]\gt ; IX .
On the Proof of the Law of Errors of Observations .
By W. CROFTON , F.R. Received Naxch 24 , \mdash ; Read Apnl 2 , 1869 .
1 .
So much has been published upon the Theory of Errors , that some apology seems to be required from a new writer who does not profess to have arrived at any results which were unknown to his predecessors .
Nevertheless , so great , as is well known , are the difficulties of the theory , whether we seek to form a correct estimate of the principles on which it rests , or to follow the subtle mathematical analysis which has been found indispensable in reasoning upon them , that any contribution which tends to simplify the process .
, without weakening their logical exactness , will probably be considered of some value .
My object in this paper is to give the mathematical proof , in its most general form , of the law of single errors of observations , on the hypothesis that an error in practice arises from the joint operation of a large number of independent sources of error , each of which , did it exist alone , would produce errors of extremely small amount as compared generally with those arising from all the other sources combined .
Now this proof is contained in a process given for a different object , namely , POISSON'S generalization of LAPLACE'S investigation of the law of the mean results of a large number of observations , to be found in his 'Recherches sir la Probabilite des jugements , ' and which is leproduced in Mr. TODHUNTER'S valuable ' History of the Theory of Probability .
' It is obvious that we should altogether restrict the generality of the proof , confining it merely to a few artificial and conventional cases , if we were to suppose each source of error to give positive and negative errors with equal facility , or to assume the law of error ( even supposing it unknown ) to be the same for all the sources .
None of the processes , therefore , contained in the 4th chapter of the Theory Analytique des Probabilites ' are of sufficient generality for our purpose , though some writers have so employed them ; nor will the method apply here which LESLIE ELLIS has given in his emoir ' On the Method of Least Squares\ldquo ; ( Camb .
Phil. Trans. 1844 ) , based upon FOIJRIER'S theorem , on account of the assumption of equal facility for positive and negative errors .
rfhe proof which follows will be found , I think , of full generality , \mdash ; the only cases excluded being incompatible with the existence of the exponential law ( see art .
7 ) , and at the same time greatly simpler than POISSON'S , dispensing with his refined and difficult analysis* .
2 .
It is remarkable that the well-known exponential function which is now pretty 2 The length of this communication may seem at variance with the statement that the proof here iven is a inppkkrr one than those of fOrmer writers .
Still I think it be found to be on examination ; the length of $ the arises from fuUer explanations being given than usually the case .
I am persuaded that the doubts and misconoeptions which have prevailed extensively with relation to this subject have been in great part wiQned by the extreme brevity and scanty explanation of the great writers who have treated it .
generally received among mathematicians as expressing the law of frequency of single errors of observation , does not seem to have been distinctly given by any one of the three great philosophers LAPLACE , GAUSS , and POXSSON ( who may be called the founders of the Theory of Errors ) as being , in their opinion , the expression of that law .
It has been erroneously supposed , as LESLIE ELLIS points out , that GAUSS'S and LAPLAr'S proofs of the method of Least Squares depend upon that assumption .
It is true that GAUSS'S first method , in the ' Theoria Motus , ' does require it ; but he does not present that method as other than tentative and hypothetical : and later , in the ' Theoria Combinationis Observationum , ' he says , speaking of the law of single errors , " " plerumque incognita est As , however , this law of error seems in our day to have been adopted by general consent , some inquiry into the grounds on which its .
validity rests will be appropriate here .
And first I would remark that it can scarcely be maintained that any attempt hitherto made to establish this law independently of the hypothesis I have named in art .
1 has been successful .
We may pass by GAUSS'S proof in the ' Theoria Motus , ' which shows that the law must hold if we take as an axiom that the arithmetical mean of several observations is the most probable result .
Now this really is not an axiom , but only a convenient rule which is generally near the truth : thiswe see by considering any case in which we are certain that the errors do not follow the exponential law ; does the mind see here a priori that the rule does not give the most probable result ?
It seems certain that we should have just the same confidence in it here as in any case ; yet GAUSS'S proof shows that it does not give the most probable result* .
It should indeed be stated that GAUSS himself ( as might have been expected from that acute and accurate mind ) is very far from asserting the above assumption to be an axiom ; consequently he does not give his proof as more than bypothetical .
He only states that the rule is generally accepted\mdash ; " " axiomatis loco haberi solet hypothesis A method of remarkable simplicity was given by Sir J. HERSCHEL in a very interesting review of QUETELET'S ' Letters on Probability , ' which conducts to the same law of error by means of no or two bold assumptions ; but striking as the coincidence is , it can hardly be seriously viewed as a demonstration ; nor is it formally so presented by its distinguished author .
