Lars Nielsen. Abstract. In this paper, a general technique for evaluation of measurements by the method of

Transcription

1 Evaluation of measurements by the method of least squares Lars Nielsen Danish Institute of Fundamental Metrology (DFM), Lyngby, DK LNdfmdtudk Abstract In this paper, a general technique for evaluation of measurements by the method of Least Squares is presented The input to the method consist of estimates and associated uncertainties of the values of measured quantities together with specied constraints between the measured quantities and any additional quantities for which no information about their values are known a priori The output of the method consist of estimates of both groups of quantities that satisfy the imposed constraints and the uncertainties of these estimates Techniques for testing the consistency between the estimates obtained by measurement and the imposed constraints are presented It is shown that linear regression is just a special case of the method It is also demonstrated that the procedure for evaluation of measurement uncertainty that is currently agreed within the metrology community can be considered as another special case in which no redundant information is available The practical applicability of the method is demonstrated by two examples Introduction In 787, the French mathematician and physicist Laplace (749{827) used the method of Least Squares to estimate 8 unknown orbital parameters from 75 discrepant observations of the position of Jupiter and Saturn taken over the period 582{745 Since then, the method of Least Squares has been used extensively in data analysis Like Laplace, most people use a special case of the method, known as unweighted linear regression The calculation of the average and the standard deviation of a repeated set of observations is the most simple example of that The unweighted regression analysis is based on the assumptions that the observations are independent and have the same (unknown) variance In addition, the linear regression is based on the assumption that the observations can be modelled by a function that is linear in the unknown quantities to be determined by the regression analysis For most measurements carried out in practice, none of these assumptions can be justied In order to evaluate the result of a general measurement, in which some redundant information has been obtained, one therefore has to apply the method of Least Squares in its general form This paper describes how measurements can be evaluated by the method of Least Squares in general The paper is based on an earlier work of the author [2] but includes Address: uilding 37, Matematiktorvet, DK-28 Lyngby, Denmark

2 2 Lars Nielsen several new features not published before as well as practical examples from the daily work at DFM An alternative approach is described in [6] 2 Measurement model In a general measurement, a number m>ofquantities is either measured directly using measuring instruments or known a priori, for example from tables of physical constants etc The (exact) values of these m quantities are denoted =( ::: m ) T : Due to measurement uncertainty, the values z obtained by the measurement (or from tables etc) z =(z ::: z m ) T are only estimates of the values The standard uncertainties of the estimates z i, u(z i ) i = ::: m are determined in accordance with the GUM [] and depend on the accuracy of the instruments and the reliability ofany tabulated value used In general, some of the estimates z i may be correlated If r(z i z j ) is the correlation coecient between the estimates z i and z j then the covariance u(z i z j )between these two estimates is given by u(z i z j )=u(z i )r(z i z j )u(z j ): ecause of the uncertainty, the estimates z can be considered as an outcome of a m- dimensional random variable Z with expectation (the exact values of the quantities) and covariance matrix = u(z z T )= u 2 (z ) u(z z 2 ) u(z z m ) u(z 2 z ) u 2 (z 2 ) u(z 2 z m ) A : u(z m z ) u(z m z 2 ) u 2 (z m ) In addition to the m quantities for which prior information is available either from direct measurement or from other sources, a general measurement may involve a number k of quantities for which no prior information is available The values of these quantities are denoted by =( ::: k ) T : In general, the values and are constrained by anumber n of physical or empirical laws These constraints may be written in terms of an n-dimensional function f( ) = f ( ) f 2 ( ) f n ( ) A = A k n<m+ k: (2) It is assumed that f i :! R i=:::n, are dierentiable functions (with con-

