Professional Documents
Culture Documents
OF OBSERVATIONS
Naser El-Sheimy
This page is intentionally blank
1. INTRODUCTION TO ADJUSTMENT OF OBSERVATIONS
1.1. Why Observations?
Geomatics Engineers are usually faced with the problem of estimating some unknown
quantities (parameters). This is done through collecting several measurements of some
kind known as observations – and then adopting the appropriate mathematical model
relating both observation and unknowns.
Page 1
1.2. Sources of Errors
Sources of errors
Page 2
2) Systematic Errors
Characteristics: Occur according to so deterministic system, when known,
can be exposed by some functional relationship.
Source: Instrumental of natural personal of all
Effect: Shifting all the observation it can be constant if it’s magnitude and
sign remain the same throughout the measuring process.
Example:
Collimation
error
∆ ∆
L L
3) Random Errors
Characteristics: The remaining errors after gross and systematic errors
have been removed.
They have no functional relationship based upon a deterministic system,
usually modelled by stochastic model (probability theory)
Source: Personal, natural and instrumental.
Cannot be generally eliminated however it can be minimised by taking
redundant observations and applying the so called “Method of Least
Squares”
This process is referred to as the “Adjustment of Observations” or
“Adjustment Computations” which is our main concern in this course.
Based on the above fact, we can’t seek the ‘true’ value, all we can get is an
‘estimate’ for the ‘true’ value.
Page 3
The small variations between the measurements and the ‘true’ value or its
‘estimate’ are regarded as “errors.”
1.4. Notation used in Least Squares
x = [x1 ... xu ]
T
x2
u... number of unknowns
♦ Observations:
L = [l1 ... l n ]
T
l2
n... number of observations (measurements)
0 = f(x, l)
The function that relates x and l.
Types of Observations
Direct Indirect
Page 4
xˆ = Mean ( L)
Page 5
1.6. Behaviour of Random Errors
♦ For simplicity we’ll consider the direct case under the same conditions.
♦ Assumption: All measurements are free of gross errors and corrected for all
systematic errors.
-ε ε
1
-v 0 v +v
Probability Distribution Function
Range = vmax - vmin (PDF)
Page 6
1.7. Understanding the Meaning of Residual Errors (v)
♦ Given L = (l1 l2, … ln) and accepting, for the time being, the fact that the
arithmetic mean Error! Objects cannot be created from editing field codes. of all
lI is the best estimate Error! Objects cannot be created from editing field codes.
for the measured quantity – i.e.,
Σl i
xˆ = x =
n
we can compute vi as vi = xˆ − l i
♦ The residuals express the degree of closeness of the repeated measurements of the
same quantity to each other and therefore the (v) values can be used in expressing the
precision of xˆ (and also the precision of the observer who take/made these
measurements.)
e.g. two observers A and B measure the same angle.
A B
V1 = +2″ V1 = +5″
V2 = -1″ V2 = -3″
V3 = +1″ V3 = +2″
V4 = 0 V4 = -3″
V5 = -2″ V5 = -4″
V6 = +3″
Now, we define the range over which the residuals change as,
We conclude that the angle computed from the data set of (A) is more
precise if it is computed from the data set (B)
Page 7
1.8. Determining the Probability Distribution Histogram (PDH)
1) Calculate the estimated parameter:
Σl i
xˆ = x =
n
2) Calculate the residuals for each observation:
vi = xˆ − l i
3) Calculate the range of the residuals:
nj RF
fj =
n⋅∆ j
n j ... the number of residuals
that fall within the fj
boundaries of ∆ j
-v +v
∆j
Range
Page 8
Notes:
The area under the histogram = 1
k
APDH = ∑ Ai RF
j =1
k
= ∑ fi ⋅ ∆
j =1
k nj
=∑ ⋅∆
j =1 n⋅∆
A1 A2 A3 A4 A5 A6
k nj
=∑ -v
∆ ∆ ∆ ∆ ∆ ∆
+v
j =1 n
Range
k
nj
1
= ∑ n =1
n j =1 =
n
The histogram can be used for probability computation.
v1 ≤ Ρ ≤ v2 = Α ( v1 to v2 ) (v1 to v2)
A
v1 v2
Page 9
1.9. Characteristics of PDF
♦ Many scholars throughout the history of statistics have tried to describe the PDF
curve
♦ A commonly accepted model was given by Gauss (Gauss PDF)
h
G (ε ) =
2ε 2
e −h e = 2.71828
π
G(ε)
G(εi)
-ε εi +ε
Note: h is the only parameter that completely describes the shape of the Gauss
PDF.
h
♦ G (0 ) =
π
♦ h is usually called “precision index”. Precision α h
♦ Since the area under the curve = 1, then the higher the PDF (larger the h), the
narrower the curve must become – i.e. less range (more precise)
high precision
h1 low precision
π
h2
π
Range 1
Range 2
Page 10
Properties of Gauss PDF
1) Area under the curve = unity
∞
∫−∞
G (ε )dε = 1
-ε ε
max
-ε 0 +ε
Page 11
1
5) Two points of inflection at ε = ± . These points occur where second
h 2
∂ 2 G (ε )
derivative of function is equal to zero - = 0.
∂ε 2
h
π
1 1
h 2 h 2
6) Statistical properties
ε
( 1 ) = −∫∞1G(ε )dε = A1
Ρε ≤ε
Ρ (ε ≥ ε ) = ∫ G (ε )dε = A
∞
2 ε 2
2
ε2 A1 A2
= 1 − ∫ G (ε )dε
−∞
ε2
( 1 2 ε
)
Ρ ε ≤ ε ≤ = ∫ G (ε )dε
−∞ ε1 ε2 +∞
1
(More when we discuss statistical analysis)
Page 12
1.10. Reliability of Measurements
♦ Recall that the true value t of certain unknown quantity can never be obtained.
However, an estimate xˆ can be determined.
♦ xˆ will not be satisfactory from the client point of view because it is influenced by
random error.
♦ Therefore, we need a certain “measure” of the existing random errors to describe its’
“goodness”, “reliability”, and “repeatability”
♦ This measure should help as well in accepting or rejecting certain observations
depending on the desired precision.
♦ Three terms are commonly used in expressing the reliability of measurements:
1. Precision: the degree of closeness of repeated measurements of the
same quantity to each other. Precision is affected only by random
errors.
-ε +ε -ε +ε
closely clustered widely clustered
-ε +ε
systematic error
Page 13
♦ Consider a circular target, where t is the centre of the target
precise but not not precise but not precise & precise and
accurate accurate not accurate accurate
Measures of Precision
Random Sample
t = unknown t = known
Σ vi Σεi
ae = ∂e =
(n − 1) n
vi = xˆ − l i εi = t − li
Σl i
xˆ = x =
n
Random Variable
∞
1
∂e = ∫ ε G(ε )dε = h
−∞ π
Page 14
2) Probable Error (Pe)
Pe has the following property: Half the resulting errors are smaller in
magnitude than Pe and the other half are larger than Pe.
50% of vi > Pe
50% of vi < Pe
How: Arrange the absolute values of the errors into either ascending or
descending order.
For odd numbers Pe = V n +1
2
e.g. v1 v2 v3 v4 v5
Pe = v3
1 n
For even numbers Pe = V + V (n 2 + 1)
2 2
e.g. v1 v2 v3 v4 v5 v6
v3 + v 4
Pe =
2
Problems with Pe
1″ 1″ 3″ 5″ 6″ 7″ 11″
1″ 1″ 3″ 5″ 6″ 7″ 100″
2 ∫ G (ε )dε = 0.5
0
0.4769 +∞
Pe → −∞
-Pe +P
h
Page 15
3) Standard Deviation (σ)
σ is defined as the square-root of the arithmetic mean of the sum of
square of the errors.
The square of the standard deviation – σ2 – is known as the “variance” or
“mean square error” and consequently σ is sometimes referred to as the
Root Mean Square Error (RMSE)
Error! Objects cannot be created from editing field codes.
Random Sample
t = unknown t = known
1 n 2 1 n 2
σ2 = ∑ v1
n − 1 i =i
σ2 = ∑εi
n i =i
We have only one unknown, and hence we σ : Root mean square error
only need a single observation to determine (RMSE)
it. The remainder (n-1) is usually known as
the “Degrees of Freedom” or “Redundancy”,
which we will refer to as “r”
r = n−u
r... redundancy
n... number of observations
u... number of unknowns
1
σ =
h 2
Page 16
Unlike the previous two measures of precision, σ has some distinguished properties
which make it a preferred measure of precision, they are:
G(ε)
-1 1
h 2 h 2
G(ε)
-3σ -2σ -σ σ 2σ 3σ
Ρ(− σ ≤ ε ≤ σ ) = 0.683
Ρ(− 2σ ≤ ε ≤ 2σ ) = 0.954 called confidence interval
Ρ(− 3σ ≤ ε ≤ σ ) = 0.997
Page 17
1.11. Review of Uni-variate Statistics
Given: L = l1 , l2 , l3 , ……… ln
∑l
xˆ = x = i =i
♦ Variance of x – σ x2
σ 2
=
∑v 2
σ2x square units of x (m2 or degree2)
x
n −1
σ x2
σ x2 =
n
Note: Since x̂ is the best estimate of the unknown parameter out of a group of
measurements (having a variance of σ x2 ), therefore σ x2 should be better than σ x2
σx <σx
σ x ... precision of the mean
σ x ... precision of a single measurement
σ x2
♦ Proof that σ = 2
x :
n
To determine the standard deviation of the mean, we begin with the expression for
the mean:
Page 18
n
∑x i
x1 x 2 x
x= i =i
= + +................. + n
n n n n
e.g. x = x1 + x 2 = f ( x1 , x 2 )
2 2
∂f 2 ∂f 2
σ =
2
x
σ x1 + σ x 2
∂ x1 ∂ x 2
♦ Applying the law of propagation of variances to the equation for the mean
2 2 2
1 1 1
σ = σ x21 + σ x22 = ............... + σ xn2
2
x
n n n
2
1
(
= σ x21 + σ x22 + ............... + σ xn2 )
n
1
(
= 2 nσ 2 )
n
σ2
=
n
σ
=
n
♦ This yields the final expression for the standard deviation of the mean
σx =
∑v 2
i
(n − 1)n
Page 19
1.12. Direct Observation with Different Conditions
♦ Consider that two observers A and B are measuring the same quantity X (say a
distance or angle) and that it is required to estimate the best estimate of X by making
use of the two sets of observations
Observer A Observer B
α1 α1
α2 α2
α3 α3
…
…
αn αm
Best estimate: ∑α i ∑α i
αA = i =i
αB = i =i
n m
σ αA =
∑v 2
i
i = i...n σ αB =
∑v 2
i
i = i...m
Precision: n −1 m −1
vi = α A − α i vi = α B − α i
♦ Now, α A and α B can be considered as measurements – they are the best estimate of
two groups of measurements for the same quantity (X).
♦ Therefore our original problem can be formulated as follows:
Given: α A , σ α A and α B , σ α B
αA +αB
♦ Could we say α = ? NO - because this will not take into account
n
difference in precision expressed by σ α A and σ α B
♦ To account for the difference in accuracy between the two sets of observations, we
introduce a new quantity called the weight:
1
Pα
σ2
Why?
Page 20
Recall
∑ vi
2
σx =
∑v 2
i
(n − 1)n
Therefore, a measurement of high precision will have a small variance and
vice versa.
Since the value of variance goes in opposite direction to that of precision,
another measure of precision is often used.
We call this quantity “Weight”.
∑ Piα i p α + p Bα B
α= = A A
∑ Pi p A + pB
1 1
pA = pB =
σ α2A σ α2B
♦ In general:
x̂ =
∑ Px Weighted mean
∑P
σ xˆ =
∑Pv 2
1 1
(n − 1)∑ P
♦ Note: is all the observations have the same variance (i.e. equal weights), the above
formulas will be the same as the mean and variance equations discussed before
Page 21
ENGO361 Naser El-Sheimy
1. MULTIVARIATE STATISTICS
♦ Geomatics problems (surveying problems) normally include the measurement of
several quantities. In turn, these measurements are used to determine several
unknown parameters.
