NMI TR 8 Uncertainties in Colour Measurements James L. Gardner First edition — August 2005 Bradfield Road, Lindfield, NSW 2070 PO Box 264, Lindfield, NSW 2070 Telephone: (61 2) 8467 3600 Facsimile: (61 2) 8467 3610 Web page: http://www.measurement.gov.au © Commonwealth of Australia 2005 This document was originally prepared for the Korean Institute of Standards and Science in 2003 while James Gardner was there as a guest researcher. It had minor revisions up to November 2003. This document is that latter revision (circulated as Revision 2.7) with text added to the section dealing with distribution temperature. CONTENTS 1 Propagation of Uncertainty ....................................................................................................... 1 2 Tristimulus Values .................................................................................................................... 2 3 Sources versus Surfaces ............................................................................................................ 3 4 Transfer Measurement and Component Uncertainties .............................................................. 3 5 Uncertainties (Variances) and Covariances of the Tristimulus Values ..................................... 5 6 Covariance of a Colour Triplet.................................................................................................. 7 7 Methods of Calculation for Colour Triplets .............................................................................. 7 7.1 (x,y,Y) Colour Coordinates ............................................................................................... 7 7.2 (u,v,Y) Colour Coordinates ............................................................................................... 8 7.3 (u′,v′,Y) Colour Coordinates ............................................................................................. 8 7.4 (L*,a*,b*) Colour Coordinates......................................................................................... 9 7.5 (L*,u*,v*) Colour Coordinates ......................................................................................... 9 7.6 (L*h*c*) Colour Coordinates ......................................................................................... 10 8 Uncertainty Components in Colour Measurement .................................................................. 10 8.1 Base Uncertainty of the Reference Standard .................................................................. 11 8.2 Random Measurement Noise in the Transfer ................................................................. 11 8.3 Random Wavelength Accuracy in the Transfer ............................................................. 11 8.4 Wavelength Offset Error in the Transfer........................................................................ 12 8.5 Constant Relative Error in Absolute Value of the Reference or the Transfer Ratio ...... 12 8.6 Constant Offset in the Reference or Sample Channels .................................................. 12 8.7 Scaling of the Reference Value or Transfer Ratio Linear with Wavelength ................. 13 8.8 Source Noise .................................................................................................................. 13 8.9 Bandwidth Effects .......................................................................................................... 14 8.10 Sample Uniformity ......................................................................................................... 14 8.11 Multiple Transfers from the Reference .......................................................................... 14 9 Representative Examples for Surface Colour Uncertainties ................................................... 15 9.1 Typical Uncertainty Components................................................................................... 15 9.2 (x,y,Y) by Component ..................................................................................................... 16 9.3 (L*a*b*) by Component ................................................................................................ 17 9.4 Combined Uncertainty for the Various Surface Colour Quantities ............................... 19 10 Methods for Source Parameters and Uncertainties ................................................................. 20 10.1 Dominant Wavelength.................................................................................................... 20 10.2 Source Correlated Colour Temperature ......................................................................... 21 10.3 Source Distribution Temperature ................................................................................... 22 11 Representative Examples for Source Colour Uncertainties .................................................... 25 11.1 Typical Uncertainty Components................................................................................... 25 11.2 Broadband Sources ......................................................................................................... 26 11.3 Narrow-band Sources ..................................................................................................... 27 12 Conclusion ............................................................................................................................... 29 13 References ............................................................................................................................... 29 iii SUMMARY Techniques are presented to estimate uncertainties in the various colour coordinates that may be reported from measurements of surface reflectance or source spectral power distributions. The measurements are treated as a transfer from a reference reflector or source. Uncertainties are then calculated for effects arising in either the base reference or in the transfer measurement. Uncertainties of the various colour quantities are propagated through uncertainties in the tristimulus values, which are in turn propagated from uncertainties in the spectral distribution of the reflectance or source power. Effects contributing to the uncertainty of a given colour value are treated separately and then combined into the final uncertainty. The form of the tristimulus integrals leads to simple expressions to calculate their uncertainties. Many systematic effects can be described by a single parameter. Then for fullycorrelated effects, the uncertainty of each tristimulus value is a linear sum of the signed spectral uncertainty, weighted by the appropriate colour-matching function. Further, the tristimulus values themselves are fully correlated and the covariances of the tristimulus values are products of the uncertainty sums. For effects random between wavelengths the covariances are the product of the same linear sums and the variances are linear sums of the spectral variances, again weighted by the appropriate colour-matching functions. For surface colour uncertainties, the colour-matching functions are replaced by their convolution with the illuminant spectral power distribution. Sensitivity coefficients for the dependence of colour quantities on the tristimulus values are then used to propagate uncertainty from the tristimulus values to those quantities, considered as colour triplets. Some colour quantities are combinations of simpler quantities; for these, expressions are given for the propagation of the covariance from the simpler values to the more complex ones. Representative calculations are made for both sources and surfaces. Uncertainties vary with position in colour space and should be estimated for each spectral measurement. In addition to the various colour triplets, methods for estimating uncertainties in correlated colour temperature and dominant wavelength of sources are given, as is that for source distribution temperature (which is not strictly a colour parameter, as it does not depend on the tristimulus values). iv 1 PROPAGATION OF UNCERTAINTY