2005vol50 64no2 PDF

Redactor ef / Editor in chief
Prof.dr.ing. Ioan Naforni

Colegiul de redacie / Editorial Board:
Prof. dr. ing. Virgil Tiponu
Prof. dr. ing. Alexandru Isar
Conf. dr. ing. Dorina Isar
Prof. dr. ing. Traian Jurc
Prof. dr. ing. Aldo De Sabata
Prof. dr. Ing. Mihail Tnase
As. ing. Kovaci Maria - secretar de redacie
Colectivul de recenzare / Advisory Board:

Prof. dr. ing. Monica Borda, UT Cluj-Napoca
Conf. dr. Romeo Negrea, UP Timioara
Prof. dr. ing. Virgil Tiponu, UP Timioara
Prof. dr. ing. Alexandru Isar, UP Timioara
Prof. dr. ing. Ioan Naforni, UP Timioara
Buletinul tiinific
al
Universitii "Politehnica" din Timioara
Seria ELECTRONIC I TELECOMUNICAII
TRANSACTIONS ON ELECTRONICS AND COMMUNICATIONS
Tom 50(64), Fascicola 2, 2005

CONTENTS
Mircea Ciugudean, Beniamin Drgoi, Aurel Filip:
" Simulated-Inductance High-Stability Sine Oscillator"............................................... 3
Dominique Pastor:
"On the coverage probability of the Clopper-Pearson confidence interval".................7
Bado Pascal:
"Speech denoising using Diversity Enhanced Discrete Wavelet Transform and MAP
filters associations" .....11
Iulian Voicu, Monica Borda:
"Color Image Processing Using Complex Wavelet Transform.....18
Miranda Nafornita, Florin Chis:
"Modern Network Algorithms "....................................................................................24
Raul Mluan, Monica Borda, Pedro Gmez Vilda, Francisco Daz:
"Hybridization modelling in estimation of oligonucleotide microarray
expression".............................................................................................................................. 29
Mircea Coer:
"Patents in the framework of TRIZ "........................................................................... 34
Mirella Amelia V. Mioc:
"A comparison between a linear feed-back shift register and a multiple input shift
register".................................................................................................................................. 39
Buletinul tiinific al Universitii Politehnica din Timioara

Seria ELECTRONIC i TELECOMUNICAII
TRANSACTIONS on ELECTRONICS and COMMUNICATIONS
Tom 50(64), Fascicola 2 , 2005
Simulated-Inductance High-Stability Sine Oscillator

Mircea Ciugudean1, Beniamin Drgoi1, Aurel Filip1
Abstract The theory of a new sine LC oscillator with
simulated inductance using two operational amplifiers is
developed. The resulting parallel resonant LC circuit may
provide a high quality factor which assures very good
frequency stability and very small harmonic distortions.
The circuit oscillation condition is derived, the qualityfactor equation and the voltage phase diagrams are
established, then the performances are presented .
Key words: simulated inductance, sine oscillator, RC
oscillator, high frequency stability.
1.
vo1
A1
vi
R1
vi
D2
vo2
vi
Leq
vi
INTRODUCTION
A2
A double-simulation sine oscillator based on a

resonant parallel circuit LeqCeq is known, whose
components are simulated by the help of OAs [1]. This
oscillator has been named the electronic quartz
because of their very high quality factor and frequency
stability. Using a similar principle, we proposed a new
LeqC oscillator whose only inductance is simulated by
the help of two OAs [2], [3], [4].
2.
R3
C2
D1
C1
R2
R4
Fig.1. Oscillator scheme
The limiting device does not actually deteriorate the

quality factor of the oscillating parallel C1-Leq circuit
because of the very slight diode opening, whose
dynamic equivalent resistance is moreover amplified
by the quality factor [1].
This new-type oscillator [3, 4] is of the same
category as the presented in the work [1] one, which
may be obtained by a similar procedure from another
Antoniou-type band-pass filter [2].
We have established both the circuits pass from the
filter function to oscillator function when the quality
factor exceeds a certain minimum value. When
eliminates the generator and its output resistance the
oscillating parallel C1-Leq circuit may attain a very
high quality factor.
In order to derive the circuit oscillation condition
one uses the open-loop technique. Thus, removing the
limiting diodes and supposing ideal OAs and
capacitors, the circuit open loop becomes as in Fig.2.
For the A1 amplifier output one may write the
voltage equation:
OSCILLATOR PRINCIPLE
The circuit in Fig.1, without the D1 and D2 diodes,

fed by vi voltage from a generator with medium-value
output resistance (less than 1 M) is a band-pass
Antoniou-type filter [2].
The circuit attached to C1 capacitor, including the A1
and A2 operational amplifiers and the C2, R1, R2, R3, R4
passive components represents a simulated small-loss
inductance [2]:
R R R
L eq = C 2 1 3 4
R2
Thus, the C1 capacitor and the Leq inductance achieve a
resonant parallel circuit with a high quality factor
which is deteriorated by the generator output resistance.
When the generator is eliminated and low resistances
are used, the circuit in Fig.1 becomes a sine oscillator
having the output-voltage amplitude limited by the
amplifier supply sources.
To obtain small output-voltage amplitude it is
necessary to use a limiting device such as a simple
diode, two counter-parallel diodes, a serial group
involving a Zener diode and a simple diode or two back
to back Zener diodes.
v o1 = v i
R4
R2 +R4
Z
Z
1 + 2 v o 2 2
R
R
3
3
(1)
Likewise, for the A2 amplifier output one may write:

R
R
Z1
1 + 3 v o1 3
(2)
v o 2 = v o1
R 1 + Z1
Z2
Z2
Deriving vo1 from equation (2) and replacing him in
the equation (1) one obtains:
------------------------------------------------------------------------------------------------------------------------------
Facultatea de Electronic i Telecomunicaii, Departamentul

Electronic Aplicat, Bd. V. Prvan, Nr.2, 300223 Timioara,
e-mail: Mircea.Ciugudean@etc.upt.ro
C2 (Z2)
vi
R1
A1
vo1
the diodes are missing, from the R1-C1 divider in Fig.1

one may find the phase difference between voltages vi
and vo1, for C1=C2 as:
R3
A2
tg =
vo2
R4
Fig.2. Oscillator open loop
R4 Z2
1
Z2
1+ = vo2
(3)
vi
+
Z R R R
R2 + R4 R3
3
1
3
3
1+
R1 + Z1 Z2 Z2
Here, to carry out the oscillation condition, must be
taken the equality vi = vo2. So, after some simple
calculus one finds the characteristic equation:
R R R
Z1 Z 2 + 1 3 4 = 0
(4)
R2
1
1
and Z 2 =
jC 1
jC 2
the characteristic equation becomes:
( j)2 C1C 2 + R 2 = 0
R 1R 3 R 4
(9)
This may be modified through simultaneous

adjustment of R1 and R2 resistors whose ratio must be
kept invariable so that the frequency does not change.
Unfortunately, as may see below, in this case the
output-voltage amplitudes modify. It is yet possibly, if
limiting diodes exist, to use vi as oscillator constant
output voltage (by the help of an AO repeater, Fig.5)
instead of vo2. Thus, the oscillator may be used as two
voltage generator with adjustable phase displacement.
With relation (9), the output voltages vo1 and vo2
have the following amplitudes relative to vi amplitude
(vim):
C1
R2
R 1R 2
R 3R 4
v o1m = v im 1 +
R
R 1R 2
, v o 2 m = v im 1 + 2
R 3R 4
R4
vR2=vR3
0 vi= vR4=vC1
Replacing here Z1 =
=45
(10)
vo2
vC2=vR1
or
vo1
R2
C1 C 2
=0
(5)
R 1R 3 R 4
Heaving always in the left part of this equation a real
expression, this mean the amplitude oscillation
condition is invariably fulfil. Thus, the circuit in Fig.1
is an oscillator. From equation (5) the oscillation
frequency may be written:
2
R2
C1 C 2 R 1 R 3 R 4
=
1
f=
2
a) R1=R3
0 vi= vR4=vC1
=60
vC2=vR1
(6)
In the case of equalities C1=C2=C, R1=R3=R the

frequency will be:
1
2RC
vo2
or
R2
C1 C 2 R 1 R 3 R 4
fo =
vR2=vR3
R2
R4
b) R1=3R3
vo1
(7)
Fig.3. Voltage phase diagrams with C1=C2 and R2=R4.
Thus, one may observe the frequency-adjusting

possibility by the help of a potentiometer replacing both
the resistors R2 and R4 (Fig.6). For yet R2=R4 the
frequency has the simplest expression:
1
(8)
fo =
2RC
The circuit simulation established this oscillates also
with both diode amplitude limiter and small-loss
capacitors.
For the case of equalities: R1=R3 and R2=R4 the

diagram in Fig.3a is valid, =45o and the outputvoltage amplitudes are:
v o1m = 2 v im
and
v o 2 m = 2v im . For R1=3R3 and R2=R4 the diagram in
Fig.3b is valid, =60o and the output-voltage
amplitudes are: v o1m = v o 2 m = 2v im . This last
situation may be exploited to generate, with the help of
another two OAs, a three-phase sine voltage [5]. If
R2R4 (for example the case of frequency adjusting)
the phase diagrams have the form of Fig.4.
3. VOLTAGE PHASE DIAGRAMS

The output and input-voltage phase diagrams are
provided in Fig.3 and Fig.4 for different situations. If
vR2=vR3
0 vi= vR4=vC1
<45
point at vi sine voltage amplitude [1].

One imposes a value of 1k for the rdp resistance if
a single diode is being used, 2k for two diodes and
3k in the case of a Zener diode. This value assures a
good compromise between the output voltage
limitation efficiency and spectral purity respectively.
Due to the Q quality-factor great value (as we seen it
above) rdeq is great too, and the limiting diode slightly
influences the output-voltage THD factor. A
0.010.02% THD value may be obtained [1].
Consequently, the output voltage form is extremely
close to a sinusoidal one.
If we use a limiting branch with Zener and a simple
diode in series, the vim voltage amplitude is close to Vz
voltage. So, we may obtain the wanted output voltage
amplitude. It is also possible to use as a limiting device
one or two (not lighting) LEDs, when the vi voltage
amplitude is vim1.3V. LED gives also the advantage
of a better voltage thermal stability.
The most successful limiting solution is the use of
one or two thermo-stabilized-transistor collector
junctions from A726 (Fairchild) integrated circuit. In
this case the temperature influence on the outputvoltage amplitude and frequency is minimized [1].
vo2
o
vC2=vR1
vo1
a) R2>R4
0 vi= vR4=vC1 vR2=vR3 vo2

>60o
vC2=vR1
vo1
b) R2<R4
Fig.4. Voltage phase diagrams with C1=C2 and R2R4.
4. QUALITY FACTOR AND AMPLITUDE

LIMITING
The calculus of the resonant circuit LeqC1 quality
factor, including the regulating diode, OA input
resistance and capacitor losses effects, for C1=C2 and
R1=R2=R3=R4=R, gives the formula:
rdeq
Q
R i 0.5R pC
(11)
R 2R + rdeq
)(
Fig.5 shows an oscillator with frequency adjusting

by the help of the P potentiometer. This changes the
R2/R4 ratio and, unfortunately, also changes the output
voltage amplitudes apart from vim, if limiting diodes
exist. To use vi as approximately-constant output
voltage relative to the frequency the scheme needs an
AO repeater.
here: - rdeq is the diode equivalent dynamic resistance

(established below),
- RpC is the parallel loss resistance of each
capacitor,
- Ri is the OA input resistance.
Usually heaving rdeq >>R the above formula may be
C2
simplified as:
Q
R i 0.5R pC
vo1
(12)
vi
A2
R1
vo3
vi
Leq
vi
A3
vo2
R
Using low-value resistors R, low-loss capacitors and
high-input resistance OAs a quality factor of
10002000 may be achieved at low frequency.
In fact, the simplest amplitude limitation may be
obtained by means of one or two diodes (with counterparallel connection) as shown in Fig.1. The
symmetrical amplitude limitation using two devices
assures better spectral purity. In accordance with [1] the
diode (diodes) gives, in parallel with the resonant
circuit Leq-C1, a dynamic equivalent resistance:
rdeq kQrdp
A1
R3
D2 D1
C1
R2
R4
Fig.5. Adjusted-frequency oscillator
(11)
5. EXPERIMENT RESULTS
where: - k is a coefficient depending upon the limiting

diode type and number; for a single diode k=2.2 and for
two counter-parallel diodes k=0.5 ; for a limiting
branch with a Zener diode and an ordinary diode in
series, back to back, k=1 and for two this last type
counter-parallel branches k=0.25;
- Q is the quality factor of the resonant C1Leq
circuit and it has a great value;
- rdp is the dynamic "peak" resistance of a diode,
defined as the v-i characteristic slope in the working
In Fig.6 a practical example of this type oscillator

with simulated inductance is given, providing a sine
signal of 2.34 KHz. The quality factor of the resonant
circuit is Q=1580, obtained by using JFET-input OAs,
low R resistances and low-loss capacitors. The
oscillator has a frequency temporal instability of 2.10-6
and a THD of 0.02%.
To avoid the temperature effect it is necessary to
use R-C reciprocally-thermal-compensated pairs and
C2
V+
TL08
vo1m=
0.7V
A1
68nF
vim=0.5V
VR1 1k,1%
BA243
D2
V+
TL082
A2
VR2
vim=0.5V
D1
C1
68nF
high quality factor (>1000), which assures very good

frequency stability and very small harmonic distortions
(THD of 0.02%).
In order to derive the circuit oscillation condition
one used the open-loop technique. So, after calculus,
one finds the characteristic equation and the frequency
relation. One may observe the frequency-adjusting
possibility by the help of a single potentiometer.
The phase-diagram of output voltages shows a
phase displacement of 45o or 60o. This last situation
may be exploited to generate, with the help of another
two OAs, a three-phase sine voltage. The oscillator
may be used as two-voltage generator with adjustable
phase displacement.
Using R-C reciprocally-thermal-compensated pairs
and collector-junctions from A726 IC (with thermostabilized regime) as limiting diodes one may achieve
a relative frequency instability of 2.10-6/oC and a
relative voltage-amplitude instability of 2.10-4/oC.
The simulations confirmed the oscillator theory and
quality performances. This new circuit enriches the
high-stability simulated-inductance oscillator family.
R3 1k,1%
vo2m=
1V
1k,1%
R4
1k,1%
Fig.6. Experimented oscillator scheme.
collector junctions from A726 IC with thermostabilized regime as limiting diodes [1]. Thus, it may
obtain the relative frequency instability of 2.10-6/oC and
relative voltage-amplitude instability of 2.10-4/oC.
6. SIMULATION RESULTS
REFERENCES
The simulations where developed with usual

amplitude-limiting diodes (missing the A726 IC
model) and confirmed the oscillator theory and quality
performances.
The time diagram of the input and output voltages is
shown in Fig.7.
[1] M. Ciugudean, Ph. Marchegay, Double-Simulation New

Quadrature Sine Oscillator the Electronic Quartz, Buletinul
tiinific al Universitii Politehnica din Timioara, Tom 49(63),
Fascicola 1, 2004, p.142-146.
[2] A. Antoniou, "Realisation of Gyrators Using Operational
Amplifiers and their Use in Active Network Synthesis", Proceedings
of IEE, Vol.116, November 1969, pp.1838-1850.
[3] M. Ciugudean, M. Panti, Oscilator sinusoidal RC de nalt
calitate, Patent OSIM, Bucharest, No.109492/1993.
[4] M. Ciugudean, A. Filip, A. Avram, M. Panti, A new Highstability
Sine
Oscillator
with
Simulated
Inductance,
Microelectronics and Computer Science,Communications and
Electronics Systems, Chiinu, 2005, pp. 356-359.
[5]. M. Ciugudean, L. Jurca, I. Lie, "Three-Phase Sine Generators",
Electronics and Telecommunications Symposium, Timioara,
September 1996, Vol.I, pp. 8-73.
7. CONCLUSIONS
The theory of a new sine LC oscillator, with
simulated inductance, using two operational amplifiers
is developed. Using R-C components for inductance
simulation, the oscillator becomes a RC one. The
resulting parallel resonant Leq-C circuit may provide a
Fig.7. Time diagram of the input and output voltages
Buletinul Stiintific al Universitatii Politehnica din Timisoara

si TELECOMUNICATII
Seria ELECTRONICA

On the coverage probability of the
Clopper-Pearson confidence interval
Dominique Pastor
We start by giving a description of the 100(1 )% CPCI,

that is the CPCI with level of confidence (1 ) where
(0, 1). This description is purposely brief for it focuses only
on the material needed for stating proposition 1 below. For
further details on the construction of the CPCI, the reader can
refer to numerous papers and textbooks on statistics ( [2], [3],
[5], [6] amongst many others).
Let (`k )k{0,...,n} be the sequence defined as follows. We
put
`0 := 0.
(1)
Abstract Given (0, 1), the coverage probability of the

100(1 )% Clopper-Pearson Confidence Interval (CPCI) for
estimating a binomial parameter p is proved to be larger than or
equal to 1 /2 for sample sizes less than a bound that depends
on p. This is a mathematical evidence that, as noticed in recent
papers on the basis of numerical results, the CPCI coverage
probability can be much higher than the desired confidence
level and thence, that the Clopper-Pearson approach is mostly
inappropriate for forming confidence intervals with coverage
probabilities close to the desired confidence level.
I. I NTRODUCTION
NTERVAL estimation of a binomial proportion p is one of
the most basic and methodologically important problems in
practical statistics. A considerable literature exists on the topic
and manifold methods have been suggested for bracketing a
binomial proportion within a confidence interval.
Among those methods, the Clopper-Pearson Confidence
Interval (CPCI), originally introduced in 1934 by Clopper and
Pearson [4], is often referred as the exact procedure by some
authors for it derives from the binomial distribution without
resorting to any approximation to the binomial. However,
despite this exacteness, the CPCI is basically conservative
because of the discreteness of the binomial distribution: given
(0, 1), the coverage probability of the 100(1 )% CPCI
is above or equal to the nominal confidence level (1 ).
In this paper, it is proved that this coverage probability is
actually larger than or equal to 1/2 if the sample size is less
than ln(/2)/ ln(max(p, 1p)). This bound basically depends
on the proportion itself. Thence, as suggested by numerical
results presented in [2] and [3], the CPCI is not suitable
in practical situations when getting a coverage probability
close to the specified confidence level is more desirable than
guaranteeing a coverage probability above or equal to the said
confidence level.
For k {1, . . . , n}, `k is defined as the unique solution in

(0, 1) for in the equation

n
X
n
i (1 )ni = /2.
(2)
i
i=k
Similarly, the sequence (uk )k{0,...,n} is defined as follows.

