UC-NRLF *C Its IDb i U fe Wi "i iTHEORY OF OBSERYATIONS LIBRARY OF THE UNIVERSITY OF CALIFORNIA. Class wm^ THEORY OF OB SE RVATION S THEORY OF OBSERVATIONS BY T. N. THIELE DIRECTOR OF THE COPENHAGEN OBSERVATOHY -►•<>>»3i5Kc>-.-. Y o! rue '>■ f UNIVERSITY ; Of . PUBLISHED HY CHARLES & EDWIN LAYTON 56 FARRINGDON STREET LONDON 1903 t4- COPENHAGEN. — PRINTED BT BIANCO LUNO CONTENTS Numbers of I. The Law of Causality. formulae Page § 1. Belief in Causality 1 § 2. The Observations and their Circumstances 2 § 3. Errors of Observations 3 § 4. Theoretical and Empirical Science , 4 II. Laws of Errors. § 5. On Rei)etitions 6 § 6. Laws of Actual Errors and Laws of Presumptive Errors 5 § 7. The Law of Large Numbers of Repetitions 6 § 8. Four Different Forms of Laws of Errors 7 III. Tabular Arrangements. § 9. Frequency and Probability 8 § 10. Repetitions with Qualitative Differences between the Results 8 § 11. Repetitions with Quantitative Differences between the Results 9 IV. Curves of Errors. § 12. Curves of Actual Errors of Observations in Discontinued Values 10 § 13. Curves of Actual Errors for Rounded Observations 10 § 14. Curves of Presumptive Errors 11 § 1.5. Typical Curves of Errors 14 § 16. Particular Measures of Curves of Errors 14 V. Functional Laws of Errors. § 17. 1. Their Determination by Interpolation 15 § 18. 2—8. The Typical or Exponential Law of Errors 16 Problems 19 § 19. 9—13. The Binomial Functions 20 § 20. 14. Some of the more general Functional Laws of Errors. Series 21 111731 Numbers of VI. Laws of Errors expressed by Symmetrical Functions. formulae Page § 21. 15—16. Coefficients of the Equation of Errors. Suras of Powers 22 § 22. 17—24. Half-luvariants 24 § 23. 25—27. Mean Values, Mean Deviations, Mean Errors 27 Examples 29 VII. Relations between Functional Laws of Errors and Half-Invariants. § 24. 28- 29. Their Kelations 30 Examples 31 § 25. 30—31. A very general Series by Half-Invariants 83 VIII. Laws of Errors of Functions of Observations. § 26. Functions of One Single Observation 35 § 27. 32—33. Half-Invariants of Linear Functions 36 § 28. Functions of Two or More Observations. Bonds 37 § 29. 34—35. Linear Functions of Unbound Observations 38 Examples 39 § 30. 36. Non-Linear Functions 41 Examples 41 §31. 37. Laws of Errors of the Mean Value. Approximation to the Typical Form 41 § 32. 38—46. Laws of Errors of the Mean Deviation and Higher Half-Invariants 44 § 33. 47. Transition between Laws of Actual and of Presumptive Errors. Rules of Prediction . . 47 Examples 50 §34. 48—50. Determination of the Law of Presumptive Errors when the Presumptive Mean Value is known beforehand 51 §35. 51. Weights of Observations. Probable Errors and other Dangerous Notions 52 IX. Free Functions. § 36. 52 - 58. Conditions of Freedom 53 Examples 55 § 37. Possibility of regarding certain Bound Observations as free 56 §38. 59—61. Every Function of Observations can be divided into a Sum of Two, which belong to Two Mutually Free Systems of Functions .56 Example 58 § 39. Every Single Observation likewise .59 § 40. Complete Sets of Free Functions 59 § 41. 62—66. Orthogonal Transformation 59 §42. 67. Schedule of Liberation 60 Example 62 General Remarks about Computation with Observed or Inexactly Given Values 63 X. Adjustment. §43. Can Laws of Errors be Determined by Means of Observations whicli are not Repeti- tions ? 64 § 44. The Principle of Adjustment 67 § 45. 68—72. Criticism ; the Method of the Least Squares 68 § 46. Adjustment by Correlates and by Elements 70 Numbers of XI. Adjustment by Correlates. formulae Page § 47. 73—81. General Solution of the Problem 71 §48. 82—84. Summary and Special Criticism 72 § 49. Schedule for Adjustment by (Correlates 73 § .'lO. Modifications of this Method 75 Examples 75 XII. Adjustment by Elements. § 51. 85 —86. One Equation for each Observation. Normal Equations 77 § 52. 87—89. Special Case of Free Normal Equations 80 § 53. 90—102. Transformation of the General Case into the Preceeding Special Case 80 § 54. 103-108. The Minimum Sum of Squares 83 § 55. Criticism 84 § .56. Under- Adjustment and Over- Adjustment 85 Examples 86 XIII. Special Auxiliary Methods. § 57. Lawful and Illegitimate Facilities 94 §58. 109—110. Adjustment upon Differences from Preceeding Computations 94 §69. Ill — 113. How to obtain Approximate Freedom of the Normal Equations 96 § 60. The Method of Fabricated Observations. Example 98 § 61. The Method of Partial Eliminations .' 99 Example 100 §62. 114—116. Rules for the General Transformation of the System of Elements 103 Example 105 §63. 117-120. The Method of Normal Places 106 Example HO § 64. Graphical Adjustment 112 XIV. The Theory of Probability. § 65. 121—123. Relation of Probability to the Laws of Errors by Half-Invariants 116 §66. 124—125. Laws of Errors for the Frequency of Repetitions. Obliquity of these Laws of Errors 118 XV. The Formal Theory of Probability. § 67. Addition and Multiplication of Probabilities 119 Examples 121 § 68. 126. Use of the Polynomial Formula for Probabilities by Repetitions 123 Examples 123 §69. 127 — 129. Linear Equations of Differences. Oppermann's Transformation 124 Examples 126 XVI. The Determination of Probabilities a Priori and a Posteriori. § 70. Theory and Experience 129 § 71. Determination a Priori 130 §72. 130-133. Determination a Posteriori and its Mean Error 132 Example 134 § 73. 134-137. The Paradox of Unanimity. Bayes's Rule 134 Numbers of XVII. Mathematical Expectation and its Mean Error. formulae Page § 74. Mathematical Expectation 137 138—140. Examples 138 §75. 141—143. Mean Errors of Mathematical Expectation of Unbound Events 139 Examples 140 § 76. 144-146. Mean En-or of Total Mathematical Expectation of the Same Trial 141 147. Examples 142 § 77. The Complete Expression of the Mean Errors 142 I. THE LAW OF CAUSALITY. § 1. We start with the assumption that everything tJiuf exists, and everything that happens, exists or happens as a necessary consequence of a previous state of things. If a state of things is repeated in every detail, it must lead to exactly the same consequences. Any difference between the results of causes that are in part the same, must be explainable by some difference in the other part of the causes. This assumption, which may be called the law of causality, cannot be proved, but must be believed; in the same way as we believe the fundamental assumptions of religion, with which it is closely and intimately connected. The law of causality forces itself upon our belief. It may be denied in theory, but not in practice. Any person who denies it, will, if he is watchful enough, catch himself constantly asking himself, if no one else, why this has happened, and not that. But in that very question he bears witness to the law of causality. If we are consistently to deny the law of causality, we must repudiate all observation, and particularly all prediction based on past experience, as useless and misleading. If we could imagine for an instant that the same complete combination of causes could have a definite number of different consequences, however small that number might be, and that among these the occurrence of the actual consequence was, in the old sense of the word, accidental, no observation would ever be of any particular value. Scientific observations cannot be reconciled with polytheism. So long as the idea prevailed that the result of a journey depended on whether the power of Njord or that of Skade was the stronger, or that victory or defeat in battle depended on whether Jove had, or had not, listened to Juno's complaints, so long were even scientists obliged to consider it below their dignity to consult observations. But if the law of causality is acknowledged to be an assumption which always holds good, then every observation gives us a revelation which, when correctly appraised and compared with others, teaches us the laws by which God rules the world. We can judge of the far-reaching consequences it would have, if there were con- ditions in which the law of causality was not valid at all, by considering tiie cases in which the effects of the law are more or less veiled. 