (equal weights), we used the least squares function
n
ϕ = the sum of the squares of the residuals = ∑ vk2
k =1

If the residuals vk are elements of a (column) vector v, the function ϕ can be written as the
matrix product

⎡ v1 ⎤
n ⎢v ⎥
ϕ = ∑ vk2 = [ v1 v2 vn ] ⎢ 2 ⎥ = v T v
k =1 ⎢ ⎥
⎢ ⎥
⎣ vn ⎦

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–21

RMIT University Geospatial Science

In the Line of Best Fit (unequal weights), we used the least squares function
n
ϕ = the sum of the weighted squared residuals = ∑ wk vk2
k =1

If the residuals vk are elements of a (column) vector v and the weights are the diagonal

elements of a diagonal weight matrix W, the function ϕ can be written as the matrix product

⎡ w1 0 0 0 ⎤ ⎡ v1 ⎤
n ⎢0 w2 0 0 ⎥ ⎢v ⎥
ϕ = ∑ wk vk2 = [ v1 v2 vn ] ⎢ ⎥ ⎢ 2 ⎥ = v T Wv
k =1 ⎢0 0 0 ⎥⎢ ⎥
⎢ ⎥⎢ ⎥
⎣0 0 0 wn ⎦ ⎣ vn ⎦

Note that in this example the weight matrix W represents a set of uncorrelated measurements.

In general, we may write least squares function as a matrix equation

ϕ = vT Wv (2.36)

Note that replacing W with the identity matrix I gives the function for the case of equal
weights and that for n observations, the order of v is (n,1), the order of W is (n,n) and the
function ϕ = vT Wv is a scalar quantity (a single number).

In both examples of the Line of Best Fit an observation equation yk + vk = m xk + c was used

that if re-arranged as vk − m xk + c = − yk yields five equations for the coordinate pairs

v1 − mx1 − c = − y1
v2 − mx2 − c = − y2
v3 − mx3 − c = − y3
v4 − mx4 − c = − y4
v5 − mx5 − c = − y5

Note that these re-arranged observation equations have all the unknown quantities v, m and c
on the left of the equals sign and all the known quantities on the right.

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–22

RMIT University Geospatial Science

These equations can be written in matrix form

⎡ v1 ⎤ ⎡ − x1 −1⎤ ⎡ − y1 ⎤
⎢v ⎥ ⎢− x −1 ⎡ m ⎤ ⎢ − y 2 ⎥
⎥
⎢ 2⎥ + ⎢ 2 ⎥ =⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢⎣ c ⎥⎦ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ v5 ⎦ ⎣ − x5 −1⎦ ⎣ − y5 ⎦

and written symbolically as

v + Bx = f (2.37)

where f =d−l (2.38)

If n is the number of observations (equal to the number of equations) and u is the number of
unknown parameters
v is an (n,1) vector of residuals,
B is an (n,u) matrix of coefficients,
x is the (u,1) vector of unknown parameters,
f is the (n,1) vector of numeric terms derived from the observations,
d is an (n,1) vector of constants and
l is the (n,1) vector of observations.
Note that in many least squares problems the vector d is zero.

By substituting (2.37) into (2.36), we can obtain an expression for the least squares function

ϕ = vT Wv
= ( f − Bx ) W ( f − Bx )
T

(
= f T − ( Bx )
T
) W ( f − Bx )
= ( f T − xT BT ) W ( f − Bx )

and multiplication, observing the rule of matrix algebra gives

ϕ = f T Wf − f T WBx − xT BT Wf + x T BT WBx (2.39)

Since ϕ is a scalar (a number), the four terms on the right-hand-side of (2.39) are also scalars.
Furthermore, since the transpose of a scalar is equal to itself, the second and third terms are

equal ( f T WBx ) = xT BT Wf , remembering that W is symmetric hence W = WT , giving

T

ϕ = f T Wf − 2f T WBx + xT ( BT WB ) x (2.40)

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–23

RMIT University Geospatial Science

In equation (2.40) all matrices and vectors are numerical constants except x, the vector of
unknown parameters, therefore for the least squares function ϕ to be a minimum, its partial
derivative with respect to each element in vector x must be equated to zero, i.e., ϕ will be a
∂ϕ
minimum when = 0T . The first term of (2.40) does not contain x so its derivative is
∂x
automatically zero and the second and third terms are bilinear and quadratic forms
respectively and their derivatives are given in Appendix A, hence ϕ will be a minimum when

∂ϕ
= −2f T WB + 2xT ( BT WB ) = 0T
∂x

Cancelling the 2's, re-arranging and transposing gives a set of normal equations

( B WB ) x = B Wf
T T
(2.41)

Equation (2.41) is often given in the form

Nx = t (2.42)

where N = BT WB is a (u,u) coefficient matrix (the normal equation coefficient

matrix),
x is the (u,1) vector of unknown parameters and
t = BT Wf is a (u,1) vector of numeric terms.

The solution for the vector of parameters x is given by

x = N − 1t (2.43)

After solving for the vector x, the residuals are obtained from

v = f − Bx (2.44)

and the vector of "adjusted" or estimated observations l̂ is

ˆl = l + v (2.45)

The "hat" symbol (^) is used to denote quantities that result from a least squares process.
Such quantities are often called adjusted quantities or least squares estimates.

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–24

RMIT University Geospatial Science

These equations are the standard matrix solution for

least squares adjustment of indirect observations.

The name "least squares adjustment of indirect observations", adopted by Mikhail (1976) and
Mikhail & Gracie (1981), recognises the fact that each observation is an indirect measurement
of the unknown parameters. This is the most common technique employed in surveying and
geodesy and is described by various names, such as
parametric least squares
least squares adjustment by observation equations
least squares adjustment by residual equations

The technique of least squares adjustment of indirect observations has the following
characteristics

• A mathematical model (equation) links observations, residuals (corrections) and

unknown parameters.
• For n observations, there is a minimum number n0 required to determine the u

unknown parameters. In this case n0 = u and the number of redundant observations is

r = n − n0 .

• An equation can be written for each observation, i.e., there are n observation
equations. These equations can be represented in a standard matrix form; see equation
(2.37), representing n equations in u unknowns and solutions for the unknown
parameters, residuals and adjusted observations obtained from equations (2.41) to
(2.45).

The popularity of this technique of adjustment is due to its easy adaptability to computer-
programmed solutions. As an example, the following MATLAB program best_fit_line.m
reads a text file containing coordinate pairs (measurements) x and y and a weight w (a
measure of precision associated with each coordinate pair) and computes the parameters m
and c of a line of best fit y = mx + c .

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–25

RMIT University Geospatial Science

MATLAB program best_fit_line

function best_fit_line
%
% BEST_FIT_LINE reads an ASCII textfile containing coordinate pairs (x,y)
% and weights (w) associated with each pair and computes the parameters
% m and c of the line of best fit y = mx + c using the least squares
% principle. Results are written to a textfile having the same path and
% name as the data file but with the extension ".out"

%============================================================================
% Function: best_fit_line
%
% Author:
% Rod Deakin,
% Department of Geospatial Science, RMIT University,
% GPO Box 2476V, MELBOURNE VIC 3001
% AUSTRALIA
% email: [email protected]
%
% Date:
% Version 1.0 18 March 2003
%
% Remarks:
% This function reads numeric data from a textfile containing coordinate
% pairs (x,y) and weights (w) associated with each pair and computes the
% parameters m and c of a line of best fit y = mx + c using the least
% squares principle. Results are written to a textfile having the same
% path and name as the data file but with the extension ".out"
%
% Arrays:
% B - (n,u) coeff matrix of observation equation v + Bx = f
% f - (n,1) vector of numeric terms
% N - (u,u) coefficient matrix of Normal equations Nx = t
% Ninv - (u,u) inverse of N
% t - (u,1) vector of numeric terms of Normal equations Nx = t
% v - (n,1) vector of residuals
% W - (n,n) weight matrix
% weight - (n,1) vector of weights
% x - (u,1) vector of solutions
% x_coord - (n,1) vector of x coordinates
% y_coord - (n,1) vector of y coordinates
%
%
% Variables
% n - number of equations
% u - number of unknowns
%
% References:
% Notes on Least Squares (2003), Department of Geospatial Science, RMIT
% University, 2003
%
%============================================================================

%-------------------------------------------------------------------------
% 1. Call the User Interface (UI) to choose the input data file name
% 2. Concatenate strings to give the path and file name of the input file
% 3. Strip off the extension from the file name to give the rootName
% 4. Add extension ".out" to rootName to give the output filename
% 5. Concatenate strings to give the path and file name of the output file
%-------------------------------------------------------------------------
filepath = strcat('c:\temp\','*.dat');
[infilename,inpathname] = uigetfile(filepath);
infilepath = strcat(inpathname,infilename);
rootName = strtok(infilename,'.');

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–26

RMIT University Geospatial Science

MATLAB program best_fit_line

outfilename = strcat(rootName,'.out');
outfilepath = strcat(inpathname,outfilename);

%----------------------------------------------------------
% 1. Load the data into an array whose name is the rootName
% 2. set fileTemp = rootName
% 3. Copy columns of data into individual arrays
%----------------------------------------------------------
load(infilepath);
fileTemp = eval(rootName);
x_coord = fileTemp(:,1);
y_coord = fileTemp(:,2);
weight = fileTemp(:,3);

% determine the number of equations

n = length(x_coord);

% set the number of unknowns

u = 2;

% set the elements of the weight matrix W

W = zeros(n,n);
for k = 1:n
W(k,k) = weight(k);
end

% form the coefficient matrix B of the observation equations

B = zeros(n,u);
for k = 1:n
B(k,1) = -x_coord(k);
B(k,2) = -1;
end

% for the vector of numeric terms f

f = zeros(n,1);
for k = 1:n
f(k,1) = -y_coord(k);
end

% form the normal equation coefficient matrix N

N = B'*W*B;

% form the vector of numeric terms t

t = B'*W*f;

% solve the system Nx = t for the unknown parameters x

Ninv = inv(N);
x = Ninv*t;

% compute residuals
v = f - (B*x);

% open the output file print the data

fidout = fopen(outfilepath,'wt');

fprintf(fidout,'\n\nLine of Best Fit Least Squares Solution');

fprintf(fidout,'\n\nInput Data');
fprintf(fidout,'\n x(k) y(k) weight w(k)');
for k = 1:n
fprintf(fidout,'\n%10.4f %10.4f %10.4f',x_coord(k),y_coord(k),weight(k));
end

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–27

RMIT University Geospatial Science

MATLAB program best_fit_line

fprintf(fidout,'\n\nCoefficient matrix B of observation equations v + Bx = f');

for j = 1:n
fprintf(fidout,'\n');
for k = 1:u
fprintf(fidout,'%10.4f',B(j,k));
end
end

fprintf(fidout,'\n\nVector of numeric terms f of observation equations v + Bx =

f');
for k = 1:n
fprintf(fidout,'\n%10.4f',f(k,1));
end

fprintf(fidout,'\n\nCoefficient matrix N of Normal equations Nx = t');

for j = 1:u
fprintf(fidout,'\n');
for k = 1:u
fprintf(fidout,'%12.4f',N(j,k));
end
end

fprintf(fidout,'\n\nVector of numeric terms t of Normal equations Nx = t');

for k = 1:u
fprintf(fidout,'\n%10.4f',t(k,1));
end

fprintf(fidout,'\n\nInverse of Normal equation coefficient matrix');

for j = 1:u
fprintf(fidout,'\n');
for k = 1:u
fprintf(fidout,'%16.4e',Ninv(j,k));
end
end

fprintf(fidout,'\n\nVector of solutions x');

for k = 1:u
fprintf(fidout,'\n%10.4f',x(k,1));
end

fprintf(fidout,'\n\nVector of residuals v');

for k = 1:n
fprintf(fidout,'\n%10.4f',v(k,1));
end

fprintf(fidout,'\n\n');

% close the output file

fclose(fidout);

MATLAB program best_fit_line

Data file c:\Temp\line_data.dat

% data file for function "best_fit_line.m"

% x y w
-40.0 -24.0 2
-15.0 -24.0 5
10.0 -12.0 7
38.0 15.0 3
67.0 30.0 3

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–28

RMIT University Geospatial Science

MATLAB program best_fit_line

Running the program from the MATLAB command window prompt >> opens up a standard
Microsoft Windows file selection window in the directory c:\Temp. Select the appropriate
data file (in this example: line_data.dat) by double clicking with the mouse and the program
reads the data file, computes the solutions and writes the output data to the file
c:\Temp\line_data.out

MATLAB command window

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–29

RMIT University Geospatial Science

MATLAB program best_fit_line

Output file c:\Temp\line_data.out

Line of Best Fit Least Squares Solution

Input Data
x(k) y(k) weight w(k)
-40.0000 -24.0000 2.0000
-15.0000 -24.0000 5.0000
10.0000 -12.0000 7.0000
38.0000 15.0000 3.0000
67.0000 30.0000 3.0000

Coefficient matrix B of observation equations v + Bx = f

40.0000 -1.0000
15.0000 -1.0000
-10.0000 -1.0000
-38.0000 -1.0000
-67.0000 -1.0000

Vector of numeric terms f of observation equations v + Bx = f

24.0000
24.0000
12.0000
-15.0000
-30.0000

Coefficient matrix N of Normal equations Nx = t

22824.0000 230.0000
230.0000 20.0000

Vector of numeric terms t of Normal equations Nx = t

10620.0000
-117.0000

Inverse of Normal equation coefficient matrix

4.9556e-005 -5.6990e-004
-5.6990e-004 5.6554e-002

Vector of solutions x
0.5930
-12.6691

Vector of residuals v
-12.3878
2.4363
5.2605
-5.1363
-2.9403

The data in this example is taken from section 2.4.2 Line of Best Fit (unequal weights)

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–30

RMIT University Geospatial Science

MATLAB program best_fit_line

By adding the following lines to the program, the Line of Best Fit is shown on a plot together
with the data points.
%--------------------------------------
% plot data points and line of best fit
%--------------------------------------

% copy solutions from vector x

m = x(1,1);
c = x(2,1);

% find minimum and maximum x coordinates

xmin = min(x_coord);
xmax = max(x_coord);

% create a vector of x coordinates at intervals of 0.1

% between min and max coordinates
x = xmin:0.1:xmax;

% calculate y coordinates of Line of Best Fit

y = m*x + c;

% Select Figure window and clear figure

figure(1);
clf(1);
hold on;
grid on;
box on;

% plot line of best fit and then the data points with a star (*)
plot(x,y,'k-');
plot(x_coord,y_coord,'k*');

% anotate the plot

title('Least Squares Line of Best Fit')
xlabel('X coordinate');
ylabel('Y coordinate');

Least Squares Line of Best Fit

30

20

10
Y coordinate

-10

-20

-30

-40
-40 -20 0 20 40 60 80
X coordinate

Figure 2.3 Least Squares Line of Best Fit

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–31

RMIT University Geospatial Science

2.7. Least Squares Curve Fitting

The general matrix solutions for least squares adjustment of indirect observations (see
equations (2.37) to (2.45) of section 2.6) can be applied to curve fitting. The following two
examples (parabola and ellipse) demonstrate the technique.

2.7.1. Least Squares Best Fit Parabola

Consider the following: A surveyor working on the re-alignment of a rural road is required to
fit a parabolic vertical curve such that it is a best fit to the series of natural surface Reduced
Levels (RL's) on the proposed new alignment. Figure 2.2 shows a Vertical Section of the
proposed alignment with Chainages (x-values) and RL's (y-values).

Natura
l Surface

Datum RL 50.00
46.20

36.62

38.96

47.42

57.72
63.48

Red. Level
250

300

350
150

200

Chainage
100

Figure 2.2 Vertical Section of proposed road alignment

The general equation of a parabolic curve is

y = ax 2 + bx + c (2.46)

This is the mathematical model that we assume our data accords with and to account for the
measurement inconsistencies, due to the irregular natural surface and small measurement
errors we can add residuals to the left-hand-side of (2.46) to give an observation equation

yk + vk = axk2 + bxk + c (2.47)

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–32

RMIT University Geospatial Science

Equation (2.47) can be re-arranged as

vk − axk2 − bx − c = − yk (2.48)

n = 6 equations in u = 3 unknown parameters a, b, c can be written in matrix form v + Bx = f

as

⎡ v1 ⎤ ⎡ − x12 − x1 −1⎤ ⎡ − y1 ⎤
⎢v ⎥ ⎢− x 2 ⎥ ⎡a ⎤ ⎢ − y ⎥
⎢ 2⎥ + ⎢ 2 − x2 −1⎥ ⎢ ⎥ ⎢ 2 ⎥
b =
⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ 2 ⎥ ⎢⎣ c ⎥⎦ ⎢ ⎥
⎣ v6 ⎦ ⎢⎣ − x6 − x6 −1⎥⎦ ⎣ − y6 ⎦

where

⎡ v1 ⎤ ⎡ − x12 − x1 −1⎤ ⎡ − y1 ⎤
⎢v ⎥ ⎢ 2 ⎥ ⎡a ⎤ ⎢− y ⎥
− x2 − x2 −1⎥
v (6,1) = ⎢ 2⎥, B ( 6,3) = ⎢ , x (3,1) = b⎥ ,
⎢ f(6,1) = ⎢ 2⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ 2 ⎥ ⎢⎣ c ⎦⎥ ⎢ ⎥
⎣ v6 ⎦ ⎣⎢ − x6 − x6 −1⎥⎦ ⎣ − y6 ⎦

Considering all the measurements to be of equal precision, i.e., W = I, the least squares
solution for the three parameters in the vector x is given by the following sequence of
operations

• form the normal coefficient matrix: N = BT WB

• form the vector of numeric terms: t = BT Wf

• compute the matrix inverse: N −1

• compute the solutions: x = N − 1t

• compute the residuals: v = f − Bx

This is the identical series of operations to solve for the parameters of the Line of Best Fit,
except in this case u = 3. With minor modifications to the MATLAB program best_fit_line.m
another MATLAB program best_fit_parabola.m can be created to determine the parameters a,
b, c of the best fit parabola. The relevant modifications are shown below.

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–33

RMIT University Geospatial Science

MATLAB program best_fit_parabola

Making the following changes to the MATLAB program best_fit_line, a new program
best_fit_parabola can be created.

Changes to function name and help instructions

function best_fit_parabola
%
% BEST_FIT_PARABOLA reads an ASCII textfile containing coordinate pairs (x,y)
% and weights (w) associated with each pair and computes the parameters
% a, b and c of a best fit parabola y = a(x*x) + bx + c using the least
% squares principle. Results are written to a textfile having the same
% path and name as the data file but with the extension ".out"

Changes to function remarks in documentation section

% Remarks:
% This function reads numeric data from a textfile containing coordinate
% pairs (x,y) and weights (w) associated with each pair and computes the
% parameters a, b, and c of a best fit parabola y = a(x*x) + bx + c using
% the least squares principle. Results are written to a textfile having
% the same path and name as the data file but with the extension ".out"

Changes to formation of coefficient matrix B

% form the coefficient matrix B of the observation equations

B = zeros(n,u);
for k = 1:n
B(k,1) = -(x_coord(k)^2);
B(k,2) = -x_coord(k);
B(k,3) = -1;
end

Changes to data plotting section

%------------------------------------------
% plot data points and Parabola of best fit
%------------------------------------------

% copy solutions from vector x

a = x(1,1);
b = x(2,1);
c = x(3,1);

% find minimum and maximum x coordinates

xmin = min(x_coord);
xmax = max(x_coord);

% create a vector of x coordinates at intervals of 0.1

% between min and max coordinates
x = xmin:0.1:xmax;

% calculate y coordinates of Parabola of Best Fit

y = a*(x.*x) + b*x + c;

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–34

RMIT University Geospatial Science

MATLAB program best_fit_parabola

Using the data from Figure 2.2 a data file c:\Temp\parabola_data.dat was created

% data file for function "best_fit_parabola.m"

% x y w
100.0 63.48 1
150.0 46.20 1
200.0 36.62 1
250.0 38.96 1
300.0 47.42 1
350.0 57.72 1

Running the program from the MATLAB command window generated the following output
file c:\Temp\parabola_data.out and a plot of the Least Squares Parabola of best Fit

Parabola of Best Fit Least Squares Solution

Input Data
x(k) y(k) weight w(k)
100.0000 63.4800 1.0000
150.0000 46.2000 1.0000
200.0000 36.6200 1.0000
250.0000 38.9600 1.0000
300.0000 47.4200 1.0000
350.0000 57.7200 1.0000

Coefficient matrix B of observation equations v + Bx = f

-10000.0000 -100.0000 -1.0000
-22500.0000 -150.0000 -1.0000
-40000.0000 -200.0000 -1.0000
-62500.0000 -250.0000 -1.0000
-90000.0000 -300.0000 -1.0000
-122500.0000 -350.0000 -1.0000

Vector of numeric terms f of observation equations v + Bx = f

-63.4800
-46.2000
-36.6200
-38.9600
-47.4200
-57.7200

Coefficient matrix N of Normal equations Nx = t

29218750000.0000 97875000.0000 347500.0000
97875000.0000 347500.0000 1350.0000
347500.0000 1350.0000 6.0000

Vector of numeric terms t of Normal equations Nx = t

16912600.0000
64770.0000
290.4000

Inverse of Normal equation coefficient matrix

4.2857e-009 -1.9286e-006 1.8571e-004
-1.9286e-006 8.9071e-004 -8.8714e-002
1.8571e-004 -8.8714e-002 9.3714e+000

Vector of solutions x
0.001500
-0.688221
116.350000

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–35

RMIT University Geospatial Science

MATLAB program best_fit_parabola

Vector of residuals v
-0.948
0.676
2.103
-0.889
-2.498
1.555

Least Squ ares P arabo la of B est Fit

65

60

55
Y coordinate

50

45

40

35
1 00 150 2 00 250 300 3 50
X c oord inate

Figure 2.4 Least Squares Parabola of Best Fit

2.7.2. Least Squares Best Fit Ellipse

In November 1994, a survey was undertaken by staff of the Department of Geospatial Science
at the Melbourne Cricket Ground (MCG) to determine the dimensions of the playing surface.
This survey was to decide which of two sets of dimensions was correct, those of the
Melbourne Cricket Club (MCC) or those of the Australian Football League (AFL). The MCC
curator Tony Ware and the AFL statistician Col Hutchison both measured the length of the
ground (Tony Ware with a 100-metre nylon tape and Col Hutchison with a measuring wheel)
and compared their distances with the "true" distance determined by Electronic Distance
Measurement (EDM) with a Topcon 3B Total Station. Their measurements were both
reasonably close to the EDM distance and it turned out that the "official" AFL dimensions

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–36

RMIT University Geospatial Science

were incorrect. After this "measure-off", observations (bearings and distances) were made to
seventeen points around the edge of the playing surface to determine the Least Squares
Ellipse of Best Fit and to see if the major axis of this ellipse was the actual line between the
goals at either end. The Total Station was set-up close to the goal-to-goal axis and 20-25
metres from the centre of the ground. An arbitrary X,Y coordinate system was used with the
origin at the Total Station and the positive X-axis in the direction of the Brunton Avenue end
of the Great Southern Stand (approximately west). The table of coordinates is given below;
point numbers 1 to 6 were not points on the edge of the ground.

Point No. X-coordinate Y-coordinate

7 -54.58 17.11
8 -45.47 36.56
9 -28.40 53.22
10 -2.02 63.72
11 28.12 63.44
12 57.49 52.55
13 80.85 34.20
14 98.08 9.14
15 105.69 -17.30
16 103.83 -46.96
17 88.42 -71.50
18 61.26 -86.84
19 26.47 -91.07
20 -6.59 -81.37
21 -34.55 -59.24
22 -51.51 -29.28
23 -56.30 -2.31

Table 2.4 Arbitrary coordinates of points around the

perimeter of the playing surface of the MCG
(date of survey November 1994)

To develop an observation equation for the Least Squares Ellipse of Best Fit and to determine
the lengths and directions of the axes of the ellipse the following derivation of the general
equation of an ellipse is necessary.

Figure 2.5 shows an ellipse whose axes are aligned with the u-v axes. The semi-axes lengths
are a and b ( a > b) , the centre of the ellipse is at X 0 , Y0 and the ellipse axes are rotated by an

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–37

RMIT University Geospatial Science

angle β , measured positive anti-clockwise from the x-axis. The x-y axes are parallel to the X-
Y axes and pass through the centre of the ellipse.

Y
y
v
u
θ

r
β
Y0
a x

b
X0 X

Figure 2.5

The u,v Cartesian equation of the ellipse is

u2 v2
+ =1 (2.49)
a 2 b2

Equation (2.49) can be expressed in matrix form as

⎡1 a 2 0 ⎤ ⎡u ⎤
[u v] ⎢ 2⎥⎢ ⎥
=1 (2.50)
⎣ 0 1 b ⎦ ⎣v ⎦

The u,v axes are rotated (positive anti-clockwise) by an angle β from the x,y axes and the
relationship between coordinates is shown in Figure 2.6

y sin β
y x sin β
v •
u
y cos β x cos β
β
x

Figure 2.6

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–38

RMIT University Geospatial Science

Inspection of Figure 2.6 shows

u = x cos β + y sin β
(2.51)
v = − x sin β + y cos β

Replacing cos β and sin β with the letters c and s the coordinate relationships can be
represented as a matrix equation

⎡u ⎤ ⎡ c s⎤ ⎡ x ⎤
⎢ v ⎥ = ⎢ −s c ⎥ ⎢ y ⎥ (2.52)
⎣ ⎦ ⎣ ⎦⎣ ⎦

Transposing this equation (remembering the reversal rule with the transpose of matrix
products) gives

⎡ c −s ⎤
[u v] = [ x y] ⎢ ⎥ (2.53)
⎣ s c⎦

Substituting (2.52) and (2.53) into (2.50) and multiplying the matrices gives

⎡c −s ⎤ ⎡1 a 2 0 ⎤ ⎡ c s⎤ ⎡ x ⎤
[x y] ⎢ ⎥ ⎢ 2⎥⎢ ⎥⎢ ⎥ =1
⎣ s c ⎦ ⎣ 0 1 b ⎦ ⎣ −s c ⎦ ⎣ y ⎦

⎡ ⎛ c 2 s2 ⎞ ⎛ cs cs ⎞ ⎤
⎢⎜ 2 + 2 ⎟ ⎜ 2 − 2 ⎟ ⎥
⎝a b ⎠ ⎝a b ⎠⎥ ⎡ x ⎤
[x y] ⎢ =1
⎢ ⎛ cs cs ⎞ ⎛ s2 c 2 ⎞ ⎥ ⎢⎣ y ⎥⎦
⎢⎜ 2 − 2 ⎟ ⎜ 2 + 2 ⎟⎥
⎢⎣ ⎝ a b ⎠ ⎝a b ⎠ ⎥⎦

Replacing the elements of the square matrix with the symbols A, B and H, noting that the top-
right and lower-left elements are the same, this equation may be written in a general form as

⎡A H ⎤ ⎡ x⎤
[x y] ⎢ =1
⎣H B ⎥⎦ ⎢⎣ y ⎥⎦

or Ax 2 + 2 Hxy + By 2 = 1 (2.54)

Equation (2.54) is the equation of an ellipse centred at the coordinate origin but with axes
rotated from the x,y axes. The semi axes lengths a and b, and the rotation angle β can be
determined from (2.54) by the following method.

Letting x = r cos θ and y = r sin θ in equation (2.54) gives the polar equation of the ellipse

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–39

RMIT University Geospatial Science

1
A cos2 θ + 2 H cosθ sin θ + B sin 2 θ = (2.55)
r2

r is the radial distance from the centre of the ellipse and θ is the angle measured positive
anti-clockwise from the x-axis. Equation (2.55) has maximum and minimum values defining
the lengths and directions of the axes of the ellipse. To determine these values from (2.55),
consider the following

1
Let 2
= f = A cos2 θ + 2 H cosθ sin θ + B sin 2 θ
r
= A cos2 θ + H sin 2θ + B sin 2 θ

and aim to find the optimal (maximum and minimum) values of f and the values of θ when
these occur by investigating the first and second derivatives f ′ and f ′′ respectively, i.e.,

⎧ max ⎫ ⎧ f ′ = 0 and f ′′ < 0 ⎫

f is ⎨ ⎬ when ⎨ ⎬
⎩ min ⎭ ⎩ f ′ = 0 and f ′′ > 0 ⎭

f ′ = ( B − A) sin 2θ + 2 H cos 2θ
where (2.56)
f ′′ = 2 ( B − A) cos 2θ − 4 H sin 2θ

Now the maximum or minimum value of f occurs when f ′ = 0 and from the first member of
(2.56) the value of θ is given by

2H
tan 2θ = (2.57)
A− B

But this value of θ could relate to either a maximum or a minimum value of f. So from the
second member of equations (2.56) with a value of 2θ from equation (2.57) this ambiguity
can be resolved by determining the sign of the second derivative f ′′ giving

⎧ f max ⎫ ⎧ f ′′ < 0 ⎫
⎨ ⎬ when ⎨ ⎬
⎩ f min ⎭ ⎩ f ′′ > 0 ⎭

In the polar equation of the ellipse given by equation (2.55) f max coincides with rmin and f min

coincides with rmax so the angle β (measured positive anti-clockwise) from the x-axis to the
major axis of the ellipse (see Figure 2.5) is found from

⎧ rmax ⎫ ⎧ f ′′ > 0 ⎫ ⎧β = θ ⎫
⎨ ⎬ when ⎨ ⎬ and ⎨ ⎬ (2.58)
⎩ rmin ⎭ ⎩ f ′′ < 0 ⎭ ⎩β = θ − 2 π ⎭
1

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–40

RMIT University Geospatial Science

These results can be verified by considering the definitions of A, B and H used in the
derivation of the polar equation of the ellipse, i.e.,

cos2 β sin 2 β sin 2 β cos2 β cos β sin β cos β sin β

A= 2
+ 2
, B = 2
+ 2
, H= −
a b a b a2 b2

⎛ 1 1⎞ ⎛ 1 1⎞
and A − B = ⎜ 2 − 2 ⎟ cos 2 β , 2 H = ⎜ 2 − 2 ⎟ sin 2 β
⎝a b ⎠ ⎝a b ⎠

2H
giving tan 2 β =
A− B

Noting that the values of θ coinciding with the maximum or minimum values of the function
2H
f are found from equation (2.57) then tan 2 β = = tan 2θ or
A− B

tan 2θ = tan 2 β

whereupon

2θ = 2 β + nπ or θ = β + 12 nπ where n is an integer

Also, from the second member of equations (2.56)

f ′′ = 2 ( B − A) cos 2θ − 4 H sin 2θ

Now, for n = 0
⎛ 1 1⎞
θ = β , f ′′ θ = β = −2 ⎜ − 2 ⎟ and since a > b , f ′′ θ = β > 0
⎝a b ⎠
2

So θ = β makes f minimum and so r is maximum and

f min = A cos2 β + 2 H sin β cos β + B sin 2 β

( cos β + sin 2 β )
2 2
1
= 2
=
a a2

So rmax = a

When n = 1
θ = β + 12 π , sin 2θ = − sin 2 β , cos 2θ = − cos 2 β and so

⎛ 1 1⎞
f ′′ θ = β + 1π = 2 ⎜ 2 − 2 ⎟ and since a > b , f ′′ θ = β + 1π < 0
2
⎝a b ⎠ 2

So θ = β + 12 π makes f maximum and so r is minimum and

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–41

RMIT University Geospatial Science

(sin β + cos2 β )
2 2
1
f max = 2
=
b b2

So rmin = b

When n = 2
θ = β + π , sin 2θ = cos 2 β , cos 2θ = cos 2 β and f ′′ θ = β +π > 0

1
So θ = β + 12 π makes f min = and rmax = a
a2
When n = 3
θ = β + 23 π , sin 2θ = − cos 2 β , cos 2θ = − cos 2 β and f ′′ θ = β + π < 0 3
2

1
So θ = β + 23 π makes f max = and rmin = b
b2
All other even values of n give the same result as n = 2 and all other odd values of n give the
same result as n = 1

Now consider Figure 2.5 and the general Cartesian equation of the ellipse, re-stated again as

aX 2 + 2hXY + bY 2 + dX + eY = 1 (2.59)

where the translated x,y coordinate system is related to the X,Y system by

X = x + X0 and Y = y + Y0

Substituting these relationships into (2.59) gives

a ( x + X 0 ) + 2h ( x + X 0 )( y + Y0 ) + b ( y + Y0 ) + d ( x + X 0 ) + e ( y + Y0 ) = 1
2 2

Expanding and gathering terms gives

ax 2 + 2hxy + by 2 + ( 2aX 0 + 2hY0 + d ) x

+ ( 2aY0 + 2hX 0 + e ) y
+ aX 02 + 2hX 0Y0 + bY02 + dX 0 + eY0 = 1

Inspection of the left-hand-side of this equation reveals three parts:

(i) ax 2 + 2hxy + by 2 is the left-hand-side of the equation of an ellipse, similar in form
to equation (2.54) ,
(ii) coefficient terms of x and y; ( 2aX 0 + 2hY0 + d ) and ( 2aY0 + 2hX 0 + e ) ,

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–42

RMIT University Geospatial Science

(iii) a constant term aX 02 + 2hX 0Y0 + bY02 + dX 0 + eY0

Now when the coefficients of x and y are zero the ellipse will be centred at the origin of the
x,y axes with an equation of the form

ax 2 + 2hxy + by 2 = c (2.60)

where c = 1 − ( aX 02 + 2hX 0Y0 + bY02 + dX 0 + eY0 ) (2.61)

2aX 0 + 2hY0 + d = 0
and (2.62)
2hX 0 + 2bY0 + e = 0

Equations (2.62) can be written in matrix form and solved (using the inverse of a 2,2 matrix)
to give X 0 and Y0

⎡ d ⎤ ⎡ −2 a − 2 h ⎤ ⎡ X 0 ⎤
⎢ e ⎥ = ⎢ −2 h −2b ⎥ ⎢ Y ⎥
⎣ ⎦ ⎣ ⎦⎣ 0 ⎦
⎡X0⎤ 1 ⎡ −b h ⎤ ⎡ d ⎤
⎢ Y ⎥ = 2ab − 2h 2 ⎢ h − a ⎥ ⎢ e ⎥
⎣ 0⎦ ⎣ ⎦⎣ ⎦

eh − bd dh − ae
giving X0 = and Y0 = (2.63)
2 ( ab − h 2 ) 2 ( ab − h 2 )

Dividing both sides of (2.60) by c gives

Ax 2 + 2 Hxy + By 2 = 1 (2.64)

a h b
where A= , H= , B=
c c c

Equation (2.64), identical to equation (2.54), is the equation of an ellipse centred at the x,y
coordinate origin whose axes are rotated from the x,y axes by an angle β . The rotation angle
β and semi-axes lengths a and b of the ellipse can be determined using the method set out
above and equations (2.58), (2.57), (2.56) and (2.55). Thus, we can see from the development
that the general Cartesian equation of an ellipse is given by

aX 2 + 2hXY + bY 2 + dX + eY = 1 (2.65)

Note that the coefficients a and b in this equation are not the semi-axes lengths of the ellipse.

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–43

RMIT University Geospatial Science

Returning to the problem of the Least Squares Ellipse of Best Fit for the MCG, the size, shape
location and orientation of this ellipse can be determined from a set of observation equations
of the form

vk + aX k2 + 2hX kYk + bYk2 + dX k + eYk = 1 (2.66)

This observation equation is the general Cartesian equation of an ellipse with the addition of
the residual vk . The addition of vk to the left-hand-side of (2.65) is simply a convenience and

reflects the fact that the measured coordinates X k , Yk are inconsistent with the mathematical
model. For each of the 17 measured points around the perimeter of the MCG an equation can
be written and arranged in the matrix form v + Bx = f

⎡a⎤
⎡ v1 ⎤ ⎡ X 12 X 1Y1 Y12 X1 Y1 ⎤ ⎢ ⎥ ⎡1⎤
⎢v ⎥ ⎢X2 X 2Y2 Y22 X2
⎥ 2h
Y2 ⎥ ⎢ ⎥ ⎢⎢1⎥⎥
⎢ 2⎥+⎢ 2 ⎢b⎥=
⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ 2 ⎥ d ⎢ ⎥
⎣ v17 ⎦ ⎢⎣ X 17 X 17Y17 Y172 X 17 Y17 ⎥⎦ ⎢ ⎥ ⎣1⎦
⎢⎣ e ⎥⎦

The vector x contains the parameters a, h, b, d and e of the general equation of the ellipse and
with a weight matrix W = I (i.e., all observations of equal precision) the solution for x is
given by the following sequence of operations

• form the normal coefficient matrix: N = BT WB

• form the vector of numeric terms: t = BT Wf

• compute the matrix inverse: N −1

• compute the solutions: x = N − 1t

• compute the residuals: v = f − Bx

This is the identical series of operations to solve for the parameters of the Line of Best Fit,
and for the Parabola of Best Fit except in this case the residuals v have little practical meaning
because they are not connected to quantities such as distances or coordinates. In the case of
the Least Squares Ellipse of Best Fit, it is better to compute the offsets h (perpendicular
distances) from the ellipse to the data points rather than the residuals v.

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–44

RMIT University Geospatial Science

To compute offsets h the following preliminary sequence of operations is required:

(i) compute the parameters a, h, b, d and e using the Least Squares process set out
above.
(ii) compute the coordinates of the origin X 0 , Y0 and the constant c using equations
(2.63) and (2.61).
(iii) compute coefficients A, H and B of the ellipse given by (2.64) which can then be
used to compute the rotation angle β and the semi-axes lengths a and b from
equations (2.55) to (2.58) and.
(iv) compute the u,v coordinates of the data points using equations (2.51).

Now, having the u,v coordinates, the offsets h can be computed. Consider the sectional view
of a quadrant of an ellipse in Figure 2.7. The u,v axes are in the direction of the major and
minor axes respectively (a and b are the semi-axes lengths) and P is a point related to the
ellipse by the normal, which makes an angle φ with the major axis, and the distance h = QP
along the normal. The u,v coordinates of P are the distances LP and MP respectively. From
the geometry of an ellipse, the normal intersects the minor axis at H and the distance QH = ν

(where ν is the Greek symbol nu) and the distances DH and OH are ν e 2 and ν e 2 sin φ
respectively. e is the eccentricity of the ellipse and the eccentricity and flattening f of an
ellipse are related to the semi-axes a and b.

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–45

RMIT University Geospatial Science

tan
ge
nt
L • P

h
Q

b ν

)
e2
1−
al
rm
ν(
no
O D φ M u
a
νe 2

Figure 2.7 u,v coordinates of P related to an ellipse (a,b)

The equations for f, e 2 and ν are

a−b
f =
a
e = f (2 − f )
2

a
ν=
1 − e 2 sin 2 φ

Using these relationships, the angle φ and perpendicular distance h are given by

v + ν e 2 sin φ
tan φ = (2.67)
u

u
h= −ν (2.68)
cos φ

Inspecting these equations; if the semi-axes a,b and the u,v coordinates of P are known, the
perpendicular offset h can be determined from (2.68) and (2.67). It should be noted that
functions of φ appear on both sides of the equals sign of equation (2.67) and φ must be
solved by iteration.

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–46

RMIT University Geospatial Science

To determine the Least Square Best Fit Ellipse for the playing surface of the MCG a
MATLAB program best_fit_ellipse.m operating in the same way as the MATLAB programs
best_fit_line and best_fit_parabola with a data file (in this example: MCG_ellipse_data.dat)

Data file c:\Temp\MCG_ellipse_data.dat

% Data file for MCG Survey, November 1994

% Coordinates of 17 boundary points (edge of concrete)
%
% point X Y weight
7 -54.58 17.11 1
8 -45.47 36.56 1
9 -28.40 53.22 1
10 -2.02 63.72 1
11 28.12 63.44 1
12 57.49 52.55 1
13 80.85 34.20 1
14 98.08 9.14 1
15 105.69 -17.30 1
16 103.83 -46.96 1
17 88.42 -71.50 1
18 61.26 -86.84 1
19 26.47 -91.07 1
20 -6.59 -81.37 1
21 -34.55 -59.24 1
22 -51.51 -29.28 1
23 -56.30 -2.31 1

gives the following results (contained in an output file having the same name and path as the
data file but with the extension .dat) and plot on the screen.

