Mathematical Analysis of ILI MFL Data: January 2003

Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/273257472

Mathematical Analysis of ILI MFL Data

Technical Report · January 2003


DOI: 10.13140/2.1.1050.9606

CITATIONS READS
0 124

1 author:

Peter Robert Massopust


Technische Universität München, Munich, Gemany
170 PUBLICATIONS   1,895 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Approximation and Interpolation Theory with Splines and Fractals View project

Fractal Geometry View project

All content following this page was uploaded by Peter Robert Massopust on 08 March 2015.

The user has requested enhancement of the downloaded file.


Mathematical Analysis of ILI MFL Data
Peter R. Massopust
Senior Research Scientist

Tuboscope Pipeline Engineering


Houston, TX 77051
USA
[email protected]
December 2, 2002

Abstract
Formulae for the probability of detection of interior and exterior defects in pipelines are derived taking
into account generic parameters relating to the defect geometry, the tool and the pipe. In addition, a
sampling theorem of ILI MFL data is derived and the most frequent types of errors associated with such
sampling are investigated. Bounds on these errors are also presented.
Keywords and phrases: probability, detection, MFL, bandlimited function, Shannon Sampling Theorem,
error analysis

1 Introduction
In order to assess the performance of an ILI tool and the underlying defect characterization and grading
algorithms, the data, i.e., the samples of the MFL signature response, that was collected by the tool needs
to be carefully examined and analyzed. The data samples are effected by numerous controllable and un-
controllable factors, which include the geometry defect, the tool parameters, and the pipe material. In the
following we will assume that the data has been properly denoised and compensated for tool velocity.
One immediate question that arises in the above context, involves the probability of detecting a defect,
either on the interior or exterior of a pipe. Clearly, this probability depends on several parameters, some of
which are not controllable, such as defect geometry and pipe material properties. In Section 2, we present
a formula for the probability of detection of a defect taking into account controllable and uncontrollable
parameters. As for any model, we need to make some assumption before we derive these formulae. An error
analysis for the probability of detection is also presented.
When working with sampled values of a signal, one naturally has to ask to what degree the samples
represent the signal and whether the entire signal can be reconstructed from its sampled values. An answers
to these questions is given by the celebrated Whittaker-Shannon-Kotel’nikov (WSK) sampling theorem,
which states that if a signal does not contain frequencies above a certain value, it can be reconstructed in its
entirety from an infinite number of samples. (Since most signals do have compact support, i.e., are identically
equal to zero outside an interval of finite length, or fall off to infinity at a rate that allows to effectively treat
them as compactly supported, the requirement of having an infinite number of samples is not to restrictive.)
A more restrictive assumption is that the signal does not contain frequencies above a certain value, that
it is band-limited. (In mathematical terms this means that the Fourier transform is compactly supported.)

1
2 Peter R. Massopust

Fortunately, there exist sampling theorems for signal not being necessarily band-limited. Since ILI MFL
signals are not band-limited, we need to present these sampling theorems and investigate their conclusions
in order to derive a sampling theorem for ILI data. This is done in Section 3 of this paper.
The next section deals with errors associated with sampling theorems. The three most prolific errors are
truncation error, amplitude error, and position-jitter error. The first deals with the case that only a finite
number of samples is used to reconstruct the original signal. The amplitude error involves inaccuracies in
the measurement of the sampled values due to quantization, noise (white and non-white), and round-off. It
is assumed that these inaccuracies are bounded above by no more than ε. The last error involves inexact
position measurements and are also assumed to be bounded by no more than a given δ. Finally, we present
an estimate for the combined error.

2 Probability of Defect Detection


In this section we present a formula for the probability of detecting a defect in a pipe. The variables
entering into this formula depend on several parameters some of which are beyond control. The assumptions
under which the probability of detection formula is valid are clearly identified. An error formula involving
the parameters on which the probability of detection depends is also presented.