1 In inanimate nature the relation of cause and effect is so clear that the effects are determined by observable causes belonging to the condition immediately preceding, so that the problem, within this domain, may be solved by a tabular arrangement of the several observed results according to the causing circumstances, and the transformation of the tables into laws by means of interpolation. When, however, living beings are the object of our observations, the case immediately becomes more complicated. It is the prerogative of living beings to hide and covertly to transmit the influ- ences received, and we must therefore within this domain look for the influencing causes throughout the whole of the past history. A difference in the construction of a single cell may be the only indication present at the moment of the observation that the cell is a transmitter of the still operative cause, which may date from thousands of years back. In consequence of this the naturalist, the physiologist, the physician, can only quite ex- ceptionally attain the same simple, definite, and complete accordance between the observed causes and their effects, as can be attained by the physicist and the astronomer within their domains. Within the living world, communities, particularly human ones, form a domain where the conditions of the observations are even more complex and difficult. Living beings hide, but the community deceives. For though it is not in the power of the com- munity either to change one tittle of any really divine law, or to break the bond between cause and effect, yet every community lays down its own laws also. Every community tries to give its law fixity, and to make it operate as a cause; for instance, by passing it oft' as divine or by threats of punishment; but nevertheless the laws of the community are constantly broken and changed. Statistical Science which, in the case of communities, represents observations, has therefore a very difficult task; although the observations are so numerous, we are able from them alone to answer only a very few questions in cases where the intellectual weapons of historical and speculative criticism cannot assist in the work, by independently bringing to light the truths which the communities want to conceal, and on the other hand by re- moving the wrong opinions which these believe in and propagate. § 2. An isolated sensation teaches us nothing, for it does not amount to an ob- servation. Observation is a putting together of several results of sensation which are or are supposed to be connected with each other according to the law of causality, so that some represent causes and others their effects. By virtue of the law of causality we must believe that, in all observations, we get essentially correct and true revelations; the difficulty is, to ask searchingly enough and to understand the answer correctly. In order that an observation may be free from every other assumption or hypothesis than the law of causality, it must include a perfect description of all the circumstances in the world, at least at the instant preceding that at which the phenomenon is observed. But it is clear that this far surpasses what can be done, even in the most important cases. Real observations have a much simpler form. By giving a short statement of the time and place of observation, we refer to what is known of the state of things at the instant; and, of the infinite multiplicity of circumstances connected with the observation we, generally, not only disregard everything which may be supposed to have little or no influence, but we pay attention only to a small selection of circumstances, which we call essential, because we expect, in virtue of a special hypothesis concerning the relation of cause and effect, that the observed phenomenon will be effect of these circumstances only. Nay, we are often compelled to disregard certain circumstances as unessential, though there is no doubt as to their influencing the phenomenon; and we do this either because we cannot get a sufficient amount of trustworthy information regarding them, or because it would be impracticable to trace out their connection with the eftect. For instance in statistical observations on mortality, where the age at the time of death can be regarded as the observed phenomenon, we generally mention the sex as an essential circumstance, and often give a general statement as to residence in town or country, or as to occupation. But there are other things as to which we do not get sufficient information: whether the dead person has lived in straitened or in comfortable circumstances, whether he has been more or less exposed to infectious disease, etc. ; and we must put up with this, even if it is certain that one or other of these things was the principal cause of death. And analogous cases are frequently met with both in scientific observations and in everyday occurrences. In order to obtain a perfect observation it is necessary, moreover, that our sensations should give us accurate information regarding both the phenomenon and the attendant circumstances; but all our senses may be said to give us merely approximate descriptions of any phenomenon rather than to measure it accurately. Even the finest of our senses recognizes no difference which falls short of a certain finite magnitude. This lack of accuracy is, moreover, often greatly increased by the use of arbitrary round numbers for the sake of convenience. The man who has to measure a race-course, may take into account the odd metres, but certainly not the millimetres, not to mention the microns. § 3. Owing to all this, every actual observation is affected with errors. Even our best observations are based upon hypothesis, and often even on an hypothesis that is cer- tainly wrong, namely, that only the circumstances which are regarded as essential, influence the phenomenon; and a regard for practicability, expense, and convenience makes us give approximate estimates instead of the sharpest possible determinations. Now and then the observations are affected also by ijross errors which, although 1* not introduced into them on purpose, are yet caused by such carelessness or neglect that they could have been, and ought to have been, avoided. 1 contradistinction to these we often call the more or less unavoidable errors accidetdal. For accident (or chance) is not, what the word originally meant, and what still often lingers in our ordinary acceptation of it, a capricious power which suffers events to happen without any cause, but only a name for the unknown element, involved in some relation of cause and effect, which pre- vents us from fully comprehending the connection between them. When we say that it is accidental, whether a die turns up "six" or "three", we only mean that the circumstances connected with the throwing, the fall, and the rolling of the die are so manifold that no man, not even the cleverest juggler and arithmetician united in the same person, can suc- ceed in controlling or calculating them. In many observations we reject as unessential many circumstances about which we really know more or less. We may be justified in this; but if such a circumstance is of sufficient importance as a cause, and we arrange the observations with special regard to it, we may sometimes observe that the errors of the observations show a regularity which is not found in "accidental" errors. The same may be the case if, in computations dealing with the results of observations, we make a wrong supposition as to the operation of some circumstance. Such errors are generally called systematic. § 4. It will be found that every applied science, which is well developed, may be divided into two parts, a theoretical (speculative or mathematical) part and an empirical (observational) one. Both are absolutely necessary, and the growth of a science depends very much on their influencing one another and advancing simultaneously. No lasting divergence or subordination of one to the other can be allowed. The theoretical part of the science deals with what we suppose to be accurate determinations, and the object of its reasonings is the development of the form, connection, and consequences of the hypotheses. But it must change its hypotheses as soon as it is cle£lr that they are at variance with experience and observation. The empirical side of the science procures and arranges the observations, compares them with the theoretical propositions, and is entitled by means of them to reject, if necessary, the hypotheses of the theory. By induction it can deduce laws from the obser- vations. But it must not forget — though it may have a natural inclination to do so — that, as shown above, it is itself founded on hypotheses. The very form of the observation, and especially the selection of the circumstances which are to be considered as essential and taken into account in making the several observations, must not be determined by rule of thumb, or arbitrarily, but must always be guided by theory. Subject to this it must as a rule be considered best, that the two sides of the science should work somewhat independently of one another, each in its own particular way. In what follows the empirical side will