3 Evaluation of Measurements 3 tinuous derivatives) dened in a region R k+m around ( ) As indicated in (2), the number n has to be larger than or equal to the number k of quantities for which no prior information is available otherwise some of the values cannot be determined In addition, the number n of constraints has to be smaller than the total number k + m of quantities involved otherwise the values of and would be uniquely determined by the constraints and no measurements would be needed The estimates z, the covariance matrix and the n-dimensional function f( ) are the input to the general Least Squares method It should be stressed that no probability distribution has to be assigned to the input estimates z On the contrary, if a probability distribution has been assigned to an estimate, it should be used to calculate the mean value and the variance of the estimate which should then serve as input to the Least Squares method Like any other covariance matrix, the covariance matrix u(z z T )= is positive semi-denite Otherwise, at least one linear combination x T z of the estimates z would have negative variance u(x T z z T x)=x T x In the following it is assumed that is positive denite and therefore non-singular 3 Normal equations Least Squares estimates ^ and ^ of the values and are found by minimizing the chi-square function 2 ( z) =(z ; ) T ; (z ; ) subject to the constraints f( ) =: It is convenient to solve this minimization problem by using Lagrange multipliers [5]: If a solution ( ^ ^) to the minimization problem exists, the solution satises the equation where r ( ) ( ^ ^ z) = ( z) =(z ; ) T ; (z ; )+2 T f( ) for a particular set of Lagrange multipliers =( ::: n ) T y taking the gradient of the function, the following n + m + k equations in ( ^ ^ ) evolve: where r f = f f n r f( ^ ^) T = ; ; (z ; ^)+r f( ^ ^) T = f( ^ ^) = f k f n k A and r f = f f n f m f n m A : The equations (3) are called the normal equations of the Least Squares problem (3)

4 4 Lars Nielsen 4 Solving the normal equations If ( l l l )isanapproximate solution to the normal equations, a rened solution ( l+ l+ l+ ) can be found by the iteration l+ l+ l+ A = l l The step ( l l l+ ) is given by where D( l l )= D( l l ) l l l+ A + l l l+ A = A l = ::: : (k ) ; (z ; l ) ;f( l l ) (k k) (k m) r f( l l ) T (m k) ; r f( l l ) T r f( l l ) r f( l l ) (n n) A (4) A (42) is a symmetric matrix This iteration procedure is similar to Newton iteration except that the second order partial derivatives of the functions f i have been neglected as it is practice to do in non-linear Least Squares estimation [4] In order to reduce the eects of numerical rounding errors, it is recommended to calculate the step ( l l l+ )by solving the linear equations (4) by Gauss-Jordan elimination with full pivoting [4] This algorithm also provides the inverse matrix D( l l ) ; which is needed at the nal stage for estimating the covariance matrix of the solution as shown in Section 5 If proper starting values ( ) are selected, the iteration is expected to converge towards the solution (^ ^) ^ ^ A =lim l! l l l Since the solutions ^ are expected to be close to the estimates z of available a priori, the estimates z are obviously the proper starting values to be selected for the iteration The selection of proper starting values is more dicult in general If, however, f( ) are linear functions in the variables, the iteration process will converge after a few iterations, independent ofthechoice of Most dierentiable functions f( ) can be handled by the described method In order to get reliable standard uncertainties, it is required, however, that the function can be approximated by a rst order Taylor expansion, ie A : f( ) = f( ^ ^)+r f( ^ ^)( ; ^)+r f( ^ ^)( ; ^) when the values and are varied around the solution ^ and ^ on a scale comparable to the standard uncertainties of the solution If this vaguely formulated criterion is met, the function f( ) issaidtobelinearizable Note that almost any dierentiable function

5 Evaluation of Measurements 5 will be linearizable if the standard uncertainties are suciently small On the other hand, if the uncertainties are suciently high, all non-linear functions will no longer be linearizable The requirement that f( ) is linearizable is considered to be the only major limitation of the method of Least Squares! It should be mentioned that the minimization using Lagrange multipliers will fail in case the gradients r f i and r f i of one of the constraint functions f i are both equal to zero at the point of the solution (^ ^) This gives some restrictions on how a constraint maybeformulated A function f i dening a constraint may, for example, be replaced by the square of that function, fi 2 ut since f i(^ ^) =, the gradient offi 2 will be zero at the point of the solution (^ ^) although the gradient off i is not 5 Properties of the solution Since the solution ( ^ ^ ) depends on the estimates z, which are considered as the realization of the multivariate random variable Z, the solution (^ ^ ) canalsobe regarded as a multivariate random variable If the functions f i (^ ^) are linearizable, the estimates (^ ^) are linear in Z ^ ^ A = A + D( ) ; In that case, the expectation of the solution is E ^ ^ A = (k ) ; (Z ; ) (n ) A A : (5) which means that (^ ^) are central estimators of the values ( ) Under the same ; assumption, the covariances of the solution are given by the symmetric matrix D( ^ ^) provided by the Gauss-Jordan elimination algorithm 2 u( ^ ^ T ) u( ^ ^ T ) () (k n) u(^ ^ T ) u(^ ^ T ) () (m n) () (n k) () (n m) ;u( T ) A = D( ) ; = D( ^ ^) ; : (52) This relation can be derived as follows Partition the symmetric matrix D ; into nine sub-matrices similar to the left hand side of (52) or similar to the partitioning of D according to the denition (42) Express the covariance matrix of the solution (5) in terms of the covariance matrix of the random variable Z and the matrix D ; Insert the partitioned D ; into the resulting matrix double product and express the covariances of the solution in terms of ; and the sub-matrices of D ; Reduce these expressions to the nal result by multiple use of the nine relations between the sub-matrices of D ; and D derived from the identity DD ; = I 2 The empty brackets in the left hand matrix indicates the parts of D ; that do not contain information about covariances