Output file c:\Temp\MCG_ellipse_data.out

Ellipse of Best Fit Least Squares Solution

Input Data
point x(k) y(k) weight w(k)
7 -54.5800 17.1100 1.0000
8 -45.4700 36.5600 1.0000
9 -28.4000 53.2200 1.0000
10 -2.0200 63.7200 1.0000
11 28.1200 63.4400 1.0000
12 57.4900 52.5500 1.0000
13 80.8500 34.2000 1.0000
14 98.0800 9.1400 1.0000
15 105.6900 -17.3000 1.0000
16 103.8300 -46.9600 1.0000
17 88.4200 -71.5000 1.0000
18 61.2600 -86.8400 1.0000
19 26.4700 -91.0700 1.0000
20 -6.5900 -81.3700 1.0000
21 -34.5500 -59.2400 1.0000
22 -51.5100 -29.2800 1.0000
23 -56.3000 -2.3100 1.0000

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–47

RMIT University Geospatial Science

General Equation of Ellipse with X,Y origin not at centre of ellipse

aXX + 2hXY + bYY + dX + eY = 1
a = 1.720717e-004
h = 2.690541e-005
b = 1.865607e-004
d = -7.743828e-005
e = 3.729881e-005

Equation of Ellipse with x,y origin at centre of ellipse

Axx + 2Hxy + Byy = 1
A = 1.535544e-004
H = 2.401002e-005
B = 1.664842e-004

Ellipse parameters
semi-major axis a = 86.017
semi-minor axis b = 73.544

Bearing of major axis

beta(degrees) = -37.465030
beta(DMS) = -37 27 54.11
Brg(degrees) = 127.465030
Brg(DMS) = 127 27 54.11

Coordinates of centre of ellipse

X(centre) = 24.620
Y(centre) = -13.547

Data and offsets to ellipse of best fit

pt offset X Y x y u v
7 0.129 -54.580 17.110 -79.200 30.657 -81.511 -23.842
8 0.159 -45.470 36.560 -70.090 50.107 -86.111 -2.863
9 0.162 -28.400 53.220 -53.020 66.767 -82.696 20.744
10 0.164 -2.020 63.720 -26.640 77.267 -68.145 45.124
11 0.060 28.120 63.440 3.500 76.987 -44.051 63.236
12 -0.090 57.490 52.550 32.870 66.097 -14.116 72.457
13 -0.223 80.850 34.200 56.230 47.747 15.588 72.101
14 -0.216 98.080 9.140 73.460 22.687 44.507 62.691
15 -0.123 105.690 -17.300 81.070 -3.753 66.630 46.334
16 0.936 103.830 -46.960 79.210 -33.413 83.195 21.661
17 0.284 88.420 -71.500 63.800 -57.953 85.891 -7.191
18 -1.181 61.260 -86.840 36.640 -73.293 73.664 -35.887
19 -0.224 26.470 -91.070 1.850 -77.523 48.624 -60.407
20 0.627 -6.590 -81.370 -31.210 -67.823 16.483 -72.817
21 0.554 -34.550 -59.240 -59.170 -45.693 -19.171 -72.259
22 -0.438 -51.510 -29.280 -76.130 -15.733 -50.856 -58.796
23 -0.703 -56.300 -2.310 -80.920 11.237 -71.063 -40.303

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–48

RMIT University Geospatial Science

Least Squares Ellipse of Best Fit

60
10 11
Punt Rd
goals 9 12
40
8 13
20
7
14
Y coordinate

0
23 MCG
-20 15
22
-40 16
21
-60
17
20 Brunton Ave
-80 18 goals
19

-60 -40 -20 0 20 40 60 80 100 120

X coordinate

Figure 2.8 Plot of Least Squares Best Fit Ellipse and data points for the MCG

The MATLAB program best_fit_ellipse.m calls two other MATLAB functions ellipse.m (a
function to compute the coordinates of points on an ellipse) and DMS.m (a function to convert
decimal degree to degrees, minutes and seconds). A copy of these programs is shown below.

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–49

RMIT University Geospatial Science

MATLAB program best_fit_ellipse

function best_fit_ellipse
%
% BEST_FIT_ELLIPSE reads an ASCII textfile containing point numbers of
% coordinate pairs (X,Y) and weights (W) associated with each pair and
% computes the cordinates of the origin, the lengths of the axes and
% the rotation angle of the Best Fit Ellipse using the least squares
% principle. Results are written to a textfile having the same path
% and name as the data file but with the extension ".out"

%============================================================================
% Function: best_fit_ellipse
%
% Author:
% Rod Deakin,
% School of Mathematical and Geospatial Sciences, RMIT University,
% GPO Box 2476V, MELBOURNE VIC 3001
% AUSTRALIA
% email: [email protected]
%
% Date:
% Version 1.0 22 March 2003
% Version 1.1 10 May 2003
% Version 1.2 9 November 2005
%
% Functions Required:
% [X,Y] = ellipse(a,b,theta)
% [D,M,S] = DMS(DecDeg)
%
% Remarks:
% The general equation of an ellipse is
% aXX + 2hXY + bYY + dX + eY = 1
% This function computes the parameters a,h,b,d,e of a Least Squares Best Fit
% Ellipse given a set of X,Y coordinate pairs and weights (w) associated with
% each pair. The centre of the best fit ellipse is at Xo = (eh-db)/(2ab-2hh)
% and Yo = (dh-ea)/(2ab-2hh).
% A constant c = 1 - (aXoXo + 2hXoYo + bYoYo + dXo + eYo) is divided into a,
% h and b giving A = a/c, H = h/c and B = b/c which are the parameters of an
% ellipse Axx +2Hxy + Byy = 1. The major axis of this ellipse is rotated
% from the coordinate axes by an angle beta which can be determined from the
% polar equation of an ellipse
% A*cos_squared(theta) + 2H*cos(theta)*sin(theta) + B*sin_squared(theta) =
1/r_squared
% The maximum and minimum values of this function occur for theta given by
% tan(2*theta) = 2H/(A-B) and the angle beta is determined by evaluating
% the sign of the second derivative of the polar equation. This angle is
% substituted into the polar equation to determine the length of the semi-major
% axis length a. Beta - 90 degrees will give the length of the semi-minor
% axis length b
% Note that the semi-axes lengths a,b are not the same as the parameters
% a and b in the general equation of the ellipse.
% Results are written to a textfile having the same path and name as the
% data file but with the extension ".out"
%
% References:
% Notes on Least Squares (2005), Geospatial Science, RMIT
% University, 2005
%

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–50

RMIT University Geospatial Science

MATLAB program best_fit_ellipse

% Arrays:
% B - coeff matrix of observation equation v + Bx = f
% f - vector of numeric terms
% N - coefficient matrix of Normal equations Nx = t
% Ninv - inverse of N
% p - vector of perpendicular distances from ellipse to points
% point - vector of point numbers
% t - vector of numeric terms of Normal equations Nx = t
% u,v - vectors of u,v coords of ellipse
% W - weight matrix
% weight - vector of weights
% x - vector of solutions
% x,y - vectors of x,y coords of ellipse
% x_coord - vector of X coordinates
% y_coord - vector of Y coordinates
% xpt,ypt - vectors of coords for point number locations on plot
% Xpt,Ypt - vectors Xpt = xpt + Xc, Ypt = ypt + Yc
%
%
% Variables:
% A,B,H - parameters of ellipse Axx + 2Hxy + Byy = 1
% a,h,b, - parameters of ellipse aXX + 2hXY + bYY + dX + eY = 1
% d,e
% a1,b1 - semi-major and semi-minor axes of ellipse
% beta - angle between x-axis and major axis of ellipse (degrees)
% brg - bearing of major axis (u-axis) of ellipse (degrees)
% c - constant of translated ellipse or cos(x)
% d2r - degree to radian conversion factor = 180/pi = 57.29577951...
% e2 - eccentricity squared
% flat - flattening of ellipse
% f_dd - second derivative of the function "f" where f is the polar
% equation of an ellipse
% lat - latitude (radians) of point related to an ellipse
% n - number of equations
% new_lat - new latitude in iteration
% nu - radius of curvature in prime meridian
% pion2 - 90 degrees or pi/2
% u - number of unknowns
% s - sin(x)
% s1,s2 - sin(lat) and sin_squared(lat)
% scale - scale factor to reduce size of numbers in normal equations
% theta - angle for which polar equation of ellipse gives max/min
% values
% two_theta - 2*theta
% Xc,Yc - coords of centre of ellipse X = x + Xc, Y = y + Yc
% X0,Y0 - scaled coords of centre of ellipse
%
%============================================================================

%
% Set program constants
d2r = 180/pi;
pion2 = pi/2;
scale = 100;

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–51

RMIT University Geospatial Science

MATLAB program best_fit_ellipse

%-------------------------------------------------------------------------
% 1. Call the User Interface (UI) to choose the input data file name
% 2. Concatenate strings to give the path and file name of the input file
% 3. Strip off the extension from the file name to give the rootName
% 4. Add extension ".out" to rootName to give the output filename
% 5. Concatenate strings to give the path and file name of the output file
%-------------------------------------------------------------------------
filepath = strcat('c:\temp\','*.dat');
[infilename,inpathname] = uigetfile(filepath);
infilepath = strcat(inpathname,infilename);
rootName = strtok(infilename,'.');
outfilename = strcat(rootName,'.out');
outfilepath = strcat(inpathname,outfilename);

%----------------------------------------------------------
% 1. Load the data into an array whose name is the rootName
% 2. set fileTemp = rootName
% 3. Copy columns of data into individual arrays
%----------------------------------------------------------
load(infilepath);
fileTemp = eval(rootName);
point = fileTemp(:,1);
x_coord = fileTemp(:,2);
y_coord = fileTemp(:,3);
weight = fileTemp(:,4);

% Determine the number of equations and set the number of unknowns

n = length(point);
u = 5;

% Set the elements of the weight matrix W

W = zeros(n,n);
for k = 1:n
W(k,k) = weight(k);
end

% Form the coefficient matrix B of the observation equations.

% Note that the coordinates are scaled by a factor 1/100 to
% reduce the size of equations.
B = zeros(n,u);
for k = 1:n
B(k,1) = (x_coord(k)/scale)^2;
B(k,2) = (x_coord(k)/scale)*(y_coord(k)/scale);
B(k,3) = (y_coord(k)/scale)^2;
B(k,4) = x_coord(k)/scale;
B(k,5) = y_coord(k)/scale;
end

% Form the vector of numeric terms f

f = ones(n,1);

% Form the normal equation coefficient matrix N

% and the vector of numeric terms t
N = B'*W*B;
t = B'*W*f;

% Compute the inverse and solve the system Nx = t

Ninv = inv(N);
x = Ninv*t;

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–52

RMIT University Geospatial Science

MATLAB program best_fit_ellipse

% Copy the results into the variables a,h,b,d,e

a = x(1,1);
h = x(2,1)/2;
b = x(3,1);
d = x(4,1);
e = x(5,1);

% Compute the coordinates of the centre of the ellipse

X0 = (e*h - b*d)/(2*(a*b - h*h));
Y0 = (d*h - a*e)/(2*(a*b - h*h));
Xc = X0*scale;
Yc = Y0*scale;

% Compute the variables A,H,B and then the lengths

% of the axes and the rotation angle beta
c = 1-(a*X0*X0 + 2*h*X0*Y0 + b*Y0*Y0 + d*X0 + e*Y0);
A = a/c;
H = h/c;
B = b/c;

% compute the angle theta for max or min

two_theta = atan2(2*H,(A-B));
% compute second derivative
f_dd = 2*(B-A)*cos(two_theta) - 4*H*sin(two_theta);
% test the second derivative to determine max or min
theta = two_theta/2;
if f_dd < 0
beta = theta - pion2;
else
beta = theta;
end
% compute semi-major axis length
c = cos(beta);
s = sin(beta);
a1 = sqrt(1/(A*c*c + 2*H*c*s + B*s*s))*scale;
% compute semi-minor axis length
c = cos(beta+pion2);
s = sin(beta+pion2);
b1 = sqrt(1/(A*c*c + 2*H*c*s + B*s*s))*scale;
% convert beta to degrees
beta = beta*d2r;

% Calculate bearing of major axis noting that the rotation angle

% beta is considered positive anti-clockwise from the X-axis to
% the major axis of the ellipse
brg = 360+90-beta;
if(brg>360)
brg = brg-360;
end

%-----------------------------------------------------------------------
% Compute perpendicular distances from points to the ellipse of best fit
%-----------------------------------------------------------------------
% Create a set of u,v coordinates by first reducing the X,Y coords
% to x,y coordinates and then rotating these coordinates by the
% rotation angle beta. The u-axis is the major axis of the ellipse.
x = x_coord-Xc;
y = y_coord-Yc;
for k=1:n
u(k,1) = x(k)*cos(beta/d2r) + y(k)*sin(beta/d2r);
v(k,1) = -x(k)*sin(beta/d2r) + y(k)*cos(beta/d2r);
end

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–53

RMIT University Geospatial Science

MATLAB program best_fit_ellipse

% Compute the flattening and eccentricity squared for the ellipse

flat = (a1-b1)/a1;
e2 = flat*(2-flat);

% Compute the distance along the normal to the ellipse passing

% through the point
for k=1:n
% Compute the latitude of the normal to the ellipse through
% the point by iteration. nu is the radius of curvature of
% the prime vertical normal section
lat = pi/2;
new_lat = atan2(v(k,1),u(k,1));
while(abs(new_lat-lat)>1e-10)
lat = new_lat;
s1 = sin(lat);
s2 = s1*s1;
nu = a1/sqrt(1-e2*s2);
new_lat = atan2((v(k,1)+nu*e2*s1),u(k,1));
end
% p is the distance along the normal from the ellipse to the point
p(k,1) = (u(k,1)/cos(lat))-nu;
end

%----------------------------------------------------
% Compute the coordinate locations for a point number
% to be shown on the plot. These locations used in
% in the plot routines below.
%----------------------------------------------------
for k=1:n
theta = atan2(x(k),y(k));
if theta<0
theta = theta + 2*pi;
end
r = sqrt(x(k)^2 + y(k)^2)-10;
xpt(k) = r*sin(theta);
ypt(k) = r*cos(theta);
end
Xpt = xpt + Xc;
Ypt = ypt + Yc;

%-----------------------------
% print the data to the screen
%-----------------------------
fprintf('\n Ellipse of Best Fit\n');
fprintf('\n General Equation of Ellipse with X,Y origin not at centre of ellipse');
fprintf('\n aXX + 2hXY + bYY + dX + eY = 1');
fprintf('\n a = %14.6e',a/scale^2);
fprintf('\n h = %14.6e',h/scale^2);
fprintf('\n b = %14.6e',b/scale^2);
fprintf('\n d = %14.6e',d/scale^2);
fprintf('\n e = %14.6e\n',e/scale^2);
fprintf('\n Equation of Ellipse with x,y origin at centre of ellipse');
fprintf('\n Axx + 2Hxy + Byy = 1');
fprintf('\n A = %14.6e',A/scale^2);
fprintf('\n H = %14.6e',H/scale^2);
fprintf('\n B = %14.6e\n',B/scale^2);
fprintf('\n Ellipse parameters');
fprintf('\n semi-major axis a = %8.3f',a1);
fprintf('\n semi-minor axis b = %8.3f\n',b1);
fprintf('\n Bearing of major axis');
fprintf('\n beta(degrees) = %12.6f',beta);
[D,M,S] = DMS(beta);
fprintf('\n beta(DMS) = %4d %2d %5.2f',D,M,S);
fprintf('\n Brg(degrees) = %12.6f',brg);

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–54

RMIT University Geospatial Science

MATLAB program best_fit_ellipse

[D,M,S] = DMS(brg);
fprintf('\n Brg(DMS) = %4d %2d %5.2f\n',D,M,S);
fprintf('\n Coordinates of centre of ellipse');
fprintf('\n X(centre) = %12.3f',Xc);
fprintf('\n Y(centre) = %12.3f\n',Yc);
fprintf('\n Data and offsets to ellipse of best fit');
fprintf('\n pt offset X Y x y u
v');
for k=1:n
fprintf('\n %3d %10.3f %10.3f %10.3f %10.3f %10.3f %10.3f
%10.3f',point(k),p(k,1),x_coord(k),y_coord(k),x(k),y(k),u(k,1),v(k,1));
end

fprintf('\n\n');

%----------------------------------
% print the data to the output file
%----------------------------------

% Open the output file

fidout = fopen(outfilepath,'wt');

fprintf(fidout,'\n\nEllipse of Best Fit Least Squares Solution');

fprintf(fidout,'\n\nInput Data');
fprintf(fidout,'\n point x(k) y(k) weight w(k)');
for k = 1:n
fprintf(fidout,'\n%3d %12.4f %12.4f
%12.4f',point(k),x_coord(k),y_coord(k),weight(k));
end

fprintf(fidout,'\n\n General Equation of Ellipse with X,Y origin not at centre of

ellipse');
fprintf(fidout,'\n aXX + 2hXY + bYY + dX + eY = 1');
fprintf(fidout,'\n a = %14.6e',a/scale^2);
fprintf(fidout,'\n h = %14.6e',h/scale^2);
fprintf(fidout,'\n b = %14.6e',b/scale^2);
fprintf(fidout,'\n d = %14.6e',d/scale^2);
fprintf(fidout,'\n e = %14.6e\n',e/scale^2);
fprintf(fidout,'\n Equation of Ellipse with x,y origin at centre of ellipse');
fprintf(fidout,'\n Axx + 2Hxy + Byy = 1');
fprintf(fidout,'\n A = %14.6e',A/scale^2);
fprintf(fidout,'\n H = %14.6e',H/scale^2);
fprintf(fidout,'\n B = %14.6e\n',B/scale^2);
fprintf(fidout,'\n Ellipse parameters');
fprintf(fidout,'\n semi-major axis a = %8.3f',a1);
fprintf(fidout,'\n semi-minor axis b = %8.3f\n',b1);
fprintf(fidout,'\n Bearing of major axis');
fprintf(fidout,'\n beta(degrees) = %12.6f',beta);
[D,M,S] = DMS(beta);
fprintf(fidout,'\n beta(DMS) = %4d %2d %5.2f',D,M,S);
fprintf(fidout,'\n Brg(degrees) = %12.6f',brg);
[D,M,S] = DMS(brg);
fprintf(fidout,'\n Brg(DMS) = %4d %2d %5.2f\n',D,M,S);
fprintf(fidout,'\n Coordinates of centre of ellipse');
fprintf(fidout,'\n X(centre) = %12.3f',Xc);
fprintf(fidout,'\n Y(centre) = %12.3f\n',Yc);
fprintf(fidout,'\n Data and offsets to ellipse of best fit');
fprintf(fidout,'\n pt offset X Y x y
u v');
for k=1:n
fprintf(fidout,'\n %3d %10.3f %10.3f %10.3f %10.3f %10.3f %10.3f
%10.3f',point(k),p(k,1),x_coord(k),y_coord(k),x(k),y(k),u(k,1),v(k,1));
end

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–55

RMIT University Geospatial Science

MATLAB program best_fit_ellipse

fprintf(fidout,'\n\n');

% Close the output file

fclose(fidout);

%-------------------------------------------------------------------
% Call function 'ellipse' with parameters a,b,theta and receive back
% X,Y coordinates whose origin is at the centre of the ellipse
%-------------------------------------------------------------------
[X,Y] = ellipse(a1,b1,beta);
X = X + Xc;
Y = Y + Yc;

%-------------------------------------------------------------------
% Set the X,Y coordinates of the major and minor axes of the ellipse
%-------------------------------------------------------------------
aX = [X(180) X(360)];
aY = [Y(180) Y(360)];
bX = [X(90) X(270)];
bY = [Y(90) Y(270)];

%-------------------------------------------------
% plot the ellipse of Best Fit and the data points
%-------------------------------------------------
figure(1);
clf(1);
plot(X,Y,'r-',aX,aY,'b-',bX,bY,'b-');
hold on;
plot(x_coord,y_coord,'k.');
axis equal;
box off;

% plot the point numbers inside the ellipse

point_string=int2str(point);
text(Xpt,Ypt,point_string);

% anotate the plot

title('Least Squares Ellipse of Best Fit')
xlabel('X coordinate');
ylabel('Y coordinate');

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–56

RMIT University Geospatial Science

MATLAB program ellipse

function [X,Y] = ellipse(a,b,theta)

% ELLIPSE[X Y] = (A,B,THETA) Function to compute the X,Y coordinates of

% an ellipse given semi-axes A and B and a rotation angle THETA.
% The angle THETA is considered to be positive anti-clockwise from the x-axis

% set degree to radian conversion factor

d2r = pi/180;

% Calculate u,v coordinates of ellipse using parametric equations

% u = a*cos(psi)
% v = b*sin(psi)
% where the u-axis is the major axis, the v-axis is the minor axis
% and psi is the auxiliary angle measured positive anti-clockwise
% from the u-axis to a point moving around the auxiliary circle of
% radius a. The x,y coordinates are computed by rotating the ellipse
% axes by an angle theta, considered as positive anti-clockwise from the x-axis.
% x = u*cos(theta) - v*sin(theta)
% y = u*sin(theta) + v*cos(theta)

for k=1:360
u = a*cos(k*d2r);
v = b*sin(k*d2r);
X(k) = u*cos(theta*d2r) - v*sin(theta*d2r);
Y(k) = u*sin(theta*d2r) + v*cos(theta*d2r);
end
return

MATLAB program DMS

function [D,M,S] = DMS(DecDeg)

% [D,M,S] = DMS(DecDeg) This function takes an angle in decimal degrees
% and returns Degrees, Minutes and Seconds

val = abs(DecDeg);
D = fix(val);
M = fix((val-D)*60);
S = (val-D-M/60)*3600;
if(DecDeg<0)
D = -D;
end
return

© 2005, R.E. Deakin Notes on Least Squares (2005) 2–57

RMIT University Geospatial Science

3. PROPAGATION OF VARIANCES

In least squares problems, where measurements (with associated estimates of variances and
covariances) are used to determine the best estimates unknown quantities it is important to be
able to determine the precisions of these estimated (or calculated) quantities. To do this
requires an understanding of propagation of variances so that certain rules and techniques can
be developed.

3.1. Some Statistical Definitions

Students studying Least Squares must become familiar with statistical definitions,
terminology and rules. Some of these rules and definitions have been introduced in earlier
sections of these notes, e.g., in Chapter 2 the definition and classification of measurements
and measurement errors was discussed as well as the rules for computing means and variances
for finite and infinite populations. In addition, Chapter 2 contains sections explaining matrix
representations of variances and covariances, known as variance-covariance matrices Σ and
the related cofactor matrices Q and weight matrices W. The following sections in this chapter
repeat some of the rules and definitions already introduced as well as expanding on some
concepts previously mentioned.

3.1.1. Experiments, Sets, Sample Spaces, Events and Probability

The term statistical experiment can be used to describe any process by which several chance
observations are obtained. All possible outcomes of an experiment comprise a set called the
sample space and a set or sample space contains N elements or members. An event is a subset
of the sample space containing n elements. Experiments, sets, sample spaces and events are
the fundamental "tools" used to determine the probability of certain events where probability
is defined as

n
P ( Event ) = (3.1)
N

For example, if a card is drawn from a deck of playing cards, what is the probability that it is
a heart? In this case, the experiment is the drawing of the card and the possible outcomes of
the experiment could be one of 52 different cards, i.e., the sample space is the set of N = 52

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–1

RMIT University Geospatial Science

possible outcomes and the event is the subset containing n = 13 hearts. The probability of
drawing a heart is

n 13
P ( Heart ) = = = 0.25
N 52

This definition of probability is a simplification of a more general concept of probability that

can be explained in the following manner (see Johnson & Leone, 1964, pp.32-3).

Suppose observations are made on a series of occasions (often termed trials) and
during these trials it is noted whether or not a certain event occurs. The event can
be almost any observable phenomenon, for example, that the height of a person
walking through a doorway is greater than 1.8 metres, that a family leaving a
cinema contains three children, that a defective item is selected from an assembly
line, and so on. These trials could be conducted twice a week for a month, three
times a day for six months or every hour for every day for 10 years. In the
theoretical limit, the number of trials N would approach infinity and we could
assume, at this point, that we had noted every possible outcome. Therefore, as
N → ∞ then N becomes the number of elements in the sample space containing
all possible outcomes of the trials. Now for each trial we note whether or not a
certain event occurs, so that at the end of N trials we have noted nN events. The
probability of the event (if it in fact occurs) can then be defined as

⎛n ⎞
P ( Event ) = lim ⎜ N ⎟
N →∞
⎝N ⎠

Since nN and N are both non-negative numbers and nN is not greater than N then

nN
0≤ ≤1
N

Hence

0 ≤ P {Event} ≤ 1

If the event occurs at every trial then nN = N and nN N = 1 for all N and so

P ( Event ) = 1 . This relationship can be described as: the probability of a certain

(or sure) event is equal to 1.

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–2

RMIT University Geospatial Science

If the event never occurs, then nN = 0 and nN N = 0 for all N and so

P ( Event ) = 0 . This relationship can be described as: the probability of an

impossible event is zero.

The converse of these two relationships need not hold, i.e., a probability of one
need not imply certainty since it is possible that lim nN N = 1 without nN = 1 for
N →∞

all values of N and a probability of zero need not imply impossibility since it is
possible that lim nN N = 0 even though nN > 0 . Despite these qualifications, it
N →∞

is useful to think of probability as measured on a scale varying from (near)

impossibility at 0 to (near) certainty at 1. It should also be noted that this
definition of probability (or any other definition) is not directly verifiable in the
sense that we cannot actually carry out the infinite series of trials to see whether
there really is a unique limiting value for the ratio nN N . The justification for
this definition of probability is utilitarian, in that the results of applying theory
based on this definition prove to be useful and that it fits with intuitive ideas.
However, it should be realized that it is based on the concept of an infinitely long
series of trials rather than an actual series, however long it may be.

3.1.2. Random Variables and Probability Distributions of Random Variables

A random variable X is a rule or a function, which associates a real number with each point in
a sample space. As an example, consider the following experiment where two identical coins
are tossed; h denotes a head and t denotes a tail.

Experiment: Toss two identical coins.

Sample space: S = {hh, ht , th, tt} .

Random Variable: X, the number of heads obtained, may be written as

X ( hh ) = 2
X ( ht ) = 1
X ( th ) = 1
X ( tt ) = 0

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–3

RMIT University Geospatial Science

In this example X is the random variable defined by the rule "the number of heads obtained".
The possible values (or real numbers) that X may take are 0, 1, 2. These possible values are
usually denoted by x and the notation X = x denotes x as a possible real value of the random
variable X.

Random variables may be discrete or continuous. A discrete random variable assumes each
of its possible values with a certain probability. For example, in the experiment above; the
tossing of two coins, the sample space S = {hh, ht , th, tt} has N = 4 elements and the

probability the random variable X (the number of heads) assumes the possible values 0, 1 and
2 is given by

x 0 1 2
P ( X = x) 1
4
2
4
1
4

Note that the values of x exhaust all possible cases and hence the probabilities add to 1

A continuous random variable has a probability of zero of assuming any of its values and
consequently, its probability distribution cannot be given in tabular form. The concept of the
probability of a continuous random variable assuming a particular value equals zero may
seem strange, but the following example illustrates the point. Consider a random variable
whose values are the heights of all people over 21 years of age. Between any two values, say
1.75 metres and 1.85 metres, there are an infinite number of heights, one of which is 1.80
metres. The probability of selecting a person at random exactly 1.80 metres tall and not one
of the infinitely large set of heights so close to 1.80 metres that you cannot humanly measure
the difference is extremely remote, and thus we assign a probability of zero to the event. It
follows that probabilities of continuous random variables are defined by specifying an interval
within which the random variable lies and it does not matter whether an end-point is included
in the interval or not.

P ( a < X ≤ b) = P (a < X < b) + P ( X = b)

= P ( a < X < b)

It is most convenient to represent all the probabilities of a random variable X by a formula or

function denoted by f X ( x ) , g X ( x ) , hX ( x ) , etc, or by FX ( x ) , G X ( x ) , H X ( x ) , etc.

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–4

RMIT University Geospatial Science

In this notation the subscript X denotes that f X ( x ) or FX ( x ) is a function of the random

variable X which takes the numerical values x within the function. Such functions are known
as probability distribution functions and they are paired; i.e., f X ( x ) pairs with FX ( x ) ,

g X ( x ) pairs with G X ( x ) , etc. The functions with the lowercase letters are probability

density functions and those with uppercase letters are cumulative distribution functions.

For discrete random variables, the probability density function has the properties

1. f X ( xk ) = P ( X = xk )

∞
2. ∑ f (x ) =1
k =1
X k

and the cumulative distribution function has the properties

1. FX ( xk ) = P ( X ≤ xk )

2. FX ( x ) = ∑ f (x )
xk ≤ x
X k

As an example consider the probability distribution functions f X ( x ) and FX ( x ) of the sum

of the numbers when a pair of dice is tossed.

Experiment: Toss two identical dice.

⎧ 1, 1 1, 2 1, 3 1, 4 1, 5 1, 6 ⎫
⎪ 2, 1 ⎪
⎪ ⎪
2, 2 2, 3 2, 4 2, 5 2, 6

⎪ 3, 1 3, 2 3, 3 3, 4 3, 5 3, 6 ⎪
Sample space: S=⎨ ⎬
⎪ 4, 1 4, 2 4, 3 4, 4 4, 5 4, 6
⎪
⎪ 5, 1 5, 2 5, 3 5, 4 5, 5 5, 6 ⎪
⎪ ⎪
⎪⎩ 6, 1 6, 2 6, 3 6, 4 6, 5 6, 6 ⎪
⎭
Random Variable: X, the total of the two numbers

The probability the random variable X assumes the possible values x = 2, 3, 4, …, 12 is given
in Table 3.1

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–5

RMIT University Geospatial Science

x 2 3 4 5 6 7 8 9 10 11 12
1 2 3 4 5 6 5 4 3 2 1
P ( X = x)
36 36 36 36 36 36 36 36 36 36 36

Table 3.1 Table of probabilities

Note that the values of x exhaust all possible cases and hence the probabilities add to 1

The probability density function f X ( x ) can be deduced from Table 3.1

6− x−7
f X ( x) = , x = 2,3, 4,… ,12
36

Probability distributions are often shown in graphical form. For discrete random variables,
probability distributions are generally shown in the form of histograms consisting of series of
rectangles associated with values of the random variable. The width of each rectangle is one
unit and the height is the probability given by the function f X ( x ) and the sum of the areas of

all the rectangles is 1. Figure 3.1 shows the Probability histogram for the random variable X,
the sum of the numbers when a pair of dice is tossed.

Probability histogram
0.18

0.16

0.14

0.12
Probability f(x)

0.1

0.08

0.06

0.04

0.02

0
0 2 4 6 8 10 12 14
x

Figure 3.1 Probability histogram

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–6

RMIT University Geospatial Science

The MATLAB function dice_pdf.m was used to create the Probability histogram of Figure 3.1

function dice_pdf
% Function DICE_PDF calculates the probability of a random variable
% taking the sum of the values when two dice are tossed and
% plots the probability density function as a histogram.

% create an array of the possible outcomes

x = 2:12;

% calculate the probability

y = (6-abs(x-7))/36;

% Clear Figure 1 and plot the histogram

figure(1);
clf(1);
box on;
grid on;
bar(x,y,1,'-w');

% anotate the plot

title('Probability histogram')
xlabel('x');
ylabel('Probability f(x)');

Figure 3.2 shows the cumulative distribution function FX ( x ) = ∑ f (x )

xk ≤ x
X k for the random

variable X, the sum of the numbers when a pair of dice is tossed.

Cumulative Distribution Function

0.8

0.6
F(x)

0.4

0.2

0
0 2 4 6 8 10 12 14
x

Figure 3.2 Cumulative distribution function. [The dots at the left ends
of the line segments indicate the value of FX ( x ) at those values of x.

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–7

RMIT University Geospatial Science

For continuous random variables, the probability distribution functions f X ( x ) and FX ( x )

are curves, which may take various forms depending on the nature of the random variable.
Probability density functions f X ( x ) that are used in practice to model the behaviour of

continuous random variables are always positive and the total area under its curve, bounded
by the x-axis, is equal to one. These density functions have the following properties

1. f X ( x ) ≥ 0 for any value of x

+∞
2. ∫ f X ( x ) dx = 1
−∞

The probability that a random variable X lies between any two values x = a and x = b is the
area under the density curve between those two values and is found by methods of integral
calculus
b
P ( a < X < b ) = ∫ f X ( x ) dx (3.2)
a

The equations of the density functions f X ( x ) are usually complicated and areas under their

curves are found from tables. In surveying, the Normal probability density function is the
usual model for the behaviour of measurements (regarded as random variables) and the
probability density function is (Kreyszig, 1970, p. 107)

1 ⎛ x−μ ⎞
2

1 − ⎜
σ ⎟⎠
f X ( x) = e 2⎝ (3.3)
σ 2π

μ and σ are the mean and standard deviation respectively of the infinite population of x and
Figure 3.3 shows a plot of the Normal probability density curve for μ = 2.0 and σ = 2.5 .

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–8

RMIT University Geospatial Science

Normal Probability Density curve

0.16

0.14

0.12

0.1
f(x)

0.08

0.06

0.04

0.02

0
-10 -8 -6 -4 -2 0 2 4 6 8 10
x

Figure 3.3 Normal probability density function for μ = 2.0 and σ = 2.5

The MATLAB function normal_pdf.m was used to create the Normal probability density
curve of Figure 3.3

function normal_pdf(mx,sx)
% Function NORMAL_PDF(MX,SX) calculates the probability of a random variable X
% having a NORMAL distribution with mean MX and standard deviation
% SX and plots the probability density function

% create an array of x-values between -10 and 10 at 0.01 intervals

x = -10:0.01:10;

a = (x-mx)./sx;

% calculate the density function

y = 1/(sx*sqrt(2*pi)).*exp(-0.5.*a.^2);

% Clear Figure 1 and plot the probability density curve

figure(1);
clf(1);
box on;
plot(x,y,'k');
grid on;
% anotate the plot
title('Normal Probability Density curve')
xlabel('x');
ylabel('f(x)');

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–9

RMIT University Geospatial Science

For continuous random variables X, the cumulative distribution function FX ( x ) has the

following properties
x
1. FX ( x ) = P ( X ≤ x ) = ∫ f X ( x ) dx
−∞

d
2. FX ( x ) = f X ( x )
dx

In surveying, the Normal distribution is the usual model for the behaviour of measurements
and the cumulative distribution function is (Kreyszig, 1970, p. 108)

1 ⎛ x−μ ⎞
2

1 − ⎜ ⎟
FX ( x ) =
x
∫
2⎝ σ ⎠
e dx (3.4)
σ 2π −∞

The probability that X assumes any value in an interval a < X < b is

1 ⎛ x−μ ⎞
2

1 − ⎜ ⎟
P ( a < X < b ) = FX ( b ) − FX ( a ) =
b
∫
2⎝ σ ⎠
e dx (3.5)
σ 2π a

Figure 3.4 shows a plot of the Normal cumulative distribution curve for μ = 2.0 and σ = 2.5 .

Normal Cumulative Distribution curve

1

0.9

0.8

0.7

0.6
F(x)

0.5

0.4

0.3

0.2

0.1

0
-10 -8 -6 -4 -2 0 2 4 6 8 10
x

Figure 3.4 Normal cumulative distribution function for μ = 2.0 and σ = 2.5

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–10

RMIT University Geospatial Science

3.1.3. Multivariate Probability Density Functions

For multiple random variables, X 1 , X 2 , X 3 , … the term multivariate probability density

function f X ( x ) is used to define a function whose integral gives the probability of X 1 lying

in the range a1 < X 1 < b1 , X 2 lying in the range a2 < X 2 < b2 , X 3 lying in the range

a3 < X 3 < b3 , etc. This probability is

b1 b2 b3

P=∫∫∫ f X1 X 2 ( x1 , x2 , x3 , ) dx1 dx2 dx3

a1 a2 a3 (3.6)
= ∫ f X ( x ) dx
b

where x = [ x1 x2 x3 ]T , a = [a1 a2 a3 ]T and b = [ b1 b2 b3 ]T

Although least squares adjustment theory does not require the random variables
(measurements) to have particular probability distributions, the Normal distribution is the
usual model assumed to represent measurements and associated errors and corrections
(residuals). The Multivariate Normal distribution of random variables has a density function
of the following form (Mikhail, 1976, p. 27)

⎧⎪ 1 ⎫⎪ ⎧ 1 ⎫
f X ( x1 , x2 , , xn ) = f X ( x ) = ⎨ ⎬ × exp ⎨ − ( x − μ x ) Σ ( x − μ x ) ⎬
−1
T
(3.7)
⎪⎩ ( 2π ) Σ ⎪⎭ ⎩ 2 ⎭
n 2

with mean vector μ x and variance-covariance matrix Σ . For the case of two random
variables X and Y the Bivariate Normal probability density function has the following form

1 ⎧⎪ σ x2σ y2
f XY ( x, y ) = × exp ⎨ −
2π σ x2σ y2 − σ xy2 ⎪⎩ 2 (σ xσ y − σ xy )
2 2 2

(3.8)
⎡ ( x − μ )2 ( x − μ x ) ( y − μ y ) ( y − μ y ) ⎤ ⎫⎪
2

⎢ x
− 2σ xy + ⎥⎬
⎢ σx σ x2σ y2 σ y2 ⎥ ⎪
2
⎣ ⎦⎭

where μ x , μ y are the means, σ x2 , σ y2 are the variances of the random variables X and Y

respectively and σ xy is the covariance.

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–11

RMIT University Geospatial Science

Figure 3.5 shows a 3-dimensional plot of a Bivariate Normal probability density function with
μ x = 0.8, μ y = −0.2 , σ x = 1.5, σ y = 1.2 and σ xy = −0.5 over a range of possible values x, y of
the random variables X and Y.

Bivariate Normal Probability Density surface

0.1
f(x,y)

0.05

0
5

y 0
5

0
x
-5 -5

Figure 3.5 Bivariate Normal probability density surface

Planes having values f XY ( x, y ) = constant that cut the density surface will create elliptical

curves of intersection.

The MATLAB function bivariate_normal.m was used to create the Bivariate Normal
probability density surface of Figure 3.5. The equation of the Bivariate Normal probability
density function f XY ( x, y ) used in the function is a modified form of (3.8) where the

correlation coefficient

σ xy
ρ xy = (3.9)
σ xσ y

is used and equation (3.8) becomes

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–12

RMIT University Geospatial Science

1 ⎧⎪ 1
f XY ( x, y ) = × exp ⎨ −
2π σ xσ y 1 − ρ xy2 ⎪⎩ 2 (1 − ρ xy )
2

(3.10)
⎡ ( x − μ )2 ( ) ( ) ( ) ⎤⎫
2
− μ − μ − μ
⎥ ⎪⎬
x y y
⎢ x
− 2 ρ xy
x y
+
y

⎢ σx σ xσ y σ y ⎥⎪
2 2
⎣ ⎦⎭

function bivariate_normal(mx,my,sx,sy,sxy)
% Function BIVARIATE_NORMAL(MX,MY,SX,SY,SXY) calculates the bivariate
% normal density function f(x,y) of two random variables X and Y having
% NORMAL distributions with means MX, MY, standard deviations SX, SY and
% covariance SXY and plots the probability density surface.