2.1 Derivation of the Probability of Detection


The following assumptions on the pipe and the defects are made. The term MFL signature refers to the
signal output of the sensors. It is an indirect response to the presence of the magnetic induction B around
a defect.
1. The pipe is describable as a finite number of right circular cylindrical shells with fixed inner and outer
radii ri and re , respectively. The difference T := re − ri is called the wall thickness of the pipe. In
general, T depends on the axial position along the pipe and is constant between pipe joints.
2. Following pipeline convention, we employ a cylindrical coordinate system for which the y axis coincides
with the pipe axis. The polar angle θ is then measured from the positive x-axis towards the positive
z-axis and the coordinate lines r = const. are rays in the xz-plane.
3. Defects are of measurable size and have well-defined boundary. Let Θ denote the smallest angle such
that the defect D lies entirely between the half-planes θ = θ0 and θ = θ0 + Θ. The geometric width of
D is defined as w := r0 Θ, where r0 denotes either the inner or outer radius depending on whether the
defect is interior or exterior to the pipe. Similarly, let ` be the smallest value so that D lies entirely
between the planes y = y0 and y = y0 + `. Then ` is called the length of D. The depth of an interior,
respectively, exterior defect D is the smallest values d for which D lies entirely between the cylinders
ri and ri + d, respectively, re − d and re .
4. The defect is sufficiently extensive in size to produce a recordable MFL signature. This implies that
the amplitude of the MFL signature must be above the noise level in order to be detected.
5. There are g groups of n sensors situated around the tool. Within each group, the sensor are a distance
d apart. The groups themselves have a circumferential spacing of D > d ≥ 0.
To compute the probability of detecting a defect D satisfying the above assumptions, we assume that
D generates a flux leakage field B whose circumferential extend is wB . Note that because of the so-called
booming effect, wB > w, the geometric width of D. Since B decays quadratically to zero and the tool
sensors have limited finite detection range, only a field of circumferential width w ≤ wB will be detected.
In this paper, we refer to w as the magnetic width.
We can describe the location of a defect D of geometric width w by an arc A of length w on a circle of
radius r0 , where r0 denotes either the inner or outer radius depending on whether D is interior or exterior to

2
Mathematical Data Analysis 3

the pipe. The position of A on the circle of radius r0 is completely determined by the position of the center
of A. As the defect can be situated anywhere along the pipe wall, we describe the position of the center of
A by a random variable X with uniform distribution f (θ) = (2π)−1 .
The defect D will be detected if at least one sensor passes under it. We distinguish three cases.
Case I: 0 < w < d. In this situation, the defect D will be detected by at least one sensor in a particular
group of sensors if the center of A lies within a distance of ± w/(2r0 ) of the point (r0 , θνk ), where θνk
is the angular position of the kth sensor in the νth group, k = 1, . . . , n and ν = 1, . . . , g. Denoting the
probability of detection by the kth sensor in the νth group by
  
w w
P X ∈ θνk − , θνk + ,
2r0 2r0
we obtain for the probability PI that the defect will be detected by at least one of the n g sensors
g X
n   
X w w
PI = P X ∈ θνk − , θνk +
ν=1 k=1
2r0 2r0
g X n Z w
θνk + 2r
X 0 dθ
=
ν=1 k=1
w
θνk − 2r 2π
0
ngw
= .
2πr0
If we denote by R the radial distance1 of the sensor groups from the center of the tool, then

g (n − 1) d + g D = 2πR, (2.1)

and we may express the above probability as


  
w R g D
PI = + 1− (2.2)
r0 d 2π d

Case II: w ≥ d. Since the magnetic width of D is larger than d, the defect will be detected once the
center of A lies within a distance of ± w/(2r0 ) of the arcs subtended by the groups of sensors. Each
such arc has a length of (n − 1) d and thus subtends an angle of size (n − 1) d/r0 . Hence, the probability
PII of detecting D in this case equals
g  
X w w
PII = X ∈ θν,` −
P , θν,u +
ν=1
2r0 2r0
g  
X w 1
= + [θν,u − θν,` ]
ν=1
r0 2π
gw g (n − 1) d
= + .
2πr0 2πr0
Here we denoted by θν,` and θν,u the angle corresponding to the beginning and end of the arcs of length
(n − 1) d, respectively. Employing Eqn. (2.1) again, we find for P

R g  
PII = + w−D (2.3)
r0 2π r0
1 We assume that R is constant for all sensor groups.