Uncertainty propagation is described in detail in the ISO Guide to the Expression of Uncertainty in Measurement [1]. The uncertainty in a quantity X formed by combining measured quantities xi through the relationship X f ( x1 , x2 ,..xN ) is given by 2 N 1 N f 2 f f u (X ) u ( x ) 2 u ( xi , x j ) (1) i i 1 xi i 1 j i 1 xi x j where u ( xi ) is the uncertainty in xi and u ( xi , x j ) is the covariance of xi and x j . For uncorrelated input quantities, the covariance of pairs of variables is zero and Eq. (1) reduces to N 2 2 f 2 u (X ) (2) u ( xi ) i 1 xi the ‘sum of squares’ commonly applied. The derivatives f xi are sensitivity coefficients for the dependence of X on the various measured quantities. Given that u 2 ( xi ) u ( xi , xi ) , Eq. (1) can be expressed as N 2 u 2 ( X ) f x U xf x T (3) where f f f fx .. xn x1 x2 is a column vector of sensitivity coefficients ( T indicates the transpose) and U x u ( xi , x j ) (4) (5) is the N N variance–covariance matrix of squares-of-uncertainty (variance) in diagonal elements, covariance values elsewhere. If we form another quantity Y by combining the measured quantities xi through the relationship Y g ( x1 , x2 ,..xN ) , the uncertainty in Y is given by an expression similar to that of Eq. (1), but now the quantities X and Y are correlated through dependence on the common set xi. The covariance between X and Y is given by N N f g u ( X , Y ) u ( xi , x j ) (6) i 1 j 1 xi x j In matrix form, this is simply expressed as u( X , Y ) f x U xg xT (7) where g g g (8) gx .. xn x1 x2 Correlation coefficients are normalised covariance values, defined as u( X , Y ) (9) r( X ,Y ) u ( X )u (Y ) We note that if r ( xi , x j ) 1 for all pairs, Eq. (1) reduces to N u( X ) i 1 NMI TR 8 f u( xi ) xi (10) 1 In section 5 we see that this linear sum can be used to calculate the uncertainty of many systematic effects even where the correlation coefficient changes sign through the spectrum. Colour quantities can all be expressed in terms of the tristimulus values. Uncertainties in the derived quantities are then found by propagating uncertainties through sensitivity coefficients and the variance–covariance matrix for the tristimulus values. 2 TRISTIMULUS VALUES Tristimulus values are integrals representing convolution of the CIE colour-matching functions [2], shown in Figure 1, and the spectral power distribution of the light reaching the detector (the eye). The set of colour-matching functions chosen (2 or 10) depends on the application. 2.0 x y z Response 1.5 1.0 0.5 0.0 400 500 600 700 Wavelength /nm Figure 1. CIE colour-matching functions (the solid lines with symbols are the 1931 values for a 2º field of view; the dashed lines are for the 1964 10º observer) The X,Y, Z tristimulus values are carried in the vector N N N (11) T X Y Z xi Si yi Si zi Si i 1 i 1 i 1 where Si is the spectral power distribution reaching the eye (at the ith of N wavelengths) and xi , yi , zi are the CIE colour-matching functions defined over the wavelength range from 360 nm to 830 nm. Chromaticity values involve ratios of tristimulus values and luminance or lightness is usually expressed as a ratio to that of a reference illuminant; hence only relative spectral power distributions are important and we can delete from Eq. (11). (But note that when we calculate component uncertainties for the tristimulus values using different wavelength steps we must include a scaling factor for when adding components.) NMI TR 8 2 3 SOURCES VERSUS SURFACES In reflectance colorimetry we measure reflectance under a specified geometry as a function of wavelength in the visible range. The spectral power distribution reaching the eye is then a convolution of the reflectance spectrum and a source power distribution, usually taken as CIE Illuminant A or CIE Illuminant D65 [2], depending on the application. Illuminant A represents typical incandescent lighting and Illuminant D65 represents a phase of daylight. These agreed source distributions are shown in Figure 2 – their values carry no uncertainty. Spectral Power Distribution 120 100 80 60 40 A D65 20 0 400 500 600 700 Wavelength /nm Figure 2. Spectral power distributions for CIE Illuminants A and D65 All the colour quantities for surfaces can be derived using the same expressions as those for sources provided the colour-matching functions are replaced by their convolution with the reference illuminant spectrum xi ' SiIll xi , yi ' SiIll yi , zi ' SiIll zi (12) where SiIll is the value of the illuminant spectral power distribution at the ith wavelength. Values of the spectral power distribution, S i , in Eq. (11) are then replaced by Ri , the value of the reflectance at the ith wavelength. In the sections below we take S i to refer to the measured spectrum (source or reflectance) and the colour-matching functions to be amended as in Eqs (12) if we are referring to a surface measurement (with the prime notation dropped). 4 TRANSFER MEASUREMENT AND COMPONENT UNCERTAINTIES Spectral reflectance may be required for various geometric conditions and is measured against a reference standard. Similarly, measurements of spectral power distribution are traced back to a reference standard. Hence the spectral value Si can be written as a transfer from that of a reference standard S iRef Si ti SiRef (13) Uncertainties in the spectral value Si arise both from those of the reference value and those introduced by the spectral transfer. Colour values are formed by combining the spectral measurements at different wavelengths. Systematic effects in a spectral measurement have a cooperative relationship between wavelengths, that is, the values NMI TR 8 3 at different wavelengths are correlated. Correlations also exist between the reference values at different wavelengths, arising from systematic effects in the methods used to derive them. Table 1. Uncertainties ui , j in spectral values Si due to different effects j, combined to estimate the uncertainty u c in the value of a colour quantity; the total uncertainty of Si at wavelength i is uic and the uncertainty in the colour value due to effect j is u c | j Wavelength Effect 1 Effect 2 Effect m Combined 1 u1,1 u1,2 u1,m u1c 2 u2,1 u2,2 u2,m u 2c n un ,1 un ,2 un , m u nc Combined u c |1 u c |2 u c |m uc Individual sources of uncertainty can be considered as independent effects. Table 1 shows the two ways that uncertainties in spectral values due to these individual effects can be combined when the spectral values themselves are combined. Because the individual effect are not correlated, we can form the uncertainties of the combined effects at a given wavelength by sum-of-squares, Eq. (2). However, these spectral values are now most often partially correlated, and we must use the full expressions of Eqs (1) or (7) to calculate the final combined uncertainty, u c . The final combined uncertainty u c is also found by combining the u c | j values by sum-of-squares. The advantage of this form of combination is that many of the systematic effects that can contribute to uncertainty across wavelengths are either uncorrelated (random) – which means that the uncertainty for the combination across wavelengths is found by sum-of-squares – or fully correlated, for which the simple linear sums of Eq. (10) can be used. For uncertainty components of the reference spectrum, from Eq. (13) we have u ( Si ) ti u ( SiRef ) (14) whereas for components in the transfer we have u ( Si ) SiRef u (ti ) (15) In both cases we calculate the transfer value as the ratio ti Si SiRef . The reference spectrum and the measured spectrum generally are defined at different wavelengths. To avoid introducing correlations through interpolation procedures [3], we use different procedures for reference and transfer uncertainty components. When dealing with reference uncertainty components, we interpolate the measured spectrum (which has no uncertainty if we are only considering a single component of the reference spectrum) to the wavelengths of the reference spectrum. When dealing with transfer