% calculate correlation coefficient r (rho)

r = sxy/(sx*sy);
r2 = r*r;

% create arrays of x and y values between -3 and 3 at 0.2 intervals

[x,y] = meshgrid(-5:0.25:5);

a = (x-mx)./sx;
b = (y-my)./sy;
c = a.^2 - (a.*b).*(2*r) + b.^2;

% calculate the density function z = f(x,y)

z = 1/(2*pi*sx*sy*sqrt(1-r2)).*exp(-1/(2*(1-r2)).*c);

% Clear Figure 1 and plot the probability density surface

figure(1);
clf(1);
box on;
mesh(x,y,z,'EdgeColor','black');
% anotate the plot
title('Bivariate Normal Probability Density surface')
xlabel('x');
ylabel('y');
zlabel('f(x,y)');

3.1.4. Expectations

The expectation E { X } of a random variable X is defined as the average value μ X of the

variable over all possible values. It is computed by taking the sum of all possible values of
X = x multiplied by its corresponding probability. In the case of a discrete random variable
the expectation is given by
N
E { X } = μ X = ∑ xk P ( xk ) (3.11)
k =1

Equation (3.11) is a general expression from which we can obtain the usual expression for the
arithmetic mean

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–13

RMIT University Geospatial Science

N
1
μ=
N
∑x
k =1
k (3.12)

If there are N possible values xk of the random variable X, each having equal probability

P ( xk ) = 1 N (which is a constant), then the expectation computed from (3.11) is identical to

the arithmetic mean of the N values of xk from (3.12).

In the case of a continuous random variable the expectation is given by

+∞
E {X } = μX = ∫ x f ( x ) dx
X (3.13)
−∞

This relationship may be extended to a more general form if we consider the expectation of a
function g ( X ) of a random variable X whose probability density function is f X ( x ) . In this

case
+∞
E {g ( X )} = ∫ g ( x ) f ( x ) dxX (3.14)
−∞

Extending (3.14) to the case of two random variables X and Y

+∞ +∞
E {g ( X , Y )} = ∫ ∫ g ( x, y ) f ( x, y ) dx dy
XY
−∞ −∞

Similarly for n random variables

+∞ +∞ +∞
E {g ( X 1 , X 2 ,… , X n )} = ∫ ∫ ∫ g ( x , x ,…, x ) f ( x , x ,…, x ) dx dx … dx
1 2 n X 1 2 n 1 2 n (3.15)
−∞ −∞ −∞

Expressing (3.15) in matrix notation gives a general form of the expected value of a
multivariate function g ( X ) as

+∞ +∞ +∞
E {g ( X )} = ∫ ∫ ∫ g ( x ) f ( x ) dx X (3.16)
−∞ −∞ −∞

where f X ( x ) is the multivariate probability density function.

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–14

RMIT University Geospatial Science

There are some rules that are useful in calculating expectations. They are given here without
proof but can be found in many statistical texts, e.g., Walpole, 1974. With a and b as
constants and X and Y as random variables
E {a} = a

E {aX } = a E { X }

E {aX + b} = a E { X } + b

E {g ( X ) ± h ( X )} = E {g ( X )} ± E {h ( X )}

E {g ( X , Y ) ± h ( X , Y )} = E {g ( X , Y )} ± E {h ( X , Y )}

3.1.5. Special Mathematical Expectations

The mean of a random variable

+∞
μX = E {X } = ∫ x f ( x ) dx
X (3.17)
−∞

The mean vector m X of a multivariate distribution is

⎡ μ X1 ⎤ ⎡ E ( X 1 ) ⎤ ⎡ X1 ⎤
⎢μ ⎥ ⎢ ⎥ ⎢X ⎥
⎢ X2 ⎥
⎢ E ( X 2 )⎥
mX = = = E ⎢ 2 ⎥ = E {X} (3.18)
⎢ μX3 ⎥ ⎢ E ( X 3 )⎥ ⎢ X3⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣⎢ ⎦⎥ ⎣ ⎦ ⎣ ⎦

m X can be taken as representing the mean of a multivariate probability density function.

The variance of a random variable

+∞

{
σ X2 = E ( X − μ X ) =
2
} ∫ (x − μ )
−∞
x
2
f X ( x ) dx (3.19)

The covariance between two random variables X and Y is

+∞ +∞
σ XY = E {( X − μ X )(Y − μY )} = ∫ ∫ ( x − μ ) ( y − μ ) f ( x, y ) dx dy
−∞ −∞
x y XY (3.20)

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–15

RMIT University Geospatial Science

Equation (3.20) can be expanded to give

σ XY = E {( X − μ X )(Y − μY )}
= E { XY − X μY − Y μ X + μ X μY }
= E { XY } − E { X μY } − E {Y μ X } + E {μ X μY }
= E { XY } − μY E { X } − μ X E {Y } + μ X μY
= E { XY } − μY μ X − μ X μY + μ X μY
= E { XY } − μ X μY

If the random variables X and Y are independent, the expectation of the product is equal to the
product of the expectations, i.e., E { XY } = E { X } E {Y } . Since the expected values of X and Y

are the means μ X and μY then E { XY } = μ X μY if X and Y are independent. Substituting this

result into the expansion above shows that the covariance σ XY is zero if X and Y are
independent.

For a multivariate function, variances and covariances of the random variables X is given by
the matrix equation

{
Σ XX = E [ X − m X ][ X − mY ]
T
} (3.21)

Σ XX is a symmetric matrix known as the variance-covariance matrix and its general form can
be seen when (3.21) is expanded

⎧ ⎡ X 1 − μ X1 ⎤ ⎫
⎪⎢ ⎥ ⎪
⎪⎢ X 2 − μX2 ⎥ ⎪
Σ XX = E⎨ ⎡ X − μ X1 X 2 − μ X2 X n − μ X n ⎤⎦ ⎬
⎪⎢ ⎥⎣ 1 ⎪
⎪⎢ ⎥
⎩ ⎣⎢ X n − μ X n ⎦⎥
⎪
⎭

⎡ σ X2 1 σ X1 X 2 σX X ⎤
⎢ ⎥
1 n

⎢σ X X σ X2 1 σX ⎥
giving Σ XX =⎢ 2 1 2Xn
⎥ (3.22)
⎢ ⎥
⎢⎣σ X n X1 σ X n X 2 σ X2 1
⎥⎦

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–16

RMIT University Geospatial Science

3.2. Law of Propagation of Variances for Linear Functions

Consider two vectors of random variables x = [ X 1 X n ] and

T
X2

y = [Y1 Y2 Yn ] that are linearly related by the matrix equation

T

y = Ax + b (3.23)

where A is a coefficient matrix and b is a vector of constants. Then, using the rules for
expectations developed above we may write an expression for the mean mY using (3.18)

mY = E {y}
= E {Ax + b}
= E {Ax} + E {b}
= A E {x} + b
= Am X + b

Using (3.21), the variance-covariance matrix Σ yy is given by

{
Σ yy = E ( y − m y )( y − m y )
T
}
{
= E ( Ax + b − Am x − b )( Ax + b − Am x − b )
T
}
= E {( Ax − Am )( Ax − Am ) }
T
x x

{
= E A ( x − mx ) ( A ( x − mx ))
T
}
{
= E A ( x − m x )( x − m x ) A T
T
}
{
= A E ( x − m x )( x − m x )
T
}A T

= AΣ xx A T

or

If y = Ax + b and y and x are random variables linearly related then

Σ yy = AΣ xx A T (3.24)

Equation (3.24) is known as the Law of Propagation of Variances.

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–17

RMIT University Geospatial Science

Using the relationships previously established between variance-covariance matrices and

cofactor matrices, i.e., Σ = σ 02Q , the Law of Propagation of Variances may also applies to
cofactor matrices

If y = Ax + b and y and x are random variables linearly related then

Q yy = AQ xx A T (3.25)

3.3. Law of Propagation of Variances for Non-Linear Functions

In many practical applications of variance propagation the random variables in x and y are
nonlinearly related, i.e.,

y = f (x) (3.26)

In such cases, we can expand the function on the right-hand-side of (3.26) using Taylor's
theorem.

For a non-linear function of a single variable Taylor's theorem may be expressed in the
following form

(x − a) ( x − a)
2 3
df d2 f d3 f
f ( x ) = f (a ) + ( x − a) + 2 + 3 +
dx a dx a
2! dx a
3!
(3.27)
(x − a)
n −1
d n −1 f
+ + Rn
dx n −1 a
( n − 1) !

where Rn is the remainder after n terms and lim Rn = 0 for f ( x ) about x = a and
n →∞

df d2 f
, etc are derivatives of the function f ( x ) evaluated at x = a .
dx a dx 2 a

For a non-linear function of two random variables, say φ = f ( x, y ) , the Taylor series

expansion of the function φ about x = a and y = b is

∂f ∂f
φ = f ( a, b ) + ( x − a) + ( y − b)
∂x a ,b ∂y a ,b
(3.28)
1 ⎧⎪ ∂ 2 f ∂2 f ∂f ∂f ⎫
+ ⎨ 2 ( x − a) + 2 ( y − b) + ( x − a )( y − b ) ⎪⎬ +
2 2

2! ⎪⎩ ∂x a ,b
∂y a ,b
∂x a ,b ∂y a ,b ⎪⎭

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–18

RMIT University Geospatial Science

∂f ∂f ∂2 f
where f ( a , b ) is the function φ evaluated at x = a and y = b , and , ,
∂x a ,b ∂y a ,b
∂x 2 a ,b

etc are partial derivatives of the function φ evaluated at x = a and y = b .

Extending to n random variables, we may write a Taylor series approximation of the function
f ( x ) as a matrix equation

∂f
f ( x ) = f ( x0 ) + ( x − x 0 ) + higher order terms (3.29)
∂x x0

∂f
where f ( x 0 ) is the function evaluated at the approximate values x 0 and are the partial
∂x x0

derivatives evaluated at approximations x 0 .

Replacing f ( x ) in (3.26) by its Taylor series approximation, ignoring higher order terms,

gives

∂f
y = f ( x ) = f ( x0 ) + ( x − x0 ) (3.30)
∂x x0

Then, using the rules for expectations

m y = E {y}
⎧ ∂f ⎫
= E ⎨ f ( x0 ) + ( x − x0 )⎬
⎩ ∂x x0 ⎭
⎧ ∂f ⎫
{ }
= E f ( x0 ) + E ⎨
⎩ ∂x x0
( x − x0 )⎬
⎭
∂f
= f ( x0 ) +
∂x x0
{
E ( x − x0 ) }
∂f
= f ( x0 ) +
∂x x0
(
E {x} − E {x 0 } )
∂f
= f ( x0 ) + ( mx − x0 )
∂x x0

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–19

RMIT University Geospatial Science

And

⎡ ∂f ⎤ ⎡ ∂f ⎤
y − m y = ⎢ f ( x0 ) + ( x − x 0 )⎥ − ⎢ f ( x 0 ) + ( m x − x 0 )⎥
⎣ ∂x x0 ⎦ ⎣ ∂x x0 ⎦
∂f
= (x − mx ) (3.31)
∂x x0
= J yx ( x − m x )

J yx is the (m,n) Jacobian matrix of partial derivatives, noting that y and x are (m,1) and (n,1)

vectors respectively

⎡ ∂y1 ∂x1 ∂y1 ∂x2 ∂y1 ∂xn ⎤

⎢ ∂y ∂x ∂y ∂x ∂y2 ∂xn ⎥
J yx = ⎢ 2 1 2 2 ⎥ (3.32)
⎢ ⎥
⎢ ⎥
⎣∂ym ∂x1 ∂ym ∂x2 ∂ym ∂xn ⎦

Using (3.21), the variance-covariance matrix Σ yy is given by

{
Σ yy = E ( y − m y )( y − m y )
T
}
= E {( J ( x − m x ) ) ( J yx ( x − m x ) ) }
T
yx

{
= E J yx ( x − m x )( x − m x ) J Tyx
T
}
{
= J yx E ( x − m x )( x − m x )
T
}J T
yx

= J yx Σ xx J Tyx

Thus, in a similar manner to above, we may express the Law of Propagation of Variances for
non-linear functions of random variables as

If y = f ( x ) and y and x are random variables non-linearly related then

Σ yy = J yx Σ xx J Tyx (3.33)

This rule also applies to cofactor matrices

If y = f ( x ) and y and x are random variables non-linearly related then

Q yy = J yx Q xx J Tyx (3.34)

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–20

RMIT University Geospatial Science

3.4. The Special Law of Propagation of Variances

The Law of Propagation of Variances is often expressed as an algebraic equation. For

example, if z is a function of two random variables x and y, i.e., z = f ( x, y ) then the variance

of z is
2
⎛ ∂z ⎞ ⎛ ∂z ⎞ ∂z ∂z
2

σ = ⎜ ⎟ σ x2 + ⎜ ⎟ σ y2 + 2
2
σ xy (3.35)
⎝ ∂x ⎠ ⎝ ∂y ⎠ ∂x ∂y
z

Equation (3.35) can be derived from the general matrix equation (3.33) in the following
⎡ x⎤
manner. Let z = f ( x, y ) be written as y = f ( x ) where y = [ z ] , a (1,1) matrix and x = ⎢ ⎥
⎣ y⎦
⎡ σ x2 σ xy ⎤
is a (2,1) vector. The variance-covariance matrix of the random vector x is Σ xx = ⎢ 2 ⎥
,
⎣σ xy σ y ⎦
⎡ ∂z ∂z ⎤
the Jacobian J yx = ⎢ and the variance-covariance matrix Σ yy which contains the
⎣ ∂x ∂y ⎥⎦

single element σ z2 is given by

⎡ ∂z ⎤
⎡ ∂z ∂z ⎤ ⎡ σ
2
σ xy ⎤ ⎢ ∂x ⎥
Σ yy = ⎡⎣σ z2 ⎤⎦ = ⎢ ⎥⎢ ⎥
x
⎢
⎣ ∂x ∂y ⎥⎦ ⎣σ xy σ y2 ⎦ ⎢ ∂z ⎥
⎢ ∂y ⎥
⎣ ⎦

Expanding this equation gives (3.35).

In the case where the random variables in x are independent, i.e., their covariances are zero;
we have the Special Law of Propagation of Variances. For the case of z = f ( x, y ) where the

random variables x and y are independent, the Special Law of Propagation of Variances is
written as

If z = f ( x, y ) and x and y are independent random variables then

2
⎛ ∂z ⎞ ⎛ ∂z ⎞
2

σ = ⎜ ⎟ σ x2 + ⎜ ⎟ σ y2
2
(3.36)
⎝ ∂x ⎠ ⎝ ∂y ⎠
z

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–21

RMIT University Geospatial Science

3.5. Applications of Propagation of Variances

Propagation of variances, using either the Law of Propagation of Variances for linear
functions, equations (3.24) or (3.25) and non-linear functions, equations (3.33) or (3.34), or
the Special Law of Propagation of Variances, equation (3.36) where the variables are
independent are important "tools" for assessing the precision of computed quantities arising
from measurements. Implicit in every application of variance propagation is an a priori
knowledge of the precision of the measurements. For example, if quantities are derived from
Total Station EDM distances, then some knowledge of the precision of those distances is
assumed; if height differences are computed from a combination of Total Station EDM
distances and zenith angles, then the precisions of distances and zenith angles is assumed.
These a priori precision estimates may come from information supplied by the equipment
manufacturer, statistical analysis of observations, prior knowledge or simply educated
guesses. Whatever the source of knowledge, it is assumed that these precisions are known
before any variance propagation is made. The following sections set out some useful
techniques that are applicable to many surveying operations.

3.5.1. Variances of Total Station Horizontal Distances and Height Differences

Total Stations are modern surveying instruments combining an electronic theodolite (for
angle measurement) and an EDM (for distance measurement to a reflecting prism). EDM is
an abbreviation of Electronic Distance Measurement. The primary Total Station
measurements are horizontal and vertical circle readings α and β respectively (from which
angles may be obtained) and slope distances D. Total Stations have "on board" computers
and may display (at the push of a button) computed quantities such as vertical height
differences ±V and horizontal distances H, between the instrument and the prism (the
sighting target).

A vertical circle reading β made with a Total Station is a clockwise angle from the zenith
(defined by the vertical axis of the Total Station) measured in a vertical plane. This vertical
plane is swept out by the telescope rotating about the Total Station's horizontal axis (the
trunnion axis). The horizontal distance H and the vertical component ±V of a measured
slope distance S are

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–22

RMIT University Geospatial Science

H = D sin β

V = D cos β

H and V are functions of D and β which we may write as H = H ( D, β ) and V = V ( D, β ) .

Treating H, V, D and β as random variables and assuming that D and β are independent the
Special Law of Propagation of Variances (3.36) can be used to compute the variances of H
and V
2
⎛ ∂H ⎞ 2 ⎛ ∂H ⎞ 2
2

σ =⎜
2
⎟ σD + ⎜ ⎟ σ β = sin 2 β (σ D2 ) + D 2 cos2 β (σ β2 )
⎝ ∂D ⎠ ⎝ ∂β ⎠
H

2
⎛ ∂V ⎞ 2 ⎛ ∂V ⎞ 2
2

σ =⎜
2
⎟ σD + ⎜ ⎟ σ β = cos2 β (σ D2 ) + D 2 sin 2 β (σ β2 )
⎝ ∂D ⎠ ⎝ ∂β ⎠
V

σ D2 , σ β2 are the variances of the slope distance and zenith angle respectively. For any
properly calibrated Total Station EDM and prism combination, the standard deviation of a
distance can be expressed in the form σ D = x + y ppm (ppm is parts per million) and the

standard deviation of a zenith angle as σ β = x seconds of arc .

As an example: say D = 60.000 m, β = 85 36′ 00′′ and σ D = 8 mm + 5 ppm = 0.0083 m ,

σ β = 10′′ = 4.8481 × 10−5 radians ; the variances in H and V are

σ H2 = ( 0.9971) ( 0.0083) + ( 60 ) ( 0.0767 ) ( 4.8418 × 10−5 )

2 2 2 2 2

= 6.85 × 10−5 m 2

σ V2 = ( 0.0767 ) ( 0.0083) + ( 60 ) ( 0.9971) ( 4.8418 × 10−5 )

2 2 2 2 2

= 8.82 × 10−6 m 2

The standard deviations are of H and V are

σ H = 6.85 × 10−5 = 0.0083 m

σ V = 8.82 × 10−6 = 0.0030 m

It is interesting to note that the computed quantities H and V are not independent, even though
they have been computed from independent quantities. This can be seen by computing the

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–23

RMIT University Geospatial Science

variances in the following manner. We may write the computation of the components H and
V from the observations D and β as the vector equation

y = f (x)

⎡H ⎤ ⎡D⎤
y = ⎢ ⎥ , x = ⎢ ⎥ and y and x are non-linearly related. Using (3.33) we may write
⎣V ⎦ ⎣β ⎦

⎡σ2 σ HV ⎤
Σ yy = ⎢ H ⎥ = J yx Σ xx J yx
T

⎣σ HV σ V2 ⎦
⎡∂H ∂D ∂H ∂β ⎤ ⎡σ D 0 ⎤ ⎡∂H ∂D ∂V ∂D ⎤
2

=⎢ ⎥⎢ 2⎥⎢ ⎥
⎣ ∂V ∂D ∂V ∂β ⎦ ⎣ 0 σ β ⎦ ⎣ ∂H ∂β ∂V ∂β ⎦
⎡ sin β D cos β ⎤ ⎡σ D2 0 ⎤ ⎡ sin β cos β ⎤
=⎢ ⎢ ⎥
⎣cos β − D sin β ⎥⎦ ⎣ 0 σ β2 ⎦ ⎢⎣ D cos β − D sin β ⎥⎦

4.6031 ⎤ ⎡( 0.0083) ⎤ ⎡ 0.9971

2
⎡0.9971 0 0.0767 ⎤
=⎢ ⎥⎢ ⎥⎢ ⎥
− ⎦ ⎢⎣ ( 4.8481 × 10−5 ) ⎥⎦ ⎣4.6031 −59.8232⎦
2
⎣ 0.0767 59.8232 0
⎡ 6.85 × 10−5 4.62 × 10−6 ⎤
=⎢ −6 ⎥
⎣ 4.62 × 10 8.82 × 10−6 ⎦

The diagonal elements of Σ yy are σ H2 = 6.85 × 10−5 m 2 and σ V2 = 8.82 × 10−6 m 2 which are

the same values as computed above and the off-diagonal elements are the covariance
σ HV = 4.62 × 10−6 m 2 . These elements, which are non-zero, indicate that computed quantities
are correlated.

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–24

RMIT University Geospatial Science

3.5.2. Variances of Height Differences from Spirit Levelling

rP level line rQ
d d
staff

Δh
P

Figure 3.1

Figure 3.1 shows a schematic diagram of a spirit level and a single height difference ΔhPQ

between positions of the levelling staves at P and Q. The backsight and foresight staff
readings are rP and rQ and the length of the sights to the levelling staff are the same and equal

to d. The height difference is

ΔhPQ = rP − rQ

Considering the backsight and foresight staff readings to be independent and of equal
precision, the Special Law of Propagation of Variances (3.36) gives the variance of a single
height difference from spirit levelling as
2
⎛ ∂Δh ⎞ 2 ⎛ ∂Δh ⎞ 2
2

σ 2
Δh =⎜ ⎟ σ rP + ⎜⎜ ⎟⎟ σ rQ = σ rP + σ rQ = 2σ r
2 2 2
(3.37)
⎝ ∂rP ⎠ ⎝ ∂rQ ⎠

σ r is the standard deviation of a staff reading.

Suppose a flight of levels is run between two points A and B that are a distance D apart and
that at every set up of the level, the backsight and foresight distances are the same and equal
to d and the staff readings are all of equal precision σ r . There will be n = D 2d set ups and

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–25

RMIT University Geospatial Science

the height difference ΔH AB will be the sum of the individual height differences of the level
run

ΔH AB = Δh1 + Δh2 + + Δhn

Considering each Δh to be independent and of equal precision, the variance of the total
height difference ΔH AB is given by the Special Law of Propagation of Variances

σ Δ2H = σ Δ2h + σ Δ2h +

AB 1 2
+ σ Δ2hn = nσ Δ2h (3.38)

It is usual practice in spirit levelling to "close the level run" by returning to the start, therefore
the mean height difference of a closed level run (or a levelling loop) is

ΔH AB + ΔH BA
ΔH MEAN =
2

Again, considering ΔH AB and ΔH BA to be independent and of equal precision, the Special

Law of Propagation of Variances gives, bearing in mind (3.37) and (3.38) the variance of the
mean height difference of closed level run as

1 1 1 n
σ Δ2H = σ Δ2H AB + σ Δ2H BA = σ Δ2H = σ Δ2h = nσ r2 (3.39)
MEAN
4 4 2 2

n = D 2d is the number of set ups in the level run between A and B, that are distance D apart
(d is the length of the backsight/foresight distance) hence the variance in the mean height
difference is proportional D. Since weights are inversely proportional to variances, it is
common to express precisions of spirit levelled height differences as weights that are defined
as being inversely proportional to distances.

1
wΔH ∝ (3.40)
distance

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–26

RMIT University Geospatial Science

3.5.3. Variance of Trigonometric Heights

C′
γ

A C″
α b

A′ B′ δ a

c β

Figure 3.2

Figure 3.2 is a schematic diagram of vertical angle observations γ and δ at A and B to a

distant point C. Horizontal angles α and β are measured at A and B and the horizontal
distance c between A and B is measured. AC ′B ′ and A′C ′′B are horizontal planes. If the
heights of A and B are known then the height of C can be computed by plane trigonometry.
Heights computed in this manner are known as trigonometric heights.

If the observations, angles α , β , γ ,δ and distance c, have particular precisions then the
computed height of C will also have a precision that can be determined by propagation of
variances. The method of propagation will be demonstrated by the following example.

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–27

RMIT University Geospatial Science

Example. Referring to Figure 3.2, the following observations and standard deviations are
known together with the Reduced Levels of A and B

α = 75 33′ 25′′ σ α = 10′′ RLA = 105.450 m

β = 87 40′ 00′′ σ β = 20′′ RLB = 92.330 m
γ = 10 42′10′′ σ γ = 15′′
δ = 11 28′ 05′′ σ δ = 30′′
c = 505.450 m σ c = 0.050 m

Method of Computation of RL of C

(i) Compute sides a and b using the Sine Rule:

a b c c
= = =
sin α sin β sin (180 − (α + β ) ) sin (α + β )

giving

c sin α
a= = 1695.816217 m (3.41)
sin (α + β )

c sin β
b= = 1749.707936 m (3.42)
sin (α + β )

(ii) Compute the mean RL of C

RLC from A : RLC = RLA + b tan γ = 436.1486 m

RLC from B : RLC = RLB + a tan δ = 436.3632 m

RLA + RLB + b tan γ + a tan δ

RLC = = 436.256 m (3.43)
2

Method of Computation of Variance of RL of C

(i) Inspection of equations (3.41) and (3.42) shows that a and b are non-linear functions of
the variables α , β and c which we may write as

y = f (x)

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–28

RMIT University Geospatial Science

where y = [ a b] , x = [α β c ] and y and x are non-linearly related. Using (3.33)

T T

we may write

Σ yy = J yx Σ xx J Tyx (3.44)

⎡σ α2 0 0⎤
⎡σ 2
σ ab ⎤ ⎡ ∂a ∂α ∂a ∂β ∂a ∂c ⎤ ⎢ ⎥
where Σ yy = ⎢ a
2 ⎥
, J yx = ⎢ ⎥ and Σ xx = ⎢ 0 σ β2 0⎥
⎣σ ba σb ⎦ ⎣ ∂b ∂α ∂b ∂β ∂b ∂c ⎦
⎢0 0 σ c2 ⎥⎦
⎣
noting that the observations α , β and c are considered as independent random
variables, hence Σ xx is diagonal. The elements of the Jacobian J yx are the partial

derivatives of equations (3.41) and (3.42); they are obtained as follows.

(ii) partial derivative ∂a ∂α :

d ⎛ u ⎞ v ( du dx ) − u ( dv dx )
Using the rule for derivatives ⎜ ⎟= and the relation
dx ⎝ v ⎠ v2
d
sin (α + β ) = cosα we may write
dα

∂a sin (α + β ) c cos α − c sin α cos (α + β )

=
∂α sin 2 (α + β )
c ⎡sin (α + β ) cos α − cos (α + β ) sin α ⎤⎦
= ⎣
sin 2 (α + β )

Using the trigonometric function sin ( A − B ) = sin A cos B − cos A sin B where

A = α + β and B = α then A − B = β and we may write

∂a c sin β
= 2
∂α sin (α + β )

c sin β
Noting that b = the partial derivative is
sin (α + β )

∂a b
= (3.45)
∂α sin (α + β )

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–29

RMIT University Geospatial Science

(iii) partial derivative ∂a ∂β :

Similarly to the above derivation, we may write

∂a sin (α + β )( 0 ) − c sin α cos (α + β )

=
∂β sin 2 (α + β )
− c sin α cos (α + β )
=
sin 2 (α + β )

c sin a cos (α + β ) 1
Noting that a = and that = the partial derivative is
sin (α + β ) sin (α + β ) tan (α + β )

∂a −a
= (3.46)
∂β tan (α + β )

(iii) partial derivative ∂a ∂c :

∂a sin α a
= = (3.47)
∂c sin (α + β ) c

(iv) partial derivatives ∂b ∂α , ∂b ∂β , ∂b ∂c

Similarly to the above derivations

∂b −b
= (3.48)
∂α tan (α + β )

∂b a
= (3.49)
∂β sin (α + β )

∂b sin β b
= = (3.50)
∂c sin (α + β ) c

(v) Substitute numeric values into the equations for the partial derivatives and set the
elements of the Jacobian J yx

⎡ b −a a⎤
⎢ sin (α + β ) tan (α + β ) c ⎥ ⎡6061.961236 5625.191025 3.355062 ⎤
J yx = ⎢ ⎥=
⎢ −b a b ⎥ ⎢⎣5803.955218 5875.250353 3.461684 ⎥⎦
⎢ tan (α + β ) sin (α + β ) c ⎥⎥
⎢⎣ ⎦

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–30

RMIT University Geospatial Science

(vi) Substitute numeric values into the elements of the variance matrix Σ xx

⎡ ( 4.8481 × 10−5 ) 2 0 0 ⎤
⎡σ α2
0 0⎤ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
Σ xx = ⎢ 0 σ β2 0 ⎥=⎢ 0 ( 9.6963 × 10 )
−5 2
0 ⎥
⎢0 σ c ⎥⎦ ⎢
2
2⎥
⎣ 0
⎢⎣ 0 0 ( 0.050 ) ⎥
⎦

(vii) Perform the matrix multiplications in (3.44) to give the variance-covariance matrix of
the computed distances a and b

⎡ σ a2 σ ab ⎤ ⎡0.412012 0.422455⎤
Σ yy = ⎢ 2 ⎥
=⎢ ⎥ (3.51)
⎣σ ba σ b ⎦ ⎣ 0.422455 0.433671⎦

Note that the off-diagonal terms are not zero, indicating that the computed quantities
are correlated. The leading-diagonal elements are the variances of a and b, hence the
standard deviations are

σ a = 0.412012 = 0.642 m
σ b = 0.433671 = 0.659 m

(viii) Inspection of equation (3.43) shows that the mean RL of C is non-linear function of the
correlated variables a, b and the independent variables γ ,δ that we may write as

y = f (x)

y = [ RLC ] , x = [ a b γ δ ] and y and x are non-linearly related. Using (3.33) we

T

may write

Σ yy = J yx Σ xx J Tyx (3.52)

where Σ yy = ⎡⎣σ RL
2
⎤ , J yx = [ ∂RLC ∂a ∂RLC ∂b ∂RLC ∂γ
C ⎦
∂RLC ∂δ ] and

⎡ σ a2 σ ab 0 0⎤
⎢ ⎥
σ σ b2 0 0⎥
Σ xx = ⎢ ba
⎢ 0 0 σ γ2 0⎥
⎢ ⎥
⎣⎢ 0 0 0 σ δ2 ⎦⎥

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–31

RMIT University Geospatial Science

Note that Σ yy is a matrix containing a single element only, the variance of the mean RL

of C and Σ xx is a partitioned matrix where the upper-left part is the variance-covariance

matrix in equation (3.51) and the lower-right part contains the variances of the vertical
angles γ and δ . The Jacobian J yx is a row-vector containing the partial derivatives of

equation (3.43) that are given as follows.

(ix) partial derivatives ∂RLC ∂a , ∂RLC ∂b, ∂RLC ∂γ , ∂RLC ∂δ

d 1
Noting that tan x = sec 2 x = the partial derivatives are
dx cos2 x

∂RLC tan δ ∂RLC tan γ ∂RLC b ∂RLC a

= , = , = , =
∂a 2 ∂b 2 ∂γ 2cos2 γ ∂δ 2 cos2 δ

(x) Substitute numeric values into the equations for the partial derivatives and set the
elements of the Jacobian J yx

⎡ tan δ tan γ b a ⎤
J yx = ⎢
⎣ 2 2 2cos γ 2cos2 δ ⎥⎦
2

= [ 0.101436 0.094501 906.105346 882.805418]

(xi) Substitute numeric values into the elements of the variance matrix Σ xx

⎡ 0.412012 0.422455 0 0 ⎤
⎡ σ a2 σ ab 0 0⎤ ⎢ ⎥
⎢ ⎥ 0.422455 0.433671 0 0
⎢σ ba σ b2 0 0 ⎥ ⎢ ⎥
Σ xx = =⎢
⎢ 0 0 σ γ2 0 ⎥ ⎢ 0 0 ( 7.2722 × 10 )
−5 2
0
⎥
⎥
⎢ ⎥
⎢⎣ 0 0 0 σ δ2 ⎥⎦ ⎢ (1.4544 × 10−4 ) ⎥⎦⎥
2

⎣⎢ 0 0 0

(xii) Perform the matrix multiplications in (3.52) to give the variance-covariance matrix of
the computed mean RL of C.

Σ yy = [0.037040]

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–32

RMIT University Geospatial Science

Hence, the variance of the mean RL of C is

σ RL
2
C
= 0.037040 m 2

and the standard deviation of the mean RL of C is

σ RL = 0.037040 = 0.192 m
C

3.6. Propagation of Variances of Multiple Functions of Random Variables

In some applications of surveying and least squares, we must deal with multiple functions of
random variables that may be correlated. To handle these cases, Mikhail (1976, pp.83-87)
develops general rules and techniques that are repeated in the following sections. Note that
cofactor matrices Q replace variance-covariance matrices Σ in the following developments.

Consider x ( n ,1) and t ( m ,1) to be two correlated vectors with cofactor matrices Q xx , Q xt and Qtt .

Q xt is an (n,m) crosscofactor matrix containing estimates of covariances between the

elements of the x and t vectors. Two other vectors y ( q ,1) and z ( p ,1) are functions of x and t (y

is functionally independent of t and z of x).

y = y (x)
(3.53)
z = z (t)

The Jacobians of these functions are

∂y
J yx =
∂x
(3.54)
∂z
J zt =
∂t

⎡y⎤ ⎡x ⎤
Letting r = ⎢ ⎥ and s = ⎢ ⎥ equation (3.53) can be written as
⎣z⎦ ⎣t⎦

r = f (s) (3.55)

The Law of Propagation of Variances (3.34) can be applied to give

Q rr = J rs Q ss J Trs (3.56)

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–33

RMIT University Geospatial Science

with the Jacobian

⎡J J yt ⎤ ⎡ J yx 0⎤
J rs = ⎢ yx =
J zt ⎥⎦ ⎢⎣ 0 J zt ⎥⎦
(3.57)
⎣ J zx

Note that J yt and J zx are zero since y is independent of t and z is independent of x.

Equation (3.56) can be expanded to give

T
⎡ Q yy Q yz ⎤ ⎡ J yx 0 ⎤ ⎡Q xx Q xt ⎤ ⎡ J yx 0⎤
⎢Q =
⎣ zy Q zz ⎥⎦ ⎢⎣ 0 J zt ⎥⎦ ⎢⎣ Qtx Qtt ⎥⎦ ⎢⎣ 0 J zt ⎥⎦
⎡J 0 ⎤ ⎡Q xx Q xt ⎤ ⎡ J Tyx 0⎤
= ⎢ yx
⎣ 0 J zt ⎥⎦ ⎢⎣ Qtx Qtt ⎥⎦ ⎢⎣ 0 ⎥
J Tzt ⎦
⎡ J yx Q xx J Tyx J yx Q xt J Tzt ⎤
=⎢ T ⎥
⎣ J zt Qtx J yx J zt Qtt J Tzt ⎦

From this expanded equation, we may write the following symbolic equation and four general
relationships

⎡y⎤ ⎡ y ( x )⎤
If ⎢ ⎥ == ⎢ ⎥ and x and t are correlated
⎣z⎦ ⎢⎣ z ( t ) ⎥⎦

Q yy = J yx Q xx J Tyx
Q yz = J yx Q xt J Tzt
(3.58)
Q zy = J zt Qtx J Tyz
Q zz = J zt Qtt J Tzt

The cofactor matrices where the subscripts are different letters, i.e., Q xt , Qtx , Q yz and Q zy

are crosscofactor matrices and their form is easily constructed, for example

⎡ qx1 ,t1 qx1 ,t2 qx1 ,tn ⎤

⎢ ⎥
Q xt = ⎢ ⎥ (3.59)
m, n
⎢ q x ,t qxm ,t2 qxm ,tn ⎥⎦
⎣ m1

This directly leads to the fact that crosscofactor matrices are not necessarily square and
symmetric, as cofactor matrices are, and that

Q xt = QTtx
m, n

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–34

RMIT University Geospatial Science

3.6.1. Symbolic Multiplication in Propagation Using Matrices

To assist the practitioner in variance propagation (or cofactor propagation) of linear and non-
linear functions of random variables a technique known as symbolic multiplication can be
used. This mnemo-technical rule was originally devised by Tienstra (1966) to obtain
covariances of random variables related by systems of linear (or linearized) equations. His
rule, developed before the extensive use of matrix algebra, can be employed with matrix
equations in the following manner. For example, for linear functions

y = Ax + a (a)

z = Bt + b (b)

The constant vectors a and b play no part in variance propagation and can be ignored, and the
crosscofactor matrix Q yz is

Q yz = AQ xt BT (c)

Equation (c) can be obtained by symbolic multiplication by:

(i) writing the cofactor matrix Q on the left-hand-side of the equals sign with
subscripts y and z representing the vectors y and z on the left-hand-sides of
equations (a) and (b) in the order of (a) first and (b) second, then
(ii) writing the coefficient matrix of the random variable in equation (a) on the right-
hand-side of the equals sign, then
(iii) multiplying by the cofactor matrix Q with the subscripts x and t representing the
random vectors x and t on the right-hand-sides of equations (a) and (b) in the
order of (a) first and (b) second, then
(iv) multiplying by the transpose of the coefficient matrix of the random variable on
the right-hand-side of equation (b).

Note the in the case of non-linear functions, the coefficient matrices J yx and J zt replace A

and B respectively, and equation (c) becomes identical to the 2nd equation of (3.58).

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–35

RMIT University Geospatial Science

3.6.2. Further Remarks on Propagation Using Matrices

The rules and techniques of propagation of variances (and cofactors) given in the preceding
sections allow for propagation of variances through several transformations. These
propagations can be carried out in two ways, (i) substitution and (ii) stepwise. To
demonstrate these we consider the following three relations

y = Ax + a
z = By + b (3.60)
r = Cz + c

Let the random vector x be known, with its cofactor matrix Q xx and it is desired to obtain the
cofactor matrices of z and r. The vectors a, b and c contain constants and A, B and C are
coefficient matrices.

Propagation Through Substitution

We express z and r in terms of x by substitution:

y = Ax + a
z = By + b = B ( Ax + a ) + b
= ( BA ) x + ( Ba + b )
r = Cz + c = C ( By + b ) + c
= ( CB ) y + ( Cb + c )
= ( CB )( Ax + a ) + ( Cb + c )
= ( CBA ) x + ( CBa + Cb + c )

Noting that the last terms in the equations for z and r evaluate to vectors, say d and e we may
rewrite the equations and apply propagation using symbolic multiplication to give

y = Ax + a Q yy = AQ xx A T
z = ( BA ) x + d Q zz = ( BA ) Q xx ( BA ) = BAQ xx A T BT
T

r = ( CBA ) x + e Q rr = ( CBA ) Q xx ( CBA ) = CBAQ xx A T BT CT

T

(3.61)
Q yz = AQ xx ( BA ) = AQ xx A T BT
T

Q yr = AQ xx ( CBA ) = AQ xx A T BT CT
T

Q zr = ( BA ) Q xx ( CBA ) = BAQ xx A T BT CT
T

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–36

RMIT University Geospatial Science

Stepwise Propagation
The same result in equations (3.61) can be obtained by applying propagation in steps as
follows:

y = Ax + a Q yy = AQ xx A T
z = By + b Q zz = BQ yy BT = BAQ xx A T BT
r = Cz + c Q rr = CQ zz CT = CBAQ xx A T BT CT
(3.62)
Q yz = AQ xy BT
Q yr = AQ xz CT
Q zr = BQ yz CT

The last three (crosscofactor) relations of equations (3.62) do not correspond to those in
equations (3.61), particularly because of the absence of the matrices Q xy , Q xz and Q yz .

However, these matrices can be derived if equations (3.60) are supplemented by simple
identities x = Ix and y = Iy giving the following three pairs of equations from which the
matrix relationships can be obtained by symbolic multiplication.

x = Ix Q xy = IQ xx A T = Q xx A T
y = Ax + a
x = Ix Q xz = IQ xy BT = Q xx A T BT
(3.63)
z = By + b
y = Iy Q yz = IQ yy BT
z = By + b

Substituting these relations into the last three relations in (3.62) leads directly to equations
(3.61). This demonstrates that propagation through substitution is equivalent to stepwise
propagation.

© 2005, R.E. Deakin Notes on Least Squares (2005) 3–37

RMIT University Geospatial Science

4. APPROXIMATE VALUES

In many least squares problems, the unknown quantities being sought may be quite large
numbers and or the coefficients of these quantities may be large numbers. This can lead to
numerical problems in the formation of normal equations where large numbers are multiplied
and summed. To overcome this problem, approximate values of the unknown quantities may
be used and small, unknown corrections to the approximate values become the quantities
being sought.