We set
un := 1
(3)
and, for k {0, . . . , n 1}, uk is the unique solution in (0, 1)
for in the equation

k
X
n
i (1 )ni = /2.
(4)
i
i=0
Let X henceforth stand for a binomial variate for sample size

n and binomial parameter p [0, 1]. In other words, X stands
for the total number of successes in n independent trials with
constant probability p of success.
Since X is valued in {0, 1, . . . , n}, the random variables `X
and uX are well-defined. The 100(1 )% CPCI for size n
is then the random interval (`X , uX ) whose lower and upper
endpoints are `X and uX respectively. The CPCI coverage
probability is defined as the probability P ({`X < p < uX })
that the CPCI actually brackets the true value p of the
proportion. The 100(1)% CPCI is known to be conservative
in the sense that its coverage probability is always above or
equal to the confidence level ([2], [3], [4]). This standard result
is refined by the following proposition whose proof is given
in section 3.
II. T HEORETICAL RESULTS

Throughout the rest of this paper, stands for some real
value in the interval (0, 1) and n for some natural number.
1 ENST Bretagne, Technopole de Brest-Iroise, CS - 83818, 29238 Brest
Cedex 3, FRANCE, e-mail: dominique.pastor@enst-bretagne.fr
Proposition 2.1: With the same notations as above,

(i)
If n < ln(/2)/ ln(2), then

P ({`X < p < uX })
(ii)
0
for p = 0,
1/n
1 /2 for 0 < p (/2)
,
1/n
1/n
1
for (/2)
< p < 1 (/2)
,
1/n
1 /2 for 1 (/2)
p < 1,
0
for p = 1.
0
for p = 0,
1 /2 for 0 < p < 1 (/2)1/n ,
1
for 1 (/2)1/n p (/2)1/n ,
1 /2 for (/2)1/n < p < 1,
0
for p = 1.
If n ln(/2)/ ln(2), then

P ({`X < p < uX })
The following result derives from the proposition above.
0.995
Lemma 2.2: With the same notations as above, for every

proportion p (0, 1) and every natural number n less than
ln(/2)/ ln(max(p, 1 p)), the CPCI coverage probability is
larger than or equal to 1 /2.
0.99
Coverage probabilities
0.985
P ROOF: Since max(p, 1 p) is larger than or equal to

1/2, ln(/2)/ ln(2) is less than ln(/2)/ ln(max(p, 1
p)). If n < ln(/2)/ ln(2), the result then follows from
proposition 1, statement (i). If ln(/2)/ ln(2) n <
ln(/2)/ ln(max(p, 1 p)), p is either larger than (/2)1/n
or smaller than 1 (/2)1/n and the result follows from
proposition 1, statement (ii).
0.98
0.975
0.97
0.965
0.96
0.955
0.95
The theoretical results above are thus mathematical evidences that, as suggested in [2] and [3], the CPCI is inaccurate
in the sense that [...]its actual coverage probability can be
much larger than 1 unless the sample size n is quite large.
( [3, Sec. 4.2.1] ).
Figures 1, 2 and 3 illustrate the foregoing by displaying the
coverage probability of the 95% CPCI for p = 0.1, p = 0.05
and p = 0.01 respectively and sample sizes ranging from 1 to
500. In each figure, the value of ln(/2)/ ln(max(p, 1 p))
is represented by a vertical red line. On the left hand side of
this line, coverage probabilities are all larger than or equal
to 97.5%; on the right hand side of this same line, coverage
probabilities can be less than 97.5% and even close to 95%.
50
100
150
200
250
300
350
400
ln(/2)/ln(max(p,1p)) = 35.0120
450
500
Sample sizes
The 95% CPCI Coverage probabilities for proportion p = 0.1
Fig. 1.
1
0.995
0.99
0.985
0.98
0.975
0.97
0.965
III. P ROOF OF PROPOSITION 2.1
0.96
Given [0, 1], let P stand for the distribution of a

binomial variate for sample size n with binomial parameter
and let F be the distribution function defined for every real
value x by F (x) = P ((, x]). It is then convenient to
set G (x) = P ([x, )) for every real value x. According to
the definitions of F and G , the left hand sides in (2) and
(4) are equal to G (k) and F (k) respectively. Therefore, by
definition of `k and uk ,
G`k (k) = /2 for every k {1, . . . , n}
0.955
0.95
50
100
150
ln(/2)/log(max(p,1p)) = 71.9174
Fig. 2.
200
250
300
350
400
450
500
Sample sizes
and
Fuk (k) = /2 for every k {0, . . . , n 1}.
(5)
(6)
For every [0, 1], G (n) = n . It then follows from

(5) that `nn = /2, which completes the proof of statement (i).
0.995
Proof of statement (ii). The strict increasingness of the

sequence (uk )k{0,...,n} derives from the same type of
arguments as those used for proving statement (i). Given
k {0, . . . , n 1}, uk belongs to (0, 1) and is therefore less
than un = 1. It then suffices to show that uk < uk+1 for
every k {0, . . . , n 2} to establish the strict increasingness
of the sequence.
Let k {0, . . . , n 2}. We have that Fuk+1 (k + 1) >
Fuk+1 (k). According to (6), Fuk+1 (k + 1) = Fuk (k).
Therefore, Fuk (k) > Fuk+1 (k). The result then follows from
lemma 1, statement (i).
Given (0, 1), F (0) = (1 )n . Therefore, by (6), we
obtain that Fu0 (0) = (1 u0 )n = /2.
0.99
0.985
0.98
0.975
0.97
0.965
0.96
0.955
50
100
150
200
250
Sample sizes
Fig. 3.
300
350
400
450
500
ln(/2)/ln(max(p,1p)) = 367.0404
Proof of statement (iii). According to (1), (3) and the values

of `n and u0 , statement (iii) holds true for k = 0 and k = n.
Consider any k {1, . . . , n1}. We can write that F`k (k
1) = 1 G`k (k) = 1 (/2) > /2. This follows from the
definition of G , (5) and the fact that < 1. Since F`k (k) >
F`k (k 1) and /2 = Fuk (k) according to (6), we finally
obtain that F`k (k) > Fuk (k). The fact that `k < uk then
straightforwardly derives from lemma 1, statement (i).
According to [1, Eq. 6.6.4, p. 263], for every given

[0, 1],
G (k) = I (k, n k + 1) for k {1, . . . , n},
(7)
and
Lemma 3.3: With the same notations as above,

(i)
if p = 0, P `X p =
1,

(ii)
if 0 < p (/2)1/n , P `X p /2
(iii) if (/2)1/n < p 1, P `X p = 0.
F (k) = 1 I (k + 1, n k) for k {0, . . . , n 1}. (8)

where Ix (a, b) stands for the Incomplete Beta Function ( [1,
Eq. 6.2.1, p. 258, Eq. 6.6.2, p. 263, Eq. 26.5.1, p. 944] ). The
maps [0, 1] 7 I (k + 1, n k) and [0, 1] 7
I (k, n k + 1) are strictly increasing. Thereby, we can
straightforwardly state the following result.
P ROOF: Statement (i) is straightforward since `X 0.

As far as statements (ii) and (iii) are concerned, it is convenient to introduce the set E = {k {0, . . . , n} : `k p} .
Statement (i) of lemma 2 implies the following facts. First,
E is non empty if and only if p (/2)1/n ; second, if
0 < p (/2)1/n , E = {m, . . . , n} where m 1 according
to (1).
Thereby, `X p if and only if X m. We therefore
have that P ({`X p}) = Gp (m). Now, since p `m , it
follows from lemma 1, statement (ii), that Gp (m) G`m (m).
According to (5), the right hand side in this inequality equals
/2. Therefore, we obtain that P ({`X p}) /2.
If (/2)1/n < p 1, E is empty. Therefore, `X < p and
statement (iii) follows.
Lemma 3.1: Given [0, 1],

(i)
for every k {0, . . . , n 1}, the map [0, 1] 7
F (k) [0, 1] is strictly decreasing;
(ii)
for every k {1, . . . , n}, the map [0, 1] 7
G (k) [0, 1] is strictly increasing.
The subsequent lemma states useful properties concerning
the real values `k and uk .
(i)
the sequence (`k )k{0,...,n} is strictly increasing and
`n = (/2)1/n ;
(ii)
the sequence (uk )k{0,...,n} is strictly increasing and
u0 = 1 (/2)1/n .
(iii) for every given k {0, . . . , n}, `k < uk .

(i)
if 0 p < 1 (/2)1/n , P uX p = 0,
(ii)
if 1 (/2)1/n
p < 1, P uX p /2,
(iii) if p = 1, P uX p = 1.
P ROOF:
P ROOF: Set F = {k {0, . . . , n} : uk p}. It follows from

statement (ii) of lemma 2 that F is empty if and only if 0
p < 1 (/2)1/n . Therefore, if 0 p < 1 (/2)1/n , uX is
larger than p and statement (i) holds true.
Under the condition 1 (/2)1/n p < 1, F is not empty.
According to statement (ii) of lemma 2 and (3), we obtain that
F = {0, . . . , L} with L n 1. Therefore, the event {uX
p} is the event {X L} so that P ({uX p}) = Fp (L).
Since uL p, it follows from statement (i) of lemma 1 that
Proof of statement (i). Given k {1, . . . , n}, `k (0, 1)

and is therefore larger than `0 = 0. It then suffices to prove
that `k < `k+1 for k {1, . . . , n 1} to establish the strict
increasingness of the sequence (`k )k{0,...,n} .
Let k {1, . . . , n 1}. We have that G`k+1 (k + 1) <
G`k+1 (k). According to (5), G`k+1 (k + 1) = G`k (k). It thus
follows that G`k (k) < G`k+1 (k). According to lemma 1,
statement (ii), we can conclude that `k < `k+1 .
Fp (L) FuL (L). According to (6), the right hand side in this
inequality equals /2 and statement (ii) follows.
Statement (iii) holds true since uX 1.
We now complete the proof of proposition 2.1. According
to lemma 3.2, statement (iii), `X < uX . Thereby, {p `X }
{p < uX } so that
P ({`X < p < uX })
= P ({p < uX }) P ({p `X })

= 1 P ({p uX }) P ({p `X })
Since the condition 1 (/2)1/n (/2)1/n is equivalent to

n ln(/2)/ ln(2), proposition 1 straightforwardly follows
from lemmas 3.3 and 3.4.
IV. C ONCLUSION
Because of the discreteness of the binomial distribution, the
coverage probability of the 100(1 )% CPCI for estimating
a binomial parameter p is larger than or equal to 1 . In
addition to this standard result, this paper proves that this
coverage probability is, in fact, larger than or equal to 1 /2
when the sample size is less than a specific bound. Because
this bound depends on the proportion to estimate, the coverage
probability of the CPCI can be much larger than the confidence
level. This is a major drawback of the CPCI. Therefore, the
CPCI is not suitable for applications where it is important to
form a confidence interval whose coverage probability is close
to the specified confidence level.
R EFERENCES
[1] A BRAMOWITZ , M. and S TEGUN , I. (1972). Handbook of Mathematical
Functions, Dover Publications, Inc., New York, Ninth printing.
[2] AGRESTI , A. and C OULL , B. A. (1998). Approximate is better than
exact for interval estimation of binomial proportions, The American
Statistican, 52, NO. 2, 119-126.
[3] B ROWN , L. D. and C AI , T. and DAS G UPTA , A. (2001). Interval Estimation for a Binomial Distribution, Statist. Sci., 16, 101-133.
[4] C LOPPER , C. J. and P EARSON , E. S. (1934). The use of confidence
of fiducial limits illustrated in the case of the binomial, Biometrika, 26,
404-413.
[5] J OHNSON , N. L. and KOTZ , S. (1970). Distributions in Statistics,
Discrete Distributions, John Wiley & Sons.
[6] S TUART, A. and O RD , J. K. (1987). Kendalls Advanced Theory of
Statistics, Fifth Edition, Charles Griffin.
10
Buletinul tiinific al Universitii "Politehnica" din Timioara

Tom 50 (64), Fascicola 2, 2005
Speech denoising using Diversity Enhanced Discrete

Wavelet Transform and MAP filters associations
BADO Pascal1
i)
the bishrink filter and the Wiener filter

on the one hand and
ii)
the marginal MAP filter and the Wiener
filter on the other hand.
The structure of this paper is the following. Section II
consists in modeling the speech in order to estimate
its pdf. Section III explains the proposed denoising
method. Some variants, based on the use of different
filters in the wavelets domain are described. The
results obtained using these different variants are
compared in section IV. The final section is dedicated
to concluding remarks.
Abstract: In this paper, we present new speech denosing

methods based on some types of maximum a posteriori
(MAP) filters applied in the wavelet domain. These
methods take into account the statistical properties of
the discrete wavelets transform (DWT) coefficients of
speech signals. The experimental results are compared
with those obtained in another paper published in this
magazine, [1].
Keywords: wavelets, bishrink, MAP filters, Bayes,
Wiener.
I. INTRODUCTION
The wavelets constitute a powerful tool for
mathematical analysis which revolutionized the world
of signal processing. The wavelets are functions with
null average and finite energy. Their field of
application is very large: compression, segmentation,
denosing etc. The term denoising was introduced by
David Donoho, [2]. He treated the case of signals
perturbed by additive white Gaussian noise. The aim
of this paper is the denoising of additively perturbed
speech. Several speech denosing methods already
exist. Some of them are based on the use of the
wavelets transform, [1]. They differ by the type of
filter used in the wavelets domain but all follows the
Donohos algorithm. It is based on the following three
steps:
1. The computation of the wavelet transform, (WT),
2. The filtering of the obtained result,
3. The computation of the inverse wavelet transform,
(IWT).
Donoho used the DWT. Another WT, namelly the
diversity enhanced discrete wavelet transform,
DEDWT, was preffered in [1].
For filtering in the DEDWT domain we will use in the
following, MAP filters. Some of them will differ of
the bishrink filter, used in [1]. The knowledge of the
type of noise to be eliminated and of the probability
density function (pdf) of the useful signal are
essential for the second step of the denoising method.
The success of the denosing procedure depends on the
selection of the noise and useful signal pdfs. In the
following, we present a speech denoising method
which uses two types of MAP filters associations:
II. STATISTICAL MODELING OF THE

SPEECH SIGNAL
The various MAP filters, which we will study and
compare, require the knowledge of the pdf of the
useful signal. In this section we will identify the pdf
of speech on the basis of the comparison of the
histogram of a speech signal with the elements of a
family of usual pdfs. We consider the family of
gamma generalized distributions, [3], defined for N
random variables, X = X 1 , X 2 ,..., X N } by:
f X (x ) =
x
exp x
2()
(1)
where () is the gamma function and , , are

positive real-valued parameters; is related to and
by:
(2)
1
=
1
N
i =1
xi
For certain particular values of and , the following

models are found:
If : = 1, the (general) gamma pdf;
= 1, the generalized Gaussian pdf;
= 2 and = 0.5, the Gaussian pdf;
= 1 and = 1, the Laplacian pdf.
Analyzing this class and the histogram of speech
signal, we can observe that this signal is close to the
Laplacian pdf. This observation is highlighted in
figure 1. The Laplacian pdf is defined by:
1. ENSIETA-Brest, France
11
Laplace
speech signal
with k 0 and
k =1
k =1
Z = Yk = k X k
variables Yk , k = 1, K ,
k = 1, K , then:
fZ ( z) = fY ( z) fY ( z ) fY ( z)
K
1
2
0.8
and if the random variables
0.6
Yk , k = 1, K , represent
changes of the random variables X k , k

following the relations
0.4
(5)
= 1, K ,
Yk = k X k , then the pdf of
the random variable Z, becomes:
1
z
z
z
(6)
f X1 f X 2 f X K
This relation can be used to compute the pdf of

wavelet coefficients, f Y ( y ) of a process X, on the
fZ (z) =
0.2
0
-100
-80
-60
-40
-20
20
40
60
80
100
basis of the pdf of this process f X ( x) , if we consider
Fig. 1: The histogram of the speech segment feets well with a