be treated exclusively, and it will be treated on a general plan, investigating not the particular way in which statistical, chemical, phy- sical, and astronomical observations are made, but the common rules according to which they are all submitted to computation. II. LAWS OF EEROES. § b. Every observation is supposed to contain information, partly as to the phenomenon in which we are particularly interested, partly as to all the circumstances, connected with it, which are regarded as essential. In comparing several observations, it makes a very great difference, whether such essential circumstances have remained unchanged, or whether one or several of them have changed between one observation and another. The treatment of the former case, that of repetitions, is far simpler than that of the latter, and is therefore more particularly the subject of our investigations; nevertheless, we must try to master also the more difficult general case in its simplest forms, which force them- selves upon us in most of the empirical sciences. By repetitions then we understand those observations, in which all the essential circumstances remain unchanged, in which therefore the results or phenomena should agree, if all the operative causes had been included among our essential circumstances. Further- more, we can without hesitation treat as repetitions those observations, in which we assume that no essential circumstance has changed, but do not know for certain that there has been no such change. Strictly speaking, this would furnish an example of observations with systematic errors; but provided there has been no change in the care with which the essential circumstances have been determined or checked, it is permissible to employ the simpler treatment applicable to the case of repetitions. This would not however be per- missible, if, for instance, the observer during the repetitions has perceived any uncertainty in the records of a circumstance, and therefore paid greater attention to the following repetitions. ^ '^ § 6. The special features of the observations, and in particular their degree of accuracy, depend on causes which have been left out as unessential circumstances, or on some overlooked uncertainty in the statement of the essential circumstances. Consequently no speculation can indicate to us the accuracy and particularities of observations. These must be estimated by comparison of the observations with each other, but only in the case of repetitions can this estimate be undertaken directly and without some preliminary work. The phrase late of errors is used as a general name for any mathematical expres- sion representing the distribution of the varying results of repetitions. Lmvs of actual errors are such as correspond to repetitions actually carried out. But observations yet unmade may also be erroneous, and where we have to speak hypo- thetically about observations, or have to do with the prediction of results of future repe- titions, we are generally obliged to employ the idea of "laws of errors". In order to pre- vent any misunderstanding we then call this idea "Zfrns of presumptive errors". The two kinds of laws of errors cannot generally be quite the same thing. Every variation in the number of repetitions must entail some variations in the corresponding law of errors; and if we compare two laws of actual errors obtained from repetitions of the same kind in equal number, we almost always observe great differences in every detail. In passing from actual repetitions to future repetitions, such differences at least are to be expected. More- over, whilst any collection of observations, which can at all be regarded as repetitions, will on examination give us its law of actual errors, it is not every series of repetitions that can be used for predictions as to future observations. If, for instance, in repeated measure- ments of an angle, the results of our first measurements all fell within the first quadrant, while the following repetitions still more frequently, and at last exclusively, fell within the second quadrant, and even commenced to pass into the third, it would evidently be wrong to predict that the future repetitions would repeat the law of actual errors for the totality of these observations. In similar cases the observations must be rejected as bad or mis- conceived, and no law of presumptive errors can be directly based upon them. § 7. Suppose, however, that on comparing repetitions of some observation we have several times determined the law of actual errors in precisely the same way, employing at first small numbers of repetitions, then larger and still larger numbers for each law. If then, on comparing these laws of actual errors with one another, we remark that they be- come more alike in proportion as the numbers of repetitions grow greater, and that the agreements extend successively to all those details of the law which are not by necessity bound to vary with the number of repetitions, then we cannot have any hesitation in using the law of actual errors, deduced from the largest possible number of repetitions, for pre- dictions concerning future observations, made under essentially the same circumstances. This, however, is wholly legitimate only, when it is to be expected that, if we could obtain repetitions in indefinitely increasing numbers, the law of errors would then approach a single definite form, namely the laiv of presumptive errors itself, and would not oscillate between several forms, or become altogether or partly indeterminate. (Note the analogy with the difference between converging and oscillating infinite series). We must therefore distinguish between good and bad observations, and only the good ones, that is those which satisfy the above mentioned condition, the Uno of large numbers, yield laws of presumptive errors and afford a basis for prediction. As we cannot repeat a thing indefinitely often, we can never be quite certain that a given method of observation may be called good. Nevertheless, we shall always rely on laws of actual errors, deduced from very large numbers of concordant repetitions, as suffi- ciently accurate approximations to the law of presumptive errors. And, moreover, the purely hypothetical assumption of the existence of a law of presumptive errors may yield some special criteria for the right behaviour of the laws of actual errors, corresponding to the increasing number of the repetitions, and establish the conditions necessary to justify their use for purposes of prediction. We must here notice that, when a series of repetitions by such a test proves bad and inapplicable, we shall nevertheless often be able, sometimes by a theoretical criticism of the method, and sometimes by watching the peculiarities in the irregularities of the laws of errors, to find out the reason why the given method of observation is not as good as others, and to change it so that the checks will at least show that it has been improved. In the case mentioned in the preceding paragraph, for instance, the remedy is obvious. The time of observation is there to be reckoned among the essential circumstances. A4id if we do not attain our object, but should fail in many attempts at throwing light upon some phenomenon by means of good observations, it may be said even at this stage, before we have been made acquainted with the various means that may be employed, and the various forms taken by the laws of errors, that absolute abandonment of the law of large numbers, as quite inapplicable to any given refractory phenomenon, will generally be out of the question. After repeated failures we may for a time give up the whole matter in despair; but even the most thorough sceptic may catch himself speculating on what may be the cause of his failure, and, in doing so, he must acknowledge that the error is never to be looked for in the objective nature of the conditions, but in an insuffi- cient development of the methods employed. From this point of view then the law of large numbers has the character of a belief. There is in all external conditions such a harmony with human thought that we, sooner or later, by the use of due sagacity, parti- cularly with regard to the essential subordinate circumstances of the case, will be able to give the observations such a form that the laws of actual errors, with respect to repetitions in increasing numbers, will show an approach towards a definite form, wliich may be con- sidered valid as the law of presumptive errors and used for predictions. § 8. Four different means of representing the law of errors must be described, and their respective merits considered, namely: Tabular arrangements. Curves