6 6 Lars Nielsen From equation (5) and (52) the covariance matrices between ( ^ ^) and the estimates z are found to be u(z ^ T )=u(^ ^ T ) u(z ^ T )=u(^ ^ T ): From the last of these two relations, a relation of particular interest is derived, u(z ; ^ z T ; ^ T )=u(z z T ) ; u(^ ^ T ): For the diagonal elements, this relation reads u 2 (z i ; ^ i )=u 2 (z i ) ; u 2 (^ i ) i = ::: m: That is, the variance of the dierence between the initial estimate z i of i and the rened estimate ^ i is equal to the variance of z i minus the variance of ^ i This relation is useful when testing if the dierence z i ; i is signicantly dierent from its zero expectation 6 2 test for consistency When the estimates (^ ^) have been found, the minimum 2 value 2 (^ z) =(z ; ^) T ; (z ; ^) can be used to test if the measured values z are consistent with the measurement model (2) within the uncertainties dened by thecovariance matrix If the model is linearizable, the expectation of the random variable 2 (^ Z) is equal to the number m of measured quantities, minus the number m + k of adjusted quantities, plus the number n of constraints, that is h i E 2 (^ Z) = m ; (m + k)+n = n ; k = : If, in addition, the random variables Z are assumed to follow amultivariate normal distribution with mean values and covariance matrix, the random variable 2 (^ Z) will follow a 2 () distribution with = n ; k degrees of freedom In that case, the probability p of nding a 2 value larger than the value 2 (^ z) actually observed can be calculated from the 2 () distribution n o p = P 2 () > 2 (^ z) =; P f 2 () 2 (^ z)g: If this probability p is smaller than a certain value, the hypothesis that the measured values are consistent with the measurement model has to be rejected at a level of signicance equal to As the result of measurements are normally quoted at a 95% level of condence, an = 5% level of signicance is a reasonable choice for the consistency test Although the assumption of a normal distribution of Z may not be fullled, it is suggested to carry out the test of consistency as described above anyway This is justied by the fact that a value of 2 (^ z) signicantly higher than the expectation indicates inconsistency no matter what the distribution of Z might be The calculated probability

7 Evaluation of Measurements 7 p simply describes how unlikely the observed 2 assigned to Z value is if a normal distribution is 7 Normalized deviations If the test described in the previous section leads to a rejection of the measurements, a tool for identifying the outlying measurements is desirable A measured value z i is dened as an outlier if the dierence z i ; ^ i is signicantly dierent from zero taking into account the standard uncertainty u(z i ; ^ i ) of that dierence This leads to the introduction of the normalized deviation d i dened by 3 d i = z i ; ^ i u(z i ; ^ i ) = z i ; q ^ i u 2 (z i ) ; u 2 (^i ) i = ::: m: The normalized deviation d i has zero expectation and variance A normalized deviation with jd i j larger than 2 or 3 is therefore rather unlikely no matter what the distribution of the random variable d i might be Ifamultivariate normal distribution is assigned to Z and the model function f( ) is linearizable, the normalized deviation d i is normally distributed, In that case d i 2 N( ) i = ::: m: P fjd i j > 2g =5% and a measurement with jd i j > 2 is therefore identied as an outlier at a 5% level of signicance It is suggested to use the criteria jd i j > 2 to identify potential outliers even if the distribution assigned to Z is not normal 8 Adjustment ofavariance 2 If some values z i have a common but unknown variance u 2 (z i )= 2,thisvariance can be estimated by adjusting 2 by an iterative procedure until the \observed" 2 value becomes equal to its expectation value 2 (^ z) =(z ; ^) T ; (z ; ^) = where the covariance matrix is a function of the unknown variance 2 As the estimates ^ depends on the value assigned to 2, these estimates have to be updated together with the estimates ^ each time the value of 2 is changed during the iteration This way of estimating the unknown variance 2 leads to the well-known expression for the standard deviation in the case of a repeated measurement of a single quantity as shown in Section 3 3 If u(z i ; ^i ) =, the dierence z i ; ^i will be zero as well and d i may be set equal to zero This situation occurs whenever there is no redundant information available regarding the value of the quantity i