3
4 Peter R. Massopust

Note that PI (w = d) = PII (w = d) and that the probability of detection PII equals one only if

2π  
w≥D+ r0 − R (2.4)
g

Denote by λ := ri − R the lift-off of the sensors. (Here it is assumed that all sensors have the same lift-off.)
For an interior defect, r0 = ri = R + λ and for an exterior defect, r0 = re = R + λ + T . If we introduce the
blooming constant α by α := w/w then the above formulae become

Case I: 0 < w < d.

Interior Defect Exterior Defect


     
αw R g D αw R g D
PI = + 1− PI = + 1−
R+λ d 2π d R+λ+T d 2π d

Case II: w ≥ d.

Interior Defect Exterior Defect

R g   R g  
PII = + αw − D PII = + αw − D
R + λ 2π (R + λ) R+λ+T 2π (R + λ + T )

Assigning a probability of defect detection and knowing the tool-specific quantities R, d, D, and λ as
well as the blooming constant α, determined either through experimental means or analytic computations,
allows the calculation of the minimal geometric width w for this probability.

2.2 Error Analysis for the Probability of Detection


The probability of detection, formulas 2.1 and 2.3, of a defect D depends critically on the defect parameters
α and w, the tool parameters2 R, λ, d, and D, and the pipe parameter T .
To indicate the dependence of the probability of detection P on the three sets of parameters, we write

P = P (α, w; R, λ, d, D; T ). (2.5)

Employing the rules of error propagation we see that the uncertainty or error ∆P in P is given by

∂P ∂P ∂P ∂P ∂P ∂P ∂P
∆P =
∆α +
∆w +
∆R +
∆λ +
∆d +
∆D +
∆T,
∂α ∂w ∂R ∂λ ∂d ∂D ∂T

where ∆ denotes the error in the parameter to which it refers. The computation of the partial derivatives
yields the following results.

2 The parameter R clearly relates to the pipe diameter, but will nevertheless be counted as tool specific.

4
Mathematical Data Analysis 5

For Case I:

∂PIi [g (d − D) + 2 π R] w ∂PIe [g (d − D) + 2 π R] w
= ; = ;
∂α 2 d π (R + λ) ∂α 2 d π (R + T + λ)

∂PIi [g (d − D) + 2 π R] α ∂PIe [g (d − D) + 2 π R] α
= ; = ;
∂w 2 d π (R + λ) ∂w 2 d π (R + λ + T )

∂PIi w α [g (d − D) − 2 π λ] ∂PIe w α [g (d − D) − 2 π (T − λ)]


=− 2 ; =− 2 ;
∂R 2 d π (R + λ) ∂R 2 d π (R + T + λ)

∂PIi [g (d − D) + 2 π R] w α ∂PIe [g (d − D) + 2 π R] w α
=− 2 ; =− 2 ;
∂λ 2 d π (R + λ) ∂λ 2 d π (R + λ + T )

∂PIi (D g − 2 π R) w α ∂PIe (D g − 2 π R) w α
= ; = ;
∂d 2 d2 π (R + λ) ∂d 2 d2 π (R + λ + T )

∂PIi gwα ∂PIe gwα


=− ; =− ;
∂D 2 d π (R + λ) ∂D 2 d π (R + λ + T )