♦ The measured quantities cannot usually be treated separately. Instead, they must
be dealt with simultaneously. Both the effect of each quantity on the others and
the statistical relationship between quantities must be taken into consideration in
order to obtain a meaningful solution of the unknowns.
♦ A multivariate consists of several univariate, e.g.:
d1 sin α1 A α2 Unknowns
x = [xa, ya, xb, yb]
d1 B
α1 Observations
d1 cos α1
L = [α1 , d1 , α2 , d2]
X
Page 1
ENGO361 Naser El-Sheimy
1.1. Covariance
1 n
σ ab = ∑ vai vbi
n − 1 i =i
(units of a⋅b)
vai = a − ai
vbi = b − bi
♦ σab has the physical units of a multiplied by the physical units of b. That is, the
covariance has no specific units and can take any value between −∞ → + ∞ (i.e.
no limit)
♦ The covariance between the mean values of a and b is:
σ ab
σ ab =
n
Page 2
ENGO361 Naser El-Sheimy
♦ In practice, all the variances and covariances of a multivariate are assembled into
one matrix called the variance-covariance matrix (v-c matrix), or simply the
covariance matrix.
σ a2 σ ab σ ac
C L = σ ba σ b2 σ bc
σ ca σ cb σ c2
5 1 − 2
A = 1 3 0
Matrix is not symmetric
2 0 4
6 6
A=
Matrix is not invertable – determinant is
6 6 equal to zero ( |A| = 0 )
Page 3
ENGO361 Naser El-Sheimy
σ ab
ρ ab = ρ ba =
σ aσ b
2. Has a limits of ± 1
σ aa σ a2
ρ aa = = =1
σ aσ a σ aσ a
♦ If
ρab = 0 completely uncorrelated
ρab = +1 completely positively correlated
ρab = -1 completely negatively correlated
ρab<1 completely correlated
1 ρ ab ρ ac
ρ L = ρ ba 1 ρ bc
ρ ca ρ cb 1
NOTE:
σ a2 ρ abσ aσ b ρ acσ aσ c
Cl = σ b2 ρ bcσ bσ c
sym σ c2
Page 4
ENGO361 Naser El-Sheimy
b b b
a a a
a increases a increases Very weak
b increases b decreases correlation
ρab → +ve ρab → -ve
b b
or
a a
a no change a increases
b increases b no change
No correlation
Page 5
ENGO361 Naser El-Sheimy
Page 6
ENGO361 Naser El-Sheimy
x = [x1 ... xu ]
T
x2 number of unknowns = u
g = [g 1 ... g m ]
T
g2 number of functions = m
l = [l1 ... l n ]
T
l2 number of observations = n
The model is direct with respect to the parameters
One equation per parameter (i.e. u = m)
The parameters are expressed directly as functions of the observations
♦ Example:
Observations:
A
l = [a b]
T
(n = 2)
Unknowns:
c a x = [A c]
T
(u = 2)
Functions:
b g = [g 1 g2 ]
g1 ⇒ A = a ⋅ b (m = 2)
g2 ⇒ c = a2 + b2
Page 7
ENGO361 Naser El-Sheimy
♦ Example: Levelling between two stations (i.e. elevation difference between two
stations)
B
∆hAB
A
Observations:
l = ∆h AB (n = 1)
Unknowns:
x = hB (u = 1)
Functions:
h = h1
(m = 1)
h1 ⇒ ∆h AB = hB − h A
The model is implicit with respect to the parameters and the observations
The parameters and observations cannot be separated, and have an “interwoven”
relationship
♦ Example:
Observations:
Y
l = [x1 y3 ]
T
y1 x2 y2 x3 (n = 6)
b
2 Unknowns:
3
x = [a b] (u = 2)
T
1
a Functions:
X
f = [ f1 f2 f3 ]
f1 ⇒ 0 = a + bx1 − y1
f 2 ⇒ 0 = a + bx 2 − y 2 (m = 3)
f 3 ⇒ 0 = a + bx3 − y 3
Page 8
ENGO361 Naser El-Sheimy
Gun
Parameter Observation
Space Spaces
X Hnu L
dim X = u dim L = n
Amu Bmn
F
dim F = m
Model
Space
Linear
♦ The math models are linear w.r.t. to the parameters (i.e., when the math models
are differentiated, they yield a vector of constants)
♦ Example – A simple levelling network (Arrow pointing at the higher station)
A ∆H1
∆H 1
x = [H B ]
Unknowns: l = ∆H 2 Observations:
u=1 n=3
∆H 3
c = [H A ] Functions:
Constants m = number of parameters (unknowns)
m=u=1
Math Model
u = 1, m = 1, n = 3 ∴ unique solution
Page 9
ENGO361 Naser El-Sheimy
∆H 1
[H B ]1x1 = [1 1 1]1x 3 ∆H 2 + [H A ]1x1
∆H 3 3 x1
Non-linear
♦ e.g., 1 = 1 = 1
x 2 g 2(l) l + l 2 + l 31
♦ c is a constant, n = 3, u = m = 2
♦ We cannot come up with the g matrix
♦ To solve this problem we usually linearize the model using a Taylor Series
expansion (will be discussed later)
Conditional Model
♦ A special case of the direct model, where no parameters are expressed in the
model
0 = g m,1 (l n,1 )
Unknowns:
u = 2 (any two angles)
γ Observations:
n=3
Functions:
m = number of independent conditions
α β m=n-u=3–2=1
Math Model
α + B + γ − 180 = 0
α
[1 1 1] β − 180 = 0
γ
Page 10
ENGO361 Naser El-Sheimy
∆H3
BM2
B
∆H2
A
BM1 ∆H1 ∆H4
∆H5
∆H 1
∆H
H a 2
x = H b Unknowns: l = ∆H 3 Observations:
u=3 n=5
H c ∆H 4
∆H 5
H Functions:
c = BM 1 Constants
H BM 2 m = number of independent conditions
m=n–u=5–3=2
Math Model
H BM 1 + ∆H 1 + ∆H 2 + ∆H 3 − ∆H BM 2 =0
+ ∆H 2 + ∆H 4 − ∆H 5 =0
∆H 1
∆H
1 1 1 0 0
2
H1 − H 2
0 1 0 1 − 1 ∆H 3 + 0 =0
2 x5 2 x1
∆H 4
∆H 5
5 x1
Page 11
ENGO361 Naser El-Sheimy
n = m
Linear
♦ Example – Consider the same Levelling Network (Arrow pointing at the higher
station)
∆H3
BM2
B
∆H2
A
BM1 ∆H1 ∆H4
∆H5
∆H 1
∆H
H a 2
x = H b Unknowns: l = ∆H 3 Observations:
u=3 n=5
H c ∆H 4
∆H 5
H Functions:
c = BM 1 Constants
H BM 2 m = number of observations
m=n=5
Math Model
∆H 1 1 0 0 − H BM 1
∆H − 1 1 0 0
a
2 H
∆H 3 = 0 − 1 0 H b + H BM 2
∆
4 H 0 − 1 1 H c ux1 0
∆H 5 − 1 0 1 0
nx1 nxu nx1
Redundancy = m – u = 5 – 3 = 2
Page 12
ENGO361 Naser El-Sheimy
Non-linear
♦ Non-linear Models will be linearised using a Taylor Series expansion (will be
discussed in Chapter 6)
♦ Example – finding the co-ordinates of a point by resection
Unknowns:
Y x
x = u = 2 (any two angles)
(x2, y2) y ux1
(x1, y1)
Observations:
L2 L1
(x3, y3)
L1 l = L2 n = 3
L3
L3
Functions:
(x, y)
m = number of observations
m=n=3
redundancy = degrees of freedom = m – u = 3 – 2 = 1
Math Model:
[
L1 h1 ( x) ( x1 − x ) + ( y1 − y )
2 2
]
1
2
2 2 2 [
L = h ( x ) = ( x − x )2 + ( y − y )2
1 ]
1
2
1
[
L3 h3 ( x) (x − x ) + ( y − y )2
3
2
3 ] 2
Page 13
ENGO361 Naser El-Sheimy
♦ Example: Consider the example of line fitting and knowing that point (2) is
equidistant from points (1) and (3). This condition can be added to the
mathematical model
Observations:
Y
l = [x1 y3 ]
T
y1 x2 y2 x3 (n = 6)
b
2 Unknowns:
3
x = [a b] (u = 2)
T
1
a Functions:
X
F(x,l) = 0
f = [ f1 f2 f3 ]
f1 ⇒ 0 = a + bx1 − y1
f 2 ⇒ 0 = a + bx 2 − y 2 (m = 3)
f 3 ⇒ 0 = a + bx3 − y 3
Page 14
ENGO361 Naser El-Sheimy
♦ Example – Consider the same Levelling Network (Arrow pointing at the higher
station) and knowing that Stations A and B are at the edge of lake (i.e. Ha = Hb)
∆H3
BM2
B
∆H2
A
BM1 ∆H1 ∆H4
∆H5
C
Math Model
L = h(x)
∆H 1 1 0 0 − H BM 1
∆H − 1 1 0 0
a
2 H
∆H 3 = 0 − 1 0 H b + H BM 2
∆
4 H 0 − 1 1 H c ux1 0
∆H 5 − 1 0 1 0
nx1 nxu nx1
h(x) = 0
H a - Hb = 0
H a
[1 − 1 0] H = [ 0]
b
H c ux1
Page 15
ENGO361 Naser El-Sheimy
3. ERROR PROPAGATION
♦ Error Propagation: is the process of evaluating the errors in estimated quantities
(X) as functions of the errors in the measurements (L)
♦ Concept:
l, Cl xˆ , C xˆ
Observations Best estimate of the unknowns
Variance-Covariance matrix of Variance-Covariance matrix of
the observations the estimated unknowns
y = a + bx (1)
Y
And using the concept of true value, as
introduced in chapter (1)
b
y
yt = a + bxt (2)
yt
Defining the error of a measurement as the measured
value minus the true value , (1) – (2) a
y – yt = b(x-xt) X
xt x
dy = b⋅dx
♦ In general:
Given:
Observations: l = [l1 ln ]
T
l2 (recall l1 is univariate)
1/14
/
ENGO361 Naser El-Sheimy
Covariance information: C l
Required
Covariance information: C xˆ
With…
u = number of unknowns
n = number of observations
r = degrees of freedom
♦ Classification of problems
r=0 r >0
we have nnecessary = n no. of equations > no. of
no. of equations = no. of unknowns
unknowns requires adjustment of
unique solution observations to reach a unique
solution
We will study: We will study:
Univariate error propagation The method of Least Squares
Multivariate error propagation
Error propagation
r=0
Univariate Multivariate
dim(x) = 1
Cl = Diagonal
“Uncorrelated dim(x) = 1 dim(x) = u dim(x) = u
Observations” Cl ≠ Diagonal Cl = Diagonal Cl ≠ Diagonal
2/14
/
ENGO361 Naser El-Sheimy
3.1. Univariate
♦ Characteristics:
x = [x1 ]1x1
♦ Steps of Solution:
1. Construct the Mathematical model (direct model)
x = f (l)
Where
l = [l1 l2 ... l n ]
2. Obtain the best estimate of x
ˆ f (l)
x=
3. Estimate the precision of x̂
2 2 2
∂f ∂f 2 ∂f 2
σ = σ l21 +
2
xˆ
σ l + ... + σ l
∂l1 ∂l 2 ∂l n
2 n
2
∂f
n
2
σ = ∑
2
xˆ
σ l → Observe the units
i =1 ∂l i
i
♦ Example:
Given:
x=A
L = (a,b) a A
a =30m σ a = 0.1m
b = 40m σ b =0.2m
b
Required: Aˆ and σ Aˆ
Solution:
1. Mathematical Model
x = f (l ) → A = a ⋅ b
2. Best Estimate
xˆ = f (l ) → Aˆ = a ⋅b = 30 m⋅40 m = 120 m 2
3/14
/
ENGO361 Naser El-Sheimy
σ 2
Aˆ
(b 2
)σ 2
a + (a 2
)σ 2
b
σ A = 7.211m 2
Best estimate of A = 120 m2 ± 7.211 m2
♦ Example on Univariate Error Propagation
Given:
Mean σ n
a A
d 56.78m 2cm 4
θ 9°12'7" 30" 9
b
Required: hˆ and σ hˆ
Solution:
1. Mathematical Model
x = f (l ) → h = d tan θ
2. Best Estimate
xˆ = f (l )
4/14
/
ENGO361 Naser El-Sheimy
σd 2 2
σ dˆ = = = = 1cm
4 4 2
σθ 30 30′′
σ θˆ = = = = 10′′
a 9 3
2
∂h ∂h
= tanθ (unit less ) → σ d2ˆ → cm 2
∂d ∂d
2
∂h ∂h
∂θ
( )
= d sec 2 θ units m 2 → σ θ2ˆ → m 2 sec 2
∂θ
2
10′′
( 2 2
) ( 2
σ h2 = tan9 0 12′17′′ (1) + 56.78x100sec 2 9 0 1217 ′′ )
206265
ĥ = 9.1983 ±
x = f (l1 , l 2 ,...l n ) = f (l )
with the variance of x, σx , derived from the Law of Propagation of variances as:
2 2 2
∂x ∂x 2 ∂x 2
σ = σ l21 +
2
x
σ l2 + ...... + σ ln
∂l1 ∂l 2 ∂l n
∂x ∂ x ∂ x 2 ∂x ∂x 2 ∂x
= σ l21 + σ l2 + ...... + σ ln
∂l1 ∂l 2 ∂l 2 ∂l 2 ∂l n ∂l n
5/14
/
ENGO361 Naser El-Sheimy
∂x
σ 2
l1 0 ... 0 ∂l1
∂x
∂x ∂x ∂ x 0
σx =
2
...
