64 Dr. W. F. Sheppard .
Tor convenience , the results of this paper and its predecessor are here recapitulated .
Gas .
Weak component polarisation ( strong component = 100 ) . .
Argon 0-46 Hydrogen 3*83 Nitrogen 4-06 Air 5 -0 Oxygen 9-4 Carbon dioxide 11 -7 Nitrous oxide 15 -4 Helium Not definitely determined , but \lt ; 6-5 The total intensities of light scattered by helium and by air have been* compared .
The ratio found was 0-0170 .
Considering the difficulty of the experiment , this is considered to agree within the limit of error with the ratio of the squares of refractivities , which is 0-0144 .
Reduction of Error by Linear Compounding .
By W. F. Sheppard , Sc. D. , LL. M. ( Communicated by Prof. E. T. Whittaker , F.R.S.\#151 ; Received April 23 , 1920 .
) ( Abstract .
) If U\ , u2 , us , ... are observed values of a single quantity U\#151 ; if , in other-words , there are observed values of U containing errors u\\#151 ; U , \#151 ; U , us\#151 ; U , ... \#151 ; , we may take as an approximate value of U any function of then 's which , if each of the us in it were replaced by U , would itself become identically U. Usually this function is a linear compound ( linear function ) of the u 's , and is called a weighted mean : and we regard as the best , weighted mean that in which the constants of the function are chosen so that it shall have the least possible mean square of error .
The principle-applies also when the w 's , instead of being observations of a single quantity U , are observations of a set of quantities Ui , Us , U3 , ... ; with the modification* that the effect of replacing ( say ) ua by a function of the s will be that , if the U8 in the function were put equal to the corresponding U 's , the function ; would not become identically equal to Ui , but only approximately equal to^ Reduction of Error by Linear Compounding .
65 it .
In the particular case in which the differences of the s successively diminish , a permissible form of the function is the sum of and a linear compound of the differences of the u'\s of sufficiently high order .
The problem considered in this paper is the slightly more general one of improving a quantity by adding to it a linear compound of certain other quantities , called auxiliaries ; the improved value being taken to be the value of this sum when , by suitable choice of the coefficients of the auxiliaries , its m.s.e. ( mean square of error ) is a minimum .
Allied to this , and leading to a similar solution , is the problem of fitting to a set of observations ux , u2 , w3 , ... , by the method of least squares , a linear compound of a smaller number of functions of 1 , 2 , 3 , ... .
with coefficients to be determined .
The most simple case is that in which the errors of the w 's are independent \#151 ; or at any rate the m.p.e. ( mean product of errors ) of every pair of different \#166 ; ids is 0\#151 ; and the rn.ss.e .
are all equal to 1 .
The errors are then said to form a standard system .
In two papers published a few years ago , on " Reduction of Errors by Means of Negligible Differences " and " Fitting of Polynomial by Method of Least Squares , " I investigated the problem for this class of cases , the auxiliaries being taken to be the differences of the s beyond those of a certain order .
The scope of the present paper is wider , and the methods also are simpler .
The simplification is effected in two ways : by brief statement of general theorems , and by a theory of conjugate sets of observations .
The fundamtntal theorem for reduction of error is that the m.p.e. of the improved value and each of the auxiliaries is zero .
This leads not only to other general theorems , but also to simplification in dealing with particular problems .
Suppose , for instance , that we require the m.s.e. of an improved value , the m.ss.e .
and m.pp.e .
of the original value and the auxiliaries being known .
If we construct the m.s.e. in the usual manner , the terms will form a double series , and will involve all these m.ss.e .
and m.pp.e .
But the above theorem shows that the m.s.e. ol the improved value is equal to the m.p.e. of the improved value and the original value , so that we have only to deal with a single series , involving the m.s.e. of the original value and its m.p.e. with each of the auxiliaries .
In the case of the standard system , mentioned above , the m.s.e. of each u is 1 , and the m.p.e. of each pair of u 's is 0 .
For other cases , we construct a somewhat analogous system by taking another set of quantities y\amp ; ... , equal in number to the u 's and connected with them by linear relations , and such that it we place the two sets in linear correspondence , thus :\#151 ; iii , u % , w3 , ... , y\\gt ; y%\gt ; . . .
, W\gt ; L. XCV1II.\#151 ; A. P Reduction of Error by Linear Compounding .
the m.p.e. of corresponding quantities of the two sets will be 1 and the m.p.e. of quantities which do not correspond will be 0 .
Sets so related are called conjugate sets .
In the case of a standard system the are identical with the us , so that the latter may be called a self-conjugate set .
If there is a second set of quantities ... , also connected with the u 's by linear relations , there are converse linear relations between the conjugates of the S 's and of the % 's .
In particular , if the S 's are successive differences of the u 's , and if the conjugates of the S 's are \lt ; r2 , ... , the o- 's are sums , of successive orders , of the s. The improved values of any specified S 's , using the remaining S 's as auxiliaries , are linear compounds of the cr 's which correspond to the former .
Hence , if the u 's are a self-conjugate set , and the S 's are their differences , the improved values of any linear compounds of the u 's , using certain S 's as auxiliaries , are linear compounds of the cr 's which correspond to the other S 's .
Since the moments of the u 's , like the cr 's , are linear compounds of successive sums of the u 's , this means that the improved values are linear compounds of the moments ; and it explains the appearance of the moments in this connection .
The earlier part of the paper deals with the main theorems relating to conjugate sets and to reduction of error , and with the relation of reduction of error to fitting .
Certain general formulae are then established , by means of which the solution of the problem for any particular type of data can be completed if the values of certain quantities can be found .
The paper concludes with the application of the general formulae to the particular case of the self-conjugate set .
Some of the formulae thus obtained are , of course , equivalent to formulae obtained in the two earlier papers ; but the methods of these are to a large extent superseded by the more general and more compact methods of the present paper .
t