In general, we denote unknown values as x, approximate values as x 0 and small corrections

as Δx or δ x and

x = x0 + δ x (4.1)

In the case of a vector of unknown quantities x we have a vector of approximate values x 0

and a vector of small corrections δ x and

x = x0 + δ x (4.2)

The use of approximate values can best be explained by example and the following sections
contain worked examples of some simple least squares problems that demonstrate the use of
approximate values.

4.1. LEVEL NET ADJUSTMENT

The diagram below shows a level network of height differences observed between the fixed
stations A (RL 102.440 m) and B (RL 104.565 m) and "floating" stations X, Y and Z whose
Reduced Levels (RL's) are unknown. The arrows on the diagram indicate the direction of
rise. The Table of Height differences shows the height difference for each line of the network
and the distance (in kilometers) of each level run.

© 2005, R.E. Deakin Notes on Least Squares (2005) 4–1

RMIT University Geospatial Science

X
1 •
Line Height Diff Dist (km) 2
1 6.345 1.7 6
2 4.235 2.5
3 3.060 1.0
A ⊗
4 0.920 3.8 5
• 7
⊗B
5 3.895 1.7 Y
6 2.410 1.2
7 4.820 1.5 3
•
4 Z

The method of Least Squares can be used to determine the best estimates of the RL's of X, Y
and Z bearing in mind that the precision of the observed height differences is inversely
proportional to the distance of the level run.
The observation equation for the RL's of two points P and Q connected by an observed spirit
levelled height difference ΔH PQ can be written as

P + ΔH PQ + vPQ = Q (4.3)

where P and Q are the RL's of points P and Q and vPQ is the residual, a small unknown

correction to the observed height difference. If the RL's of P and Q are unknown but have
approximate values, say P = P 0 + δ P and Q = Q 0 + δ Q we may write a general observation
equation for an observed height difference as

P 0 + δ P + ΔH PQ + vPQ = Q 0 + δ Q (4.4)

Using this general observation equation we may write an equation for each observed height
difference

A +ΔH1 + v1 = X 0 + δ X
B +ΔH 2 + v2 = X 0 + δ X
Z 0 + δ Z +ΔH 3 + v3 = B
Z 0 + δ Z +ΔH 4 + v4 = A
A +ΔH 5 + v5 = Y 0 + δ Y
Y 0 + δ Y +ΔH 6 + v6 = X 0 + δ X
Z 0 + δ Z + ΔH 7 + v7 = Y 0 + δ Y

© 2005, R.E. Deakin Notes on Least Squares (2005) 4–2

RMIT University Geospatial Science

Rearranging these equations so that all the unknown quantities are on the left-hand-side of the
equals sign and all the known quantities are on the right-hand-side gives

v1 − δ X = X 0 − A −ΔH1
v2 − δ X = X 0 − B −ΔH 2
v3 +δ Z = B − Z 0 −ΔH 3
v4 +δ Z = A − Z 0 −ΔH 4
v5 −δ Y = Y 0 − A −ΔH 5
v6 − δ X +δ Y = X 0 − Y 0 −ΔH 6
v7 −δ Y +δ Z = Y 0 − Z 0 −ΔH 7

The approximate RL's of the unknown points X, Y and Z can be determined from the RL's of
A and B and appropriate height differences

X 0 = A + ΔH1 = 108.785 m
Y 0 = A + ΔH 5 = 106.335 m
Z 0 = A − ΔH 4 = 101.520 m

Writing these equations in the standard form v + Bx = f gives

⎡ ( X 0 − A ) − ΔH 1 ⎤
⎢ ⎥
⎡ v1 ⎤ ⎡ −1 0 0⎤
⎢ ( X 0 − B ) − ΔH 2 ⎥ ⎡
0.000 ⎤
⎢ v ⎥ ⎢ −1 0 0 ⎥ ⎢ ⎥ ⎢ −0.015 ⎥
⎢ 2⎥ ⎢ ⎥ ⎢ ⎥
1⎥ ⎡δ X ⎤ ⎢ ( B − Z ) − ΔH 3 ⎥ ⎢ −0.015 ⎥
⎢ 0
⎥
⎢ v3 ⎥ ⎢ 0 0
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ v4 ⎥ + ⎢ 0 0 1⎥ ⎢ δ Y ⎥ = ⎢ ( A − Z 0 ) − ΔH 4 ⎥ = ⎢ 0.000 ⎥
⎢ ⎥ ⎢ ⎥
⎢ v5 ⎥ ⎢ 0 −1 0⎥ ⎣⎢ δ Z ⎦⎥ ⎢ Y 0 − A − ΔH ⎥ ⎢ 0.000 ⎥
⎢ ⎥ ⎢ ( )
⎥ ⎥ ⎢ 0.040 ⎥
5
⎢
⎢ v6 ⎥ ⎢ −1 1 0⎥
⎢ ( X − Y ) − ΔH 6 ⎥ ⎢
0 0 ⎥
⎢⎣ v7 ⎥⎦ ⎢⎣ 0 −1 1⎦⎥ ⎢ ⎥ ⎣ −0.005 ⎥⎦
⎢
⎢ ( Y − Z ) − ΔH 7 ⎥
0 0
⎣ ⎦

The weight matrix for the adjustment is

⎡ 1 1 1 1 1 1 1 ⎤
W = diag ⎢
⎣ 1.7 2.5 1 3.8 1.7 1.2 1.5 ⎥⎦
= diag [0.5882 0.4000 1.0000 0.2632 0.5882 0.8333 0.6667]

© 2005, R.E. Deakin Notes on Least Squares (2005) 4–3

RMIT University Geospatial Science

The least squares solution for the vector of corrections x can be obtained from the MATLAB
function least_squares.m with the following data file c:\Temp\Level_Net_Data.dat

% Data file for Level Net Adjustment

%
% dX dY dZ f weight
-1 0 0 0.000 0.5882
-1 0 0 -0.015 0.4000
0 0 1 -0.015 1.0000
0 0 1 0.000 0.2632
0 -1 0 0.000 0.5882
-1 1 0 0.040 0.8333
0 -1 1 -0.005 0.6667

Running the program from the MATLAB command window created the following output file
c:\Temp|Level_Net_Data.out

Least Squares Adjustment of Indirect Observations

Input Data

Coefficient matrix B of observation equations v + Bx = f

-1.0000 0.0000 0.0000
-1.0000 0.0000 0.0000
0.0000 0.0000 1.0000
0.0000 0.0000 1.0000
0.0000 -1.0000 0.0000
-1.0000 1.0000 0.0000
0.0000 -1.0000 1.0000

Vector of numeric terms f and weights w of observation equations v + Bx = f

0.0000 0.5882
-0.0150 0.4000
-0.0150 1.0000
0.0000 0.2632
0.0000 0.5882
0.0400 0.8333
-0.0050 0.6667

Coefficient matrix N of Normal equations Nx = t

1.8215 -0.8333 0.0000
-0.8333 2.0882 -0.6667
0.0000 -0.6667 1.9299

Vector of numeric terms t of Normal equations Nx = t

-0.0273
0.0367
-0.0183

Inverse of Normal equation coefficient matrix

6.9073e-001 3.0981e-001 1.0703e-001
3.0981e-001 6.7720e-001 2.3394e-001
1.0703e-001 2.3394e-001 5.9898e-001

Vector of solutions x
-0.0095
0.0121
-0.0053

© 2005, R.E. Deakin Notes on Least Squares (2005) 4–4

RMIT University Geospatial Science

Vector of residuals v
-0.0095
-0.0245
-0.0097
0.0053
0.0121
0.0184
0.0124

The adjusted RL's of X, Y and Z are

X = X 0 + δ X = 108.785 − 0.0095 = 108.776 m

Y = Y 0 + δ Y = 106.335 + 0.0121 = 106.347 m
Z = Z 0 + δ Z = 101.520 − 0.0053 = 101.515 m

The adjusted height differences are

Line Observed ΔH Residual v Adjusted ΔH

1 6.345 -0.0095 6.336
2 4.235 -0.0245 4.211
3 3.060 -0.0097 3.050
4 0.920 0.0053 0.925
5 3.895 0.0121 3.907
6 2.410 0.0184 2.428
7 4.820 0.0124 4.832

© 2005, R.E. Deakin Notes on Least Squares (2005) 4–5

RMIT University Geospatial Science

5. PROPAGATION OF VARIANCES APPLIED TO LEAST SQUARES

ADJUSTMENT OF INDIRECT OBSERVATIONS

A most important outcome of a least squares adjustment is that estimates of the precisions of
the quantities sought, the elements of x, the unknowns or the parameters, are easily obtained
from the matrix equations of the solution. Application of the Law of Propagation of
Variances demonstrates that N −1 , the inverse of the normal equation coefficient matrix is
equal to the cofactor matrix Q xx that contains estimates of the variances and covariances of
the elements of x. In addition, estimates of the precisions of the residuals and adjusted
observations may be obtained. This most useful outcome enables a statistical analysis of the
results of a least squares adjustment and provides the practitioner with a degree of confidence
in the results.

5.1. Cofactor matrices for adjustment of indirect observations

The observation equations for adjustment of indirect observations is given by

v + Bx = f (5.1)

f is an (n,1) vector of numeric terms derived from the (n,1) vector of observations l and the
(n,1) vector of constants d as

f = d−l (5.2)

Associated with the vector of observations l is a variance-covariance matrix Σl l as well as a

cofactor matrix Ql l and a weight matrix Wl l = Ql−l1 . Remember that in most practical

applications of least squares, the matrix Σl l is unknown, but estimated a priori by Ql l that

contains estimates of the variances and covariances and Σ l l = σ 02Ql l where σ 02 is the

reference variance or variance factor.

Note: In the derivations that follow, the subscript "ll" is dropped from Ql l and Wl l

If equation (5.2) is written as

f = ( −I ) l + d (5.3)

© 2005, R.E. Deakin Notes on Least Squares (2005) 5–1

RMIT University Geospatial Science

then (5.3) is in a form suitable for employing the Law of Propagation of Variances developed
in Chapter 3; i.e., if y = Ax + b and y and x are random variables linearly related and b is a

vector of constants then Q yy = AQ xx A T . Hence, the cofactor matrix of the numeric terms f is

Q ff = ( − I ) Q ( − I ) = Q
T

Thus the cofactor matrix of f is also the a priori cofactor matrix of the observations l.

The solution "steps" in the least squares adjustment of indirect observations are set out
Chapter 2 and restated as

N = BT W B
t = BT W f
x = N −1 t
v = f − Bx
ˆl = l + v

To apply the Law of Propagation of Variances, these equations may be re-arranged in the
form y = Ax + b where the terms in parenthesis ( ) constitute the A matrix.

t = ( BT W ) f (5.4)

x = ( N −1 ) t (5.5)

v = f − Bx
= f − BN −1t
= f − BN −1BT Wf
= ( I − BN −1BT W ) f (5.6)

ˆl = l + v
= l + f − Bx
= d − Bx
= ( −B ) x + d (5.7)

Applying the Law of Propagation of Variances to equations (5.4) to (5.7) gives the following
cofactor matrices

© 2005, R.E. Deakin Notes on Least Squares (2005) 5–2

RMIT University Geospatial Science

Qtt = ( BT W ) Q ff ( BT W ) = N
T
(5.8)

Q xx = ( N −1 ) Qtt ( N −1 ) = N −1
T
(5.9)

Qvv = ( I − BN −1BT W ) Q ff ( I − BN −1BT W )

T

= Q − BN −1BT (5.10)

Qllˆˆ = ( − B ) Q ( − B )
T

= BN −1BT
= Q − Qvv (5.11)

Variance-covariance matrices for t, x, v and l̂ are obtained by multiplying the cofactor matrix
by the variance factor σ 02 .

5.2. Calculation of the quadratic form v T Wv

The a priori estimate of the variance factor may be computed from

vT Wv
σˆ 02 = (5.12)
r

where v T Wv is the quadratic form, and

r = n−u is the degrees of freedom where n is the number of observations and u
is the number of unknown parameters. r is also known as the number
of redundancies.

A derivation of equation (5.12) is given below. The quadratic form v T Wv may be computed
in the following manner.

Remembering, for the method of indirect observations, the following matrix equations

N = BT WB
t = BT Wf
x = N −1t
v = f − Bx

then

© 2005, R.E. Deakin Notes on Least Squares (2005) 5–3

RMIT University Geospatial Science

v T Wv = (f − Bx )T W(f − Bx )
= (f T − x T BT ) W (f − Bx )
= (f T W − x T BT W )(f − Bx )
= f T Wf − f T WBx − xT BT Wf + x T BT WBx
= f T Wf − 2f T WBx + x T BT WBx
= f T Wf − 2t T x + x T Nx
= f T Wf − 2x T t + x T t

and

v T Wv = f T Wf − xT t (5.13)

5.3. Calculation of the Estimate of the Variance Factor σˆ 02

The variance-covariance matrices of residuals Σ vv , adjusted observations Σ llˆˆ and computed

parameters Σ xx are calculated from the general relationship

Σ = σ 02Q (5.14)

Cofactor matrices Qv v , Q x x and Qlˆ lˆ are computed from equations (5.9) to (5.11) and so it

remains to determine an estimate of the variance factor σˆ 02 .

The development of a matrix expression for computing σˆ 02 is set out below and follows
Mikhail (1976, pp.285-288). Some preliminary relationships will be useful.

1. If A is an (n,n) square matrix, the sum of its diagonal elements is a scalar quantity
called the trace of A and denoted by tr ( A ) The following relationships are useful

tr ( A + B ) = tr ( A ) + tr ( B ) for A and B of same order (5.15)

tr ( A T ) = tr ( A ) (5.16)

and for the quadratic form xT A x where A is symmetric

xT A x = tr ( x x T A ) (5.17)

© 2005, R.E. Deakin Notes on Least Squares (2005) 5–4

RMIT University Geospatial Science

2. The variance-covariance matrix Σ xx given by equation (3.21) can be expressed in the

following manner, remembering that x is a vector of random variables and m x is a

vector of means.

Σ xx = E {( x − m x )( x − m x )T }
= E {( x − m x )( xT − mTx )}
= E {xx T − xmTx − m x x T + m x mTx }
= E {xx T } − E {xmTx } − E {m x xT } + E {m x mTx }
= E {xx T } − E {x} mTx − m x E {x T } + m x mTx

Now from equation (3.18) m x = E {x} hence

Σ xx = E {xx T } − m x mTx − m x mTx + m x mTx

= E {xxT } − m x mTx (5.18)

or E {xx T } = Σ xx + m x mTx (5.19)

3. The expected value of the residuals is zero, i.e.,

E {v} = mv = 0 (5.20)

4. By definition (see Chapter 2) the weight matrix W, the cofactor matrix Q and the
variance-covariance matrix Σ are related by

W = Q−1 = σ 02Σ −1 (5.21)

Now, for the least squares adjustment of indirect observations the following relationships are
recalled

v + Bx = f , N = BT WB, t = BT Wf
Q ff = Q, Qtt = N, Q xx = N −1

Bearing in mind equation (5.21), the following relationships may be introduced

1
Σ −1 = W, M = BT Σ −1B
σ 02

and from these follow

© 2005, R.E. Deakin Notes on Least Squares (2005) 5–5

RMIT University Geospatial Science

Σ ff = Σ, Σ tt = σ 04M, Σ xx = M −1

In addition, the expectation of the vector f is the mean m f and so we may write

m f = E {f } = E {v + Bx} = E {v} + B E {x}

Now since E {x} = m x and E {v} = 0

m f = Bm x (5.22)

Now the quadratic form

v T Wv = σ 02 ( vT Σ −1v ) (5.23)

and from equation (5.13)

v T Wv = f T Wf − xT t
= f T Wf − xT Nx

Using the relationships above

v T Σ −1v = f T Σ −1f − xT Mx

Now the expected value of this quadratic form is

E {v T Σ −1v} = E {f T Σ −1f − xT Mx}

= E {f T Σ −1f } − E {xT Mx}

Recognising that the terms on the right-hand-side are both quadratic forms, equation (5.17)
can be used to give

{ } { }
E {v T Σ −1v} = E tr ( ff T Σ −1 ) − E tr ( xx T M )

= tr ( E {ff Σ }) − tr ( E {xx M})

T −1 T

= tr ( E {ff }Σ ) − tr ( E {xx }M )
T −1 T

Now using equation (5.19)

( ) (
E {v T Σ −1v} = tr ⎡⎣Σ ff + m f mTf ⎤⎦ Σ −1 − tr ⎡⎣Σ xx + m x mTx ⎤⎦ M )
= tr ( I nn + m f mTf Σ −1 ) − tr ( I uu + m x mTx M )
= tr ( I nn − I uu ) − tr ( m f mTf Σ −1 + m x mTx M )
= ( n − u ) − mTf Σ −1m f + mTx Mm x

© 2005, R.E. Deakin Notes on Least Squares (2005) 5–6

RMIT University Geospatial Science

From equation (5.22) m f = Bm x hence using the rule for matrix transpose

mTf = ( Bm x ) = mTx BT , then

T

E {v T Σ −1v} = ( n − u ) − mTx BT Σ −1Bm x + mTx Mm x

= ( n − u ) − mTx Mm x + mTx Mm x
= (n − u)

Thus according to equation (5.23) and the expression above

E {v T Wv} = σ 02 E {vT Σ −1v}

= σ 02 ( n − u )

from which follows

E {v T Wv}
σ 02 =
n−u

Consequently, an unbiased estimate of the variance factor σˆ 02 can be computed from

v T Wv vT Wv
σˆ 02 = = (5.24)
n−u r

r = n − u is the number of redundancies in the adjustment and is known as the degrees of

freedom

Using equation (5.13) an unbiased estimate of the variance factor σˆ 02 can be computed from

f T Wf − xT t
σˆ 02 = (5.25)
r

© 2005, R.E. Deakin Notes on Least Squares (2005) 5–7

RMIT University Geospatial Science

6. LEAST SQUARES ADJUSTMENT OF OBSERVATIONS ONLY

In Chapter 2 the least squares technique of adjustment of indirect observations was introduced
using the example of fitting a straight line through a series of data points. The "observations"
in this example were the x,y coordinates that were indirect measurements of the unknown
parameters m and c, the slope and intercept of the line on the y-axis respectively. Subsequent
examples of curve fitting (parabola and ellipse) demonstrated this technique and in Chapter 4
adjustment of indirect observations was applied to a level network. An alternative to this
technique, known as least squares adjustment of observations only, will be introduced in this
chapter using the level network example of Chapter 4.

6.1. Adjustment of a Level Network using Least Squares Adjustment of

Observations Only

Figure 6.1 shows a diagram of a level network of height differences observed between the
fixed stations A (RL 102.440 m) and B (RL 104.565 m) and "floating" stations X, Y and Z
whose Reduced Levels (RL's) are unknown. The arrows on the diagram indicate the direction
of rise. The Table of Height differences shows the height difference for each line of the
network and the distance (in kilometers) of each level run. The height differences can be
considered as independent (uncorrelated) and of unequal precision, where the weights of the
height differences are defined as being inversely proportional to the distances in kilometres
(see Chapter 3, Section 3.5.2)

X
1 •
Line Height Diff Dist (km) 2
1 6.345 1.7 6
2 4.235 2.5
3 3.060 1.0
A ⊗
4 0.920 3.8 5
• 7
⊗B
5 3.895 1.7 Y
6 2.410 1.2
7 4.820 1.5 3
•
4 Z

Figure 6.1 Level network

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–1

RMIT University Geospatial Science

The measured height differences do not accord with the simple principle that they should sum
to zero around a "closed loop", i.e., there are misclosures. For example:
in the loop AXYA ΔH1 − ΔH 6 − ΔH 5 = +0.040 m

in the loop XBZYX −ΔH 2 − ΔH 3 + ΔH 7 + ΔH 6 = −0.065 m

in the loop AYZA ΔH 5 − ΔH 7 + ΔH 4 = −0.005 m

Hence it is required to determine the adjusted height differences (that will sum to zero) and
the RL's of X, Y and Z.

There are n = 7 observations (the measured height differences) and a minimum of n0 = 3

observations are required to fix the RL's of X, Y and Z. Hence there are r = n − n0 = 4
redundant measurements, which equals the number of independent condition equations.
Denoting the observed height differences as l1 , l2 etc , residuals as v1 , v2 etc and the RL's of
A and B as A and B, these condition equations are

( l1 + v1 ) − ( l6 + v6 ) − ( l5 + v5 ) = 0
− ( l2 + v2 ) − ( l3 + v3 ) + ( l7 + v7 ) + ( l6 + v6 ) = 0
(6.1)
( l5 + v5 ) − ( l7 + v7 ) + ( l4 + v4 ) = 0
( l1 + v1 ) − ( l2 + v2 ) = B − A
The first 3 equations of (6.1) are the loop closure conditions and the last equation is a
condition linking the RL's of A and B.

Since the measurements are of unequal precision, there is an associated weight wk with each
observation and the application of the least squares principle calls for the minimization of the
least squares function ϕ as

n
ϕ = the sum of the weighted squared residuals = ∑ wk vk2 (6.2)
k =1

or ϕ = w1v12 + w2v22 + w3v32 + + w7v72

Considering equation (6.1) it is clear that separate expressions for residuals cannot be derived
and substituted into ϕ , as was possible in the technique for adjustment of indirect
observations (see Chapter 2). Therefore another approach is needed to ensure that ϕ is a

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–2

RMIT University Geospatial Science

minimum as well as satisfying equations (6.1). This is accomplished by using a method of

function minimization developed by Lagrange 1 and set out in the following manner.

(i) Gather the terms in equations (6.1) together

v1 − v6 − v5 = 0 − ( l1 − l6 − l5 ) = f1
−v2 − v3 + v7 + v6 = 0 − ( −l2 − l3 + l7 + l6 ) = f 2
(6.3)
v5 − v7 + v4 = 0 − ( l5 − l7 + l4 ) = f3
v1 − v2 = ( B − A) − ( l1 − l2 ) = f4

(ii) Rewrite equations (6.3) in normal form (zero on the right-hand-side)

v1 − v6 − v5 − f1 = 0
− v2 − v3 + v7 + v6 − f 2 = 0
(6.4)
v5 − v7 + v4 − f 3 = 0
v1 − v2 − f 4 = 0

(iii) Now form an augmented function ϕ ′ of the form

ϕ ′ = w1v12 + w2v22 + w3v32 + + w7 v72

−2k1 ( v1 − v6 − v5 − f1 ) − 2k2 ( − v2 − v3 + v7 + v6 − f 2 ) (6.5)
−2k3 ( v5 − v7 + v4 − f 3 ) − 2k4 ( v1 − v2 − f 4 )

where k1 , k2 , k3 and k4 are Lagrange multipliers and there are as many multipliers
as there are conditions. The introduction of −2 preceding each multiplier is for
convenience only. Inspection of equations (6.5), (6.4) and (6.2) show that ϕ and
ϕ ′ are equal since the additional terms in ϕ ′ equate to zero.

(iv) The unknowns in equation (6.5) are the residuals v1 , v2 , … , v7 and the Lagrange

multipliers k1 , k2 , k3 and k4 , and so for ϕ ′ to be a minimum, the partial

derivatives of ϕ ′ with respect to each of the unknowns must be zero. Setting the

1
Joseph Louis LAGRANGE (1713-1813), a great French mathematician whose major work was in the calculus of variation,
celestial and general mechanics, differential equations and algebra. Lagrange spent 20 years of his life in Prussia and then
returned to Paris where his masterpiece, Mécanique analytique, published in 1788, formalized much of Newton's work on
calculus.

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–3

RMIT University Geospatial Science

partial derivatives of ϕ ′ with respect to the residuals leads to the following

equations

∂ϕ ′ 1
= 2 w1v1 − 2k1 − 2k4 = 0 or v1 = ( k1 + k4 )
∂v1 w1
∂ϕ ′ 1
= 2 w2 v2 + 2k2 + 2k4 = 0 or v2 = ( − k2 − k4 )
∂v2 w2
∂ϕ ′ 1
= 2 w3 v3 + 2k2 =0 or v3 = ( − k2 )
∂v3 w3
∂ϕ ′ 1
= 2 w4 v4 − 2k3 =0 or v4 = k3
∂v4 w4
∂ϕ ′ 1
= 2 w5 v5 + 2k1 − 2k3 = 0 or v5 = ( −k1 + k3 )
∂v5 w5
∂ϕ ′ 1
= 2 w6 v6 + 2k1 − 2k2 = 0 or v6 = ( −k1 + k2 )
∂v6 w6
∂ϕ ′ 1
= 2 w7 v7 − 2k2 + 2k3 = 0 or v7 = ( k 2 − k3 ) (6.6)
∂v7 w7

and when ϕ ′ is differentiated with respect to the Lagrange multipliers and equated
to zero

∂ϕ ′
= −2 ( v1 − v6 − v5 − f1 ) =0 or v1 − v6 − v5 = f1
∂k1
∂ϕ ′
= −2 ( −v2 − v3 + v7 + v6 − f 2 ) = 0 or − v2 − v3 + v7 + v6 = f 2
∂k2
∂ϕ ′
= −2 ( v5 − v7 + v4 − f 3 ) =0 or v5 − v7 + v4 = f 3
∂k3
∂ϕ ′
= −2 ( v1 − v2 − f 4 ) =0 or v1 − v2 = f 4 (6.7)
∂k4

the original condition equations (6.4) result. This demonstrates that the
introduction of Lagrange multipliers ensures that the conditions will be satisfied
when ϕ ′ is minimized.

(v) Now, substituting equations (6.6) into (6.7) gives four normal equations

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–4

RMIT University Geospatial Science

⎛ 1 1 1 ⎞ 1 1 1
⎜ + + ⎟ k1 − k2 − k3 + k4 = f1
⎝ w1 w6 w5 ⎠ w6 w5 w1
1 ⎛ 1 1 1 1 ⎞ 1 1
− k1 + ⎜ + + + ⎟ k2 − k3 + k4 = f 2
w6 ⎝ w2 w3 w6 w7 ⎠ w7 w2
1 1 ⎛ 1 1 1 ⎞
− k1 − k2 + ⎜ + + ⎟ k3 = f 3
w5 w7 ⎝ w4 w5 w7 ⎠
1 1 ⎛ 1 1 ⎞
k1 + k2 + ⎜ + ⎟ k4 = f 4 (6.8)
w1 w2 ⎝ w1 w2 ⎠

Equations (6.8) can be solved to give the Lagrange multipliers k1 , k2 , k3 and k4 ,

which can be substituted back into equations (6.6) to give the residuals
1
v1 , v2 , … , v7 . Note that the coefficient terms in equations (6.8) are known as
wk
weight reciprocals and in the case of levelling are simply the distances of the level
runs in kilometres.

Using the data from Figure 6.1 the weight reciprocals are the distances (in kilometres)

1
= {1.7 2.5 1 3.8 1.7 1.2 1.5}
wk

the numeric terms f are given by equations (6.3)

f1 = − ( l1 − l6 − l5 ) = −0.040 m
f 2 = − ( −l2 − l3 + l7 + l6 ) = 0.065 m
f3 = − ( l5 − l7 + l4 ) = 0.005 m
f4 = ( B − A) − ( l1 − l2 ) = 0.015 m

and the normal equations (in matrix form) are

⎡ 4.6 −1.2 −1.7 1.7 ⎤ ⎡ k1 ⎤ ⎡ −0.040⎤

⎢ −1.2 6.2 −1.5 2.5⎥ ⎢ k2 ⎥ ⎢ 0.065⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥ (6.9)
⎢ −1.7 −1.5 7.0 0⎥ ⎢ k3 ⎥ ⎢ 0.005⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎣ 1.7 2.5 0 4.2 ⎦ ⎣ k4 ⎦ ⎣ 0.015⎦

The solution of equations (6.9) for the Lagrange multipliers gives

k1 = −0.005700, k2 = 0.009671, k3 = 0.001402, k4 = 0.000122

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–5

RMIT University Geospatial Science

1
Substituting these values ( k1 , k2 , k3 and k4 ) together with the weight reciprocals into
wk

equations (6.6) gives the residuals v1 , v2 , … , v7 . The height differences, residuals and the
adjusted height differences (observed value + residual) of the level network are shown below.

Line Observed ΔH Residual v Adjusted ΔH

1 6.345 -0.0095 6.336
2 4.235 -0.0245 4.211
3 3.060 -0.0097 3.050
4 0.920 0.0053 0.925
5 3.895 0.0121 3.907
6 2.410 0.0184 2.428
7 4.820 0.0124 4.832

These are identical results to those obtained by least squares adjustment of indirect
observations set out in Chapter 4.

6.2. Some Comments on the Two Applications of the Method of Least Squares

1. The method of least squares has been applied in two examples:

(a) determining the parameters of a "line of best fit" through a number of data points
(see Chapter 2) and

(b) determining the adjusted height differences in a level network.

2. Consider the first example: the line of best fit.

• A mathematical model (equation) was established linking observations, residuals

(corrections) and unknown parameters.

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–6

RMIT University Geospatial Science

• For n observations, there is a minimum number n0 required to determine the u

unknown parameters. In this case n0 = u and the number of redundant
observations is r = n − n0
• An equation was written for each observation, i.e., there were n observation
equations. The observation equations were recast as residual equations.
• Since there were n equations in u unknowns (n 〉 u) there is no unique solution and
the least squares principle was used to determine the u normal equations from
which the best estimates of the u unknown parameters were calculated.

This technique of least squares "adjustment" is known by various names, some of which are

parametric least squares,

least squares adjustment by observation equations,
least squares adjustment by residual equations, and
least squares adjustment of indirect observations.

The last of these is perhaps the most explicit since each observation is in fact an indirect
measurement of the unknown parameters. Least squares adjustment of indirect
observations is the name adopted for this technique by Mikhail (1976) and Mikhail &
Gracie (1981) and will be used in these notes.

3. Consider the second example: the level network.

• A relationship or condition that the observations (and residuals) must satisfy was
established. In this case, the condition to be satisfied was that observed height
differences (plus some unknown corrections or residuals) should sum to zero
around a closed level loop.
• The minimum number of observations n0 required to fix the heights of X, Y and Z
and satisfy the condition between the fixed points A and B was determined giving
the number of independent condition equations equal to the number of redundant
observations r = n − n0 .
• There were r equations in n unknown residuals, and since r = n − n0 was less than
n, there was no unique solution for the residuals. The least squares principle was

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–7

RMIT University Geospatial Science

used to determine a set of r normal equations, which were solved for r Lagrange
multipliers which in turn, were used to obtain the n residuals.
• The residuals were added to the observations to obtain the adjusted observations
which were then used to determine the heights of points X, Y and Z.

This technique of least squares "adjustment" is known by various names, two of which are

least squares adjustment by condition equations, and

least squares adjustment of observations only.

The second of these is the more explicit since equations involve only observations. No
parameters are used. Least squares adjustment of observations only is the name adopted for
this technique by Mikhail (1976) and Mikhail & Gracie (1981) and will be used in these
notes.

It should be noted that in practice, the method of adjustment of observations only is seldom
employed, owing to the difficulty of determining the independent condition equations
required as a starting point. This contrasts with the relative ease of the technique of
adjustment of indirect observations, where every observation yields an equation of fixed form.
Computer solutions of least squares problems almost invariably use the technique of
adjustment of indirect observations.

6.2.1. A Note on Independent Condition Equations.

C
Consider the level network shown in Figure 6.2. The 1 • 4
RL of A is known and the RL's of B, C and D are to
be determined from the observed height differences.
3
•B
A ⊗
The arrows on the diagram indicate the direction of 6
rise. 5
2
•
D
Figure 6.2 Level network

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–8

RMIT University Geospatial Science

There are n = 6 observations with a minimum of n0 = 3 required to fix the RL's of B, C and
D with respect to A. Hence there are r = n − n0 = 3 redundant measurements, which equal
the number of independent condition equations. Omitting the residuals, these equations are

l1 + l3 − l2 = 0
l4 − l5 − l3 = 0 (6.10)
l1 + l4 − l6 = 0

Alternatively, here is another set of independent condition equations

l1 + l3 + l5 − l6 = 0
l1 + l4 − l6 = 0 (6.11)
l1 + l4 − l5 − l2 = 0

But, here is a further set of condition equations, which are not independent

l1 + l3 − l2 = 0
l4 − l5 − l3 = 0 (6.12)
l1 + l4 − l5 − l2 = 0

where the third equation of (6.12) is obtained by adding the first two.

Care needs to be taken in determining independent equations and it is easy to see that this
could become quite difficult as the complexity of the adjustment problem increases.

6.3. Matrix Methods and Least Squares Adjustment of Observations Only

Matrix methods may be used to develop standard equations and solutions for this technique of
least squares adjustment.

Consider again the example of the level net shown in Figure 6.1. The independent condition
equations, (reflecting the fact that height differences around closed level loops should sum to
zero and the condition between the known RL's of A and B), are

( l1 + v1 ) − ( l6 + v6 ) − ( l5 + v5 ) = 0
− ( l2 + v2 ) − ( l3 + v3 ) + ( l7 + v7 ) + ( l6 + v6 ) = 0
(6.13)
( l5 + v5 ) − ( l7 + v7 ) + ( l4 + v4 ) = 0
( l1 + v1 ) − ( l2 + v2 ) = B − A

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–9

RMIT University Geospatial Science

These equations could be expressed in matrix form as

⎡ l1 + v1 ⎤
⎢l + v ⎥
−1 − 1 0 ⎤ ⎢
2 2⎥
⎡1 0 0 0 ⎡ 0 ⎤
⎢ 0 −1 − 1 ⎥ ⎢ l3 + v3⎥ ⎢ 0 ⎥
0 0 1 1 ⎢
⎢ ⎥ l4 + v 4 ⎥ = ⎢ ⎥ (6.14)
⎢0 0 0 1 1 0 −1⎥ ⎢ ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ l5 + v5 ⎥ ⎢ ⎥
⎣ 1 −1 0 0 0 0 0 ⎦⎢ ⎥ ⎣ B − A⎦
⎢ l6 + v6 ⎥
⎢⎣ l7 + v7 ⎥⎦

or Al + Av = d (6.15)

which can be written as

Av = f (6.16)

where f = d − Al (6.17)

and
n is the number of measurements or observations,
n0 is the minimum number of observations required,
r = n − n0 is the number of redundant observations (equal to the number of
condition equations,
v is an (n,1) vector of residuals,
l is the (n,1) vector of observations,
A is an (r,n) matrix of coefficients,
f is an (r,1) vector of numeric terms derived from the observations,
d is an (r,1) vector of constants. Note that in many least squares
problems the vector d is zero.

Now if each observation has an a priori estimate of its variance then the (n,n) weight matrix
of the observations W is known and the least squares function ϕ is
n
ϕ = the sum of the weighted squared residuals = ∑ wk vk2
k =1

In matrix form, the least squares function is expressed as

ϕ = vT Wv (6.18)

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–10

RMIT University Geospatial Science

Now ϕ is the function to be minimised but with the constraints imposed by the condition
equations (6.16). This is achieved by adding an (r,1) vector of Lagrange multipliers k and
forming a new function ϕ ′ .

ϕ ′ = v T Wv − 2k T ( Av − f ) (6.19)

Note that the second term of (6.19) equals zero, since A v − f = 0 .

Minimising ϕ ′ is achieved by differentiating with respect to the unknowns, v and k and

equating these differentials to zero

∂ϕ ′
= −2 v T A T + 2 f T = 0 T (6.20)
∂k

∂ϕ ′
= 2 v T W − 2 k T A = 0T (6.21)
∂v

Dividing by two, re-arranging and transposing equations (6.20) and (6.21) gives

Av = f (6.22)

Wv − A T k = 0 (6.23)

Note that equations (6.22) are the original condition equations and also that W = W T due to
symmetry.

From (6.23), the (n,1) vector of residuals v is

v = W −1A T k = QA T k (6.24)

which, when substituted into (6.22), gives

A ( QA T k ) = ( AQA T ) k = f (6.25)

The matrix AQA T is symmetric and of order (r,r) and equations (6.25) are often termed the
normal equations. The solution of the (r,1) vector of Lagrange multipliers k is

k = ( AQA T ) f
−1
(6.26)

Now the term AQA T in equations (6.25) and (6.26) can be "simplified" if an equivalent set of
observations le is considered, i.e.,

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–11

RMIT University Geospatial Science

le = A l (6.27)

Applying the general law of propagation of variances (cofactors) to (6.27) gives

Qe = AQll A T = AQA T (6.28)

We = Qe−1 = ( AQA T )
−1
and (6.29)

Substituting (6.29) into (6.26) gives another expression for k

k = Qe−1f = Wef (6.30)

After computing k from either (6.26) or (6.30) the residuals v are computed from (6.24) and
the vector of adjusted observations l is given by

l = l+v (6.31)

This is the standard matrix solution for least squares adjustment of observations only.

6.4. Propagation of Variances for Least Squares Adjustment of Observations

Only

In this technique of least squares adjustment, the condition equations in matrix form are

Av = f (6.32)

with f = d − Al (6.33)

Similarly to Chapter 5, equation (6.33) can be expressed in a form similar to equation (3.23)
and the general law of propagation of variances applied to give the cofactor matrix of the
numeric terms f.

f = − Al + d

and Q f f = ( − A )Q( − A )T = AQA T = Qe (6.34)

Thus the cofactor matrix of f is also the cofactor matrix of an equivalent set of observations.

The solution "steps" in the least squares adjustment of observations only are set out above and
restated as

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–12

RMIT University Geospatial Science

Qe = A Q AT
We = Q e−1
k = We f
v = Q AT k
l = l+v

Applying the law of propagation of variances (remembering that cofactor and weight matrices
are symmetric) gives the following cofactor matrices

Qkk = ( We )Q f f ( We )T = We (6.35)

Qv v = (QA T )Qk k (QA T )T = QA T We AQ (6.36)

and

ˆl = l + v
= l + QA T k
= l + QA T Wef
= l + QA T We ( d − Al )

from which follows

ˆl = ( I − QA T W A ) l + QA T W d (6.37)
e e

Applying the law of propagation of variances to (6.37) gives

Qlˆ lˆ = ( I − QA T We A ) Q ( I − QA T We A )
T

which reduces to

Qlˆ lˆ = Q − QA T We AQ = Q − Qv v (6.38)

Variance-covariance matrices for k, v and l are obtained by multiplying the cofactor matrix
by the variance factor σ 20 - see equation (2.32).

The a priori estimate of the variance factor may be computed from

v T Wv
σ 20 = (6.39)
r

where

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–13

RMIT University Geospatial Science

v T Wv is the quadratic form, and

r is the degrees of freedom.

A derivation of equation (6.39) is given in Chapter 5. The quadratic form v T W v may be

computed in the following manner.

Remembering, for the method of observations only, the following matrix equations

Qe = AQA T
We = Qe−1
k = Wef
v = QA T k

then

v T Wv = ( QA T k ) W ( QA T k )
T

= k T AQWQA T k
= k T AQA T k
= k T Qek
= k T We−1k
= f T We We−1k

and

v T Wv = f T k (6.40)

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–14

RMIT University Geospatial Science

6.5. Adjustment of a Single Closed Traverse using the method of Least

Squares Adjustment of Observations Only

The basic component of many surveys is a traverse whose bearings have been determined by
theodolite or total station observations and distances measured by EDM. If careful
observations are made with well maintained equipment, the measurements are usually free of
systematic errors and mistakes and the surveyor is left with small random errors which, in the
case of a closed traverse, reveal themselves as angular and linear misclosures. If the
misclosures are within acceptable limits, it is standard practice to remove the misclosures by
adjusting the original observations to make the traverse a mathematically correct figure. In
this section, only single closed traverses are considered and such traverses may begin and end
at different fixed points or close back on the starting point. Traverse networks, consisting of
two or more single traverses with common junction points, are not considered here; such
networks are usually adjusted by a method commonly known as Variation of Coordinates,
based on Least Squares Adjustment of Indirect Observations.

6.5.1. Some single traverse adjustment methods and their deficiencies

A traverse adjustment method should be based on sound mathematical principles related to

the measurement techniques with due allowance made for independence (or dependence) of
those measurements and also allow for differing measurement precisions.

