Z7S 3e UC-NRLF $C IbS 112 v^ '^..^-^ ON THE USE OF EQUIVALENT NUMBERS IN THE METHOD OF LEAST SQUARES. BY GEORGE P. BOND, A.M., ASSISTANT AT THE OBSERVATOBY OF HARVAKD COLLEGE. [From the Memoirs of the American Academy of Arts and Sciences, New Series, Vol. YI.] CAMBRIDGE: METCALF AND COMPANY, PEINTEBS TO THE UNIVERSITT. 1856. ON ^^/y :/^^ /' THE USE OF EQUIVALENT NUMBERS IN THE METHOD OF LEAST SQUARES. BT GEORGE P. BOND, A.M., ASSISTANT AT THE OBSEEVATORY OF HAKVAED COLLEGE. [From the Memoirs of the American Academy of Arts and Sciences, New Series, Vol. VI.] CAMBRIDGE: METCALF AND COMPANY, PBINTEBS TO THE UNIVEESITT. 1856. ON THE USE OF EQUIYALENT NUMBERS METHOD OF LEAST SQUAHES. One of the most important applications which has been made of mathematics to investigations in physical science has for its object to ascertain the best manner of combining data affected by unknown errors of observation, so that the probable effect of these errors shall be the least possible. The method of least squares proposes to accomplish this, by reducing to a minimum value the sum of the squares of the out- standing errors, and, by conforming to this single criterion, to fulfil the condition, so desirable in the prosecution of thorough and exact research, of reducing to its least possible amount the influence of errors in the data employed. The investigations here presented have been entered upon with the design of deter- mining the degree of numerical exactness proper to be observed in making use of the method of least squares, in order to secure its peculiar advantages with the least out- lay of labor. Some detail in the discussion seems to be called for from the prevalence of a practice, almost universal among computers, of adhering to the letter of the method of least squares with a strictness which implies a misapprehension of its true spirit. It is impossible to adduce any valid reasons to justify such a course when it must be fol- lowed at a serious expense of time and labor in the computations. It has not escaped the observation of Gauss, in his original exposition of the method, that some freedom of interpretation may be allowed when its theoretical results are applied in practice, as the following passage, referring to the solution of equations by least squares, will show : — I/) 9 1.5 9 4 ON THE USE OF EQUIVALENT NUMBERS " When the number of functions or equations proposed for solution is considerable, the computations become laborious, the more so from the circumstance that the co- efficients by which the primitive equations are to be multiplied are almost always com- plicated decimal fractions. If it is not thought worth the trouble in such a case to calculate the products with exactness by means of logarithms, it will generally be suf- ficient to substitute for them (i. e. for the multiplying factors) more simple numbers differing but slightly from them." * In his subsequent researches, it does not appear that Gauss has given any further development to the suggestion here put forth. Indeed, the introduction of modifications of a like nature, however desirable in a practical point of view, would have deprived a purely theoretical discussion of much of its elegance and symmetry. Yet the passage above quoted lends the support of the highest authority to the leading proposition which we shall have occasion most frequently to insist upon ; namely, the propriety of allowing some relaxation of theory in applying the calculus of probabilities to the dis- cussion of data affected by ordinary errors of observation, whenever the modification conduces to convenience and the saving of labor at the sacrifice of no appreciable ad- vantages. Even an unqualified admission of the superior probability of results which exactly fulfil the criterion proposed in the method of least squares, does not relieve us from the necessity of restricting it to examples which never actually occur, that is, if the ques- tion be made a rigorous one ; f — to such, for instance, as involve the discussion of obser- vations which are entirely free from unknown constant errors, or errors foUomng any law of facility which does not imply the assumption that the mean error is proportional to the square root of the mean of the squares of the individual errors. But we know that this proposition, which lies at the foundation of the whole subject, is not suscepti- ble of absolute demonstration by any process of mathematical reasoning. Further than this, we know from constant experience that the law of distribution of errors recognized in the method of least squares practically faUs, in extreme cases, both for very large and for very small errors. If any illustration of the failure of the assumed law be needed, it will be found in the familiar instance of computing by it the probable error of the arithmetical mean of a very large number of observations, where common sense assures us that the theoretical probable errors of the result are invariably smaller than they should be. Why, then, should an implicit adherence to its minutest details be required as essen- * Theoria Motus, § 185. t Theor. Comb., % 17. IN THE METHOD OF LEAST SQUARES. 