uncertainty components, we interpolate the reference spectrum (which has no uncertainty if we are only considering a single component of the transfer) to the wavelengths of the measured spectrum. NMI TR 8 4 5 UNCERTAINTIES (VARIANCES) AND COVARIANCES OF THE TRISTIMULUS VALUES The reference spectrum values will in general be partially correlated. The reference contribution to the variance–covariance of the tristimulus values is then calculated directly using Eq. (7); this is discussed in detail in section 8.1Ошибка! Источник ссылки не найден.. Many effects systematic across wavelength can each be described by a single parameter p as (16) Si Si ( p) Noting that uncertainties are always positive, the uncertainty of Si due to this effect alone is S (17) u ( Si ) i u ( p ) p For an effect that is random from wavelength to wavelength, the uncertainty in the X tristimulus value is propagated through Eq. (2), with the variance given by N u 2 ( X ) | p xi 2u 2 ( Si ) (18) i 1 with similar expressions for the variance of the Y, Z tristimulus values. Even though the spectral values are uncorrelated, the common parameter correlates the tristimulus values. We have u ( Si , S j ) 0 for i j , u ( Si , Si ) u 2 ( Si ) and the covariance between X and Y, for example, is N u ( X , Y ) | p xi yi u 2 ( Si ) (19) i 1 For non-random effects, the covariance of spectral values at different wavelengths is given by S S j 2 (20) u (Si , S j ) i u ( p) p p Hence u ( Si , S j ) u ( Si )u ( S j ) and the spectral values are fully correlated for the effect we are considering. The value of the correlation coefficient r ( Si , S j ) can be either +1 or –1 at different parts of the spectrum (for example, an offset in wavelength may cause such an effect) and so we cannot necessarily apply Eq. (10). However we can rewrite Eq. (20) using the sign function sgn( ) as S S S j 2 S (21) u (Si , S j ) sgn i sgn j i u ( p) p p p p The covariance of the tristimulus values, due to our single-parameter effect, is given by expressions of the form X Y 2 u ( X , Y )| p u ( p) p p N N S S j 2 xi y j i u ( p) p p i 1 j 1 N N S S S S xi sgn i i u ( p) y j sgn j j u ( p) i 1 j 1 p p p p NMI TR 8 (22) 5 These are the diagonal terms required for the tristimulus variance–covariance matrix. If we treat the uncertainty as carrying the sign of the sensitivity component S S S (23) us ( Si ) sgn i i u ( p) i u ( p) p p p we have N N i 1 j 1 u( X , Y )| p xi u s (Si ) y j u s (S j ) (24) where the s subscript denotes a signed uncertainty. Using Eqs (1), (11), (17) and (21), the variance of the X tristimulus value is given by 2 N N 1 N S j Si S j 2 Si 2 Si 2 u ( X )| p xi u ( p) u ( p) 2 xi x j sgn sgn p i 1 i 1 j i 1 p p p p or 2 N S S (25) u ( X )| p xi sgn i i u 2 ( p) p p i 1 with similar expressions for Y and Z. In terms of the signed uncertainties we have 2 u ( X )| p N x u (S ) i 1 i s i (26) Comparing Eqs (24) and (26) shows u ( X , Y ) u ( X )u (Y ) (27) that is, the tristimulus values are fully correlated. The sum terms in Eq. (24) are in fact the signed uncertainties of the X and Y tristimulus values; the covariance values are a product of these signed uncertainties. The complete variance–covariance matrix for the tristimulus values, for our effect being considered, is u 2 ( X ) u( X , Y ) u( X , Z ) U XYZ | p u ( X , Y ) u 2 (Y ) u (Y , Z ) (28) 2 u ( X , Z ) u (Y , Z ) u ( Z ) and the uncertainty in any colour quantity is found by propagating these values through Eq. (7) and the appropriate sensitivity coefficients. An important consequence of systematic effects is that all colour quantities calculated from fully-correlated spectral values are also fully correlated; the correlation is positive for systematic effects that are positively correlated, but may vary for systematic effects whose correlation can be positive or negative. The variances and covariances of the tristimulus values for all effects are found as sums of those of the independent individual effects. Here we must take into account the wavelength step used in calculating the tristimulus values. In calculating the base reference spectrum uncertainties, the wavelength step may be different from that of the measured spectrum used to calculate the systematic effects. If so, the tristimulus variances and covariances from the base reference spectrum must be scaled by Ref spectrum . NMI TR 8 2 6 Metameric colours are those for which their spectral distribution is different but their tristimulus values and hence appearance are identical. The uncertainties of the tristimulus values depend on the spectrum itself, and so uncertainties will be different for the parameters of metameric colours. 6 COVARIANCE OF A COLOUR TRIPLET Three variables are required to describe colour. In the case of the simple chromaticity values (x,y) or (u,v), the third variable is the Y tristimulus value itself (luminance), often quoted as a ratio to that of an illuminant. These values may be combined to form other colour quantities, and we then need to know not only their uncertainty (or variance), but also the relation between them (covariance), which may be quoted as a matrix of correlation coefficients. Instead of calculating individual variance or covariance values using Eq. (7), we form a 3x3 matrix of sensitivity coefficients for each desired quantity in terms of the X,Y and Z tristimulus values. If S is such a matrix with colour quantities ordered by columns and tristimulus values by rows, the variance–covariance matrix for the colour quantities is given by T U S U XYZ S (29) Using x,y,Y as an example of a colour-triplet, the matrix S xyY is given by x X x Y x Z 7 y X y Y y Z Y X Y Y Y Z (30) METHODS OF CALCULATION FOR COLOUR TRIPLETS With the exception of source temperature parameters and dominant wavelength, the quantities of interest are colour-triplets, for which uncertainty is propagated from the variance–covariance matrix of the tristimulus values, the appropriate sensitivity matrices and Eq. (29). In the following sections we derive the sensitivity matrices for the various colour triplets. Some of these require a second stage of propagation, because they in turn depend on other colour quantities. 7.1 (x,y,Y) Colour Coordinates The (x,y) chromaticity values are given simply as Y X , y , with Txy X Y Z (31) x TXY Txy The reported value of Y is usually Y/YN for a source, where YN is the Y tristimulus value of the reference illuminant, or 100 Y/YN for reflectance. In the source case, the sensitivity matrix is NMI TR 8 7 Y Z 2 Txy X S xyY 2 Txy X Txy2 with the obvious modification for reflectance. Y Txy2 X Z Txy2 Y Txy2 0 1 YN 0 (32) The linear dependence of the numerator for the x,y sensitivity coefficients on the tristimulus values has two important consequences. Any spectral uncertainty component which is a constant multiplied by the spectral value leads to zero uncertainty in the x and y chromaticity values (and also u, v, u and v). This is true for an uncertainty in the relative value of the reference or in the transfer, as expected since the chromaticity values are ratios of the tristimulus values to their sum. In the case of reflectance colorimetry, often the reference spectrum has a value that is approximately constant; in such cases, we have from Eq. (13) Si u Si Ref u SiRef Si u SiRef (33) Si and uncertainty in the offset in the reference spectrum also leads to zero uncertainty in the (x,y) chromaticity values. A further general conclusion can be drawn for red LED x,y chromaticites. Here the Z tristimulus value is effectively zero and the chromaticity is close to the monochromatic boundary. Hence x y 1 and any sensitivity coefficients for x are the negative of those for y and any uncertainties for x and y will be equal. 