Laplacian distribution.
f X (x ) =
x
1
exp
2b
b
that X 1 = X 2 = = X K = X . In our case, the

process X is a speech signal. So, the relation between
the wavelet coefficients of a speech signal and its own
pdf is:
(3)
y
1
f X
k =1 k
k
K
fY ( y ) =
where 2b is the variance and the mean. In the case

considered in the experiment already reported the
mean is null. The results obtained come to confirm the
assumption that certain authors make in the treatment
of the speech signals [1].
where
In the previous section is given a model for the speech

signal pdf. The aim of this section is the study of the
pdf transformation realized by the DWT computation.
The DEDWT is a colection of DWTs. First, we
determine the connection between the pdfs of the
input signal and of the wavelet coefficients. Second,
the relation already obtained is verified by
simulations.
u
f
X
K
k1=1 hk
1
hk1
1
k1
K
u
fU ( u ) = Ph f X
hk
k1 =1
1
and Ph =
Let us consider a random variable X with a pdf

f X (x) and the change of random variable Y =X
with a constant. The relation connecting the pdfs of
X and Y is:
y
)
= 1 f (y)
X
k represent the coefficients of the filters (h
( 7)
K 1
u
fX =
fU ( u) =
hk
k1=1 h
k1
1
A. The derivation of the wavelets transform

coefficients pdf according to the speech signal pdf
fX (
(7)
or g) used for the computation of the wavelet

transform. This relation can be applied to each filter
of the DWT computation diagram, represented in
figure 2. For the first iteration of the DWT, the pdf
obtained after the first low-pass filter h is:
III. MAP FILTERS ASSOCIATIONS
f (x)
=
fY ( x ) = X
dy
dx
(8)
hk1
k1
In the same way, the pdf obtained after the second

high-pass filter g becomes:
(7)
y
1 K y
fU =
fU
gk K

k2=1 g
k2
2 gk k2=1 gk2
K
fY ( y) =
(4)
k2
If we consider the following relation between the

random variables Z and the independent random
Fig. 2. The DWT first iteration computation diagram.
12
fY ( y) = Pg fU
k2 =1 gk
2
The central limit theorem stipulates that any

sum of independent and identically distributed
random variables will tend to be distributed according
a normal law when their number tends towards the
infinite [4]. Then we can affirm that the speech signal
pdf converges towards a Gaussian after a certain
number of convolutions, or equivalently after a certain
number of DWT iterations. The problem is how fast
this convergence is. In other words it is interesting to
know if after a finite number of DWT iterations, the
output pdf can be considered Gaussian. The number
of convolutions required for one DWT iteration
depends on the type of mother wavelets used, because
a different number of coefficients K correspond to
different types of mother wavelets. The smallest
convolutions number per DWT iteration is obtained
when the Haar wavelet mother is used. So, the
smallest convergence speed is obtained using this
mother wavelets. We have simulated the convergence
process, using the Haar mother wavelets, for which
we have the following values:
1
1
1
1
K = 2 ; h1 =
; h2 =
; g1 =
; g2 =
2
2
2
2
Starting from a Laplace pdf we obtain for the first and
the second iterations, the detail wavelet coefficients
pdf curves presented by figure 3 and figure 4. It can
be observed that the curve of figure 4 is close to a
Gaussian. So, we can assume that using Daubechies
mother wavelets (all those mother wavelets have a
higher corresponding K value), the pdf of the wavelet
coefficients of a speech signal can be considered
Gaussian, after some iterations of the DWT. So, the
starting conditions, used for the MAP filter
construction, that will be presented in the following,
concerning the distribution of the wavelet coefficients,
cannot be kept unchanged throughout the entire
denoising process in the wavelets domain. For
different iterations different models for the
distribution of the wavelet coefficients of the clean
speech must be considered.
and Pg =
(9)
g k2
k2
Using the last two relations, we can compute the

detail coefficients pdf after the first iteration:
( 8) ,( 9)
fY ( y ) = Pg
y
f
k2 =1 k1 =1 hk1 gk2
K K
y
K
fY ( y ) = Pg ( Ph ) f X
k2 =1k1 =1 hk1 gk2

K
Ph
(10)
While proceeding in the same way, the following

result is obtained for the second iteration:
fY ( y) = Pg ( Ph )
K +K 2
K K K
(11)
fX
hk gk 2hk
k =1k =1k =1
2
Finally, after the Nth iteration, the pdf of the detail

coefficients becomes:
K K
fY ( y) =P
k1=1k2=1
y
f
X
k+1=1 hk1hk2 hkN gkN+1
K
with P = Pg (Ph )
(12)
K + K 2 ++ K N
We have proved in section II, that the pdf of the

speech signal can be approximated with a Laplace
law:
f X (x ) =
x
1
exp
2b
b
So, if we compute the DWT of a speech signal, the

model of the corresponding wavelet coefficients
becomes:
fY ( y ) = P
k1 =1 k 2 =1
F (y)
k + 1 =1
(13)
B. Histograms of DWT coefficients of the speech

signal
with:
1
1
y
exp

2b
b hk1 hk 2 hk N g k N +1
If we compute the DWT of a zero mean white noise,

which has a Gaussian distribution, whose formula is:
1 x 2
1
f X (x ) =
exp ,
2
2
the output pdf is given by:

F1 ( y ) =
fY ( y ) = P
F2 ( y )
k1 =1 k 2 =1
k +1 =1
The aim of this section is to give a practical

verification of the theoretical results obtained in the
previous section. We have calculated and represented
the histograms of the DWT coefficients of the same
segment of speech signal used in section II, in order to
study their pdfs. The histograms are represented on
figure 5. The Laplace distribution is heavy tailed. Its
speed of descent towards zero is smaller then the
speed of descent towards zero of a Gaussian
distribution. Analyzing figure 5, it can be observed
that the speed of descent towards zero of the curves
increases with the level of decomposition. This means
that the pdf of the DWT coefficients converges
asymptotically towards a Gaussian. Intermediate
simulation results proved that the difference between
the speeds of descent towards zero of the curves
(14)
and:
F2 ( y) =

1
y
exp 2
2 2 hk1 hk2 hk N gk N +1
13
corresponding to the fourth and the fifth iteration is

not important; so we could think that convergence
is reached after four iterations. This conclusion is in
agreement with the results of the previous section.
output pdf
Laplace
0.8
C. Filters associations
We have just shown that the starting conditions
concerning the distribution of the wavelet coefficients
cannot be kept unchanged throughout the entire
denoising process in the wavelets domain. This is why
we propose an association of filters: a first-one when
the pdf of the wavelet coefficients is supposed
Laplacian and a second association when this law
becomes Gaussian. In this section we were
particularly interested in the choice of the second
filter. If we consider y a noisy wavelet coefficient, w
the true coefficient and n the noise, [5], we can write
that:
y=w+n
(15)
0.6
0.4
0.2
0
-50
-40
-30
-20
-10
10
20
30
40
50
Fig. 3: A comparison of the input (Laplacian) distribution with the

output distribution after the first iteration.
The classical MAP estimator for (15) is:
output pdf
Laplace
Gauss
w( y ) = arg max Pw y (w y )
w
Using Bayes rule, one gets:
0.8
w( y ) = arg max Py w ( y w) Pw (w)
0.6
(17)
0.4
w( y ) = arg max[Pn ( y w) Pw (w)]
w
Equation (17) is equivalent to:
w( y ) = arg max ln Pn ( y w) + ln Pn (w)

w
0.2
0
-40
(16)
-30
-20
-10
10
20
30
40
Fig. 4: A comparison of the input (Laplacian) distribution with the

output distribution after the second iteration and with the Gaussian
distribution.
(18)
If the signal and the noise have both a Gaussian pdf,

then:
2
Pn (n ) =
signal before DWT

it1
it5
1 n
exp
2 n
n 2
(19)
0.8
and
0.6
Pw (w) =
0.4
(20)
With the aid of the last two relations, the argument of

the right hand side of (18) becomes:
0.2
0
-40
1 w 2
1
exp
2
2
-30
-20
-10
10
20
30
40
Fig. 5: A comparison of the histograms of the speech signal before

the calculation of the DWT and those of the DWT coefficients after
the first and the fifth iteration.
14
ln Pn ( y w ) + ln Pw ( w ) =
1 ( y w) 2
ln
2
2n2
n
The solution of (22) is in this case:
w( y w) 22n w = 0
(21)
or:
sgn(w)( y w) 2 2n = 0
1 ( w)
+ ln
2
2 2
This equation has two possible solutions:

If w >0, the solution is:
2
= y 2 n
To maximize (21), the following equation is solved:
d ln Pn ( y w ) + ln Pw ( w )
dw
=0
Using (21) it becomes:
( y w ) 2 (w )
2
22n
or:
2 2
=0
+ n
If we define:
(g )+ = g0 ,,
if g > 0
otherwise ,
the relation (26) can be written in the final form:
w = sign ( y ) y
(24)

+
2
(27)
The right hand side of equation (27) is the classical

soft shrinkage function. This is the input-output
relation of the MAP filter still called soft-thresholding
Supposing that the wavelet coefficients corresponding

to the noise n are Gaussian distributed, the MAP
equation becomes:
filter. It requires to estimate
ln Pn ( y w) + ln Pw (w) =
2n
y> 2
, the solution of the MAP
equation can be written in the following compact

form:
n2
(
)
(
)
=
sign
y
y
2
sign
y
(26)
w
Then the second filter to be considered is necessarily

the Wiener filter. For the first filter, we have two
possibilities:
a) If we do not take into account the inter-scale
dependency of the wavelet coefficients, for the first
iterations, where the coefficients of the DWT of the
speech signal follows a Laplace law, one obtains:
1 ( y w) 2
1
+ ln
ln
2
2
2n
n 2
Therefore, if
[ n]
w
1
Pw (w) =
exp 2

2
then w < 0, and the solution of the
MAP equation becomes:
2
= y+ 2 n
This relation corresponds to the zero order Wiener

filter whose impulse response is:
2
If y < 2 n
So, the solution of the MAP equation is in this case:
2
(23)
w= y 2 2
+ n
2
2 n then
w > 0, and the solution of the MAP equation is:
2
= y 2 n
We can notice that if y >
w 2n + 2 + 2 y = 0
h [ n] =
If w<0, the solution is:
2
= y+ 2 n
(22)
n2 , the noise variance

of the useful signal.
and the standard deviation

The estimate of the noise variance is done by using
the sequence of the details obtained after the DWT
first iteration and the estimate of the standard
deviation of the useful signal coefficients is done
locally in a moving window. In short, if the scale
inter-dependency of the DWT coefficients is not
considered, the filter which should be used is the soft-
(25)
2w
15
thresholding filter with local threshold. In the

following, this filter will be called marginal MAP
filter.
b) On the other hand if the estimate takes into
account two successive iterations, we can write:
w = (w1 , w2 ) and the pdf proposed by Levent
Sendur and Ivan W. Selesnick [3] for the wavelet
coefficients, is:
Pw (w) =
3
2
2
+
exp
w
w
1
2

22
2
2
y1 + y22 3n

+ y
1 =
w
1
y12 + y22
This relation gives place to the use of a filter called

bishrink filter.
IV.
Section III enables us to consider two

associations of possible filters:
1. In the case where the inter-scale dependency of the
DWT coefficients is not considered, for the first
iterations, we use the marginal MAP filter and for the
last iterations, we use the Wiener filter. We called this
association of filters, the MAP_Wiener filter.
2. When we take into account the inter-scale
dependency of the coefficients of two successive
iterations, the association that we called
Bishrink_Wiener is used.
We simulated the denoising method proposed and
compared the signal to noise ratio (SNR)
enhancement obtained for these two associations of
filters for the marginal MAP filter and for the bishrink
filter, proposed in [1], called in the following
Bishrink_f, for different values of the input SNR. The
results are represented on figure 6.
A first conclusion obtained analyzing this
figure is that the Bishrink_Wiener gives a better result
than the bishrink filter and the Bishrink_f filter. This
comes to confirm the theoretical results obtained in
section III, highlighting the importance of
commutation between models, especially when the
inter-scale dependency is considered. Afterwards,
there is not any difference between the MAP_Wiener
filter and the marginal MAP filter. The two
corresponding curves are identically. So, if the interscale dependency is not considered, the commutation
between models is not required. These curves have the
larger increasing speed, reason that makes us think
that for the speech signals, the inter-scale dependency
between the DWT coefficients of two consecutive
iterations can be neglected in the denosing process.
We have repeated the experiments with colored noise.
It was generated by filtering a zero mean white
Gaussian noise with the aid of a first order IIR lowpass filter. We have simulated two such filters, having
different cut-off frequencies. We have obtained each
time practically the same results like those represented
in figure 6. So, the denoising method proposed in this
paper, is robust against the nature of the power
spectral density of the perturbing noise. This is, of
course, an important advantage of the proposed
denoising method. It can be also observed that the
proposed method has an excellent performance for
low input SNRs. At 5 dB, its SNR gain is of 10 dB.
becomes:
w( y ) =
( y w )2 ( y w )2
arg max 1 21 2 2 2 + f (w)
2 n
2 n
w
(29)
This is equivalent to solve the system composed by
the following two equations, if f (w) is assumed to
be strictly convex and differentiable:
y1 w1
f
(30)
+
1 w = 0
n2

and:
y2 w 2
f
(31)
+
2 w = 0
2n

where f 1 and f 2 represent the derivative of f (w)
with respect to w1 and w2 , respectively. Taking into
account the relation (28), it can be written:
3
3
f (w) = log
w12 + w22 (32)
2
2
and the solutions of the system already mentioned are:
3w1
w12
+ w22
(33)
and:
f 2 (w) =
3 w2
w12 + w22
EXPERIMENTAL RESULTS
(28)
Let us find the MAP estimator corresponding to the

model given in (28), if the noise is again considered to
be Gaussian distributed.
If we define f (w) = log(Pw (w)) , then (18)
f1(w) =
(35)
(34)
Solving (30) and (31) by using (33) and (34), the

MAP estimator can be written as:
16
REFERENCES:
15
output
SNR
(dB)
[1]
10
[2]
RSB map
RSB map wiener
RSB bishrink f
RSB bishrink wiener
[3]
[4]
[5]
-5
-5
10
15
input SNR (dB)
Fig. 6: Comparison of the output SNR for various MAP filters.
V. CONCLUSION
In this article, we proposed new methods of speech
enhancement. At the beginning, on the basis of the
statistical properties of the speech signals and those of
the DWT coefficients, we showed that it is necessary
to combine different MAP filters at different DEDWT
decomposition levels for an efficient denoising.
Combining different types of filters, deduced in this
paper, we have built two MAP filters associations. To
follow, we have tested, by simulation, these
associations. The results obtained by comparing the
output SNR, are better than those of the Bishrink filter
and those of the Bishrink_f filter, used in [1]. So, the
method proposed is better than the noise spectral
subtraction method or the pure statistical denoising
method, that has inferior SNR improvements, versus
the association DEDWT-Bishrink_f, like is proved in
[1]. Moreover, these results show that the association
that does not take into account the inter-scale
dependency of the DWT coefficients is the best for
denoising speech signals. The method proposed is
efficient especially for low input SNR signals (see
figure 6). This is the reason why it can be used in
applications where the quality of speech is very bad.
For example, the proposed denoising method can be
applied for the enhancement of the robustness of a
Voice Activity Detector. When the SNR of the speech
is high enough, the distortion introduced by the
proposed denoising method is important, being better
to avoid its utilization in applications where a high
quality of the speech signal is required.
17
Cdric Clolus, Speech enhancement using Enriched

Diversity wavelet Transform and Bishrink filter,
Buletinul Stiintific al Universtatii Politehnica din
Timisoara, Tom 50(64), Fascicola 1, 2005.
D. L Donoho, De-nosing by soft-thresholding, IEEE
Trans. Inform. Theory, vol.41, pp. 613-627, May 1995.
Jong Won Shin, Joon-Hyuk Chang, Nam Soo Kim,
Statistical modeling of speech signals based on
generalized gamma distribution, IEEE Signal
Processing Letters, vol. 12 no. 3 pp.258-261, Mar. 2005.
http://en.wikipedia.org/wiki/Central_limit_
theorem.
L. Sendur and I. W. Selesnick, Bivariate shrinkage
functions for wavelet-based denoising exploiting
interscale dependency, IEEE Trans. on Signal
Processing. 50(11): 2744-2756, November 2002.