of Errors, Functional Laws of Errors, Symmetric Functions of the Repetitions. In comparing these means of representing the laws of errors, we must take into consideration which of them is the easiest to employ, and neither this nor the description of the forms of the laws of errors demands any higher qualification than an elementary knowledge of mathematics. But we must take into account also, how far the diflerent forms are calculated to emphasise the important features of the laws of errors, i. e. those which may be transferred from the laws of actual errors to the laws of presumptive errors. On this single point, certainly, a more thorough knowledge of mathematics would be desirable than that which may be expected from the majority of those students who are obliged to occupy themselves with observations. As the definition of the law of presumptive errors presupposes the determination of limiting values to infinitely numerous approximations, some propositions from the differential calculus would, strictly speaking, be necessary. III. TABULAR ARRANGEMENTS. § 9. In stating the results of all the several repetitions we give the lajv of errors in its simplest form. Identical results will of course be noted by stating the number of the observations which give them. The table of errors, when arranged, will state all the various results and the fre- quency of each of them. The table of errors is certainly improved, when we include in it the relative fre- quencies of the several results, that is, the ratio which each absolute frequency bears to the total number of repetitions. It must be the relative frequencies which, according to the law of large numbers, are, as the number of observations is increased, to approach the constant values of the law of presumptive errors. Long usage gives us a special word to denote this transition in our ideas: probability is the relative frequency in a law of pre- sumptive errors, the proportion of the number of coincident results to the total number, on the supposition of infinitely numerous repetitions. There can be no objection to con- sidering the relative frequency of the law of actual errors as an approximation to the corresponding probability of the law of presumptive errors, and the doubt whether the relative frequency itself is the best approximation that can be got from the results of the given repetitions, is rather of theoretical than practical interest. Compare § 73. It makes some difference in several other respects — as well as in the one just mentioned — if the phenomenon is such that the results of the repetitions show qualitative differences or only differences of magnitude. § 10. In the former case, in which no transition occurs, but where there are such abrupt differences that none of the results are more closely connected with one another than with the rest, the tabular form will be the only possible one, in which the law of errors can 9 be given. This case frequently occurs in statistics and in games of chance, and for this reason the theory of probabilities, which is the form of the theory of observations in which these cases are particularly taken into consideration, demands special attention. All pre- vious authors have begun with it, and made it the basis of the other parts of the science of observation. I am of opinion, however, that it is both safer and easier to keep it to the last. § 11. If, however, there is such a diiference between the results of repetitions, that there is either a continuous transition between them, or that some results are nearer each other than all the rest, there will be ample opportunity to apply mathematical methods; and when the tabular form is retained, we must take care to bring together the results that are near one another. A table of the results of firing at a target may for instance have the following form : 1 foot to the left 1 foot too high 3 Central 13 1 foot too low 4 Total ... 20 134 26 180 Central 1 foot to the right Total 17 6 2G 109 19 141 8 1 13 If here the heading "1 foot to the left" means that the shot has swerved to the left between half a foot and one foot and a half, this will remind us that we cannot give the exact measures in such tables, but are obliged to give them in round numbers. The number of results then will not correspond to such as were exactly the same, but dis- regarding small differences, we gather into each column those that approach nearest to one another, and which all fall within arbitrarily chosen limits. In the simple case, where the result of the observation can be expressed by a single real number, the arranged table not only takes the extremely simple form of a table of functions with a single argument, but, as we shall see in the following chapters, leads us to the representation of the law of errors by means of curves of errors and functional laws of errors. It is an obvious course to fix the attention on the two extreme results in the table, and not seldom these alone are given, instead of a law of error, as a sort of index of the exactness of the whole series of repetitions, and as the higher and lower limits of the observed phenomenon. This index of exactness, however, must be rejected as itself too inexact for the purpose, for the oftener the observations are repeated, the farther we must expect the extremes to move from one another: and thus the most valuable series of observations will appear to possess the greatest range of discrepancy. a 10 On the other hand, if, in a table arranged according to the magnitude of the values, we select a single middle value, preceded and followed by nearly equal numbers of values, we shall get a quantity which is very well fitted to represent the whole series of repetitions. If, while we are thus counting the results arranged according to their magnitude, we also take note of those two values with which we respectively (a) leave the first sixth part of the total number, and (b) enter upon the last sixth part (more exactly we ought to say 16 per ct.), we may consider these two as indicating the limits between great and small deviations. If we state these two values along with the middle one above referred to, we give a serviceable expression for the law of errors, in a way which is very convenient, and although rough, is not to be despised. Why we ought to select just the middle value and the two sixth-part values for this purpose, will appear from the following chapters. IV. CURVES OF ERRORS. § 12. Curves of actual errors of repeated observations, each of which we must be able to express by one real number, are generally constructed as follows. On a straight line as the axis of abscissae, we mark ofl' points corresponding to the observed numerical quantities, and at each of these points we draw an ordinate, proportional to the number of the repetitions which gave the result indicated by the abscissa. We then with a free hand draw the curve of errors through the ends of the ordinates, making it as smooth and regular as possible. For quantities and their corresponding abscissae which, from the nature of the case, might have appeared, but do not really appear, among the repetitions, the ordinate will be = 0, or the point of the curve falls on the axis of abscissae. Where this case occurs very frequently, the form of the curves of errors becomes very tortuous, almost discontinuous. If the observation is essentially bound to discontinuous numbers, for instance to integers, this cannot be helped. § 13. If the observation is either of necessity or arbitrarily, in spite of some in- evitable loss of accuracy, made in round numbers, so that it gives a lower and a higher limit for each observation, a somewhat different construction of the curve of errors ought to be applied, viz. such a one, that the area included between the curve of error, the axis of abscissae, and the ordinates of the limits, is proportional to the frequency of repetitions within these limits. But in this way the curve of errors may depend very much on the degree of accuracy involved in the use of round numbers. This construction of areas can be made by laying down rectangles between the bounding ordinates, or still better, trapezoids with their free sides approximately parallel to the tangents of the curve. If the 11 limiting round numbers are equidistant, the mean heights of the trapezoids or rectangles are directly proportional to the frequencies of repetition. In this case a preliminary con- struction of curve-points can be made as in § 12, and may often be used as sufficient. It is a very common custom, but one not to be recommended, to draw a broken line between the observed points instead of a curve. § 