∂PIi ∂PIe [g (d − D) + 2 π R] w α
= 0; =− 2
∂T ∂T 2 d π (R + T + λ)

and for Case II:

i e
∂PII gw ∂PII gw
= ; = ;
∂α 2 π (R + λ) ∂α 2 π (R + λ + T )

i e
∂PII gα ∂PII gα
= ; = ;
∂w 2 π (R + λ) ∂w 2 π (R + λ + T )

i e
∂PII g (D − w α) + 2 π λ ∂PII g (D − w α) + 2 π (T + λ)
= 2 ; = 2 ;
∂R 2 π (R + λ) ∂R 2 π (R + λ + T )

i e
∂PII g (D − w α) − 2 π R ∂PII g (D − w α) − 2 π R
= 2 ; = 2 ;
∂λ 2 π (R + λ) ∂λ 2 π (R + λ + T )

i e
∂PII ∂PII
= 0; = 0;
∂d ∂d

5
6 Peter R. Massopust

i e
∂PII g ∂PII g
=− ; =− ;
∂D 2 π (R + λ) ∂D 2 π (R + λ + T )

i e
∂PII ∂PII g (D − w α) − 2 π R
= 0; = 2
∂T ∂T 2 π (R + T + λ)

The superscripts i and e refer to inner and outer defects, respectively.


The absolute error for an inner and outer defect in both cases is then given by
Case I:

[g (d − D) + 2 π R] w [g (d − D) + 2 π R] α w α [g (d − D) − 2 π λ]
∆PIi = ∆α + ∆w + ∆R

2 d π (R + λ) 2 d π (R + λ) 2
2 d π (R + λ)


[g (d − D) + 2 π R] w α (D g − 2 π R) w α gwα
+ ∆λ + ∆d + ∆D;

2 2
2 d π (R + λ) 2 d π (R + λ)
2 d π (R + λ)


[g (d − D) + 2 π R] w [g (d − D) + 2 π R] α w α [g (d − D) − 2 π (T − λ)]
∆PIe = ∆α +
2 d π (R + λ + T ) ∆w + ∆R

2 d π (R + T + λ) 2
2 d π (R + T + λ)


[g (d − D) + 2 π R] w α (D g − 2 π R) w α gwα
+ ∆λ + 2 ∆d + ∆D

2 d π (R + λ + T )2 2 d π (R + λ + T ) 2 d π (R + λ + T )


[g (d − D) + 2 π R] w α
+ ∆T.

2 d π (R + T + λ)2
(2.6)

Case II:

g w g α g (D − w α) + 2 π λ
i

∆PII = ∆α +
2 π (R + λ) ∆w + ∆R

2 π (R + λ) 2
2 π (R + λ)


g (D − w α) − 2 π R g
+ ∆λ + ∆D;

2 2 π (R + λ)
2 π (R + λ)

(2.7)

g w g α g (D − w α) + 2 π (T + λ)
e

∆PII = ∆α +
2 π (R + λ + T ) ∆w + ∆R

2 π (R + λ + T ) 2
2 π (R + λ + T )


g (D − w α) − 2 π R g g (D − w α) − 2 π R
+ ∆λ + ∆D + ∆T.

2 π (R + λ + T )2 2 π (R + λ + T ) 2 π (R + T + λ)2

Therefore, we obtain for the relative error in the two cases:

6
Mathematical Data Analysis 7

Case I:

∆PIi

∆α ∆w g(D − d) + 2π λ ∆R λ ∆λ
i
= + + + +
PI α w [g (d − D) + 2π R](R + λ) R
R+λ λ


D g − 2π R ∆d g ∆D
d [(g(d − D) + 2π R)] d + g (d − D) + 2π R D ;

(2.8)
∆PIe

∆α ∆w g(D − d) + 2π (T + λ) ∆R λ ∆λ
e = + + + +
PI α w [g (d − D) + 2π R](R + λ + T ) R
R+λ+T λ


D g − 2π R ∆d g ∆D T ∆T
d [(g(d − D) + 2π R)] d + g (d − D) + 2π R D + R + T + λ T .

For the second case we obtain the following.