Color Image Processing Using Complex Wavelet

Transform
Iulian Voicu1, Monica Borda2
Abstract Discrete wavelet transform became a
powerful method for signal processing in the last
decades. One of the most important applications of this
tool is compression, for the bi-dimensional signals, JPEG
2000 being an example. Complex wavelet transform
offers improved directionality compared with classical
discrete wavelet transform but it is redundant. The
paper presents an experimental study that refers to R,
G, B images compressed with complex or classical
wavelet transform. This study wants to emphasize the
tradeoffs
between
redundancy
and
improved
directionality of complex wavelet transform in
compression field.
For this study we used for
compression Embedded Zero-Tree Wavelet (EZW)
algorithm proposed by Shapiro. The experimental
results discussed at the end of the paper have been
realized for lower and higher bit rates for both
transforms. The whole work is intended to find reliable
methods for applying Dual-Tree Complex Wavelet
Transform (DTCWT) in JPEG 2000 standard.
Keywords: complex wavelets, compression, JPEG 2000
Figure 1. DTCWT for one-dimensional case
more directions that will hopefully perform better in

the compression field. After a tremendous work
Kingsbury [6], [7], [8] propose a new transform with
I. INTRODUCTION
improved directionality.
He called it Dual Tree Complex Wavelet
The discrete wavelet transform (DWT) [1], [2], [14],
Transform (DTCWT), because uses two DWT in
have been developed in the last decades and became
parallel trees. Two branches decompose the signal
an important method for signal processing.
independently as it can be seen in figure 1. Finally,
It was applied successfully in compression
the subbands from both trees are combined in a proper
applications, including image compression, in this sense
manner attaining six subbands with complex
a new standard JPEG2000 [3], [4], [5] being developed
coefficients oriented at 15 0 , 45 0 and 75 0 .
by the researchers community.
Figure 1 presents the block scheme for the oneThe heart of the JPEG2000 standard is DWT,
dimensional DTCWT case, where H ( z ) represents
which is the most important stage in image
compressing chain. JPEG 2000 standard specify for the Z-domain of the filters used in decomposing the
DWT the 9/7 taps filters for lossy compression or 5/3 signal. More details about how the filters can be
taps filters for lossless compression. It is known that choused can be read in [9], [10].
images present features or details along many
II. COMPRESSION SYSTEM
directions. Related to this property, DWT emphasizes
the details only after three directions. This fact makes
DWT to have poor directionality. To eliminate this,
the researchers started to develop new transforms with
The new transform, DTCWT, detects the details

over more than three directions comparing with DWT,
1
Facultatea de Electronic i Telecomunicaii, Centru de Cercetari pentru
Prelucrarea si Securizarea Datelor, Calea Dorobantilor 71-73, Cluj Napoca
e-mail Iulian.Voicu@com.utcluj.ro
2
Facultatea de Electronic i Telecomunicaii, Centru de Cercetari pentru
Prelucrarea si Securizarea Datelor, Calea Dorobantilor 71-73, Cluj Napoca
e-mail Monica.Borda@com.utcluj.ro
18
Figure 2. Bloc scheme of compression system

which is an important advantage for compression
applications.
The process of coding and decoding an image is
showed in the figure 2. For the beginning images
suffer a decomposition process into R, G, B plans.
Then, DWT or DTCWT are applied to every color
plane while for coding we used (EZW) algorithm,
proposed by Shapiro [12] specifically to code the
wavelet coefficients. It is well known the fact that
after the discrete wavelet decomposition is
performed, for bi-dimensional signals, four
subbands are resulting (LL, LH, HL, HH) at every
level of decomposition.
For every coefficient from a low subband
corresponds a set of four coefficients in the next
higher subband, all those characterizing the same
area of the image but at a different scale. These
coefficients can form a zero-tree [12] if the root
and all its descendants are insignificant related to a
threshold. A zero-tree is labelled on only two bits
in EZW algorithm thus a large number of
coefficients being coded for high values of the
threshold. After the image is decomposed in
subbands, it is necessary to scan the coefficients in
a predetermined way to reduce the information
needed to the decoder. There are two methods for
doing this: raster and Morton scan, presented in
figure 3.
Although the quality of final
reconstructed image depends on the way that
coefficients are scanned, in our study we used only
the Morton scanning method. Some problems
appeared when we tried to apply EZW algorithm to
subbands
resulted
from
the
DTCWT
decomposition. This is because we have no order
relationship between complex coefficients and real
numbers that are the thresholds.
Real subbands can be made from complex
coefficients of DTCWT. This is based on the way
that complex subbands are obtained [6], [7].

In [13] a detailed 2D decomposition scheme is
described. Thus from complex subbands oriented at
15 0 a real subband can arise combining these
coefficients. The algorithm is presented in figure 4.
Complex subband 1 refers to the + 15 0 while
Complex suuband 2 to the 15 0 . The same is
with the other subbands oriented at 45 0 or
75 0 . The process is repeated at every level of
decomposition and ensures that three real subbands
arise from the six complex subbands. After the
transformation from complex to equivalent real
subbands is complete, we can apply EZW
algorithm similar with the case of DWT.
Combining complex subbands in this way, the
property of ascendant and descendent coefficients
is preserved. Next operation is to control the budget
of bits used to code the wavelet coefficients, an
increased number of bits meaning that almost or all
the coefficients have been coded while less bits
used lead to sending only a part of them. Having
the received coefficients, IDWT or IDTCWT is
applied depending on the transform used in the
coding part. In this way we get the reconstructed
image.
Figure 3: Scanning methods: A) Morton; B)

raster
19
Figure 4. Illustration of the method in order to obtain a real subband

important and undoubtedly advantage of DTCWT
over the DWT in compression.
Another important reason that leads to our
studies and make us trying DTCWT in
compression is that this transform retains the
details over the six directions. Compared with the
three directionality of DWT, the new transform
offers more important information.
Knowing that in an image the details are oriented
at different degrees we think that preserving details
over the six directions leads to an improved image.
Now that the motivations for applying a
redundant transform in compression applications
were presented lets see the results obtained in our
study.
We have done many tests and simulations of
our compression system presented in figure 2. We
thought that before replacing DWT with DTCWT
in JPEG 2000, it is better to simulate this easier
system, having in this way an idea of what is being
to happen.
The system was tested for a large variety of
standard test images in two situations: with DWT or
with DTCWT in order to compare these two
processings. Our principal goal was to see how the
system works at different bit rates. Second goal was
to see if the different filters used in both DWT and
DTCWT influence the quality of processed image.
We started our experiments with low bit rates
Note that the quality of reconstructed image

will be affected by the bit budget scheduled in our
experiments. As this budget is increased, more
coefficients will arrive at the decoder and
reconstructed image will approximate better the
original.
III. EXPERIMENTAL RESULTS
Some of our experimental results regarding
compression of colored images with complex
wavelet transform are presented next.
In our previous paper [11] was shown that using
complex wavelet transform in compression
applications could lead to satisfactory results for
some bit rates.
DTCWT transform is a redundant transform
[6], which is an undesired property for compression
applications where redundancy is not wanted and at
a first look most researchers dropped out their
studies of using redundant transforms in
compression field. Indeed, redundancy is a
disadvantage against classical transform DWT.
On the other hand a redundant transform
provides more coefficients non-zero compared with
a non-redundant transform (DWT). Also, in
compression applications, it is desired to have a lot
of coefficients that are non-zero instead of having
zero coefficients. This fact can be viewed as an
20
As it can be seen in figures 5 and 6, first two

images in each figure, at bit rates sized between
0.5bpp and 1bpp, DWT perform better than
DTCWT.
Here the differences and the qualities of

reconstructed images are not good comparing with
the original, no matter how transform is used, but
especially when we used DTCWT.
Figure 5. Baboon: from left to right: first row: original image; image compressed using DWT, at: 0.5bpp; 1bpp;
1.5bpp second row: original image; image compressed using DTCWT, at: 0.5bpp; 1bpp; 1.5bpp
Figure 6. Peppers: from left to right: first row: original image; image compressed using DWT, at: 0.5bpp; 1bpp;
1.5bpp second row: original image; image compressed using DTCWT, at: 0.5bpp; 1bpp; 1.5bpp
procedure. As future works we will focus on
improving even these results by considering this
procedure. Figures 7 and 8 presents for the same
baboon and peppers images the R, G, B color
planes of compressed images at this bit rate.
As we have written before, the second goal was
to verify if different pair of filters influences the final
results.
In this direction we tested 5/3 and 9/7 pairs of
linear phase filters for the DWT, while for the
DTCWT we used in the first level the same 5/3 and
9/7 and for the following levels q-shift filters
designed by Kingsbury [8].
We found that the differences are not
significant, for every option finding the same
critical bit rate, or slightly around, where DTCWT
start to perform better than DWT.
This is an expected result and it is explained by

the redundancy of DTCWT. This situation is
encountered around of a 1.3 - 1.5bpp values of bit
rate.
Starting with 1.5bpp value of the bit rate, the
images processed with DWT or DTCWT are
visually very closed. For the baboon image in
figure 5 the differences are almost invisible. In case
of peppers the final result with DTCWT is more
blurred than image compressed with DWT.
The study showed us that around this value
colored images decomposed with DTCWT are
visually slightly worst than those decomposed with
DWT but at higher bit rates indicates no significant
differences when it is used DWT or DTCWT
decomposition. It is worth to mention that we
havent take into consideration bit allocation
21
Unfortunately, although DTCWT delivers

details along six directions, at lower bit rates the
redundancy property affects the performances of

the compressed images.
Figure 7. Baboon: from left to right: first row: original image and original R, G, B color planes. Second row:
compressed image and compressed R, G, B color planes at 1.5bpp using DWT; Third row: compressed image
and compressed R, G, B color planes at 1.5bpp using DTCWT
Figure 8. Peppers: from left to right: first row: original image and original R, G, B color planes; second row:
compressed image and compressed R, G, B color planes at 1.5bpp using DWT; third row: compressed image and
compressed R, G, B color planes at 1.5bpp using DTCWT
22
IV. CONCLUSIONS
We would like to thank professors A. Isar and I.

Nafornita (Politehnica University of Timisoara)
for the fruitful discussions that we had during the
writing of this paper.
The paper presents a study of applying complex

wavelet transform in compression applications. For
this, was simulated a compression system that used
EZW algorithm to compress wavelet coefficients.
In case of complex wavelet transform that supply
complex coefficients there is a need of a preprocessing stage where coefficients are mixed for
getting real numbers. Combining complex
coefficients in the way showed in figure 4, the
property of ascendant and descendent coefficients
is preserved.
Our study focused on finding the point when we
can use DTCWT instead of DWT in decomposition
stage, in image compression application. For this
we found out that there is an interval around 1.5bpp
where the quality of reconstructed images is very
bosom no matter what transform is used in
decomposition step. Below this value we cannot
apply
DTCWT
successfully
because
of
redundancy, which seems to restrict the useful of it
only at higher bit rates. We can say that the
redundancy covers the advantage offered by
DTCWT of those six oriented subbands.
Also, these partially satisfactory results are
obtained without taking into account bit allocation
procedure that could improve the quality of
compressed image.
Another point in our study was to see if different
pairs of filters that decompose the signal modify the
final results. In our tests we tried 5/3 and 9/7 taps
linear phase filters for DWT, while for DTCWT we
used the same linear filters for level one of
decomposition and q-shift filters for higher levels.
We found that the differences are not significant,
for every option finding the same critical bit rate,
or slightly around.
REFERENCES
[1] G. Strang, T.Q. Nguyen, Wavelets and Filter Banks,
Wesley-Cambridge Press, 1996
[2] P. Vaidyanathan, "Multirate Signal Preocessing, Prentice
Hall, 1993
[3] K. Varma, Amy Bell, JPEG2000 Choices and Tradeoffs
for Encoders, IEEE Processing Magazine, Nov. 2004
[4] A. Skodras, C. Christopoulos and T. Ebrahimi, The
JPEG2000 Still Image Compression Standard, IEEE
Processing Magazine, September 2001
[5] Bryan E. Usevitch, A Tutorial on Modern Lossy Wavelet
Image Compression: Foundations of JPEG2000, IEEE Signal
Processing Magazine, September 2001
[6] N.G. Kingsbury, Image Processing with Complex
Wavelets, Phil. Trans. Royal Society London A, no. 357, pp.
2543-2560, Sept. 1999
[7] N.G. Kingsbury, Complex Wavelets for Shift Invariant
Analysis and Filtering of Signals, Journal of Applied and
Computational Harmonic Analysis, May 2001
[8] N.G. Kingsbury, Design of Q-Shift Complex Wavelets for
Image Processing Using Frequency Domain Energy
Minimization, Proc. IEEE Conf. on Image Processing,
Barcelona, Sept 15-17, 2003
[9] I.W. Selesnick, The Design of Approximate Hilbert
Transform Pairs of Wavelet Bases, IEEE Transactions on Signal
Processing, vol. 50, no. 5, May 2002
[10] I.Voicu, M.Borda, New Method of Filters Design for Dual
Tree Complex Wavelet Transform, Proceedings ISSCS, Iasi,
Romania, vol. 2, July 2005
[11] I.Voicu, M.Borda, Complex Wavelet Transform in
Compression Field, Proceedings EUROCON, Belgrade, Serbia &
Montenegro, November 2005
[12] J.M. Shapiro, Embedded imaging coding using zerotrees of
wavelet coefficients, IEEE Trans. Signal Processing, vol. 41, pp.
3445-3462, December 1993.
[13] J.Neumann, G. Steidl, Dual-Tree Wavelet Transform in the
Frequency Domain and an Application to Signal Classification,
Dept. of Mathematics and Computer Science, September, 2003
[14] A. Isar, I. Nafornita, Reprezentari timp-frecventa, Ed.
Politehnica, Timisoara 1998.
V. FUTURE WORK
Future work will be focused on finding reliable
methods for applying DTCWT in JPEG 2000
standard.
We think that complex wavelet transforms with
improved directionality can perform better than
classical discrete wavelet transform in compression
field.
Of course, redundancy is not good for compression
but we hope to find a link between the, in
coefficients of both trees of DTCWT this case
coding only a part of the coefficients being an
acceptable solution. In this way the redundancy can
be removed and we would exploit only the
advantage of improved directionality.
VI. ACKNOWLEDGEMENT
The first author was supported by a grant from the
National Council of Scientific Research and
Education (CNCSIS), Romania, code 245 BD.
23

Modern Network Algorithms

Miranda Nafornita1, Florin Chis2
IP addresses among multiple computers, using
schemes such a network address translation, or NAT.
Thus CIDR and NAT have helped the Internet
handle exponential growth with a finite 32-bit address
space. While there are plans for a new IP (IPv6) with
a 128-bit address, the effectiveness of NAT in the
short run and the complexity of rolling out a new
protocol have slowed down the IPv6 deployment.
Despite this, a new world containing billions of
wireless sensors may lead to an IPv6 resurgence.
Abstract In the future, the routing will be based not

only on destination host numbers but also on source host
or even source users, as well as destination URLs
(Universal Resource Locators) and specific business
policies. Modern routers and switches will have to use
class of services and QoS to determine the path to
particular applications for particular end users. All this
requires the use of layers 4, 5 and above.
Keywords: routing, prefix lookup, packet classification
I. INTRODUCTION
II. PREFIX LOOKUP
Most users deal with domain names, but these

names are translated to an IP address by a directory
service called DNS before packets are sent. Internet
prefixes are defined using bits and not alphanumerical
characters, of up to 32 bits in length. Internet began
with a simple hierarchy in which 32-bit addresses
were divided into a network address and a host
number; routers only stored entries for the networks.
For flexible address allocation, the network address
came in variable sizes: Class A (8 bits), Class B (16
bits), and Class C (24 bits).
IP prefixes are often written in dot-decimal
notation; so the notation is equivalent to a binary
string. To use more efficiently address space, Internet
prefixes are variable length so the second way to
denote a prefix is by slash notation A/L (A denotes a
32 IP address in dot-decimal notation and L denotes
the length of prefixes). Shorter prefixes can be
assigned to areas with a large number of endpoints
and larger prefixes to those with a few endpoints. A
third common way to describe prefixes is to use a
mask in place of an explicit prefix length. The last two
ways are more compact for writing down large
prefixes [10].
To cope with exhaustion of Class B addresses,
the Classless Internet Domain Routing (CIDR)
scheme assigns for new organizations multiple
contiguous Class C addresses that can be aggregated
by a common prefix. This reduces core router table
size. The potential depletion of address space has led
Internet registries to be very conservative in the
assignment of IP addresses. Many organizations are
coping with these sparse assignments by sharing a few
The decision to deploy CIDR helped save the

