Vous êtes sur la page 1sur 7

1.

5 Uncertainty Calculations
R. H. DIECK (2003)

The purpose of this section is to outline the fundamental meth- error with some confidence. That is, the average, plus or minus
ods of measurement uncertainty analysis for use as an objective 2σ divided by the square root of the number of data points,
estimator of data quality. These methods apply to all test, eval- would contain the true average, µ, 95% of the time.
uation, and process data secured by a measurement instrument However, in test measurements, one typically cannot sam-
or system. Some examples are given to clarify the application ple the entire population and must make do with a sample of
of the principles presented. data points. The sample standard deviation, SX, is then used
to estimate σX. For a large data set (defined as having 30 or
1
more degrees of freedom ) ±2SX divided by the square root
UNCERTAINTY AND ERROR of the number of data point e reported average contains the
true average, µ, 95% of the time. That SX divided by the square
Measurements are made so that the resulting data may be root of the number of data points in the reported average, M,
used for decision-making. In fact, the most fundamental def- is called the standard deviation of the average (sometimes
inition of “good” data is “data that are applicable, or useful, also called the random uncertainty) and is written as
for drawing conclusions or making decisions.” Because of
this, no test or evaluation data should be presented or used N

without including its measurement uncertainty. It is a prop-


erly evaluated measurement uncertainty that provides the infor-
∑(X − X )
i =1
i
2

SX = M = SX / M 1.5(2)
mation needed to properly assess the usefulness of data. For N −1
data to be useful, it is necessary that their measurement errors
where
be small in comparison to the changes or effect under eval-
SX = standard deviation of the average; the sample standard
uation. The actual measurement error is unknown and unknow-
deviation of the data divided by the square root of M
able. Measurement uncertainty estimates its limits with some
SX = sample standard deviation
confidence.
Therefore, measurement uncertainty may be defined as X = sample average, that is,
the limits to which a specific error or system error may extend
M

∑ ( X /N )
with some confidence. The most commonly used confidence
in uncertainty analysis is 95%, but other confidences may be X= i 1.5(3)
i =1
employed where appropriate. In this section, all examples
will be at 95% confidence.
Xi = ith data point used to calculate the sample
Error is most often defined as the difference between the standard deviation and the average
measured value of one data point and the true value of the N = number of data points used to calculate the
measurand. That is: sample standard deviation
E = (measured) − (true) 1.5(1) (N − 1) = degrees of freedom of SX and SX
M = the number of data points in the reported
where average test result
E = measurement error
measured = value obtained by a measurement Note in Equation 1.5(3) that, usually, N = M. This is not
true = true value of the measurand a requirement, however. N does not necessarily equal M. It is
possible to obtain SX from historical data with many degrees
It is possible to estimate only the expected limits to an of freedom ((N – 1) greater than 30) and to take only M data
error at some confidence. The most common method for points in a specific test. The test result, or average, would
1
estimating those limits is to use the normal distribution. therefore be based on M measurements, and the standard devi-
For an infinite population (N = ∞), the standard deviation, ation of the average could still be calculated with Equation
σ, would be used to estimate the expected limits of a particular 1.5(3). In that case, there would be two averages. One average

86
© 2003 by Béla Lipták
1.5 Uncertainty Calculations 87

would be from the historical data used to calculate the sample where
standard deviation, and the other would be the average test U Ai = standard deviation (based on data) of the average
result for M measurements. for uncertainty source i of Type A each with its own
In summary, the error for any particular error source is degrees of freedom (UA is in units of the test or
unknown and unknowable. To use data for decisions, it is measurement result. It is an SX ).
necessary to estimate the limits to which a particular error may NA = number of parameters of Type A uncertainty
extend with some confidence. This estimate is called the uncer- θi = sensitivity (or influence coefficient) of the test or
tainty. Many are familiar with the term accuracy. However, measurement result, R, to the ith Type A uncertainty
accuracy is sometimes ambiguous. For example, if an exper-
imenter needed an instrument with twice the accuracy of ±1%, The uncertainty of each error source is in units of that
would that be ±0.5% or ±2%? For this reason, this section will source. When multiplied by the sensitivity for that source, it
use the term uncertainty throughout to describe the quality of is converted to that uncertainty in result units. The effects of
test data. several error sources may be estimated by root-sum-squaring
their uncertainties, as they are now all in the same units. The
sensitivities, θi, are obtained for a measurement result, R,
CLASSIFYING ERROR SOURCES AND THEIR which is a function of several parameters, Pi. The equations
UNCERTAINTIES follow.