9 tial to its successful application, or to the attainment of all the advantages which its employment may confer upon the discussion of any practical problem ? It is true that no other system can be proposed which is free from similar objections, or which can be mathematically demonstrated to be exclusively the best, without quali- fication, and therefore the arguments above stated are of no force whatever, if employed as reasons for the rejection of the method of least squares. They nevertheless greatly weaken the position of those who Avould insist upon a strict compliance with its pre- cepts, and effectually preclude all alignments of a purely theoretical character in support of such a course. Still it is desirable that the force of any objections which may be made to an attempt to modify the theoretical conditions for eiFecting the most favorable combination of equations should be appreciated at their true value. We therefore pro- ^ pose to show that the spirit of the method of least squares, rightly apprehended, in reality rather invites than discountenances a liberal construction of its rules. Admitting that the best possible solution is attained when the sum of the squares of the outstanding errors, represented by S2, is a minimum, it is evident that /2 is a minimum relatively to the manner in which the original equations have been treated. And since the peculiarity of the solution consists in the employment of a system of factors, a, a, &c., by which the original equations are multiplied before combination, the first differential coefficient of fl relatively to either of these factors, in the case of the least-square solution, must have the value for each factor, dJl a a When, therefore, the factors are varied by small amounts, Ba, B a, &c., the conse- quent variations of fl developed in a series, will contain only terms multiplied by the second and higher powers of 8 a ; or, in general terms, if we deviate from the exact (1.) precepts of the method of least squares hy small variations of the first order, we shall fail to satisfy its fundamental criterion hy small terms of the second order only. Looking thus at the most elementary principle of the method, we find a warrant for some degree of liberty in applying it, — a liberty which we can scarcely hesitate to avail ourselves of, if we further consider the peculiar circumstances attending its actual employment in the discussion of data furnished directly by observation. Among its first requirements is the assignment of weights to the original observations ; but it is one which it is not possible to fulfil correctly, for we are provided neither with a theory nor with data for the purpose. All that can be done is to accept, as indices of the relative value of the different observations, certain numbers depending either proxi- mately or remotely upon no other authority than the mere exercise of the judgment 2 b ON THE USE OF EQUIVALENT NUMBERS alone. No one can pretend that this is a process susceptible of strict accuracy ; yet an error here is as fatal as if we had disregarded any other of the precepts of the method. This step being an arbitrary one, although one of fundamental importance, we may properly appeal to it as a precedent for the modification of others suggested by con- siderations of convenience, though they may not, like this, be justified on the plea of actual necessity. In this view of the subject, we find support for the modification suggested by Gauss, in the passage we have quoted above. Each of the complicated factors which it is there proposed to simplify is itself a product of two other factors, one of which is the weight of the equation under treatment ; if one of these, that is, the number representing the weight, is erroneous, the product is of course errone- ous, with whatever accuracy the other is expressed. Again, as a matter of convenience, it is usual to express the conditional equations proposed for solution in a linear form, by reducing the indeterminates to small quanti- ties and neglecting the terms multiplied by their second and higher powers, and to con- struct, from them normal equations, as they are called, previously to applying the method of least squares. Both of these may be practices perfectly allowable under the circumstances, but since they are almost always theoretically incorrect, their admis- sion is a virtual relinquishment of all pretensions to a rigorous course of computation, and cannot be compensated for by any subsequent refinements. We will now proceed to examine the limits of accuracy appropriate to the arithmeti- cal operations required in the combination of conditional equations by the method of least squares, and afterwards to develop in detail some proposed modifications of that method, having for their object the reduction to its minimum value of the amount of labor requisite for its successful application. It is scarcely necessary to remark, that the subject .is plainly one which is in its nature somewhat vague and insusceptible of rigorous treatment, though it is at the same time interesting from its practical bearings. If no very precise or definite rules for regulating the degree of numerical exactness suited to the discussion of any given problem can be arrived at, it may still be of service to point out the principles which ought to guide the computer in the choice of such limits as shall perfectly meet all reasonable requirements of accuracy, without imposing upon him the unprofitable labor of multiplying the extent and difficulties of calculation, to no useful purpose, and without the remotest prospect of sensibly improving the real value of the results. Let us suppose a series of equations, a X -\- b y-j- -\- m =e, a' X -|- J' y -j- -\- m' = e'l IN THE METHOD OF LEAST SQUARES. 