However , the methods both of GAUSS and Sir J. HERSCHEL are of great interest to the natural philosopher , as showing that certain a priori mathematical assumptions of a very simple kind lead to the same law of error which reasoning based on a study of the facts which surround us also points out as expressing , at least approximately , what generally does occur in rerum naturd : though we can see no necessity that the facts * See ELLIS , loc. .
p. 207 .
nburgh Review , July 1850 .
See a criticism by LBSLIII ELLIS in the Philosophical Magazine , xvu .
Ako boon ( Edinb .
Trans. vol. ) and TmONSON and TAIT ( Natural Philosophy ) , who speak more favourably .
M. 's ' Lettres ' will amply repay a perusal ; in connexion with our present inquiry , he points out that not only errors of observations , but the vanations of many other fluotuating magnitudes , such as the of men , the of the weather , \amp ; c. \amp ; om their mean values , seem to follov the same law .
If thim be the inference seems timate that these divergences ffom the mean types , or errors of herself , as thoy may be called , are produced in each case , not by one or two , but by a vast of hidden coexkting causel !
THB LAW OF ERRORS OF OBSERVATIONS .
should be so , it being quite easy to eonceive a different economy of nature in which no such accordance would subsist* .
It is possible a priori to conceive that the law of single errors of observation might be of any form whatever , varying with each kind of observation : how far it is true that in practice one general law will be found to prevail , is essentially a question of facts\mdash ; an inquiry , not into what might be , but what .
Now the hypothesis above mentioned , \mdash ; namely , that errors in rerum naturd result from the superposition of a large number of minuter errors arising from a number of independent sources , \mdash ; when submitted to mathematical analysis , leads to the law which is generally received ; as far therefore as this hypothesis is in accordance with fact , so far is the law practically true .
Fully to decide how far this hypothesis does agree with facts is an extremely subtle question in philosophy , which would embrace not only an extended inquiry into the laws of the material universe , but an examination of the senses and faculties of man , which form an important element in the generation of error .
Still , without pretending to enter on a demonstration of the truth of this hypothesis , a few reflections upon the facts , especially in the case of Astronomy ( which is par excellence the science of observation , and where accordingly the lessons of experience are the clearest and most complete ) , will , I think , at least convince us of its reasonableness in certain large classes of errors of observations .
Now if we attend to what has taken place in the history of astronomical observation , we find that the gross errors of the earlier observers proceeded mainly from three or four principal causes\mdash ; for instance , refraction , imperfect measurement of time , and the use of the naked eye in pointing to objects .
When these few capital occasions of error were removed ( at least approximately ) , refraction being discovered and allowed for , and the pendulum and telescopic sights introduced , it was found that observations at once attained a high order of accuracy , showing that the principal sources of error had been eliminated .
It would seem , in fact , that in coarse and rude observations the errors proceed from a very few principal causes ; and in this case , consequently , our hypothesis will probably represent the facts only imperfectly , and the frequency of the errors will only approximate roughly and vaguely to the law which follows from .
But when * The extreme simplicity of the exponential relation itself , whether considered as expressing the law of single errors , or that of the mean results of a large number of observations , as contrasted with the long and difficult methods by which it was established , has naturally led to several attempts to dispense with or simplify the latter in some the hypothesis we here adopt is taken as a basis ; but , so far as the present writer is aware , every process given , except POISSON'S , fails in generality .
In recent Memoir on the Law of Frequency of Error by Professor TAIT ( Edin .
Trams .
vol. xxiv .
) ( where , it should be stated , the learned author speaks with hesitation , and only gives his method as an attempt ) , it is assumed that each of the errors which are combined can be assimilated to the deviation from its probable value of the number of white among a given large number of baUs drawn from an urn , which contains white and black in a given proportion .
It is then shown ( as indeed is done in IAPIiAcn 's 3rd chapter ) that this error follows the exponential law .
Thus the proof only applies to the combination of a number of elementary orrors , each of which follows that law .
But it quite certain that many simple errors do not that law ; hence the method is ether deflcient in generality .
We cannot , however , assert this positively , if there is reason to believe that the error which arises from each pnncipal cause is itself a composite error , which certainly is often the case .