7.2 (u,v,Y) Colour Coordinates In a manner similar to the treatment for (x,y,Y) above 4X 6Y u , v with Tuv X 15Y 3Z Tuv Tuv In the source case reporting Y/YN, the sensitivity matrix is 60Y 12 Z 6Y 0 2 2 Tuv Tuv 60 X 6 X 18Z 1 SuvY Tuv2 Tuv2 YN 12 X 18Y 0 2 2 Tuv Tuv 7.3 (34) (35) (u′,v′,Y) Colour Coordinates These are given as a simple scaling of (u,v), u ' u, v ' 3v / 2 . This scaling is applied to the middle column of Eq. (35). NMI TR 8 8 7.4 (L*,a*,b*) Colour Coordinates L*,a* and b* are calculated as 1 Y 3 L* 116 16 YN X 1 3 Y 1 3 a* 500 X N YN Y 13 Z 13 b* 200 ZN YN (36) where Xn, Yn and Zn are the tristimulus values for a perfect reflector (i.e. the illuminant distribution alone – these carry no uncertainty). The sensitivity matrix is S L*a*b* 7.5 0 116 13 2 3 Y Y 3 N 0 500 13 2 3 XN X 3 500 13 2 3 YN Y 3 0 200 13 2 3 YN Y 3 200 13 2 3 ZN Z 3 0 (37) (L*,u*,v*) Colour Coordinates These are defined as 1 Y 3 L* 116 16 , u* 13L *(u ' u ' N ) , v* 13L *(v ' vN ') YN (38) where u′, v′ are CIE 1976 chromaticity coordinates and u′N, v′N are similar quantities for the illuminant alone. First we calculate the covariance matrix U Lu ' v ' for the quantities L*, u′ and v′, for which the sensitivity matrix in terms of the tristimulus values is 60Y 12Z 9Y 0 Tuv2 Tuv2 116 13 2 3 60 X 9 X 27 Z (39) S L*u ' v ' YN Y 2 2 3 T T uv uv 12 X 27Y 0 Tuv2 Tuv2 where Tuv X 15Y 3Z . The covariance matrix U L*u ' v ' for the quantities L*, u′ and v′ is then U L*u ' v ' S TL*u ' v ' U XYZ S L*u ' v ' NMI TR 8 (40) 9 The sensitivity matrix for the quantities L*, u* and v* in terms of L*, u′ and v′ is 1 13(u ' uN ') 13(v' vN ') (41) S L*u*v*' 0 13L * 0 0 0 13L * and the final uncertainties and correlations are carried in the variance–covariance matrix U L*u*v* S TL*u*v*U Lu ' v 'S L*u*v* (42) 7.6 (L*h*c*) Colour Coordinates The quantities hue and chroma are calculated from either (a*,b*) or (u*,v*) chromaticity pairs. Taking the (a*,b*) example b* (43) h* tan 1 , c* a *2 b *2 a* We first calculate the variance–covariance matrix U L*a*b* for L*a*b* from that of the tristimulus values using the sensitivity matrix in section 7.4. U L*a*b* S TL*a*b*U XYZ S L*a*b* (44) The sensitivity matrix for the quantities L*, h* and c* in terms of L*, a* and b* is 1 0 0 b * a * (45) S L*h*c*' 0 2 c* c* a* b* 0 c *2 c* and the final uncertainties and correlations are carried in the variance–covariance matrix U L*h*c* S TL*h*c*U La*b*S L*h*c* (46) Hue and chroma are also calculated from (u*,v*) chromaticities; uncertainties in these are found by substituting (u*,v*) for (a*,b*) in the above equations. Saturation s* may be required in place of chroma; s*=c*/L* and it is a simple matter to modify the sensitivity matrix S L*h*v* to accommodate this change. 8 UNCERTAINTY COMPONENTS IN COLOUR MEASUREMENT For each independent component we calculate the transfer ratio at each wavelength and hence the spectral uncertainties using Eq. (14) or (15). We then propagate those uncertainties to the tristimulus values, and from those to the desired colour quantities as a full variance–covariance matrix. The component matrices are then summed to obtain the combined uncertainties and correlations. Such correlations are required for estimating uncertainties of combinations of the colour quantities, such as in determining colour-differences, correlated colour temperature and dominant wavelength. The tristimulus values, and hence their uncertainties, vary strongly throughout colour space; it is not possible to provide accurate colour uncertainties as a single value applicable over the whole of colour space. NMI TR 8 10 8.1 Base Uncertainty of the Reference Standard The spectral reference standard is likely to be measured at a limited number of wavelengths, with some systematic errors in the process. The reference values then will be at least partially correlated. In this case the full expressions of Eq. (1) or its matrix form Eq. (7) must be used to calculate the tristimulus uncertainties and correlations. If the correlation coefficient for the reference values is constant then this calculation can be split into fully-correlated and uncorrelated parts and the simpler expressions of Eqs (2) and (10) used to calculate the variance. At the highest levels of accuracy, reference spectra uncertainties are dominated by systematic effects and they are highly correlated. 8.2 Random Measurement Noise in the Transfer Assuming that the monochromator efficiency and the source spectral power distribution are significant at all wavelengths through the visible range, the uncertainty in the transfer can generally be defined in terms of two components. The first is a fixed fraction p of the transfer value, which can be related to source noise. In the formalism of section 5Ошибка! Источник ссылки не найден., we have at any wavelength, ti pti' with p=1 (if it is not, a correction should be applied), where ti' is the measured transfer ratio. Hence Si pti ' SiRef pSi ' , where S i' is the measured spectral value, and u (ti ) ti u ( p) . From Eq. (13) we then have (47) us (Si ) Si u( p) The second is fixed offsets in the sample and reference signal channels, which can be related to electronic noise in the measurement – offsets due to scattered light are S p treated in section 8.6. Here ti Refi with p, pR 0 but u ( p), u ( pR ) 0 . The Si pR offsets being treated here are considered uncorrelated, specified as fractions of the respective maximum sample or reference signal value. Hence 2 2 1 S u (ti ) Ref u 2 ( p) Refi 2 u 2 ( pR ) Si ( Si ) 2 (48) and 2 Si 2 (49) us ( Si ) u ( p) Ref u ( pR ) Si As both these effects are uncorrelated between wavelengths we use Eq. (18) to calculate the variance of the tristimulus values and Eq. (19) to calculate the covariances. 2 8.3 Random Wavelength Accuracy in the Transfer We assume that the same wavelength setting is used to measure the sample and reference signals. If we have incorrectly set the wavelength by an amount p, the transfer function becomes NMI TR 8 11 ti Si p Srefi p 1 Si Si 1 p Si 1 Si 1 Siref t ' 1 p p i ref ref S S S 1 i i i Siref 1 ref p Si (50) to first order. We now have p 0 , u ( p ) 0 and 1 Si 1 S ref (51) us ( Si ) ref i Siref u ( p) Si Si The derivatives are calculated numerically. As these values are uncorrelated between wavelengths we use Eq. (18) to calculate the uncertainties and Eq. (19) to calculate the covariance of the tristimulus values. 8.4 Wavelength Offset Error in the Transfer This can arise if the spectral lamp(s) used for calibrating the wavelength scale have different alignment to the broad-spectrum lamp used for measurement. The uncertainty treatment for each spectral value is identical to that of the previous section but now the uncertainties of different wavelengths are fully correlated. Hence we use Eq. (26) to calculate the uncertainties and Eq. (27) to calculate the covariances of the tristimulus values. 8.5 Constant Relative Error in Absolute Value of the Reference or the Transfer Ratio These two effects are equivalent. They may rise from a number of instrumental effects. In a double-beam spectrometer the beam alignment on the detector may be different in the two internal paths, for example. Variation in the distance setting of the reference and sample sources may also produce such an effect. The magnitude of this effect can be estimated by reversing the reference and sample positions. (A better estimate of the relevant spectral value is then made by taking the geometric mean of the two sets of readings.) Whether the effect is in the reference or measured spectrum, we have Si pSi' (52) where Si' is the uncorrected value. We have p 1 (if it is not a correction must be applied) and we estimate its uncertainty u(p). This effect is fully correlated between wavelengths and we use Eq. (26) to calculate the uncertainties and Eq. (27) to calculate the covariance of the tristimulus values. In section 7.1 we note the consequence of such an uncertainty on (x,y) chromaticity values. 