14. There can be no doubt that the curve of errors, as a form for the law of errors, has the advantage of perspicuity, and were not the said uncertainty in so many cases a critical drawbacic, this would perhaps be sufficient. Moreover, it is in practice quite possible, and not very difficult, to pass from the curve of actual errors to one which may hold good for presumptive errors ; though, certainly, this transition cannot be founded upon any positive theory, but depends on siiill, which may be acquired by working at good examples, but must be practised judiciously. According to the law of large numbers we must expect that, when we draw curves of actual errors according to relative frequency, for a numerous series of repetitions, first based upon small numbers, afterwards redrawn every time as we get more and more repe- titions, the curves, which at first constantly changed their forms and were plentifully furnished with peaks and valleys, will gradually become more like each other, as also simpler and more smooth, so that at last, when we have a very large but finite number of observations, we cannot distinguish the successive figures we have drawn from one an- other. We may thus directly construct curves of errors, which may be approved as pictures of curves of presumptive errors, but in order to do so millions of repetitions, rather than thousands, are certainly required. If from curves of actual errors for small numbers we are to draw conclusions as to the curve of presumptive errors, we must guess, but at the same time support our guess, partly by an estimate of how great irregularities we may expect in a curve of actual errors for the.given number, partly by developing our feeling for the form of regular curves of that sort, as we must suppose that the curves of presumptive errors will be very regular. In both respects we must get some practice, but this is easy and interesting. Without feeling tied down to the particular points that determined the curve of actual errors, we shall nevertheless try to approach them, and especially not allow many large deviations on the same side to come together. We can generally regard as large deviations (the reason why will be mentioned in the chapter on the Theory of Probabilities) those that cause greater errors, as compared with the absolute frequency of the result in question, than the square root of that number (more exactly ]//t ^~ , where /; is the frequency of the result, n the number of all repetitions). But even deviations two or three times as great as this ought not always to be avoided, and Ave may be satisfied, if only one third of the deviations of the determining points must be called large. We may use 12 the word '•adjustment" (graphical) to express the operation by which a curve of presumptive errors is determined. (Comp. §64). The adjustment is called an over- adjustment, if we have approached too near to some imaginary ideal, but if we have kept too close to the curve of actual errors, then the curve is said to be tinder-adjusted. Our second guide, the regularity of the curve of errors, is as an aesthetical notion of a somewhat vague kind. The continuity of the curve is an essential condition, but it is not sufficient. The regularity here is of a somewhat diflerent kind from that seen in the examples of simple, continuous curves with which students more especially become acquainted. The curves of errors get a peculiar stamp, because we would never select the essential circumstances of the observation so absurdly that the deviations could become indefinitely large. Nor would we without necessity retain a form of observation which might bring about discontinuity. It follows that to the abscissae which indicate very large deviations, must correspond rapidly decreasing ordinates. The curve of errors must have the axis of abscissae as an asymptote, both to the right and the left. All frequency being positive, where the curve of errors deviates from the axis of abscissae, it must exclusively keep on the positive side of the latter. It must therefore more or less get the appearance of a bow, with the axis of abscissae for the string. In order to train the eye for the apprehension of this sort of regularity, we recommend the study of figs. 2 & 3, which represent curves of errors of typical forms, exponential and binomial (comp. the next chapter, p. 16, seqq.), and a comparison of them with figures which, like Nr. 1, are drawn from actual observations without any adjustment. The best way to acquire practice in drawing curves of errors, which is so important that no student ought to neglect it, may be to select a series of observations, for which the law of presumptive errors may be considered as known, and which is before us in tabular form. We commence by drawing curves of actual errors for the whole series of observa- tions; then for tolerably large groups of the same, and lastly for small groups taken at random and each containing only a few observations. On each drawing we draw also, besides the curve of actual errors, another one of the presumptive errors, on the same scale, so that the abscissae are common, and the ordinates indicate relative frequencies in proportion to the same unit of length for the total number. The proportions ought to be chosen so that the whole part of the axis of abscissae which deviates sensibly from the curve, is between 2 and 5 times as long as the largest ordinate of the purve. Prepared by the study of the differences between the curves, we pass on at last to the construction of curves of presumptive errors immediately from the scattered points of the curve which correspond to the observed frequencies. In this construction we must not consider ourselves obliged to reproduce the curve of presumptive errors which we may 13 know beforehand; our task is to represent the observations as nearly as possible by means of a curve which is as smooth and regular as that curve. The following table of 500 results, got by a game of patience, may be treated in this way as an exercise. Actual frequency for groups of: ^ errors hod, ted ~w o ° a^ III ■3 02 3Q ^ 25 repetitions 100 repetitions ] 11 III IV V I II III IV V !^ 7 1 1 1 110 1 3 0(X)03 0-0019 7 8 1 2 '^ 110 14 2 7 0-0071 00192 8 9 1 3 ] 1 5 3 1 1 3 2 2 3 1 3 1 1 2 1 6 10 7 7 5 35 00392 0-0636 9 10 9 2 9 5 6 6 6 4 5 4 8 3 3 3 5 6 4 6 3 4 25 22 20 17 17 101 0-0859 0-1005 10 11 3 6 3 3 3 6 4 4 5 5 3 5 3 7 2 5 5 6 3 8 15 17 18 17 22 89 0.1054 0-1021 11 12 8 5 3 4 3 3 2 8 3 7 4 6 5 -4 6 5 3 3 5 7 20 16 20 20 18 94 00934 0-0823 12 13 2 4 4 3 6 3 3 1 4 113 6 4 3 6 7 3 6 1 13 13 9 18 17 70 0-0705 0-0.591 13 14 1 2 2 4 1 2 3 2 12 4 3 5 4 4 2 4 9 6 9 12 10 46 0-0485 00387 14 15 1 2 2 1 1 3 '? 3 2 2 3 1 1 3 2 1 5 7 10 5 3 30 0-0298 0-0216 15 16 1 2 1 1 1 10 1 2 3 2 4 2 2 2 5 15 0-0145 0-0088 16 17 1 2 1 10 2 1 4 00046 0-0020 17 18 1 1 110 1 112 1 5 0-0007 0-0002 18 19 1 10 1 0-0000 19 Total 25 25 25 25 26 25 25 25 25 25 25 25 25 25 25 25 25 25 25 25 laj 100 100 100 100 500 0-9999 The law of presumptive errors here given is not the direct result of free-hand con- struction; but the curve so got has been improved by interpolation of the logarithms of its statements of the relative frequencies, together with the formation of mean numbers for the deviations, a proceeding which very often will give good results, but which is not strictly necessary. By this we can also determine the functional law of errors (Comp. the 14 next chapter). The equation of the curve is: Log2/ = 2-0228 + 0-0030(a'— 11) — 0-6885(«— ll)2+0-01515(j;— 11)3— 0-001678(^— 11)« § 15. By the study of many curves of presumptive errors, and especially such as represent ideal functional laws of errors, we cannot fail to get the impression that there exists a typical form of curves of errors, which is particularly distinguished by symmetry. Familiarity with this form is useful for the construction of curves of presumptive errors. But we must not expect to get it realised in all cases. For this reason I have considered it important to give, alongside of the typical curves, an example taken from real observa- tions of a skew curve of errors, which in consequence of its marked want of symmetry deviates considerably from the typical form. Fig. 4 shows this last mentioned law of presumptive errors. Deviation from the typical form does not indicate that the observations are not good. But it may become so glaring that we are forced by it to this conclusion. If, for instance, between the extreme values of repetitions — abscissae — there are intervals which are as free from finite ordinates as the space beyond the extremes, so that the curve of errors is divided into two or several