Internet, but it has introduced the complexity of
longest-matching-prefix lookup. Three kinds of
matches are allowed: exact match, prefix match and
range match. Exact-match lookups: the header field of
the packet should exactly match the field and they
represent the simplest form of database query. A
query specifies a key K and the response returns the
state information associated to this key. Exact-match
queries are easily implemented using well-known
techniques, such as binary search and hash tables. But
in the networking context, the models and metrics for
lookups are different. The lookups must be completed
in the time to receive a packet, the use of memory
references rather than processing as measure of speed,
and the potential use of hardware speedups. Exactmatch lookups are crucial for a very important
networking function called bridging, often integrated
within routers [8].
Prefixmatch: the field should be a prefix of the
header field. A packet arriving on an input link of a
router carries a 32-bit Internet (IP) address. The
processor consults a forwarding table to determine the
output link for the packet, which contains a set of
prefixes with their corresponding output link. The
packet is matched to the longest prefix that matches
the destination address in the packet, and the packet is
forwarded to the corresponding output link. The task
of determining the output link is called address
lookup, and must be implemented at very high speeds
(until gigabit or terabit per second).
In a range match, the header values should lie in
the specified range [8].
1
Politehnica University of Timisoara, Communications dept.
Bd. V. Parvan no. 2, 300223 Timisoara, e-mail miranda.nafornita@etc.upt.ro
2
Politehnica University of Timisoara, Communications dept.
24
was to require ordinary IP lookup at domain entry

points. Tag switching, developed today in a more
general form, is multiprotocol label switching, MPLS.
Neither tag switching nor MPLS completely avoid the
need for ordinary IP lookups and each adds a
complexity.
Flow switching is a second proposal to finesse
lookup and also relies on a previous hop router to pass
an index into the next hop router. Unlike tag
switching, these indexes are computed on demand
when data arrives, and then are cached. Eventually, IP
forwarding can be completely avoided in the switched
portion of a sequence of flow-switched routers. Flowswitching seems likely to work poorly in the
backbone, because backbone flows are short lived and
exhibit poor locality. On the other hand the current
use of circuit-switched optical switches to link core
routers, the underutilization of backbone links and
increase of optical bandwidth make possible the
resurrection of flow switching based on TCP
connections.
The most important lookup algorithms are as
follows. If memory is not an issue, the fastest scheme
is one called recursive flow classification (RFC). If
memory is an issue, a simple scheme that works well
for classifiers up to around 5000 rules is the Lucent
bit vector scheme. For larger classifiers, the best
trade-off between speed and memory is provided by
decision tree scheme. Unfortunately, all these
algorithms are based on heuristics and cannot
guarantee performance on all databases. If guaranteed
performance is required for more than two field
classifiers, there is no alternative but to consider
hardware schemes such as ternary CAM (content
addressable memory).
Traffic distributions, memory trends and database

sizes motivate the requirements for lookup schemes.
A study of backbone traffic shows over 250.000
concurrent flows of short duration and this number is
increasing, that means caching solutions do not work
well. Roughly half the packets received by a router
are minimum-size TCP acknowledgements. Hence,
the router must be able to perform prefix lookups in
the time to forward a minimum-size packet. Assuming
wire speed forwarding, forwarding a 40-byte packet
should take no more than 320 nsec at 1 Gbps, 32 nsec
at 10 Gbps, and 8 nsec at 40 Gbps.
Clearly, the most crucial metric for a lookup
scheme is lookup speed. Cost of computation is
dominated by memory access time, so the simplest
measure of lookup speed is the worst-case number of
memory access. The possible use of host routes (full
32-bit addresses) and multicast routes means that the
backbone routers will have prefix databases till 1
million
prefixes.
Unstable
routing-protocol
implementation can lead to requirements for updates
(add or delete a prefix) on the order of milliseconds, a
several orders of magnitude below the lookup
requirement, allowing to pre-compute information to
speed up lookup, at the cost of longer update times.
Memories can be cheap and slow (DRAM access
time : 60 nsec) or faster and expensive (SRAM off /
on-chip memory, 1-10 nsec). So, the interesting
metrics for lookup schemes are lookup speed,
memory, and update time [5].
The longest matching prefix is a very complex
problem. In virtual circuit networks, like ATM [13],
when a source wishes to send data to a destination, a
call is set up. The virtual circuit identifier VCI at each
router is a moderate-size integer that is easy to lookup
but the cost is a round-trip delay for call setup. The
ATM has a previous hop switch pass an index into
next hop switch, pre-computed just before data is send
by the previous hop switch. The same idea was used
in datagram networks such as Internet to finesse the
need for prefix lookup: tag switching and flow
switching.
Tag switching is a first proposal to finesse
lookup. In threaded indices each router passes an
index into the next routers forwarding table, thereby
avoiding prefix lookup. The indexes are pre-computed
by the routing protocol whenever the topology
changes. The main differences between threaded
indices and virtual circuit indices are: 1) threaded
indices are per destination and not per active sourcedestination pair as in virtual circuit networks such as
ATM; 2) threaded indexes are pre-computed by the
routing protocol whenever the topology changes.
Cisco later introduced tag switching, which is similar
in concept to threaded indices, except tag switching
also allows router to pass a stack of tags (indices) for
multiple routers downstream. Both schemes do not
deal well with hierarchies. The last backbone router
has only one aggregated routing entry for the entire
destination domain and can thus pass only one index
for all subnets in that domain. The adopted solution
III. PACKET CLASSIFICATION

Future routers and switches will have to use class
of services and QoS to determine the path to
particular applications for particular end users. All
this requires the use of layers 4, 5 and above. This
new vision of forwarding is called packet
classification or layer 4 switching, because routing
decision can be based on headers available at layer 4
or higher in the OSI architecture. Other fields a router
may need to examine include source address (to
forbid or provide different service to some source
networks), port fields (to discriminate between traffic
types), and even TCP flags (to distinguish between
externally and internally initiated connections).
Besides security and QoS, other functions that require
classification include network address translation
(NAT), metering, traffic shaping, policing, and
monitoring.
The packet classification problem is to determine
the lowest-cost matching rule for each incoming
message at a router. The information relevant to a
lookup is contained in K distinct header fields in each
message. The classifier, or rule database, consists of
a finite set of N rules, each rule being a combination
25
quality of service (QoS). Another major feature of

MPLS is the ability to place IP traffic on a defined
path through the network providing bandwidth
guarantees and differentiated service (DiffServ)
features for specific user application or flow [8,10,
11].
Since a VCI provides a simple label to quickly
distinguish a flow, a label allows a router to easily
isolate a flow for special service. MPLS uses labels to
finesse the need for packet classification, a much
more difficult problem than prefix lookup. MPLS is
used for the core router today, although prefix
matching is still required. IP networks with
connectionless operation, for traffic engineering
purposes, becomes
more connection-oriented
network, where the path between the source and the
destination is pre-calculated based on user specifics.
To speed up the forwarding schemes, an MPLS
device uses labels rather than address matching to
determine the next hop for a received packet. To
provide traffic engineering, tables are used, tables that
represent the levels of QoS the network can support.
The tables and the labels are used together to establish
an end-to-end path called a label switched path (LSP).
Traditional IP routing protocols (e.g.: OSPF-open
shortest path first, IS-IS intermediate system to
intermediate system), and extension to existing
signaling protocols (RSVP-resource reservation
protocol and CR-LDP-constraint-based routing label
distribution protocol) comprise the suite of MPLS
protocols.
MPLS is based on the following ideas: a)
Forwarding information (label) separate from the
content of IP header; b) A single forwarding
paradigm (label swapping), multiple routing
paradigms; c) multiple link-specific realizations of
the label swapping forwarding paradigm shim,
virtual
connection/path
identifier
(VCI/VPI),
frequency slot (wavelength), time-slot; d) The
flexibility to form forwarding equivalence classes
(FECs); e) A forwarding hierarchy via label stacking
[14].
The separation of forwarding information from
the content of the IP header allows MPLS to be used
with devices such as OXCs (optical cross-connect),
whose data plane cannot recognize the IP header.
LSRs (label switch routers) forward data using the
label carried by the data. This label, combined with
the port of the switch where data was received, is used
to determine the output port and outgoing label for the
data. The MPLS control plane operates in terms of
label swapping and forwarding paradigm abstraction.
At the same time, the MPLS control plane allows
multiple link-specific realizations of this abstraction.
For example [15], a wavelength could be viewed as an
implicit label. Finally, the concept of a forwarding
hierarchy via label stacking enables interaction with
devices that can support only a small label space. This
property of MPLS is essential in the context of OXCs
and
DWDM
(Dense Wavelength
Division
of K values, one of each header field. Each field in a

rule is allowed the three kinds of matches. Exact
match, where the header field of the packet should
exactly match the rule field, is useful for protocol and
flag fields. Prefix match, where the rule field should
be a prefix of the header field could be useful for
blocking access from a certain subnetwork. Range
match, where the header values should lie in the range
specified by the rule, can be useful for specifying port
number range.
Each rule has an associated directive, which
specifies how to forward the packet matching this
rule. The directive specifies if the packet should be
blocked or forwarded. If the packet should be
forwarded, the directive specifies the outgoing link to
which the packet is sent, and, perhaps, also a queue
within that link if the message belongs to a flow with
bandwidth guarantees. A packet is said to match a
rule, if each field of a packet matches the
corresponding field of the rule. The match type is
implicit in the specification of the field. Since a
packet may match multiple rules in the database, each
rule in the database is associated with a nonnegative
number, the cost. Ambiguity is avoided by returning
the least-cost rule matching the packets header. The
cost function generalizes the implicit precedence rules
that are used in practice to choose between multiple
matching rules.
Several variants of packet classification have
already established on Internet. First, many routers
implement firewalls at trust boundaries, such as the
entry and exit points of a corporate network. A
firewall database consists of a series of packet rules
that implement security policies. Rules are placed in
the database in a specific linear order, where each rule
takes precedence over a subsequent rule. Thus, the
goal there is to find the first matching rule. The same
effect can be achieved by making cost equal to the
position of rule in database. A general firewall could
arbitrarily interleave rules that allow packets with
rules that drop packets. Second, the need for
predictable and guaranteed service has led to
proposals for reservation protocols, such as DiffServ,
that reserve bandwidth between source and
destination. Third, routing based on traffic type has
become more stringent recently.
Packet classification unifies the forwarding
function required by firewalls, resource reservation,
QoS routing, unicast routing and multicast routing. In
classification, the forwarding database of a router
consists of a potentially large number of rules on key
header fields. A given packet header can match
multiple rules. Each rule has a given cost and the
packets are forwarded using the least-cost matching
rule.
IV. MPLS - MULTIPROTOCOL LABEL
SWITCHING
Originally introduced to speed up lookups, MPLS
is now a mechanism for providing flow differentiation
26
originally introduced in IP networks to improve

forwarding performance by eliminating timeconsuming longest prefix matching, has shifted
toward improved resiliency, enhanced QoS, and
traffic engineering capabilities [1], [2], [3], [12].
GMPLS has been specifically designed to extend
capabilities offered by MPLS network elements that
have non-packet-based forwarding engines (from
packet and frame/cell switching technologies to
circuit
switching
technologies,
including
SONET/SDH). It encompasses the entire range from
packet-switching-capable devices up to fiberswitching-capable devices. The GMPLS architecture
specifies all the protocol capabilities in term of
signaling (RSVP-TE Resource-Reservation-ProtocolTraffic-Engineering), routing (OSPF-TE), and linkmanagement (LMP Link-Management-Protocol) [4].
The main idea of GMPLS is to extend the
original MPLS scheme that needs to recognize the
packet boundaries and extract the label information
before switching the packet, to a scheme that can
perform switching without depending on recognizing
the packet boundaries or the header information. Such
a scheme must depend on some other optical
properties (label mapping space) to find out the
forwarding class for a packet before switching it to
destination. The application of GMPLS can affect the
future of IP-over-WDM networks resulting in
reducing the number of layers, which can also reduce
the cost, complexity and processing overhead in
optical backbones. The current standards for GMPLS
define five label mapping spaces, namely: Packet
Switch Capable (PSC), Layer 2 Switch Capable
(L2SC), Time-slot Switch Capable (TSC),
Wavelength Switch Capable (LSC), and Fiber Switch
Capable (FSC). Only the last two layers (LSC and
FSC) can be utilized in all-optical switching device.
The remaining label mapping spaces require optical to
electrical conversion, which limits the device speed to
a level much less than achievable through an alloptical switching device.
IP-based backbone networks are gradually
moving towards a network model consisting of highspeed routers that are flexibly interconnected by
lightpaths set up by an optical transport network
consisting of WDM links and optical cross-connects.
Recovery mechanisms at both network layers will be
crucial to reach the high availability requirements of
critical services. GMPLS protocol suite can provide a
distributed control plane that can be used to deliver
rapid and dynamic circuit provisioning of end-to-end
optical lightpaths.
Multiplexing) since the number of wavelengths

(which acts as labels) is not very large.
MPLS fast path forwarding is as follows. A
packet with an MPLS header is identified, a 20-bit
label is extracted, and the label is looked up in a table
that maps the label to a forwarding rules. The
forwarding rule specifies a next hop and also specifies
the operations to be performed on the current set of
labels in the MPLS packet. These operations can
include removing or adding labels. Router MPLS
implementations have to impose some limits to
guarantee wire speed forwarding. The label space may
be dense, supporting a smaller number of labels than
220 (this allows a smaller amount of lookup memory,
avoiding a hash table) and limiting the number of
label-stacking operations that can be performed on a
single packet.
The MPLS framework includes significant
applications such a constraint-based routing.
Constraint-based routing is a combination of
extensions to existing IP link-state routing (e.g., OSPF
and IS-IS) with RSVP or CR-LDP like the MPLS
control plane, and the Constrained Shortest-Path-First
(CSPF) heuristic. The extensions to OSPF and IS-IS
allow nodes to exchange information about network
topology, resource availability and even policy
information. This information is used by the CSPF
heuristic to compute paths subject to specified
resource and/or policy constraints. For example, either
RSVP-TE (TE Traffic Engineering) or CR-LDP is
used to establish the label forwarding state along the
routes computed by CSPF-based algorithm; this
creates the LSP. The MPLS data plane is used to
forward the data along the established LSPs.
Constraint-based routing is used today for two main
purposes: traffic engineering and fast reroute. With
suitable network design, the constraint-based routing
of IP/MPLS can replace ATM as the mechanism for
traffic engineering. Fast reroute offers an alternative
to SONET as a mechanism for protection/restoration.
V. GENERALIZED MPLS (GMPLS)
GMPLS extends MPLS to provide the control
plane (signaling and routing) for devices that switch
in any of these domains: packet, time, wavelength and
fiber [7]. This common control plane promises to
simplify network operation and management by
automating end-to-end provisioning of connections,
managing network resources and providing the level
of QoS that is expected in the new applications.
GMPLS is promoted as a major technology for the
automation of network operations [3].
Because of the large installed base of
SONET/SDH network infrastructure and circuitoriented TDM-based networks there is a need for
efficient interworking between traditional circuitoriented networks and IP/MPLS networks. The most
promising technology able to meet these requirements
is the GMPLS protocol suite. GMPLS extends the
label switching introduced in MPLS. The latter,
VI. REMARKS
In recent years the main focus of transport
network evolution has been on increasing transport
capacities and introducing data networking
technologies and interfaces (e.g. Gigabit Ethernet).
This evolution is complemented by outgoing
initiatives to reduce the operational effort and
27
GMPLS Provider Operational Expenditures: A Quantitative Study,

IEEE Comm. Magazine, July, 2005, pp. 28-34.
[5] A. Bianco, M. Atiquzzaman, G.S. Kuo, Next Generation
Switching and Routing, IEEE Comm. Magazine, Jan 2005, pp. 8687.
[6] Bart Puype et al., Benefits of GMPLS for Multilayer
Recovery, IEEE Comm. Magazine, July, 2005, pp. 51-59.
[7] T. Khattab, H. Alnuweiri, Optical GMPLS Networks with
Code Switch Capable Layer for Sub-Wavelength Switching, IEEE
2004, {tkhattab, hussein}@ece.ubc.ca.
[8] Generalized Multiprotocol Label Switching, GMPLS,
http://www.iec.org.
[9] G. Varghese, Network Algorithmics, Morgan Kaufman
Publishers, 2005.
[10] T.Li, MPLS an Evolving Internet Architecture, IEEE
Comm. Magazine, vol.14, no 2, March/April, 2002.
[11] T.M.Chen, Reliable Services in MPLS, IEEE Comm.
Magazine, vol.37, no 12, Dec, 2002.
[12] X. Xiao, A. Hannan, et al., Traffic Engineering with MPLS
in the Internet, IEEE Comm. Magazine, vol.14, no 2, March/April,
2002.
[13] M. Bocci, J. Guillet, Operation, Administration and
Maintenance in MPLS Networks, IEEE Comm.Magazine, vol.42,
no 10, Oct, 2004.
[14] D. Cavendish, H. Ohta, H. Rakotorano, Framework for
MPLS-based Control of Optical SDH/SONET Networks, IEEE
Comm. Magazine, vol.14, no 2, March/April,2002.
[15] G. Bernstein, E. Mannie, V. Sharma, MPLS an Evolving
Internet Architecture, IEEE Comm. Magazine, vol.15, no 4,
July/Aug, 2001.
accordingly the costs of network operations. GMPLS

together with standardized interfaces like usernetwork and network-network interfaces (UNI/NNI)
are automating the operation of telecom networks.
They allow services to be provided efficiently and
improve the resilience of networks. For service
provisioning (or switched connections), the new
approach is that the connections are being set up by
the client without operator interaction. This speeds up
the provisioning process and reduces effort for the
network operator.
REFERENCES
[1] J.T. Park, Resilience in GMPLS Path management: Model
and Mechanism, IEEE Comm. Magazine, July, 2004, pp. 128-135.
[2] D. Papadimitriou, D.Verchere, GMPLS-Network Interface in
Support of End-to-End Rerouting, IEEE Comm. Magazine, July,
2005, pp. 35-43.
[3] A. Banerjee, J. Drake, J.P. Lang, B. Turner, K. Kompelia, Y.
Rekhter, Generalized Multiprotocol Label Switching: An
Overview of Routing and Management Enhancements, IEEE
Comm. Magazine, Jan, 2001, pp. 144-150.
[4] S. Pasqualini, A. Kirstader,
A. Iselt, R. Chahine,
S.Vervrugge, D.Colle, M.Pickavet, P.Demeester, Influence of
28

Hybridization modelling in estimation of oligonucleotide

microarray expression
Raul Mluan1,2 Monica Borda1 Pedro Gmez Vilda2 Francisco Daz2
interaction are still fields of important research and is
effects are not taken into account in the estimation of
genetic expression by the algorithms currently in use.
The present work will present the technical factors
that affect expression measurement and are related
with hybridization process. Base-pairs interactions,
cross-hybridization, kinetic and thermodynamic
process are most significant problems relating with
microarray assay.
Abstract Oligonucleotide microarray technology has

become widely spread in genetic probing for different
purposes. This technology has shown to be highly
reliable and dependable, although not free from
problems regarding expression estimation, as it relies on
the differential or contrasted hybridization of test
probes. In many cases hybridization results do not show
the expected behavior, thus affecting to the reliability of
expression estimation. In the present paper a modeling
of the hybridization process is proposed which can be
used to predict and correct expression levels on a
specific test probe, improving the reliability of the
results. Some examples from real microarray processing
are shown and discussed.
Keywords: oligonucleotide, hybridization, dynamics
II. NUCLEIC ACIDS