8.6 Constant Offset in the Reference or Sample Channels These two effects are not equivalent. Scattered light in either reference or sample beam can cause such an offset, as can electronic offset in an amplifier. For an offset S p in the sample channel, ti i Ref and Si u ( Si ) SiRef u (ti ) u ( p) S For an offset in the reference beam we have ti Ref i , hence Si p NMI TR 8 (53) 12 u (ti ) Si S Ref i 2 u ( p ) and Si u ( p) (54) SiRef Note that if the reference spectral value is constant, a common occurrence in reflectance spectroscopy, and the source intensity is approximately constant, an uncertainty of the offset in the reference spectrum is equivalent to an uncertainty in the relative value of the transfer or reference value, discussed in the last section. The conclusions of section 7.1 then also apply. u ( Si ) ti u ( SiRef ) The effects are fully correlated between wavelengths and we use Eq. (26) to calculate the uncertainties and Eq. (27) to calculate the covariance of the tristimulus values. An offset that is common to both channels is correlated. An example is incorrect background subtraction of room light reaching a detector from outside the signal paths. In such a case the two effects should be treated together, with a correlation coefficient of –1. 8.7 Scaling of the Reference Value or Transfer Ratio Linear with Wavelength Such an effect may be due to an alignment that shifts with wavelength, or a reference standard which ages at different rates at different wavelengths. If we assume that known effects have been corrected and that the uncertainty of the correction is linear with wavelength, we have Si Si 1 p(i 1 ) , where p 0 is the fractional differential error, and (55) us (Si ) (i 1 ) Siu( p) Again we use Eqs (26) and (27) to calculate the tristimulus variance–covariance matrix. The resultant uncertainties (in all except (x,y) or (u,v) chromaticities) depend on the wavelength chosen for the reference. This should be where the true values are known for the reference or the transfer ratio. 8.8 Source Noise Source noise is treated through random and correlated effects in the transfer. The treatment depends on the time scale of the fluctuations, and is generally different for source and reflectance colorimetry. Reflectance colorimetry usually is carried out in a double-beam instrument, with relatively rapid switching between sample and reference paths. In the visible range the source is usually a thermal one, with relatively slow fluctuations; hence source fluctuations are likely to correlate the transfer between sample and reference measurements, but not the transfer ratio between wavelengths. Short-term source noise is likely to be important if a discharge lamp (e.g. a deuterium source) is used, leading to increased random fluctuations in the transfer. Long-term drift may be important for source colorimetry, because the reference and sample beams are usually measured sequentially. Long-term drift or fluctuation leads to correlations between spectral values at different wavelengths, if spectra are recorded on a time scale short compared with the rate of the fluctuation. These correlations can be measured experimentally by recording a number of repeat spectra and calculating the correlation between values. NMI TR 8 13 Correlation between wavelengths may also occur if the gain of a detector changes randomly but slowly; such effects have been seen in a scanning system where the detector was a photomultiplier driven from a DC–DC converter and a power supply of only moderate stability. 8.9 Bandwidth Effects Unresolved or poorly resolved spectral shapes can affect the calculated values of colour coordinates – this is important in measuring LED colour parameters, for example. Such effects are systematic, but not covered here. We take the reference value as the one applicable for the resolution being used in our measurement. In most cases the spectral distribution of the reference is relatively smooth and bandwidth effects are small. For reflectance colorimetry we also take the source spectral power distribution to be constant over the bandwidth and consider any fluctuation effects to have been treated in the transfer uncertainties. The measurement equation is R S d Ref Si Si (56) Ref R S d where R is the spectral response function of the instrument and the integration is over the bandwidth. If we also consider the reported spectral value for the sample to be one applicable for the resolution being used, that is, an average over the bandwidth of the measurement, this becomes Si R d Si Ref SiRef ti SiRef (57) Si R d as in Eq. (13). Bandwidth effects are then treated as correction factors and not uncertainty components. Ignoring the bandwidth effects can lead to correction factors that are larger than the propagated uncertainties – it is important to state the measurement conditions when reporting results. 8.10 Sample Uniformity This component is estimated from repeat measurements on a given sample. 8.11 Multiple Transfers from the Reference The spectral result of one measurement may be the reference for a subsequent one, as we progress from base standards through various levels of working standards. Here the base reference spectrum may be used to calculate uncertainties provided the transfer uncertainty components are combined for all the transfers. The usual rules for such combination apply; if the uncertainties for a given component are correlated, the combined transfer uncertainty is found by sum-of-squares, Eq. (2). If the effects are correlated, they are added linearly, Eq. (10). Care must be taken here. For example, if the wavelength scale was recalibrated between transfers, and the lamp repositioned, the offset applicable for the second transfer is not correlated to that of the first, whereas it is if the system is undisturbed between transfers. NMI TR 8 14 9 REPRESENTATIVE EXAMPLES FOR SURFACE COLOUR UNCERTAINTIES 9.1 Typical Uncertainty Components Uncertainty components for a double-beam reflectance measurement are shown in Table 2. These are chosen to be large enough to show relativity between the components, yet be representative of a routine measurement. The reference reflectance value is taken to be unity and strongly correlated between wavelengths. Scattered light can produce reasonable offsets in either the reference or sample beams, and these components have been set to 0.1% of the maximum signals, taken to be independent effects. An uncertainty of 1% in the transfer absolute value can be caused by misalignment of the sample and reference beams, or a non-uniform detector response if there is a lateral or angular shift of the beams at the detector. The random wavelength setting of 0.03 nm can be achieved by careful calibration using a number of spectral lines. The wavelength offset of 0.1 nm represents 1/50 of the typical 5 nm bandwidth used for spectral measurements. Table 2. Uncertainty components used for representative calculations for surfaces (standard uncertainties); the component number identifies the uncertainty source for the subsequent plots Component Identification Uncertainty 1 Base standard 0.5% of the value Correlation coefficient 0.8 20 points used 2 Reference slope uncertainty 1% over 400 nm 3 Reference offset uncertainty 0.1% of the maximum 4 Transfer random uncertainty (relative) 1% 5 Transfer random uncertainty (signal offsets) 0.1% of the signal maxima 6 Transfer scaling factor uncertainty 1.0% 7 Spectrum signal offset uncertainty 0.1% of the maximum 8 Wavelength random uncertainty (nm) 0.03 nm 9 Wavelength offset uncertainty (nm) 0.1 nm 10 Sum of all components All sample spectra were calculated over a 5 nm grid from 360 nm to 830 nm, using the 10 colour-matching functions and Illuminant D65. A grey sample surface was simulated as one with a uniform reflectance of 50%. A green surface with broad spectral distribution was simulated by the V( distribution scaled by 0.9 on a 10% grey background; red and blue surfaces were simulated by shifting the V( distribution by 75 nm and –150 nm, respectively, both scaled by 0.9 and on a 10% grey background. Examples of surface colour uncertainty estimates are also given in [4]. NMI TR 8 15 9.2 (x,y,Y) by Component Grey surface: x = 0.3138, y = 0.3310, Y = 50.00 Grey 0.0005 0.6 u(x) u(y) 0.5 Uncertainty Uncertainty 0.0004 Grey 0.0003 0.0002 0.0001 u(Y) 0.4 0.3 0.2 0.1 0.0 0.0000 