7 in which m is the element demed from observation, and e the unknown error of the equation, to be solved by the method of least squares, giving for x the value x, with its probable error, £, obtained from a compai-ison between the observed and the computed values of m, after substituting x, y, &c. in the primitive equations. If Xq be the true value of x, we may represent by x a quantity such that it is an even chance whether Xq — x is comprised between the limits s ~\- x and f — x. The magnitude of the limit defined by x has an evident relation to the question how far the simplification of the arithmetical processes may be carried without detriment to the results. For instance, the solution of the above equations may be repeated with small varia- tions from the process at first applied, giving for x a new value Xi with a probable error E^ , differing but little from a. If we were in entire ignorance of the relative amount of the probable errors f, and s, there would be no reason at all for giving the preference to x rather than to Xi . If only the single circumstance were known that fi exceeded £ by a given small amount, we should be equally at a loss, while the value of s remained unknown, to state the relative weight of Xi compared with X, and should, in fact, be again obliged to resort to the hypothesis that £ and fj were sensibly equal. And in general, the greater the uncertainty of f, or, in other words, the larger the value of x, the less reason would there be for excluding from competition with X any other determination of x, such as x^, of which the probable error e^ dif- fered but little from s. In order to employ the limit x as here proposed, its value must be known before the computations have reached an advanced stage. That this is not ordinarily practicable will readily appear. On the other hand, it must be left entirely to the judgment of the computer to decide as to the precise manner in which x is to be applied in limiting the allowable amount of difference fj — £. Objections of a similar character apply equally to other standards which might be proposed for the same object. As has before been remarked, the question must be treated, if at all, upon a somewhat arbitrary basis, and we must be content with sug- gestions addressed to the judgment or common sense of the computer, in cases where no fixed rule is admissible. Viewed in this light, there will ordinarily be no difficulty in recognizing the point at which there will be danger of compromising accuracy in the attempt to simplify the computations, nearly enough at least for practical purposes, if we are prepared to admit, at least in its general spirit and tendency, the truth of the following propo- sition : — O ON THE USE OF EQUIVALENT NUMBERS The application of the method of least squares to the discussion of observations of physi- cal phenomena, with the exception of a few special cases of rare occurrence, requires the (2.) use of such numbers only, in the arithmetical processes peculiar to it and characteristic of the method, as may be designated by one of the numerals 0, 1, 2 9, or of the fractions — , — —, or by a product of one of these numbers by an integral power of 10. An idea may be formed of the amount of the intentional errors occasioned by these substitutions, by noticing that if by N is represented any number whatsoever, and by N' a number chosen from the proposed series which most nearly coincides with N, we shall have N— isr 1 /„ , The maximum value of — — := - nearly. \o.) iV " JV — iV' 1 The probable value of — zr^ — <^ — . Before proceeding to a detailed investigation of the consequences of the changes pro- posed, it will be useful to point out the degree of insecurity attaching to the values which must ordinarily be adopted to represent the probable error of x ; the different sources which may be supposed to contribute to the increase of s ; and their relative importance in connection with the question of the comparative accuracy of the two results X and Xi. s may be referred to the combined influence of two mutually independent errors r) and rj, T) being the probable value of Xg — x which would result from the errors of observation alone, supposing the theory of the method of least squares and its applica- tion to the data to be rigorously exact, and V the probable amount of error in x having its origin in errors necessarily committed in the discussion of the observed data, sup- posing the mode of discussion, although the best practicable, to fall short of strict conformity with the theory. tjI represents the value of rf when the same data have been reduced, by a process made intentionally still less exact, to a small extent, both in its theory and in its arithmetic, than that which gives the error •??'