The ' error in time for astronomers , not content with the degree of accuracy.they had reao their researches into the remaining sources of error , they found that , not three or , but'a great number of minor sources of error , of nearly coordinate importance , began to reveal themselves , having been till then masked and overshadowed by the g1aver errors .
whioh had been now approximately removed* .
It was as if a small number of trees had been cut down , leaving an innumerable growth of shrubs and brushwood at their remaining to be cleared .
There were errors of graduation , and many others , in construction of instruments ; other errors of their adjistments ; errors ( technically so called ) of observation ; errors from changes of temperature , of weather , from sIiglit ular motions and vibrations ; in short , the thousand minute disturbing influences with which modern astronomers are familiar , and which it is superfluous to recapitula* here .
Many of these are known and allowed for , or eliminated , at least approximately , in practical astronomy ; still we seem to be justified in considering the error which remains as the result of a great numb of yet minuter errors , each iuconsiderable itself .
Thus a cursory view of the nature of astronomical errors , and the light which this throws on various cognate classes of observations , seeru to lead to the conclusion that the above hypothesis will be found to hold , generally , in the case of refined .
delicate observations .
No doubt much morewould be necessary to justify us in \mdash ; instance , is certainly not a simple error , but one resulting from the joint action of several causes , one or more of which we can conceive detected and allowed for , leaving the others in operation .
An error may thus anse from the superposition of only three or four component errop , which at first sight are of simple origin , but in reality represent each a group of minor errors ; and the hypothesis would then hold .
It is questionable whether , among the causes which in practice vitiate any observation , any simple error ever does enter , of considerable magnitude and importance as compared with the others combiued ; such , for instance , as would be the error produced in the time ( or through the time on some astronomical magnitude by the pendulum being of an inch too long or too short , every thing else being pretty accurate .
If it be said that ignorance or negligence might produce such a result , we answer that such negligence or ignorance would make itself felt in other ways also : one such error would not stand alone .
Isolated acts of neglect by a careful Would under the head of occasional errors , as explained further on .
It seems very difficult to discern , a priori , the nature of the errors incurred in estimating magnitudes the eye , or of errors arising from the imperfection of our enses , such as those incurred in poinhng to a star with the naked eye .
It is quite possible that such errors may anse each from several sources , though their natmek hidden from our view .
*A similar law to that mentioned above seems to prevail in many kindred cases .
Thus in uocessive improvements in artillery , machinery , \amp ; c. , in proportion as the goeater souroes of and inacouxecy are understood and remedied , the number of minor disturbing influences which are thus rendered and still vitiate the results , though to a less extent , increases rapidly .
We may evem traoo a sort of analogy here in various phenomena both of the moral and material universe , which ntly have no on point we are considering .
Thus the pnncipal wants of human nature , the necessariee of life in fact , are few ; and so long as these are supplied with culty , minor wants are roely felt , .
we re in communities : but when the greater wants are satisfied , the numbQr and variety of the ondary nquin of our nature visible in the multitudinous productions of oivihzed.lifa The whioh } in shortening human existence are very few in number ; but could they be extirpated , the number of mnor of nearly coordinate importance , which still would influence rate of mortality would be larga The statistics of cnme , and many other phenomena , would give rise to remarks of a similar natu1w. .
Tffl LAW OF ERRORS OF OBMnVATIONS .
$his absolutely ; thus it is not.enough for our purpose to show , could we do so conclusively , ' that each error in practioe is .
compounded of a large number of smaller errors ; we must .
show that they are }endent , at least for the most part .
Thus we may conceive one of the minute errors affecting an astronomical magnitude to be an error in the r.efraction proceeding from a rise in the general temperature , and another affecting the same observation to be an error of time arising from the expansion of the pendulum through the same cause ; now these two minute errors are not independent , and would have to he mathematicaPly combined in quite a different way from two that were imdependent ; and , indeed , such a change of temperature would influence the actual error of the observation in other ways also .
However , we may at least safely conclude that the hypothesis in question is not a mere arbitrary assumption , but a reasonable and probable ccount of what does in fact take place in the case of careful and refined observations .
3 .
In proceeding to submit this hypothesis to mathematical analysis , the minute simple errors which go to form the observed compound error will be assumed to follow each its 0W11 unknown law , expressed by different unknown functions of the utmost generality* : positive and negative values of each error will not be assumed equally possible ; on the contraly , the cases will be included , as obviously ought to be done , of minute disturbing ihfluences which always cause the observed magnitude to err in excess , and of others which cause it to err only in defect .