σ l22
∂l1 ∂l 2 ∂l n M O ∂l 2
M
2
0 σ ln ∂x
∂l n
♦ Now if we have more than one unknown, say a vector x of u unknowns, that are
related to the n observations as follows:
x1 = f1 (l1 , l 2 ,..., l n )
x 2 = f 2 (l1 , l 2 ,..., l n )
M
x u = f u (l1 , l 2 ,..., l n )
♦ In this case, σx will be a matrix Cx with dimensions u x u that takes the following
form:
♦ More over, the Jl matrix will contain all the partial derivatives
∂x j
j = 1,2,...u i = 1,2,...n
∂l i
6/14
/
ENGO361 Naser El-Sheimy
♦ Taking all the above characteristics of the multivariate into account, the final
covariance law will be:
C x = J l ⋅ C l ⋅ J Tl
♦ Summary of steps for solving a Multivariate Error Propagation:
Given: l nx1 , C l nxn
Required: xˆ ux1 , C xˆ uxu
1. Form the direct (explicit) mathematical model
x = f(l)
2. Establish the variance-covariance matrix of the observations Cl
7/14
/
ENGO361 Naser El-Sheimy
♦ Example:
Given:
HA = 20 m
HA = 20 m (BM)
∆h1 = ?
∆h 5
l = 1 = m H1 = ?
∆h2 8
4 0
Cl = mm
2
∆h2 = ?
0 9 H2 = ?
Required:
H
x = 1 and Cx
H 2
8/14
/
ENGO361 Naser El-Sheimy
C x = J l ⋅ C l ⋅ J Tl
σ H2 1 σ H1H 2 1 0 4 0 1 1
=
σ H 2 H1 σ H2 2 1 − 1 0 9 0 − 1
4 0 1 1 4 4
= mm 2
4 − 9 0 − 1 4 13
From which we get
σ H1 = 4 mm 2 = 2 mm σ H 2 = 13 mm 2 = 3.6 mm
σ H1 H 2 4
ρ H1H 2 = = = +0.55 ( Signifcant )
σ H1 σ H 2 2 ⋅ 3.6
9/14
/
ENGO361 Naser El-Sheimy
♦ Types of analysis
σ x or C x σ x or C x
Pre-specified accuracy Actual accuracy
L hr
L sin θ
B θ
∆h2 ∆h
hI
∆h1
A A
HB = HA + Σ∆hi HB = HA + hI + L sin θ - hr
(Spirit leveling) OR (Trigonometric levelling)
10/14
/
ENGO361 Naser El-Sheimy
Recall:
2 2 2
∂x σ l2i ∂x σ l 2 σ
2 2
σ =
2
+ + ....... + ∂x l n
x
∂l i ni ∂l 2 n2 ∂l n nn
14444444444244444444443
n −terms
2
n
∂x σ l2i
σ 2
= ∑
x
i = i ∂l i
ni
∂x
… Effect of the math model
∂l i
σ l2i
… Effect of the instrument and the number of
ni
observations
Note:
∂x
… will depend on the mathematical model
∂l i
11/14
/
ENGO361 Naser El-Sheimy
∂x
♦ Usually and σ li are related and easy to decide upon and therefore the number
∂l i
of observations ni are the main quantities we are interested in.
♦ Since we have only a single equation and n unknowns, we can not estimate ni
unless we have some additional information or make some assumptions (e.g., the
n terms equally contribute to the total error budget).
♦ Example:
The (d) distance can’t be measured
directly, however, we can measure
s1 s2 and α.
We know the following: d
Measurement Standard deviation
s1 = 136 m σ s1 = 1.5 cm s2
α
s 2 = 115 m σ s2 = 1.5 cm s1
α = 50° σ α = 10"
Note: the value of the
measurements can be obtained from
a map (i.e. approximate values)
Required
Solution:
1. Mathematical Model
(
d = s12 + s 22 − 2s1 s 2 cos α ) 1
2
= 107.77 m (cosine law)
2. Error Model
2 2 2
∂d 2 ∂d 2 ∂d 2
σ =
2
σ s1 + σ s2 + σα
∂s1 ∂s 2 ∂α
d
∂d 2 s1 − 2 s 2 cos α
= = 0.576 (unitless)
∂s1 2d
∂d 2 s 2 − 2 s1 cos α
= = 0.255 (unitless)
∂s 2 2d
∂d 2 s1 s 2 cos α
= = 111.17 m = 11117 cm
∂α 2d
σ α~2
(0.5 cm) 2
= (0.576) σ + (0.255) σ
2 2 2 2
+ (11117 cm )
2
s1 s2
( ρ )2
12/14
/
ENGO361 Naser El-Sheimy
σ s21 σ s22 σ α2
0.25cm 2 = 0.332 + 0.065 + 0.003
ns1 n s2 nα
0.25
= 0.332
(1.5) 2
→ n s1 = 9
3 n s1
0.25
= 0.065
(1.5) 2
→ n s2 = 2
3 n s2
0.25
= 0.003
(10) 2
→ nα = 4
3 nα
n s2 n s2 nα
↓ ↓ ↓
ns1 = 5 ns2 = 3 nα = 6
13/14
/
ENGO361 Naser El-Sheimy
♦ Example:
The height h of a survey
station (A) above the reflector
instrument at (B) is
required with an accuracy
hr
of 0.01m.
S
A
h = ssin α − hr
B α h
s = 400m α = 30o
ρ ′′
s
(0.0056)2 m 2 = (0.5)2 σ s2 + σ t2
Balancing the accuracies of the two terms
0.0056 / 2
σs = = 0.008m (8mm ) → choose an EDM
0.5
0.0056 / 2
σt = = 0.004m (4mm )
1
14/14
/
Naser El-Sheimy ENGO361
4. ADJUSTMENT OF OBSERVATIONS
4.1. The concept of adjustment
♦ In the previous chapters, we dealt with the case where the number observables (n)
is just enough to provide the necessary number of equations (m or nnecessary) to
estimate (u) number of unknowns. This usually results in a unique solution for
the unknowns.
♦ In practice, however, a unique solution is dangerous, as an error in a single
observation can radically effect the final solution for the unknowns. Therefore, in
Geomatics we typically have redundant observations, i.e. having (taking) more
observations than is necessary for a unique solution.
u = 1, n = 3, r = u – n = 3 – 1 = 2
Due to the different uncertainty (errors) in the observations ∆hi, the elevation of
point p is not identical.
Page 1 of 17
Naser El-Sheimy ENGO361
2. Area of Triangle
u = 1, nnecessary = 3, n = 6
1 B
∆1 = a b sin C
2
1
∆ 2 = cbsin A \
2
c a
1
∆ 3 = a csin B
2
Again, identical results may not
(most probably) be obtained, due to
the errors in the observation vector A b C
L.
Where ˆl = l + vˆ
n,1 n,1 n,1
Note: The estimated residuals v̂ are unknown and must be determined before the
observations can be estimated.
Page 2 of 17
Naser El-Sheimy ENGO361
Estimated
v = v1 v2 … vn
given
l = l1 l2 … ln Mathematical
Model
required x̂
n>u Unique
x = x1 x2 … xn i.e, r > 0 solution
♦ Principle:
In addition to the fact that the adjusted observation must satisfy the mathematical
model exactly, the corresponding residuals must satisfy the Least Squares
criterion.
n
φ = ∑ v i2 = v12 + v 22 + .......... + v n2 = min
i =i
x –––––––– u (unknowns) = 1
Given:
l 15.12
l = 1 = m –––––– n (observations) = 2
l 2 15.14
r (redundancy/extra observations) = n – u = 2 – 1 = 1
The final value of x̂ (best estimate) can be obtained from the observation
equations.
Page 3 of 17
Naser El-Sheimy ENGO361
For For
l1 l2 lˆ1 lˆ2 ∑v 2
Note that φ2 is the smallest, but is it the very minimum value when all possible
combinations of corrections are considered?
♦ Geometric Interpretation of ∑v 2
= min :
The adjusted observations lˆ1 and lˆ2 are related to each other by lˆ1 - lˆ2 = 0 or lˆ1 =
lˆ which is a line with 45o inclination (condition line)
2
A1, A2, and A3 correspond to φ1, φ2 and φ3 respectively, and all the 3 points satisfy
the condition line that is lˆ1 = lˆ2 where,
15.13
[0.01 − 0.01] =0.
15.13
Page 4 of 17
Naser El-Sheimy ENGO361
This means that the vector v2 is perpendicular to the condition line and therefore it
is the minimum function.
♦ Proof that the mean is the Least Squares estimate for a group of measurements of
certain parameters:
[
Given: l1 l 2 ... l n ]
Required: x
l1 + v1 = xˆ
l 2 + v2 = xˆ X
X
l n + vn = xˆ
Using the least squares condition:
φ
n
φ = ∑ vi2
i =i
X
X̂
∂φ
=0
∂xˆ
∂φ
= 2( xˆ − l1 ) + 2( xˆ − l 2 ) + .... + 2( xˆ − ln ) = 0
∂xˆ
0 = 2 n xˆ − 2(l1 + l 2 +........ + l n )
n
0 = nxˆ − ∑ li
i =i
xˆ =
∑l i
n
If the observations are unequal in precision, we have to consider the weight of the
observations, that is
xˆ =
∑ρl i i
⇒ Weighted mean
∑ρ i
Page 5 of 17
Naser El-Sheimy ENGO361
♦ Proof that the weighted mean is the Least Squares estimate for a group of
observations, with different precision, of the same parameters:
l i + v i = xˆ
φ = ∑ρv i i
2
= min where
v i = xˆ − l i
subsitute by v
φ = ∑ ρ (xˆ − l ) = min
2
i i
φ = ∑ ρ (xˆ − 2 xˆ l + l ) = min
2 2
i i i
φ = xˆ ∑ ρ − 2 xˆ ∑ ρ l + ∑ ρ l
2
i i i i
2
i = min
For φ to be minimum
∂φ
=0
∂xˆ
∂φ
= 2 xˆ ∑ ρ i − 2∑ ρ i li = 0
∂xˆ
xˆ =
∑ ρ i li
∑ ρi
It is very important to note that for
∂φ
φ = ∑ v 2 = min = 2∑ v i = 0 → ∑ v = 0
∂v
and
∂φ
φ = ∑ ρ i v i2 = min = 2∑ ρ i v i = 0 → ∑ ρ i v i = 0
∂v
(check chapter 1 and lab 1)
♦ Based on the above discussion, the basic concept of adjustment is, therefore, to
allow the observables l to change slightly while solving for x.