. tjI will bear to Xi a relation similar to that which ■>/ bears to x. j; cannot be completely eliminated, so long as the errors of observation remain unknoAvn, by any treatment, and the same may be said of rf ; but 7)1 can always be reduced to its least limit, •»?', by suitable refine- ments of theory and of computation. In view of the fact that ij and rj' must have always sensible, but very uncertain values, it will be of but little consequence that ij'i should be reduced to its utmost limit without regard to the labor and inconvenience which it may cost. At all events, the attempt will be ineffectual as a means of im- proving the substantial accuracy of the results, as we shall presently see. IN THE METHOD OF LEAST SQUARES. 9 Since ■;; is independent of ij and Vi » we have, assuming ?;' to be the least attainable value of 9;'i, If rl' be used to designate the probable value of x — x^ vphich would result from small intentional deviations from that treatment of the data which is recognized to be the best, we have (4.) ,;a = ,'^ + ,"2, ,2 ^,,2 + ,'3^ ,,2 = ,2 ^ ,'2 _|_ ,"2. As regards the uncertainty of £, some estimate of its extent may be obtained in the following manner. If it is an even chance that the error of which the probable value is 17 is comprised" somewhere between the limits ?? -f- \ and 17 — X, «? having been derived from compari- sons of a given system of equations with observation, the number of individual equa- tions thus compared being represented by n, and the number of unknown quantities entering into them by «', \ may be found from the expression* (5.) X = 0.477 ^/n — nf Any value of « — w' less than 100 gives ^ 21 ' The scale of substituted numbers (2) admits, as we have before stated, of represent- ing 17 within the probable amount of — ??; hence, for any value of ?i — «' less than 100, the series will afford numbers representing t? with a probable error less than \. A slight examination M'ill show that a similar remark applies still more decisively to £. The considerations which oblige us to attribute a sensible value to ij are too many and too obvious to require to be specified in detail. It will be sufficient to cite one or two which have already been alluded to. The existence of unknown constant errors in the data will render the application of the method of least squares, strictly speaking, inexact. From this source »?' will inevitably acquire some influence. Again, the un- certainty incident to any attempt to assign to the original data their proper relative weights, will have a similar effect. No process more loose and arbitrary can well be conceived, than that by which the relative precision of the elements afforded directly by observation is graduated. Yet, imperfect as it is, improvement in this particular is scarcely to be hoped for. Exact conformity with a theory which requires a previous knowledge of the relative weight of observations is quite impossible. • Gauss, Zeitschrift fiir Astr., B. I. Theor. Comb.,§ 40. 10 ON THE USE OF EQUIVALENT NUMBERS At the same time, theij, that the existence and influence of i?' are admitted, its amount is altogether uncertain, to an extent suiRcient at least to make the uncertainty of e which is dependent on that of 17 and 7?' not less in proportional amount than that of ij ; consequently we shall obtain from (5) the expression (5 a.) x> 0.477 ■v/n — n' by which to measure the uncertainty of e. If n — n' < 100, (55.) K>^.. When £ is represented by a number chosen from the series (2), the probable error of the representation is, by (3), less than — £ ; in other words, it is more than an even chance that this number will fall within the limits s -\- —£ and £ — — £ ; and since I 25 25 the inherent uncertainty of £ makes it more than an even chance that its actual value is outside of the limits £ -f- ^ « and £ — — f , in accordance with the above determination of X, we conclude that £ can be represented by one of the series of numbers 0, 1, 2 9, or of the fractions -, - -, or by a product of one of these numbers by an integral power of 10, with more accuracy than we can determine its amount by one hundred comparisons between the observed and the computed values of m. It would be easy to show, from the probable existence of constant errors alone, that an indefinite increase of the number of comparisons with observation would not sensibly diminish the uncertainty of £ below the amount stated. The proposition (2) would thus be sustained, as far as relates to all expressions for probable errors and weights, since they must depend upon conditions similar to those limiting the accuracy of f . An immediate consequence of this admission will be the extension of the proposition in question, in the qualified sense, at least, in which alone it is to be understood, to all other arithmetical expressions required in the application of the method of least squares, since the peculiar province of the latter is restricted entirely to the solution of equations of the form (5 c.) a (a; — ar,) -f J (y — y,) 4- -f (m — m,) = e, in which each separate term and factor may be defined as proposed in (2). To illustrate this, let us suppose for the moment that 0.477 , or usually x > — £. -^ sfn — n' 21 «, will thus be comprised between the limits £ -f- ^ x and t — ^ x, when 17" has such a value that \ £ J 10 n — n'' The relative