Bowditch's Rule and the Transit Rule, both of which adjust lengths and bearings of traverse
lines and Crandall's method, which adjusts the lengths only of the traverse lines, are three
popular adjustment methods that fail to meet the general guidelines above. Although
Crandall's method, which is explained in detail in later sections, does have mathematical
rigour if it assumed that the bearings of a traverse close and require no further adjustment.

Bowditch's Rule and the Transit Rule for adjusting single traverses are explained below by
applying the rules to adjust a four-sided polygon having an unusually large misclose. The
polygon, shown in Figure 6.3, does not reflect the usual misclosures associated with traverses
using modern surveying equipment.

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–15

RMIT University Geospatial Science

Bowditch's Rule
Nathaniel Bowditch (1773-1838) was an American mathematician and astronomer (see
citation below). In 1808, in response to a prize offered by a correspondent in The Analyst 2 ,
Bowditch put forward a method of adjusting the misclose in a chain and compass survey
(bearings measured by magnetic compass and distances measured by surveyor's chain). His
method of adjustment was simple and became widely used. It is still used today for the
adjustment of a figure prior to the computation of the area, where the area-formula assumes a
closed mathematical figure.

Prior to the advent of programmable calculators and computers, Bowditch's Rule was often
used to adjust traverses that did not close due to the effects of random errors in the
measurement of bearings and distances. Its use was justified entirely by its simplicity and
whilst it had theoretical rigour – if the bearings of traverse lines were independent of each
other, as they are in compass surveys – it is incompatible with modern traversing techniques.
Bowditch's rule cannot take into account different measurement precisions of individual
traverse lines nor can it accommodate complicated networks of connecting traverses.
Nevertheless, due to its long history of use in the surveying profession, its simplicity and its
practical use in the computation of areas of figures that misclose, Bowditch's Rule is still
prominent in surveying textbooks and is a useful adjustment technique.
Bowditch, Nathaniel (b. March 26, 1773, Salem, Mass., U.S. – d. March 16, 1838, Boston, Mass., U.S.), self-
educated American mathematician and astronomer, author of the best book on navigation of his time, and
discoverer of the Bowditch curves, which have important applications in astronomy and physics. Between
1795 and 1799 Bowditch made four lengthy sea voyages, and in 1802 he was put in command of a merchant
vessel. Throughout that period he pursued his interest in mathematics. After investigating the accuracy of The
Practical Navigator, a work by the Englishman J.H. Moore, he produced a revised edition in 1799. His
additions became so numerous that in 1802 he published The New American Practical Navigator, based on
Moore's book, which was adopted by the U.S. Department of the Navy and went through some 60 editions.
Bowditch also wrote many scientific papers, one of which, on the motion of a pendulum swinging
simultaneously about two axes at right angles, described the so-called Bowditch curves (better known as the
Lissajous figures, after the man who later studied them in detail). Bowditch translated from the French and
updated the first four volumes of Pierre-Simon Laplace's monumental work on the gravitation of heavenly
bodies, Traité de mécanique céleste, more than doubling its size with his own commentaries. The resulting
work, Celestial Mechanics, was published in four volumes in 1829-39. Bowditch refused professorships at
several universities. He was president (1804-23) of the Essex Fire and Marine Insurance Company of Salem
and worked as an actuary (1823-38) for the Massachusetts Hospital Life Insurance Company of Boston. From
1829 until his death, he was president of the American Academy of Arts and Sciences. Copyright 1994-1999
Encyclopædia Britannica

2
The Analyst or Mathematical Museum was a journal of theoretical and applied mathematics. In Vol. I, No. II,
1808, Robert Patterson of Philadelphia posed a question on the adjustment of a traverse and offered a prize of
$10 for a solution; the editor Dr Adrian appointed as the judge of submissions. Bowditch's solution was
published in Vol. I, No. IV, 1808, pp. 88-93 (Stoughton, H.W., 1974. 'The first method to adjust a traverse
based on statistical considerations', Surveying and Mapping, June 1974, pp. 145-49).

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–16

RMIT University Geospatial Science

Bowditch's adjustment can best be explained by considering the case of plotting a figure
(using a protractor and scale ruler) given the bearing and distances of the sides.

Consider Figure 6.3, a plot that does not close, of a four-sided figure ABCD. The solid lines
AB, BC, CD and DE are the result of marking point A, plotting the bearing AB and then
scaling the distance AB to fix B. Then, from point B, plotting the bearing and distance BC to
fix C, then from C, plotting the bearing and distance CD to fix D and finally from D, plotting
the bearing and distance DA. However, due to plotting errors, the final line does not meet the
starting point, but instead finishes at E. The distance EA is the linear misclose d, due to
plotting errors, i.e., errors in protracting bearings and scaling distances.

B'

x'

A
misclose
d
E
C'
x C

D'

A'
C' D'
B' d
A B C D E
L

Figure 6.3 Graphical plot of polygon ABCD with misclose d

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–17

RMIT University Geospatial Science

To adjust the figure ABCDE to remove the misclose d the following procedure can be used.

1. Draw lines parallel to the line xx' (the misclose bearing) through points B, C and D.
2. Draw a right-angled triangle AEA'. The base of the triangle is L, equal to the sum of the
lengths of the sides and the height is the linear misclose d.
3. Along the base of the triangle, mark in proportion to the total length L, the distances AB,
BC and CD. These will be the points B, C and D.
4. Draw vertical lines from B, C and D intersecting the hypotenuse of the triangle at B', C'
and D'. These distances are then marked off along the parallel lines of the main figure.
5. The adjusted figure is AB'C'D'A.

This adjustment is a graphical demonstration of Bowditch's Rule; i.e., the linear misclose d is
apportioned to individual sides in the ratio of the length of the side to the total length of all the
sides in the direction of the misclose bearing.

Bowditch's Rule as it is normally applied to the adjustment of traverses can be deduced by

again considering Figure 6.3. The linear misclose d has easting and northing components
ΔEm and ΔN m , the subscript m referring to the misclose. The distances BB', CC' and DD'

each have easting and northing components, say dEB , dN B , dEC , dN C and dED , dN D , the

east misclose dEm = dEB + dEC + dED and the north misclose dN m = dN B + dN C + dN D .

Thus, we may express Bowditch's Rule for calculating adjustments dEk , dN k to individual

easting and northing components ΔEk , ΔN k of line k of a traverse whose total length is L as

⎛ dE ⎞
dEk = distk ⎜ m ⎟
⎝ L ⎠
(6.41)
⎛ dN ⎞
dN k = distk ⎜ m ⎟
⎝ L ⎠

As an example of a Bowditch adjustment, Table 6.1 shows the bearings and distances of the
polygon in Figure 6.3.

( −3.173) + ( −8.181) = 8.775 and the

2 2
The linear misclose, which is quite large, is d =

length L, equal to the sum of the four sides, is L = 51.53 + 53.86 + 36.31 + 54.71 = 196.41

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–18

RMIT University Geospatial Science

The corrections to the easting and northing components of the line CD are

3.173
dE = 36.31 × = 0.587
196.41
8.181
dN = 36.31 × = 1.512
196.41

Note: (i) Easting and northing misclosures dEm and dN m used in equations (6.41) have
opposite signs to the misclosures in the tabulation,
(ii) The sums of the corrections are equal and of opposite sign to the misclosures
and
(iii) The sums of the adjusted easting and northing components are zero.

adjusted
components corrections
components
Line Bearing Dist ΔE ΔN dN dN ΔE ΔN
AB 52º 31' 51.53 40.891 31.358 0.832 2.146 41.723 33.504
BC 152º 21' 53.86 24.995 -47.709 0.870 2.243 25.865 -45.466
CD 225º 30' 36.31 -25.898 -25.450 0.587 1.512 -25.311 -23.938
DA 307º 55' 54.71 -43.161 33.620 0.884 2.280 -42.277 35.900
misclose -3.173 -8.181 3.173 8.181 0.000 0.000

Table 6.1. Bowditch Rule adjustment of polygon ABCD

Transit Rule
The Transit Rule has no theoretical basis related to surveying instruments or measuring
techniques. Its only justification is its mathematical simplicity, which is no longer a valid
argument for the method in this day of pocket computers. The Transit Rule for calculating
adjustments dEk , dN k to individual easting and northing components ΔEk , ΔN k of line k of a

traverse whose east and north misclosures are dEm and dN m is

⎛ ⎞ ⎛ ⎞
⎜ dE ⎟ ⎜ dN ⎟
dEk = ΔEk ⎜ n m ⎟ dN k = ΔN k ⎜ n m ⎟ (6.42)
⎜ ⎟ ⎜ ⎟
⎜ ∑ ΔE j ⎟ ⎜ ∑ ΔN j ⎟
⎝ j =1 ⎠ ⎝ j =1 ⎠

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–19

RMIT University Geospatial Science

n
ΔEk is the absolute value of the east component of the kth traverse leg and ∑ ΔE
j =1
j is the

sum of the absolute values of the east components of the traverse legs and similarly for ΔN k
n
and ∑ ΔN
j =1
j .

As an example of a Transit Rule adjustment, Table 6.2 shows the bearings and distances of
the polygon in Figure 6.3. The east and north misclosures are dEm = 3.173 and dN m = 8.181 ,
and the sums of the absolute values of the east and north components of the traverse legs are
n n

∑ ΔE j = 134.945 and
j =1
∑ ΔN
j =1
j = 138.137

The corrections to the easting and northing components of the line CD are

3.173
dE = 25.898 × = 0.587
134.945
8.181
dN = 25.450 × = 1.512
138.137

Note: (i) Easting and northing misclosures dEm and dN m used in equations (6.42) have
opposite signs to the misclosures in the tabulation,
(ii) The sums of the corrections are equal and of opposite sign to the misclosures
and
(iii) The sums of the adjusted easting and northing components are zero.

adjusted
components corrections
components
Line Bearing Dist ΔE ΔN dN dN ΔE ΔN
AB 52º 31' 51.53 40.891 31.358 0.961 1.857 41.852 33.215
BC 152º 21' 53.86 24.995 -47.709 0.588 2.826 25.583 -44.883
CD 225º 30' 36.31 -25.898 -25.450 0.609 1.507 -25.289 -23.943
DA 307º 55' 54.71 -43.161 33.620 1.015 1.991 -42.146 35.611
misclose -3.173 -8.181 3.173 8.181 0.000 0.000

Table 6.2 Transit Rule adjustment of polygon ABCD

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–20

RMIT University Geospatial Science

6.5.2. Crandall's method. A semi-rigorous single traverse adjustment method

Suppose that the angles of a traverse – either beginning and ending at the same point or
between two known points with starting and closing known bearings – have been adjusted so
that the traverse has a perfect angular closure and the resulting bearings are considered as
correct, or adjusted. We call this a closed traverse. A mathematical closure, using the
adjusted bearings and measured distances, will in all probability, reveal a linear misclose, i.e.,
the sums of the east and north components of the traverse legs will differ from zero (in the
case of a traverse beginning and ending at the same point) or certain known values (in the
case of a traverse between known points). Crandall's method, which employs the least
squares principle, can be used to compute corrections to the measured distances to make the
traverse close mathematically. The method was first set out in the textbook Geodesy and
Least Squares by Charles L. Crandall, Professor of Railroad Engineering and Geodesy,
Cornell University, Ithaca, New York, U.S.A. and published by John Wiley & Sons, New
York, 1906.

N
φ2
ΔE1 ΔE 2

2
φ1 s2 2
s1

1 3 φ3

s3

Figure 6.4 Schematic traverse diagram

Figure 6.4 shows a schematic diagram of a traverse of k = 1, 2, … , n legs where φk , sk are the
adjusted bearing and measured distance respectively of the kth leg. The east and north

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–21

RMIT University Geospatial Science

components of each traverse leg are ΔEk = sk sin φk and ΔN k = sk cos φk respectively. If the

adjusted distance of the kth traverse leg is ( sk + vk ) where vk is the residual (a small unknown

correction) then the two conditions that must be fulfilled by the adjusted bearings and
adjusted distances in a closed traverse are

( s1 + v1 ) sin φ1 + ( s2 + v2 ) sin φ2 + + ( sn + vn ) sin φn = DE

(6.43)
( s1 + v1 ) cos φ1 + ( s2 + v2 ) cos φ2 + + ( sn + vn ) cos φn = DN

where DE = EEND − ESTART and DN = N END − N START are the east and north coordinate
differences respectively between the terminal points of the traverse. Note that in a traverse
beginning and ending at the same point DE and DN will both be zero.

Expanding equation (6.43) gives

v1 sin φ1 + v2 sin φ2 + + vn sin φn + S E = DE

(6.44)
v1 cos φ1 + v2 cos φ2 + + vn cos φn + S N = DN

n
S E = s1 sin φ1 + s2 sin φ2 + + sn sin φn = ∑ ΔE k
k =1
where n
(6.45)
S N = s1 cos φ1 + s2 cos φ2 + + sn cos φn = ∑ ΔN k
k =1

S E , S N are the sums of the east and north components, ΔEk , ΔN k respectively, of the

k = 1, 2, … , n traverse legs.

Equations (6.44) can be expressed in matrix form as

⎡ v1 ⎤
⎢v ⎥
⎡ sin φ1 sin φ2 sin φ3 sin φn ⎤ ⎢ 2 ⎥ ⎡ DE − S E ⎤
⎢ cos φ ⎢v ⎥ = (6.46)
⎣ 1 cos φ2 cos φ3 cos φn ⎥⎦ ⎢ 3 ⎥ ⎢⎣ DN − S N ⎥⎦
⎢ ⎥
⎢⎣ vn ⎥⎦

or Av = f

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–22

RMIT University Geospatial Science

The solution for the vector of residuals v is given by equations (6.24) and (6.26) re-stated
again as

v = W −1A T k = QA T k
(6.47)
k = ( AQA T ) f
−1

where k is the vector of Lagrange multipliers, Q = W −1 is the cofactor matrix and W is the
weight matrix, A is a coefficient matrix containing sines and cosines of traverse bearings and
f is a vector containing the negative sums of the east and north components of the traverse
legs.

In Crandall's method, weights are considered as inversely proportional to the measured

distances and the measured distances are considered to be independent. Hence the weight
matrix W is diagonal

⎡ w1 0 0 0 ⎤ ⎡1 s1 0 0 0 ⎤
⎢0 w2 0 0 ⎥ ⎢ 0 1 s2 0 0 ⎥
⎢ ⎥ ⎢ ⎥
W=⎢0 0 w3 ⎥=⎢ 0 0 1 s3 ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢⎣ 0 wn ⎥⎦ ⎢⎣ 0 1 sn ⎥⎦

and since Q = W −1 then

⎡ s1 0 0 0⎤
⎢0 s2 0 0⎥
⎢ ⎥
Q = ⎢0 0 s3 ⎥
⎢ ⎥
⎢ ⎥
⎢⎣ 0 sn ⎥⎦

⎡ s1 0 0 0 ⎤ ⎡ sin φ1 cos φ1 ⎤ ⎡ ΔE1 ΔN 1 ⎤

⎢0 s2 0 0 ⎥ ⎢sin φ2 cos φ2 ⎥ ⎢ ΔE2 ΔN 2 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
and QA T = ⎢ 0 0 s3 ⎥ ⎢ sin φ3 cos φ3 ⎥ = ⎢ ΔE3 ΔN 3 ⎥ (6.48)
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢⎣ 0 sn ⎥⎦ ⎢⎣sin φn cos φn ⎥⎦ ⎢⎣ ΔEn ΔN n ⎥⎦

⎡ n n
⎤
⎢ ∑ sin φk ΔEk ∑ sin φ ΔN k ⎥ k

AQA = ⎢ n ⎥
T k =1 k =1

⎢ n
⎥
⎢ ∑ cos φk ΔEk ∑ cos φk ΔN k ⎥
⎣ k =1 k =1 ⎦

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–23

RMIT University Geospatial Science

ΔEk ΔN k
Now, since sin φk = and cos φk = then AQA T can be written as
sk sk

⎡ n ( ΔEk ) 2 n
ΔEk ΔN k ⎤
⎢∑ ∑ ⎥
AQA =
T ⎢ k =1 sk k =1 sk ⎥ = ⎡a c ⎤ (6.49)
⎢ n ⎢ ⎥
( ΔN k ) ⎥⎥ ⎣ c b⎦
2
⎢ ∑ ΔE k ΔN k
n

⎢⎣ k =1 sk
∑k =1 sk ⎥⎦

1 ⎡ b −c ⎤
( AQA )
−1
and T
= ⎢ −c a ⎥
ab − c 2 ⎣ ⎦

giving the Lagrange multipliers from equations (6.47) as

b ( DE − S E ) − c ( DN − S N )
k1 =
ab − c 2
(6.50)
a ( DN − S N ) − c ( DE − S E )
k2 =
ab − c 2

The residuals v (corrections to the measured distances) are given as

⎡ ΔE1 ΔN 1 ⎤ ⎡ k1ΔE1 + k2 ΔN1 ⎤

⎢ ΔE ΔN 2 ⎥ ⎢ k ΔE + k Δ N ⎥
⎢ 2 ⎥ ⎡ k1 ⎤ ⎢ 1 2 2 2 ⎥
v = QA T k = ⎢ ΔE3 ΔN 3 ⎥ ⎢ ⎥ = ⎢ k1ΔE3 + k2 ΔN 3 ⎥ (6.51)
⎢ ⎥ ⎣ k2 ⎦ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢⎣ ΔEn ΔN n ⎥⎦ ⎢⎣ k1ΔEn + k2 ΔN n ⎥⎦

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–24

RMIT University Geospatial Science

6.5.3. Example of Crandall's method

Figure 6.5 shows a closed traverse between stations A, B, C, D and E. The linear misclose
(bearing and distance) of the traverse is 222º 57' 31" 0.2340 and the components of the
misclose are −0.1594 m east and −0.1712 m north. It is required to adjust the distances
using Crandall's method.

B
96° 49′ C
86.430
7. 7′
12 2° 2
0
47
4

13 2.3
7° 70
16

16
′
A

229 D
295 .600
° 40 0
′ .42 35′
8
9 °
9
22

E
Figure 6.5 Closed traverse between stations ABCDE

The adjusted bearings and measured distances and the traverse leg components are shown in
Table 6.3 below. S E and S N are the summations of east and north components and since this

traverse begins and ends at the same point then DE and DN will both be zero.

Line Bearing Distance traverse leg components

k φk sk ΔEk ΔN k
1 42º 27′ 127.470 86.035437 94.055858
2 96º 49′ 86.430 85.819028 -10.258619
3 137º 16′ 162.370 110.182189 -119.264002
4 229º 35′ 98.420 -74.932042 -63.809760
5 295º 40′ 229.600 -206.945175 99.447747
S E = 0.159438 S N = 0.171224

Table 6.3 Traverse components and sums

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–25

RMIT University Geospatial Science

Table 6.4 shows the functions of the components for each line and their summations.

( ΔE k ) ( ΔN k ) ΔEk ΔN k
2 2

Line
sk sk sk
1 58.069322 69.400678 63.482677
2 85.212376 1.217624 -10.186101
3 74.768213 87.601787 -80.931014
4 57.049491 41.370509 48.581545
5 186.525721 43.074279 -89.635154
a = 461.625124 b = 242.664876 c = -68.688048

Table 6.4 Functions of traverse components

The Lagrange multipliers k1 and k2 are computed from equations (6.50) using a, b, c from

Table 6.4, S E and S N from Table 6.3, and since this traverse begins and ends at the same

point then DE and DN will both be zero.

( ΔE k )
2
5
a =∑ = 461.625124 DE − S E = −0.159438 k1 = −4.7018E − 04
k =1 sk
( ΔN k )
2
5
b =∑ = 242.664876 DN − S N = −0.171224 k2 = −8.3868E − 04
k =1 sk
5
ΔEk ΔN k
c =∑ = −68.688048
k =1 sk

Table 6.5 shows the original traverse data, the residuals and adjusted traverse distances.

Line Bearing Distance Traverse leg components Residual Adjusted

k φk sk ΔEk ΔN k vk = k1 ΔEk + k 2 ΔN k Distance
1 42º 27′ 127.470 86.035437 94.055858 -0.119 127.351
2 96º 49′ 86.430 85.819028 -10.258619 -0.032 86.398
3 137º 16′ 162.370 110.182189 -119.264002 0.048 162.418
4 229º 35′ 98.420 -74.932042 -63.809760 0.089 98.509
5 295º 40′ 229.600 -206.945175 99.447747 0.014 229.614

Table 6.5 Adjusted traverse distances: Crandall's method

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–26

RMIT University Geospatial Science

6.5.4. A rigorous single traverse adjustment method

A traverse is a combination of two basic survey measurements, distances and directions.

Ignoring the physical fact that the same measuring equipment is likely to be used on each leg
of the traverse, distances and directions are independently determined quantities. Bearings φ ,
angles α and coordinates E,N are derived quantities and in general, cannot be considered as
mathematically (or statistically) independent. Restricting the adjustment method to single
traverses, means angles at traverse points, derived from directions at those points, can be
considered as mathematically independent quantities.

Three conditions, expressing the mathematical relationship between traverse measurements

and derived coordinates, may be deduced from Figure 6.6 below, in which P1 and Pn are

"fixed stations" whose east and north coordinates are known and P2 , P3 , P4 ,… , Pn −1 are

"floating stations" whose coordinates are to be determined from the traverse angles α and
distances s. The starting bearing φ0 and the finishing bearing φn are known.

N
Fi

P2
α2
xe

φ2
db

φ1
°
ea
rin

α1 s1
g

ΔE1
φ0

s2
2 αn-1 φn-1

°
P1 φn φn
ΔE 2 sn-2
°P
3
Pn-1
sn-1
αn
db
ear
ing

ixe
F
N
Pn

Figure 6.6 A closed traverse between two fixed stations

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–27

RMIT University Geospatial Science

These three conditions are:

(i) The starting bearing φ0 plus all the measured angles should equal the known

finishing bearing φn ,
(ii) The starting east coordinate plus all the east components of the traverse legs
should equal the known east coordinate at the end point and
(iii) The starting north coordinate plus all the north components of the traverse legs
should equal the known north coordinate at the end point.

These conditions apply to all single traverses whether they start and end at different fixed
points or close back on the starting point and can be expressed mathematically as

φ0 + α1 + α 2 + α 3 + + an = φn
E1 + ΔE1 + ΔE2 + ΔE3 + + ΔEn −1 = En (6.52)
N1 +ΔN1 + ΔN 2 + ΔN 3 + + ΔN n −1 = N n

Equations (6.52) are relationships between adjusted quantities α1 , α 2 ,…, α n or functions of

adjusted quantities ΔE1 , ΔE2 ,… , ΔEn −1 and ΔN1 , ΔN 2 ,… , ΔN n −1 .

Traverses will generally misclose due to the small random errors in the angles (derived from
the measured directions) and the measured distances. To make the traverse mathematically
correct, small corrections must be applied to the measurements to give adjusted quantities.
These adjusted quantities are:

s = s′ + v s

α = α ′ + vα

where s and α are adjusted distance and angle respectively, s′ and α ′ are the measured angle
and distance, and vs and vα are small corrections. Replacing the adjusted quantities with
measurements and corrections allows the first member of equations (6.52) to be expressed as

φ0 + (α1 + vα ) + (α 2 + vα ) + (α 3 + vα ) +
1 2 3
( )
+ α n + vαn = φn

and summing the measured angles and rearranging gives a simple expression for the
summation of corrections to measured angles as

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–28

RMIT University Geospatial Science

vα1 + vα2 + vα3 + + vαn = f1 (6.53)

where, apart from some multiple of 180º

⎛ n
⎞
f1 = φn − ⎜ φ0 + ∑α k′ ⎟ = φn − φn′ (6.54)
⎝ k =1 ⎠

f1 is the angular misclose in the traverse and equation (6.54) simply states that the sum of the
corrections to the measured angles is equal to the angular misclose.

The second and third members of equations (6.52) can also be expressed as a summation of
corrections by considering the following

ΔE = s sin φ and ΔN = s cos φ

where ΔE , ΔN are east and north components of a traverse leg and

s = s′ + vs and φ = φ ′ + vφ

where φ ′ and vφ are "measured" bearing and correction respectively, hence we express the

east and north components as

ΔE = ( s′ + vs ) sin (φ ′ + vφ )
ΔN = ( s′ + vs ) cos (φ ′ + vφ )

Using the trigonometric expansions for sin ( A + B ) and cos ( A + B ) , and the approximations

sin vφ vφ and cos vφ 1 since vφ is a small quantity gives

ΔE = ( s′ + vs ) {sin φ ′ cos vφ + sin vφ cos φ ′} = s′ sin φ ′ + s′vφ cos φ ′ + vs sin φ ′ + vs vφ cos φ ′

ΔN = ( s′ + vs ) {cos φ ′ cos vφ − sin φ ′ sin vφ } = s′ cos φ ′ − s′vφ sin φ ′ + vs cos φ ′ − vs vφ sin φ ′

and since vs and vφ are both small then their product vs vφ 0 , hence

ΔE = s′ sin φ ′ + vφ s′ cos φ ′ + vs sin φ ′

ΔN = s′ cos φ ′ − vφ s′ sin φ + vs cos φ ′

Finally, the east and north components of a traverse leg computed using the measured
quantities are ΔE ′ = s′ sin φ ′ and ΔN ′ = s′ cos φ ′ , and we may write

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–29

RMIT University Geospatial Science

ΔE = ΔE ′ + vφ ΔN ′ + vs sin φ ′
(6.55)
ΔN = ΔN ′ − vφ ΔE ′ + vs cos φ ′

Substituting equations (6.55) into the second and third members of equations (6.52) gives

(
E1 + ΔE1′ + vφ1 ΔN1′ + vs1 sin φ1′ )
+ ( ΔE ′ + v ΔN ′ + v
2 φ2 2 s2 sin φ2′ )
+
(
+ ΔEn′−1 + vφn −1 ΔN n′−1 + vsn −1 sin φn′−1 = En )
(
N1 + ΔN1′ − vφ1 ΔE1′ + vs1 cos φ1′ )
(
+ ΔN 2′ − vφ2 ΔE2′ + vs2 cos φ2′ )
+
(
+ ΔN n′−1 − vφn −1 ΔEn′−1 + vsn −1 cos φn′−1 = N n )
Letting the misclose in the east and north coordinates be

⎧ n −1
⎫
f 2 = En − ⎨ E1 + ∑ ΔEk′ ⎬ = En − En′
⎩ k =1 ⎭
(6.56)
⎧ n −1
⎫
f 3 = N n − ⎨ N1 + ∑ ΔN k′ ⎬ = N n − N n′
⎩ k =1 ⎭

n −1
and recognising that vφ1 = vα1 , vφ2 = vα1 + vα2 , vφ3 = vα1 + vα2 + vα3 etc, and vφn −1 = ∑ vαk then we
k =1

may write

( ) (
vα1 ΔN1′ + vs1 sin φ1′ + vα1 + vα2 ΔN 2′ + vs2 sin φ2′ + vα1 + vα2 + vα3 ΔN 3′ + vs3 sin φ3′ + )
⎛ n −1 ⎞
+ ⎜ ∑ vαk ⎟ ΔN n′−1 + vsn −1 sin φn′−1 = f 2
⎝ k −1 ⎠

( ) (
−vα1 ΔE1′ + vs1 cos φ1′ − vα1 + vα2 ΔE2′ + vs2 cos φ2′ − vα1 + vα2 + vα3 ΔE3′ + vs3 cos φ3′ − )
⎛ n −1 ⎞
− ⎜ ∑ vαk ⎟ ΔEn′−1 + vsn −1 cos φn′−1 = f 3
⎝ k −1 ⎠

Gathering together the coefficients of vα1 , vα2 , vα3 , etc and rearranging gives

( sin φ1′) vs + ( sin φ2′ ) vs + ( sin φ3′ ) vs + + ( sin φn′−1 ) vs

1 2 3 n −1

+ ( N n′ − N1 ) vα + ( N n′ − N 2′ ) vα + ( N n′ − N 3′ ) vα + + ( N n′ − N n′ −1 ) vα
1 2 3 n −1
= f2 (6.57)

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–30

RMIT University Geospatial Science

( cos φ1′) vs + ( cos φ2′ ) vs + ( cos φ3′ ) vs + + ( cos φn′−1 ) vs

1 2 3 n −1

− ( En′ − E1 ) vα − ( En′ − E2′ ) vα − ( En′ − E3′ ) vα − − ( En′ − En′ −1 ) vα

1 2 3 n −1
= f3 (6.58)

Equations (6.53), (6.57) and (6.58) are the three equations that relate corrections to angles and
distances, vα and vs respectively to angular and coordinate misclosures f1 , f 2 and f 3 given by
equations (6.54) and (6.56). In equation (6.53) the coefficients of corrections to angles are all
unity, whilst in equations (6.57) and (6.58) the coefficients of the corrections are sines and
cosines of bearings and coordinate differences derived from the measurements. Equations
(6.53), (6.57) and (6.58) are applicable to any single closed traverse.

6.5.5. Application of Least Squares Adjustment of Observations Only to Particular Single

Closed Traverses

There are three types of single closed traverses.

Type I Traverses that begin and end at different fixed points with fixed
orienting bearings at the terminal points. Figure 6.7(a).

Type II Traverses that begin and end at the same point with a single
fixed orienting bearing. Figure 6.7(b)

Type III Traverses that begin and end at the same point with a fixed
datum bearing. Figure 6.7(b)
Fi

α2
xe
db

°
ea
rin

α1 s1
P2
g
φ0

s2
α3
P1 α4
°
P3 s3 °
P4 ar
ing
α5 e
s4 db
ixe
F

P5

Figure 6.7(a) Type I traverse

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–31

RMIT University Geospatial Science

N
Fi
xe

α2
db

α2
°
ea

α1 rin g
°
rin

ed b ea P2
g

s1 Fix
α1 s1
φ0

P
°
2
P1
s2 α5
α5 P1
s2

P3
°
s4 s4
α3
P3
°
P4 s3
s3 α3

α4 ° P4

α4°
Figure 6.7(b) Type II traverse Figure 6.7(c) Type III traverse

Figures 6.7(a), 6.7(b) and 6.7(c) show three types of closed traverses. In each case, the
traverse consists of four(4) distances s1 to s4 and five(5) angles α1 to α5 . Traverse points

shown with a triangle ( Δ ) can be regarded as fixed with known coordinates.

In Figures 6.7(a) and 6.7(b) the bearing of the traverse line P1 → P2 is found by adding the

observed angle α1 to the fixed bearing φ0 . In both of these traverses five angles must be
observed to "close" the traverse.

In Figure 6.7(c) the bearing of the traverse line P1 → P2 is fixed and only four angles need be

observed to close the traverse. The angle α1 at P1 , clockwise from north to P2 , is the bearing

of the traverse line P1 → P2 . α1 is used in the adjustment as an observation with a standard

deviation of zero.

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–32

RMIT University Geospatial Science

For any single closed traverse, the method of adjustment is as follows:

(i) Calculate the coordinates of the traverse points by using the observed bearings and
distances beginning at point P1 .
(ii) Calculate the angular and coordinate misclosures. In each case, the misclose is the
fixed value minus the observed or calculated value. These three values are the
elements f1 , f 2 and f 3 in the vector of numeric terms f
(iii) Calculate the coefficients of the correction (or residuals) in equations (6.53), (6.57) and
(6.58). These coefficients are either zero or unity for equation (6.53), or sines and
cosines of observed bearings together with coordinate differences in equations (6.57)
and (6.58). These values are the elements of the coefficient matrix A
(iv) Assign precisions (estimated standard deviations squared) of the observations. These
will be the diagonal elements of the cofactor matrix Q
Note: In Type III traverses where the bearing P1 → P2 is fixed, the angle α1 (which is
not observed) is assigned a variance (standard deviation squared) of zero.
(v) Form a set of three(3) normal equations ( AQA T ) k = f

(vi) Solve the normal equations for the three(3) Lagrange multipliers k1 , k2 and k3 which

are the elements of the vector k from k = ( AQA T ) f and then compute the vector of
−1

residuals (corrections) from v = QA T k

(vii) Calculate the adjusted bearings and distances of the traverse by adding the corrections
to the observed angles and distances.

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–33

RMIT University Geospatial Science

6.5.6. Example of Traverse Adjustment using Least Squares Adjustment of Observations

Only

Figure 6.8 is a schematic diagram of a traverse run between two fixed stations A and B and
oriented at both ends by angular observations to a third fixed station C.

The bearings of traverse lines shown on the diagram, unless otherwise indicated, are called
"observed" bearings and have been derived from the measured angles (which have been
derived from observed theodolite directions) and the fixed bearing AC. The difference
between the observed and fixed bearings of the line BC represents the angular misclose. The
coordinates of the traverse points D, E and F have been calculated using the observed
bearings and distances and the fixed coordinates of A. The difference between the observed
and fixed coordinates at B represents the coordinate misclosures.

In this example estimated standard deviations of measured angles α are sα = 5′′ and for

measured distances s are ss = 10mm + 15ppm where ppm is parts per million.
2 75

2034.785 E FIXED
8776.030 N
°0
2′ 4

″4
9′ 1
7″

138° 18

A 5248.853 E
°2

110 8321.726 N
274

°1
3″

240 5′ 20″ E
′ 58″

°
12′ 3

1.60 ″
9 8″ 18
3 4′ 140 4 6′
195°

°
68 032.3 0°
° 1 13 5411.746 E
F 7786.963 N
D
° 1 13
°
1 5 6 49 ′ 50
D

4287.873 E
163°
FI XE

559.03′ 32″

7944.574 N 4. 68 ″
″
3 48′ 11
302°
0
22

″
01
3 8′ 46″ B
C 6° 7′
23 6° 3
23 6843.030 E
OBSERVED
ED 7154.779 N
E RV
S D
OB FIXE 6843.085 E
FIXED
7154.700 N
C

Figure 6.8 Traverse diagram showing field measurements, derived values and fixed values.

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–34

RMIT University Geospatial Science

Step 1: Calculation of misclosures and formation of vector f

From equations (6.54) and (6.56) the angular and coordinate misclosures are the elements
f1 , f 2 and f 3 of the vector of numeric terms f. These misclosures may be characterised as

misclose = fixed - observed

angular misclose: f1 = φn − φn′

= 236 37′ 46′′ − 236 38′ 01′′
= −15′′

east misclose: f 2 = En − En′

= 6843.085 − 6843.030
= 0.055 m
= 5.5 cm

north misclose: f 3 = N n − N n′
= 7154.700 − 7154.779
= −0.079 m
= −7.9 cm

⎡ −15 ⎤ sec
vector of numeric terms: f = ⎢ 5.50 ⎥ cm
⎢ ⎥
⎢⎣ −7.9 ⎥⎦ cm

Note that the units of the numeric terms are seconds of arc (sec) and centimetres (cm)

Step 2: Form the coefficient matrix A of the equations (6.16) Av = f

The first row of A contains coefficients of zero or unity from equation (6.53)

vα1 + vα2 + vα3 + + vαn = f1

⎛ 100 ⎞
The second row of A contains the coefficients sin φk′ and ( N n′ − N k′ ) ⎜ ⎟ from equation
⎝ ρ ′′ ⎠
(6.57).

( sin φ1′) vs + ( sin φ2′ ) vs + ( sin φ3′ ) vs + + ( sin φn′−1 ) vs

1 2 3 n −1

+ ( N n′ − N1 ) vα + ( N n′ − N 2′ ) vα + ( N n′ − N 3′ ) vα + + ( N n′ − N n′ −1 ) vα
1 2 3 n −1
= f2

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–35

RMIT University Geospatial Science

Note that the coefficients of the distance residuals are dimensionless quantities and the
180
coefficients of the angle residuals have the dimensions of sec/cm where ρ ′′ = × 3600 is
π
the number of seconds in one radian.

⎛ 100 ⎞
The third row of A contains the coefficients cos φk′ and − ( En′ − Ek′ ) ⎜ ⎟ from equation
⎝ ρ ′′ ⎠
(6.58).

( cos φ1′) vs + ( cos φ2′ ) vs + ( cos φ3′ ) vs + + ( cos φn′−1 ) vs

1 2 3 n −1

− ( En′ − E1 ) vα − ( En′ − E2′ ) vα − ( En′ − E3′ ) vα − − ( En′ − En′ −1 ) vα

1 2 3 n −1
= f3

Note that the coefficients of the distance residuals are dimensionless quantities and the
180
coefficients of the angle residuals have the dimensions of sec/cm where ρ ′′ = × 3600 is
π
the number of seconds in one radian. The equation Av = f is

⎡v ⎤ ↑
⎢ v ⎥ angles
⎢ ⎥
⎢v ⎥
⎢ ⎥ ↓
⎡ 0 0 0 0 1 1 1 1 1⎤ v
⎢ ⎥ ⎡ −15 ⎤
⎢ 0.9382 0.9309 0.2914 0.9147 −0.7860 −0.3829 −0.5658 −0.3065 0⎥ ⎢v ⎥ ↑ = ⎢ 5.50 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎣⎢ −0.3462 0.3653 −0.9566 −0.4040 −2.3311 −1.2388 −0.7729 −0.6939 0⎦⎥ v ⎢ −7.90 ⎥⎦
⎣
⎢ ⎥
4.6 cm 2.5 cm 1.8 cm 3.3 cm 5′′ 5′′ 5′′ 5′′
⎢ ⎥
5′′ v distances
⎢ ⎥
⎢v ⎥
⎢⎣ v ⎥⎦ ↓

Note that the numbers below the columns of A are the estimates of the standard deviations of
the distances or angles associated with the coefficients.