8 8 Lars Nielsen 9 Example: alibration of an analytical balance An analytical balance with capacity Max=22 g, resolution d= mg, and built-in adjustment weight was calibrated by DFM in October 999 during an inter-laboratory measurement comparison piloted by DFMTwo mass standards were used as reference standards One of them was a traditional 2 g weight (named R2g) of known conventional mass value 4 m R and density R The other reference standard was a specially designed 2 g stack of weights consisting of four discs (named g, 5g, 25g and 25g*) machined from the same metal bar of known density The conventional mass values m, m 2, m 3, m 4 respectively of these four discs were not known a priori only the conventional mass value m S = m + m 2 + m 3 + m 4 of the stack was known The calibration was performed by placing a weight combination at the weighing pan of the balance and by recording the corresponding average indication I in the display A total of 8 weight combinations were used Each weight combination was weighed 3 times from which the average indication was calculated The calibration was repeated 4 times during a period of days in which the inter-laboratory comparison took place From these four calibrations, a grand average indication I i i= ::: 8 was calculated for each of the 8 weight combinations specied in Table The standard uncertainty of the grand average was estimated from the observed variation in indication over the four calibrations I I 2 I 3 I 4 I 5 I 6 I 7 I 8 I 9 g g g g g g g g g 5g 5g 5g 5g 5g 25g 25g* 25g 25g 25g 25g 25g* 25g* 25g* 25g* I I I 2 I 3 I 4 I 5 I 6 I 7 I 8 5g 5g 5g 5g 25g 25g 25g* R2g R2g 25g 25g 25g* 25g* 25g* Tab The weight combinations corresponding to the 8 balance indications I i Due to the eect of air buoyancy, the balance indication depends not only on the mass of the weighed body, but also on the density ofthebodyaswell as the density of the air When calibrated in air of known density a, the reference indication I R of the balance corresponding to a load generated by aweight with conventional mass value m and density is given by I R = m ; (a ; a ) ; where a =2 kg/m 3 and =8 kg/m 3 are the reference densities of the air and the weight respectively to which the conventional value of mass refers As a model for the 4 The conventional massvalue of a body is dened as the mass of a hypothetical weight of density 8 kg/m 3 that balances the body when weighed in air of density 2 kg/m 3 and temperature 2

9 Evaluation of Measurements 9 m S m R R a I [g] [g] [kg/m 3 ] [kg/m 3 ] [kg/m 3 ] [div] z u(z) ^ u(^) d I 2 I 3 I 4 I 5 I 6 I 7 [div] [div] [div] [div] [div] [div] z u(z) ^ u(^) d I 8 I 9 I I I 2 I 3 [div] [div] [div] [div] [div] [div] z u(z) ^ u(^) d I 4 I 5 I 6 I 7 I 8 [div] [div] [div] [div] [div] z u(z) ^ u(^) 3 d f A m m 2 m 3 m 4 [g/div] [/div] [g] [g] [g] [g] ^ 86-44E u( ^) 9 E-9 Tab 2 Measured and estimated values and associated standard uncertainties calibration curve of the balance, a second order polynomial through zero is assumed I R = f ; I + AI 2 where f and A are unknown quantities to be determined from the calibration data In this example, there are m = 23 quantities for which prior information is available from the measurements performed: =(m S m R R a I ::: I 8 ) T whereas there are k = 6 quantities for which no prior information is available: =(f A m m 2 m 3 m 4 ) T :