accuracy of x and x^ will now be investigated for some special exam- ples of deviation from a strict compliance with the method of least squares. Let the equations proposed for solution be the following : — ax-j-iy-j- 4-m =e, weight = w, (7.) a' x' + J' y + + ?n' = e', " = w', where e is the difference between the observed and computed value oi m ; m being the element derived from observation. In solving these by least squares, the final equation for x is formed by taking the sum of all the equations after multiplying the first by a w, the second by a w', and so on, and then making a w e -\- a' w' e' -\- =0, and for y h w e -\- I' w' e' -\- =0, • continuing in succession to form new equations until a final equation is obtained for each unknown quantity. We shall compare the results of two solutions of the above equations (7), in one of which (I.) the factors aw, a w' conform strictly to the method of least squares. In the other (II.}, these factors are replaced respectively hy a, al ; a being that one of the numbers 0, 1, 2 9, or of the fractions j, y ^1 or of their products by an in- tegral power of 10, which approaches most nearly to a given ratio with aw, and a that which approaches most nearly to the same ratio with a w, 8cc. In a similar manner, )8, y are used in the place of bw, cw The true values of x and y we will indicate by Xq, y^ Those deduced by (I.) will be denoted hy x,y , and those deduced by (II.) will be denoted by Xi, yi For the final equation for x, we make a w e + o! w' e' + := 0. IN THE METHOD OF LEAST SQUARES. 18 For the final equation for Xi « ei + «' ej' + = 0. For the corresponding final equations for c^o, which must be rigorous, we make either a w ec, -{- a' ra' e'o -\- ^ a w e,, -\- a' w' e'o -\- or o Co + «' e'o + = a Co + ce' e'o + according as the first (I.) or the second (II.) form of combination is adopted. Co, e, and Si are the values of e when the indeterminates Xq, y^ , x, y Xi,yi , &c., replace x, y in (7). The final equations for the combination (I.) are : — P x-\- P' y-{- P" Z+ + I, = 0, . P'x+Q y-j-Q'z-\- + M=0, ^ '' P"a;4- Q'2/+ Q"«4- + iV = 0, P = waa -{- w' a' a' -{- , Q = to i b -{- w' b' h' -\- , R = w c c -{- w' c' c' -\- (9.) P' = wab-\-w' a'b' -}- , Q — ivb c -\- w' b' c' -\- , R' = w c d -\- w'c'd' -\- From the conditions I. and II. applied to the original equations (7), if we make Aa =^ a W -{- da , Bj3=&«)-f-8i3, Cy — c w -\- 8y , Aa' = a'w' -\-8a' , B/3' = 5'w' + 8/3' , Cy' = c'^mj' + Sy' , may be obtained P{x-x,)-\-P{y-y,)-\- -^L+Px,-\-Py, + = L + Pa;, + P'3/', + — — e^da — e\da' — Hence, P (x — xi) + P' (y — y,) + = 8« e, + 8a' e'. -(- (10.) P'{x — x,) + Q {y — y,) + = 8^e. + 8^' e'. + And in a similar manner, P (xo — x) -\- P' {yo — y) -\- = a w e, + a' w' e'o + (11.) P' (a;„ _ a;) + Q (y„ _ y) + = bwe,-\-b'to'e'o-{- Since e is the probable value of Xq — x, and 17" the probable value of (x — Xi), to ob- 3 14 ON THE USE OP EQUIVALENT NUMBERS n tain the ratio — , we will compare the probable values of Xq — x and x — Xi, having, as above, /■in\ n" Probable value of {x — a;,) f Probable value of {xg — x) Xq — OS and x — w^ must be derived from a solution of equations (10) and (11), but since (II.) differs from (I.) by small variations only, we have, very nearly, (13.) w e,= _|_ «,/ e',2 _(- = ^e'' + w'e" + For the second member of (13) is a minimum relatively to the mode of solution, and, as has already been shown, (1), it differs from the first member by small terms of the second order only, those of the first order vanishing with the first differential coefficient oi SI = w e^ -\- vf e'^ -\- If, then, /^o, /i, and ^i represent the probable values of Bq, e, and Ci corresponding to the unit of weight of the equations (7), we may assume, for the purpose of deter- mining X — ,2?!, that /i — /lii is a small quantity compared with /i, since we have a w e'' -I- tti' e" -I- — = ri 7-^-7 = 1, very nearly. /ti w e, -|- w' e'c -\- Moreover, in the absence of exact knowledge of the magnitude of the errors of e^, e'o, , it is necessary to admit that they are best represented by the errors e, e, ; hence we have — = 1, and consequently - = 1, very nearly. The conditions of the solution (II.) give for the probable value of either of the ratios 0' w' aw ^ bw ^ 8 a' 8 fi' ' a'w'' b'w'' (14.) 8a — = g^ aw I'-g..... ow 8 a' a'w' = g> Because a being by (II.) nearly proportional to a w, the probable value of S a will also be proportional to a iv ; and a similar remark applies equally to S fl, S y, &c. The probable values of the second numbers of (10) and (11) are then, respectively, 8aei-\-8a' e'i-\- = g Mi V? ' a w e^ -\- a' w' e' ^ -\- = ,j^ ^p^ 8 ^ e, + S ^' e'l + = g' Fi \/Q ' J to 60 + i' M!' e'o + = ^0 \/Q, Hence, in consequence of the identity of the coefficients P,!^ , Q, Q' , &c. in the two systems (10) and (11), we obtain IN THE METHOD OF LEAST SQUARES. 