Case II:
i

∆PII gwα ∆α gwα ∆w (D g − g w α + 2 π λ) R ∆R
i
= +
+

PII −g D + 2 π R + g w α α
−g D + 2 π R + g w α w
(D g − 2 π R − g w α) (R + λ) R


λ ∆λ gD ∆D
+ + ;
R+λ λ D g − 2 π R − g w α D

e

∆PII gwα ∆α gwα ∆w (D g − g w α + 2 π (T + λ)) R ∆R
e =
+
+

PII −g D + 2 π R + g w α α −g D + 2 π R + g w α w (D g − 2 π R − g w α) (R + λ + T ) R


λ ∆λ gD ∆D T ∆T
+ + + .
R+λ+T λ Dg − 2πR − gwα D
R+T +λ T
(2.9)

3 A Sampling Theorem for ILI Data


In this section a sampling theorem for ILI data is presented that provides the basis for the next section
in which the errors associated with finite numerical approximations are discussed. Particular emphasis is
placed on the type of function, i.e., its regularity and decay at infinity, an MFL signal naturally represents.
Before analyzing the issues surrounding sampling theorems for ILI data, we need some definitions and
related results.

Definition 3.1. A function f ∈ L2 (R) is called band-limited if its Fourier transform

f 7→ fb (3.1)
Z +∞
1
fb(ω) = √ f (x) e−ixω dx (3.2)
2π −∞

has compact support, i.e.,


|fb(ω)| = 0, for |ω| > π W . (3.3)

7
8 Peter R. Massopust

In engineering terminology, a function f is band-limited if its (frequency) spectrum contains only fre-
quencies between −π W and +π W .
Besides functions in L2 (R), there is need to consider functions in more general Lebesgue spaces Lp (R),
for 1 ≤ p ≤ ∞. More precisely, Lp (R) is the space of all functions f : R → R such that the norms
Z +∞ 1/p
kf kp := |f (x)|p dx and kf k∞ := ess sup |f (x)| (3.4)
−∞ x∈R

are finite3 . We denote the class of band-limited functions f ∈ Lp (R) with bandwidth W by Bπp W .
For band-limited functions, the celebrated Whittaker-Shannon-Kotel’nikov (WSK) sampling theorem [8]
holds.4

Theorem 3.1 (WSK Sampling Theorem). If a signal f (x) contains no frequencies higher than π W > 0,
then it is completely determined by its values f (k/W ) on a discrete set of points with spacing 1/W and can
be reconstructed via the formula

+∞  
X k sin(W x − k)
f (x) = f (3.5)
W W x−k
k=−∞

Moreover, the above series is absolutely and uniformly convergent.

Remark 3.1. In the above theorem we assume that the signal is centered at x = 0 and sampled in both
the negative and positive direction. This can easily be achieved by a simple translation of the signal and a
redefinition of the summation indices.

Note that the theorem is only valid for band-limited functions. There are many instances, however, where
this assumption fails to be true. In order to establish an sampling theorem for not necessarily band-limited
functions f , one assumes that the series in (3.5) will not equal f but converges to f as W → ∞:
+∞  
X k
f (x) = lim f sinc (W x − k), (3.6)
W →∞ W
k=−∞

where we set
sin x
sinc x :=
.
x
In other words, the bandwidth W needs to be chosen large enough so that the aliasing error

+∞  
X k
(RW f )(x) := f (x) − f sinc (W x − k) (3.7)
W
k=−∞

becomes sufficiently small to be computationally negligible. The next result gives an estimates on the error
RW in terms of the regularity of the underlying signal f . For this purpose, we need the following definition.

Definition 3.2. Let 0 < α ≤ 1 and suppose f is a function with bounded rth order derivative f (r) . Then
f is said to belong the the Lipschitz class LiprL (α) if there exists a positive constant L, called a Lipschitz
constant, such that
|f (r) (x + h) − f (r) (x)| ≤ L hα , for all h > 0.. (3.8)
For r = 0 we write LipL (α) instead of Lip0L (α).
3A function f for which kf k∞ is finite is bounded everywhere except on a set of measure zero.
4 For a historical account of this sampling theorem, we refer the interested reader to [3, 6, 9].