1 2 3 4 5 6 7 8 9 1 10 2 3 4 5 6 7 8 9 10 Component Component Red surface: x = 0.5169, y = 0.3689, Y = 34.91 0.0008 Red Red 0.4 u(Y) Uncertainty Uncertainty 0.0006 u(x) u(y) 0.0004 0.0002 0.3 0.2 0.1 0.0000 0.0 1 2 3 4 5 6 7 8 9 10 1 2 3 4 Component 5 6 7 8 9 10 Component Green surface: x = 0.3880, y = 0.4924, Y = 71.72 0.0004 Uncertainty Green Green 0.8 u(Y) u(x) u(y) Uncertainty 0.0005 0.0003 0.0002 0.6 0.4 0.2 0.0001 0.0 0.0000 1 2 3 4 5 6 Component NMI TR 8 7 8 9 10 1 2 3 4 5 6 7 8 9 10 Component 16 Blue surface: x = 0.1963, y = 0.1430, Y = 14.83 0.0008 0.20 Blue u(x) u(y) 0.15 Uncertainty Uncertainty 0.0006 Blue 0.0004 u(Y) 0.10 0.05 0.0002 0.00 0.0000 1 2 3 4 5 6 7 8 9 1 10 2 3 4 5 6 7 8 9 10 Component Component For the routine uncertainty components chosen, the uncertainty of the reference scale is relatively minor in all cases. In part this is due to the high level of correlation between reference values. If the correlation coefficient is reduced from 0.8 to 0.3, the reference uncertainty component approximately doubles in value. The effect of wavelength errors varies significantly with the surface colour. For the grey surface, wavelength errors are zero, as expected. The effect of random wavelength error is negligible, but wavelength offset is highly significant for the green surface, less so for red and blue, and the effect is quite different for the x and y chromaticities – the effect is strongest where the overlap between the spectrum and colour-matching function is strongest, i.e. where the colour-coordinate itself is strongest. Offsets in the signals have a significant effect on chromaticity. Here the reference spectrum is constant and we have considered the source strength to be constant, and so it is the offset in the sample signal that is important. The effect of the offset uncertainty is more pronounced as the 10% general background set for the calculations above is reduced, particularly for the red and blue surfaces; the effect of the offset (scaled from the maximum signal value) is more significant if the reflectance is small for much of the visible range. Scaling of the transfer value doesn’t affect the chromaticity, but is highly significant for the luminance. 9.3 (L*a*b*) by Component Grey surface: L* = 76.07, a* = 0, b* = 0 Grey 0.35 a* b* 0.15 Grey 0.30 Uncertainty Uncertainty 0.20 0.10 0.05 0.25 L* 0.20 0.15 0.10 0.05 0.00 0.00 1 2 3 4 5 6 7 Component NMI TR 8 8 9 10 1 2 3 4 5 6 7 8 9 10 Component 17 Red surface: L* = 65.68, a* = 48.97, b* = 47.75 0.4 Red Red 0.3 a* b* Uncertainty Uncertainty 0.3 0.2 L* 0.2 0.1 0.1 0.0 0.0 1 2 3 4 5 6 7 8 9 1 10 2 3 4 5 6 7 8 9 10 Component Component Green surface: L* = 87.83, a* = –26.77, b* = 69.92 0.4 Green 0.4 Uncertainty Uncertainty L* a* b* 0.3 Green 0.2 0.1 0.3 0.2 0.1 0.0 0.0 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 8 9 10 Component Component Blue surface: L* = 45.41, a* = 34.73, b* = –66.36 0.3 Blue a* b* Uncertainty Uncertainty 0.3 Blue 0.2 0.1 0.0 1 2 3 4 5 6 7 Component 8 9 10 0.2 L* 0.1 0.0 1 2 3 4 5 6 7 Component In contrast to the (x,y) chromaticities, the uncertainty components for (a*,b*) are more even in their effect – no component particularly dominates. Wavelength effects are generally small, except for the green surface, where the signal is concentrated more towards the central part of the visible range and the overlap with the colourmatching functions is larger. Scaling uncertainties affect both a* and b* through the non-linear dependence on the Y tristimulus value. The uncertainty in L* is important in determining the accuracy of colour differencing. NMI TR 8 18 9.4 Combined Uncertainty for the Various Surface Colour Quantities Blue Blue y x Green Red Red Grey Grey 0.0000 v' u' Green 0.0005 0.0010 0.0000 0.0005 Uncertainty Blue Blue Green Green Red Red b* a* Grey 0.0 0.1 0.2 0.3 0.4 v* u* Grey 0.5 0.0 0.2 Uncertainty Blue Green Green h*(u*v*) h*(a*b*) 0.005 0.8 c*(u*v*) c*(a*b*) Grey Uncertainty 0.6 Red Grey 0.000 0.4 Uncertainty Blue Red 0.0010 Uncertainty 0.010 0.0 0.2 0.4 0.6 0.8 Uncertainty The uncertainty in chromaticities for all components combined is most evenly distributed over the different colours for (a*,b*), although even here the difference between uncertainty in a* and that in b* can be significant. Note that for a grey surface, hue is infinite and its uncertainty has no meaning. The main conclusions to be drawn for surface colour uncertainties are: Uncertainty in the base reference reflectance spectrum is relatively unimportant for routine measurements. It becomes a significant component for measurements of the highest accuracy, close to the primary standards. Uncertainties can vary strongly with colour, and may be significantly different for the members of a chromaticity pair. They should be estimated for each measurement. NMI TR 8 19 10 METHODS FOR SOURCE PARAMETERS AND UNCERTAINTIES 10.1 Dominant Wavelength Dominant wavelength of a source is given as that of the monochromatic locus on a 1931 (x,y) chromaticity diagram at its intersection by a line passing through the (x,y) chromaticity of the source and that of the equal-energy source (xE = yE = 0.3333). By equating gradients of lines of the point on the locus (xm,ym) and that of the source, we have ym y E y y E (58) xm xE x xE and we can find the point on the monochromatic locus by interpolating and searching on a grid. However, the gradients are discontinuous for x xE (which occurs between 550 nm and 555 nm for the monochromatic locus) and it is better to search for a minimum in the value of ym yE x xE y yE xm xE . This is carried out on a relatively coarse regular grid, typically 5 nm using the tabulated colourmatching functions x , y , z to calculate (xm,ym) at each grid point. To avoid finding the complementary wavelength, the search end is limited to 560 nm if xm xE and is begun at 545 nm if xm xE . Once the minimum is found, we fit xm,ym as quadratic functions, using the minimum point and the point either side. Using xm as an example, we have xm a bd cd 2 (59) where d is the wavelength referenced to that at the grid minimum point and scaled by the wavelength interval. Hence the values of d for the three points are –1,0 and +1 and it is a simple matter to find the coefficients a, b and c. We repeat this process to find a similar fit for ym. We now use Newton–Raphson iteration to refine the grid-minimum dominant wavelength to a more accurate one. We have (60) f ym yE x xE y yE xm xE dy dx df x xE m y y E m dd dd dd (61) with dxm b 2cd dd and a similar expression for the ym derivative from its quadratic fit. Beginning with d0=0, an improved value of d is found as d d 0 (62) f ; repeated df dd iteration rapidly finds d to the required accuracy. The dominant wavelength is then found by adding d to the wavelength of the grid-search minimum, where is the wavelength increment during the grid search. The sensitivity coefficients required to calculate the uncertainty of the dominant wavelength are obtained from Eq (62) and its ym analogue at the final value of dominant wavelength as d dx m x dd NMI TR 8 1 (63) 20 The uncertainty of the dominant wavelength is given by 2 (64) u (d ) d u 2 ( x) d u 2 ( y ) 2 d d u ( x, y ) x y x y Note that for near-monochromatic sources in the long-wavelength region where the Z tristimulus value is zero, x y 1 and the sensitivity coefficients have the same magnitude but opposite sign; for fully, positively correlated components the covariance term in Eq. (64) effectively cancels the first two terms and the uncertainty in dominant wavelength is small. 