15 Probable value of (a; — a;,) _ /i, ^ _ Probable value of {y — y,) _ ^i Probable value of (a?o — a;) ims And from (12) the general expression '' Probable value of (a?o — a;) im, Probable value of (^o — y) p-a (16.) e Giving to g the value (3) ^ 1 we shall have in the present case and by (6) V 1 ^ 1250 In other words, by using the foi-m of solution (II.) in the place of a rigorous appli- cation of the method of least squares, the probable errors of the concluded results will not be increased by one one-thousandth part, — a difference entirely too small to be sensible. The two processes, as far as regards accuracy, therefore, may be consid- ered as perfectly identical. On the other hand, the advantages of simplicity and convenience are altogether in favor of the second, in which all the operations of mul- tiplication and division required in the construction of the final equations are reduced to their simplest arithmetical forms. The necessity of distinguishing between the probable error of x^, that is, the proba- ble value of (a?o — x^, and the difference between x^ and x, or ( I T> I 1 AT « I Equations formed by I eliminating Xi. (c,y) Pi »j, Zi -|- + -^i^j, = •» Equations formed by . , . . j eliminating Si and y^. 1 IF Probable error of the equation (22) (a) = (1 ± - ^) M ^ — i (( u t( (t (23.) (c,,) = (1 ± i^^'^Jf"^' To demonstrate these results, (23), it is to be observed that the probable errors of the second members of the equations (22), (a), (6^), are the probable sums of the second members of the equations i'l (^0 — a;,) + P', (yo — 2/,) + = « eo + a' e'o + Q,.(y«-!/,) + = (i3-^ «)fo+(^'-^«')e'„ + The probable value of (j?e\ is, by (13) and (14), . , . «''ej = — - a ej := w ej 1 -^ — = /* (1 iff) — , 18 ON THE USE OF EQUIVALENT NUMBERS and the probable value of the sum of the terms aco -\- a' e'^ -\- is «e„+«'e'„ + = _^ ^(1 ± g) V« a + «' a' + = ^ {1 ± I g) \^ , which is the probable error of the equation (22), (a). Again, since Au = aw(l + —'), Bp^bw(l + ^-^\, \ a id) \ b loj and Pi = a a -{- a' a' -\- , Q, = J + 0' J' + P\ = a b -\- a' b' -\- , 'Q,=pa-]-^'a' -\- or, substituting the probable values \ aw J \ bwj we have Moreover, The sum of all the terms ('^ „a—'-^^a\ = ('-^ p, _ !^ /Q \ = o, Therefore the probable sum of the terms 0-^„)e„ + 0'-^a')e'„ + will be which is the probable error of the equation (22), (b^). The other probable errors in (23) are readily supplied by analogy. If we neglect - g, of which the probable value is less than — , the probable errors of (22), (a), (b^) become Probable error of equation (22) (a) = /* — ,• Probable error of equation (22) (b^) = fi -^ • We shall now proceed to explain a third form of solution, (III.)- IN THE METHOD OF LEAST SQUARES. 19 Returning to the equations (10), P (x — ic,) + -P' (y — yi) + = e,8« +e'.8«' + (24.) p/(^_^,)_^Q (y-y.) + = e,d^+e',8^'4-, we find, for the probable values of their second numbers, g li \/w a a -\- w' a' a' -\- = g' F V^" > (25.) g /* Vwb J 4- w' J' J' + = g- ^ V Q , It is evident that the probable sum CiS a-j-e\S a,' -^ , being proportional to the square root of the sum of the squares of the individual terms, depends mainly upon the large terms ; or, since e^Sa = ei -v/mT— ^ and e^ \/w'= fi, this sum will be +^'+ w' S a J. , . Sa ^ i a' If any two or more of the coefiicients -7=, as, for instance, -7=. and -7==, were equal, any small change increasing the former and diminishing the latter by equal amounts would not alter the coefficient of u: but if -7-^ were much smaller than —7=^, we shoidd have, very nearly, www' and a small change in -7^ would affect the coefficient of /x by an amount insensible compared with the effect of an equal change in —r^. Let (P) represent the sum of a certain number of the largest of the terms composing the series P =^ w aa-\- w' a' a' -\- and (p) the sum of a number of the smallest of the terms of the same series. Let also (S P) be the sum of the terms — ^ corresponding to the series (P), and (Bp) the sum of the terms — , corresponding to the series (p). Then we have the probable values For the large terms, (8 P) = g' (P), For the small terms, (^p) =^ g' (p), g representing the general probable value of — for all the terms, whether of large, small, or medium value. ^ Or irti if UNIVERSITY OF J 20 ON THE USE OF EQUIVALENT NUMBERS Let US suppose the mode of solution (II.) to be itself varied by changing the factors «, a , &c., corresponding to the large and small terms, so that for the large tenns, g, or the probable value of — , -7^, , for these particular terms, becomes ■^ a 10 a IB g=H, and for the small terms, We shall then have the probable values, For the large terms (8 P,) = IP (P), ^ '' For the small terms (S;?,) = A' {p). (8 P) and (pp) becoming (S Pj) and {^Pi) when g becomes H and h. In order that the probable sum of the second member of the equation. P{x — x,) + P'{y — y,)^ = e, 8 « + e'l S «' + should not be increased by the proposed changes of S a, we must have (8P) +{hp) >(8P,)+(Si.,)> or, by (26) and (27), W (P) + h- (p)< g= (P) + g^ (p). We shall assume, for the terms corresponding to {p), that the probable value of A is A = —1. This condition involves only small changes in the factors a, a' , because, for the terms corresponding to {p), a w being small, B a ^ aw h = — aw, will also be small ; we then have H^(P)<^(P) + (^-l)(p), or, since we can put ^^ — 1 = — 1 very nearly, g being small compared with imity, we obtain (28.) g^ _ H^ > — , H^ < g^ — — , s ^ (P) ^ ® (P) representing the condition to be observed in order that the second members of (24) should not be increased by the changes made in the large and small values of a. This, it will be remembered, can be applied only when the condition h ^ — 1 in- volves only small changes in the factors a, a of the order of the mean value of S a for all the factors. ^^ being necessarily a positive quantity, H must always be less than g. IN THE METHOD OF LEAST SQUARES. 