10 Lars Nielsen f( ) = f A m m 2 m 3 m 4 f A m m m m Tab 3 orrelation coecients of the estimated ^ values etween these quantities, there are n = 9 constraints: (m + m 2 + m 3 + m 4 ) ; (a ; a ) m R ; (a ; a ) ; ; f ; I + AI 2 R ; ; f ; I 8 + AI 2 8 m S ; (m + m 3 + m 3 + m 4 ) A = : The measured values z and associated standard uncertainties are given in Table 2 under the row headings z and u(z) All measured values are assumed to be uncorrelated y solving the normal equations, the estimates ^ and ^ and associated standard uncertainties given in Table 2 under the row headings ^, u(^), ^ and u( ^) are obtained Selected correlation coecients derived from D( ^ ^) ; are given in Table 3 The observed minimum 2 value is (^ z) =8:6 which should be compared to the expectation value = n ; k =9; 6 = 3 Since P 2 (3) > 8:6 =8:3%, it is concluded that the measured values are consistent with the specied constraints taking into account the measurement uncertainties This conclusion is conrmed by the calculated normalized deviations given in Table 2 under the row heading d all normalized deviations satisfy the criterion jdj < 2 From the estimates of the quantities f and A and the associated covariance matrix, the error of indication E, dened as E = I ; I R = I ; f ; I + AI 2 and the associated standard uncertainty u(e) can be calculated as a function of the indication I The result is shown in Figure as the full lines representing E ; u(e), E, and E +u(e) The measured points E i i= ::: 8 shown in the gure are the observed average balance indications I i minus the corresponding reference values I R The error bars of the measured points indicate the standard uncertainties u(e i )thathave been calculated taking into account the covariance between I i and I R Example: Evaluation of calibration history Aweight (named Rmg) of nominal mass mg has been calibrated 39 times in the period For calibration number i, themass m i of the weight at the time t i and the associated standard uncertainties u(m i )andu(t i ) are given The calibration history of the weight isshown in Figure 2 as dots with error bars indicating the standard

11 Evaluation of Measurements alance P22S Error of indication E /mg Fit Fit+u(Fit) Fit-u(Fit) Measured alance indication I /g Fig Error of indication of the calibrated balance uncertainties the scale mark 992- on the time axis indicates the position of the date January 992 etc Due to wear and changes in the amount of dirt adsorbed to the surface, the mass of the weight is expected to change in time A reasonable model of the change in mass as a function of time is a superposition of a deterministic linear drift and a random variation m i = a + a 2 t i + m i i = ::: 39 where m i isarandomvariable with zero expectation and variance 2 The drift parameters a, a 2 and the associated covariance matrix as well as the variance 2 are unknown a priori and are to be estimated from the calibration history available Once the estimates ^a and ^a 2 have been found, it is possible to predict a value ^m of the mass of the weight as a function of time t ^m =^a +^a 2 t + ^ m where m ^ = with standard uncertainty u( m) ^ = The standard uncertainty ofthe predicted mass value is given by u 2 (^m) =u 2 (^a )+t 2 u 2 (^a 2 )+2tu(^a ^a 2 )+ 2 : The measurement model used for evaluating the calibration history is =(m ::: m 39 t ::: t 39 m ::: m 39 ) T =(a a 2 ) T

12 2 Lars Nielsen Rmg (efore adjustment of σ ) 75 Mass Value [mg] ν = 37, χ 695 obs = 66 6, σ = µ g P{ χ ( 37) > 66 6} = 2% Time of alibration Result Ignore Fit + u Fit Fit - u Normalized Deviation Time of alibration Deviation Fig 2 Evaluation of the calibration history of a mg weight assuming that = Rmg (After adjustment of σ) 75 Mass Value [mg] ν = 37, χobs = 37, σ = 92 µ g P{ χ ( 37) > 37 } = 46 7% Time of alibration Result Ignore Fit + u Fit Fit - u Normalized Deviation Time of alibration Deviation Fig 3 Evaluation of the calibration history of the mg weight with adjusted to 92 g f( ) = m ; (a + a 2 t + m ) m 39 ; (a + a 2 t 39 + m 39 ) A = : The measured values z are given by the calibration history, except for the values of