A. Deoxyribonucleic acid structure
Deoxyribonucleic acid or DNA can be defined as an
antiparallel double helix of nucleotides [8], having
deoxyribose as their sugars, linked by sugarphosphate bonds to adjacent nucleotides in the same
chain and by hydrogen bonds to complementary
nucleotides in the opposite chain. Also it represents
the fundamental substance of which genes are
composed. DNA is a nucleic acid that contains the
genetic instructions monitoring the biological
development of all cellular forms of life, and many
viruses. The main role of DNA is to produce protein.
DNA it is organized as two complementary strands
[12], head-to-toe, with the hydrogen bonds between
them. Each strand of DNA is a chain of chemical
building blocks, called nucleotides, of which there
are four types: adenine (A), cytosine (C), guanine (G)
and thymine (T). Between the two strands, each base
can only bond with one single predetermined other
base: A with T, T with A, C with G, and G with C, are
the only possible combinations.
Besides DNA there is another nucleic acid called
Ribonucleic acid (RNA), also with four bases, but
instead of thymine it uses a molecule called uracil
(U). Just like T pairs with A, U will pair with A by
hydrogen bonding. RNA is used by the cell as
temporary copies of portions of DNA. Unlike DNA,
RNA is almost always a single-stranded molecule and
has a much shorter chain of nucleotides.
I. INTRODUCTION
DNA microarrays make the use of hybridization
properties of nucleic acids to monitor DNA or RNA
abundance on a genomic scale in different types of
cells. Hybridization process take place between
surface-bound DNA sequences, the probes, and the
DNA or RNA sequences in solution, the targets.
Hybridization is the process of combining
complementary, single-stranded nucleic acids into a
single molecule. Nucleotides will bind to their
complement under normal conditions, so two
perfectly complementary strands will bind to each
other readily. Conversely, due to the different
geometries of the nucleotides, a single inconsistency
between the two strands will prevent them from
binding.
In oligonucleotide microarrays hundreds of
thousands of oligonucleotides are synthesized in situ
by means of photochemical reaction and mask
technology. Probe design in these microarrays is
based upon complementarity to the selected gene. An
important component in designing an oligonucleotide
array is ensuring that each probe binds to its target
with high specificity.
The dynamics of the hybridization process
underlying genomic expression is complex as
thermodynamic factors influencing molecular
1
Facultatea de Electronic, Telecomunicaii si Tehnologia Informaiei, Departamentul

Comunicaii Str. G. Bariiu Nr. 26-28 Cam. 364, 400027 Cluj-Napoca, e-mail raul.malutan@com.utcluj.ro
2
Facultad de Informtica, Departamento de Arquitectura y Tecnologa de Sistemas Informticos,
Campus de Montegancedo s/n, 28660, Boadilla del Monte, Madrid, Spain
29
Hybridization process has been studied from point

of view of interaction between base pairs, the
interaction with unintended targets and also from its
kinetics processes. Because in practice the DNA chips
are immersed in the target solution for a relatively
short time, the attainment of equilibrium is not
guaranteed. Yet full analysis of the reaction kinetics
requires knowledge of the equilibrium state. An
understanding of the equilibrium state is also
necessary to identify the relative importance of kinetic
controls of the performance of the DNA microarrays.
The effect of the cross-hybridization on probe
intensity is predictable in the oligonucleotide
microarrays, and models for avoiding this have been
developed [1], [9], [10], and they will be described
further in the paper. Last section of this work will
describe a method of modeling the hybridization
process based on existing models and unexploited
parameters.
An important type of RNA is messenger RNA

(mRNA) that encodes and carries information from
DNA during transcription to sites of protein synthesis
to undergo translation in order to yield a gene product.
DNA can be synthesized from a mature mRNA
template, and in this case the obtained product it will
be a complementary DNA (cDNA). These cDNA
sequences are most important components involved in
the microarray hybridization process.
B. DNA hybridization
Hybridization refers to the annealing of two nucleic
acid strands following the base pairing rule. At high
temperatures approximately 90C to 100C the
complementary strands of DNA separate, denature,
yielding single-stranded molecules. Two single
strands under appropriate conditions of time and
temperature e.g. 65C, will renature to form double
stranded molecule. Nucleic acids hybrids can be
formed between two strands of DNA, two strands of
RNA or one strand of DNA and one of RNA. Nucleic
acids hybridization is useful in detecting DNA or
RNA sequences that are complementary to any
isolated nucleic acid.
Finding the location of a gene or gene product by
adding specific radioactive or chemically tagged
probes for the gene and detecting the location of the
radioactivity or chemical on the chromosome or in the
cell after hybridization is called in situ hybridization.
In the micrroarray technology [14], hybridization is
used in comparing mRNA abundance in two samples,
or in one sample and a control. RNA from the sample
and control are extracted and labeled with two
different fluorescent labels, e.g. a red dye for the RNA
from the sample population and green dye for that
from the control population. Both extracts are washed
over the microarray and gene sequences from the
extracts hybridized to their complementary singlestrand DNA molecule previously attached to the
microarray. Then, to measure the abundance of the
hybridized RNA, the array is excited by a laser.
In the oligonucleotide microarrays [13], the
hybridization process occurs in the same way, the
only difference here is that the sequences to be laid
over the chip are sequences of 25 nucleotides length,
perfect complementary to same length sequence of the
gene, PM perfect match, and sequences of same
length, designed to correspond to PM, but having the
middle base, 13th one, changed by its complementary
base, MM mismatch. The MM probes give some
estimates of the random hybridization and cross
hybridization signals. One thing to be followed in the
design of oligonucleotide arrays is ensuring that the
probes bind to their target with high accuracy. When
the two strands are completely complementary they
will bind by a specific hybridization, on contrary if
there are mismatches between the nucleotides of the
strands and they bind a process called non-specific
hybridization or cross-hybridization occurs [7].
III. TECHNICAL FACTORS AFFECTING GENE

EXPRESSION MEASUREMENTS
A. Interaction between pairs
The nucleic acids duplex stability can be endanger
by the interaction between the nucleotide bases.
Thermodynamics for double helix formation of
DNA/DNA, RNA/RNA or DNA/RNA can be
estimated with nearest neighbor parameters. Enthalpy
change, H, entropy change, S, free energy
change, G, and melting temperature, Tm, were
obtained on the basis of the nearest-neighbor model.
The nearest-neighbor model for nucleic acids assumes
that the stability of a given base pair depends on the
identity and orientation of neighboring base pairs [5].
In the nearest-neighbor model, sequence dependent
stability is considered in terms of nearest-neighbor
doublets. In duplex DNA there are 10 such unique
internal nearest-neighbor doublets. Listed in the 5-3
direction, these are AT/AT TA/TA AA/TT AC/GT
CA/TG TC/GA CT/AG CG/CG GC/GC GG/CC.
Dimmer duplexes are represented with a slash
separating strands in antiparallel orientation e.g.,
AC/TG means 5-AC-3 WatsonCrick base-paired
with 3-TG-5. The G37 can be computed from H
and S parameters using the equation:

G37o = H o TS o
(1)
As described in [6] the melting temperature Tm is

defined as the temperature at which half of the strands
are in double helical and half are in the random-coil
state. A random-coil state is a polymer conformation
where the monomer subunits are oriented randomly
while still being bonded to adjacent units.
For self-complementary oligonucleotides, the Tm for
individual melting curves was calculated from the
fitted parameters using the following equation:
30
Tm = H o / (S o + R ln CT )
(2)
Where R is the general gas constant, i.e. 1.987 cal/K
mol, the CT is the total strand concentration, and Tm is
given in K. For non-self-complementary molecules,
CT in (2) was replaced by CT/4.
The observed trend in nearest-neighbor stabilities at
37 C [5] is GC/CG = CG/GC > GG/CC > CA/GT =
GT/CA = GA/CT = CT/GA > AA/TT > AT/TA >
TA/AT. This trend suggests that both sequence and
base composition are important determinants of DNA
duplex stability. It has long been recognized that
DNA stability depends of the percent G-C content.
The fraction of probe sites occupied by probe-target

duplexes is then given by the differential equation:
d (t )
= k f x(1 (t )) kb (t )
dt
(3)
For initial condition (0) = 0 , (3) has the following

solution:
(t ) =
x
( x + K )k f t
1 e
x+K
(4)
where K = kb k f .
B. Interactions with unintended tags
Using (4) Burden et al. estimate the measured

fluorescence intensity I, with I0 as the background
intensity at zero concentration, to be:
As seen in previous sections the major issue in

microarray oligonucleotide technology is the selection
of probes sequences with high sensitivity and
specificity. Because the use of MM probes for
assessment of non-specific binding is unreliable [2], a
model based on previous mentioned nearest neighbor
model has been developed. The positional dependent
nearest neighbor model [9] has some modification to
the nearest neighbor model, first was to assign
different weigh factors at each nucleotide position on
a probe with the scope of reflecting that the binding
parts of a probe may contribute differently to the
stability of bindings, and secondly they took into
account two different modes of binding the probes:
gene specific binding, i.e. formation of DNA-RNA
duplexes with exact complementary sequences, and
non-specific binding, i.e., formation of duplexes with
many mismatches between the probe and the attached
RNA molecule.
The model appears to indicate that the two ends of a
probe contribute less to binding stability according to
their weight factors, and due the mismatch which
destabilizes the duplex structure a dip in the gene
specific binding weight factors of MM probes around
the mismatch position occurs.
I ( x, t ) = I 0 +
bx
( x + K ) k f t
1 e
x+K
(5)
At equilibrium, intensity I(x) at target concentration

x follows Langmuir Isotherm (Fig. 1):
I (x ) = I 0 +
bx
x+K
(6)
I0 + b
1
I0 + b
2
I(x)
I0
C. Dynamic adsorption model
Fig. 1.
The above model, together with the nearest neighbor

model solves the problem of binding on microarrays,
but still there are factors that affect the gene
expression measuring. One of them affects the process
of competing adsorption and desorption of target
RNA to from probe-target duplexes at the chip
surface.
Burden et al. [1] develop a dynamic adsorption
model based on Langmuir isotherm. If x is the
concentration of mRNA target and (t) is the fraction
of sites occupied by probe-target duplex, then in the
forward absorption, target mRNA attaches to probe at
a rate k f x(1 (t )) , proportional to concentration of
Hyperbolic response function for the intensity I(x)

according to the Langmuir isotherm
IV. HYBRIDIZATION DYNAMICS

COMPENSATION
A. Modeling hybridization by thermodynamics
It is well known that hybridization processes may be
seen under the point of view of general
thermodynamic conditions [4], meaning that the
hybridization probability of a given test segment will
be defined by its thermodynamic conditions, i.e. by its
hybridization temperature. Regarding this, one can
state that hybridization process will respond to the
dynamic equation.
specific target mRNA and fraction (1 (t )) of

unoccupied probes; and in the backward desroption
reaction, target mRNA detaches from probes at a rate
kb (t ) proportional to fractions of occupied probes.
Kf
P + T C
Kb
31
(7)
y = a(1 e bx )
where P represents the number of oligonucleotides

available for hybridization, T the concentration of free
RNA target, C the number of bound complexes, kf
and kb are the respective forward and backwards rate
constants for the reaction. This equation has as a
natural solution the following expression in the time
domain:
(9)
where a and b are parameters to be estimated

adaptively using least square fitting and gradient
method.
Vertical least square fitting proceeds by finding the
sum of squares of the vertical deviations R2 of
parameters a and b:
T
[1 exp( t )]
(8)
T +K
where K defined as in (4) is an equilibrium
1
denotes a
dissociation constant, and =
k f (T + K )
C (t ) =
[y a(1 e )]
R2 =
bxi
(10)
Further lets denote:
characteristic time over which the system reaches

equilibrium.
Recent studies [3], [15] confirm the hypothesis that
hybridization process for the each of the probe pairs
follows a time model according to the one from Fig.
2. This model of evolution predicts that the
probability of hybridization will be almost zero if not
enough time interval is provided for the experiment to
take place, and that in the limit, if enough time is
allowed saturation will take place.
i = yi a(1 e bxi )
(11)
With this notation (10) will become
R2 =
2
i
(12)
The condition of R2 to be a minimum is that:

(R 2 )
=0
a
(R 2 )
=0
b
(13)
From (12) and (13) we will get

(R 2 )
=
a
(R 2 )
=
b
Fig. 2. Theoretical model for perfect match hybridization.

Intensity of perfect match versus hybridization time. Figure adapted
from [3]
i
=
a
i i =
b
(1 e ) = 0
bxi
ax e
i
(14)
=0
bxi
A solution for equations in (14) can be found using

the gradient method. In this case the parameters are
going to be computed adaptively:
Practical solutions to the different hybridization

dynamics could include some of the following
procedures:
Estimate the co-linear and orthogonal
components in the PM-MM vectors, and use
the co-linear components in estimating the
expression.
Use Independent Component Analysis to
correct the PM and MM vectors [11].
Use multiple regression to convey PM-MM
probe pairs to equivalent thermodynamic
conditions
by
processing
diachronic
hybridization experiments.
The last procedure will be treated more detailed in the
following section of the present work.
ak +1 = ak a
bk +1 = bk b
(R 2 )
= ak + a
a
(R 2 )
= bk + b
b
(1 e )
(15)
(16)
bk xi
i ,k
i ,k
ak e bk xi
where i,k is defined as in (11) and is a parameter

used as an adjust step.
V. RESULTS AND DISCUSSION
Having this in mind a data base of hybridization
records have been created from experimental data
fitted by the before mentioned models. In Fig. 3 the
results of hybridization experiments for the same
probe test set obtained after two different time
intervals have been presented.
B. Exponential regression model

From (8) we assume that a model to solve the
multiple regression problem will have the following
form:
32
[8]
[9]
[10]
[11]
[12]
Fig. 3.
[13]
Time dynamics of hybridization
Fig. 4 shows the regression parameters obtained for

time constants. There were extracted the profile of the
perfect and mismatch for two different time values so
that it can be underline the fact that if enough time is
allowed to some probes, the mismatch ones can be
corrected
[14]
[15]
Fig. 4. Perfect and mismatch profiles for time constants. Top

template shows the profiles after 30 minutes, and bottom template
shows the profile after 100 minutes
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
C. Burden, Y.E. Pittelkow, S.R. Wilson, Statistical Analysis

of Adsorption Models for Oligonucleotide Microarrays,
Statistical Applications in Genetics and Molecular Biology,
vol. 3, no. 1, article 35, 2004
C. Li, W.H Wong, Model-based analysis of oligonucleotide
arrays: Expression index computation and outlier detection,
Proc. Natl. Acad. Sci. USA, vol. 98, no. 1, January 2001, pp.
31 36
H. Dai, M. Meyer, S. Stepaniants, M. Ziman, R. Stoughton,
Use of hybridization kinetics for differentiating specific
from non-specific binding to oligonucleotide microarrays,
Nucleic Acids Research, vol. 30, no. 16, 2002, pp. e86.1
e86.8
H. El Samad, M. Khammash, L. Petzold, D. Gillespie,
Stochastic Modelling of Gene Regulatory Networks, Int.
Journal of Robust and Nonlinear Control, vol. 15, issue 15,
2005, pp. 691 711
J. SantaLucia Jr, A unified view of polymer, dumbbell, and
oligonucleotide DNA nearest-neighbor thermodynamics ,
Proc. Natl. Acad. Sci. USA, Biochemistry, vol. 95, February
1998, pp. 1460 1465
J. SantaLucia Jr, H.T. Allawi, P.A. Seneviratne, Improved
Nearest-Neighbor Parameters for Predicting DNA Duplex
Stability, Biochemistry, vol. 35, no. 11, 1996, pp. 3555
3562
J.C. Huang, Q.D. Morris, T.R. Hughes, and B.J. Frey,
GenXHC: a probabilistic generative model for crosshybridization compensation in high-density genome-wide
microarray data, Bioinformatics, vol. 21, suppl. 1, 2005, pp.
i222 i231
33
J.D. Watson, F.H.C. Crick, A structure of DNA, Nature,

April 1953, pp. 737
L. Zhang, M.F. Miles, K.D. Aldape, A model of molecular
interactions on short oligonucleotide microarrays, Nature
Biotechnology, vol. 21, no. 7, July 2003, pp. 818 821
N. Sugimoto et al., Improved thermodynamic parameters
and helix initiation factor to predict stability of DNA
duplexes, Nucleic Acids Research, vol. 24, no. 22, 1996, pp.
4501 4505
P. Gmez et al., Robust Preprocessing of Gene Expression
Microarrays for Independent Component Analysis, to appear
in Proc. Independent Component Analysis and Blind Signal
Separation: 6th International Conference, 2006
P.P. Vaidyanathan, Genomics and Proteomics: A signal
Processors Tool, IEEE Circuits and System Magazine,
Fourth Quarter 2004, pp. 6 29
R.J. Lipshutz, S.P.A. Fodor, T.R. Gingeras, D.J. Lockhart,
High density synthetic oligonucleotide arrays, Nature
genetics supplement, vol. 21, January 1999, pp. 20 24
S.K. Moore, Making chips, IEEE Spectrum, March 2001,
pp. 54 60
Y. Zhang, D.A. Hammer, D.J. Graves, Competitive
Hybridization Kinetics Reveals Unexpected Behavior
Patterns, Biophysical Journal, vol. 89, 2005, pp. 2950
2959