21 (28) may easily be extended to the analogous cases of the second members of the equations P' {x — X,) -\- Q {y - Vr) + = e.«j3 + e',8^' 4- P" (x — a;.) + Q' (y — 2/i) + = ei8y + e'l 8 y + so that (?) (r) w «'-^>^- ''-•''>wy give the limits within which the proposed changes of the factors ^, ^ 7, 7' will not increase the probable sums eiB^-{-e\B^ -\- and ejBy-\-e\Sy-\- For the factors a, a' , /3, /3' corresponding to the equations most important in their influence upon the final determination of a?, y respectively, if we use num- JV N' . 1 hers chosen from a series for which — — — is only — as large as it is for the series (2), we shall have And if at the same time we omit altogether a certain number of the unfavorable equations by making in these instances a = 0, /S = , that is, S a = — aw, S /3 ^ — b IV, ov h = — 1, we find ^631 We shall therefore keep within the limits (28) and (29) as long as the coefiicients in the omitted equations satisfy the conditions (P)^68l' (Q)^68l' The probable values of a? — a?!, j/ — i/^, will not have been increased, and conse- quently the solution may be accepted as equivalent to II. A general method, III., of adjusting the degree of numerical accuracy which should be observed in the expression of the factors a, a' , P, ^' > inay be derived from the following considerations. In II. the adjustment is evidently not so favorable as it might be, since the limit of the intentional inaccuracies S a, S a' ,B^,B0 has been fixed by the relations ha — aw g, 6 a' =^ a' w' g', 8^ = hwg, 8^' = b' w' g', g having the same average value whether aw, bw be large or small ; thus the 4 22 ON THE USE OF EQUIVALENT NUMBERS largest iaaccuracies are committed when aw,bw are largest, that is, when the equation has most influence upon the final result for any particular indeterminate. In order to secure a more advantageous distribution, it will be necessary to recur again to the equations (10). It appears that, for a given limit of inaccuracy in the expression of the factors, the probable values of 07 — <3?i, 3/ — 2/i • ^'^^^ ^^ least when the separate terms of the second members of these equations, irrespective of their signs, are equal to each other, or e,8a = e',fi«'= Ci 8 /3 = e'l 8 /3' = S a, S /3 ought therefore to be inversely proportional to the probable errors of the primitive equations, or directly as the square roots of their weights. We shall, then, define III. hy the relations ba—kgs/w, 8fi=BgVw, (30.) da' = A.gVw', 8^'=BgV w A, B and g heing constant quantities. To secure the equivalency of II. and III., the values of A, B must depend on the condition that the probable values of the second members of (10) should remain unchanged, or g/i Va? + A^ + = gM Vp, g/. Vb' + b" + = ^^ Vq, lL.\/n= VP, BV'»= V'Q, Hence it is easy to conclude, that, if we make in (30) A = mean value of a \^w, B = mean value of b Vw, the means being in every instance taken without regard to signs, the probable values of X — cVi, t/ — y, will be smaller in III. than in II., while III. in point of facility has a decided advantage over II. ; since by making a = 0, ;S = in all cases in which a Vw <. Ag, b \/w = S « + S' a' + , + 'Qi + Q. + + M, = ^ a + ^ J + (32.) _|- ^' a' + iS' 5' + } = S^ + S'/3' + "Ri + 'Ri-\- + JV. = y a + y J + + ya' + yj' + } =Sy + S'y' + , The sum of all the equations (32) is (A + P', + -{- i,) + CQ. + Q. 4. + ilf.) + = S-?+S'-S'+ or (33.) Jr(Pi,Qt,Ki Ly,M„N, ) = ^(S^). When this is not fulfilled, the particular equation at fault may be detected by using the partial tests, Pi + P'. + X. = Sa + S'«' + (^^•^ 'Q1+Q1+ Jii;=s/3 + s'j3' + •.. If all the numbers are not carried out to the same number of figures beyond the decimal point, it will be advisable, in order to apply the test to the best effect, to alter arbitrarily the decimal point, making, for the time being, the number of figures to the right hand of it the same in all the equations. The successive stages in the computations by (II.) and (III.) will be as follows : — a) To assign the values of the equivalent factors a, /3 , which can readily be done by simple inspection. The factors once adopted must be used rigorously in mul- tiplying every term of the equations to which they belong. 