13 Evaluation of Measurements 3 m i i= ::: 39 which are set equal to the expectation value zero The associated covariance matrix u(z z T )= is built up from the uncertainties u(m i )andu(t i )available from the calibration history and a negligible but nite 5 initial value of the unknown variance 2 Since the standard uncertainties u(m i ) are of the order g, the value =E-7 g is considered negligible and is selected as a starting point y solving the normal equations, estimates ^a and ^a 2 of the drift parameters and the associated covariance matrix are found after a few iterations The predicted value ^m of the mass of the weight and the associated standard uncertainty u(^m) as a function of time are shown in Figure 2 as solid lines The normalized deviations d associated with the mass values m i are shown in Figure 2 as well 6 The observed minimum chi-square value is 2 =66:6 which is large compared to the expectation value =39; 2=37 Since P f 2 (37) > 66:6g =:2%, the hypothesis =,ornorandomvariation in the mass, is rejected at a 2% level of signicance The value of is therefore increased as described in Section 8 until the calculated minimum 2 value becomes equal to its expectation value = 37 In this way the standard uncertainty reecting the random variation of the mass of the weight is found to be =92 g The result of the evaluation of the calibration history after adjustment of is shown in Figure 3 Note the signicant increase in the standard uncertainty of the predicted value of the mass of the weight and the decrease in the absolute value of the normalized deviations d The calibration history can also be evaluated by an iterativetechnique based on linear regression [3] The results obtained are identical to the results presented in this section ase I: Univariate output quantity, Y = h(x ::: X N ) In this section it is shown that the evaluation of measurements by the method of least squares is consistent with the generally accepted principles for evaluating measurement uncertainty as described in the GUM [] Using the nomenclature of the GUM, a univariate output quantity Y is assumed to be related to N input quantities X ::: X N through a specied function h, Y = h(x ::: X N ): The values assigned to the input and output quantities are denoted x ::: x N and by y respectively In the nomenclature of this paper, the measurement model is =(X ::: X N ) T =(Y ) The measured values are f( ) =(Y ; h(x ::: X N )) = : z =(x ::: x N ) T 5 If the variance 2 is assumed to be exactly zero, the quantities m i have toberemoved from the model Otherwise the covariance matrix will be singular 6 The absolute value of normalized deviations of t i and m i is equal to the absolute value of the normalized deviation of m i

14 4 Lars Nielsen with the known covariance matrix = u(z z T )= u 2 (x ) u(x x N ) u(x N x ) u 2 (x N ) The coecient matrix D of the normal equations is where D( ^ ^) = ( N) (N ) ; ;r x h(x) T ;r x h(x) r x h = h h ::: : X X N In the present case, the solution to the normal equations is found after one iteration, A : A y = ^ = h(x ::: x N ) ^ =(x ::: x N ) T =: The associated covariances are given by u2 (y) u(y ^ T ) () ( ) u(^ y) u(^ ^ T ) () (N ) () ( ) () ( N) ;u 2 () In other words, ; A = D( ^ ^) = u 2 (y) =r x h(x) r x h(x) T = r xh(x) r x h(x) T r x h(x) r x h(x) T NX NX i= j= c i u(x i x j )c j c i h X i (x i ) which is identical to the linear variance propagation formula given in the GUM 2 ase II: Linear regression, Y = Xa Linear regression is applied when there is a linear relationship Y = Xa between some observed quantities Y and some unknown quantities a The design matrix X is made up of known elements that maybegiven as specied functions of one or several independent variables In the notation of this paper, the measurement model for the linear regression problem is = Y =(Y ::: Y n ) T = a =(a ::: a k ) T f( ) =Y ; Xa = where X (n k) is the known design matrix The measured values are z = y =(y ::: y n ) T A :