2 2 10.2 Source Correlated Colour Temperature Given (u,v), correlated colour temperature is defined as the blackbody temperature which minimises the distance d in (u,v) space from the blackbody locus (uT,vT), where d 2 (uT u ) 2 (vT v) 2 (65) At the minimum, du dv f (uT u ) T (vT v) T 0 (66) dT dT Although this expression is strictly true only at the minimum, we can use Newton– Raphson iteration to solve for T. The fitting is much more reliable in (u,v) space rather than directly in terms of temperature. The function f is then modified to dv g (uT u ) (vT v) T 0 (67) duT Given an initial estimate (T = 3000 K for practical cases encountered in photometry), we calculate uT , vT and their derivatives with respect to T. Then an improved value of T is given as g T T duT dT (68) dg du T where d 2 vT dvT dg 1 (vT v) duT duT 2 duT 2 (69) with dvT dvT duT d 2 vT d 2vT d 2uT , duT dT dT duT 2 dT 2 dT 2 Repeated iteration until the steps are below the required accuracy, yields the correlated colour temperature. This method is described in [5]. (70) The uncertainty in CCT is given by T T T 2 T 2 u 2 (T ) u (u )u (v) (71) u (u ) u (v) 2ruv u v u v At the minimum, f = 0 and implicit differentiation [6] yields the required sensitivity coefficients T f f T f f , (72) u T v v T u 2 NMI TR 8 2 21 where 2 d 2uT duT d 2vT dvT f (ut u ) ( v v ) t T dT 2 dT dT 2 dT 2 (73) and du dv f f T , T (74) u dT v dT The procedure to calculate the derivatives is as follows. Given the colour-matching functions xi , yi , zi at air wavelengths i , and the relative spectral radiance on the blackbody locus given by Planck’s law 1 (75) Pi c2 i5 (e n T 1) a i (where na is the refractive index of air), we have chromaticity values 4 xi Pi 6 yi Pi , vT uT ti Pi ti Pi (76) where ti xi 15 yi 3zi . Hence duT dT 4 xi Pi P P P uT ti i 6 yi i vT ti i dv T T , T T T t P dT t P i i i i (77) and 2 Pi 2 Pi u 4 xi uT ti 2 T 2 2 d 2uT T T T dT 2 t P i i t i Pi T (78) 2 Pi 2 Pi Pi vT 6 yi vT ti 2 ti T d 2 vT T 2 T 2 T dT 2 ti Pi The derivatives of the Planck function are Pi c2 Pi (79) c 2 T na iT 2 (1 e na iT ) and Pi 2 Pi c2 c2 1 (80) Pi c2 c2 2 T T T na iT 2 (1 e na iT ) na iT 2 (1 e na iT ) To be consistent with the CIE Illuminant values, we must use the value of c2 as 1.4388 10–2 m.K and ignore the correction for the refractive index of air. 10.3 Source Distribution Temperature Distribution temperature is defined as the value of T that minimises the integral 2 S ( ) 1 aP(t , T ) d NMI TR 8 (81) 22 where St ( ) is the spectral irradiance, and wavelengths are taken in the range 400 to 750 nm. CIE in its definition give the value of c2 as 1.4388 10–2 m.K and do not clearly define whether air- or vacuum-wavelengths are used. The current value of c2 is 1.4387752 10–2 m.K, and we can take na = 1.00029 to convert from the airwavelengths at which measurements are made to the vacuum wavelengths needed for the calculation of photon energies assumed in the derivation of Planck’s law. However to be consistent with the CIE Illuminant values, we must use the value of c2 as 1.4388 10–2 m.K and ignore the correction for the refractive index of air. Distribution temperature only has meaning for sources close to the Planckian locus. Woeger [7] demonstrated a general method for least-squares fitting and uncertainty estimation, based on a generalised Newton–Raphson technique [8]. Determination of distribution temperature was used as an example. In that example, he considered a weighted fit, where the weighting included correlations in the spectral points. The strict definition of distribution temperature requires an unweighted fit; the method below follows Woeger, but removes the weighting and includes the constant 1 in the model definition. If the spectral data are present at regular intervals, the integral can be written in sum form over the N spectral points 2 aS n 1 f n 2 (82) Pn (T ) where the normalisation constant a has been moved to the numerator for convenience. Note If the data are not present at regular intervals, the separation between points must be included in the summation – distribution temperature is not found by merely fitting a Planckian form to the spectral energy distribution. Minimisation with respect to a and T leads to the set of two equations (83) F M y TM 0 where f1 f1 a T f 2 f 2 M y a T (84) f n f n a T is a N 2 matrix of the derivatives of the function with respect to the parameters being fitted and f1 f M 2 (85) fn NMI TR 8 23 is a N 1 matrix containing the model form being fitted. The derivatives are P Sn n f n Sn f n T , (86) 2 a Pn T Pn P where Pn , n are given above in (75) and (79), respectively. T The generalised Newton–Raphson technique is an iterative procedure to solve Eq. (83) by expanding the equations F in a Taylor series about the initial point, retaining only first-order derivatives, then setting the value at the new point (a better approximation to the solution) to zero; that is (87) F(y y) F(y 0 ) J. y 0 Here J is the Jacobian matrix F J i (88) y j where i =1,2 for our two equation rows and j=1,2 for our two parameters. Hence f f f J fn n n n (89) y j n yi n y j yi where only first-order terms are retained, or (90) J M y TM y a It follows that, given the parameter matrix with some initial estimates, y 0 , T improved estimates are given by y y 0 (M y TM y )1 F(y 0 ) y0 (M y TM y )1 M y TM This process is iterated until a required accuracy is reached. (91) To calculate the uncertainties in the parameters a and T (input parameters y) from those in the spectral values Si (input parameters x) we need the sensitivity coefficients f f a f i T a T and . Given that i i , the sensitivity coefficients are S i S i Si a Si T Si carried in the matrix Q where M x M y Q and f1 S 1 0 Mx 0 NMI TR 8 0 f 2 S 2 0 0 0 f n S n (92) 24 is a diagonal sensitivity matrix with f n a . Hence Q My1M x where M y1 is the S n Pn non-unique generalised inverse of M y (= M y T M y M y T ). The uncertainty matrix 1 for the parameters y is given by U y QU x Q T M y1M x U x (M y1M x )T M y1M x U x M x TM y1T (93) M y1KM y1T The inverse is given by U y 1 M y1KM y1T 1 M y T K 1M y (94) where K is square and so we can calculate its inverse, form the last product and take its inverse to get the covariance matrix between the fit parameters. Distribution temperature is calculated using spectral values only in the wavelength range 400 to 750 nm. For a given set of input spectral values and their correlations, only the subset in the required spectral range is used for this calculation. Note Added in Revision 2.8 The variance–covariance matrix for highly correlated data is ill-conditioned to the point where the inverse matrices in Eq. (94) may not be calculated accurately in a routine procedure, particularly for large numbers of spectral data points. Also the inverse M y M y T 1 required for the sensitivity coefficients can not be reliably calculated. However, we can directly calculate sensitivity coefficients for the dependence of distribution temperature on the spectral value at each wavelength. The generalised Newton–Raphson method rapidly converges. Using the calculated value of distribution temperature found for the true spectrum, we increment each spectral value in turn by 1%, then recalculate the distribution temperature and hence the sensitivity coefficient as the change in distribution temperature divided by the change in spectral value. We can then propagate both random and systematic uncertainty components in the spectral value as quadratic and linear sums, respectively. Note that the linear sum may be negative because it contains the sign of correlation. For example, a background level added to blackbody curve reduces the distribution temperature. 11 REPRESENTATIVE EXAMPLES FOR SOURCE COLOUR UNCERTAINTIES 11.1 Typical Uncertainty Components The reference for colorimetry of broad-band sources is traced to spectral irradiance or radiance standards, usually derived by estimating the temperature of a blackbody. For the calculations here, a blackbody at 2800 K is assumed, with a temperature uncertainty of 1 K. All values of spectal irradiance are assumed fully correlated. The remaining components were set as in Table 2, as representing a routine measurement. NMI TR 8 25 11.2 Broadband Sources Representative calculations were made for a blackbody at 2856 K (CIE Illuminant A), one at 5500 K and for the distribution CIE Illuminant D. 11.2.1 (x,y) by Component 0.0010 0.0010 Illuminant D 0.0008 x y 0.0004 Uncertainty Uncertainty x y 0.0006 0.0005 0.0002 x y 0.0006 0.0004 0.0002 0.0000 0.0000 1 2 3 4 5 6 7 8 9 0.0000 10 1 2 3 4 Component 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 Component Component Again the contribution of the reference is small – systematic effects in the measurement process dominate. The effect of offsets in the transfer (component 7) is significant for Illuminant A, because its spectral power is low at the shorter wavelengths. Similarly, an offset in the 2800 K reference (component 3) is significant at the higher source temperatures. Random errors are also significant in all cases. Because the distributions are relatively smooth, wavelength errors have small effect. Scaling errors (component 6) do not affect (x,y) chromaticity. 