smaller curves of errors beside one another, there can scarcely be any doubt that we have not a series of repetitions proper, but a combination of several; that is to say, different methods of observation have been used and the results mixed up together. In such cases we cannot expect that the law of large numbers will remain in force, and we had better, therefore, reject such observations, if we cannot retain them by tracing out the essential circumstances which distinguish the groups of the series, but have been overlooked. § 16. When a curve of presumptive errors is drawn, we can measure the magnitude of the ordinate for any given abscissa; so far then we know the law of errors perfectly, by means of the curve of errors, but certainly in the tabular form only, with all its copious- ness. Whether we can advance further depends on, whether we succeed in interpolating in the table so found, and particularly on, whether we can, either from the table or direct from the curve of errors, by measurement obtain a comparatively small number of constants, by which to determine the special peculiarities of the curve. By interpolating, by means of Newton's formula, the logarithms of the frequencies, or by drawing the curves of errors with the logarithms of the frequencies as ordinates, we often succeed, as above mentioned, in giving the curve the form of a parabola of low (and always even) degree. Still easier is it to make use of the circumstance that fairly typical curves of errors show a single maximum ordinate, and an inflexion on each side of it, near which the curve for a short distance is almost rectilinear. By measuring the co-ordinates of the maximum point and of the points of inflexion, we shall get data sufficient to enable us to 15 draw a curve of errors which, as a rule, will deviate very little from the original. All this, however, holds good only of the curves of presumptive errors. With the actual ones we cannot operate in this way, and the transition from the latter to tlie former seems in the meantime to depend on the eye's sense of beauty. V. FUNCTIONAL LAWS OF ERROES. § 17. Laws of errors may be represented in such a way that the frequency of the results of repetitions is stated as a mathematical function of the number, or numbers, expressing the results. This method only differs from that of curves of errors in the circumstance that the curve which represents the errors has been replaced by its mathema- tical formula; the relationship is so close that it is difficult, when we speak of these two methods, to maintain a strict distinction between them. In former works on the theory of observations the functional law of errors is the principal instrument. Its source is mathematical speculation; we start from the properties which are considered essential in ideally good observations. From these the formula for the typical functional law of errors is deduced; and then it remains to determine how to make computations with observations in order to obtain the most favourable or most probable results. Such investigations have been carried through with a high degree of refinement; but it must be regretted that in this way the real state of things is constantly disregarded. The study of the curves of actual errors and the functional forms of laws of actual errors have consequently been too much neglected. The representation of functional laws of errors, whether laws of actual errors or laws of presumptive errors founded on these, must necessarily begin with a table of the results of repetitions, and be founded on interpolation of this table. We may here be content to study the cases in which the arguments (i. e. the results of the repetitions) proceed by constant differences, and the interpolated function, which gives the frequency of the argument, is considered as the functional law of errors. Here the only difficulty we en- counter is that we cannot directly employ the usual Newtonian formula of interpolation, as this supposes that the function is an integral algebraic one, and gives infinite values for infinite arguments, whether positive or negative, whereas here the frequency of these infinite arguments must be = 0. We must therefore employ some artifice, and an obvious one is to interpolate, not the frequency itself, y, but its reciprocal, — . This, however, turns out to be inapplicable; for — will often become infinite for finite arguments, and will, at any rate, increase much faster than any integral function of low degree. 16 But, as we have already said, the interpolation generally succeeds, when we apply it to the logarithm of the frequency, assuming that Log y = a -\- hx ^ cx^ -\- . . . -\- gx"", where the function on the right side begins with the lowest powers of the argument x, and ends with an even power whose coefficient must be negative. Without this latter condition the computed frequency, would again become infinitely great for ,?; = -j- co . That the observed frequency is often = 0, and its logarithm — oo like — , does no harm. Of course we must leave out these frequencies of the interpolation, or replace them by very small finite frequencies, a few of which it may become necessary to select arbitrarily. As a rule it is possible to succeed by this means. In order to represent a given law of actual errors in this way, we must, according to the rule of interpolation, determine the coefficients a, b, c, . . . g, whose number must be at least as large as that of the various results of repetitions with which we have to deal. This determination, of course, is a troublesome business. Here also we may suppose that the law of presumptive errors is simpler than that of the actual errors. And though this, of course, does not imply that log y can be ex- pressed by a small number of terms containing the lowest powers of x, this supposition, nevertheless, is so obvious that it must, at any rate, be tried before any other. § 18. Among these, the simplest case, namely that in which Log y is a function of X of the second degree Log 2/ ^ a-\-bx — cx^, gives us the typical form for the functional law of errors, and for the curve of errors, or with other constants y = he-^\—) =-- /dO-°""n-^) , (2) where ^ = ^+T + n2 + rf:3 + --- = 2-71828. The function has therefore no other constants than those which may be interpreted as unit for the frequencies /(, and as zero m and unit n for the observed values; the corresponding typical curve of errors has therefore in all essentials a fixed form. The functional form of the typical law of errors has applications in mathematics which are almost as important as those of the exponential, logarithmic, and trigonometrical functions. In the theory of observations its importance is so great that, though it has been over-estimated by some writers, and though many good observations show presumptive as well as actual laws of errors that are not typical, yet every student must make himself perfectly familiar with its properties. (4) 17 Expanding the index we get so that the general function resolves itself into a product of three factors, the first of which is constant, the second an ordinary exponential function, while the third remains a typical functional law of errors. Long usage reduces this form to e "'; but this form cannot be recommended. In the majority of its purely mathematical applications e-'™' is preferable, unless (as in the whole theory of observations) the factor ^ in the index is to be preferred on account of the resulting simplification of most of the derived formulae. The differential coefficients of e~^\n) with regard to x are D^e^lU) = — «-«(,j;^ — 3n''a;)e~TW D^e" 2^1 ») = — w- " (x'^ — oti^ ■2x^-\-3- bn*x) e~~iu} X>8g- 2 (^) = n- 12 (a^e -^ bn^ 3x* + 3 • 5w* • Sa;^ — 1 . 3 • 5n^) e'^i^)' The law of the numerical coefficients (products of odd numbers and binomial numbers) is obvious. The general expression of D'e 2\n) can be got from a comparison of the coefficients to ( — m)'' of the two identical series for equation (3), one being the Taylor series, the other the product of e 2Vn/ and the two exponential series with m^ and m as arguments. It can also be induced from the differential equation n^iy+i^ + a; /)••+' ^ + (»• + 1) D^
V - ^ ^V + c^f- (6) where 1 IX — my
1-18133
•1645
- -313
•43
-•2
2-0
1-19629
01353
-0-271
0-41
-•3
2-1
1-20853
•1103
- -232
•38
-•3
2-2
1-21846
■0889
- -1%
•34
-•4
2-3
1-22643
-0710
- -163
•30
-•4
dz*
-2
-2
-2
-2
-2
-2
-1
— 1
— 1
— 1
-0
-0
—
2-4
25
26
£
dz
2-7 1
2-8 1
2-9 1
3-0 1
3^1 1
3-2 1
3-3 1
3-4 1
3-6 1
3-6 1
3-7 1
3-8 1
3-9 1
4-0 1
4-1 1
4-2 1
4-3 1
4-4
45
23277
23775
24163
25089
25159
25210
25247
25273
25292
25304
25313
25319
25326
25328
25329
26330
25.331
i = (?