♦ This means that in the over determined model f(x,l)=0, we consider (l) as
approximate values for the observables which need to be corrected by certain
small amount, denoted by v so as to yield a unique solution. (v is the vector of
residuals)
♦ The mathematical model becomes:
( )
f xˆ u,1 , ˆl n,1 = f (xˆ u,1 , l n,1 + v n,1 ) = 0
(unknowns…) x̂ is the best estimate (adjusted) valves of the parameters
l̂ is the adjusted values of the observables
l is the original observation
(unknowns…) v is the residuals vector
Page 6 of 17
Naser El-Sheimy ENGO361
♦ The above model cannot be solved for x and v simultaneously because we have
more unknowns (u “for x” + n “for v”) in n equations (that is we need extra u
equations)
♦ Recall, the residuals v are very small and will behave according to Gauss Law of
Random errors (Σv = 0). As a result we can find several conditions that can be
used to provide us with the required u extra equations which will enable us to
solve for both x and v simultaneously.
♦ The condition of Least Squares of the residuals (Σv2 = min) was found to satisfy
the properties of the best estimate:
1) Maximum
Likelihood
most precise not precise
(most
probable)
2) Minimum
variance 0 0
(most
precise)
3) Unbiased accurate not accurate
(most
accurate)
0 0
Page 7 of 17
Naser El-Sheimy ENGO361
General Model
Implicit Model
f(xu,1, ln,1) = 0
• X and L can not be written
as an explicit function
♦ These math models can be either linear or non-linear. Non-linear models should
be linearised first before conducting the Least Squares Adjustment.
4.4. Linearisation of Non-linear Models
l = f(x)
∂f
∂f
l = f(x o ) +
∂x x 0 (x − x )
0 L + V = f(X0+δ) tanθ =
∂x
f(X0)
1 ∂ 2f
+
2! ∂x 2 x0
(
x − x0
2
)
X
+ higher order terms X0
X = X0 +
Page 8 of 17
Naser El-Sheimy ENGO361
∂f
( )
l = f x0 +
∂x x0
.δ
∂f
(( ) )
0 = f x0 − l +
∂x x0
δ
w + A δ =0
n×1 n×u u×1
2) Multivariate Functions
fm,1 (xu,1 , ln,1) = 0
POE:
xo – approximate values of the unknowns can be estimated from u
group of observations ( use the number of observations necessary
to solve for x)
lobs – observed values
f (x, l ) ≈ f (x 0 , l obs ) +
∂F
∂x x 0 Lobs
(xˆ − x ) + ∂f
0
∂l x0l
(ˆl − l )
obs
In matrix form
w + A δ + B v = 0
m×1 m×u u×1 m×n n ×1
Page 9 of 17
Naser El-Sheimy ENGO361
♦ From the multivariate linearisation, we can drive the two general adjustments
1) Parametric Model
l n,1 = f n,1 (x u,1 ) or f n,1 (x u,1 ) − l n,1 = 0 n,1
POE − x 0 ,l obs
∂f ∂L
( )
f x 0 − l obs + δ −
14243 ∂x ∂ L
v=0
w { {
A I
2) Conditional Model
f r,1 (l n,1 ) = O r,1
POE − l obs
∂f
( )
f l obs +
∂L
v=0
x = [hB hC ] C
T
∆hˆ1 = hˆB − hA
Page 10 of 17
Naser El-Sheimy ENGO361
h1 + v1 = hBo + (1)δ hB
1
424
(
v1 = hBo − h1 + δ hB
3
)
w
h2 + v2 = hˆB − hˆC
( )
= hBo − hCo + δ hB − δ hC
δ h
( )
v2 = hBo − hCo − h1 + [1 − 1] B
δ hC
1424 43 4
w
A similar equation can be derived for h3, and together the three equations can be
written in matrix form:
v1 1 0 hBo − h A − h1
δ
v = − 1 1 hB + h o − h o − h
2 δ h B C 2
v3 0 1 C
hC − h A − h3
o
v n,1 = A n,u δ u,1 + w n,1
Unknowns – u = 2 dAP
x = [x P yP ]
T
(xP, yP)
Observation Equation – ˆl = f (xˆ )
Page 11 of 17
Naser El-Sheimy ENGO361
v dAρ = w + A δ
∂d Aρ ∂d Aρ δx ρ
( ( )
v dAρ = f x 0 − l obs + )
∂x ρ ∂y ρ δy ρ
∆x 0 ∆y 0 δx ρ
( )
v dAρ = d A0 ρ − d Aobsρ + 0 0
d Aρ d Aρ δy ρ
∂d Aρ 2( y ρ0 − y A )(1) ∆y A0 ρ
= = 0
∂y ρ 2 (x ρ0 − y A ) + ( y ρ0 − y A ) d Aρ
2 2
∆x A0 ρ ∆y A0 ρ
0
d d A0ρ
vdAρ Aoρ d 0 − d Aobsρ
∆x Bρ ∆y B0ρ δx A A0ρ
v dBρ = d0 + d Bρ − d Bobs
d Bρ δy A
0 ρ
vdCρ B0ρ d C0ρ − d Cobs
∆x ∆yCρ0 ρ
0Cρ
d Cρ d C0ρ
v n,1 = A n,u δ u,1 + w n,1
h1 A
L = h2 , n = 3 h1
h3 B
h3
h
x = B , u = 2
hc h2
r = n −u = 3− 2 =1
C
Page 12 of 17
Naser El-Sheimy ENGO361
()
g ˆl = 0
hˆ1 − hˆ2 − hˆ3 = 0
Expand into differential form:
∂f 0 obs
∂l f(l ) = f(l )
B v + w = 0
r×n n×1 r×1 r×1
v1
[1 − 1 − 1]v2 + [h1 h2 h3 ] = 0
v3
• The number of equations is fewer than in the parametric model, but the condition
model is more difficult to program
• If there is more than one condition, they must be independent (that is you cannot
get one equation from the other condition equations … one of the problems with
conditional adjustment)
• Following the estimation of the v n,1 vector, we can estimate the best estimate
value of the unknown parameters.
• ˆl = l + v
n,1 n,1 n,1
• Since there are more observations than unknowns, there are several possibilities
for the direct model equations.
• All possibilities will a be equivalent if the adjusted observations are used, i.e
xˆ = f(ˆl)
hˆ
hˆB 1 0 0 ˆ1 hA
ˆ = h2 +
hc 0 0 1 ˆ hA
h3
Another possibility is:
hˆB = hA + h1
Page 13 of 17
Naser El-Sheimy ENGO361
hˆc = hA + h1 − h2
hˆ
hˆB 1 0 0 ˆ1 hA
ˆ = h2 +
hc 1 − 1 0 hˆ hA
3
xi Y
y
i
x2 (x1, y1)
xc
y •
l8,1 = 2 x = y c r
x3 (x4, y4) •
γ (xc, yc)
y3
x • • (x2, y2)
4 (x3, y3)
y 4
X
n=8 u=3
Math Model f(x,l) = 0:
Page 14 of 17
Naser El-Sheimy ENGO361
(x
1 )
− xc0
2
+ (y 1 − y c0 ) 2
− r 02
) (x −x ) (y )
0 2 2
( ) (
w 4,1 = f x 0 , l 0 = f x 0 , l obs = 2 c + 2 − y c0 − r 02
(x
3 −x ) 0 2
c + (y 3 − y c0 ) 2
− r 02
(x
4 −x ) 0 2
c + (y 4 − y c0 ) 2
− r 02
∂f1 ∂f1
∂x 0 0 0 0 0 0
∂y1
1
0 ∂f 2 ∂f 2
0 0 0 0 0
∂f ∂x2 ∂y 2
B 4,8 = =
∂l 0 ∂f 3 ∂f 3
0 0 0 0 0
∂x3 ∂y3
∂f 4 ∂f 4
0 0 0 0 0 0
∂x4 ∂y 4
eg. ( x1 − xc ) + ( y1 − y c ) − r o 2 = 0
2 2
∂f1
∂x1
(
= 2 x1 − xc0 )
∂f1 1st row of the B matrix
∂y1
(
= 2 y1 − y co )
∂f1
∂xc
(
= −2 x1 − xc0 )
∂f1
∂y c
(
= −2 y1 − y c0 )
1st row of the A Matrix
∂f1
= −2r 0
∂r
T
v 1,n [
= v x1 v y1 vx2 vy2 vx3 v y3 vx 4 vy4 ]
xˆ = x o + δˆ
Page 15 of 17
Naser El-Sheimy ENGO361
♦ As discussed before, the best estimate is the one that satisfies the Least-Squares
Condition:
φ = ∑ v 2 = min or ∑ pv 2
= min
φ = v 1,n
T
Pn,1 v n,1 = min
Note that: δ 1,Tu A u,T n Pn,n Wn,1 = W1,Tn Pn,n A n,u δ u,1
φ = δ T (A T PA )δ + 2(w T PA )δ + w T Pw = min
♦ The condition for φ to be minimum is that it derivative with respect to all
variables must be zero. The only variable in the minimisation (variation) function
φ is the vector δ.
∂φ
♦ Thus for φ = min → =0
∂δ
∂φ
∂δ
(
= 2δ T A T PA + 2 w T PA + 0 = 0 ) ( )
♦ Transpose the whole equation (note: PT = P symmetric matrix)
(1A44P24A43) δˆ
T
u,n n,n n,u u,1 + A u,T n Pn,n w n,1 = 0
14243
u,u u,1
Page 16 of 17
Naser El-Sheimy ENGO361
δˆ = −N −1u
(
δˆ = − A T PA )−1
A T Pw
xˆ = x o + δˆ
↓
adjusted parameters
♦ It is interesting to note here that the normal equations system can be derived
directly from the variation function in a more direct way:
φ = v T Pv = min
Where
v = Aδ + w .
And therefore,
∂φ ∂φ ∂v
= ⋅
∂δ ∂v ∂δ
(
= 2v T P ⋅ (A ) = 0 )
Which, after transposing, yields the following
A T Pv = 0
A T P(Aδ + w ) = 0
(A T
)
PA δ + A T Pw = 0
N δ + u =0
♦ Solution for the residuals v
vˆ = Aδ + w
= − A A T PA( ) −1
A T Pw + w
( ( ) A P + I )w
= − A A T PA
−1 T
vˆ = (I − A(A PA ) A P )w
T −1 T
♦ Adjusted observations
ˆl = l obs + vˆ
n,1 n,1 n,1
Page 17 of 17
5. PARAMETRIC LEAST-SQUARES ADJUSTMENT
♦ This chapter provides more detail on the Least Squares solution equations for the
parametric method (also known as Observation Equation Least Squares).
5.1. Estimated Parameters and Adjusted Observations
xˆ = x o + δˆ
♦ As discussed before, the best estimate is the one that satisfies the Least-Squares
Condition:
φ= ∑v 2
= min or ∑ pv 2
= min
In matrix form, this is:
Note that: δ 1,Tu A u,T n Pn,n Wn,1 = W1,Tn Pn,n A n,u δ u,1 = scalar quantity,
therefore:
δ 1,T u A Tu,n Pn,n Wn,1 + W1,Tn Pn,n A n,uδ u,1 = 2( W1,Tn Pn,n A n,uδ u,1 ) = 2(δ 1,T u A Tu,n Pn,n Wn,1 )
φ = δ T (A T PA )δ + 2(w T PA )δ + w T Pw = min
♦ The condition for φ to be a minimum quantity means its derivative with respect to
all variables within the equation must be zero. The only variable in the
minimisation (variation) function φ is the vector δ.
∂φ ∂ ( X T CX )
♦ Therefore, for φ = min → = 0 (recall from Lab 1: = 2X TC )
∂δ ∂X
∂φ
= 2δ T (A T PA ) + 2(w T PA ) + 0 = 0
∂δ
♦ Transpose the whole equation (note because P is symmetric matrix, then: PT = P)
(1A44P24A43) δˆ
T
u,n n,n n,u u,1 + A u,T n Pn,n w n,1 = 0
14243
u,u u,1
δˆ = −N −1u
(
δˆ = − A T PA )
−1
A T Pw
xˆ = x o + δˆ
↓
adjusted parameters
♦ It is interesting to note here that the normal equations system can be derived
directly from the variation function in a more direct way:
φ = v T Pv = min
Where
v = Aδ + w .