I will exclude all mention of the term and will consider solely the frequency or density of the error , viewed as a function of its magnitude .
Let any magnitude which has to be determined by observations a.fected with some one cause of error ( simple or compound ) be represented by the line ; \mdash ; let a large number of such observations be de , and let the observed values be represented by a number of lengths BA ' , measured from : it will be found in general that in the neighbourhood of A the line will be dotted over with a multitude of points the distance AA ' being the error in case .
These dots will at some point and end at some point ; which generally are on opposite sides of , but may both be at the same side .
Between and the dots will be distributed over CD with a valiable density : this density , at any point , will represent .
the frequency or of errors of magnitude AA ' .
If at every point we erect an ordinate representing the density at , we shall trace out a.locus or curve , whose equation we may call , taking A as origin , .
( 1 ) regard to the limits or amplitudes of errors , see note on art .
7 .
MR. W. CROFTON ON Tlffl PROOF This we may call the curve or function of Error* .
It is of course generally discontinuous , : as it is only to include values of between the points , D. The function stnctly speaking should vanish for all values of beyond and ; however , we shall not require any consideration of the analytical methods of expressing such functions .
If be the number of observations taken , and if we put , then as denotes the number of errors between and .
( 2 ) It is well to notice that , if be any constant , the equation really is the same function of Error as ( 1 ) , the number of observations only being altered .
4 .
In order not unduly to limit the generality of the investigation , it is necessary further to study the nature of the possible ways in which the dots we have spoken of as representing the observations may be scattered along the line , in the case of various unknown simple causes of error ; noting also what becomes of the function and the curve in each case .
And first , in many cases the dots will be distributed continuously along , thus giving a curve without gaps or intervals .
It is by no means necessary that this curve should descend towards CD at its two extremities more than in the middle ; in other words , the extreme values of a simple Error are not always less probable than the intermediate ones .
There may be cases where the extreme values are the most probable ; for instance , the Error occasioned by supposing a point fixed , which is in reality performing extremely minute and slow oscillations about its mean position But besides the cases of continuous distribution , there are others , not only conceivable , but which we may be sure do actually occur , in which a function or curve does not assist our conceptions , and we shall do better merely to consider the points or dots themselves .
There may be what is called a constant Error ; that is , some cause which gives the observatiOn always too great ( or too small ) by the same fixed minute amount : the distribution here is simply a group of coincident points somewhere on .
Or a certain cause may only admit of two or more definite values for the error ; the distribution will be two or more groups of coincident points , the numbers in each group being equal or unequal .
Again , an important class of Errors are those which may be called oceasional Errors , that is , produced by intermittent causes not always in In such a case , if observations be made , a certain number of them ( say ) are unaffected by the Error ; the remaining , made when the cause is in operation , we may suppose represented by dots continuously or discontinuously distributed ; we have then a group of coincident points at , besides a number distributed in somb way over Errors of mistake or forgetfulness , and many others also , are of this description .
*The word ' ' error\ldquo ; is sometimes used for shortness to express a sourco of error .
To avoid eonfusion may write it with a capital , when used in this sense .
an Error\ldquo ; mean a source of error .
or the assemblage of actual erroro ( or the curve or function symbohring which that source rucoe in a large number of trials , and which form a visible manifestation or ntabon of .
In error\ldquo ; will mean a uCular magnitude .
THE LAW OF ERRORS OF OBSERVATIONS .
6 .
If we alter the ordinate and abscissa of every point in the curve in a given ratio , changing the limits of the Error in the same ratio , we find the curve represented by , . . .
( 3 ) which may be called a similar Error to ( 1 ) or .
The number of observations will be different in the two cases , represented by the areas of the two figures .
We may find it convenient to suppose the number of observations the same ; if so ( 4 ) will be a similar function of Error to , the number of observations being the same for both , the limits of the error in ( 3 ) and ( 4 ) being 6 .
To find the function of Error resultingfrom the combination of a given Error whose equation is . . . . .
( 5 ) ( the limits ) with another independent Error , . . .
( 6 ) .whose limits are We shall do this most clearly by help of a geometrical construction .
Let the ( N ) values of the first Error be measured from A according to their signs along the indefinite line ; likewise measure the ( n ) values of the second Error along , where .
Take any two values , of the first , and of the second ; they give a value of the compound Error , to which will correspond a point of the plane , whose coordinates are .
The number of such points contained within the element , each point corresponding to a compound error , will be or being the element of the area .