-J''
dij
dz
(VTj
0-0561 -0-1.35 0-27
•04.39 — 110
•0340 - 089
•23
•20
d^
dz"
-0-4
- -4
— -3
24462 ^0261 - ^071
24691 -0198 - -0.56
24864 -0149 - 043
■0082
•0060
-0043
•0031
-0022
•0015
■025
•019
•014
•Oil
•008
•006
■07
•06
•04
•03
•02
•02
•2
•0011 - -(XH 01
■0007 - •OOS -01
-0005 — -002 01
•0
-0
-0
25323 0-0003 —0-001 001 -
-0002 - -001 •OO
•0001 - •OOl 00
•0001 - 000 00
■0001 — -000
■axx) — ■OCX)
■0000 - 000
•00
■(X)
-00
dz*
-16 — -3
•14 — -3
-11 - -2
24993 0-0111 —0-0.33 009 -02 03
19
dH d
Here jy, '^\, -r^ are, each of them, the same for positive and negative values
of z\ the other columns of the table change signs with z.
The interpolations are easily worked out by means of Taylor's theorem :
„..=, = ^ + S-C + J3-C' + 5S-C- + Ag-C- + ... (7)
and
The typical form for the functional law of errors (2) shows that the frequency is
always positive, and that it arranges itself symmetrically about the value x = >«, for which
the frequency has its maximum value ij = h. For x^m^n the frequency is z/ = A • 0-60653.
The corresponding points in the curve of errors are the points of inflexion. The area
between the curve of errors and the axis of abscissae, reckoned from the middle to a; = m 4^ w,
will be nh • 0"85562 ; and as the whole area from one asymptote to the other is nh l/2;r
= nh • 2-50663, only nh • 0-39769 of it falls outside either of the inflexions, consequently
not quite that sixth part (more exactly 16 per ct.) which is the foundation of the rule,
given in § 11, as to the limit between the great and small errors.
The above table shows how rapidly the function of the typical law of errors de-
creases toward zero. In almost all practical applications of the theory of observations
e~~i'^ = 0, if only z>b. Theoretically this superior assymptotical character of the function
is expressed in the important theorem that, for 2; = ^j^ oo , not only e"i'^ itself is =
but also all its differential coefficients; -and that, furthermore, all products of this function
by every algebraic integral function and by every exponential function, and all the differential
quotients of these products, are equal to zero.
In consequence of this theorem, the integral \e 2'^ dz ^ V2n can be computed
as the sum of equidistant values of e 2'' multiplied by the interval of the arguments
without any correction. This simple method of computation is not quite correct, the
underlying series for conversion of a sum into an integral being only semiconvergent in
this case; for very large intervals the error can be easily stated, but as far as intervals
of one unit the numbers taken out of our table are not sufficient to show this error.
If the curve of errors is to give relative frequency directly, the total area must be
1 = nhV27r; h consequently ought to be put = "^^•
Problem 1. Prove that every product of typical laws of errors in the functional
1 /x— m\2
form = he''~2\'~ir) ,with the same independent variable x, is itself a typical law of errors.
How do the constants h, m, and n change in such a multiplication?
20
Problem 2, How small are the frequencies of errors exceeding 2, 3, or 4 times
the mean error, on the supposition of the typical law of errors?
Problem 3. To find the values of the definite integrals
Sr = \x''e 2\ » / dx.
J— 00
Answer: sn+i = and $2, = 1 • 3 • 5 . . . (2« — 1) 9r"+'\/27c.
§ 19. Nearly related to the typical or exponential law of errors in functional form
are the binomial functions, which are known from the coefficients of the terms of the «"■
power of a binomial, regarded as a function of the number x of the term.
X
==
n
1
2
3
4
5
6
7
1
1 1
2
1 2
1
3
1 3
3
1
4
1 4
6
4
1
5
1 5
10
10
5
1
6
1 6
15
20
15
6
1
7
1 7
21
35
35
21
7
1
8
1 8
28
56
70
56
28
8
9
1 9
36
84
126
126
84
36
10
1 10
45
120
210
252
210
120
11
1 11
55
165
330
462
462
330
12
1 12
66
220
495
792
924
792
13
1 13
78
286
715
1287
1716
1716
14
1 14
91
364
1001
2002
3003
3432
For integral values of the argument the binomial function can be computed directly
by the formula
M^)
1 . 2 • 3 . . . n
1.2-3...a;.l-2-3...(«-
n{n — l)...{n-x+ 1)
-X)
fin (n — x)
(9)
1.2. ..a;
When the binomial numbers for n are known, those for m + 1 are easily found
by the formula
t3n+,(x) ^ fi„(x) + fin{X-l). (10)
By substitution according to (9) we easily demonstrate the proposition that, for
21
any integral values of w, r, and t
Mi)fin-t{r) = Mr)-fin-r(t), (11)
which means that, when the trinomial {a + b + c)" is developed, it is indifferent whether
we consider it to be ((a + b)-\- c)" or (a + (6 + c))".
For fractional values of the argument x, the binomial function j3„{x) can be taken
in an infinity of different ways, for instance by
. , , sin;:a;
This formula results from a direct application of Lagrange's method of interpolation, and
leads by (10) to the more general formula
^,,_ l'2., .n sin^ra; .^„,
'^"'^^ ~ (l-x){2-x)...{n^x) ~^tor' ^ '
This species of binomial function may be considered the simplest possible, and has
some importance in pure mathematics; but as an expression of frequencies of observed
values, or as a law of errors, it is inadmissible because, for x > n or x negative, it gives
negative values alternating with positive values periodically.
This, however, may be remedied. As /9o {x) has no other values than and 1,
when X is integral, we can put for instance
^o(^) = (
sin Tzx \^
by (10) then
TIX ) '
sin^ nx
^.(^) = [^.
l)^
2
+
(13)
(x—\f ' (x— 2)V ?r^
Here the values of the binomial function are constantly positive or 0. But this
form is cumbersome ; and although for x = co the function and its principal coefficients
are == 0, this property is lost here, when we multiply by integral algebraic or by exponen-
tial functions.
These unfavourable circumstances detract greatly from the merits of the binomial
functions as expressions for continuous laws of errors.
When, on the contrary, the observations correspond only to integral values of the
argument, the original binomial functions are most valuable means for treating them. That
Pn(x) = 0, if j; > w or negative, is then of great importance. But this case must be referred
to special investigations.
§ 20. To represent non-typical laws of errors in functional form we have now
the .choice between at least three different plans:
22
1) the formula (1) or
y ^ (,a+IJX+yXi+...-XX^'' ,
2) the products of integral algebraic functions by a typical function or (6)
y
I- k k l/I— m\'
3) a sum of several typical functions
y = l^lh^e-^ri^l . (14)
This account of the more prominent among the functional forms, which we have at our
disposal for the representation of laws of errors, may prove that we certainly possess good
instruments, by means of which we can even in more than one form find general series
adapted for the representation of laws of errors. We do not want forms for the series,
required in theoretical speculations upon laws of errors; nor is the exact representation of
the actual frequencies more than reasonably difficult. If anything, we have too many forms
and too few means of estimating their value correctly.
As to the important transition from laws of actual errors to those of presumptive
errors, the functional form of the law leaves us quite uncertain. The convergency of the
series is too irregular, and cannot in the least be foreseen.