And therefore,
∂φ ∂φ ∂v
= ⋅
∂δ ∂v ∂δ
= 2v T P ⋅ (A ) = 0 ( )
Which, after transposing, yields the following
A T Pv = 0
A T P(Aδ + w ) = 0
(A T
)
PA δ + A T Pw = 0
N δ + u =0
♦ Solution for the residuals vector (the correction to the observations):
vˆ = Aδ + w
(
= − A A T PA )−1
A T Pw + w
( ( ) A P + I )w
= − A A T PA
−1 T
vˆ = (I − A(A PA ) A P )w
T −1 T
ˆl = l obs + vˆ
n,1 n,1 n,1
5.2. Estimated Variance-Covariance Matrices for the Adjusted
Parameters and the Adjusted Observables
♦ The general only way to compute C xˆ and Cˆl is through the use of covariance
law.
δˆ = − N −1u
= −(A T PA ) A T Pw
−1
= −(A T PA ) A T P(f (x 0 ) − l )
−1
(x 0 ) + (1A4
= −(A T PA ) A T P f{ PA ) A T P {
−1 T −1
l
144244 3 4244 3
CONSTANT
14 4424443 14
CONSTANT
4244
CONSTANT
3 VARIABLE
K2 K1
= ( N −1 A T P) σ P −1 (N −1 A T P )
2 T
0
= N −1 A Tσ P P −1 (N −1 A T P )
2 T
0
123
I
=σ 2
0
N −1 A T I (N −1
A T P)
T
=σ 2
0
N −1 A T (P A N −1 )
=σ 2
0
N −1 A T ( P A N −1 )
=σ 2
N −1 A T P A N −1
0
1424 3
N
=σ 2
0
−1
N NN −1
Cδ = σ 2
0
N −1
That is, the variance-covariance matrix of the solution vector δ
Cδ = σ 02 N −1
xˆ = x{o + δˆ
constant
C xˆ = JC δ J T
T
∂xˆ ∂xˆ
= Cδ
∂δ ∂δ
σ 02 ≠ 1, then
∴ C xˆ = Cδ = N −1 If C xˆ = σˆ 02 N −1
2
Note: we have introduced a new quantity σ̂ 0 which is called the a posteriori variance
v T Pv
factor, where
σ̂ =
2
0 (degree of freedom (d.o.f.) = n – u)
d.o.f
5.2.3. C ˆl - The V-C matrix of the adjusted observations
Function Model
ˆl = f (xˆ ) = Axˆ
T
∂ˆl ∂ˆl
C ˆl = C xˆ = JC xˆ J T
∂xˆ ∂xˆ
C ˆl n.n = A n,u C xˆ u,u A u,T n
We can also derive an expression for the variance-covariance matrix for v (
the residuals vector )
C v = C l − C ˆl
♦ Steps:
1. Identify the elements of x u,1 and l n,1
2. Form the observation equations l = f(x)
3. Find approximate values for xo ( using u observation equations) and use
them to evaluate the design linearised equations:
v n,1 = A n,u δ u,1 + w n,1
Where
∂f
A=
∂x x = xo
( )
w n,1 = f n,1 x o − l n,1
sym 6 0 sym 1
6 0
9 1
9
x 02 = xˆ 1
• Calculate A2, W2 with x 02 and l
• Solve for δˆ 2 = − N 2 u 2 (step 6)
−1
♦ The opposite sketch shows a leveling network abcd, in which point (a) is assumed
to be fixed with zero elevation.
Section
b
length
Observations:
h1 6.16 4 km
h 12.57 h4 h6
2 2 km
h5
h3 6.41 2 km
l 6,1 = = m d
h4 1.09 4 km
h5 11.58 a c
2 km h2 h3
h6 5.07 4 km Ha = 0.0
Parameters (unknowns):
hb h1
x 3,1 = hc
hd
♦ Required
given that the variance of the elevation differences = 1 cm2/km estimate the
elevation of point b, c, and their variance-covariance matrix C xˆ .
Note: n = 6, u = 3
♦ Solution
hˆ1 = hˆc − ha
hˆ2 = hˆd − ha
ˆh = ˆ ˆ
− hc + hd
3
hˆ4 = hˆb − ha
hˆ5 = − hˆb + hd
ˆ ˆ
h6 = − hb + hc ˆ
{ 1444424444 3
l f ( xˆ )
1.09
x o
3,1 = 6.16 m
12.57
0
w1 h0c − ha h 0.0
1
w h − h h 0.0
2 d0 a
2
w3 − hc + hd h3 0.0
= 0 − = m
w4 hb − ha h4 0.0
w5 0 0 h5 − 0.1
− hb + hd
w6 0 0 h6 0.0
− hb + hc {
1 l
44244 3
( )
f x0
Note: Equations 4, 1, and 2 will, definitely, have zero elements for the w vector
4. Write down the linearised equations – v 6,1 = A 6,3 δ 3,1 + w 6,1
0 1 0
0 0 0
0 −1 1
A 6,3 = unitless
1 0 0
− 1 0 1
− 1 1 0
C l = 10 −4 [ 4 2 2 4 2 4 ]m 2
And recalling that P matrix α C l−1 P = σ 02 C l−1
If we assume σ 02 = 10 −4
∴ P = 10 −4 ⋅ 10 −4 [1 4 1
2
1
2
1
4
1
2
1
4 ] 1m 2
∴ P = [1 4 1
2
1
2
1
4
1
2
1
4 ] 1m 2
7. Adjusted observations ˆl = l + v
0.00 6.16 0.00 6.16
0.02 12.75 0.02 12.59
0.02 6.41 0.02 6.43
vˆ = Aδˆ + w = m ˆl = l + vˆ = + + m
− 0 . 04 1.09 − 0.04 1.05
− 0.04 11.58 − 0.04 11.54
0.04 5.07 0.04 5.11
σ h = 10 −4 x10.67 = 3.27cm
b
σ h = 10 −4 x10.67 = 3.27cm
c
σ h = 10 −8 x8 = 2.83cm
d
0.0
Homework: Try to solve this problem with x = 0.0 , you should end up
o
0.0
with the same results. In general, for any linear model, you can always
assume x o = 0
6. CONDITIONAL LEAST SQUARES ADJUSTMENT
♦ The general form of the Conditional mathematical model was given before as:
( )
f r ,1 ˆl n ,1 = 0 r ,1 where ˆl = l + vˆ
l̂ is the vector of adjusted observations
v is the vector of adjusted residuals
r (degrees of freedom) = n –u
♦ Two basic properties must be satisfied for the conditional model:
1) Number of equations = Number of degrees of freedoms. This means that each
redundant observation provides one independent condition equation.
2) The equations describe the functional relationship among the observations
only. This obviously indicates that the unknown parameters x will not be
among the direct output of the conditional adjustment. Thus the adjusted
parameters x̂ and their covariance matrix C xˆ , have to be computed after the
()
adjustment using the direct model ( xˆ = f ˆl ) and the law of propagation of
variances (this is a disadvantage when comparing the conditional and the
parametric adjustment).
h1 + h2 + h3 + (H A − H B ) = 0
h1 h2
♦ Example:
h3
h1 h2
b
a
h4
u=3 h5
n=5
r=2
h1 0
raw(1) 1 − 1 1 0
0 h2 H − H b 0
a
raw(2 ) 0 − 1 0 1 1 h3 + 0 = 0
raw(3) 1 0 1 − 1 − 1 h4 H a − H b 0
h5 0
Note:
Raw (3) = Raw (1) – Raw (2)
(i.e. Raw (3) is linearly dependent on Raw (1) and Raw (2))
2. Traverse Networks
i) Traverse connecting two fixed
points
1 α2 b α3
é α1
α1
a d1
é x1 ù ê
ê
d1 fixed d2 d3
point 2
ê
y1 úú α 2
x= ê l= 2
x2 ú d 2 ∑ ∆x − (x − x a ) = 0 where ∆xi = d i sin α i
ê
i b
ê ú
y2 û α 3 i =i
ë
2
ë 3
d ∑ ∆y − ( y
i =1
i b − ya ) = 0 ∆yi = d i cos α i
n = 4, u = 6
r =6−4= 2
n=8 u=6 α2 α3
r=8–6=2 d2
d1 d3
∑ ∆x i =0 α1
∑ ∆y i =0 α4
d4
4
θ4
∑θ
i =1
i −K =0 θ1
where K = (2S − 4) ⋅ 90 0
θ2
and S = no of stations
e.g. if s = 4 → K = 3600 θ1
θ3
e.g. Triangle θ1 + θ 2 + θ 3 − 180 = 0
6.2. linearized (or reduced) form of the condition equation:
♦ Notes:
After adjustment, the misclosure vector (w) must become zero.
Both matrices B and W are numerically known, the only unknown here is
the residual (v) vector.
♦ Variation function φ in matrix form:
φ = vT P v = min.
♦ In the parametric model derivation, an explicit substitution was made for v. This
cannot be done here since v is not isolated in the functional model.
♦ Could we do the following in order to come up with an expression for (v)?
1. Starting with B v̂ + w = 0
pre-multiply by B-1
B-1 B v̂ = - B-1 w
I v̂ = - B-1 w
This is not possible since B is not square matrix (r < n always)
2. Starting with B v̂ + w = 0
pre-multiply by BT
BT B v̂ = - BT w
v̂ = - (BT B) -1 BT w
To examine the validity of this possibility, look at the rank and dimension
of B and BT B.
Provided that all the condition equations are independent, rank (B) = r (its
smallest dimension).
Since rank (BT B) = rank (B) = r < n and dimension of (BT B) = n, n
Therefore, BT B is singular (i.e., (BT B) -1 does not exist)
ϕ = vˆ 1,n
T
Pn,n vˆ n,1 + 2kˆ 1,r
T
(B r,n v n,1 + w r,l ) = min
Where k is the vector of Lagrange multipliers (one-multiplier/condition)
(The factor 2 is introduced for convenience only)
Note: the second term equal to zero (B v̂ + w = 0) and therefore does not
change the value of φ
∂φ
= 2vˆ T P + 2kˆ T B = 0
∂v
∂φ
= 2vˆ T B T + 2w T = 0
∂k
♦ Transpose each and divide by 2
P B T vˆ 0 0
B 0 kˆ + w = 0
♦ Solution possibilities to the hyper-matrix system:
1. Invert hyper-matrix
−1
vˆ P BT 0
= −
ˆ
k B 0 w
This solution requires inversion of a large matrix
2. Elimination
For the hyper-matrix system
A B x u 0
+ =
C D y v 0
Ax+By+u=0 (i)
Cx+Dy+v=0 (ii)
Eliminate x from (i)
x = - A-1 (B y + u) (provided A-1 exists!) (iii)
Now substitute (iii) into (ii)
- C A-1 (B y + u) + D y + v = 0
- C A-1 B y - C A-1 u + D y + v = 0
(D - C A-1 B) y + (v - C A-1 u) = 0
In this case
A = P, B = BT, C = B, D = 0, x = v̂ , y = k, u = 0, v = w
substitute into
(D - C A-1 B) y + (v - C A-1 u) = 0
(0 - B P-1 BT) k + w = 0
(B P-1 BT) k = w
∴ k = (B P-1 BT)-1 w = (B Cl BT)-1 w
Note: Since the P matrix is not included in the computation of K or the V vectors,
it is usually at the case to use σ o2 =1
v̂ = - P-1BT k = - Cl BT (B Cl BT)-1 w
♦ Adjusted observation
ˆl = l + vˆ
♦ Estimated variance factor
v T Pv
σ̂ 02 =
r
vˆ = − C l B T (BC l B T ) −1 w
= − C l B T (BC l B T ) −1 f(l) = − Kf(l)
Apply the law of Variance-covariance propagation
T
∂vˆ ∂vˆ
C vˆ = C l
∂l ∂l
∂vˆ ∂ ( Kf ( l )) ∂( f ( l ))
=− = −K = − KB = −C l B T (BC l B T ) −1 B
∂l ∂l ∂l
∴ C vˆ = Cl B T (BCl B T ) −1 BCl
♦ Covariance matrix for the Adjusted Observations – C ˆl
Functional model
ˆl = l + vˆ
= l − C l B T (BC l B T ) −1 w
= l − C l B T (BC l B T ) −1 f(l)
Covariance propagation
T
∂ˆl ∂ˆl
C ˆl = C l
∂l ∂l
∂ˆl
= I − Cl B T (BCl B T ) −1 B
∂l
∴ C ˆl = C l − C l B T (BC l B T ) −1 BC l
∴ Cˆl = C l − C vˆ
♦ Linearised model
Bv+w=0
♦ Initial Computation
∂f
B (1) =
∂l l obs
w (1) = f(l obs )
♦ Iteration (1):
Solve for v̂ (1) ,
∂f
B ( i +1) =
∂l ˆl
(i )
♦ Calculate
♦ The condition model adjustment will give the residual vector (v) and the adjusted
observations along with their respective covariance matrices.