15 Evaluation of Measurements 5 with known covariance matrix = u(z z T )= u 2 (y ) u(y y n ) u(y n y ) u 2 (y n ) The coecient matrix D of the normal equations is D( ^ ^) = (k k) (k n) ;X T (n k) ; I (n n) ;X I (n n) (n n) A : A : Again, the solution to the normal equations is found after one iteration, ^a = ^ = X T ; y ^Y = ^ = X^a = ; ; ( ^Y ; y) where (X T ; X) ; The associated covariances are given by u(^a ^at ) u(^a ^Y T ) () (k n) u( ^Y ^a T ) u( ^Y ^Y T ) () (n n) () (n k) () (n n) ;u( T ) that is, = ; A = D( ^ ^) X T ;X T ; X XX T I ; XX T ; ; ; X I ; ; XX T ; XX T ; ; ; ^a = X T ; y u(^a ^a T )= =(X T ; X) ; as is known from the theory of linear regression 3 ase III: Repeated observations of a single quantity Assume that a quantity X is measured n times with the same uncertainty Sucha measurement can be modelled by n quantities X ::: X n having a common value The measured values are = X =(X ::: X n ) T =() f( ) = X ; X n ; A = : z = x =(x ::: x n ) T and under the assumption that the measurement results are mutually independent, the associated covariance matrix is given by = u(z z T )= 2 2 A : A

16 6 Lars Nielsen The coecient matrix D of the normal equations is D( ^ ^) = ( n) ; ( n) (n ) ;2 I (n n) I (n n) ; (n ) I (n n) (n n) ( ) where denotes a matrix with all elements equal to The solution of the normal equations is found after one iteration, ^ = ^ = n nx i= The associated covariances are given by u2 (^) u(^ ^X T ) () ( n) u( ^X ^) u( ^X ^X T ) () (n n) () (n ) () (n n) ;u( T ) As expected, = A x i ^X = ^ =^ (n ) = ; ;2 ( ^X ; x): ; A = D( ^ ^) 2 n ; 2 n ; ( n) n ; ( n) 2 n ; (n ) 2 n ; (n n) I (n n) ; n ; (n n) n ; (n ) I (n n) ; n ; (n n) ;2 (n ; (n n) ; I (n n) ) ^ = n nx i= x i u 2 (^) = 2 n: If 2 is not known a priori, it can be estimated by solving the equation ie, 2 (^ z) = nx i= 2 = n ; (x i ; ^) 2 2 nx i= (x i ; ^) 2 = n ; which is the well known expression for the experimental standard deviation s 4 onclusion A general technique for evaluation of measurements by the method of Least Squares has been presented The applicability of the method has been demonstrated by two examples It has been shown that the method is fully compatible with the generally accepted principles for evaluation of measurement uncertainty laid down in the GUM and that ordinary linear regression is just a special case of the method The input to the method consists of An estimate of the value of each measured quantity, including anyrelevant inuence quantity The covariance matrix of these estimates formed by the standard uncertainties of the estimates and the correlation coecients between the estimates A :

17 Evaluation of Measurements 7 A measurement model describing all the known relations between the measured quantities and some additional quantities (if needed) for which no prior information is available The output of the method consists of An adjusted estimate of the value of each measured quantity and an estimate of each additional quantity introduced in the measurement model The covariance matrix of all these estimates from which the standard uncertainties and correlation coecients can be calculated Achi-square value which is a measure of the degree of consistency between the measurement model, the input estimates, and the covariances of the input quantities The adjusted estimate of the value of a measured quantity diers from the input estimate only if the measurement model imposes additional information regarding the value of that particular quantity In that case the standard uncertainty of the adjusted estimate will be smaller than the standard uncertainty of the input estimate For a good measurement, the dierence between the adjusted estimate and the input estimate of a measured quantity should not be large compared to the standard uncertainty of that dierence It has therefore been suggested that the ratio d of the dierence to its standard uncertainty is calculated and assessed against a selected criterion, eg jdj < 2 y plotting the d values of the adjusted estimates it is possible to assess whether a too high chi-square value is caused by a few poor input estimates or is due to a poor model ibliography IPM, IE, IF, ISO, IUPA, IUPAP, OIML, Guide to the expression of uncertainty in measurement, ISO, L Nielsen, Least-squares estimation using Lagrange multipliers, Metrologia 35 (998), 5{8 Erratum, Metrologia 37 (2), 83 3 L Nielsen, Evaluation of the calibration history of a measurement standard, DFM report DFM--R25, 2, {6 4 W H Press, S A Teukolsky, W T Vetterling and P Flannery, Numerical Recipies in, 2nd ed, ambridge, ambridge University Press, 992, 36{4 and 68{688 5 T L Saaty and J ram, Nonlinear Mathematics, Dover Publications, New York, 98, 93{95 6 K Weise and W Woger, Uncertainty and Measurement Data Evaluation, Wiley- VH, 999, 83{224 [in German]