When estimating the limits of errors, their sources may be R = the measurement result where
grouped into classifications to ease their understanding. The R = f(P1, P2, P3...PN)
most common error source and uncertainty classifications P = a measurement parameter used to calculate the
now utilized are the International Standards Organization result, R
2
(ISO) classifications and the United States/American Soci- ∂R
3
ety of Mechanical Engineers (US/ASME) classifications. θi =
∂Pi
The first groups error sources and their uncertainties by a
type designation depending on whether there are data avail- Obtaining the θi is often called error propagation or uncer-
able to calculate the sample standard. The second classifica- tainty propagation.
tion groups errors and their uncertainties by their effects on
the experiment or test data. That is, the US/ASME classifi- ISO Type B Uncertainties and Errors For Type B, where
cation groups error sources and uncertainties by random and there are no data to calculate a standard deviation,
systematic types with subscripts used to denote whether there
are data to calculate a standard deviation. For this reason, 1/ 2
 NB

∑ (θ U )
2
the US/ASME, or engineering, classification system groups UB =   1.5(5)
i Bi
usually are more useful and recommended for engineering  i =1 
applications.
where
U Bi = standard deviation of the average for uncertainty
The ISO Classification of Errors and Uncertainties source i of Type B. (UB is in units of the test or
measurement result, R. It is an SX . This standard
The ISO error and uncertainty classification system is not deviation of the average is an estimate and not based
recommended in this section. However, the total uncertainty on data.)
it yields is in complete agreement with the recommended NB = number of parameters with a Type B uncertainty
classification system, the engineering classification system. θi = sensitivity of the test or measurement result to the
In this system, errors and uncertainties are classified as Type ith Type B uncertainty R
A if there are data to calculate a sample standard deviation
2
and Type B if there are not. For these uncertainties, it is assumed that the U Bi represent
The impact of multiple sources of error is estimated by one standard deviation of the average for one uncertainty source
root-sum-squaring their corresponding uncertainties. The with an assumed normal distribution of errors. Here, the degrees
equations follow. of freedom associated with this standard deviation (also stan-
dard deviation of the average) is assumed to be infinity.
ISO Type A Uncertainties and Errors For Type A, where Note that θi, the sensitivity of the test or measurement result
data exist for the calculation of a standard deviation, to the ith Type B uncertainty, is actually the change in the result,
R, that would result from a change of the size of the Type B
1/ 2
 NA
 uncertainty in the ith input parameter used to calculate that result.
∑ (θ U )
2
UA =  i Ai  1.5(4) The degrees of freedom from both the U Ai and the U Bi are
 i =1  needed to compute the degrees of freedom of the combined

© 2003 by Béla Lipták


88 General Considerations

total uncertainty. The degrees of freedom for this total uncer-


tainty are needed to select a proper Student’s t for the con- TABLE 1.5a
Student’s t Statistic for 95% Confidence, t95, Degrees of Freedom,
fidence of interest, usually 95%. Those degrees of freedom
ν. This is Frequently Written as t95,ν
are calculated with the Welch–Satterthwaite approximation.
4
The general formula for degrees of freedom is ν t95 ν t95 ν t95

1 12.706 11 2.201 21 2.080


2 2 4.303 12 2.179 22 2.074
 N 2


∑( )SX 
i

3 3.182 13 2.160 23 2.069

dfR = ν R =  i =1 1.5(6) 4 2.776 14 2.145 24 2.064


 N
( ) 
4
SX  5 2.571 15 2.131 25 2.060

 ∑
 i =1
i

(ν i ) 

6 2.447 16 2.120 26 2.056
7 2.365 17 2.110 27 2.052
 
8 2.306 18 2.101 28 2.048
9 2.262 19 2.093 29 2.045
where
10 2.228 20 2.086 ≥30 2.000
dfR = νR = degrees of freedom for the result
νi = the degrees of freedom of the ith standard deviation
of the average related. That is, if it were possible to know the error in a
measurement from one source, one could calculate or predict
For the ISO model, Equation 1.5(6) becomes an error magnitude and direction from the other nonindepen-
dent error source. These are sometimes called dependent
error sources. Their degree of dependence may be estimated
2
 NA NB
 with the linear correlation coefficient. If there are some non-
∑( ) + ∑( )
2 2
 θiU Ai θiUBi  independent errors, whether Type A or Type B, Equation
 i =1 i =1  1.5(8) becomes
6
dfR, ISO = ν R, ISO = 1.5(7)
( )+ ( )
 NA θ U θiUBi 
4 4
NB