♦ The direct model, then, uses the adjusted observations from the condition model
solution to calculate the parameters (if any)
Functional model
xˆ = g(ˆl)
variance propagation
T
∂xˆ ∂xˆ
C xˆ = Cˆl ˆ
∂ˆl ∂l
(see the next example)
n = 6, u = 3 4
6 5
r = 6-3 = 3
C
Angle Observed value
α1 89 59 58.3
α2 180 00 01.4
270 00 00.2 All the six angles have the same standard
α3
89 59 59.8 deviation σ = 1", that is Cl = I arcsec2
α4
α5 179 59 57.0
α6 90 00 03.1
♦ Solution
1. The three independent condition equations – f(ˆl) = 0
(α̂1 + α̂ 4 ) − α̂ 2 = 0
(α̂1 + α̂ 5 ) − α̂ 3 = 0
(α̂ 2 + α̂ 6 ) − α̂ 3 = 0
Note: we wrote the condition equations in terms of the adjusted
observations.
2. The linearised condition equations – B 3,6 v 6,1 + w 3,1 = 0
1 − 1 0 1 0 0
∂f(l)
B 3,6 = = 1 0 − 1 0 1 0 unitless
∂l
0 1 − 1 0 0 1
(α1 + α 4 ) − α 2 − 3.3
w 3,1 = f(l) = (α1 + α 5 ) − α 3 = − 4.9 arcsec
(α 2 + α 6 ) − α 3 4.3
v=− C
{l B T
{ ( BC lB )
T −1
w
{
unitless
123 arc sec
arc sec2 sec2 arcsec
1arce
42 43
1
arc sec2
3 1 − 1
BC l B = BB = 1 3
T T
1 (B Cl BT should be symmetric)
− 1 1 3
− 0.65
(BB ) w = 2.7
T −1
− 2.55
2.05
− 1.90
− 0 . 15
∴ v = −C l B T (BC l B T ) −1 w = arcsec
− 0.65
2.70
− 2.55
5. The adjusted observations – l̂
αˆ1 90 0 0.35
αˆ 179 59 59.50
2
αˆ 3 270 0 0.05
ˆl = l + vˆ = =
αˆ 4 89 59 59.15
αˆ 5 179 59 59.70
αˆ 6 90 0 0.55
check f(ˆl) = 0
6. The variance-covariance matrix of the residual vector – C vˆ
Cv = Cl BT (B Cl BT)-1 B Cl
Remember: These matrices have been computed before in the computation
of v̂
7. The variance-covariance matrix of the adjusted observations – C ˆl
C ˆl = C l − C vˆ
x̂ = g(l̂)
α̂ 1
α̂
α̂ 1 1 0 0 0 0 0 2
α̂
x̂ = α̂ 4 = 0 0 0 1 0 0 ⋅ 3
α̂ 6 0 0 0 0 0 1 α̂ 4
α̂
5
α̂ 6
C xˆ = JC ˆl J T
6.6. Summary of Parametric and Condition Least Squares
C xˆ = σ̂ 02 (A T PA) −1
C ˆl = AC xˆ A T
v T Pv
where σ̂ 02 =
r
End of adjustment.
No additional x̂ = g(l̂) More computations
computations are
C xˆ = JC ˆl J T
needed
7. COMBINED (IMPLICIT) MODEL
♦ Linearised functional model (m equations)
∂ϕ
= 2vˆ T P + 2kˆ T B =0
∂vˆ
∂ϕ
= 2kˆ T A =0
∂δˆ
∂ϕ
= 2δˆ T A T + 2vˆ T B T + 2w T =0
∂kˆ
Dividing by 2 and transposing:
Pvˆ + B Tkˆ =0
Tˆ
A k =0 ②
Aδˆ + Bvˆ + w = 0
♦ The above equation can be written in hyper-matrix notation with the following
conditions:
1) The upper left matrix of the hyper-matrix must be invertible (the P matrix
is invertible)
2) The hyper-matrix should be symmetric (arrange the equations to achieve
this condition)
P BT 0 vˆ 0 0
B 0 A kˆ + w = 0
0 AT 0 δˆ 0 0
A B x u 0
♦ Partition into + = and eliminate x ( v̂ in this case) using
C D y v 0
(D - C A-1 B) y + (v - C A-1 u) = 0
substitute for A, B, C, D, x, y, u, and v
0 A B −1 T ˆ
k w B 0
T − P (B 0 ) + − P −1 (0 ) =
A 0 0 δˆ 0 104 243 0
0
0 A BP −1 B T 0 kˆ w 0
T − + =
A 0 0 0 δˆ 0 0
144444244444 3
⇓
− BP −1B T A kˆ w 0
+ =
AT 0 δˆ 0 0 ③
♦ Partition this hyper-matrix to eliminate K̂
[0 - (AT) (-B P-1 BT) -1 A] δ̂ + [0 - (AT) (-B P-1 BT)-1 (w)] = 0
∴ N-1 δ̂ + u = 0
-1
δ̂ = - N u recall P-1 = Cl
Note: if B = - I (i.e. Parametric Model)
N = AT [(-I) P-1 (-I)T)] -1 A = AT (P-1) -1 A = AT P A
That is
T T -1
v̂ = - Cl B (B Cl B ) (A δ̂ + w)
= - Cl BT k̂
♦ Therefore, the adjusted quantities are:
xˆ = x 0 + δˆ
ˆl = l + vˆ
∂δˆ ∂f(l, x 0 )
= −constant = −constant B
∂l ∂l
Therefore:
Cδˆ = { [A T
(BC l B T ) −1 A]−1 A T (BC l B T ) −1 B }C l
{B T
(BC l B T ) −1 A[A T (BC l B T ) −1 A]−1 }
Cδˆ = N −1 A T (BC l B T ) −1 BC l B T (BC l B T ) −1 AN −1
= N −1 A T (BC l B T ) −1 AN −1
= N −1 NN −1 = N −1
= [A T (BC l B T ) −1 A]−1
Cδˆ = [A T M −1A]−1
2. C xˆ
xˆ = x o + δˆ
C xˆ = Cδˆ
3. C vˆ without proof
where
M = BC l B T
4. C ˆl
ˆl = l + vˆ
Cˆl = Cl − C vˆ
7.1. Iterative Solution of the combined model
♦ Linearised model
Aδ + Bv + w = 0
Where
Y V
U
TX θ
TY
X
X a − b U
=
Y i b a V i
Therefore, the unknowns are a an b
In order to estimate a and b, observations are required. In this case, the
following table gives 3 points of known co-ordinates in both systems.
i U V X Y
1 0.0 1.0 -2.1 1.1
2 1.0 0.0 1.0 2.0
3 1.0 1.0 -0.9 2.8
For all three given points, the Cl matrix of the (U, V) co-ordinates is
0.01 0
C li = = 0.01 I 2,2
0 0.01
∴ C l = 0.01 I 6,6
2. Mathematical model
f1 → a Ui - b Vi - Xi = 0
f2 → b Ui + a Vi - Yi = 0 ⇒ a Vi + b Ui - Yi = 0
3. Linearised equations
U1 − V1 0 − 1
V U1 1 0
1
∂f U − V2 1 0
A= = 2 =
∂x x=x0 ,l =lobs V2 U 2 0 1
U3 − V3 1 − 1
V3 U 3 1 1
Note that A is not a function of x or l.
U1 V1 U 2 V2 U 3 V3
a 0
− b0 0 0 0 0
0 0
b a 0 0 0 0
∂f
=0 a −b
0 0
B= 0 0 0
∂l x = x 0 ,l =l obs
0 0
0 0 b a 0 0
0 0 0 0 a0 − b0
0 0 0 0 b0 a 0
Therefore, ao and bo are needed to evaluate B. They (ao and bo) can be
evaluated by simple computation through the use of 2 equations of the
math model.
a0 = 1, b0 = 2 using the equations of point 2
X1 − a 0 U1 + b 0 V1 − 0.1
Y1 − a V1 − b U1 0.1
0 0
X − a 0 U + b 0 V 0.0
w = f(x 0 , l) = 2 0 2 0 2 =
Y2 − a V2 − b U 2 0.0
X − a 0 U + b 0 V 0.1
3 0 3 0 3
Y3 − a V3 − b U 3 − 0.2
4. The δ̂ vector
δˆ = −N −1u = −[A T (BP −1B T ) −1 A]−1 A T (BP −1B T ) −1 w
0.0
δ̂ =
− 0.05
1.0
Check: make uses of x̂ == to calculate the values of the (X,Y) coordinates
1.95
making use of the (U, V) coordinates and the math model:
X a − b U
=
Y i b a V i
8. COMBINATIONS OF MODELS
♦ Assume that observations are made in two groups, with the second group consisting of one or
several observations. Both groups have a common set of parameters, i.e.
♦ Given: 2 sets of observations collected at different times for the same parameters
n1 n2
647 48 647 48
i.e. L1 and CL1 and L2 and CL 2
♦ Required: Xu,1 The best estimate for a group of parameters from a group of measurements
that have been taken at two different times ( e.g. L1 at t1 and L2 at t2 )
♦ Functional model:
F2 ( x , L2 ) = 0 A2 m uδ u ,1 + B2 m 2, n 2 V2 n 2,1 + W2 m 2 ,1 = 0
2
♦ Note: For each group of observations, there’s a quadratic form and a Lagrange Multiplier
♦ To minimize φ, differentiate φ w.r.t. all variables ( δ, V1, V2, K1, K2 ) and equate to zero.
∂φ
= 2 V1T P1 + 2 K1T B1 =0
∂V1
∂φ
= 2 V2T P2 + 2 K 2T B2 =0
∂V2
∂φ
= 2 δ T A1T + 2V1T B1T + 2W1T = 0
∂K1
∂φ
= 2δ T A2T + 2V2T B2T + 2W2T =0
∂K 2
P1 0 B1T 0 0 V1 0 0
0 P2 0 B2T 0 V2 0 0
B1 0 0 0 A1 K1 + W1 = 0
0 B2 0 0 A2 K 2 W2 0
0 0 A1T A2T 0 δ 0 0
♦ Required: X
♦ Functional Models:
( )
Lˆ1 = F1 Xˆ with the linear model
→ A1n1, u δ u ,1 + W1n1,1 = V1n1,1
♦ Variation function
( ) ( )
= δ T A1T + W1T P1 ( A1 δ + W1 ) + δ T A2T + W2T P2 ( A2 δ + W2 )
∂φ
= 2 δ T A1T P1 A1 + 2 W1T P1 A1 + 0
∂δ
2 δ T A2T P2 A2 + 2 W2T P2 A2 +0 =0
♦ Transpose and divide by 2
(14
A P A + A P A )δ + (A PW + A P W ) = 0
24
T
1 1
3 1424 1
3
T
2 2
1424 2
3 1424 3
T
1 1 1
T
2 2 2
u ,u u ,u u ,1 u ,1
1
42 4 3 142 4
3 1
42 4 3 142 4
3
N1 N2 u1 u2
∴ δ = −( N1 + N 2 ) (U 1 + U 2 )
−1
♦ Note: The δ vector involves addition of the normal equation matrices and vectors
corresponding to each set of observations.