Draw through a line UV equally inclined to the axes , then is constant along this line ; put , take , and draw parallel to ; take , then the number of points within the elementary parallelogram POpq will be Hence ) whole number of points between the parallels and ( that is , the of compound errors whose magnitudes lie between and ) will be The total number of compound errors thus obtained will be Nn ; however , for 01iformity , we will suppose the number of observations taken , affected with the compound , to be , the same as for ( 5 ) .
This will oblige us to divide by Thus if we represent the compound Error by a curve whose coordinates are , it will be .
( 7 ) Thus if we wish to find the Error resulting from the combination of the two whose equations are we have from formula ( 7 ) denoting numbers observations ) , whence Hence it is easy to see that if any number of Errors of the forms \amp ; c. be combined , the resultant Error will be . .
( 8 ) .
Expandin in formula ( 7 ) , it becomes THE LAW OF ERRORS OF OBSERVATIONS .
where $ , \amp ; o. , being the mean value of the Error , its mean square , its mean cube , \amp ; c. 7 .
In the problem of finding the law of error resulting from the superposition of a great number of Errors , each of very small importance by itself , we will consider each component Error as the diminutive of some Error of finite importance* ( see art .
5 ) .
Thus if be some possible finite Error , and we reduce its dimensions in the ratio , where is infinitesimal , the diminished Error will be ; and if the mean value , mean square , mean cube , \amp ; c. of the former be called , . .
, it is easy to see that the same means , the reduced Error , will be .
Now adoptin the usual axiom that no function can represent a finite Error unless , . . .
are finite , it follows that the mean cube , mean 4th power , \amp ; c. of the * Thus all conceivable cases of Errors whose extreme limits , or amplitude , are very ] , are contained in the above method of proof ; also those small Errors which , though their extreme amplitude be not very small , are merely possible finite Errors ( of great or infinite amplitude ) on a reduced scale .
It is necessary , however , to observe , in examining the nature of all the minute simple Errors which our hypothesis in its generality comprises , that there are cases quite conceivable , and involving no absurdity , of simple Errors of trivial or infinitesimal importance which come under neither of these categories , and to which the method in the text will not apply .
To give a simple instance , imagine an occasional source of Error , which rarely operates , but which , when it does , gives a fixed finite error ( thus we may conceive an observer to mistake , once in a thousand times , the succeeding division of his instrument for the true one ) .
Let this happen on an once for times that the cause is not in operation ( being supposed very great ) ; then the mean value of the Error is , its mean square is , \amp ; c. It is therefore of infinitesimal importance whether , with LAPLACE , we estimate the rtance of an Error by its mean value ( irrespective of sign ) , or , with GAUSS , by its mean square ; but as its mean \amp ; c. cannot be rejected in comparison with the mean square , the above analysis cannot be applied to it .
Minute simple Errors of such a description must then be excepted from those which are supposed to enter into the composition of the actual errors of observations .
If an appreciable number of them did cntcr , the received exponential law could not hold for the compound Error .
Thus were we to combine a number of small Errors of the nature of the simple instance just cited , the resultant Error would be of a discontinuous nature , represented by groups of coincident points , with finite intervals between them .
Though it is necessary clearly to understand that the full generality of the hypothesis is restricted by the exceptions explained in this note , yet there seems every reason to uppose that such cases are too rare in practice to cause any sensible deviation from the exponential law of error , the great majority of the minute component Errors which jointly affect any observation in rerum naturd having each , it is natural to suppose , a very minute range or amplitude .
diminutive Error may be rejected in comparison with its mean .
We infer , therefore , represent any Error of indefinite amplitude , and if a Error , , of indefinitely small importance as compared with it , be superposed , the resulting compound Error will be represented the equation , ( 9 ) where are inJinitesimal constants , viz. the mean value of the new Error and the mean value of its square , the number of observations being supposed the same for the Error ( 9 ) as for the former , .
If we now conceive ) in the above to be one of a large number of independent infinitesimal Errors , and to be the compound finite Error resulting from the combination of all the others , we from ( 9 ) that each elementary Error affects the law of the combined Errors in a manner which only involyes the mean value of the elementary Error , and ( ) its mean square .
But if this be so , we may , for our present purpose , substitute for any other Error whatever which has the same mean value and mean square ( provided of course its mean cube \amp ; c. may be neglected in comparison with its mean square ) .
We may therefore for our purpose replace by , ( 10 ) which fulfils these conditions .