8
Mathematical Data Analysis 9

It is worthwhile noting that if f ∈ LiprL (α) with α > 1, then f is a polynomial of degree less than r.
We are now able to state an estimate for the error RW . In the following we denote by C(R) the space of
all continuous functions on the real line R endowed the the norm kf k := supx∈R |f (x)|.
Theorem 3.2. Suppose that f ∈ C(R) and that

|f (x)| ≤ Mf |t|−γ , for |x| ≥ 1 and some 0 < γ ≤ 1, (3.9)

for a positive constant Mf depending only on the signal f . If f (r) ∈ Lip1L (α), 0 < α ≤ 1 and r ∈ Z+
0 , then

log W
sup |(RW f )(x)| ≤ K(f, r, α, γ) , (3.10)
x∈R W r+α

provided W ≥ exp[2/(r + α + γ)] and with


r+α+γ h γ  i
K(f, r, α, γ) := 7 L (2/π)r+α + 5γ/2 e + 2γ e (4 Mf + 3 kf k) . (3.11)
γ 2

Proof. See [1] or [3].


In connection with the above theorem the question naturally arises as to whether an MFL signal f is
band-limited, and – if not – to which Lipschitz class f belongs so that the aforementioned error estimate may
be employed. The signal f is the response measured due to a change in the induced magnetic field B and
this response is directly proportional to B x , the axial component of B. Hence the signal f belongs to the
same function space as B x . Since B satisfies a second order elliptic differential equation with appropriate
boundary conditions on a domain Ω, the usual theorems for the existence and uniqueness of solution to such
differential equations apply (Cf., for instance, [5].), resulting in the fact that

B = (B x , B y , B z ) ∈ (H 2 (Ω))3 .

Here, the Sobolev space H 2 (Ω) denotes the space of all functions g : Ω → R whose second derivative is in
L2 (Ω). By the celebrated Sobolev Embedding Theorem, H 2 (Ω) embeds continuously into Cb1 (Ω), the space
of all bounded continuously differentiable functions on Ω. Functions g that are in Cb1 (Ω) are clearly also in
the Lipschitz class Lip1L (1) with  
∂g
L = sup : x ∈ Ω =: kg 0 k.
∂x
Since an MFL signal f has only finitely many zeros, it cannot be band-limited. (This result follows from
complex analysis as band-limited functions are naturally associated with entire functions of exponential type.
For more details, we refer to [9].)
Now we can apply Theorem 3.2 to an arbitrary MFL signal f with r = α = 1. What remains to be
determined is the exponent γ in (3.9) and a procedure to compute the signal-specific constant Mf . We
employ the length ` of a defect D as the reference for the determination of Mf and γ. As the MFL signal
crosses the baseline, or in our present terminology, has its zeros at x = ±`, we can use the value of f at
the side lobes, which are located at positions ±` ± δ`,δ`  `, to estimate Mf . (In case the lobes are not
symmetrically positioned with respect to the center x = 0, we choose the larger of the two values.) It follows
from experimental and theoretical considerations that we can increase |f (±` ± δ`)| – if necessary – so that
|f (x)| ≤ max{|f˜(` + δ`)|, |f˜(−` − δ`)|} |x|1 |` + δ`|1 , where the ˜ indicates a possibly larger f -value. Hence,
we choose
Mf := max{|f˜(` + δ`)|, |f˜(−` − δ`)|} and γ := 1. (3.12)
Therefore, we obtain for the constant Kf := K(f, 1, 1, 1) the value

14 kf 0 k  √ 
 
Kf = 3 1 + 2 e 5 + 2 e (4Mf + kf k) (` + δ`) ≈ (55.9875) kf 0 k + (23.1352) (4Mf + kf k)(` + δ`).
π2

9
10 Peter R. Massopust

The above estimate includes the term ` + δ`, which must be included since the growth rate (3.9) applies
only to |x| ≥ (` + δ`). Using (3.10), we can choose W large enough so that the aliasing error is less than any
prescribed value ϑ > 0. More precisely, let W ∗ be the unique solution5 of

log W ϑ
= , for W ≥ e[2/(r+α+γ)] ' 1.95,

W2 Kf r=α=γ=1

then RW ∗ (f ) < ϑ, for all x ∈ R.