♦ The same procedure can be applied for 3(or more) set of observations with the combined
144
( 42444
(
3 144424443
)
δ = −(N1 + N 2 )−1 A1T C L−11 f (x )0 − L1 + A2T C L−1 ( f (x 0 ) − L2 ) ) 2
u1 u2
0
x : non-stochastic and therefore, δ = F(L)
∴Cδ = J C L J T
T
∂δ ∂δ
∴ Cδ = CL
∂L ∂L
Where
L CL 0
L(n1 + n2 ) = 1 C L(n = 1
L2 1 + n2 )( n1 + n2 )
0 C L2
∂δ ∂δ ∂δ
=
∂L ∂L1 ∂L2
∂δ
= (N1 + N 2 )−1 A1T C L−1 = N −1 A1T C L−11
∂L1
∂δ
= (N1 + N 2 )−1 A2T C L− 2 = N −1 A2T C L−21
∂L2
0 C L−1 A1 N −1
∴Cδ = ( C L1
N −1 A1T C −1L1
1
N −1 A2T C L−21
1444442444443 0 C L2 C L−1 A2 N −1
)
↓ 144444 22444443
C L C −1 A1 N −1
1 L1
C C −1 A2 N −1
1 L2 L2
424 3
I
( )
= N −1 A1T C L−11 A1 + A2T C L−21 A2 N −1
= N −1 (N1 + N 2 )N −1 = N −1 NN −1
0
∴ Xˆ = X + δ
∴C Xˆ = Cδ
8.2. Sequential Least Squares-Parametric Model
• In the previous section (summation of normals), it has been shown that two sets of
observations ( L1 and L2 ) for the same set of parameters can be combined to get a new
solution
N1 = A1Tu ,n1 P1n1,n1 A1n1,u , and N 2 = A2Tun 2 P2n 2,n 2 A2n 2,u
1442443 1442443
u ,u u ,u
• What if n2 << u (e.g. n2 = 1 and u= 4). With the summation of normals method, a (u,u) matrix
must be inverted again to add the new (single) observation (the assumption here the solution
has been obtained already for the 1st group of observations)
• This can be a significant computational burden specially when observations are being added
at a regular interval in real time, example in GPS positioning.
• Sequential Least squares provides a method where only (n2,n2) inversion is required for
updating the solution with new n2 observations.
Derivation
To derive the sequential expressions, the summation of normal solution is re-written as:
(
δ = − N1 + A2T C L−21 A2 ) (U
−1
1 + A2T C L−21W2 )
(
δ = − N1 + A2T C L−21 A2 )
−1
(
U1 − N1 + A2T C L−21 A2 )−1
A2T C L−21W2
(S −1
+ T T R −1T )
−1
(
= S − ST T R + TST T )
−1
TS.............(i )
(S −1
+ T T R −1T )
−1 T
T R −1 = ST T R + TST T ( ) −1
...........(ii )
Apply lemma (i) to the 1st term and lemma (ii) to the 2nd term of the δ equation with:
N1−1 = S T = A2 R = C L2
1st term:
(
− N1 + A2T CL−21 A2 )−1 (
U1 = − N1−1 − N1−1 A2T CL + A2 N1−1 A2T A2 N1−1
)
−1
U1
144 42444 3 3
2
LHS of Lemma(i)
1 4444444 4244444444
RHS of Lemma(i)
2nd term
(
14444244443
)−1
− N1 + A2T C L−21 A2 A2T C L−21 (
W2 = − N1−1 A2T C L + A2 N1−1 A2T
144442244443
)−1
W2
LHS of Lemma(ii) RHS of Lemma(ii)
(
δ = − N1−1 − N1−1 A2T C L2 + A2 N1−1 A2T
)−1 A2 N1−1 U1 − N1 A2T (CL 2
+ A2 N1−1 A2T )−1W2
(
δ = − N1−1U1 + N1−1 A2T C L2 + A2 N1−1 A2T )−1
A2 N1−1U1 (
− N1−1 A2T C L2 + A2 N1−1 A2T ) −1
W2
δ= − N1−1U1
1424 3
+ N1−1 A2T (C L2 +
−1
)
A2 N1−1 A2T A2 N1−1U1
123
− W2
δ (− ) −δ ( − )
Now set
(
K = N1−1 A2T C L2 + A2 N1−1 A2T )−1
K is known as the Gain Matrix, which quantifies how much each new observations will
contribute to the corrections to the parameters
(
K = N1−1 A2T C L2 + A2 N1−1 A2T )
−1
• If the 2nd set of observations, L2, is imprecise, i.e. CL2 has large elements. The Gain Matrix
will generally have small elements. Thus the new observations will not greatly contribute to
solution update.
• As the precision of L2 increases, CL2 element decreases, L2 tends to contribute more the
solution update.
• δ (+ ) = δ (− ) − K ( A2δ (− ) + W2 )
14 4244 3
of the form Aδ +W =V
• That is
• V2 (+ ) = A2δ (+ ) + W2
Covariance Matrices
Note: the Cδ(+) < Cδ(−) due to the subtraction of KA2Cδ(−) and thus the covariance matrix ins
improved by adding observations.
Step 1:
Model: L1 = F1(X)
• Linearized Model:
A1 δ + W1 = V1
n1 ,u u ,1 n1 ,1 n1 ,1
0
• P.O.E. : X
• Initial solution with n1 observations (n1 ≥ u ) ,
where
P = σ 02C L−11
Cδ (− ) = N1−1
Step 2:
Model: L2 = F2 ( X )
A2 δ + W2 = V2
Linearized model:
n2 ,u u ,1 n2 ,1 n2 ,1
Xˆ = X 0 + δ (+ )
Cδ (+ ) = C Xˆ = Cδ (− ) − KA2 Cδ (− )
Step 3:
Addition of 3rd set of observations
L3 = F3 ( X )
δ(+) and Cδ(+) from step (2) become δ(−) and Cδ(−) respectively, for step (3).
The combination of models discussed so far have mainly derived for the parametric model
The derivation of the summation of normals and sequential LS for the implicit models follow the
same scheme as the parametric case.
PARAMETRIC IMPLICIT
δ = −(N1 + N 2 )−1 (U 1 + U 2 )
δ = −( N1′ + N 2′ ) −1 U 1′ + U 2′
N i = AiT Pi Ai
U i = AiT PiWi ′
( −1
N i = AiT B i Pi BiT ) −1
Ai
Cδ = C Xˆ = ( N i + N 2 )
−1
′
( −1
U i = AiT Bi Pi −1 BiT Wi
14243
)
Mi
N i as before
Pi = σ 02 C L−i1 & C Li = σ 02 Pi −1
−1
′ ′
Cδ = C Xˆ = N i + N 2
δ (+ ) = δ (− ) − K ( A2δ (− ) + W2 ) Change
(
P1 by Bi P1−1B1T ) −1
(
K = N1−1 A2T C L2 + A2 N1−1 A2T ) −1
AND
C L2 → B2C L2 B2T
8.5. Parameter Observations
• Parameter observation is a method which can be used in cases where a priori information
about the parameters is available.
• ( )
For example, station coordinates (or elevations) and their covariance matrix Xˆ and C xˆ and
may be available from a previous adjustment.
• In this case xˆ can be considered as a direct observations (with C xˆ ) along with some
observations vector to estimate better estimate of xˆ .
Functional Model:
Xˆ uobs ˆ
,1 = X u ,1
X obs = V X = X 0 + δ
0
V X = δ + X − X obs
V X u ,1 = δ u ,1 + W X u ,1 (V = Aδ + W )
• Note: a variable (observation or parameter) that has infinite variance, σ2→∞, has a
corresponding weight of P = 1 2 = 0 . In this case, the variable becomes an unknown
σ
parameters.
• In between these two extreme cases there are an infinite number of possibilities for
weighting parameters.
Functional Models (Linearized)
An , uδ u ,1 + Wn ,1 = Vn ,1
I u ,uδ u ,1 + W X n ,1 = Vu ,1
C L = P −1 (P = C )
−1
L
C X = PX−1 (P = C )
X
−1
X
Variation function
φ = V T PV + V XT PX V X = min
( ) (
= δ T AT + W T P ( Aδ + W ) + δ T + W XT PX (δ + W X ) ) = min
= δ T AT P A δ + δ T AT P W + W T P A δ + W T P W
+ δ T PX δ + δ T PX W X + W XT PX δ + W XT PX W X = min
= δ T AT PA δ + 2 δ T AT P W + W T P W
+ δ T PX δ + 2 δ T PX W X + W XT PX W X = min
Minimize φ
∂φ
= 2δ T AT PA + 2W T PA + 2δ T PX + 2WXT PX = 0
∂δ
(A T
) ( )
PA + PX δ + AT PW + PX W X = 0
(
δ u ,1 = − AuT,n Pn,n An,u + PX u ,u
) (A
−1 T
P Wn ,1 + PX u ,u WX u ,1
u ,n n ,n )
Variance-Covariance matrix of Estimated Correction Cδ:
Functional Model
−1
δ = −(AT PA + PX ) AT P F X − L + PX X − X obs
0
0
T T
∂δ ∂δ ∂αδ ∂δ
∴ Cδ = C L + obs C X obs
∂L ∂L ∂X ∂X
−1
T
Cδ = 1
A23 + PX
PA
N
• If the observed parameters are very precise, i.e. Cx has small elements, therefore C X−1 will
have large elements, so as (AT C L−1 A + C X−1 ) and therefore its inverse will be very small.
Finally δˆ will be smaller than if it calculated without the additional parameter observations.
• If the parameters observations are not precise, Cx will has large elements, C X−1 will have
small elements. Thus the parameter observation will have little contribution to solution
vector.
ENGO361, Dr. Naser El-Sheimy 1/15
9. Statistical Analysis
In Chapter 2, we defined the one–dimensional Gaussian Probability Distribution Function (PDF)
of normally distributed random variable as:
( x − µ )2
− x
2σ 2
1 x
f ( x) = e ,
σ 2π
x
where µ and σ are the mean and standard deviation of variable (x). And the Normal Distribution
Function (NDF) as:
( x − µ )2
x
−
t 2σ 2
1
F ( x) = ∫ e x dx
− ∞ σ x 2π
Where (t) is the upper bound of the integration as shown is the Figure below.
f(x)
t
• As stated before that the area under the curve represents the probability of occurrence.
Furthermore, the integration of this function yields the area under the curve.
• Unfortunately, the integration of the Equation (1) cannot be carried out in closed form, and
thus numerical integration techniques must be used to tabulate values of the this function.
This has been done for the function when the mean is Zero (µ = 0) and the variance is 1 (σ2 =
1).
ENGO361, Dr. Naser El-Sheimy 2/15
The result of this integration is shown in the following table. The tabulated values represent the
areas under the curve from –∞ to t.
t Decimals of (t)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
-3.2 .
-3.1 .
. .
. .
. .
-0.2 .
-0.1 .
0.0 .
0.1 .
0.2 .
. .
.
1.6 . . . . . . . . 0.95352
.
3.2
Example:
9 To determine the area under the curve from –∞ to 1.68 : Scan along the raw 1.6 and
under the column 8. At the intersection of raw 1.6 and the column 0.8 (1.68), the value is
0.95352 occurs.
Meaning:
• Since the area under the curve represents probability, and its maximum area is 1, this means
that there is 95.352% (0.95352 x 100) probability that (t) is less that or equal to 1.68.
• Alternatively, it can be stated that there is a 4.648% (1-0.95352) x 100 probability that (t) is
greater than 1.68.
9 Once available, this table can be used to evaluate the distribution function of any mean, µ,
and variance σ2.
9 For example, if y is a normal random variable with a mean, µy, and standard deviation σ y, an
equivalent normal random variable z = (y - µ y)/σ y can be defined that has µ = 0 and σ2 = 1.