Step 3: Form the normal equations

The normal equations are given by equations (6.25) as ( AQA T ) k = f

where Q = W −1 is the cofactor matrix containing estimates of the variances of the

measurements. Q and the weight matrix are W are square diagonal matrices, i.e., all off-
diagonal elements are zero and since weights are inversely proportional to the estimates of the

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–36

RMIT University Geospatial Science

⎧ 1 1 1 1 1 1 1 1 1 ⎫
variances, the diagonal elements of Q = ⎨ 2 ⎬
⎩ sα 1 sα2 2 sα2 3 sα2 4 ss21 ss22 ss23 ss24 ss25 ⎭

where the first 4 elements relate to the angles and the remaining 5 elements relate to the
distances. Now consider a diagonal matrix that denoted Q whose diagonal elements are the

square-roots of the elements of Q and Q = Q Q and another matrix A = A Q . Each

element of A is the original element of A multiplied by the estimate of the standard deviation
associated with the particular element and the normal equations are given by ( AA T ) k = f

where

⎡ 0 0 0 0 5 5 5 5 5⎤
A = 4.3155 2.3272 0.5245 3.0186 −1.9145 −1.9145 −2.8288 −1.5324 0⎥
⎢
⎢ ⎥
⎣⎢ −1.5926 0.9134 −1.7219 −1.3333 −6.1939 −6.1939 −3.8644 −3.4696 0 ⎥⎦

⎡ 125 −51.0285 −125.9168⎤

and ⎢
AA = −51.0285 62.8875
T
64.2375⎥
⎢ ⎥
⎢⎣ −125.9168 64.2375 209.2995⎥⎦

Step 4: Solve the normal equations for the vector of Lagrange multipliers k

From equations (6.47) and with the modification mentioned above

⎡ 0.0219 0.0063 0.0112 ⎤ ⎡ −15 ⎤

k = ( AQA T ) f = ( AA T ) f=⎢ 0.0250 −0.0039 ⎥ ⎢ 5.5 ⎥
−1 −1

⎢ ⎥⎢ ⎥
⎢⎣symmetric 0.0127 ⎥⎦ ⎢⎣ −7.9 ⎥⎦

⎡ −0.3825⎤
and k = ⎢ 0.0738⎥
⎢ ⎥
⎣⎢ −0.2906⎦⎥

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–37

RMIT University Geospatial Science

Step 5: Calculation of residuals and adjusted traverse dimensions

The residuals are obtained from equation (6.24) v = QA T k

Since the cofactor matrix Q is diagonal, the individual residuals can be calculated from

v j = s 2j ( a1 j k1 + a2 j k2 + a3 j k3 ) (6.59)

where

a1 j , a2 j , a3 j are elements of the coefficient matrix A

k1 , k2 , k3 are the elements of the vector k

s 2j is the estimate of the variance of the jth measurement

For example, the residual for the second distance ( j = 2 ) is

( 2.5) {( 0 )( −0.3825) + ( 0.9309 )( 0.0738) + ( 0.3653)( −0.2906 )} = −0.23 cm

2

and the residual for the third measured angle ( j = 7 ) is

( 5) {(1)( −0.3825) + ( −0.5658)( 0.0738) + ( −0.7729 )( −0.2906 )} = −4.99′′

2

Exactly the same result can be obtained by using the estimate of the standard deviations s j

and the elements of the matrix A

v j = s j ( a1 j k1 + a2 j k2 + a3 j k3 ) (6.60)

Both methods give

⎡ 3.59 cm ⎤ ↑
⎢ −0.23 ⎥
⎢ ⎥
⎢ 0.97 ⎥ distances
⎢ ⎥
⎢ 2.01 cm ⎥ ↓
v = ⎢ 5.92′′ ⎥ ↑
⎢ ⎥
⎢ −1.27 ⎥
⎢ −4.99 ⎥ angles
⎢ ⎥
⎢ −5.09 ⎥
⎢⎣ −9.56′′ ⎥⎦ ↓

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–38

RMIT University Geospatial Science

The residuals for the bearings are the cumulative residuals for the angles up to the particular
traverse line. They are

⎡ 5.92′′⎤
⎢ 4.65 ⎥
⎢ ⎥
vφ = ⎢ −0.34 ⎥
⎢ ⎥
⎢ −5.43 ⎥
⎢⎣ −14.99′′⎥⎦

Applying these residuals (or corrections) to the measured quantities gives the adjusted
traverse dimensions as

Line Bearing Distance

k φk sk
1 110º 27′ 25.9′′ 2401.645
2 68º 34′ 22.6′′ 1032.338
3 163º 03′ 31.7′′ 559.032
4 113º 49′ 44.6′′ 1564.703

Table 6.6 Adjusted traverse distances

2034.785 E
A
8776. 030 N 5248.869 E
8321.627 N
110
° 15 ″
22.6
3″

240 ′ 25.9 34′ 338

E
°
12′ 3

1.64 8°
D 6 1032
5 .
559.0 03′ 31.7″

5411.766 E
16 3°
195 °

° F
7786.854 N
32

4287.883 E
7944.497 N
° 113
° 49
156 ′ 44.6 6843. 085 E
4.70 ″
3 B 7154.700 N

″
46
C 7′
6°3
2 3

Figure 6.9 Traverse diagram showing adjusted measurements.

© 2005, R.E. Deakin Notes on Least Squares (2005) 6–39

RMIT University Geospatial Science

7. LINEARIZATION USING TAYLOR'S THEOREM AND THE DERIVATION OF

SOME COMMON SURVEYING OBSERVATION EQUATIONS

In many surveying "problems" the solution depends upon selection of a mathematical model
suitable to the problem, and using this, together with the observations (or measurements)
obtain a solution.

For example, a surveyor is required to determine the location (the coordinates) of a point.
From this "unknown" point, they can see three known points (i.e., points of known
coordinates). Understanding geometric principles, the surveyor measures the directions to
these three known points with a theodolite, determines the two angles α and β between the
three lines and "solves the problem". In surveying parlance, this technique of solution of
position is known as a resection; the mathematical model is based on geometric principles and
the observations are the directions, from which the necessary angles are obtained for a
solution.

Choosing a resection, as an example of a "surveying problem" is appropriate, since it

demonstrates the case of determining quantities (the coordinates of the unknown point) from
indirect measurements. That is, the surveyor's measurements of directions are indirect
measurements of coordinate differences between the unknown point and the known points.

In many surveying problems, the observations exceed the necessary number required for a
unique solution. Again, using a resection as an example, consider the case where the surveyor
(at an unknown point) measures the directions to four known points. There are now multiple
solutions for the resection point, since the four directions give rise to three angles, exceeding
the minimum geometric requirements for a unique solution. That is, there is a redundancy in
the mathematical model. In this case of the resection, and other surveying problems where
there are redundant measurements, the method of least squares can be employed to obtain the
best estimate of the "unknowns".

Least squares (as a method of determining best estimates), depends upon the formation of sets
of observation equations and their solution. The normal techniques of solution of systems of
equations require that the sets of observation equations must be linear, i.e., "unknowns"

© 2005, R.E. Deakin Notes on Least Squares (2005) 7–1

RMIT University Geospatial Science

linearly related to measurements. This is not always the case. For example, in a resection,
the measurements, directions α ik from the unknown point Pi to known points Pk , are non-
linear functions of the coordinate differences (the unknowns).

The observation equations for observed directions in a mathematical model of a resection

have the general form

⎛ Ek − Ei ⎞
αik + vik + zi = tan −1 ⎜ ⎟ (7.1)
⎝ Nk − Ni ⎠

αik are the observed directions from the resection point Pi to the known points Pk ,

vik are the residuals (small corrections) associated with observed directions,

zi is an orientation "constant"; the bearing of the Reference Object (RO) for the

set of observed directions,

Ek , N k are the east and north coordinates of the known points, and

Ei , N i are the east and north coordinates of the resection point.

Clearly, in this case, the measurements α ik are non-linear functions of the unknowns Ei , N i

and any system of equations in the form of (7.1) would be non-linear and could not be solved
by normal means. Consequently, whenever the equations in a mathematical model are non-
linear functions linking the measurements with the unknowns, some method of linearization
must be employed to obtain sets of linear equations.

The most common method of linearization is by using Taylor's theorem to represent the
function as a power series consisting of zero order terms, 1st order terms, 2nd order terms and
higher order terms. By choosing suitable approximations, second and higher-order terms can
be neglected, yielding a linear approximation to the function. This linear approximation of
the mathematical model can be used to form sets of linear equations, which can be solved by
normal means.

© 2005, R.E. Deakin Notes on Least Squares (2005) 7–2

RMIT University Geospatial Science

7.1. Taylor's Theorem

This theorem, due to the English mathematician Brook Taylor (1685–1731) enables the value
of a real function f ( x ) near a point x = a to be estimated from the values f ( a ) and the
derivatives of f ( x ) evaluated at x = a . Taylor's theorem also provides an estimate of the
error made in a polynomial approximation to a function. The Scottish mathematician Colin
Maclaurin (1698–1746) developed a special case of Taylor's theorem, which was named in his
honour, where the function f ( x ) is expanded about the origin x = a = 0 . The citations
below, from the Encyclopaedia Britannica give some historical information about Taylor and
Maclaurin.

Taylor, Brook (b. Aug. 18, 1685, Edmonton, Middlesex, Eng.– d. Dec. 29, 1731,
London), British mathematician noted for his contributions to the development of
calculus.
In 1708 Taylor produced a solution to the problem of the centre of oscillation. The
solution went unpublished until 1714, when his claim to priority was disputed by the
noted Swiss mathematician Johann Bernoulli. Taylor's Methodus incrementorum directa
et inversa (1715; "Direct and Indirect Methods of Incrementation") added to higher
mathematics a new branch now called the calculus of finite differences. Using this new
development, he was the first to express mathematically the movement of a vibrating
string on the basis of mechanical principles. Methodus also contained the celebrated
formula known as Taylor's theorem, the importance of which remained unrecognized
until 1772. At that time the French mathematician Joseph-Louis Lagrange realized its
importance and proclaimed it the basic principle of differential calculus.
A gifted artist, Taylor set forth in Linear Perspective (1715) the basic principles of
perspective. This work and his New Principles of Linear Perspective contained the first
general treatment of the principle of vanishing points. Taylor was elected a fellow of the
Royal Society of London in 1712 and in that same year sat on the committee for
adjudicating Sir Isaac Newton's and Gottfried Wilhelm Leibniz's conflicting claims of
priority in the invention of calculus.

Maclaurin, Colin (b. February 1698, Kilmodan, Argyllshire, Scot.–d. June 14, 1746,
Edinburgh), Scottish mathematician who developed and extended Sir Isaac Newton's
work in calculus, geometry, and gravitation. A child prodigy, he entered the University
of Glasgow at age 11. At the age of 19, he was elected professor of mathematics at
Marischal College, Aberdeen, and two years later he became a fellow of the Royal
Society of London. At this time he became acquainted with Newton. In his most
important work, Geometrica Organica; Sive Descriptio Linearum Curvarum Universalis
(1720; "Organic Geometry, with the Description of the Universal Linear Curves"),
Maclaurin developed several theorems similar to some in Newton's Principia,
introduced the method of generating conics (the circle, ellipse, hyperbola, and parabola)
that bears his name, and showed that certain types of curves (of the third and fourth
degree) can be described by the intersection of two movable angles. On the
recommendation of Newton, he was made professor of mathematics at the University of
Edinburgh in 1725. In 1740 he shared, with the mathematicians Leonhard Euler and
Daniel Bernoulli, the prize offered by the Académie des Sciences for an essay on tides.
His Treatisw of Fluxions (1742) was written in reply to criticisms by George Berkeley of
England that Newton's calculus was based on faulty reasoning. In this essay he showed

© 2005, R.E. Deakin Notes on Least Squares (2005) 7–3

RMIT University Geospatial Science

that stable figures for a homogeneous rotating fluid mass are the ellipsoids of revolution,
later known as Maclaurin's ellipsoids. He also gave in his Fluxions, for the first time,
the correct theory for distinguishing between maxima and minima in general and pointed
out the importance of the distinction in the theory of the multiple points of curves. The
Maclaurin series, a special case of the Taylor series, was named in his honour. In 1745,
when Jacobites (supporters of the Stuart king James II and his descendants) were
marching on Edinburgh, Maclaurin took a prominent part in preparing trenches and
barricades for the city's defense. As soon as the rebel army captured Edinburgh,
Maclaurin fled to England until it was safe to return. The ordeal of his escape ruined his
health, and he died at age 48. Maclaurin's Account of Sir Isaac Newton's Philosophical
Discoveries was published posthumously, as was his Treatise of Algebra (1748). "De
Linearum Geometricarum Proprietatibus Generalibus tractatus" ("A Tract on the General
Properties of Geometrical Lines"), noted for its elegant geometric demonstrations, was
appended to his Algebra. Copyright 1994-1999 Encyclopædia Britannica

Taylor's theorem may be expressed in the following form

( x − a) ( x − a)
2 3

f ( x ) = f ( a ) + ( x − a ) f ′( a ) + f ′′( a ) + f ′′′( a ) + "

2! 3!
(7.2)
( x − a)
n −1
( n −1)
+ f ( a ) + Rn
( n − 1) !

where Rn is the remainder after n terms and lim Rn = 0 for f ( x ) about x = a

n →∞

f ′( a ) , f ′′( a ), " etc are derivatives of the function f ( x ) evaluated at x = a .

Taylor's theorem can also be expressed as power series

(x − a)
k
n
f ( x) = ∑ f (k )
(a) (7.3)
k =0 k!

where f ( k ) ( a ) denotes the kth derivative of the function f ( x ) evaluated at x = a and

f (0) ( a ) is the function f ( x ) evaluated at x = a , and 0! = 1 .

Other forms of Taylor's theorem may be obtained by a change of notation, for example: let
x = a + h , then f ( x ) = f ( a + h ) and x − a = h . Substitution into equation (7.2) gives

h2 h3
f ( x ) = f ( a + h ) = f ( a ) + h f ′( a ) + f ′′( a ) + f ′′′( a ) + "
2! 3!
(7.4)
h n −1
+ f ( n −1) ( a ) + Rn
( n − 1)!
This may be a more convenient form of Taylor's theorem for a particular application.

© 2005, R.E. Deakin Notes on Least Squares (2005) 7–4

RMIT University Geospatial Science

Inspection of equations (7.2), (7.3) and (7.4) show that Taylor's theorem can be used to
expand a non-linear function (about a point) into a linear series. Expansions of this form, also
called Taylor's series, are a convergent power series approximating f ( x ) .

Taylor's series for functions of two variables

Say φ = f ( x, y ) then the Taylor series expansion of the function φ about x = a and y = b is

∂f ∂f
φ = f ( a, b ) + ( x − a ) + ( y − b)
∂x ∂y
1⎧ 2 ∂ f
2
2 ∂ f
2
∂f ∂f ⎫
+ ⎨( x − a ) 2 ( + y − b ) 2 + ( x − a )( y − b ) ⎬ +" (7.5)
2! ⎩ ∂x ∂y ∂x ∂y ⎭

where f ( a , b ) is the function φ evaluated at x = a and y = b

∂f ∂f ∂ 2 f
, , , etc are partial derivatives of the function φ evaluated at x = a and
∂x ∂y ∂x 2
y = b.

Taylor's series for functions of three variables

Say φ = f ( x, y , z ) then the Taylor series expansion of the function φ about x = a , y = b and

z=c

∂f ∂f ∂f
φ = f ( a, b, c ) + ( x − a ) + ( y − b) + ( z − c )
∂x ∂y ∂z
1⎧ 2 ∂ f
2
2 ∂ f
2
2 ∂ f
2
+ ⎨( x − a ) + ( y − b) + ( z − c)
2! ⎩ ∂x 2 ∂y 2 ∂z 2
∂f ∂f ∂f ∂f ∂f ∂f ⎫
+ ( x − a )( y − b ) + ( x − a )( z − c ) + ( y − b )( z − c ) ⎬ +" (7.6)
∂x ∂y ∂x ∂z ∂y ∂z ⎭

where f ( a , b, c ) is the function φ evaluated at x = a , y = b and z = c

∂f ∂f ∂f ∂ 2 f
, , , , etc are partial derivatives evaluated at x = a , y = b and z = c .
∂x ∂y ∂z ∂x 2

© 2005, R.E. Deakin Notes on Least Squares (2005) 7–5

RMIT University Geospatial Science

Extensions to four or more variables follow a similar pattern. Equations (7.5) and (7.6) show
only terms up to the 2nd order; no remainder terms are shown.

7.2. Linear Approximations to Functions using Taylor's Theorem

In the Taylor expansions of functions shown above, suppose that the variables x, y , z, " etc

are expressed as x = x 0 + Δx , y = y 0 + Δy , z = z 0 + Δz " etc where x 0 , y 0 , z 0 , " etc are

approximate values and Δx, Δy , Δz , " etc are small corrections.

The Taylor series expansion of a single variable can be expressed as

2 3
φ = f ( x ) = f ( x0 ) + ( x − x0 ) + ( x − x0 ) ( )
df 2 d f 0 3 d f
+ x − x +"
dx dx 2 dx 3
2 3
= f ( x 0 ) + Δx + ( Δ x )
df 2 d f 3 d f
+ ( Δ x ) +"
dx dx 2 dx 3
= f ( x 0 ) + Δx + higher order terms
df
dx

df d 2 f d 3 f
where the derivatives , 2
, 3
, " etc are evaluated at the approximation x 0 . If the
dx dx dx

correction Δx is small, then ( Δx ) , ( Δx ) , " etc will be exceedingly small and the higher
2 3

order terms may be neglected, giving the following linear approximation

df
For φ = f ( x ) φ = f ( x )  f ( x 0 ) + Δx (7.7)
dx

Using similar reasoning, linear approximations can be written for functions of two and three
variables.

∂f ∂f
For φ = f ( x, y ) φ = f ( x , y )  f ( x 0 , y 0 ) + Δx + Δy (7.8)
∂x ∂y

∂f ∂f ∂f
For φ = f ( x, y , z ) , φ = f ( x , y , z )  f ( x 0 , y 0 , z 0 ) + Δx + Δ y + Δz (7.9)
∂x ∂y ∂z

Similar linear approximations can be written for functions of four or more variables. In
equations (7.7), (7.8) and (7.9) the derivatives are evaluated at the approximations x 0 , y 0 , z 0 .

© 2005, R.E. Deakin Notes on Least Squares (2005) 7–6

RMIT University Geospatial Science

Generalizing this linear form gives, for φ = f ( x1 , x2 , x3 ," xn )

∂f ∂f ∂f ∂f
φ = f ( x1 , x2 , x3 ," xn )  f ( x10 , x20 , x30 ," xn0 ) + Δx1 + Δx2 + Δx3 + " + Δ xn (7.10)
∂x1 ∂x2 ∂x3 ∂xn

This equation can be written in matrix form

φ = f ( x )  f ( x 0 ) + j Δx (7.11)

where x is a vector of variables, x 0 a vector of approximate values of the variables, j is a row

vector of partial derivatives and Δx is a column vector of corrections.

Suppose this generalized form, equation (7.11), is extended to the general case of m variables
y1 , y2 , y3 , " ym and each variable yk is a function of a set of variables x1 , x2 , x3 , " xn i.e.,

y1 = f1 ( x1 , x2 , x3 ," xn )
y2 = f 2 ( x1 , x2 , x3 ," xn )
#
ym = f m ( x1 , x2 , x3 ," xn )

Expressing each variable yk in a linearized form gives

∂y1 ∂y ∂y
y1 = y10 + Δx1 + 1 Δx2 + " + 1 Δxn
∂x1 ∂x2 ∂xn
∂y2 ∂y ∂y
y2 = y20 + Δx1 + 2 Δx2 + " + 2 Δxn
∂x1 ∂x2 ∂xn (7.12)
#
∂ym ∂y ∂y
ym = ym0 + Δx1 + m Δx2 + " + m Δxn
∂x1 ∂x2 ∂xn

Equations (7.12) can be expressed in matrix notation as

y = y 0 + J Δx (7.13)

where

y is an (m,1) vector of (unknown) function values, y = [ y1 y2 " ym ]

T

T
y 0 is an (m,1) vector of approximate values of the functions, y 0 = ⎡⎣ y10 y20 " ym0 ⎤⎦

Δx is an (n,1) vector of corrections to the approx. values, Δx = [ Δx1 Δx2 " Δxn ]
T

© 2005, R.E. Deakin Notes on Least Squares (2005) 7–7

RMIT University Geospatial Science

J is the (m,n) the Jacobian matrix of partial derivatives

⎡ ∂y1 ∂y1 ∂y1 ∂y1 ⎤

⎢ ∂x "
∂x2 ∂x3 ∂xn ⎥
⎢ 1 ⎥
⎢ ∂y2 ∂y2 ∂y2 ∂y2 ⎥
"
J = ⎢ ∂x1 ∂x2 ∂x3 ∂xn ⎥
⎢ ⎥
⎢ # # ⎥
⎢ ∂y ∂ym ∂ym ∂ym ⎥
⎢ m " ⎥
⎢⎣ ∂x1 ∂x2 ∂x3 ∂xn ⎦⎥

7.3. The Derivation of some Common Surveying Observation Equations

Consider Figure 7.1. Pi is the instrument point and directions αik and distances sik have been

observed to stations P1 , P2 , P3 " Pk . P1 is the Reference Object (RO) and the direction

αi1 = 0D 00′ 00′′ . A bearing is assigned to the RO and bearings to all other stations may be
obtained by adding the observed directions to the bearing of the RO.

N'
P3 Ek - Ei Pk



Nk - Ni

sik
α i3 αik
si3 φik
E'
zi


α i2 Pi
s i2 α i1
s i1
P2 P1 (RO)



Figure 7.1 Observed directions α and distances s from Pi

αik observed direction Pi to Pk

φik bearing Pi to Pk

© 2005, R.E. Deakin Notes on Least Squares (2005) 7–8

RMIT University Geospatial Science

sik distance Pi to Pk

zi orientation constant for directions at Pi (bearing of the RO)

θ ik "observed" bearing θ ik = αik + zi

Ei , N i coordinates of Pi

Ek , N k coordinates of Pk

From Figure 7.1 the bearings φik and distances sik are non-linear functions of the coordinates

of points Pi and Pk

⎛ Ek − Ei ⎞
φik = tan −1 ⎜ ⎟ (7.14)
⎝ Nk − Ni ⎠

sik = ( Ek − Ei ) + ( Nk − Ni )
2 2
(7.15)

With E = E 0 + ΔE and N = N 0 + ΔN where E 0 , N 0 are approximate values of the

coordinates and ΔE , ΔN are small corrections, linear approximations of φik and sik can be
written as

∂φik ∂φ ∂φ ∂φ
φik = φik0 + ΔEk + ΔN k ik + ΔEi ik + ΔN i ik (7.16)
∂Ek ∂N k ∂Ei ∂N i

∂sik ∂s ∂s ∂s
sik = sik0 + ΔEk + ΔN k ik + ΔEi ik + ΔN i ik (7.17)
∂Ek ∂N k ∂Ei ∂N i

where φik0 and sik0 are approximate bearings and distances respectively, obtained by

substituting the approximate coordinates Ek0 , N k0 , Ei0 , N i0 into equations (7.14) and (7.15).

The partial derivatives in equations (16) are evaluated in the following manner.

du dv
−u
v
d 1 du d ⎛u⎞ dx dx
Using the relationships: tan −1 u = and ⎜ ⎟=
dx 1 + u 2 dx dx ⎝ v ⎠ v2

∂φik
The partial derivative
∂Ek

© 2005, R.E. Deakin Notes on Least Squares (2005) 7–9

RMIT University Geospatial Science

∂φik ⎛ E k − Ei ⎞ ( Nk − Ni )
2
1 ∂ N k − Ni
= ⎜ ⎟=
∂Ek ⎛ E − Ei ⎞ ∂Ek ⎝ N k − N i ⎠ ( N k − N i ) + ( E k − Ei ) ( N k − N i )
2 2 2 2

1+ ⎜ k ⎟
⎝ Nk − Ni ⎠

giving

∂φik N k − Ni N − N cos φik

= = k 2 i = = bik (7.18)
∂Ek ( N k − N i ) + ( Ek − Ei )
2 2
sik sik

Similarly

∂φik − ( Ek − Ei ) − ( Ek − Ei ) − sin φik

= = = = aik (7.19)
∂N k ( N k − N i ) + ( Ek − Ei )
2 2
sik2 sik

∂φik − ( Nk − Ni ) − ( N k − N i ) − cos φik

= = = = −bik (7.20)
∂Ei ( N k − N i ) + ( Ek − Ei )
2 2
sik2 sik

∂φik
=
( Ek − Ei ) ( E − E ) sin φik = −a
= k 2 i = (7.21)
∂N i ( N k − N i ) + ( Ek − Ei )
2 2 ik
sik sik

aik and bik are known as direction coefficients.

The partial derivatives of equation (7.17) are evaluated in the following manner

∂sik
The partial derivative
∂Ek

∂sik 1 ⎡ E − Ei
1
2 −
= ( Ek − Ei ) + ( N k − N i ) ⎤ 2 2 ( Ek − Ei ) = k = sin φik = d ik
2
(7.22)
∂Ek 2 ⎣ ⎦ sik

Similarly

∂sik N k − N i
= = cos φik = cik (7.23)
∂N k sik

∂sik − ( Ek − Ei )
= = − sin φik = − d ik (7.24)
∂Ei sik

∂sik − ( N k − N i )
= = − cos φik = −cik (7.25)
∂N i sik

© 2005, R.E. Deakin Notes on Least Squares (2005) 7–10

RMIT University Geospatial Science

cik and d ik are known as distance coefficients.

7.3.1. Observation equation for measured directions

An observation equation, relating observed directions to coordinates Pi and Pk can be written

as

⎛ Ek − Ei ⎞
αik + vik + zi = φik = tan −1 ⎜ ⎟ (7.26)
⎝ Nk − Ni ⎠

where vik are the residuals (small corrections) associated with observed directions. Using
equation (7.16) together with the partial derivatives given in equations (7.18) to (7.21) gives a
linear approximation of the observation equation for an observed direction

αik + vik + zi = aik ΔN k + bik ΔEk − aik ΔN i − bik ΔEi + φik0 (7.27)

− ( Ek − Ei ) − sin φik N − N cos φik

where aik = 2
= and bik = k 2 i = are the direction coefficients
sik sik sik sik

7.3.2. Observation equation for measured distances

An observation equation, relating observed distances to coordinates Pi and Pk can be written

as

sik + vik = ( Ek − Ei ) + ( Nk − Ni )
2 2
(7.28)

where vik are the residuals (small corrections) associated with observed distances. Using
(7.17) together with the partial derivatives given in equations (7.22) to (7.25) gives a linear
approximation of the observation equation for an observed distance

sik + vik = cik ΔN k + d ik ΔEk − cik ΔN i − d ik ΔEi + sik0 (7.29)

Nk − Ni E − Ei
where cik = = cos φik and d ik = k = sin φik are the distance coefficients
sik sik

© 2005, R.E. Deakin Notes on Least Squares (2005) 7–11

RMIT University Geospatial Science

7.4. An Example of Taylor's Theorem in Practice

Figure 7.2 shows a point P, whose coordinates are unknown, intersected by bearings from
stations A and B whose coordinates are known.

N Coordinates
P
φA ⊗ A: 12273.910 E B: 12875.270 E
29612.310 N 28679.600 N
A
N Bearings
φB φ A = 81D 01′ 23′′ φ B = 34D 47′ 52′′
N
B Approximate coords P: 13677 E
E 29834 N
Figure 7.2 Bearing intersection

The information given above can be used to compute the coordinates of P by using an
iterative technique employing linearized observation equations approximating the bearings
φ A and φ B . These observation equations [see equation (7.27)] have been derived using
Taylor's theorem.

In general, a bearing is a function of the coordinates of the ends of the line, i.e.,

⎛ Ek − Ei ⎞
φik = tan −1 ⎜ ⎟ = f ( Ek , N k , Ei , N i ) (7.30)
⎝ N k − Ni ⎠

where subscripts i and k represent instrument and target respectively. In this example
(intersection) A and B are instrument points and are known and P is a target point and is
unknown hence

φik = f ( Ek , N k )

is a non-linear function of the variables Ek and N k only (the coordinates of P). Using

equations (7.26) and (7.27) with modifications ΔEi = ΔN i = 0 since the coordinates of the
instrument points are known gives

φik = aik ΔN k + bik ΔEk + φik0 (7.31)

© 2005, R.E. Deakin Notes on Least Squares (2005) 7–12

RMIT University Geospatial Science

Ek = Ek0 + ΔEk , N k = N k0 + ΔN k and Ek0 , N k0 , ΔEk , ΔN k are approximate coordinates and

− ( Ek − Ei ) − sin φik N − N cos φik

small corrections respectively. aik = 2
= and bik = k 2 i = are
sik sik sik sik

direction coefficients and φik0 is an approximate bearing. Note that φik0 and the direction

coefficients aik and bik are computed using the approximate coordinates of P.

Using equation (7.31), two equations for bearings φ A and φ B may be written as

φ A = a AΔN P + bAΔEP + φ A0
φ B = aB ΔN P + bB ΔEP + φ B0

These equations can be rearranged and expressed in matrix form as

⎡aA bA ⎤ ⎡ ΔN P ⎤ ⎡φ A − φ A0 ⎤
=
⎢a
⎣ B bB ⎥⎦ ⎢⎣ ΔEP ⎥⎦ ⎢⎣φ B − φ B0 ⎥⎦

or Cx = u

where C is a matrix of direction coefficients, x is the vector of corrections to the approximate

coordinates of P and u is a vector of numeric terms (observed bearing – computed bearing).

The solution for the corrections in vector x is given by

x = C−1u

From the information given with Figure 7.2 the computed bearings (φik0 ) and distances ( sik0 )

using the approximate coordinates of P are

AP: 81º 01' 17.1" BP: 34º 46' 47.8"

1420.5 m 1405.5 m
and the numeric terms in vector u are

u A = φ A − φ A0 uB = φ B − φ B0
= 81D 01′ 23′′ − 81D 01′17.1′′ = 34D 47′ 52′′ − 34D 46′ 47.8′′
= 5.9′′ = 64.2′′

© 2005, R.E. Deakin Notes on Least Squares (2005) 7–13

RMIT University Geospatial Science

With the elements of x (the corrections to the approximate coordinates of P) in centimetres

and the elements of u (the differences in observed and computed bearings) in seconds of arc

Cx = u
↓ 2
cm's seconds

the elements of the coefficient matrix C will be computed in sec cm (seconds per centimetre)
to maintain consistency of units so that

x = C−1u
0 ↓ 2
cm cm sec sec

Note that if the units (or dimensions) of the elements of C are sec cm then the units of the

elements of the inverse C−1 are cm sec .

The elements of C are the direction coefficients and with distances sik0 in centimetres

− sin φik0 cos φik0 180

aik = × ρ ′′ and bik = × ρ ′′ where ρ ′′ = × 3600
sik0 sik0 π

− sin ( 81D 01′17.1′′)

giving aA = × ρ ′′ = −1.43426 sec cm
(1420.5)(100 )
cos (81D 01′17.1′′)
bA = × ρ ′′ = 0.22662 sec cm
(1420.5)(100)
aB = −0.83714 sec cm
bB = 1.20538 sec cm

The matrix equation Cx = u is

⎡ −1.43426 0.22662 ⎤ ⎡ ΔN P ⎤ ⎡ 5.9 ⎤

⎢ −0.83714 1.20538 ⎥ ⎢ ΔE ⎥ = ⎢ 64.2 ⎥
⎣ ⎦⎣ P⎦ ⎣ ⎦

and the solution x = C−1u is

⎡ ΔN P ⎤ ⎡ −0.78316 0.14724 ⎤ ⎡ 5.9 ⎤ ⎡ 4.83⎤ cm

⎢ ΔE ⎥ = ⎢ −0.54391 0.93187 ⎥ ⎢64.2 ⎥ = ⎢56.62 ⎥ cm
⎣ P⎦ ⎣ ⎦⎣ ⎦ ⎣ ⎦

© 2005, R.E. Deakin Notes on Least Squares (2005) 7–14

RMIT University Geospatial Science

giving the "adjusted" coordinates of P as

N P = N P0 + ΔN P = 29834.400 + 0.048 = 29834.048

EP = EP0 + ΔEP = 13677.000 + 0.566 = 13677.566

These are the "new" approximate coordinates for P. A further iteration will show that the
corrections to these values are less than 0.5 mm, hence the values above could be regarded as
exact.

© 2005, R.E. Deakin Notes on Least Squares (2005) 7–15

RMIT University Geospatial Science

8. THE STANDARD ERROR ELLIPSE

After a Least Squares Adjustment of survey data, intersection, resection or a combination of

both, using the method of Variation of Coordinates, the cofactor matrix Q xx = N −1 contains

estimates of the variances and covariances of the adjusted quantities. These precision
estimates, variances sE2 , sN2 and covariance sEN can be used to define a geometric figure

known as the Standard Error Ellipse, which is a useful graphical representation of the
precision of a position fix. Poor or "weak" fixes are indicated by narrow elongated ellipses
and good or "strong" position fixes are indicated by near circular ellipses.

Error ellipses may be computed for points before any observations are made provided the
approximate locations of points (fixed and floating) are known. Observations (directions,
bearings and distances) may be scaled from maps and diagrams and an approximate set of
normal equations formed. The inverse of the coefficient matrix N yields all the information
required for the computation of the parameters of the error ellipses. In such cases, error
ellipses are an important analysis tool for the surveyor in planning survey operations

8.1. The Pedal Curve of the Standard Error Ellipse

Consider a point whose precision estimates, variances sE2 , sN2 and covariance sEN are known.
The variance in any other direction u may be calculated by considering the projection of E
and N onto the u-axis, which is rotated anti-clockwise from the E-axis by an angle φ

u = E cos φ + N sin φ u
(8.1)
= f ( E, N )
N sin φ
os φ
Ec E
φ

Figure 8.1

© 2005, R.E. Deakin Notes on Least Squares (2005) 8–1

RMIT University Geospatial Science

Applying the law of propagation of variances to equation (8.1) gives an expression for
variance in the u-direction

⎛ ∂f ⎞ 2 ⎛ ∂f ⎞ ∂f ∂f
2 2

s =s ⎜
2 2
⎟ + sN ⎜ ⎟ + 2 sEN (8.2)
⎝ ∂E ⎠ ⎝ ∂N ⎠ ∂E ∂N
u E

∂f ∂f
The partial derivatives = cos φ , = sin φ are obtained from (8.1) to give an equation for
∂E ∂N
the variance su2 in a direction φ (positive anti-clockwise) from the E-axis.

su2 = sE2 cos2 φ + sN2 sin 2 φ + 2 sEN cos φ sin φ (8.3)

Equation (8.3) defines the pedal curve of the Standard Error Ellipse

P
•
A
• tang
ent
al
m
mino

nor

φ axis a
r

O θ E
r
majo
b is

Pedal curve
ax

Ellipse

0 1 2
Scale of units

Figure 8.2. The pedal curve of the Standard Error Ellipse

In Figure 8.2, A is a point on an ellipse. The tangent to the ellipse at A intersects a normal to
the tangent passing through O at P. As A moves around the ellipse, the locus of all points P is

© 2005, R.E. Deakin Notes on Least Squares (2005) 8–2

RMIT University Geospatial Science

the pedal curve of the ellipse. The distance OP = su2 for the angle φ . The maximum and

minimum values of su2 define the directions and lengths of the axes of the ellipse and the

following section details the equations linking variances sE2 , sN2 and covariance sEN with the

Standard Error Ellipse parameters a, b and θ .

8.2. Parameters of the Standard Error Ellipse

Equation (8.3) has maximum and minimum values defining the lengths and directions of the
axes of the error ellipse. To determine these values from (8.3) the trigonometric identities
1 − cos 2φ = 2sin 2 φ , 1 + cos 2φ = 2 cos2 φ , sin 2φ = 2sin φ cos φ can be used to give

1 2 1
su2 = sE (1 + cos 2φ ) + sN2 (1 − cos 2φ ) + sEN sin 2φ
2 2
= ( sE2 + sN2 ) + ( sE2 − sN2 ) cos 2φ + sEN sin 2φ
1 1
2 2

Letting A = (
1 2
2
sE − sN2 ) and B = sEN this expression has the general form

su2 =
2
(
1 2
sE + sN2 ) + A cos 2φ + B sin 2φ (8.4)

Equation (8.4) can be expressed as a trigonometric addition

su2 =
2
( sE + sN2 ) + R cos ( 2φ − α )
1 2

= ( sE2 + sN2 ) + R cos 2φ cos α + R sin 2φ sin α

1
(8.5)
2

Equating the coefficients of cos 2φ and sin 2φ in equations (8.4) and (8.5) gives R cos α = A
and R sin α = B from which we obtain

R= A2 + B 2

(s − sN2 ) + ( sEN )
2
= 1 2 2
4 E

1
( sE2 − sN2 ) + 4 ( sEN )
2
=
2

2
1
= W (8.6)
2

© 2005, R.E. Deakin Notes on Least Squares (2005) 8–3

RMIT University Geospatial Science

(s − sN2 ) + 4 ( sEN )
2
W= 2 2
where E (8.7)

and the angle α from

B 2s
tan α = = 2 EN 2 (8.8)
A sE − s N

Inspection of equation (8.5) shows that su2 =

2
( sE + sN2 ) + R cos ( 2φ − α ) will have a
1 2

maximum value when ( 2φ − α ) = 0 i.e., cos ( 0 ) = 1 and a minimum value when ( 2φ − α ) = π

i.e., cos (π ) = −1 or

su2 ( max ) = (
1 2
2
sE + sN2 ) + R = (
1 2
2
sE + sN2 + W )
(8.9)
su2 ( min ) = ( sE2 + sN2 ) − R = ( sE + sN2 − W )
1 1 2
2 2

Inspection of Figure 8.2 shows that the maximum and minimum values of su2 are in the
directions of the major and minor axes of the Standard Error Ellipse and the semi-axes lengths
are

a=
2
( sE + sN2 + W )
1 2
(8.10)
b=
2
( sE + sN2 − W )
1 2

The value of φ when su2 is a maximum is when ( 2φ − α ) = 0 , i.e., when α = 2φ thus from

equation (8.8), letting θ = φ when su2 is a maximum, the angle θ , measured anti-clockwise
from the E-axis to the major axis of the Standard Error Ellipse, is given by

2 sEN
tan 2θ = (8.11)
s − sN2
2
E

Note that su2 is a minimum when 2φ − α = π , i.e., when α = 2φ − π thus from equation (8.8),

letting θ = φ and recognizing that tan ( x − π ) = tan x , then tan 2θ = 2 sEN (s 2

E − sN2 ) which is

the same equation for the angle to the major axis. Hence, it is not possible to distinguish
between the angles to the major or minor axes and the ambiguity must be resolved by using
equation (8.3).

© 2005, R.E. Deakin Notes on Least Squares (2005) 8–4

RMIT University Geospatial Science

Alternatively, the parameters of the Standard Error Ellipse can be determined from equation
(8.3) by the methods outlined in Chapter 2 (Sction 2.7.2 Least Squares Best Fit Ellipse).
Consider equation (8.3) expressed as

f = sE2 cos2 φ + sN2 sin 2 φ + 2 sEN cos φ sin φ (8.12)

and the aim is to find the maximum and minimum values of f (the maximum and minimum
variances) and the values of φ when these occur by investigating the first and second
derivatives f ′ and f ′′ respectively, i.e.,

⎧ max ⎫ ⎧ f ′ = 0 and f ′′ < 0 ⎫

f is ⎨ ⎬ when ⎨ ⎬
⎩ min ⎭ ⎩ f ′ = 0 and f ′′ > 0 ⎭

f ′ = ( sN2 − sE2 ) sin 2φ + 2 sEN cos 2φ

where (8.13)
f ′′ = 2 ( sN2 − sE2 ) cos 2φ − 4 sEN sin 2φ

Now the maximum or minimum value of f occurs when f ′ = 0 and from the first member of
(8.13) the value of φ is given by

2 sEN
tan 2φ = (8.14)
s − sN2
2
E

But this value of φ could relate to either a maximum or a minimum value of f. So from the
second member of equations (8.13) with a value of 2φ from equation (8.14) this ambiguity
can be resolved by determining the sign of the second derivative f ′′ since it is known that

⎧ f max ⎫ ⎧ f ′′ < 0 ⎫
⎨ ⎬ when ⎨ ⎬
⎩ f min ⎭ ⎩ f ′′ > 0 ⎭

In the equation of the pedal curve of the Standard Error Ellipse given by equation (8.12) f max
2
coincides with smax 2
and f min coincides with smin so the angle θ (measured positive anti-
clockwise) from the E-axis to the major axis of the ellipse (see Figure 8.2) is found from

⎧ smax
2
⎫ ⎧ f ′′ < 0 ⎫ ⎧θ = φ ⎫
⎨ 2 ⎬ when ⎨ ⎬ and ⎨ ⎬
⎩ f ′′ > 0 ⎭ ⎩θ = φ − 2 π ⎭
1
⎩ smin ⎭

Substituting φ = θ and φ = θ + 12 π into equation (8.3) will give the max. and min. values of
the variance which are the lengths of the semi axes a and b of the Standard Error Ellipse.