Therefore, we arrive at the following ϑ-approximate sampling theorem for ILI data.

Theorem 3.3 (ϑ-approximate Sampling Theorem for ILI Data). Suppose that a bandwidth W has been
chosen in such a way that the aliasing error RW (f ) of an MFL signal f is computationally negligible. Then
the signal f is given by
+∞  
X k
f (x) =ϑ f sinc (W x − k) . (3.13)
W
k=−∞

Here the symbol =ϑ indicates that the equality is only correct to within ϑ accuracy.

Theorem 3.3 can now be employed to derive a ϑ-approximate sampling theorem for the derivative f 0 of
an MFL signal.

Theorem 3.4 (ϑ-approximate Sampling Theorem for the Derivative of ILI Data). For the derivative f 0 of
an MFL signal f we have
+∞
(−1)k+1 W
 
0
X k
f (x) =ϑ f x+ . (3.14)
W k
k=−∞

Proof. For the steps leading to Eqn. (3.14), we refer to [3].

Remark 3.2. The aliasing error for the derivative f 0 obeys a different estimate than (3.9). The symbol
=ϑ in Eqn. (3.14) is only reminiscent of the fact that f is band-limited to within ϑ-accuracy and is not be
interpreted as an accuracy statement for f 0 .

To use the above theorems for ILI MFL data, we need to relate the sample rate to the bandwidth W .
Since at Tuboscope, the sample rate for axial data is approximately 0.1 in or 2.54 mm, one obtains for W a
frequency of about 400 Hz. (The precise value is 393.701 Hz.) This value is above the lower bounds required
for the above and also the following estimates.

4 Error Estimates for Sampling


In the previous section, we derived a sampling theorem for an MFL signal and its derivative assuming
that we have an infinite number of exactly measured samples at your disposition. However, for all practical
purposes, the number of available samples is only finite and may not be measured exactly in two respects:
firstly, the amplitudes f (k/W ) may not be exact (amplitude error) and the positions x/W may also be not
exact (position-jitter). In this present section, we derive bounds for these three individual errors as well as
the combined error.
5 The function log W/W 2 is strictly monotone decreasing for W ' 1.65, which guarantees the uniqueness of the solution.

10
Mathematical Data Analysis 11

4.1 Truncation Error


For most practical applications, only a finite number of samples are given and this finite set needs to be
employed to reconstruct the signal f . The ensuing error from using only finitely many samples is termed
truncation error is formally defined by
X k 
(TN f )(x) := f sinc (W x − k), (4.1)
W
|k|>N
p
where f ∈ BπW and only samples from k = −N to k = +N are given. The following estimate on the
truncation error holds.
Theorem 4.1 (Estimate on Truncation Error). Suppose that f ∈ BπW 1
with fˆ(r) ∈ LipL (α), 0 < α ≤ 1.
Then
|TN f (x)| = O(N −r−α ), (4.2)
for N satisfying N + 1 ≥ W |t|/ξ, 0 < ξ < 1, and N ≥ r > 0.
Here a = O(nβ ) means that limn→∞ a/nβ = const.
Proof. We refer to [3] and [9].

4.2 Amplitude Error


Let us now assume that the samples f (k/W ) are not exactly measured but are given as quantized values
f (k/W ) with local errors εk := f (k/W ) − f (k/W ). In addition, we assume that these local errors are
uniformly bounded in the sense that |εk | ≤ ε, for all k ∈ Z. (We may think of ε are half the difference
between two consecutive quantization steps.) The total amplitude or quantization error Aε f is given by
+∞      +∞
X k k X
(Aε f )(x) := f −f sinc (W x − k) = εk sinc (W x − k). (4.3)
W W
k=−∞ k=−∞

In [3] or [9] the following result is shown.