Z2
t 1 −
N (Z ) = ∫ e 2 this is known as the Standard Normal Distribution
− ∞ 2π
Function.
-Z +Z
a b
• If –a = b = t
• P(-t < Z < t) = P(|Z| < t ) = N(t) – N(-t)
-Z +Z
-t t
• P(Z > t) = P(Z < -t)
• 1 – N(t) = N(-t)
From the above characteristic, the probability of the standard deviation can be estimated as
follows:
Looking into the table for t = 1 and t = -1, the area between -σ and σ is:
P(- σ < Z < σ ) = 0.84134 – 0.15866 = 0.68268
G(ε)
-3σ -2σ -σ σ 2σ 3σ
Ρ(− σ ≤ ε ≤ σ ) = 0.683
Ρ(− 2σ ≤ ε ≤ 2σ ) = 0.954 called confidence interval
Ρ(− 3σ ≤ ε ≤ σ ) = 0.997
Meaning:
For any group of measurements there is approximately a 68.3 % chance that any single
observation has an error ±σ. This is true for any set of measurements having normally distributed
errors.
In other words, any measurement has the same chance of becoming within these limits as it has
as falling outside them.
Therefore
From the SNDF tables it is apparent that 0.75 is between t value of 0.67 and 0.68, that is
N(0.67) = 0.7487 and N(0.68) = 0.7517
For any set of observations, therefore, the 50% probable error can be obtained by computing the
standard error and then multiplying it by 0.6745, i.e.
E50 = 0.6745 σ
• The 95% probable error, or E95, is the bound within which, theoretically, 95% of the
observation group’s errors should fall.
• Using the same reasoning as in the in the equations for the 50% probable error:
• From the SNDF tables it is determined that 0.975 occurs with a (t) value of 1.960.
• Thus, for any set of observations, therefore, the 95% probable error can be obtained by
computing the standard deviations (or error) and then multiplying it by 1.96, i.e.
E95 = 1.96 σ
E50 0.6745σ
E90 1.6449 σ
E95 1.96 σ
E99 2.576 σ
E99.7 2.965 σ
E99.9 3.29 σ
The E99.7 is often used in detecting blunders as will be explained in the following example.
Example:
The arc-second portion of 50 direction readings from 1″ instrument are listed below. Find the
mean, standard deviation (error), and the E95. Check the observations at a 99% level of
certainty for blunders.
41.9 46.3 44.6 46.1 42.5 45.9 45.0 42.0 47.5 43.2 43.0 45.7 47.6
49.5 45.5 43.3 42.6 44.3 46.1 45.6 52.0 45.5 43.4 42.2 44.3 44.1
42.6 47.2 47.4 44.7 44.2 46.3 49.5 46.0 44.3 42.8 47.1 44.7 45.6
45.5 43.4 45.5 43.1 46.1 43.6 41.8 44.7 46.2 43.2 46.8
Solution:
50
∑ Li
2252
The mean = =1 =
i = 45.04′′
n 50
50
∑ (Mean − V ) i
2
∴ E95 = ±1.96 σ = ±4.16″ (Thus 95% of the data should fall between 45.04 ± 4.16″ or in the
“40.88 - 49.20” range)
The data actually contain three values that deviate from the mean by more than 4.16 (i.e. that
are outside the range 40.88 to 49.20). They are 49.5 (2 times) and 52.0. No values are less
than 40.88, and therefore 47/50 = 94% or the measurements lie in the E95 range.
ENGO361, Dr. Naser El-Sheimy 7/15
∴ E99 = ±2.576 σ = ±5.46″ (Thus 99% of the data should fall between 45.04 ± 5.46″ or in the
“39.58 - 50.50” range).
Actually, one value is greater than 50.50, and thus 49/50 = 98% of all measurements fall in
this range.
ENGO361, Dr. Naser El-Sheimy 8/15
• The following contains a finite population of 100 values. The mean (µ) and the variance
(σ2) of that population are 26.1 and 17.5.
• By randomly selecting 10 values of this table, an estimate of the mean and the variance of
the this sample can be estimated.
• However, it would not be expected that these estimates ( X and S2) would exactly match
the mean and the variance of the population. Now if the sample size were increased, it
would be expected that the X and S2 would more nearly match µ and σ2 as shown in the
table below..
No. X S2
10 26.9 28.1
20 25.9 21.9
30 25.9 20.0
40 26.5 18.6
50 26.6 20.0
60 26.4 17.6
70 26.3 17.1
80 26.3 18.4
90 26.3 17.8
100 26.1 17.5
ENGO361, Dr. Naser El-Sheimy 9/15
• Since the mean and the variance of the sample ( X and S2) are computed from random
variable, they are also random variables. Thus it is concluded that the values computed
contain errors.
Set 1: 29.9, 18.2, 30.7, 24.4, 36.0, 25.6, 26.5, 29.9, 19.6, 27.9 X = 26.9, S2 = 28.1
Set 2: 26.9, 28.1, 29.2, 26.2, 30.0, 27.1, 26.5, 30.6, 28.5, 25.5 X = 27.9, S2 = 2.9
Set 3: 32.2, 22.2, 23.4, 27.9, 27.0, 28.9, 22.6, 27.7, 30.6, 26.9 X = 26.9, S2 = 10.9
Set 4: 24.2, 36.0, 18.2, 24.3, 24.0, 28.9, 28.8, 30.2, 28.1, 29.0 X = 27.2, S2 = 23.0
• Fluctuations in the mean and variance computed from sample sets raises questions about
the reliability of these estimates.
• A higher confidence value is likely to be placed on a sample set with small variance than
on one with large variance.
• In the above table, because of its small variance, one is more likely to believe that the
mean of the 2nd set is the most reliable estimate of the population mean. In reality this is
not the case, as the means of the other three sets are actually closer to the population
mean (µ = 26.1).
• The estimation of the mean and variance of a variable from sample data is referred as
point estimation (because it results in one value for each parameter in question)
• After having performed a point estimation, the question remains as how much the
deviation of the estimate is likely to be from the still unknown true values of the
parameters (mean and variance). In other word, we would like to have an indication of
how good the estimation is and how much it can be relied on.
• An absolute answer to this question is not possible because sampling never lead to
the true parameters.
• It is only possible to estimate probabilities with which the true value of the parameters in
question is likely to be within a certain interval around the estimate. Such probabilities
can be determined if the distribution function F(X) of the random variable is given.
Recall: The probability that a random variable Z takes values within the boundary “a and b”
is given by:
X =b
P(a < Z < b) = F(b) – F(a) = ∫ F ( X ) dX
X =a
ENGO361, Dr. Naser El-Sheimy 10/15
• The probability that the parameter does not fall in a given interval is α.
• The probability statement for the confidence interval, which is symmetric here, is
then:
X −µ
P − Z < <Z = 1−α
α /2 σ α / 2
n
or
P X − Z ⋅σ < µ < X +Z ⋅σ
= 1−α
α /2 n α /2 n
• For example for α = 5%, Zα/2 = 1.96, therefore we can write the above equation as:
• The above is an example of the so called two-sided confidence interval. On the other
hand, for a one-sided confidence interval we write:
Two-sided One-sided
α/2 α/2 α
1-α 1-α
-Zα/2 Zα/2 Zα
Example:
Suppose a distance is measured n = 8 times with the mean X = 10.1 cm. We assume that the
variance of the normal population is known to be σ2 = 0.1 cm2. Then the 95% confidence interval
on µ (which is unknown) for the two-sided confidence interval is:
0.1 0.1
P 10.1 − 1.96 < µ < 10.1 + 1.96 = 0.95
8 8
0.1
P µ < 10.1 + 1.6449 = P{µ < 10.28}= 0.95
8
• Let us know consider the case in which the standard deviation of the distribution σ is not
known and has to be repalced by the standard deviation of the, S, of the sample. Therefore,
the estimator under testing is X − µ which has different distribution than X − µ (definitely
S σ
n n
X − µ has different distribution than X − µ as well be shown in the next section)
S σ
n n
ENGO361, Dr. Naser El-Sheimy 12/15
The Gaussian or ND is the most frequently used distribution in statistical theory. Its density.
Function is given by:
( x − M x )2
−
1 2σ x2
()
f x = e which is fully described by to parameters Mx and σx.
σ x 2π
X − MX
The cumulative NDF of the standardized random variable Z = (having a zero mean
σX
and unit standard deviation, i.e. [M z = 0 & σ z = 1] ) is given by:
2
-Z
Z
1
F (Z ) = ∫e 2 dz , the table handed during the class gives the values of F(Z)
2π −∞
• The distribution of the sum of squares of independent random variables each of which is
normally distributed is known as X2 (chi square) distribution.
Examples:
• χ n2 = X 12 + X 22 + .............. + X n2 which is χ n2 distribution of “n degrees of
freedom”
• VnT,1Pn , nVn ,1 is a χ n2 distribution of “n degrees of freedom”
X α2 ,m
ENGO361, Dr. Naser El-Sheimy 13/15
∫ f (χ )dχ
2 2
=1−
0
• The distribution of the ratio of two independent random variable, each having a X2 (Chi-
square) distribution, is said to have an F-distribution.
χ m2 / m
Fm, n = 2 is F-distribution of (m) and (n) degrees of freedom.
χn / n
• The practical application of the F distribution in least squares adjustment and statistical
testing is concerned with the comparisons of variances such as those obtained from two
adjustment. In some practical case the comparison may be between a variance obtained
from the adjustment (such as σˆ 02 ) and an a priori given reference variance (such as σ 02 ),
2
this case refers to the Fr , ∞ = χ r (where r is the degrees of freedom).
∑ v 2 is χ n2−1
2
Recall: ∑v , where distribution
variance = σ 2 =
n{−1
degrees of Freedoms
ENGO361, Dr. Naser El-Sheimy 14/15
Fα , m1 , m2
Fα , m1 , m2
500
.
.
.
∞
ENGO361, Dr. Naser El-Sheimy 15/15
The t-distribution is used in connection with sampling (i.e. Testing using sample statistics
instead of population parameters)
′ ′ ′
Let X 1 , X 2 ,....... X n be n independent random (stochastic) variables of identical normal
distribution with mean M and standard deviation σ Then the random variable.
x−M
t= n is said to have a t-distribution with (n-1) degrees of freedom. Where
s
1 n 1 n
x= ∑ xi and s 2
= ∑ (xi − x )2 , where x and s 2 are the sample mean and
n i =1 n − 1 i =1
standard deviation respectively.
tα,b
Question 1: (9%)
The opposite figure shows a triangulation station (O) from which the horizontal directions d1, d2,
and d3 were measured to the three stations A, B, and C respectively
with: A
d1 = 45o 15’ 25” d1
2 0 0
d2 = 75o 25’ 35” C L = 0 2 0 arcsec2 α1 d
2 B
0 0 2 O
α2
o
d3 = 115 35’ 45” d3
C
1. Calculate the estimates of the two horizontal angels α1 and α2
and their Variance-Covariance matrix.
2. Discuss the degree of correlation between the two angles
Question 2: (3%)
If you have an Electronic Distance Measuring (EDM) device that has been calibrated to give a
standard deviation for a single measurement = 15 mm, how many times should you measure a
baseline of a length ≅ 5 km, and whose relative error is specified not to exceed 2 PPM?
Question 3: (3%)
h2 B
The opposite figure shows a leveling line which runs
from a benchmark (A) of known elevation (Ha) to
point B. The observed height differences h1 and h2 are h1
to be observed with the same precision and
uncorrelated.
Calculate how many times should h1 and h2 be A
measured to have the standard deviation of the
elevation of point B not to exceed 3 mm? Assume that the standard deviation of a single height
difference observation = 5mm.
A distance is measured five times with the following values obtained: 156.11, 156.14, 156.08,
156.05, and 156.15 m. The first two measurements have a standard deviation of 1 cm and the last
three have standard deviation of 2 cm. All measurements are uncorrelated.
1. What is the weighted least-squares estimate of the distance?
2. What is the standard deviation of the estimated distance?
1. Write down in matrix form the condition equations for the following triangulation networks:
(a) (b)
α1
α1 α3
α4 α3
α4 α2
α2
Good Luck
NE\ne