∑ ∑ 
i Ai
 (νi ) (ν i )   B Ni . T
 N j ,T
 
1/ 2

 i =1 
∑∑ ∑
i =1
U R, ISO = t95  (θiUi,T )2 + θiθ jU(i,T ),( j ,T ) (1 − δ i, j ) 
 T = A  
i =1  j =1 
The degrees of freedom calculated with Equation 1.5(7) 1.5(9)
is often a fraction. The U.S. National Standard on Test
5
Uncertainty recommends that this be truncated to the next
where
lower whole number to be conservative.
Ui,T = ith elemental uncertainty of Type T (can be
In computing a total uncertainty, the uncertainties noted
Type A or B)
by Equations 1.5(5) and 1.5(6) are combined. For the ISO
3 UR,ISO = total uncertainty of the measurement or test
model, this is calculated as
result
θi = sensitivity of the test or measurement result to
[ ] the ith Type T uncertainty
1/ 2
U R, ISO = ±t95 (U A )2 + (UB )2 1.5(8)
θj = sensitivity of the test or measurement result to
the jth Type T uncertainty
where t95 = Student’s t for νR degrees of freedom U(i,T ),( j,T ) = covariance of Ui,T on Uj,T
K

Student’s t may be obtained from Table 1.5a. = ∑U


l =1
i ,T (l )U j ,T (l ) 1.5(10)
Note that alternative confidences are permissible; 95% is
5
recommended by the ASME, but 99% or 99.7% (or any other = sum of the products of the elemental systematic
confidence) is obtained by choosing the appropriate Student’s uncertainties that arise from a common source
t. However, 95% confidence is recommended for uncertainty (l)
analysis.
5 l = an index or counter for common uncertainty
In all of the above, the errors were assumed to be inde- sources
pendent. Independent sources of error are those for which an K = number of common source pairs of uncertain-
error in a measurement in one source cannot be used to ties
predict the magnitude or direction of an error from the other, δi,j = Kronecker delta (δi,j = 1 if i = j, and δi,j = 0
6
independent, error source. Nonindependent error sources are otherwise )

© 2003 by Béla Lipták


1.5 Uncertainty Calculations 89

This ISO equation will yield the same total uncertainty where
as the engineering equation, but the ISO classification does Bi,T = ith parameter elemental systematic uncertainty
not provide insight into how to improve an experiment’s or of Type T
test’s uncertainty—that is, whether to possibly take more data BR = systematic uncertainty of the measurement or
because the random uncertainties are too high or calibrate test result
better because the systematic uncertainties are too large. N = total number of systematic uncertainties
The engineering classification presented next is therefore the θi = sensitivity of the test or measurement result to
recommended approach. the ith systematic uncertainty
θj = sensitivity of the test or measurement result to
the jth systematic uncertainty
Engineering Classification of Errors and Uncertainties
B(i,T ),( j,T) = covariance of Bi on Bj
The engineering classification recognizes that experiments and M

tests have two major types: systematic and random. Their cor-
responding estimates of the limits of those errors are the sys-
= ∑B
l =1
i ,T (l ) Bj ,T (l ) 1.5(13)