24 ON THE USE OP EQUIVALENT NUMBERS b) After the multiplications have been performed, and the sums taken, the numbers adopted in the final equations are to be tested by (33). c) The solution of the final equations. d) The determination of weights. If changes have been made in the decimal pointing, or otherwise, by introducing the constants A, B , it must be remembered that, although the final equations thus formed will give the same values of o^i , j/, , &c. that Avould have been obtained if no such alteration had been made, the determination of the weights and probable errors of Xi, ^i , &c. requires that the correct pointing be restored in the coefficients, or else that the probable errors be computed in conformity with the formulse (23). When the number of indeterminates is considerable, it will be advisable, in solving the final equations, to eliminate x, y, z, &c., in succession, and then to repeat the operation, commencing the elimination in the reverse order, s, y, x, &c. One of the advantages of so doing is a complete check upon the work by the comparison of the value of that indeterminate which is obtained last by both eliminations. It is, however, mostly recommended from its facilitating the computation of the weights. In this case, the following formulae may be used, if the number of indeterminates does not exceed six. Let these be x, y, z, f, -n, ^, and their weights, W^^y , W^y , &c. The ordinary formulae for computing the weights give (35.) TF(0 = U^ y z i,, TFw = P< , 5 J y . W^■Q is the coefficient of f in the equation resulting from the elimination of x, y, z, f , and r}, by the process indicated in (18), (19), and (20), and W^^^ the coefficient of x after ^, r}, f, z, and y have been eliminated. We have, also, (36.) W(,, = ^J-JTF(o, W^, = ^^^W,:.,. '-'X y z i ■'■ t, " i * The factors and divisors required in (36) will have been already computed during the eliminations which have preceded. From the equations containing only f , 17, ?, the latter is to be eliminated ; and from the equations containing 2;, y, and x, x is to be eliminated. We then have Trco = |^J^pr„, w,z, = ^;^^w,,,. ix y z Z Hi H i X For the weight of Xi,yi, &c., when (II.) or (III.) is used, we shall have, from (23), Weight o{xi= A T^d) , (37.) « yi = B W(y) , IN THE METHOD OF LEAST SQTJAKES. 25 The limits of effective accuracy appropriate to the numerical representation of the coefficients a, 6 may be investigated in the following manner : — If we determine cr, , j/, by the method of least squares from the equations, (38.) (a — da) Xi-{- {b — dh) i/i-\- -\- m — d m = ei, weight = w, we shall have from (8) and (38), i£ da, db , which may be employed to represent the errors of the adopted coefficients, are small, (39.) P(xi — x) -\- P' (yi — y) -\- = aw( — e -\- xda-{- y dh -{- -\- dm) -\- If fi is the probable value of dm \/w, the most suitable values of a; da, y dh . evidently fulfil the conditions (40.) xda\w =ydb\^w = dm\^w = ii', xda'Vw'^^ydb'Vw' := dm' \ w' =^ (>.' . Observing that we may substitute in the second members of (39) the probable values e <^ a X, e' <. a' X , e Khy, e' <. V y , we may conclude, by comparing (39) with (11), (12), and (6), that, if;*' is less than the limit (41.) ^ = -4^=, 'd denoting the number of indeterminates in (38), the difference Cj — i of the probable errors of x^,yi derived from (38) and of x,y derived from (8) will be (42.) e.-e. di. = — 371.09 4- 5.91569 + 7203.91— 0.09344 — 2.28516 — 0.34664— 0.18194 = —138534.00 + 2458225.00 + 62.13 — 510.58 4- 213.84 + 73.45 = — 115.81 + 0.71612 + 1.11063 — 0.06392+ 0.25868 = + 25.66 + 9.29213 — 0.36175 — 0.57384 = + 75.23 + 2.22346— 0.37766 = + 17.11 II. + 5.42383 Coef. of Coef. of Coef. of Coef. of Coef. of Coef. of II dL. dy. dn. dtf. da. di. = — 372.23 -f 5.89945 + 7208.64— 0.04729 — 2.22332 — 0.33807— 0.19932 = — 166.23 + 2229.06— 0.13790 — 0.19504 + 0.28590— 0.21945 = — 1041.71 + 7.03449 + 10.51875 + 0.43566 + 3.12386 = + 176.07 + 9.01520 — 0.49449— 0.72094 0:= + 100.43 + 2.17690— 0.46611 = + 16.00 III. + 4.24758 Coef. of Coef. of Coef. of Coef. of Coef. of Coef. of II dL. dt. dn. dip. da. di. = — 371.03 + 5.97864 + 7296.25— 0.06706 — 2.33932 — 0.37876— 0.08024 = — 112.76 + 2048.68+ 0.02871 — 0.44288 + 0.25090 — 0.09299 = — 1108.40 + 7.18029 + 11.56325 + 0.21715+ 3.07926 = + 274.45 + 9.07933 — 0.50448 — 0.43778 = + 117.10 + 2.22130— 0.43492 = + 30.76 + 5.33957 From which are finally deduced the values of the six unknown quantities : II. III. di =— 3.15 — 3.77 — 5.76 da =— 34.37 — 46.94 — 53.85 dq, =— 4.29 — 22.41 — 33.50 dTT =+ 166.44 + 186.17 + 212.41 d y = + 0.054335 + 0.08963 + 0.05115 dL = — 3.06 — 56.32 — 14.58 The following are the outstanding errors of the original equations : — IN THE METHOD OF LEAST SQUARES. 33 I. II. III. 1 — 111.00 — 155.51 — 127.23 2 — 8.31 — 3.70 — 7.29 3 + 59.18 + 59.08 + 84.73 4 — 36.67 — 47.02 — 54.55 5 + 19.92 + 21.77 + 33.12 6 + 0.07 -1- 6.92 + 19.63 7 4- 85.77 + 54.52 + 16.00 . 8 + 25.01 + 35.37 + 38.66 9 -f- 135.88 + 157.61 + 144.46 10 — 204.64 — 155.64 — 174.86 11 + 83.44 + 69.63 + 64.28 ms of the squares of these errors are I. II. III. Q = 92919 88556 85205 It is obvious that the solution I., as given by Gauss in the memoir above quoted, is incorrect, since the sum of the squares of the errors should be less by the method of least squares than by any other mode of combination.* A revised solution gives the following equations for I. Those for II. and III. are repeated for the sake of comparison. I. (Revised solution.) Coef. of dL. Coef. of dt. 7203.91 — 0.09344 Coef. of dTt. Coef. of df. 2.28516 — 446.986 Coef. of di. 0.34664— 0.18194 Coef. of dii. = — 371.09 + 5.91569 + = — 128204.00 + 2061567.00 + 64.72780 — 446.986 + 223.4856 + 78.50985 = — 115.28 + 0.71566+ 1.11176+ 0.05150+ 0.25807 = + 276.46 + 9.29666 — 0.53681 — 0.57179 0=+ 110.68 + 2.20290— 0.43212 = + 41.94 + 5.40298 II. Coef. of dL. = — 372.23 + 5.89945 + = — 166.23 + 0=z— 1041.71 = + 176.07 = + 100.43 = + 16.00 Coef. of Coef. of Coef. of Coef. of Coef. of dy. dn. d i^v". .',_<■■■•