Patents in the framework of TRIZ

Mircea Coer1
f1.1. principle of solution
f1.2. reduction rule
f2. effect(phenomena)
f2.1 natural
f2.2 technical
f3. heuristics
f3.1 engineering principle
f3.2 separation principle
f3.3 standard transformation
f3.4 prescript
f4. trend/path of evolution
f5. special cases
Using TRIZ for the study of patents one needs to be
conversant with the 8 patterns of evolution [2]
followed by every technical system or technological
process as it is shown in the table 2:
Abstract This paper uses a few patents for illustrating

some key concepts used in TRIZ. TRIZ can be defined
as a: "human-oriented knowledge-based systematic
methodology of inventive problem solving." [4].
Keywords: TRIZ, contradiction, path of evolution, SuF
analysis, inventive principle, separation principle,
effects.
I. BASICS
"Patents reveal solutions to technical problems and
they represent an inexhaustible source of information:
more than 80 percent of man's technical knowledge is
described in patent literature." - Statement of the
European Patent Office [4].
Studying patterns in the way Altshuller, the "father"
of TRIZ did, about fifty years ago when he started his
career as a promising young inventor, means using a
certain frame of work different from the usual one,
based on classification.
Analyzing around 400,000 patents, Altshuller
recognized that systems do not evolve randomly, but
along demonstrable principles called patterns of
evolution which principles are general in nature and
can be applied to any system or product. He also
made the observation that many inventive technical
problems from various fields of engineering are
solved by the same generic approaches.
Specialists in TRIZ had already analyzed almost 2
million patents from all around the globe, which
represent 10% of the total number of issued patents.
The table 1 illustrates the classification system used
by Altshuller [1] when classifying and studying
world-wide patents.
The results of the study of patents must be put in a
certain pattern abstract form in order to produce
uniform results, so they contain reusable information.
TRIZ specialists proposed [4] a Patent Abstract Form
which mainly contains the following items:
A. Legal information
B. Abstract of extracted knowledge
C. State of the art
D. Elements and functions of the prototype system
E. Resolved conflict
e1. type of contradiction
e2. structure of problem
F. Invented system
f1. contradiction
1
Table 1
Level
1.
Standard
solution
2. Change
of a
system
3. Solution
across
industries
4. Solution
across
sciences
5.
Discovery
System
changing
Trade-off,
quantitative
changes
Qualitative
change
Variants
number
A few
Used
knowledge
Someone
profession
Inventions
pool
32%
Tens
One
industry
45%
Radically
changed
system
Hundreds
Many
industries
19%
New system
created
Thousands
, tens of
thousands
Hundreds
of
thousands,
millions
Many
sciences
4%
New
knowledge
created
0.3%
New
discovery
A certain phraseology must be clarified in order to

better understand certain aspects on which TRIZ users
focus when studying technical inventive problems.
Some such terms are:
AC - administrative contradiction: something is
required but we don't know how to obtain the result
TC - technical contradiction: an action is
simultaneously useful and harmful
PC - physical contradiction: a system should have
both property A and non-A
ROMATSA RA, ARAD, e-mail:coser.m@excite.com
34
Table 2
Pattern
Technology
follows a life
cycle of birth,
growth,
maturity, and
decline.
Stage 1. A
system does not
yet exist, but
important
conditions for its
emergence are
being developed.
Stage 2. A new
system appears
due to high-level
invention, but
development is
slow.
Stage 3. Society
recognizes value
of the new
system.
Stage 4.
Resources for
original system
concept end.
Stage 5. Next
generation of
system emerges
to replace
original system.
Stage 6. Some
limited use of the
original system
may coexist with
the new system.
Increasing
Ideality.
Uneven
development
of subsystems
resulting in
contradictions.
Increasing
dynamism and
controllability.
Increasing
complexity,
followed by
simplicity
through
integration.
Example
1. Early
experiments like
"cat's whisker", the
crystal detector, etc.
Matching and
mismatching
of parts.
Early automobiles used leaf springs to

absorb vibration. These were an
assembly of unrelated or mismatched
parts borrowed from horse carriages and
whatever else was available.
Later fine tuning allowed adjustments of
the parts so that they mated into a
matched system - the shock absorber.
Purposely mismatch parts to create
additional resources from the differences.
An example of this might be using a
bimetal spring that changed spring rates
when a current is applied.
Automatic matching and mismatching as
needed. For example a computer
controlled active suspension system.
Transition
from macro
systems to
microsystems
using energy
fields to
achieve better
performance or
control.
Decreasing
human
involvement
with increasing
automation.
Development of cooking systems from

wood burning stove to electric and then
to microwave ovens.
2. Late 40s - W.
Sheckley, J.
Bardeen, W.
Brattain, inventors
of the point-contact
transistor.
3. Companies invest
in producing the
new device and
replacing the
vacuum tubes where
possible. Financial
resources available.
4. New components
based on the new
concept but with
new materials and
design; need for
more complex
function and smaller
device.
5. J. Kilby and R.
Noyce - inventors
of integrated circuit.
Development of clothes washing from

washboard to washing machine with
ringer, automatic washing machine then
automatic washing machine with
automatic bleach and softener dispensers.
Structure (topology) of a problem - obtained from the

representation of contradictions together with the
hierarchical structure; generic problem structures are:
point: have contradictions within a single
subsystem or have no contradictions at all.
pair: have a single technical contradiction
between functions in two subsystems.
linear: have chains of engineering
contradictions
network: have a loop of several dependent
contradictions
star: have a set of independent technical or
mathematical contradictions with a common
root (usually a physical contradiction)
6. Despite the
explosive
development of IC
industry, the
classical transistor
is still used in
certain electronic
circuits.
ENIAC computer in 1946 weighed

several tons, took a whole room, and did
computational functions. In modern
laptops weighing a few pounds capability
exist for text processing, mathematical
calculations, communications, graphics,
video, sound, etc.
Subsystems have different life cycle
curves. Primitive subsystems hold back
development of total system. A common
mistake is to focus on improving wrong
subsystem.
Early automobiles were controlled by
engine speed. Then manual gearbox,
followed by automatic transmissions, and
continuously variable transmissions
(CVT).
Stereo music systems have evolved from
adding separate components such as
speakers, AM/FM radio, cassette player,
CD player, etc. to integrated "boom box."
Inventive principles: the 40 principles revealed by G.

Altshuller based on the observation that technical
contradictions contained in a large number of
problems from different domains used the same
typical principle for removing them (he also identified
39 engineering parameters representing very general
features of TS and TP and placed them in a so- called
Matrix of Contradiction)
Separation Principles: a system of heuristics that
serves as a basis for solving physical contradictions;
four types of separations are used:
separation in space: possible when a
requirement exists in one place and is absent
in another
separation in time: possible when a
requirement exists at one period and is absent
at another time interval
35
conditions; it can affect quantitatively or qualitatively

the input parameter.
Large data bases with effects and phenomena were
built containing about 10,000 effects described in
scientific literature (compare with around 100 known
by an ordinary engineer). They are organized on a
functional principle: they contain a list of functions
commonly encountered in practice and a
corresponding list of effects that may be employed to
achieve these functions. [4]
separation upon condition: possible if a

requirement exists under one condition and is
absent under another
separation between parts and whole: possible
when a requirement exists at the key
subsystem level but does not exist at the subsystem, system or super system level
Standard Solutions: are based on a formalism called
SuF (Substance-Field) [3] analysis (Fig.1 - The
complete basic Su-Field model); from the two general
categories of inventive problems standard and nonstandard, the former can be solved using rules based
on Su-Field Models and on the Patterns of Evolution
of TS (which determine the modifications the system
has to undergo).
II. CASE STUDIES

We will present some patents collected from the
USPTO (United States Patent and Trademark Office)
collection [5] based on their area of application in the
field of the telecommunication system and underline
some aspects of them using TRIZ philosophy and
terminology.
Case 1:
Multi-frequency light source - USPA 20060193032,
August 31, 2006
The text says: "a multi-frequency light producing
method and apparatus multiplies the number of
optical channels present in an incident wavelength
division multiplexed (WDM) signal light source by
four-wave mixing (FWM) the WDM signal with at
least one pump lightwave at least one time." Further
on in text contradictions to be solved are set: "For a
system employing a large number of wavelengths,
using a different light source such as a laser diode to
generate each optical wavelength can be expensive
and consume a large amount of electric power." then
effects are used in solving contradictions: "Four-wave
mixing (FWM) is a phenomenon wherein three optical
lightwaves of different frequencies (wavelengths),
propagating in a nonlinear medium interact with one
another due to the nonlinear optical effect of the
conversion medium. This interaction generates
additional optical signal having different frequencies
from the three original signals." Using Physical and
Technical Effects helps inventors to obtain high level
even revolutionary solutions on the path to ideality
[2].
S2
S1
Fig. 1
Su-Field modeling helps focusing on the main system

elements and identifying the problem as one
belonging to a particular group. Deviations from
complete Su-Field triangles reflect the existence of
some problem in the group. Su-Field analysis presents
a general formula that shows the direction of solving
the problem. Three types of situations exist when
completing the model: the desired effect may not
occur; a harmful effect may occur; insufficient desired
effect occurs.
Standard Solutions: developed by G. Altshuller in
1985, consists of 76 Standard Solutions separated in
five classes.
Class 1: solutions for building and destroying
Su-Field Models.
Class 2: solutions directed toward improving
Su-Field Models
Class 3: rules for transitioning to either a
super-system or to the micro-level; they are
based on Pattern of Evolution.
Class 4: help solving the specific problems
related to measurement and detection.
Class 5: rules for applying the Standard
Solutions.
Effects and Phenomena: standard solutions include
methods but also physical effects (they do not refer to
a concrete physical effect but to the type of effect, and
for this reason they do not wear out; strong inventive
solutions may be obtained sometimes by using natural
effects that have rarely or never been used before in a
specific area; knowledge of less known effects in
physics, chemistry or geometry may be necessary)
Technical effects: transformation of energy or
substance from input to output under certain
Case 2:
Auxiliary power conservation for telecommunications
site - USPA 20060182262, August 17, 2006
The abstract says: "a method and/or system is
provided
for
managing
power
in
a
telecommunications site having an auxiliary power
supply that selectively powers the site when an outage
is experienced in connection with a primary power
supply for the site. The method includes: detecting a
power outage to the primary power supply;
determining the duration for the detected outage;
selecting a power conservation protocol based on the
determined duration of the outage; switching to the
auxiliary power supply and implementing the selected
power conservation protocol." Further on in the text
36
their potential corrosivity, whether at the time the

material is being fabricated, or while it is being
decomposed by fire." Trying to solve the folowing
aspect "ever increasing recourse is being made to fireretardant fillers that do not contain halogen, and in
particular to metal hydroxides" a new technical
contradiction is about to rise: " technical solutions of
that type present the drawback of requiring large
quantities of filler in order to achieve a satisfactory
level of effectiveness, whether in terms of ability to
retard flame propagation, or in terms of resistance to
fire." And here is the solution based on using Standard
Solutions: " the solution to the technical problem
posed consists in that the cable includes at least one
constituent part based on a fire-resistant material of
composition that comprises a synthetic polymer and a
fire-retardant filler, comprising an inorganic
compound associated with a hydrocarbon compound
selected from the group of asphaltenes, each
constituent part being selected from an insulating
covering and a protective sheath. The presence of the
inorganic compound within the fire-retardant filler
conventionally makes it possible to increase the fire
resistance of the material. To compensate for this
presence also having the drawback of strongly
degrading the mechanical properties of the
corresponding material, it is associated simply with a
compound based on malthene."
An extended study of a patent will follow a more
complex scheme like the one presented in Fig. 2.
the state of the art is detailed: " Previously, the

industry has used: multiple commercial power feeds, a
battery reserve, local power generation equipment
expensive to implement, particularly when the
demand for their deployment or use is relatively
infrequent." All these technical systems: "may not be
practical where the location prohibits the noise often
associated with running generators; the auxiliary
power supply has a limited capacity with respect to
the amount and/or duration of power it can supply to
its site; the degree to which power consumption can
be conserved during back-up emergency situations
largely determines how much back-up or emergency
power capacity a site should be provisioned with."
Here comes the reduction to a single technical
contradiction: "One option is to switch off all but the
very essential site functions when a site begins
operating on back-up or emergency power,
unfortunately: "limiting the level and/or quality of
services provided by the site; there is a desire to
maximize power conservation when operating a site
on auxiliary power, while still providing a suitable
level and/or quality of service."
Here is an example of the use of one of the 40
Inventive Principles - Principle of Extraction
(extracting the harmless functions from the system):"
after the primary power has been lost and a site
switches operations to auxiliary power, some site
functions are turned off or otherwise not supplied
power."less important or extraneous non-core
functions." Efforts are made to increase the dynamism
of the system: "To further extend the longevity of the
auxiliary power supply, additional site functions may
be shed as the reserve power supply is consumed or
depleted." so maintaining the natural evolution path.
III. CONCLUSIONS
The study of patents is of fundamental importance for
TRIZ and not only, because patents represent the
primary source of knowledge and the methodic study
of this fund can bring up new heuristics. Due to the
fact that this kind of study is very difficult and not
very productive (one new heuristic per 10,000 patents
[4]) new methods of study, perhaps based on artificial
intelligence are needed.
Case 3:
Fire-resistant power and/or telecommunication cable USPA 20060148939, July 6, 2006.
In the abstract it is stated that: "a cable capable of
withstanding extreme temperature conditions...remain
in operation during a defined length of time on being
subjected to high levels of heat and/or directly to
flames." The Administrative Contradiction is
established: "is essential to maximize the ability of a
cable to retard the propagation of flames, and also to
withstand fire." After stating that: "a cable is
constituted in general terms by at least one conductor
part extending inside at least one insulator part" the
physical contradiction appears: "amongst the best
insulating materials and/or protection materials used
in cable-making, many of them are unfortunately also
highly flammable."
The state of the art of the technique is presented
further on: " The technique that has been in the most
widespread use until now consists of implementing
halogen compounds in the form of a halogen byproduct dispersed in a polymer matrix." Safety
constraints are also present: " present regulations are
tending to ban the use of substances of that type,
essentially because of their toxicity and because of
REFERENCES
[1] Alshuller, Genrich. The Innovation Algorithm. Technical
Innovation Center, Inc.Worcester, MA, 2000
[2] Altshuller, Genrich. Creativity as an Exact Science. The Theory
of the Solution of Inventive Problems. Studies in Cybernetics: 5.
Brunel University, Gordon and Breach, Science Publishers,
Inc.1984
[3] Salamatov, Yuri. TRIZ: The Right Solution at the Right Time. A
Guide to Innovative Problem Solving. Insytec B.V., 1999
[4] Savransky, Semyon D. Engineering of Creativity. Introduction
to TRIZ Methodology of Inventive Problem Solving. CRC Press,
2000
[5] www.uspto.gov
37
Studied Patent
Yes
New TS ?
New-TS Patent
analysis
TRIZ Data Base
Contradiction
Patent Analysis
TRIZ Data Base
No
Yes
Solved
Contradiction ?
No
Effects
Procedure
Effects&Evolution
Procedures
Evolution
Procedure
Yes
Known trends of
TS evolution ?
What ?
TRIZ Data Base
No
Yes
Unknown effect
used in solution ?
If not sure, Ask Expert !
Solution more
effective than
previous ?
Describe
TRIZ Data Base
No
Solution in conflict with

known trends of TS
evolution ?
Known effect
Describe:
Natural
Phenomenon
or
Technical
Effects
Correct
decision ?
Common
usage ?
No
Waste deposit
If not sure, Ask an Expert
Yes
No
No
Yes
How ?
TRIZ Data Base
No
New line of TS
evolution ?
Yes
No
Waste deposit
No
TRIZ Data Base
START OVER
Fig. 2
38
Yes
Describe
TRIZ Data Base
Buletinul tiinific al Universitii Politehnica din Timioara

A comparison between a linear feed-back shift register

and a multiple input shift register
1
Mirella Amelia V. Mioc1
Abstract The aim of this paper is to propose

some new techniques for the analysis of Linear
Feedback Shift Registers LFSR and of Multiple
Input Shift Registers MISR, based on their
generating polynomials. First, mathematical
representations of those polynomials ranks are
found. Second, analytic expressions for their
weights are discovered. Using these parameters the
LFSR and MISR, associated to a given generating
polynomial can be designed and simulated. Some
simulation results proving the qualities of the new
design, proposed in this paper, are presented.
Keywords: cryptography, feedback shift register,
calculation.
I.
The in-tapping LFSR is usually called a MISR

(Multiple Input Shift Register).
Cycle codes belong to algebraically codes for
errors detecting.
All arithmetical operations will be developed in a
Galois group.
This paper develops an analysis of a Linear
Feedback Shift Register and a Multiple Input
Shift Register.
II. THE DESCRIPTION OF THE
EXPERIMENT
First of all, the algorithm was applied using
grade 4 polynomials.
The results were correct and accurate.
Then the next natural step was to change the
grade of the polynomial used for generating the
scheme used in the calculus process. The chosen
grade was 8, [1].
The Rijndael algorithm (chosen as replacement
for the well known DES) is based on the
implementing with shift registers of a method
based on working in a Galois field, with the
generator polynomial an irreducible grade 8
polynomial, thus substantially improving
security [1].
Out of the programs, we came to the conclusion
that the results should also rely on the previous
link (given by the equal with one coefficient of
the polynomial).
INTRODUCTION
The Shift Register Cryptosystems variant has

been developed from the evolution of the
encrypting techniques [2].
Such a cryptosystem is based upon generating a
sequence in a finite field and for obtaining it a
Feedback Shift Register is used.
The Linear Feedback Shift Registers are used in
a variety of domains [3]: sequences generators;
counters;
BIST
(Built-In-Self-Test)
[4];
encryption;
PRBS
(Pseudo-Random
Bit
Sequences).
There are two types of LFSR from the utilization
point of view:
the well-known LFSR, that is a intapping LFSR;
the out-tapping LFSR.
1
Facultatea de Automatica si Calculatoare,