tematic uncertainties and random uncertainties, respectively. = sum of the products of the elemental system-
atic uncertainties that arise from a common
Random Errors and Uncertainties The general expression source (l)
for random uncertainty is the (1 SX ) standard deviation of the l = an index or counter for common uncertainty
9
average sources
δi,j = Kronecker delta (δi,j = 1 if i = j, and δi,j = 0
1/ 2 6
 B Ni ,T
  B Ni ,T
 otherwise )
∑∑ ∑ ∑ (θ S ) 
2 2
SX , R = θ S   = Mi,T
 T = A  i Xi ,T 
  T = A
i Xi ,T
i =1 i =1  Here, each Bi,T and Bj,T are estimated as 2 SX for an assumed
1.5(11) normal distribution of errors at 95% confidence with infinite
degrees of freedom. If there are less than an infinite number
where of degrees of freedom, the appendix of Reference 5 details
SXi ,T = sample standard deviation of the ith random error that analysis.
source of Type T The random uncertainty, Equation 1.5(11), and the sys-
SX = random uncertainty (standard deviation of the aver- tematic uncertainty, Equation 1.5(12), must be combined to
i ,T
age) of the ith parameter random error source of obtain a total uncertainty,
Type T
SX , R = random uncertainty of the measurement or test
[ ]
1/ 2
result U R,ENG = t95 ( BR /2)2 + ( SX , R )2 1.5(14)
Ni,T = total number of random uncertainties, Types A and
B, combined
Note that BR is in units of the test or measurement result,
Mi,Τ = number of data points averaged for the ith error
as was SX , R .
source, Type A or B i
The degrees of freedom will need to be determined for
the engineering system total uncertainty. It is done with the
Note that SX , R is in units of the test or measurement result
Welch–Satterthwaite approximation, the general form of
as a result of the use of the sensitivities, θi. Here, the elemental
which is Equation 1.5(10). The specific formulation here is
random uncertainties have been root-sum-squared with due
consideration for their sensitivities, or influence coefficients.
Since these are all random uncertainties, there is, by definition,  B  NSXi ,T N B ,T 2 

∑ ∑( ) ∑
 
2 i
 θ j Bj ,T   
no correlation in their corresponding error data, so these can  SX ,T +   
  2  
 T = A  i =1
i
always be treated as independent uncertainty sources.  j −1  
df R = ν R = 
  θ B  
4

( )
4
Systematic Errors and Uncertainties The systematic uncer-  B  NSXi ,T S 2   
N B ,T  j j ,T 

i

tainty of the result, BR, is the root-sum-square of the elemental
∑∑ 
∑ 
Xi ,T
 +
systematic uncertainties with due consideration for those that  T = A  i =1 ν i,T j −1
νj 
6
are correlated. The general equation is   
 

1/ 2
1.5(15)
 B NT
 NT
 

BR = 
 T = A
∑∑ (θ i Bi,T ) 2 +
 ∑ θ iθ j B(i,T ),( j ,T ) (1 − δ i, j ) 
 
where
N S = number of random uncertainties of Type T
i =1 j =1  Xi ,T
1.5(12) NBj,T = number of systematic uncertainties of Type T

© 2003 by Béla Lipták


90 General Considerations

νι,Τ = degrees of freedom for the ith uncertainty of The number of degrees of freedom needed to define Stu-
Type T dent’s t for Equations 1.5(16) through 1.5(19) is calculated
νϕ,Τ = infinity for all systematic uncertainties with the Welch–Satterthwaite approximation, Equation 1.5(7)
for the ISO system and Equation 1.5(15) for the engineering
system.
TOTAL UNCERTAINTY It usually may be assumed that the degrees of freedom
equal 30 or higher. In these cases, the equations for uncertainty
ISO Total Uncertainty simplify further by setting t95 equal to 2.000. This approach is
recommended for a first-time user of uncertainty analysis pro-
The ISO total uncertainty, which ISO refers to as the
cedures, as it is a fast way to get to an approximation of the
expanded uncertainty, for independent uncertainty sources
measurement uncertainty.
(the most common) is [from Equation 1.5(8)]

[ ]
1/ 2
U ISO = ± K (U A ) 2 + (U B ) 2 1.5(16)

The ISO uncertainty with some non-independent uncer- CALCULATION EXAMPLE


tainty sources is [from Equation 1.5(9)]
In the following calculation example, all the uncertainties are
1/ 2 independent, and are all in the units of the test result, tem-
 B Ni ,T
 N j ,T
 