Likewise , if there be another elementary Error whose mean value is and mean square , we may replace it by and so on , for all the elementary Errors .
Hence ( see equation 8 ) the Error compounded of any number of them will be cannot neglect the mean square as compared with the mean 1st power , as the latter is the algebraical sum of a number of positive and negative elements , which sum may be of any amount , however small , and may sometimes vanish altogether ; whereas the former is the sum of a number of positive elements , and therefore cannot vanish .
If the new Error be such as to give any discontinuous distribution of points ( see art .
4 ) , it is easy to satisfy ourselves , by the method of art .
6 , that the above still holds good .
In fact , if the values of the new Error be , . . . .
, we shall have , instead of the formula ( 7 ) , which is true in all cases , whether the distribution be continuous or discontinuous , or a mixture of both ; and hence the formula ( 9 ) ] follow .
This suggestion is due to Professor J. C. , one of the Referees charged by the Royal Society with the duty of reporting upon the present Paper .
The remainder of the proof , which was of a different nature in the Paper as originally presented , is much simplified thereby .
THE LAW OF ERRORS OF OBSERVATIONS .
We conclude therefore that if a great number of minute independent Errors be hned , and if we write of mean Errors , quares orrors , . . . .
( 11 ) the resulting funchon of Error wilt be .
( 12 ) The Probability of an error being found to lie between and is of course If positive and negative errors.in the observation are equally probable , as generally can be secured in practice , at least approximately , then ; that is , the sum of the mean values of the elementary component rors vanishes , and the Probability is expressed by the usual value If we calculate by integration from equation ( 12 ) the mean value of the composite Error ( or , as GAUSS calls it , the constant part of the Error ) and the mean value of its square , we shall find Mean Error of mean values of component Errors , Mean Square of Frror We have thus a verification of the correctness of our analysis , as the same results may be found from independent algebraical computation 8 .
Considering the celebrity of the question , it may not be supelfluous to show how the result might have been obtained without any antecedent knowledge of the peculiar property of combination of the Errors in equation ( 8 ) .
* We may observe that is always positive ; for if we take any set of numbers , positive or negative , the mean of their squares is always greater than the square of the mean ( see TODHUNTER'S ' Algebra , ' p. 407 ) .
Therefore , also \amp ; c. Consequently This expression will be found to with POISSON'S final result in the memoir already cited .
where each of the quantities \amp ; c. may take any number ( different for each quantity ) of different independent values , adopting for shortness the symbol for ' the mean value of it is not difficult to prove , by elementary algebra , that , or Let us suppose ffi the infinitesimal simple Errors which it is propoaed to combine to be successively superposed upon some assumed function of Error ; then by equation ( 9 ) the new function arising from the first of them will be , .
If another be now superposed upon this , we shall have , and finally the function arising from the superposition of all the given Errors upon the assumed Error will be . . .
( 13 ) But as are infinitesimals , we have , retaining the square of Thus ( 13 ) wilk become , or , adopting the notation ( 11 ) . . . .
( 14 ) 9 .
Let us now take as the assumed function of Error ( 15 ) ( where is the number of observations ) , and imagine the whole given system of small Errors superposed upon it ; the resulting function is Now by a theorem in the Differential Calculus* , ; *This theorem , which is new to the present writer , may be proved in various ways .
Thus if we put and differentiate with regard to , we have : again , ; we thus obtain the partial differential equation THE OF ERROBS OF OBSERVATIONS .
hence Now we may here assume as small as we please* , \mdash ; that is , we may assume the Error ( 15 ) upon which the given system was superposed , to be of as small importance as we please .
We conclude , then , rejecting this Error altogether , that a system of very small Errors , when combined , give for the resulting function of Error as before .
the integral of which is .
To determine the arbitrary function , we remark that if hence Another proof may be obtained by employing POISSON'S ingenious transformation ( Traite de Mecanique , tom .
p. 356 ) , which gives *In order that we may retain the three first terms only in the expansion it is necessary to show that and the succeeding differential coefficients are not infinite .
Now they generally will be infinite in the case where is an infinitesimal Error , as win be of the form , where is infinitesimal ; but in the case where we may take as small as we please , and yet retain only the three first terms above , becanse the differential coefficients of do not here become infinite ; in fact it is easy to see that any differential coefficient will consist of a sel.ies of terms of the form : now by the rules in the Differential Calculus for evaluating indeterminate forms , this quantity tends to zero as diminishes .