© 2005, R.E. Deakin Notes on Least Squares (2005) 8–5

RMIT University Geospatial Science

8.3. Example Computation

In Figure 8.2, sE2 = 6.0, sN2 = 2.0 and sEN = 1.2

The lengths of the semi-axes of the Standard Error Ellipse are

(s − sN2 ) + 4 ( sEN )
2
W= 2 2
E

= 42 + 4 (1.2 )
2

= 4.6648

a=
2
( sE + sN2 + W ) = 2.5164
1 2

b=
2
( sE + sN2 − W ) = 1.2914
1 2

The angle between the E-axis and the major axis (positive anti-clockwise), noting the
quadrant signs to determine the proper quadrant of 2θ

2 sEN 2 (1.2 ) ⎛ + ⎞
tan 2θ = = =⎜ ⎟
sE2 − sN2 6−2 ⎝+⎠
2θ = 30D 57′ 50′′
θ = 15D 28′ 55′′

Substituting the values φ = θ = 15D 28′ 55′′ and φ = 90D + θ = 105D 28′ 55′′ into equation (8.3)

gives su = 2.5164 and su = 1.2914 respectively so θ = 15D 28′ 55′′ is the angle (positive anti-
clockwise) between the E-axis and the major axis. Hence the bearing of the major axis is
90D − θ = 74D 31′ 05′′

© 2005, R.E. Deakin Notes on Least Squares (2005) 8–6

RMIT University Geospatial Science

Alternatively, using the method of evaluating the second derivative we have from equation
(8.14)

2 sEN 2 (1.2 ) ⎛ + ⎞
tan 2φ = = =⎜ ⎟
s − sN
2
E
2
6−2 ⎝+⎠
2φ = 30D 57′ 50′′
φ = 15D 28′ 55′′

The second derivative, from the second member of equations (8.13) is

f ′′ = 2 ( sN2 − sE2 ) cos 2φ − 4 sEN sin 2φ

= 2 ( 2 − 6 ) cos ( 30D 57′ 50′′) − 4 (1.2 ) sin ( 30D 57′ 50′′)
= −9.3295

Now, since f ′′ < 0 then θ = φ = 15D 28′ 55′′ is the angle (positive anti-clockwise) from the E-

axis to the major axis of the ellipse. The bearing of the major axis is 90D − θ = 74D 31′ 05′′ .

The Standard Error Ellipse semi-axes lengths a and b are obtained from equation (8.3) with
φ = θ = 15D 28′ 55′′ and φ = θ + 12 π = 105D 28′ 55′′ respectively giving

φ = 15D 28′ 55′′ 2

smax = 6.3324 and a = smax
2
= 2.5164

φ = 105D 28′ 55′′ 2

smin = 1.6677 and b = smin
2
= 1.2914

© 2005, R.E. Deakin Notes on Least Squares (2005) 8–7

RMIT University Geospatial Science

8.4. Some Examples of Resections and Error Ellipses

D
B

C
A
B

Figure 8.3. Resection 1 Figure 8.4. Resection 2

In Resection 1 (Figure 8.3) the error ellipse indicates a strong position fix and the observed
stations are spread through an arc of approximately 200°.

In Resection 2 (Figure 8.4) the error ellipse indicates a poor position fix. The observed
stations lay in a small arc of approximately 35°.

© 2005, R.E. Deakin Notes on Least Squares (2005) 8–8

RMIT University Geospatial Science

8.5. Some Examples of Orientation and Shape of Ellipses

N N N

θ E E E E

1 2 3 4

N
N
N N

θ E
θ E E E
θ
θ

5 6 7 8

Ellipse sE2 sN2 sEN ρ EN a b θ

1 4 4 2 0.5 2.45 1.41 45°

2 4 4 -2 -0.5 2.45 1.41 -45°
3 16 4 0 0 4 2 0°
4 4 16 0 0 4 2 90°
5 16 4 6 0.75 4.30 1.23 22° 30′
6 4 16 6 0.75 4.30 1.23 67° 30′
7 4 16 -6 -0.75 4.30 1.23 -67° 30′
8 16 4 -6 -0.75 4.3 1.23 -22° 30′

sEN
Note: ρ EN = is the correlation coefficient −1 ≤ ρ EN ≤ 1
sE s N

© 2005, R.E. Deakin Notes on Least Squares (2005) 8–9

RMIT University Geospatial Science

8.6. References

Additional information on error ellipses can be found in the following references.

Cooper, M.A.R., 1982. Fundamentals of Survey Measurement and Analysis, Granada,

London.

Mikhail, E.M., 1976. Observations and Least Squares, IEP–A Dun-Donnelley, New York.

Mikhail, E.M. and Gracie, G., 1981. Analysis and Adjustment of Survey Measurements,
Van Nostrand Reinhold Company, New York.

Richardus, P., 1966. Project Surveying, North-Holland Publishing Company, Amsterdam.

© 2005, R.E. Deakin Notes on Least Squares (2005) 8–10

RMIT University Geospatial Science

9. LEAST SQUARES RESECTION

A resection is a mathematical technique for determining the plane coordinates of an

"unknown" point by observing theodolite (or Total Station) directions to three or more
"known" points; known points being those whose coordinates are known. Directions to three
known points are the minimum requirement for a solution, which may be obtained from
geometric principles. There are many techniques for geometric solution, two of which have
been studied in the practical projects associated with this course; (i) the Collins Point method
and (ii) the method of Auxiliary angles.

In the case of four or more observed directions to known points, the method of least squares
(least squares adjustment of indirect observations) may be employed to obtain the best
estimates of the coordinates of the resected point. This technique requires the formation of a
set of observation equations that yield "normal equations" that are solved for the best
estimates of the coordinates of the resected point. Owing to the nature of the observation
equation, which is a linearized approximation, the least squares solution process is iterative.
That is, approximate values are assumed, corrections computed and approx values updated;
with the process repeated until the corrections to approximate vales become negligible. This
least squares technique is often called Variation of Coordinates.

9.1. The Resection Observation Equation

In the case of four or more observed directions to known points, the method of least squares
may be employed to obtain the best estimates of the coordinates of the resected point. This
technique requires the formation of a set of observation equations; each equation based on a
linearized form of the following equation whose elements are shown in Figure 9.1. (see
Chapter 7, Section 7.3 for details regarding the linearization process using Taylor's theorem).

⎛ Ek − E P ⎞
α k + vk + z = tan −1 ⎜ ⎟ (9.1)
⎝ Nk − NP ⎠

© 2005, R.E. Deakin Notes on Least Squares (2005) 9–1

RMIT University Geospatial Science

N'
P3 Ek - Ei Pk




Nk - Ni
sik
α i3 αik
si3 φik
E'
zi


α i2 Pi
s i2 α i1
s i1
P2 P1 (RO)



N

Figure 9.1 Observed directions α and distances s from Pi

where
αk are the observed directions from the resection point P to the known points Pk ,

vk are the residuals (small corrections) associated with observed directions,

z is an orientation "constant"; the bearing of the Reference Object (RO) for the
set of observed directions,
Ek , N k are the east and north coordinates of the known points, and

EP , N P are the east and north coordinates of the resection point.

Equation (9.1) is a non-linear equation, which has a linear approximation of the form

α k + vk + z 0 + Δz = − ak ΔN P − bk ΔEP + φk0 (9.2)

where ak , bk are direction coefficients (see Chapter 7, Section 7.3)

− ( Ek − EP0 ) − sin φk0

ak = =
(s )
0 2
k
sk0
(9.3)
N k − N P0 cos φk0
bk = =
(s ) 0 2
k
sk0

© 2005, R.E. Deakin Notes on Least Squares (2005) 9–2

RMIT University Geospatial Science

ΔE , ΔN , Δz are small corrections to approximate coordinates EP0 , N P0 of the resected point

and the approximate orientation constant z 0 such that

EP = EP0 + ΔE
N P = N P0 + ΔN (9.4)
z = z 0 + Δz

φk0 , sk0 are approximate bearings and distances obtained by substituting the approximate

coordinates EP0 , N P0 into

⎛ Ek − EP0 ⎞
φk0 = tan −1 ⎜ 0 ⎟
(9.5)
⎝ Nk − NP ⎠

(E − EP0 ) + ( N k − N P0 )
2 2
sk0 = k (9.6)

9.2. Formation of the Observation Equations into Standard Matrix Form

The observation equation (9.2) can be re-arranged into a "standard" form

vk + ak ΔN + bk ΔE + Δz = φk0 − (α k + z 0 ) (9.7)

A set of n such equations can be represented in matrix form as

⎡φ10 − (α1 + z 0 ) ⎤
⎡ v1 ⎤ ⎡ a1 b1 1⎤
⎡ ΔN ⎤ ⎢ ⎥
⎢v ⎥ ⎢a b2 1⎥ ⎢ ⎥ ⎢φ20 − (α 2 + z 0 ) ⎥
⎢ 2⎥ + ⎢ 2 ⎥ ΔE = ⎢ ⎥
⎢#⎥ ⎢# #⎥ ⎢ ⎥ ⎢ # ⎥
⎢ ⎥ ⎢ ⎥ ⎢ Δz ⎥
bn 1⎦ ⎣ ⎦ ⎢φ 0 − α + z 0 ⎥
⎣ vn ⎦ ⎣ a n ⎢⎣ n ( n )⎥⎦
or v + Bx = f (9.8)

where n is the number of observed directions (= to the number of equations)

u is the number of "unknowns" (in the case of a resection u = 3)
v is an (n,1) vector of residuals
B is an (n,u) coefficient matrix containing the direction coefficients and ones
x is a (u,1) vector of "unknowns" which are corrections to approximate
coordinates and orientation constant

© 2005, R.E. Deakin Notes on Least Squares (2005) 9–3

RMIT University Geospatial Science

f is an (n,1) vector of numeric terms which are "computed bearing − observed

bearing"

9.3. Formation of the Normal Equations and Solution of the Unknowns

In any least squares adjustment, every measurement (or observation) has an associated
precision (a variance) and a measure of connection with every other measurement
(covariances). These statistics are contained in a covariance matrix Σ . The elements of a
covariance matrix are population statistics and in practice, the covariance matrix is estimated
a priori by a cofactor matrix Q. Covariance matrices and cofactor matrices are related by
Σ = σ 02Q where σ 02 is the variance factor. An estimate of the variance factor σˆ 02 may be
computed after the adjustment. In least squares theory it is often useful to express the relative
precision of observations in terms of weights, where a weight is defined as being inversely
proportional to a variance, this leads to the definition of a weight matrix as the inverse of a
cofactor matrix, i.e. W = Q −1 . Weight matrices, covariance matrices and cofactor matrices
are square and symmetric (see Chapter 2, Section 2.5).

Applying the least squares principle to equation (9.8) with the precisions of the observations
estimated by a weight matrix leads to a set of normal equations of the form

( B WB ) x = B Wf
T T

or Nx = t (9.9)

where W is an (n,n) weight matrix

N = BT WB is a (u,u) coefficient matrix of the set of normal equations
t = BT Wf is a (u,1) vector of numeric terms

The solution of the n = 3 unknowns is

⎡ ΔN ⎤
x = ⎢ ΔE ⎥ = N −1t (9.10)
⎢ ⎥
⎢⎣ Δz ⎥⎦

© 2005, R.E. Deakin Notes on Least Squares (2005) 9–4

RMIT University Geospatial Science

9.4. The form of the Coefficient Matrix N and the Vector of Numeric Terms t

For n observations, the coefficient matrix B and the vector of numeric terms f have the
following form

⎡φ 0 − (α1 + z 0 ) ⎤
⎡ a1 b1 1⎤ ⎡ f1 ⎤ ⎢ 1 ⎥
⎢a b2 1⎥ ⎢ f ⎥ ⎢φ20 − (α 2 + z 0 ) ⎥
⎢ 2 ⎥ ⎢ 2⎥ ⎢ ⎥
B = ⎢ a3 b3 1⎥ , f = ⎢ f 3 ⎥ = ⎢φ30 − (α 3 + z 0 ) ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢# ⎥ ⎢#⎥ ⎢ # ⎥
⎢⎣ an 1⎥⎦ ⎢⎣ f n ⎥⎦ ⎢ 0 0 ⎥
⎣⎢φn − (α n + z )⎦⎥
bn

Each observed direction has an associated variance σ 2 estimated by s 2 where σ is the

standard deviation (a population statistic) and s is its estimate. The observed directions are
assumed independent, hence the covariances between observations are zero and the
covariance matrix, cofactor matrix and weight matrix are all diagonal matrices. The cofactor
matrix Q and the weight matrix W have the following form

⎡ s12 0 0 " 0⎤
⎢ ⎥
⎢0 s22 0 " 0⎥
Q =⎢0 0 s32 " 0⎥
⎢ ⎥
⎢# # # % #⎥
⎢⎣ 0 0 0 " sn2 ⎥⎦
(9.11)
⎡1 s12 0 0 " 0 ⎤ ⎡ w1 0 0 " 0⎤
⎢ ⎥
⎢ 0 1 s2
2
0 " 0 ⎥ ⎢⎢ 0 w2 0 " 0⎥
⎥
W = Q −1 =⎢ 0 0 1 s32 " 0 ⎥=⎢0 0 w3 " 0⎥
⎢ ⎥ ⎢ ⎥
⎢ # # # % # ⎥ ⎢# # # % # ⎥
⎢⎣ 0 0 0 " 1 sn2 ⎥⎦ ⎢⎣ 0 0 0 " wn ⎥⎦

The normal equation coefficient matrix N and vector of numeric terms t have the following
form

⎡ n n n
⎤ ⎡ n ⎤
⎢ ∑ wk ak ∑ wk ak bk ∑ ⎢ ∑ wk ak f k ⎥
2
w a
k k⎥
⎢ k =1 k =1 k =1
⎥ ⎢ k =1 ⎥
⎢ n n n
⎥ ⎢ n
⎥
N = BT WB = ⎢ ∑ wk ak bk ∑w b 2
k k ∑ wk bk ⎥ , t = BT Wf = ⎢ ∑ wk bk f k ⎥ (9.12)
⎢ k =1 k =1 k =1
⎥ ⎢ k =1 ⎥
⎢ n n n
⎥ ⎢ n ⎥
⎢ ∑ wk ak ∑w b k k ∑ wk ⎥ ⎢ ∑ wk f k ⎥
⎣ k =1 k =1 k =1 ⎦ ⎣ k =1 ⎦

© 2005, R.E. Deakin Notes on Least Squares (2005) 9–5

RMIT University Geospatial Science

Note that if each element of the coefficient matrix B and vector of numeric terms f is divided
by the appropriate estimate of the standard deviation augmented matrices B and f can be
formed

⎡ (φ10 − α1 + z 0 ) s1 ⎤
⎡ a1 s1 b1 s1 1 s1 ⎤ ⎡ f1 s1 ⎤ ⎢ ⎥
⎢a s b2 s2 1 s2 ⎥ ⎢ f s ⎥ ⎢ (φ20 − α 2 + z 0 ) s2 ⎥
⎢ 2 2 ⎥ ⎢ 2 2⎥ ⎢ ⎥
B = ⎢ a3 s3 b3 s3 1 s3 ⎥ , f = ⎢ f 3 s 3 ⎥ = ⎢ (φ30 − α 3 + z 0 ) s3 ⎥ (9.13)
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ # ⎥ ⎢ # ⎥ ⎢ # ⎥
⎢⎣ an sn bn sn 1 sn ⎥⎦ ⎢⎣ f n sn ⎥⎦ ⎢ 0 ⎥
⎢⎣(φn − α n + z )
0
sn ⎥⎦

and the normal equation coefficient matrix N and vector of numeric terms t are given by

N = BT WB = BT B
(9.14)
t = BT Wf = BT f

9.5. The Iterative Solution and Assessment of Precision

The elements of the solution vector x = [ ΔN ΔE Δz ] are corrections to the approximate

T

coordinates EP0 , N P0 and the approximate orientation constant z 0 . These corrections are
added to the approximate values to obtain updated values of the approximations and another
iteration performed.

When the corrections for the kth iteration reach some desired value, say less than 0.5 mm, then
the current "approximate" values may be regarded as exact and the solution is complete.

At the end of the iterative process, the residuals v are computed from equation (9.8) and an
assessment of the "quality" of the observations can be made. Large residuals may indicate
poor observations.

An estimate of the variance factor σˆ 02 can be computed from the residuals after the
adjustment process by using the following equation (see Chapter 5, equation (5.25))

v T Wv f T Wf − xT t
σˆ 02 = = (9.15)
n−u n−u

© 2005, R.E. Deakin Notes on Least Squares (2005) 9–6

RMIT University Geospatial Science

Note that in (9.15) the variance factor can be computed without calculating the residuals.

The following pages show an example resection. Page 8 shows a diagram of the resection
observations to four known stations. Page 9 shows a geometric solution for the coordinates of
P using the Collins Point technique and pages 10 to 14 show the calculations required for a
least squares solution using all four observations.

© 2005, R.E. Deakin Notes on Least Squares (2005) 9–7

RMIT University Geospatial Science

9.6. Example Resection: Diagram and Observations

EPIPHANY

STUDLEY PARK
ST JOHNS

N
P

GOVERNMENT HOUSE

STATION COORDINATES OBSERVED DIRECTION

Government House 321862.876 E
0° 00′ 00″
(Flagpole) 5811188.930 N
Saint. Johns 322731.700 E
87° 09′ 09″
(Spire) 5815369.270 N
Epiphany 323590.140 E
134° 40′ 36″
(Flagpole) 5816974.280 N
Studley Park 325526.582 E
201° 48′ 52″
(Beacon) 5815551.657 N

© 2005, R.E. Deakin Notes on Least Squares (2005) 9–8

RMIT University Geospatial Science

9.7. Example Resection: Collins Point Solution

EPIPHANY
323590.140 E
5816974.280 N

O 9′ 03.
26″
55 1 5666
71
1740. 325526.582 E
STUDLEY PARK 5815551.657 N
324095.159 E β
5814561.142 N P
α δ
γ
δ 15 17′ 41.62″
O
95 26″

4″
.11 ′ 11.

.6
48

2 1 4
1′ 804
40 3° 30

0
° 0 2
40 97.
44
21

56
174° 41′ 57.6

6 31′ 10.38″
O
GOVERNMENT
14126.62222

γ
168

HOUSE
°1

321862.876 E
0′ 47

5811188.930 N
4″
.

3
26″

15
2
10 ° 53
90 ′ 0
1. 5.
65 64
62 ″
1

180° − (γ + δ)

COLLINS POINT
326831.627 E
5801485.445 N

© 2005, R.E. Deakin Notes on Least Squares (2005) 9–9

RMIT University Geospatial Science

9.8. Example Resection: Least Squares Solution

The data for this resection is shown on the preceding pages. The coordinates of P are
computed from the Collins Point resection and are rounded to the nearest 0.1 metres.

Observations and computed data

Coordinates Observed Directions Computed Bearing and Distance

Station East North Deg Min Sec Deg Min Sec Dist (m)
P 324095.200 5814561.100
Gov't House 321862.876 5811188.930 0 00 00 213 30 14.18 4044.107
St. Johns 322731.700 5815369.270 87 09 09 300 39 21.44 1585.015
Epiphany 323590.140 5816974.280 134 40 36 348 10 44.62 2465.466
Studley Park 325526.582 5815551.657 201 48 52 55 18 56.40 1740.706

Computation of Direction coefficients and numeric terms for the observation equation

vk + ak ΔN + bk ΔE + Δz = φk0 − (α k + z 0 ) (9.16)

Note 1: In equation (9.16) the numeric terms on the right-hand-side are computed bearing -
"observed" bearing. The observed bearings are obtained by adding the observed
directions to the computed bearing of the RO (Government House).
Note 2: The dimensions (or units) of the numeric terms on the right-hand-side of (9.16) are
seconds of arc. This means that the elements on the left-hand-side must have
consistent dimensions, i.e. vk (seconds), ak , bk (seconds/length), ΔE , ΔN (length)
and Δz (seconds). If the corrections to approximate coordinates are expressed in
centimetres (cm), then the direction coefficients have dimensions of sec/cm and
equations (9.3) are

− ( Ek − EP0 ) − sin φk0

ak = × ρ ′′ = × ρ ′′
(s )
0 2
k
sk0
(9.17)
N k − N P0 cos φk0
bk = × ρ ′′ = × ρ ′′
(s )
0 2
k
sk0

180
where distances and coordinate differences are in cm's and ρ ′′ = × 3600 (seconds
π
in one radian).

© 2005, R.E. Deakin Notes on Least Squares (2005) 9–10

RMIT University Geospatial Science

Direction coefficients Computed Bearing Observed Bearing Numeric term

Station a (sec/cm) b (sec/cm) Deg Min Sec Deg Min Sec comp-obs (sec)
Gov't House 0.281538 -0.425294 213 30 14.18 213 30 14.18 0.00
St. Johns 1.119473 0.663531 300 39 21.44 300 39 23.18 -1.74
Epiphany 0.171384 0.818873 348 10 44.62 348 10 50.18 -5.56
Studley Park -0.974383 0.674301 55 18 56.40 55 19 06.18 -9.78

Coefficient matrix B, vector of numeric terms f and normal equation coefficient matrix
N = BT WB

ΔN ΔE Δz
0.281538 -0.425294 1 0.00
1.119473 0.663531 1 -1.74
B= f=
0.171384 0.818873 1 -5.56
-0.974383 0.674301 1 -9.78

In this example, the observations are assumed to be of equal precision. In such cases the
weight matrix W can be replaced by the Identity matrix I, and the normal equation coefficient
matrix N = BT WB = BT IB = BT B and the vector of numeric terms t = BT Wf = BT If = BT f

2.311280 0.106383 0.598012 6.630754

N = BT B 0.106383 1.746384 1.731411 t = BT f = -12.302480
0.598012 1.731411 4.000000 -17.081376

Inverse of normal equation coefficient matrix N −1 and solution vector x = N −1t

0.454844 0.069562 -0.098111 3.8360 (cm)

N −1 = 0.069562 1.013710 -0.449187 x = N − 1t = -4.3372 (cm)
-0.098111 -0.449187 0.459100 -2.9665 (sec)

© 2005, R.E. Deakin Notes on Least Squares (2005) 9–11

RMIT University Geospatial Science

Corrections to approximate coordinates and correction to approximate orientation

constant

ΔE = −4.3372 cm = −0.043 m , ΔN = 3.8360 cm = 0.038 m and Δz = −2.9665 sec = −3.0 sec

Adjusted coordinates of P and residuals (after one iteration)

EP = EP0 + ΔE = 324095.200 − 0.043 = 324095.157 m

N P = N P0 + ΔN = 5814561.100 + 0.038 = 5814561.138 m

Station Residual
Government House 0.04 (sec)
St. Johns -0.19
Epiphany 0.30
Studley Park -0.15

Variance factor σ 02 and precision of computed coordinates

An estimate of the variance factor σˆ 02 can be computed from equation (9.15) with W = I,
n = 4, u = 3 and

vT v f T f − xT t
σˆ 02 = = = 129.620678 − 129.465415 = 0.155263 sec 2
n−u 1

Assuming all the observations are of equal precision and letting W = I is equivalent to
assigning an estimated standard deviation of 1 second to each observation. Inspection of the
variance factor shows that if a standard deviation of 0.39 sec ( 0.39 = 0.155263 ) was used as
an estimate of the standard deviation of the observed directions, the estimate of the variance
factor computed from the adjustment would have been approximately unity (a variance factor
of unity indicates that the estimates of variances are close to the population statistics).

© 2005, R.E. Deakin Notes on Least Squares (2005) 9–12

RMIT University Geospatial Science

From this adjustment, we may conclude that the standard deviation of the observed directions
was approximately 0.4 sec.

A most important "by-product" of a least squares adjustment is the ability to estimate the
precision of the computed quantities. Theory shows that this information is contained in the
inverse of the normal equations and the covariance matrix of the computed quantities Σ xx is
given by

⎡ 0.454844 0.069562 −0.098111⎤ ⎡ 0.070 0.011 −0.015⎤

2 2 −1 ⎢ ⎥ ⎢
Σ xx = σ Q xx = σ N = 0.155263 0.069562 1.013710 −0.449187 = 0.011 0.157 −0.070⎥
0 0 ⎢ ⎥ ⎢ ⎥
⎣⎢ −0.098111 −0.449187 0.459100 ⎦⎥ ⎣⎢ −0.015 −0.070 0.071 ⎥⎦

and the standard deviation of the adjusted coordinates are

σ E = 0.157 cm = 0.004 m
σ N = 0.070 cm = 0.003 m

Parameters of Standard Error Ellipse

From the variance matrix Σ xx above we have σ E2 = 0.157, σ N2 = 0.070 and σ EN = 0.011

Using the formulae from Chapter 8, Section 8.3 (replacing s with σ ), the lengths of the semi-
axes of the Standard Error Ellipse are

(σ − σ N2 ) + 4 (σ EN )
2
W= 2 2
E

= ( 0.157 − 0.070) + 4 ( 0.011)

2 2

= 0.0897

a = (
1 2
2
σ E + σ N2 + W ) = 0.3980 cm

b = (
1 2
2
sE + sN2 − W ) = 0.2620 cm

© 2005, R.E. Deakin Notes on Least Squares (2005) 9–13

RMIT University Geospatial Science

The angle between the E-axis and the major axis (positive anti-clockwise), noting the
quadrant signs to determine the proper quadrant of 2θ

2σ EN 2 ( 0.011) ⎛+⎞
tan 2θ = = =⎜ ⎟
σ E − σ N 0.157 − 0.070 ⎝ + ⎠
2 2

2θ = 14D 11′ 28′′

θ = 7D 05′ 44′′

Substituting the values φ = θ = 7D 05′ 44′′ and φ = 90D + θ = 97D 05′ 44′′ into equation (8.3)

gives su = 0.3980 cm and su = 0.2620 cm respectively so θ = 7D 05′ 44′′ is the angle (positive

anti-clockwise) between the E-axis and the major axis. Hence the bearing of the major axis is
90D − θ = 82D 54′16′′

Figure 9.2 shows a schematic diagram of the example resection and the Standard Error Ellipse

EPIPHANY

STUDLEY PARK
ST JOHNS

GOVERNMENT HOUSE

Figure 9.2 Schematic diagram of resection and Standard Error Ellipse

© 2005, R.E. Deakin Notes on Least Squares (2005) 9–14

RMIT University Geospatial Science

10. LEAST SQUARES BEARING INTERSECTION

Intersection is a mathematical technique for determining the plane coordinates of an

"unknown" point by observing bearings (clockwise angles from North) to the unknown point
from two or more "known" points; known points being those whose coordinates E,N are
known. Figure 10.1 shows an unknown point P intersected by bearings φ A , φ B , φC , φ D

observed from known stations A, B, C and D. Bearings from two known points are the
minimum requirement for a solution, which may be obtained from geometric principles set
out below. In the case of three or more observed bearings from known points, the method of
least squares may be employed to obtain the best estimates of the coordinates of the
intersected point P. This technique requires the formation of a set of observation equations
that yield "normal equations" that are solved for the best estimates of the coordinates of P.
Owing to the nature of the observation equation, which is a linearized approximation, the least
squares solution process is iterative. That is, approximate values are assumed, corrections
computed and approx values updated; with the process repeated until the corrections to
approximate vales become negligible.

N N

φA B
A EP − EA φB
NP − NA

P
· N

N C

φD φC

N D

Figure 10.1. Unknown point P intersected by bearings from known points A,B,C and D

© 2005, R.E. Deakin Notes on Least Squares (2005) 10–1

RMIT University Geospatial Science

10.1. Position from Two Observed Bearings

From Figure 10.1, using the bearings φ A , φ B to the unknown point P from known points A and

B with coordinates E A , N A and EB , N B the following two equations can be obtained

EP − E A
tan φ A =
NP − N A
(10.1)
E − EB
tan φ B = P
NP − NB

Expanding these equations and re-arranging gives

EP = N P tan φ A − N A tan φ A + E A
(10.2)
EP = N P tan φ B − N B tan φ B + E B

Equating equations (10.2) gives a solution for N P

N A tan φ A − N B tan φ B + EB − E A
NP = (10.3)
tan φ A − tan φ B

Having obtained a solution for N P from (10.3) then EP can be obtained from either of
equations (10.2).

It should be noted that if P lies on the line between A and B then its position is indeterminate.

10.2. The Least Squares Bearing Intersection Observation Equation

In the case of four or more observed directions to known points, the method of least squares
may be employed to obtain the best estimates of the coordinates of the resected point. This
technique requires the formation of a set of observation equation; each equation based on a
linearized form of the following equation whose elements are shown in Figure 10.1.

⎛ E P − Ek ⎞
φk + vk = tan −1 ⎜ ⎟ (10.4)
⎝ NP − Nk ⎠

where φk are the observed bearings known points Pk to the unknown point P,

vk are the residuals (small corrections) associated with observed bearings,

Ek , N k are the east and north coordinates of the known points, and

© 2005, R.E. Deakin Notes on Least Squares (2005) 10–2

RMIT University Geospatial Science

EP , N P are the east and north coordinates of the intersection point P.

Equation (10.4) is a non-linear equation, which has a linear approximation of the form

φk + vk = ak ΔN P + bk ΔEP + φk0 (10.5)

where ak , bk are direction coefficients (see Chapter 7, Section 7.3)

− ( EP0 − Ek ) − sin φk0

ak = =
(s ) 0 2
k
sk0
(10.6)
N P0 − N k cos φk0
bk = =
(s ) 0 2
k
sk0

ΔE , ΔN are small corrections to approximate coordinates EP0 , N P0 of the intersected point

such that

EP = EP0 + ΔE
(10.7)
N P = N P0 + ΔN

φk0 , sk0 are approximate bearings and distances obtained by substituting the approximate

coordinates EP0 , N P0 into

⎛ EP0 − Ek ⎞
φk0 = tan −1 ⎜ ⎟ (10.8)
⎝ NP − Nk ⎠
0

(E − Ek ) + ( N P0 − N k )
2 2
sk0 = 0
P (10.9)

10.3. Formation of the Observation Equations into Standard Matrix Form

The observation equation (10.5) can be re-arranged into a "standard" form

vk − ak ΔN − bk ΔE = φk0 − φk (10.10)

A set of n such equations can be represented in matrix form as

© 2005, R.E. Deakin Notes on Least Squares (2005) 10–3

RMIT University Geospatial Science

⎡ v1 ⎤ ⎡ −a1 −b1 ⎤ ⎡φ10 − φ1 ⎤

⎢ v ⎥ ⎢ −a −b2 ⎥ ⎡ ΔN ⎤ ⎢⎢φ20 − φ2 ⎥⎥
⎢ 2⎥ + ⎢ 2 ⎥ =
⎢#⎥ ⎢ # ⎥ ⎢⎣ ΔE ⎥⎦ ⎢ # ⎥
⎢ ⎥ ⎢ ⎥ ⎢ 0 ⎥
⎣ vn ⎦ ⎣ − a n −bn ⎦ ⎣⎢φn − φn ⎦⎥

or v + Bx = f (10.11)

where n is the number of observed bearings (= to the number of equations)

u is the number of "unknowns" (in the case of an intersection u = 2)
v is an (n,1) vector of residuals
B is an (n,u) coefficient matrix containing the direction coefficients
x is a (u,1) vector of "unknowns" which are corrections to approximate
coordinates
f is an (n,1) vector of numeric terms; "computed bearing − observed bearing"

10.4. Formation of the Normal Equations and Solution of the Unknowns

In any least squares adjustment, every measurement (or observation) has an associated
precision (a variance) and a measure of connection with every other measurement
(covariances). These statistics are contained in a covariance matrix Σ . The elements of a
covariance matrix are population statistics and in practice, the covariance matrix is estimated
a priori by a cofactor matrix Q. Covariance matrices and cofactor matrices are related by
Σ = σ 02 Q where σ 02 is the variance factor. An estimate of the variance factor σˆ 02 may be
computed after the adjustment. In least squares theory it is often useful to express the relative
precision of observations in terms of weights, where a weight is defined as being inversely
proportional to a variance, this leads to the definition of a weight matrix as the inverse of a
cofactor matrix, i.e. W = Q −1 . Weight matrices, covariance matrices and cofactor matrices
are square and symmetric (see Chapter 2, Section 2.5).

Applying the least squares principle to equation (10.8) with the precisions of the observations
estimated by a weight matrix leads to a set of normal equations of the form

(B T
WB ) x = BT Wf

or Nx = t (10.12)

© 2005, R.E. Deakin Notes on Least Squares (2005) 10–4

RMIT University Geospatial Science

where W is an (n,n) weight matrix

N = BT WB is a (u,u) coefficient matrix of the set of normal equations
t = BT Wf is a (u,1) vector of numeric terms

The solution of the n = 2 unknowns is

⎡ ΔN ⎤
x = ⎢ ⎥ = N −1t (10.13)
⎣ ΔE ⎦

10.5. The form of the Coefficient Matrix N and the Vector of Numeric Terms t

For n observations, the coefficient matrix B and the vector of numeric terms f have the
following form

⎡ −a1 −b1 ⎤ ⎡ f1 ⎤ ⎡φ10 − φ1 ⎤

⎢ −a −b2 ⎥ ⎢ f ⎥ ⎢φ 0 − φ ⎥
⎢ 2 ⎥ ⎢ 2⎥ ⎢ 2 2⎥

B = ⎢ − a3 −b3 ⎥ , f = ⎢ f 3 ⎥ = φ3 − φ3 ⎥
⎢ 0

⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ # ⎥ ⎢#⎥ ⎢ # ⎥
⎢⎣ − an −bn ⎥⎦ ⎢⎣ f n ⎥⎦ ⎢⎣φn0 − φn ⎥⎦

Each observed bearing has an associated variance σ 2 estimated by s 2 where σ is the

standard deviation and s is its estimate. The observed bearings are assumed independent,
hence the covariances between observations are zero and the covariance matrix, cofactor
matrix and weight matrix are all diagonal matrices. The cofactor matrix Q and the weight
matrix W have the following form

⎡ s12 0 0 " 0⎤
⎢ ⎥
⎢0 s22 0 " 0⎥
Q=⎢0 0 s32 " 0⎥
⎢ ⎥
⎢# # # % #⎥
⎢⎣ 0 0 0 " sn2 ⎥⎦

⎡1 s12 0 0 " 0 ⎤ ⎡ w1 0 0 " 0⎤

⎢ ⎥
⎢ 0 1 s2
2
0 " 0 ⎥ ⎢⎢ 0 w2 0 " 0⎥
⎥
and W=Q =⎢ 0
−1
0 1 s32 " 0 ⎥=⎢0 0 w3 " 0⎥ (10.14)
⎢ ⎥ ⎢ ⎥
⎢ # # # % # ⎥ ⎢# # # % # ⎥
⎢⎣ 0 0 0 " 1 sn2 ⎥⎦ ⎢⎣ 0 0 0 " wn ⎥⎦

© 2005, R.E. Deakin Notes on Least Squares (2005) 10–5

RMIT University Geospatial Science

The normal equation coefficient matrix N and vector of numeric terms t have the following
form

⎡ n n
⎤ ⎡ n ⎤
⎢ ∑ wk ak ∑ ⎢ ∑ wk ak f k ⎥
2
w a b
k k k⎥

N = BT WB = ⎢ n ⎥ , t = BT Wf = ⎢ k =1 ⎥
k =1 k =1
(10.15)
⎢ n
⎥ ⎢ n
⎥
⎢ ∑ wk ak bk ∑ wk bk2 ⎥ ⎢ ∑ wk bk f k ⎥
⎣ k =1 k =1 ⎦ ⎣ k =1 ⎦

Note that if each element of the coefficient matrix B and vector of numeric terms f is divided
by the appropriate estimate of the standard deviation augmented matrices B and f can be
formed

⎡ (φ 0 − φ1 ) s1 ⎤
⎡ −a1 s1 −b1 s1 ⎤ ⎡ f1 s1 ⎤ ⎢ 1 ⎥
⎢ −a s2 −b2 s2 ⎥ ⎢ f s ⎥ ⎢ (φ20 − φ2 ) s2 ⎥
⎢ 2 ⎥ ⎢ 2 2⎥ ⎢ ⎥
B = ⎢ − a3 s3 −b3 s3 ⎥ , f = ⎢ f 3 s 3 ⎥ = ⎢ (φ30 − φ3 ) s3 ⎥ (10.16)
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ # ⎥ ⎢ # ⎥ ⎢ # ⎥
⎢⎣ − an −bn sn ⎥⎦ ⎢⎣ f n sn ⎥⎦ ⎢ 0 ⎥
⎢⎣ (φn − φn )
sn
sn ⎥⎦

and the normal equation coefficient matrix N and vector of numeric terms t are given by

N = BT WB = BT B
(10.17)
t = BT Wf = BT f

10.6. The Iterative Solution and Assessment of Precision

The elements of the solution vector x = [ ΔN ΔE ] are corrections to the approximate

T

coordinates EP0 , N P0 . These corrections are added to the approximate values to obtain
updated values of the approximations and another iteration performed.

When the corrections for kth iteration reach some desired value, say less than 0.5 mm, then the
current "approximate" values may be regarded as exact and the solution is complete.

At the end of the iterative process, residuals v are computed from equation (10.11) and an
assessment of the "quality" of the observations can be made. Large residuals may indicate
poor observations.

© 2005, R.E. Deakin Notes on Least Squares (2005) 10–6

RMIT University Geospatial Science

The variance factor σ 02 can be computed from the residuals after the adjustment process by
using the following

v T Wv f T Wf − xT t
σ 02 = = (10.18)
n−u n−u

Note that in (10.18) the variance factor can be computed without calculating the residuals.

10.7. Example Intersection: Diagram and Observations

C φC

D
N φD
P
φB
B

N
φA

N A

STATION COORDINATES OBSERVED BEARING

12875.270 E
A 34° 47′ 52″
28679.600 N
12273.910 E
B 81° 01′ 23″
29612.310 N
14117.390 E
C 200° 40′ 18″
30999.980 N
14717.690 E
D 252° 09′ 35″
30168.700 N

© 2005, R.E. Deakin Notes on Least Squares (2005) 10–7

RMIT University Geospatial Science

10.8. Example Intersection: Least Squares Solution

The data for this intersection is shown on the previous page. The coordinates of P are
computed from equations (10.3) and (10.2) then rounded down to the nearest 0.1 metres.

Observations and computed data

Coordinates Observed Bearing Computed Bearing and Distance

Station East North Deg Min Sec Deg Min Sec Dist
P 13677.500 29834.000
A 12875.270 28679.600 34 47 52 34 47 48.08 1405.778
B 12273.910 29612.310 81 01 23 81 01 28.43 1420.990
C 14117.390 30999.980 200 40 18 200 40 12.06 1246.199
D 14717.690 30168.700 252 09 35 252 09 48.52 1092.712

Computation of Direction coefficients and numeric terms for the observation equation

vk − ak ΔN − bk ΔE = φk0 − φk (10.19)

Note 1 In equation (10.19) the numeric terms on the right-hand-side are computed bearing
minus observed bearing.
Note 2 The dimensions (or units) of the numeric terms on the right-hand-side of (10.19) are
seconds of arc. This means that the elements on the left-hand-side must have
consistent dimensions, i.e., vk (seconds), ak , bk (seconds/length) and ΔE , ΔN
(length). If the corrections to approximate coordinates are expressed in centimetres
(cm), then the direction coefficients have dimensions of sec/cm and equations (10.6)
are

− ( EP0 − Ek ) − sin φk0

ak = × ρ ′′ = × ρ ′′
(s )
0 2
k
sk0

N P0 − N k cos φk0
bk = × ρ ′′ = × ρ ′′
(s )
0 2
k
sk0

180
where distances and coordinate differences are in cm's and ρ ′′ = × 3600 (seconds
π
in one radian).