U R, ISO = ±  ∑∑
 T = A i =1
(θiUi,T )2 +


∑j =1
θiθ jU( t ,T ),( j ,T ) (1 − δ i, j ) 


perature. More detailed examples are given in many of the
references cited. Their review may be needed to assure a
more comprehensive understanding of uncertainty analysis.
1.5(17)
Of particular importance is the concept of uncertainty prop-
agation. This is needed when the uncertainties are for param-
Engineering Total Uncertainty
eters that contribute to a calculated result.
7
The engineering equation for total uncertainty for inde- It has been shown that there is often little difference
pendent uncertainty sources (the most common) is [from among the uncertainties calculated with the different models.
8
Equation 1.5(14)] The data from Table 1.5b will be used to calculate measure-
ment uncertainty with these two models. This data are all in
[ ]
1/ 2
U R,ENG = ±t95 ( BR /2) 2 + ( SX , R ) 2 1.5(18) temperature units, so the influence coefficients, or sensitivi-
ties, are all unity.
Here, just the first term of Equation 1.5(17) is needed, as all Note the use of subscripts “A” and “B” to denote where
the systematic uncertainty sources are independent. data exist to calculate a standard deviation. Also note that,
The engineering equation for uncertainty for nonindepen- in this example, all errors (and therefore uncertainties) are
dent uncertainty sources (those with correlated systematic independent and that all degrees of freedom for the system-
uncertainties) is also Equation 1.5(18), but it is necessary to atic uncertainties are infinity except for the reference junc-
also to use the full expression for BR [Equation 1.5(12)] tion, whose degrees of freedom are 12. Also note that BR is
calculated as
1/ 2
 B NT
 NT
 

BR =  ∑∑
 T = A
(θ i Bi,T ) 2 +
 ∑ θ iθ j B(i,T ),( j ,T ) (1 − δ i, j ) 
   0.06  2  0.07  2  0.1 2 
1/ 2

BR = 2 
i =1 j =1
+ + = 0.13 1.5(20)
1.5(19)  2   2.18   2  

TABLE 1.5b
Test Measurement Uncertainties, F
Systematic Degrees of Standard Number Random Degrees
Defined Measurement Uncertainty, Freedom for Deviation, of Data Uncertainty, of Freedom
Process Bi Bi SX,i Points, Ni S X ,i for S X ,i

Calibration of thermocouple 0.06A ∞ 0.3A 10 0.095A 9


Reference junction 0.07B 12 0.1A 5 0.045A 4
Data acquisition 0.10A ∞ 0.6A 12 0.73A 11
RSS BR = 0.13 N/A N/A N/A S X , R = 0.20 N/A

© 2003 by Béla Lipták


1.5 Uncertainty Calculations 91

Notice that the 0.07 is not divided by t but by its appropriate t95 is therefore 2.07. UR,ISO is then
Student’s t for 12 degrees of freedom. The use of “2” for
Student’s t is only appropriate for all degrees of freedom of 30 2 2 1/2
UR,ISO = ±2.07[(0.21) + (0.058) ] = 0.45
5
or more. for 95% confidence 1.5(25)
Each uncertainty model will now be used to derive a
measurement uncertainty.
Engineering Uncertainty Calculation Example

ISO Uncertainty Calculation Example For the engineering System, UR,ENG, model [Equation 1.5(18)],
we have the following expression:
For the UISO model we have, from Equation 1.5(13), the
following expressions: 2
UR,ENG = ±t95[(0.13/2) +(0.20) ]
2 1/2
1.5(26)