Departamentul de Calculatoare, Bd. V, Parvan
Nr. 2, 300223, Timisoara, e-mail:
mmioc@cs.utt.ro
39
This specific link that is also part of the

calculation for MISR is to be taken after the
XOR was made for that rank of the polynomial.
The mathematically representation of each rank
of the polynomial was made accordingly to these
specific links.
Also, this new revealed thing was verified with
the help of programs.
In the end, correlating the results obtained for the
grade 4 polynomials with the results obtained for
the 8 grade polynomials, we came to
mathematical relations for calculating each rank.
The knowledge of the generating polynomial
rank is very important in the estimation of the
security level of a cryptosystem.
These relations point out the previous existing
links and act similarly to a feedback calculated
also from the previous links.
The link acts after the XOR was calculated and
in this way takes the result that was previously
obtained.
In order to demonstrate that those presented
above are correct and precise, we made an
analysis for all the 30 irreducible grade 8
polynomials that were also found in another
program made by us, coming to three particular
cases.
Out of this analysis, there was made a
generalization, that led to the writing of specific
programs for each irreducible grade 8
polynomial, used for the substantially
improvement of security [1].
The simulation programs were tested in both
ways: with the method of making tables
according to the proposed circuits and also with
mathematical methods that materialize the hard
operations.
As input data sets were used several different
multiple combinations of bits, randomly
generated.
The two programs are described in this paper,
one of them simulating the functioning of a
LFSR and the functioning of an analytic MISR,
and the other one simulating the functioning of a
synthetic MISR. The programs are based on
irreducible grade 8 polynomials, allowing the
user to introduce the coefficients of the chosen
polynomial.
There have been chosen three polynomials for
testing these two programs.
The
first
one
is
the
polynomial
P(x)=x8+x6+x5+x3+1; the coefficients would be
introduced in the program as it follows: 1 0 1 1 0
1 0 0 1, and would lead to the following scheme
depicted in Fig. 1.:
Fig. 1. Scheme for the polynomial

P(x)=x8+x6+x5+x3+1
In the program the weights for each chosen
polynomial are calculated.
The knowledge of those weighs is necessary for
the design of the LFSR and MISR associated to
the corresponding generating polynomial.
For the case in the Fig. 1. Scheme, the weights
are:
S0=1 P(x)
S1=x P(x)
S2=x2 P(x)
S3=(x3+x) P(x)
S4=(x4+x2+x) P(x)
S5=(x5+x3+x2) P(x)
S6=(x6+x4+x3+x) P(x)
S7=(x7+x5+x4+x2) P(x)
The
second
one
is
the
polynomial
P(x)=x8+x4+x3+x+1; the coefficients would be
depicted in Fig. 2. :

P(x)=x8+x4+x3+x+1
For the case in the Fig. 2. scheme, the weights
are:
S0=1 P(x)
S1=x P(x)
S2=x2 P(x)
S3=x3 P(x)
S4=x4 P(x)
S5=(x5+x) P(x)
S6=(x6+x2+x) P(x)
S7=(x7+x3+x2) P(x)
The
third
one
is
the
polynomial
P(x)=x8+x7+x6+x4+1; the coefficients would be
40
void prel_SZ()//SR0 calculation

{ int i,j,xor=0;
for (j=1;j<=n;j++)
{ xor=0;
for (i=0;i<8;i++)
{ if (ax[i]==1) //there is a //connection
xor=xor^sr[i][j-1];
if (i!=0)
sr[i][j]=sr[i-1][j-1];
}
sr[0][j]=xor^col[0][j-1];
}
for (i=0;i<8;i++)
rez[i][0]=sr[i][n];
}
In this procedure, the result was calculated by
translating the elements corresponding to the
SRis, with i from 1 to 7, and by collecting the
feedback in the SR0 while executing XOR with
all the SRis that have a corresponding
connection (the power i exists in the chosen
polynomial).
The prel_SN procedure is reproduced below:
void prel_SN(int ind)
{ int i,j,xor=0;
for (i=0;i<8;i++)
for(j=0;j<51;j++)
sr[i][j]=0;
for (j=1;j<=n;j++)
{ xor=0;
for (i=0;i<8;i++)
{ if (ax[i]==1)
xor=xor^sr[i][j-1];
if (i!=0)
sr[i][j]=sr[i-1][j-1];
}
sr[ind][j]=col[ind][j-1] ^sr[ind-1][j-1];
sr[0][j]=xor;
}
for (i=0;i<8;i++)
rez[i][ind]=sr[i][n];
}
SRis, with i from 1 to 7 without ind (the
argument of the procedure), by connecting the
polynomial) and by executing XOR with the
element of the corresponding column and the
element of the corresponding SRind.
The prel_SA procedure is reproduced below:
void prel_SA(int ind)
{ int i,j,k,xor=0;
for (i=0;i<8;i++)
for(j=0;j<51;j++) sr[i][j]=0;

depicted in Fig.3.:

P(x)=x8+x7+x6+x4+1
For the case in Fig. 3. scheme, the weights are:
S0=1 P(x)
S1=x P(x)
S2=(x2+x) P(x)
S3=(x3+x2+x) P(x)
S4=(x4+x3+x2) P(x)
S5=(x5+x4+x3+x) P(x)
S6=(x6+x5+x4+x2) P(x)
S7=(x7+x6+x5+x3) P(x)
In order to get the final results from the
programs, the user has to introduce in the
program the polynomials coefficients as
described above and also the input data sets,
consisting of 8 columns, each of them having a
length of 2n elements (where n is an integer).
In the MISRS.CPP program the synthetic MISR
is calculated for the input data given by the user
and then the results are provided in the end.
Calculating the results consists of categorizing
the 8 SRi steps in which the scheme will be
treated in 3 major types of ways according to the
below described types:
1. the first type is characterized by the
existence of the coefficient of the i
power corresponding to the SRi; in
this case the procedure prel_SA is
called, having i+1 as an argument;
2. the second type is characterized by
the absence of the coefficient of the
i power corresponding to the SRi;
in this case the procedure prel_SN
is called, having i+1 as an
argument;
3. the third type is actually a
particular one: it refers to the case
of SR0, and the procedure prel_SZ
is called, having no argument.
The prel_SZ procedure is reproduced below:
41
The result obtained in this way is the result of the

current step of the calculation for the analytical
MISR.
For each of the eight cases, the result is obtained
as described above.
The final result for the analytical MISR is
obtained by making XORs with all the results of
the 8 steps on each and every column of all the
eight columns.
The interpretation of this final result is that the
ones and zeroes (eight by the number) obtained
are the coefficients of a polynomial. The grade of
this polynomial may be any of those between
seven and zero.
The whole procedure for the LFSR is simpler: it
is obtained by making XOR with all the weights
multiplied by the corresponding column, and
then, using the result obtained (whose grade
gives the number of rounds that are to be done)
as o column corresponding to the SR0 and no
other columns, translating all the elements
without 0, and collecting the feedback in the 0
element executing XOR with all the other
elements that have a corresponding connection
(the power with that rank exists in the chosen
polynomial).
for (j=1;j<=n;j++)
{ xor=0;
for (i=0;i<8;i++)
{ if ((ax[i]==1)&&
((ind-1)!=i))
xor=xor^sr[i][j-1];
if (i!=0)
sr[i][j]=sr[i-1][j-1];
}
sr[ind][j]=col[ind][j-1]
^sr[ind-1][j-1];
sr[0][j]=xor^sr[ind][j];
}
for (i=0;i<8;i++)
rez[i][ind]=sr[i][n];
}
SRis, with i from 1 to 7 without ind (the
argument of the procedure), by connecting the
polynomial) and with the current corresponding
SRind, and by executing XOR with the element
of the corresponding column and the element of
the corresponding SRind.
The difference between prel_SN and prel_SA is
given by the SRind: it is taken into consideration
in the procedure prel_SA in SR0, and does not
appear in the procedure prel_SN in SR0.
The final result is calculated by making XORs
with all the partial results obtained in the 8 steps
on each and every column of all the eight
columns.
In this case, the final result consisting of those

eight figures (ones and zeroes) represents also
the coefficients of a polynomial, just as in the
case of MISR.
The results obtained from the MISRS.CPP
program and the other two results obtained from
the MISRALF.CPP program are the same,
especially since they have been obtained from
the same input data used for calculation that are
equivalent.
These kinds of calculation for verifying the
correctness of the results have been made
primarily on paper.
III. THE RESULTS INTERPRETATION

In the MISRALF.CPP program the analytic
MISR and LFSR are calculated for the input data
given by the user and the results are provided in
the end.
For the given polynomial, the program calculates
in the procedure genr the corresponding 8
weights. These weights are a base for calculation
MISR and LFSR also.
The whole procedure for the analytic MISR
consists of 8 steps: for each step is considered
the corresponding column which is multiplied by
the corresponding weight, then it is divided by
the chosen polynomial, again it is multiplied by
x8 and, finally, divided by the chosen
polynomial.
IV. CONCLUSIONS
Analyzing the experiments already presented it
can be observed that the simulation speed for the
MISR is higher versus the simulation speed for
the LFSR.
The mathematical expressions of generating
polynomial weights, reported in this paper, were
very useful for the simulation of the
corresponding LFSR and MISR.
42
[3] Horowitz and Hill, The Art of Electronics, 2nd edition,

1989.
[4] A. AlYamani II, Logic BIST: Theory, Problems, and
Solutions, Stanford University, RATS/SUM02, 2002.
[5]http://www.mail-archive.com/cryptographydigests@senator-bedfellow.mit.edu/msg02659. html
[6]http://www-2.cs.cmu.edu/afs/cs/project/pscico
guyb/realworld/www/slidesS03/crypto2.4up.pdf
V. REFERENCES
[1]B. Schneier, Applied Cryptography: Protocols,
Algorithms, and Source Code in C, John Wiley and Sons,
New York, 1996.
[2] H. Niederreiter, A PublicKey Cryptosystem Based on
Shift Register Sequences, Proceedings of EUROCRYPT85,
Linz, Austria 1985.
43

Tom 51(65), Fascicola 1-2, 2006
Instructions for authors at the Scientific Bulletin of

Politehnica University of Timisoara, Transactions on
Electronics and Communications
First Author1 Second Author2
Abstract These instructions present a model for editing
the papers accepted for publication in the Scientific
Bulletin of Politehnica University of Timisoara,
Transactions on Electronics and Communications. The
abstract should contain the description of the problem,
methods, solutions and results in a maximum of 12 lines.
No references are allowed here.
Keywords: editing, Bulletin, author
I. INTRODUCTION
The page format is A4. The articles must be of 6
pages or less, tables and figures included.
II. GUIDELINES
The paper should be sent in this standard form. Use a
good quality printer, and print on a single face of the
sheet. Use a double column format with 0.5 cm in
between columns, on an A4, portrait oriented,
standard size. The top and bottom margins should be
of 2.28 cm, and the left and right margins of 2.54 cm.
Microsoft Word for Windows is recommended as a
text editor. Choose Times New Roman fonts, and
single spaced lines. Font sizes should be: 18 pt bold
for the paper title, 12 pt for the author(s), 9 pt bold for
the abstract and keywords, 10 pt capitals for the
section titles, 10 pt italic for the subsection titles;
distance between section numbers and titles should be
of 0.25 cm; use 10 pt for the normal text, 8 pt for
affiliation, footnotes, figure captions, and references.
III. FIGURES AND TABLES
Figures should be centered, and tables should be left
aligned, and should be placed after the first reference
in the text. Use abbreviations such as Fig.1. even at
the beginning of the sentence. Leave an empty line
before and after equations. Equation numbering
should be simple: (1), (2), (3) and right aligned:
Fig. 1. Amplitudes in the standing wave
x(t ) =
a
a
y ( t )d .
(16)
IV. ABOUT REFERENCES

References should be numbered in a simple form [1],
[2], [3], and quoted accordingly [1]. References are
not allowed in footnotes. It is recommended to
mention all authors; et al. should be used only for
more than 6 authors.
Table 1
Parameter
I
U
Value
2.4
10.0
Unit
A
V
V. REMARKS
A. Abbreviations and acronyms
Abbreviations and acronyms should be explained
when they appear for the first time in the text.
Faculty of Electronics and Telecommunications, Communications Dept.

Bd. V. Parvan 2, 300223 Timisoara, Romania, e-mail first.author@etc.upt.ro
2
Faculty of Electronics and Telecommunications, Communications Dept.
Bd. V. Parvan 2, 300223 Timisoara, Romania, e-mail second.author@etc.upt.ro
Abbreviations such as IEEE, IEE, SI, MKS, CGS, ac,

dc and rms need no further explanation. It is
recommended not to use abbreviations in section or
subsection titles.
B. Further recommendations
The International System of units is recommended.
Do not mix SI and CGS. Preliminary, experimental
results are not accepted. Roman section numbering is
optional.
REFERENCES
[1] A. Ignea, Preparation of papers for the International
Symposium Etc. 98, Buletinul Universitii Politehnica, Seria
Electrotehnica, Electronica si Telecomunicatii, Tom 43 (57), 1998,
Fascicola 1, 1998, pp. 81.
[2] R. E. Collin, Foundations for Microwave Engineering, Second
Edition, McGraw-Hill, Inc., 1992.
[3] http://hermes.etc.utt.ro/bulletin/bulletin.html

2005vol50 64no2 PDF

Transféré par

Informations du document

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

2005vol50 64no2 PDF

Transféré par

Droits d'auteur :

Formats disponibles

Redactor ef / Editor in chief

Prof.dr.ing. Ioan Naforni

Colectivul de recenzare / Advisory Board:

Tom 50(64), Fascicola 2, 2005

Buletinul tiinific al Universitii Politehnica din Timioara

Simulated-Inductance High-Stability Sine Oscillator

A double-simulation sine oscillator based on a

Fig.1. Oscillator scheme

The limiting device does not actually deteriorate the

The circuit in Fig.1, without the D1 and D2 diodes,

Likewise, for the A2 amplifier output one may write:

Facultatea de Electronic i Telecomunicaii, Departamentul

the diodes are missing, from the R1-C1 divider in Fig.1

This may be modified through simultaneous

In the case of equalities C1=C2=C, R1=R3=R the

Thus, one may observe the frequency-adjusting

For the case of equalities: R1=R3 and R2=R4 the

3. VOLTAGE PHASE DIAGRAMS

point at vi sine voltage amplitude [1].

0 vi= vR4=vC1 vR2=vR3 vo2

4. QUALITY FACTOR AND AMPLITUDE

Fig.5 shows an oscillator with frequency adjusting

here: - rdeq is the diode equivalent dynamic resistance

Fig.5. Adjusted-frequency oscillator

where: - k is a coefficient depending upon the limiting

In Fig.6 a practical example of this type oscillator

high quality factor (>1000), which assures very good

Fig.6. Experimented oscillator scheme.

The simulations where developed with usual

[1] M. Ciugudean, Ph. Marchegay, Double-Simulation New

Fig.7. Time diagram of the input and output voltages

Buletinul Stiintific al Universitatii Politehnica din Timisoara

Tom 50(64), Fascicola 2, 2005

We start by giving a description of the 100(1 )% CPCI,

Abstract Given (0, 1), the coverage probability of the

For k {1, . . . , n}, `k is defined as the unique solution in

Similarly, the sequence (uk )k{0,...,n} is defined as follows.

Let X henceforth stand for a binomial variate for sample size

II. T HEORETICAL RESULTS

Proposition 2.1: With the same notations as above,

If n < ln(/2)/ ln(2), then

If n ln(/2)/ ln(2), then

The following result derives from the proposition above.

Lemma 2.2: With the same notations as above, for every

P ROOF: Since max(p, 1 p) is larger than or equal to

The 95% CPCI Coverage probabilities for proportion p = 0.1

III. P ROOF OF PROPOSITION 2.1

Given [0, 1], let P stand for the distribution of a

The 95% CPCI Coverage probabilities for proportion p = 0.05

For every [0, 1], G (n) = n . It then follows from

Proof of statement (ii). The strict increasingness of the

Proof of statement (iii). According to (1), (3) and the values

The 95% CPCI Coverage probabilities for proportion p = 0.01

According to [1, Eq. 6.6.4, p. 263], for every given

Lemma 3.3: With the same notations as above,

F (k) = 1 I (k + 1, n k) for k {0, . . . , n 1}. (8)

P ROOF: Statement (i) is straightforward since `X 0.

Lemma 3.1: Given [0, 1],

Lemma 3.4: With the same notations as above,

P ROOF: Set F = {k {0, . . . , n} : uk p}. It follows from

Proof of statement (i). Given k {1, . . . , n}, `k (0, 1)

= P ({p < uX }) P ({p `X })

Since the condition 1 (/2)1/n (/2)1/n is equivalent to