Theorem 4.2 (Amplitude Error). Suppose that the signal f ∈ BπW satisfies (3.9) for some 0 < γ ≤ 1.
Then
4 √ √ 
kAε f k ≤ 3 ε + 2 Mf e1/4 ε log ε−1 , (4.4)
γ

for W ≥ 1 and ε ≤ min{1/W, 1/ e}.

4.3 Position-jitter Error


Finally, let us assume that the sample points themselves are not exact but taken at different positions
k/W + δk , k ∈ Z. As above, we assume that the local distortions δk are uniformly bounded: |δk | ≤ δ, for all
k ∈ Z. This jitter error Jδ is then given by
+∞     
X k k
(Jδ f )(x) := f −f + δk sinc (W x − k). (4.5)
W W
k=−∞

An estimate on the size of the position-jitter error is derived in [3] and [9] and stated below.

Theorem 4.3 (Position-jitter Error). Suppose that the signal f ∈ BπW satisfies (3.9), for some 0 < γ ≤ 1.
Then
4 √ √ 
kJδ f k ≤ 5 e kf 0 k + 8 Mf e1/4 δ log δ −1 , (4.6)
γ

provided that δ ≤ min{1/W, 1/ e} and W ≥ 1.

11
12 Peter R. Massopust

4.4 Combined Error


In practice, all three errors occur for – not necessarily – bandlimited signals f . To this end, we define
the combined error C by
N  
X k
(Cf )(x) := f (x) − f + δk sinc (W x − k). (4.7)
W
k=−N

Theorem 4.4 (Combined Error). Suppose that the signal f ∈ Cb1 (R) and satisfies (3.9) for some 0 <
gamma ≤ 1. For all bandwidth W with W ≥ e2 , we have

log W
kCf k ≤ K1 (f, γ, δ, ε, W ) , (4.8)
W

provided that N = bW 1+1/γ + 1c, and where


     
1 14 0
K1 (f, γ, δ, ε, W ) := 1+ 5e +W δ kf k + W ε + 6 e (Mf + kf k) . (4.9)
γ π

References
[1] P. L. Butzer and J. Lei, Approximation of Signals using Measured Sampled Values and Error Analysis,
Communications in Applied Analysis, 4 (2)(2000), 245 – 255.

[2] P. L. Butzer and J. Lei, Errors in Truncated Sampling Series with Measured Sampled Values for Not-
Necessarily Bandlimited Functions, Functiones et Approximatio, XXVI (1998), 25 – 39.
[3] P. L. Butzer, W. Splettstößer and R. L. Stens, The Sampling Theorem and Linear Prediction in Signal
Analysis, Jber. d. Dt. Math.-Verein., 90 (1988), 1 – 70.

[4] A. Crouch and M. Smith, Resolution Impact Accuracy I and II, Pipeline and Gas Technology, August,
September 2002, 26 – 30, 35 – 42.
[5] Y. U. Egorov and M. A. Shubin (Eds.), Partial Differential Equations, I., Encyclopedia of Mathematical
Sciences, Vol. 30, Springer Verlag, Berlin 1992.

[6] A. Garcia, Orthogonal Sampling Formulas: A Unified Approach, SIAM Review, 42 (3), 499 – 512.
[7] A. Papoulis, Probability, Random Variables, and Stochastic Processes, McGraw-Hill, Inc., New York,
1984.
[8] C. E. Shannon, A Mathematical Theory of Communication, The Bell System Technical Journal, 27
(1948), 379 – 423, 623 – 656.

[9] A. Zayed, Advances in Shannon’s Sampling Theory, CRC Press, Inc., Boca Raton, 1993.

12

View publication stats

You might also like