1/ 2 Here, the (0.13/2) is the BR/2, and the 0.20 is (as before) the
 0.07  
2
U A = (0.095) + (0.045) + (0.173) + 
2 2 2
= 0.21 random component. To obtain the proper t95, the degrees of
  2.18  
freedom need to be calculated just as in Equation 1.5(24).
1.5(21) There, the degrees of freedom were 22, and t95 = 2.07. UR,ENG
1/ 2
is then
 0.06  2  0.10  2 
U B =  + = 0.058
 2  
1.5(22)
 2 
2 2 1/2
UR,ENG = ±2.07[(0.13/2) +(0.20) ] = 0.44
2 2 1/2 2 2 1/2
for 95% confidence. 1.5(27)
UR,ISO = ±K[(UA) + (UB) ] = ±K[(0.21) + (0.058) ]
1.5(23) This is essentially identical to UR,ISO [Equation 1.5(25)]
within round-off errors and degrees of freedom approxima-
tions, as predicted.
Here, remember that the 0.21 is the root sum square of the
1 SX Type A uncertainties in Table 1.5b, and 0.058 is that
for the 1 SX Type B uncertainties. Also note that, in most
cases, the Type B uncertainties have infinite degrees of free- SUMMARY
dom and represent an equivalent 2 SX . That is why they are
divided by 2 to get an equivalent 1 SX . Where there are less Although these formulae for uncertainty calculations will not
than 30 degrees of freedom, one needs to divide by the
handle every situation, they will provide a useful estimate of
appropriate Student’s t that gave the 95% confidence interval.
test or measurement uncertainty in most cases. For a more
For the reference junction systematic uncertainty above, that
detailed treatment or specific applications of these principles,
was 2.18.
consult the references and bibliography.
If K is taken as Student’s t95, the degrees of freedom must
first be calculated. Remember that all the systematic compo-
nents of Type B have infinite degrees of freedom except for
the 0.07, which has 12 degrees of freedom. Also, all the Bi References
in Table 1.5a represent an equivalent 2 SX except for 0.07,
which represents 2.18 SX , as its degrees of freedom are 12 1. Doebelin, E. O., Measurement Systems, Application and Design, 4th
and not infinity. To use their data here, divide all of them ed., McGraw-Hill, New York, 1990, 38ff.
2. ANSI/ASME PTC 19.1–1985, Instruments and Apparatus, Part 1,
except the 0.07 by 2, and divide the 0.07 by 2.18; they all
Measurement Uncertainty, ANSI, New York, 1985, 64.
now represent 1 SX as do the random components. All Type 3. Guide to the Expression of Uncertainty in Measurement, International
A uncertainties, whether systematic or random in Table 1.5a, Standards Organization, Geneva, 1993, 23.
have degrees of freedom as noted in the table. The degrees 4. Dieck, R. H., Measurement Uncertainty, Methods and Applications,
of freedom for UISO is then 2nd ed., ISA, Research Triangle Park, NC, 1996, 45.
5. ANSI/ASME PTC 19.1–1998, Instruments and Apparatus, Part 1,
Measurement Uncertainty, ANSI, New York, 1998.
6. Brown, K. K., Coleman H. W., Steele, W. G., and Taylor, R. P.,
dfR = ν R Evaluation of correlated bias approximations in experimental uncer-
tainty analysis, in Proc. 32nd Aerospace Sciences Meeting & Exhibit,
2 Reno, NV, AIAA paper no. 94–0772, Jan. 10–13, 1996.
[( 0.095) + ( 0.045) + ( 0.173) + ( 0.06/2 ) + ( 0.07/2.18) + ( 0.10/2 ) ]
2 2 2 2 2 2

= 7. Strike, W. T. III and Dieck, R. H., Rocket impulse uncertainty; an


 ( 0.095) 
4 4 4 4 4 4 4
( 0.045) ( 0.173) ( 0.06/2 ) ( 0.07/2 ( 0.07/2.18) ) ( 0.10/2 ) uncertainty model comparison, in Proc. 41st Int. Instrum. Symp.,
+ + + + +
 9 4 11 ∞ 12 ∞  Denver, CO, May 1995.
8. Dieck, R. H., Measurement uncertainty models, in Proc. 42nd Int.
= 22.51 ≈ 22 1.5(24) Instrum. Symp., San Diego, CA, May 1996.

© 2003 by Béla Lipták


92 General Considerations

Bibliography Dieck, R. H., Measurement Uncertainty, Methods and Applications, 3rd ed.,
ISA, Research Triangle Park, NC, 2002.
Steele, W. G., Ferguson, R. A. and Taylor, R. P., Comparison of ANSI/ASME
Abernethy, R. B. et al., Handbook of Uncertainty in Gas Turbine Measure-
and ISO models for calculation of uncertainty, in Proc. 40th Int.
ment, USAF AEDC-TR-73-5, U.S. Air Force, 1973 (adopted by ISA
Instrum. Symp., 1994, 410–438.
as its Measurement Uncertainty Handbook).
Strike, W. T., III and Dieck, R. H., Rocket impulse uncertainty; an uncer-
Abernethy, R. B. and Ringhiser, B., The history and statistical development
tainty model comparison, in Proc. 41st Int. Instrum. Symp., Denver,
of the new ASME-SAE-AIAA-ISO measurement uncertainty method-
CO, May 1995.
ology, in Proc. AIAA/SAE/ASME/ASME 21st Joint Propulsion Conf.,
Webster, J. G., The Measurement, Instrumentation and Sensors Handbook,
Monterey, CA, July 8–10, 1985.
Ed., CRC Press, Boca Raton, FL, 1999.

© 2003 by Béla Lipták

Vous aimerez peut-être aussi