© 2005, R.E. Deakin Notes on Least Squares (2005) 10–8

RMIT University Geospatial Science

Direction coefficients Computed Bearing Observed Bearing Numeric term

Station a (sec/cm) b (sec/cm) Deg Min Sec Deg Min Sec comp-obs (sec)
A 0.837318 -1.204891 34 47 48.08 34 47 52 -3.92
B 1.433784 -0.226459 81 01 28.43 81 01 23 5.43
C -0.584244 1.548607 200 40 12.06 200 40 18 -5.94
D -1.796911 0.578189 252 09 48.52 252 09 35 13.52

Coefficient matrix B, vector of numeric terms f, normal equation coefficient matrix

N = BT WB and numeric terms t = BT Wf

ΔN ΔE
0.837318 -1.204891 -3.92
1.433784 -0.226459 5.43
B= f=
-0.584244 1.548607 -5.94
-1.796911 0.578189 13.52

In this example, the observations are assumed to be of equal precision. In such cases the
weight matrix W can be replaced by the Identity matrix I, and the normal equation coefficient
matrix N = BT WB = BT IB = BT B and the vector of numeric terms t = BT Wf = BT If = BT f

6.327065 -3.277288 -16.323890

N = BT B t = BT f =
-3.277288 4.235533 2.114714

Inverse of normal equation coefficient matrix N −1 and solution vector x = N −1t

0.263767 0.204092 -3.8741 (cm)

N −1 = x = N − 1t =
0.204092 0.394016 -2.4983 (cm)

Corrections to approximate coordinates

ΔE = −2.4983 cm = −0.025 m , ΔN = −3.8741 cm = −0.039 m

© 2005, R.E. Deakin Notes on Least Squares (2005) 10–9

RMIT University Geospatial Science

Adjusted coordinates of P and residuals (after one iteration)

EP = EP0 + ΔE = 13677.500 − 0.025 = 13677.475 m

N P = N P0 + ΔN = 29834.000 − 0.039 = 29833.961 m

Station Residual
A -3.68 (sec)
B 10.42
C -4.33
D 8.01

Variance factor σ 02 and precision of computed coordinates

An estimate of the variance factor σˆ 02 can be computed from equation (10.18) with W = I,
n = 4, u = 2 and

v T v f T f − xT t 262.925300 − 57.957192
σˆ 02 = = = = 102.484054 sec2
n−u 4−2 2

Assuming all the observations are of equal precision and letting W = I is equivalent to
assigning an estimated standard deviation of 1 second to each observation. Inspection of the
variance factor shows that if a standard deviation of 10.12 sec (10.12 = 102.484054 ) was
used as an estimate of the standard deviation of the observed directions, the estimate of the
variance factor computed from the adjustment would have been approximately unity (a
variance factor of unity indicates that the estimates of variances are close to the population
statistics).

From this adjustment, we may conclude that the standard deviation of the observed directions
was approximately 10.1 sec.

A most important "by-product" of a least squares adjustment is the ability to estimate the
precision of the computed quantities. Theory shows that this information is contained in the
inverse of the normal equations and the covariance matrix of the computed quantities Σ xx is
given by

© 2005, R.E. Deakin Notes on Least Squares (2005) 10–10

RMIT University Geospatial Science

⎡0.263767 0.204092 ⎤ ⎡ 27.0319 20.9162 ⎤

Σ xx = σ 02 Q xx = σ 02 N −1 = 102.484054 ⎢ ⎥=⎢ ⎥
⎣0.204092 0.394016 ⎦ ⎣ 20.9162 40.3804 ⎦

and the standard deviation of the adjusted coordinates are

σ E = 40.3804 cm = 0.064 m
σ N = 27.0319 cm = 0.052 m

Parameters of Standard Error Ellipse

From the variance matrix Σ xx above we have

σ E2 = 40.3804, σ N2 = 27.0319 and σ EN = 20.9162

Using the formulae from Chapter 8, Section 8.3 (replacing s with σ ), the lengths of the semi-
axes of the Standard Error Ellipse are

(σ − σ N2 ) + 4 (σ EN )
2
W = 2 2
E

= ( 40.3804 − 27.0319 ) + 4 ( 20.9162 )

2 2

= 43.9105

a = (
1 2
2
σ E + σ N2 + W ) = 7.4607 cm

b =
2
(σ E + σ N2 − W ) = 3.4280 cm
1 2

The angle between the E-axis and the major axis (positive anti-clockwise), noting the
quadrant signs to determine the proper quadrant of 2θ

2σ EN 2 ( 20.9162 ) ⎛+⎞
tan 2θ = = =⎜ ⎟
σ E − σ N 40.3804 − 27.0319 ⎝ + ⎠
2 2

2θ = 72D 18′ 09′′

θ = 36D 09′ 04′′

Substituting the values φ = θ = 36D 09′ 04′′ and φ = 90D + θ = 126D 09′ 04′′ into equation (8.3)

gives su = 7.4607 cm and su = 3.4280 cm respectively so θ = 36D 09′ 04′′ is the angle

© 2005, R.E. Deakin Notes on Least Squares (2005) 10–11

RMIT University Geospatial Science

(positive anti-clockwise) between the E-axis and the major axis. Hence the bearing of the
major axis is 90D − θ = 53D 50′ 56′′

Figure 10.2 shows a schematic diagram of the example intersection and the Standard Error
Ellipse

Figure 10.2 Schematic diagram of intersection and Standard Error Ellipse

© 2005, R.E. Deakin Notes on Least Squares (2005) 10–12

RMIT University Geospatial Science

REFERENCES

Crandall, K.C. and Seabloom, R.W., 1970, Engineering fundamentals in measurements,

probability, statistics and dimensions, McGraw-Hill, New York.

Cross, P.A. 1992, Advanced Least Squares Applied to Position Fixing, Working Paper No. 6,
Department of Land Information, University of East London.

Deakin, R.E. and Kildea, D.G., 1999, 'A note on standard deviation and RMS', The
Australian Surveyor, Vol. 44, No. 1, pp. 74-79, June 1999.

Gauss, K.F. 1809, Theory of the Motion of the Heavenly Bodies Moving about the Sun in
Conic Sections, a translation of Theoria Motus Corporum Coelestium in sectionibus
conicis solem ambientium by C.H. Davis, Dover, New York, 1963.

Johnson, N.L. and Leone, F.C., 1964, Statistics and Experimental Design In Engineering and
the Physical Sciences, Vol. I, John Wiley & Sons, Inc., New York

Krakiwsky, E.J. 1975, A Synthesis of Recent Advances in the Method of Least Squares,
Lecture Notes No. 42, 1992 reprint, Department of Surveying Engineering, University
of New Brunswick, Fredericton, Canada

Kreyszig, Erwin, 1970, Introductory Mathematical Statistics, John Wiley & Sons, New York.

Leahy, F.J. 1974, 'Two hundred years of adjustment of survey measurements', Two Centuries
of Surveying: Proceedings of the 17th Australian Survey Congress, Melbourne, 23 Feb.
− 4 Mar. 1974, Institution of Surveyors, Australia, pp.19-29.

Merriman, M. 1905, Method of Least Squares, 8th edn, John Wiley & Sons, New York.

Mikhail, E.M. 1976, Observations and Least Squares, IEP−A Dun-Donnelley, New York.

Mikhail, E.M. and Gracie, G. 1981. Analysis and Adjustment of Survey Measurements. Van
Nostrand Reinhold, New York, 340 pages.

© 2005, R.E. Deakin Notes on Least Squares (2005) R–1

RMIT University Geospatial Science

Rainsford, H.F., 1968, Survey adjustments and least Squares, Constable, London.

Tienstra, J.M., 1966, Theory of adjustment of normally distributed observations, Argus,

Amsterdam

Walpole, R.E., 1974, Introduction to Statistics, 2nd edn, Macmillan Publishing Co., Inc.
New York.

Wells, D.E. and Krakiwsky, E.J. 1971. The Method of Least Squares. Lecture Notes No. 18,
Department of Surveying Engineering, University of New Brunswick, May 1971,
reprinted September 1992, 180 pages.

© 2005, R.E. Deakin Notes on Least Squares (2005) R–2

RMIT University Geospatial Science

APPENDIX A MATRIX ALGEBRA

A 1 INTRODUCTION

Matrix algebra is a powerful mathematical tool, which is extremely useful in modern

computational techniques applicable to spatial information science. It is neither new nor
difficult, but prior to the development of the electronic computer was thought to be too
cumbersome for practical applications. In today's computer age large masses of data are
accumulated and matrix algebra is a convenient and concise way of expressing algorithms and
computer routines for the manipulation of data.

The advantages of matrix algebra may be set out as:

1. It provides a systematic means of storing and manipulating large arrays of data. Such
data may range from numerical coefficients of equations to characters and symbols
related to scanned digital images.
2. It provides a means of reducing large and complicated systems of equations to simple
expressions, which can be easily visualised and analysed.
3. It provides a concise method of expressing algorithms and of directing computer
execution of those algorithms via computer programs.

A 2 DEFINITIONS

A 2.1 Matrix

A matrix is a set of numbers or symbols arranged in a square or rectangular array of m rows

and n columns as

LM a11 a12 a13 a1n OP

M
= Ma
a
21 a22
a32
a23
a33
a2 n
a3n
PP
A m ,n
MM 31
PP
MNa
m1 am 2 am 3 amn PQ

© 2005, R.E. Deakin Notes on Least Squares (2005) A1

RMIT University Geospatial Science

A letter or symbol refers to the whole matrix. In many texts and references, matrices are
denoted by boldface type, ie,

A X P Q W

Matrices may also be indicated by pacing a tilde (~) under a symbol, ie,

A X P Q W
~ ~ ~ ~ ~

Example A 1

LM2 2 −6 OP LM 2 5 OP
MM2
A= 5 1 9
−8QP
P MM−4
C= 3 5
1QP
P x = −14 22 3 −8
N 2 N
A, C and x are all matrices. Note that the matrix x is a row matrix or row vector. Row
matrices or row vectors are usually denoted by lowercase letters.

A 2.2 Matrix element

Individual elements of a matrix are shown by lowercase letters ai j where the subscripts i and j

indicate the element lies at the intersection of the i th row and the j th column. The first
subscript always refers to the row number and the second to the column number.

column j
B
LM a 11 a12 a13 a1 j a1n OP
MM a 21 a22 a23 a2 j a2 n
PP
=M PP ← row i
A m ,n
MM a ai 2 ai 3 ca h ai n
PP
i1 ij

MMNa
m1 am 2 am 3 am j amn PQ
Another way of representing a matrix is by typical element, for example

i = 1, 2, … , n
A = ai j n s j = 1, 2, … , m

© 2005, R.E. Deakin Notes on Least Squares (2005) A2

RMIT University Geospatial Science

A 2.3 Matrix order

A matrix is said to be of order m by n (or m, n) where m is the number of rows and n is the
number of columns. The order of a matrix may be expressed in various ways ie,

A m ,n A a m ,n f Am × n Aa m × n f m An

Example A 2

LM2 2 −6 OP LM 2 5 OP
A3,3
MM2
= 5 1 9
−8PQ
P C3,2
MM−4
= 3 5
1PQ
P x1, 4 = −14 22 3 −8
N 2 N
Matrix A is of order (3,3), matrix C is (3,2) and x is (1,4). If a matrix is of order (1,1), it is
called a scalar.

A 3 TYPES OF MATRICES

A 3.1 Square Matrix

A square matrix is a matrix with an equal number of rows and columns. A square matrix
would be indicated by A m,m and said to be of order m. Square matrices have a principal or
leading diagonal whose elements are aij for i = j . In matrix A below, order (5,5), elements a,
g, m, s and y lie on the leading diagonal.

LM a b c d e OP
MM kf g h i j
PP
= l m n o
A 5, 5
MM p q r s t PP
MNu v w x y PQ
Special cases of square matrices are symmetric and skew-symmetric which are described
below.

© 2005, R.E. Deakin Notes on Least Squares (2005) A3

RMIT University Geospatial Science

A 3.2 Column Matrix or Column Vector

A column matrix or column vector is a matrix composed of only one column. Column
vectors are usually designated by lowercase letters, for example

LM b OP
1

M
= Mb P
b
2
P
bm,1
MM PP
3

MNb PQ
m

A 3.3 Row Matrix or Row Vector

A row matrix or row vector is a matrix composed of only one row. Row vectors are usually
designated by lowercase letters, for example

b1,n = b1 b2 b3 bn

A 3.4 Diagonal Matrix

A diagonal matrix is a square matrix with all "off-diagonal" elements equal to zero

LMd 11 0 0 0 OP
M
=M0
0 d22 0 0
0 P
P where dij = 0 for i ≠ j
MM
0 d33
D m,m
PP
MN 0 0 0 d PQ
mm

A diagonal matrix may have some diagonal elements equal to zero. A diagonal matrix is
often shown in the form

l
D = diag d1 , d2 , d3 , , dm q
A 3.5 Scalar Matrix

A scalar matrix is a diagonal matrix whose elements are all equal to the same scalar quantity

LMa 0 0 0 OP
M
A = M0
0 a
0
0
a
0
0
PP aij = 0 for i ≠ j
MM PP where
aij = a for i = j

MN0 0 0 a PQ

© 2005, R.E. Deakin Notes on Least Squares (2005) A4

RMIT University Geospatial Science

Example A 3

LM2 0 0 OP
MM0
W= 0 2 0
2 QP
P is a (3,3) scalar matrix
N 0

A 3.6 Identity or Unit Matrix

An identity or unit matrix is a diagonal matrix whose elements are all equal to 1 (unity). It is
always referred to as I where

LM1 0 0 0 OP
M
I = M0
0 1
0
0
1
0
0P
P
MM PP
MN0 0 0 1 PQ

Note that all the "off-diagonal" elements are zero and all the elements of the leading diagonal
are unity.

A 3.7 Null or Zero Matrix

A null or zero matrix is a matrix whose elements are all zero. It is denoted by boldface 0.

A 3.8 Triangular Matrix

A triangular matrix is a square matrix whose elements above, or below, but not including the
leading diagonal, are all zero. Square matrices whose elements above the leading diagonal
are zero are known as lower triangular matrices.

LM l11 0 0 0 OP
M
= Ml
l
21 l22 0 0
PP where lij = 0 for i < j
MM
l32 l33 0
L m,m 31
PP
MNl
m1 lm 2 lm 3 lmm PQ
Square matrices whose elements below the leading diagonal are zero are known as upper
triangular matrices.

LMu 11 u12 u13 u1m OP

M
=M0
0 u22 u23 u2 m
PP where uij = 0 for i > j
MM
0 u33 u3 m
U m ,m
PP
MN 0 0 0 umm PQ
© 2005, R.E. Deakin Notes on Least Squares (2005) A5
RMIT University Geospatial Science

Example A 4

LM2 1 −2 4 3 OP
M 0 5 4 −3 2
PP LM8 0 0 OP
G = M0
MM0
0 3 2 1
6 PP
MM5
H= 3 4 0
6 PQ
P
0 0 5 N −2
MN0 0 0 0 2 PQ

G and H are both triangular matrices. G is an upper triangular matrix and H is a lower
triangular matrix. Triangular matrices of order n have n2 + n 2 non-zero elements. c h
A 3.9 Unit Lower Triangular Matrix

This is a special case of a lower triangular matrix, in which all the elements of the leading
diagonal are equal to unity.

LM 1 0 0 0 OP
M
= Ml
l
21 1 0 0
PP lij = 0 for i < j
L m,m
MM31 l32 1 0
PP where
lij = 1 for i = j

MNl
m1 lm 2 lm 3 1 PQ

A 3.10 Unit Upper Triangular Matrix

This is a special case of an upper triangular matrix, in which all the elements of the leading
diagonal are equal to unity.

LM1 u12 u13 u1m OP

M
= M0
0 1 u23 u2 m
PP uij = 0 for i > j
U m ,m
MM
0 1 u3 m
PP where
uij = 1 for i = j

MN0 0 0 1 PQ

A 3.11 Banded Matrix

A banded matrix is any square matrix in which the only non-zero elements occur in a band
about the leading diagonal. Thus, if A is to be a banded matrix

aij = 0 when i − j > k

A typical banded matrix of order 4 is

© 2005, R.E. Deakin Notes on Least Squares (2005) A6

RMIT University Geospatial Science

LMa 11 a12 0 0 OP
A=M PP
a a22 a23 0
where aij = 0 for i − j > 1
21

MM 0 a32 a33 a34

PQ
N0 0 a43 a44

A 4 MATRIX OPERATIONS

A 4.1 Equality

Two matrices A and B are equal if and only if they are the same order and aij = bij for all i
and j. Matrices of different order cannot be equated.

A 4.2 Addition

The sum of two matrices A and B, of the same order, is a matrix C of that order whose
elements are cij = aij + bij for all i and j. Matrices of different order cannot be added. The
following laws of addition hold true for matrix algebra:

commutative law A+B= B+A

associative law A + (B + C) = (A + B) + C = A + B + C

Example A 5

LM8 3 −1 OP LM2 2 −5 OP LM10 5 −6 OP

A3,3
MM5
= 3 4 2
6 PQ
P B3,3
MM6
= 5 4 3
12 PQ
P MM11
A+B= 8 8 5
18 PQ
P
N −2 N −7 N −9

A 4.3 Scalar Multiplication

Multiplication of a matrix A by a scalar k is another matrix B of the same order whose

elements are bij = k aij for all i and j.

Example A 6

LM8 3 −1 OP LM4.0 1.5 −0.5 OP

MM5
A= 3 4 2
6 PQ
P B= 1
2
MM2.5
A = 1.5 2.0 1.0
3.0 PQ
P
N −2 N −1.0

The following laws relating to scalar multiplication hold true

© 2005, R.E. Deakin Notes on Least Squares (2005) A7

RMIT University Geospatial Science

k (A + B) = kA + kB
( k + q )A = kA + qA
k (AB) = ( kA)B = A( kB)
k ( qA) = ( kq )A

A 4.4 Scalar Product

In vector algebra, it is customary to denote the scalar product of two vectors

a = a1i + a2 j + a3k and b = b1i + b2 j + b3k , i, j and k being unit vectors in the direction of the
x, y and z axes respectively, as

LMb OP
1

a1 a2 a3 MMb PP = a b + a b + a b
2 1 1 2 2 3 3

Nb Q
3

In matrix algebra, a set of three simultaneous equations represented as

LMa 11 a12 OPLM

a13 x1 b1 OP LM OP
Ax = b or MMa21 a22 PPMM
a23 x2 = b2 PP MM PP
Na31 a32 a33 x3QN b3 Q N Q
means that each element of the column vector b is the scalar product of each row of A by the
column vector x, i.e.,

a11 x1 + a12 x2 + a13 x3 = b1

a21 x1 + a22 x2 + a23 x3 = b2
a31 x1 + a32 x2 + a33 x3 = b3

A 4.5 Matrix Multiplication

For three matrices A, B and C with their respective elements aij , bij and cij , then

AB = C implies cij = ∑ aik bk j

k

which states that the element in row i and column j of C is equal to the scalar product of row i
of A and column j of B

It is important to note that for matrix multiplication to be defined the number of columns of
the first matrix must be equal to the number of rows of the second matrix.

© 2005, R.E. Deakin Notes on Least Squares (2005) A8

RMIT University Geospatial Science

As a quick method of assessing whether a matrix multiplication is defined, write down the
matrices to be multiplied with their associated rows and columns, ie,

A 4 , 2 B2 , 6 and check that the "inner numbers" are the same.

A A

If they are the same, then the multiplication is defined and the product matrix has an order
equal to the "outer numbers".

A 4 , 2 B2 ,6 = C 4 , 6
A A

Remember, that in all cases, the first number of the matrix order refers to the number of rows
and the second number refers to the number of columns.

Example A 7

LM1 2 OP LM5 1 1 3 OP
A 3,2
MM6
= 3 0
4PQ
P and B2, 4 =
N2 Q
N 3 1 2

C3,4 = A 3,2 B2, 4

LM1 2OP L5 7 1 3O
= 3 9 M
MM6 4PP N2 3 8 2PQ
N Q
LM (1 × 5 + 2 × 2) (1 × 7 + 2 × 3) (1 × 1 + 2 × 8) (1 × 3 + 2 × 2) OP
MM(6 × 5 + 4 × 2) (6 × 7 + 4 × 3)
= (3 × 5 + 9 × 2) (3 × 7 + 9 × 3) (3 × 1 + 9 × 8) (3 × 3 + 9 × 2)
PP
N (6 × 1 + 4 × 8) (6 × 3 + 4 × 2) Q
LM 9 13 17 7 OP
MM38 54 38 26PP
= 33 48 75 27
N Q
In forming the product AB we say that B has been pre-multiplied by A, or that A has been
post-multiplied by B.

The following relationships regarding matrix multiplication hold:

AI = IA = A with I = the Identity matrix

A(BC) = (AB)C = ABC (associative law)
A(B + C) = AB + AC (distributive law)
(A + B)C = AC + BC (distributive law)

© 2005, R.E. Deakin Notes on Least Squares (2005) A9

RMIT University Geospatial Science

In these relationships above, the sequence of the matrices is strictly preserved. Note that in
general, the commutative law of algebra does not hold for matrix multiplication even if
multiplication is defined in both orders, ie,

AB ≠ BA in general

Example A 8
A is of order (2,3), B is of order (3,2), AB is of order (2,2) and BA is of order (3,3). Even if
both matrices are square and of the same order, the results will in general not be the same
when the order of multiplication is reversed.

LM4 6 2OP LM2 7 3OP

MM5 9 7PP
A= 3 1 8 and
MM6 8 4PP
B= 1 9 5
N Q N Q
LM26 98 50OP LM44 46 81OP
MM 61 172 88PP
AB = 55 94 46 and
MM68 80 104PP
BA = 56 60 109
N Q N Q
If the product of two matrices A and B is equal to the null matrix 0 then it does not follow that
either A or B is zero

Example A 9

LM−5 4 3 OP LM 2 −6 OP LM0 0 OP
MM 5
A= 3 6 8 ;
4 PQ
P MM−6
B= 7 −21 ;
18PQ
P MM0
AB = 0 0
0 PQ
P
N 2 N N
This differs from ordinary algebra where if, for example, a × b = 0 then either a or b or both a
and b are zero.

Some particular results involving diagonal matrices are useful. If A is a square matrix and D
is a diagonal matrix of the same order, then

1. AD causes each column A j of A to be multiplied by the corresponding element

d j j of D.

2. DA causes each row A i of A to be multiplied by the corresponding element dii of

D.

© 2005, R.E. Deakin Notes on Least Squares (2005) A10

RMIT University Geospatial Science

Example A 10

LMa 11 a12 a13 OP LMα 0 0 OP

A = Ma PP
MNa
21 a22 a33 and
MM 0
D= 0 β 0
γ PQ
P
31 a32 a33 Q N 0

LMαa 11 βa12 γ a13 OP LMαa 11 α a12 α a13 OP

AD = Mα a βa22 γ a33 PP and DA = Mβa βa22 β a33 PP
MNαa MNγ a
21 21

31 βa32 γ a33 Q 31 γ a32 γ a33 Q

If a diagonal matrix Dn,n has non-negative elements dii ≥ 0 then for p > 0

LMd p
11 0 0 OP
=M PP
p
0 d 0
MM
22
Dnp,n

N0 0 dnnp
PQ
and for p > 0 and q > 0

D p Dq = D p+q
1 1
and in particular D2D2 = D

A 4.6 Matrix Transpose

The transpose of a matrix A of order (m,k) is a (k,m) matrix formed from A by interchanging
rows and columns such that row i of A becomes column i of the transposed matrix. The
transpose of A is denoted by A T . There are various other notations used to indicate the
transpose of a matrix, such as: At , A′ , A* and A .

If B = A T then bij = a ji for all i and j

Example A 11

LM 3 2 3 OP LM3 1 4 OP
MM4
A= 1 −2 3
6QP
P A T
MM3
= 2 −2 0
6QP
P
N 0 N 3

LM4 OP
5
LM4 1 7 OP
MM7
B= 1 −2
0 PQ
P BT =
N5 −2 0 Q
N

© 2005, R.E. Deakin Notes on Least Squares (2005) A11

RMIT University Geospatial Science

The following relationships hold true

aA + Bf T
= A T + BT
(ABC… )T = …CT BT A T
( kA)T = kA T

cA hT T
=A

If D is a diagonal matrix, then DT = D

If H is a scalar matrix, then H = H T
If I is the Identity matrix, then I = I T
If x is a column vector, then x T x is a non-negative scalar that is equal to the sum of
the squares of the vector components.
If x is a row vector, then xx T is a symmetric matrix (square) of the same order as
the vector x.

A 4.7 Bilinear and Quadratic forms

If x is an (m,1) vector of variables, y an (n,1) vector of variables and A an (m,n) matrix of

constants, the scalar function

u = x T Ay

is known as a bilinear form.

If x is an (m,1) vector of variables and B an (m,m) square matrix of constants, the scalar
function

q = x T Bx

is known as a quadratic form. An example of a quadratic form is the sum of the squares of
the weighted residuals; the function ϕ = vT Wv , which is minimised in least squares.

A 4.8 Matrix Inverse

Division is not defined in matrix algebra. In place of division, the inverse A −1 of a square
matrix A is introduced. This inverse, if it exists, has the property

AA −1 = A −1A = I

This relationship defines the Cayley Inverse for square matrices only. A square matrix whose
determinant is zero is singular and a singular matrix does not have an inverse. A square

© 2005, R.E. Deakin Notes on Least Squares (2005) A12

RMIT University Geospatial Science

matrix whose determinant is non-zero is non-singular and does have an inverse. Furthermore,
if the inverse exists it is unique. Rectangular matrices have no determinants and so they are
taken to be singular but they may have an inverse (such as Moore-Penrose inverses), defined
using Generalised Matrix Algebra. These "generalised inverses" are not used in these notes.

Consider the matrix equation Ax = b . If A and b are known, then x may be determined from
x = A −1b . x is found in a sense, by "dividing" b by A, but in actual fact x is determined by
pre-multiplying both sides of the original equation by the inverse A −1 . For example

Ax = b

A −1Ax = A −1b (pre-multiply both sides by A −1 )

giving

x = A −1b (since A −1A = I and Ix = x )

The following rules regarding matrix inverses hold:

(ABC… )−1 = …C−1B−1A −1

( A −1 ) − 1 = A
( A T ) −1 = ( A − 1 ) T
1
(α A)−1 = A −1 (α is a scalar)
α

Matrix inversion plays an important part in least squares, primarily in the solution of systems
of linear equations. If the order of A is small, say (2,2) or (3,3), then manual calculation of
the inverse is relatively simple. But as the order of A increases, computer programs or
software products such as Microsoft's Excel or The MathWorks MATLAB are the appropriate
tools to calculate inverses and solve systems of equations.

For a (2,2) matrix the inverse is simple and may be computed from the following relationship

⎡A A12 ⎤ 1 ⎡ A22 − A12 ⎤

If A = ⎢ 11 ⎥ then A −1 =
⎣ A21 A22 ⎦ ( A11 A22 − A12 A21 ) ⎢⎣ − A21 A11 ⎥⎦

© 2005, R.E. Deakin Notes on Least Squares (2005) A13

RMIT University Geospatial Science

A 4.9 Matrix Differentiation

1. If every element of a matrix A of order (m,n) is a differentiable function of a (scalar)

variable u, then the derivative dA du is an (m,n) matrix of derivatives

LM da 11 da12 da1nOP
MM dadu du du
da P
P
M du P
21 da22 2n
dA
du m,n
= M du
MM
du
PP
MM dadu dam 2 da PP
du P
m1 mn

N du
Q
Example A 12

A=
LM3u 2u OP
2 3
then
dA
=
LM
6u 6u 2 OP
N u 4u Q
2 4
du N
2u 16u 3 Q
LM2u OP2
L 4u OP
dx M
x= Mu P
du M
MN12u PPQ
3
then = 3u 2

MN3u PQ
4 3

2. For the matrix product C = AB where the elements of the matrices A and B are
differentiable functions of the (scalar) variable u then dC du is given by

dC d dA dB
= (AB) = B+A
du du du du

Note that the sequence adopted in the product terms must be followed exactly, since for
example, the derivative of AB is in general not the same as the derivative of BA.

3. If a vector y m,1 represents m functions of the n elements of a variable vector x n,1 then the
total differential dy is given by

∂y
dy = dx
∂x

where the (m,1) and (n,1) vectors dy and dx contain differentials

LM dy OP
1 LM dx OP
1

=M P =M P
dy dx
MM PP MM PP
2 2
dy m,1 and dx n,1

Ndy Qm Ndx Q
n

© 2005, R.E. Deakin Notes on Least Squares (2005) A14

RMIT University Geospatial Science

and the partial derivative ∂y ∂x is an (m,n) matrix known as the Jacobian Matrix

LM ∂y 1 ∂y1 ∂y1OP
MM ∂x 1 ∂x2 ∂xn
PP
M
∂y M ∂x
∂y ∂y2 ∂y P
∂x P
2 2
∂x2
∂x M PP
= 1 n

MM P
MM ∂y ∂ym ∂y P
∂x P
m m

MN ∂x1 ∂x2
PQ
n

4. The derivative of the inverse A −1 is obtained from

AA −1 = I
d
dx
c
AA −1 =
dI
dx
h
=0

d A −1 dA −1
A +A =0
dx dx

hence

dA −1 dA − 1
= A −1 A
dx dx

A 4.10 Differentiation of Bilinear and Quadratic forms

For the bilinear form u = x T Ay where A is independent of both x and y

∂u ∂u
= yT AT and = xT A
∂x ∂y

For the quadratic form q = x T Bx where B is independent of x

∂q
= 2xT A
∂x

These differentials are given without proof, but can be verified in the following manner

LM x OP1
Ly O LMa11 a12 OP
= Mx P , = M P , and A = Ma PP
1
let x 3,1
MN x PQ
2

3
y 2,1
Ny Q
2
3,2

MNa
21

31
a22
a32 Q

© 2005, R.E. Deakin Notes on Least Squares (2005) A15

RMIT University Geospatial Science

then

LMa11 a12 OP L y O
u = x Ay = x1 MMa PP MNy PQ
T 1
x2 x3 21 a22
Na
31 a32 Q 2

= x1a11 + x2 a21 + x3a31 x1a12 + x2 a22 + x3a32

LM y OP
1

Ny Q
2

= y1 x1a11 + y1 x2 a21 + y1 x3a31 + y2 x1a12 + y2 x2 a22 + y2 x3a32

and

∂u
= y1 a11 + y2 a12 y1 a21 + y2 a22 y1 a31 + y2 a32
∂x
a f
= Ay
T

= yT AT

and

∂u
= x1a11 + x2 a21 + x3a31 x1a12 + x2 a22 + x3a32
∂y
= xT A

Using similar methods, the partial differential for the quadratic form q can also be verified.
More explicit proofs can be found in Mikhail (1976, pp.457-460) and Mikhail & Gracie
(1981, pp.322-324).

A 4.11 Matrix Partitioning

A subset of elements from a given matrix A is called a sub-matrix and matrix partitioning
allows the matrix to be written in terms of sub-matrices rather than individual elements.
Thus, the matrix A can be partitioned into sub-matrices as follows

LM a 11 a12 a13 a14 a1n OP

MM aa
21 a22
a32
a23
a33
a24
a34
a2 n
a3n
PP
=M PP
31
A m ,n
MM a
41 a42 a43 a44 a4 n
PP
MNa
m1 am 2 am 3 am 4 amn Q

© 2005, R.E. Deakin Notes on Least Squares (2005) A16

RMIT University Geospatial Science

Considering the vertical dotted line only, A could be written as

A = A1 A 2

where A1 is an (m,3) sub-matrix and A 2 is an m, (n − 3) sub-matrix. Similarly, considering

the horizontal dotted line only

A=
LM A OP
1

NA Q 2

where in this case A1 is a (2,n) sub-matrix and A 2 is an ( m − 2) ,n sub-matrix.

Considering both the horizontal and vertical lines

A=
LMA 11 A12 OP
NA 21 A 22 Q
where A11 is a (2,3) sub-matrix, A12 is a (2, (n − 3) ) sub-matrix, A 21 is an ( ( m − 2) ,3) sub-
matrix and A 22 is an ( ( m − 2) , (n − 3) ) sub-matrix.
All matrix operations outlined in the previous sections can be performed on the sub-matrices
as if they are normal matrix elements providing necessary precautions are exercised regarding
dimensions.

Example A 13
Transposing partitioned matrices

LM 1 2 3 4 OP LA A12 OP
8 =M
MM9
A= 5
P NA
11

N
6 7
10 11 12 PQ
21 A 22 Q
and

LM 1 5 9 OP
AT =
LMA T
A T
OP = M2 6 10
PP
Q MM 3
11 21

NA T
12 A T
22
7 11
P
N4 8 12 Q

© 2005, R.E. Deakin Notes on Least Squares (2005) A17

RMIT University Geospatial Science

Example A 14
Multiplying partitioned matrices

LM 1 OP LA LM 1 7OP
2 3 4
2 =M
A12 OP =M
2 4
PP LM OP
B11
MM3
= 2 −1 5
P NA =
11
A3,4
N −2 PQ
21 A 22 Q 2, 2
and B4 , 2
MM 7 6
PQ N Q
B21 2 ,1
2 1
N −2 5

the product is

AB = C =
LM(A B
11 11 + A12 B21 ) OP
N(A B
21 11 + A 22 B21 ) Q 2 ,1

where

A11B11 =
LM 1 2OPLM1 7OP = LM5 15OP ; A12 B21 =
LM3 4OPLM 3 6OP LM
=
1 38 OP
N2 −1QN2 4Q N0 10Q N5 2 Q N − 2 5 Q N
11 40 Q
= 3 2M
L1 7OP = 7 29 ; = 1 −2 M
L3 6O
5PQ
= 7 −4
A 21B11
N2 4 Q A 22 B21
N−2
noting that columns of A xx must equal rows of B xx . The product is

LM 6 53 OP
MM14
AB = C = 11 50
25PQ
P
N

A 5 SOME SPECIAL MATRICES

A 5.1 Symmetric Matrices

A symmetric matrix is defined to be a matrix that remains invariant when transposed, ie,

AT = A

Symmetric matrices are always square matrices. For any symmetric matrix A, the elements
conform to the following

aij = a j i

© 2005, R.E. Deakin Notes on Least Squares (2005) A18

RMIT University Geospatial Science

Example A 15

LMa 11 3 5 7 OP
A=M PP
3 a22 9 11
MM 5 9 a33 13
PQ
N7 11 13 a44

For any matrix A and for any symmetric matrix B the matrices

AA T , A T A , ABA T and A T BA

are all symmetric matrices.

In least squares, we are often dealing with symmetric matrices. For example, the matrix
equation Nu,u = BuT,n Wn,n Bn,u often appears. B is an (n,u) matrix of coefficients of the u
unknowns in n equations, W is an (n,n) weight matrix (always symmetric) and N is the (u,u)
symmetric coefficient matrix of the set of u normal equations.

A 5.2 Skew-symmetric Matrices

In contrast to the above, a skew-symmetric (or anti-symmetric) matrix is defined to be a

square matrix that changes sign when transposed, so that

AT = − A

and the elements conform to the rule

aij = − a j i

Note that this definition means that the elements of the leading diagonal can only be zero. An
example of a skew-symmetric matrix of order 3 is

LM 0 b c OP
MM−c
A = −b 0 d
0 PQ
P
N −d

© 2005, R.E. Deakin Notes on Least Squares (2005) A19

RMIT University Geospatial Science

Example A 16
Skew-symmetric matrices are found in some surveying and geodesy applications. For
instance, a 3D conformal transformation from one orthogonal coordinate system (x,y,z) to
another (X,Y,Z) is defined by the matrix equation

LM X OP LM x OP LMT OP X

MMYZ PP = λ Rκ φω MM yzPP + MMTT PP

Y

NQ NQ N Q Z

where λ is a scale factor, TX , TY and TZ translations between the coordinate origins and Rκ φ ω

is a rotation matrix derived by considering successive rotations ω , φ and κ about the x, y and
z axes respectively

LM c cφ κ cω sκ + sω sφ cκ OP
sω sκ − cω sφ cκ
Rκ φ ω = M− c sφ κ cω cκ − sω sφ sκ s c +c s s P
ω κ ω φ κ
MN s φ − sω cφ c cω φ
PQ
Note that cκ sφ sω = cosκ sin φ sin ω , and x, y, z and X, Y, Z refer to the axes of right-handed

orthogonal coordinate systems. Rotations ω , φ and κ are considered as positive anti-

clockwise according to the "right-hand-grip rule".

In many applications the rotation matrix Rκ φ ω can be simplified because ω , φ and κ are

small (often less than 3°). In such cases, the sines of angles are approximately equal to their
radian measures, the cosines are approximately 1 and products of sines are approximately
zero. This allows the rotation matrix Rκ φ ω to be approximated by R S

LM 1 κ −φ OP
RS
MM φ
= −κ 1 ωP
1 PQ
N −ω

This matrix is sometimes referred to as a skew-symmetric matrix. Although the elements of

the upper-triangular part are the opposite sign of those in the lower triangular part, it does not
conform to the definition above, since the leading diagonal elements are not zero.

© 2005, R.E. Deakin Notes on Least Squares (2005) A20

RMIT University Geospatial Science

Note that R S can be expressed as the sum of the identity matrix I and a skew-symmetric
matrix S

LM 1 κ −φ OP LM1 0 0 OP LM 0 κ −φ OP
RS
MM φ
= −κ 1 ωP= 0
M
1 PQ MN0
1
P M
0 + −κ
1 PQ MN φ
0 ω P = I+S
0 PQ
N −ω 0 −ω

A 5.3 Symmetry and Skew-symmetry

Every square matrix can be uniquely decomposed into the sum of a symmetric and skew-
symmetric matrix. Consider the following

A = A + 12 A T − 12 A T
= 1
2 cA + A h + cA − A h
T 1
2
T

= A Sym + A Skew

where

A Sym = 12 (A + A T ) is symmetric because

A TSym = 12 (A + A T )T = 12 (A T + A) = A Sym

and

A Skew = 12 (A − A T ) is skew − symmetric because

A TSkew = 12 (A − A T )T = 12 (A T − A) = − 12 (A − A T ) = − A Skew

A 5.4 Orthogonal Matrix

An orthogonal matrix is a square matrix satisfying the following two conditions:

1. the norms of all its rows and columns are equal to unity, and
2. any row is orthogonal to every other row in the matrix and similarly for the
columns.
These two conditions imply that if A is an orthogonal matrix, then

AA T = A T A = I (I is the Identity matrix)

and hence an orthogonal matrix has the very useful property that its inverse matrix is the same
as its transpose matrix, or

A −1 = A T (if A is orthogonal)

The terms norm and orthogonal are applicable to vector algebra. The norm of a vector is the
magnitude of the vector and is the square root of the product of the vector and its transpose.

© 2005, R.E. Deakin Notes on Least Squares (2005) A21

RMIT University Geospatial Science

Any row (or column) of a matrix has all the characteristics of a vector, and hence the norm of
any row (or column) of a matrix is the square root of the product of the row (or column) by its
transpose. Two vectors are orthogonal if, and only if, their scalar product is zero.
Considering rows and columns of the matrix as vectors, then any two matrix rows (or
columns) are orthogonal if their scalar product is zero.

Example A 17
Rotation matrices are examples of orthogonal matrices. For example, consider a point P with
a f
coordinates P e, n in the east-north coordinate system. If the axes are rotated about the
origin by an angle θ (measured clockwise from north), P will have coordinates e′, n′ in the
rotated system equal to

e′ = e cos θ − n sin θ
n′ = e sin θ + n cos θ

These equations can be written in matrix notation as

LMe′ OP = LMcosθ − sin θ OPLMeOP LMe′ OP = R LMeOP

Nn′Q N sin θ cosθ QNnQ or
Nn ′ Q N n Q θ

R =M
Lcosθ − sin θ OP
where θ
N sin θ cosθ Q is known as the rotation matrix.

The norms of the columns and rows of Rθ are unity since sin 2 θ + cos2 θ = 1 and the columns
and rows are orthogonal since sin θ cos θ − sin θ cos θ = 0 . Hence Rθ is orthogonal and its
inverse is equal to its transpose. This is useful in defining the transformation from e′, n′ to
e, n coordinates. Pre-multiplying both sides of the original transformation by the inverse Rθ−1
gives

Rθ−1
LMe′ OP = R R LMeOP
−1 LMe OP = R LMe′ OP
−1
since Rθ−1Rθ = I
Nn ′ Q θ θ
Nn Q or
Nn Q Nn ′ Q
θ

and

LMeOP = R LMe′ OP = LM cosθ

T OPLM OP
sin θ e′
since Rθ−1 = RθT
NnQ Nn′Q N− sin θ
θ
QN Q
cos θ n′

© 2005, R.E. Deakin Notes on Least Squares (2005) A22

RMIT University Geospatial Science

APPENDIX A REFERENCES

Mikhail, E.M. 1976, Observations and Least Squares, IEP−A Dun-Donelley, New York.

Mikhail, E.M. and Gracie, G. 1981, Analysis and Adjustment of Survey Measurements. Van
Nostrand Reinhold, New York.

Williams, I.P. 1972, Matrices for Scientists, Hutchinson &Co Ltd, London.

© 2005, R.E. Deakin Notes on Least Squares (2005) A23