11.2.2 (a*,b*) by Component 0.15 5500K blackbody Illuminant D a* b* Illuminant A a* b* 0.2 a* b* 0.4 0.10 Uncertainty Uncertainty 0.4 Uncertainty Uncertainty 0.0008 0.0010 5500K blackbody Illuminant A 0.05 0.2 0.00 1 0.0 1 2 3 4 5 6 7 Component 8 9 10 2 3 4 5 6 7 Component 8 9 10 0.0 1 2 3 4 5 6 7 8 9 Component Similar comments to those for uncertainty in (x,y) apply. Differences between the uncertainties for the chromaticity pair are more marked than for (a*,b*) than for (x,y). An important point to note is that uncertainties in a* and b* are dependent on the luminance of the source. The distributions for Illuminants A and D used for these calculations are those of the CIE tabulations, scaled to have a common luminance value and their L* values are 100. The scaling of the 5500 K blackbody spectrum was such that its L* value was 13.9 and the (a*,b*) uncertainties are smaller than for the other sources. While L*a*b*, and the other colour parameters normalised to the reference illuminant, provide better agreement with perception than the simple (x,y) or (u,v) chromaticities, they have little meaning based on the spectrum alone, as the values (and their uncertainties) depend on the luminance. They are useful quantities for displays, less so for general sources. NMI TR 8 26 10 10 11.2.3 Source Temperatures by Component 50 12 Illuminant A 6 4 2 5500K blackbody DT CCT 30 20 1 2 3 4 5 6 7 8 9 10 Component 20 10 10 0 0 0 DT CCT 30 Uncertainty /K DT CCT 8 40 Illuminant D 40 Uncertainty /K Uncertainty /K 10 1 2 3 4 5 6 7 Component 8 9 10 1 2 3 4 5 6 7 8 9 10 Component As the temperature of the source rises away from that of the reference 2800 K, the systematic offset in the reference signal dominates the uncertainty. Contribution from the base reference uncertainty is small. The Planckian distribution is non-linear, and offsets in the transfer are significant. Distribution temperature is only meaningful for sources whose spectral distribution is approximately Planckian; this is not true for CIE Illuminant D. Offset in the transfer is significant in all cases – elimination of stray light is important for accurate measurements. Systematic scaling factors (component 6) have no affect, as expected. While the different sensitivity coefficients (those for correlated colour temperature are weighted towards 555 nm) means that different components have slightly different effects, the overall uncertainty is similar for distribution temperature and correlated colour temperature. 11.3 Narrow-band Sources The most important narrow-band sources are light emitting diodes. Coloured LEDs have narrow-band emission spectra and their chromaticities are close to the monochromatic locus. Bandwidth for these sources is typically 35 nm (full width at half maximum) and a wavelength resolution of order 1 nm is required for the measurement of the spectrum; the typical 5 nm used for broad-band surfaces or sources leads to significant errors. Hence calculations were made using the components listed in Table 2, except that the reference was the same 2800 K blackbody with 1 K uncertainty used for broadband sources and the wavelength offset uncertainty was reduced to 0.02 nm. A Gaussian model with a sharpened peak [9] and 35 nm bandwidth was used to calculate LED spectra with peak wavelengths of 470 nm (blue), 524 nm (green), 585 nm (amber) and 605 nm (red). NMI TR 8 27 11.3.1 (x,y) by Component 0.0010 0.0010 LED y LED x 0.0008 470 nm 524 nm 585 nm 605 nm 0.0006 Uncertainty Uncertainty 0.0008 0.0004 0.0002 470 nm 524 nm 585 nm 605 nm 0.0006 0.0004 0.0002 0.0000 0.0000 1 2 3 4 5 6 7 8 9 10 1 2 Component 3 4 5 6 7 8 9 10 Component Here x and y uncertainties are plotted on the same scale in different graphs grouped by wavelength. For such narrow-band sources, the reference component uncertainty is negligible. Offsets in the transfer (stray light constant across the spectrum, or amplifier offset) dominate the uncertainty of both x and y. For the results shown here the wavelength range was restricted to 2 times the bandwidth about the central wavelength. If the range was extended to the full visible region, with the same offset uncertainty scaled to the peak signal levels, the transfer offset uncertainties totally dominate and the uncertainties for both x and y are of order 0.002, with that in y for the green led approximately 0.005. Systems for routine measurements are often programmed to record spectra for the full visible range, then perform a colour calculation. This is not good LED measurement practice, particularly if signals are weak (e.g. for fibre-coupled systems viewing a diffusing screen). It is far better practice to delete from the calculation ranges where the LED signal is known to be zero. Wavelength /nm 605 585 LED dominant wavelength 524 470 0.0 0.1 0.2 0.3 0.4 Uncertainty 11.3.2 Dominant Wavelength Uncertainty in dominant wavelength is calculated from that in (x,y). The increased value in the green comes not so much from the larger uncertainties in x and y there as the larger difference between them, because of the effectively complete correlation between the uncertainties of x and y. The values shown here are increased by a factor of 6 if measurements are made over the full visible wavelength range with the same uncertainty components NMI TR 8 28 12 CONCLUSION Systematic effects in the measurement system generally dominate the uncertainties of colour quantities. The uncertainties depend on position in colour space and on the reference spectrum and its uncertainties as well as the spectral power distribution of the source. They should be calculated for each measurement of a colour quantity. If we propagate uncertainties through those of the tristimulus values, the calculation of the systematic uncertainty components involves single sum expressions and not a full matrix multiplication with many terms. A simple 3x3 matrix calculation, or series of such calculations, can then propagate the uncertainties to those of any desired colour quantities. Uncertainties of colour quantities for a given effect scale with that of the effect. The representative examples given here can be used as a guide to uncertainties to be expected from given experimental conditions, in the different regions of colour space. 13 REFERENCES [1] Guide to the Expression of Uncertainty in Measurement (1993) International Organisation for Standardisation Colorimetry (1970) Publication CIE 15, International Commission on Illumination, Vienna JL Gardner (2003) Uncertainties in Interpolated Spectral Data J. Res. Natl Inst. Stand. Technol. 108, 69–78 EA Early and ME Nadal (2004) Uncertainty Analysis of Reflectance Colorimetry, Color Research and Application 29, 205-216 JL Gardner (2000) Correlated Colour Temperature – Uncertainty and Estimation Metrologia 37, 381–384 J Fontecha, J Campos, A Corrons, A Pons (2002) An Analytical Method for Estimating the Correlated Colour Temperature Uncertainty Metrologia 39, 531– 536 W Woeger (2001) Uncertainties in Models with more than One Output Quantity, Proceedings of the CIE Expert Symposium, pp12–17, International Commission on Illumination, Vienna WH Press et al. (1988) Numerical Methods in C: The Art of Scientific Computing, p 381, Cambridge University Press Y Ohno (personal communication) [2] [3] [4] [5] [6] [7] [8] [9] NMI TR 8 29