We ask in vain for a fixed rule, by which we can select the most important and
trustworthy forms with limited numbers of constants, to be used in predictions. And even
if we should have decided to use only the typical form by the laws of presumptive errors,
we still lack a method by which we can compute its constants. The answer, that the
"adjustment" of the law of errors must be made by the "method of least squares", may
not be given till we have attained a satisfactory proof of that method; and the attempts
that have been made to deduce it by speculations on the functional laws of errors must,
I think, all be regarded as failures.
VI. LAWS OF ERRORS
EXPEESSED BY SYMMETRICAL FUNCTIONS.
§ 21. All constants in a functional law of errors, every general property of a
curve of errors or, generally, of a law of numerical errors, must be symmetrical functions
of the several results of the repetitions, i. e. functions which are not altered by inter-
changing two or more of the results. For, as all the values found by the repetitions
correspond to the same essential circumstances, no interchanging whatever can have any
influence on the law of errors. Conversely, any symmetrical function of the values of the
23
observations will represent some property or other of the law of errors. And we must be
able to express the whole law of errors itself by every such collection of symmetrical
functions, by which every property of the law of errors can be expressed as unambiguously
as by the very values found by the repetitions.
We have such a collection in the coefficients of that equation of the «"■ degree,
whose roots are the n observed values. For if we know these coefficients, and solve the
equation, we get an unambiguous determination of all the values resulting from the repe-
titions, i.e. the law of errors. But other collections also fulfil the same requirements; the
essential thing is that the n symmetrical functions are rational and integral, and that one
of them has each of the degrees 1, 2 . . . «, and that none of them can be deduced from
the others.
The collection of this sort that is easiest to compute, is the sums of the powers.
With the observed values
Oj, 0.^, O3, ... o„
we have
So - o'^,+ol + ..^Yol =n
s> = o^-\-o,^..^o„
<-\-o\+ •■ + <>« \ (15)
s, = 0"; + 0^ + . . -^ oi;
Q
and the fractions — may also be employed as an expression for the law of errors; it is
only important to reduce the observations to a suitable zero which must be an average
value oi 0^ . . ,o„; for if the diiferences between the observations are small, as compared
with their differences from the average, then
may become practically identical, and therefore unable to express more than one property
of the law of errors.
From a well known theorem of the theory of symmetrical functions, the equations /-^^"
1 + UiW + ai(u'^ + • • • = (1 — Oi (', ==>!, — m and }! ., = k^ — m^, the computation of one set of
constants by the other can, according to (17), be made by the formulae (19) and (21). We
substitute only in these the A;,- for the s,-, and )! or ), for /i.
It will be seen that the constants m and m, and the special typical law of errors
to which they belong, are generally superfluous. This superfluity in our transformation
may be useful in special cases for reasons of convergency, but in general it must be con-
sidered a source of vagueness, and the constants must be fixed arbitrarily.
It is easiest and most natural to put
m = X^ and n"^ = X.^.
In this case we get A-, =0, k^ =0, k.^ =kf^X^, k^ ^k„X^, k. ■=- k„X^, and further
k, ^ kAK + '^OXl)
k, = k,{X, + 3bX,X,)
k. = kAX, + b6X,X, + 3bX])
The law of the coefficients is explained by writing the right side of equation (30)
35
Expressed by half-invariants in this manner the explicit I'orm of equation (6) is
1 + — (
2
>i.
l+^((^-^)'-3^(x-/,)) +
(31)
VIII. LAWS OF ERROES OF FUNCTIONS OF OBSERVATIONS.
§ 26. There is nothing inconsistent with our definitions in speaking of laws of errors
relating to any group of quantities which, though not obtained by repeated observations,
have the like property, namely, that repeated estimations of a single thing give rise, owing
to errors of one kind or other, to multiple and slightly differing results which are prima
facie equally valid. The various forms of laws of actual errors are indeed only summary
expressions for such multiplicity; and the transition to the law of presumptive errors
requires, besides this, only that the multiplicity is caused by fixed but unknown circum-
stances, and that the values must be mutually independent in that sense that none of the
circumstances have connected some repetitions to others in a manner which cannot be
common to all. Compare § 24, Example 6.
It is, consequently, not difficult to define the law of errors for a function of one
single observation. Provided only that the function is univocal, we can from each of the
observed values Oj, o.^ ... o„ determine the corresponding value of the function, and
f{o^), f{0,), ...f(On)
will then be the series of repetitions in the law of errors of the function, and can be
treated quite like observations.
With respect, however, to those forms of laws of errors which make use of the
idea of frequency (probability) we must make one little reservation. Even though o,- and
Oi are different, we can have /"(o,) = f (Oi), and in this case the frequencies must evidently
be added together. Here, however, we need only just mention this, and remark that the
laws of errors when expressed by half-invariants or other symmetrical functions are not
influenced by it.
Otherwise the frequency is the same for /"(o,) as for o,-, and therefore also the
probability. The ordinates of the curves of errors are not changed by observations with
discontinuous values; but the abscissa o,- is replaced by /"(o,), and likewise the argument
in the functional law of errors. In continuous functions, on the other hand, it is the
areas between corresponding ordinates which must remain unchanged.
5*
36
In the form of symmetrical functions the law of errors of functions of observations
may be computed, and not only when we know all the several observed values, and can there-
fore compute, for each of them, the corresponding value of the function, and at last the
symmetrical functions of the latter. In many and important cases it is sufficient if we
know the symmetrical functions of the observations, as we can compute the symmetrical
functions of the functions directly from these. For instance, if f{o) = o^ ; for then the
sums of the powers s'„ of the squares are also sums of the powers s^ of the observations,
if only constantly m = 2m ; s'g = Sy, s', = s^ , s'„ = s^, etc.
§ 27. The principal thing is here a proposition as to laws of errors of the linear
functions by half-invariants.
It is almost self-evident that if o' = ao-\-h
/ii = a//
(32)
etc.
;j.'r = a>r (»•>!)
For the linear functions can always be considered as produced by the change of
both zero and unity of the observations (Compare (24)).
However special the linear function ao -\-b may be, we always in practice manage
to get on with the formula (32). That we can succeed in this is owing to a happy
circumstance, the very same as, in numerical solutions of the problems of exact mathematics,
brings it about that we are but rarely, in the neighbourhood of equal roots, compelled to
employ the formulae for the solution of other equations than those of the first degree.
Here we are favoured by the fact that we may suppose the errors in good observations
to be small, so small — to speak more exactly — that we may generally in repetitions
for each series of observations Oj, Oj, ... o„ assign a number c, so near them all that
the squares and products and higher powers of the differences
0, — 0,0^ — 0, . . . o„ — c
without any perceptible error may be left out of consideration in computing the function:
i. e., these differences are treated like differentials. The differential calculus gives a definite
method, in such circumstances, for transforming any function f(o) into a linear one
m = f(c) + f'{c)-(o-c).
The law of errors then becomes
(^ . ifio)) = f(c] + r (C) {[I , (0) - c) = /-(/i I (0)) \
37
But also by quite elementary means and easy artifices we may often transform
functions into others of linear form. If for instance f{o] = — , then we write
J_ _ 1 _ c-jo -c) _ J_ _ J_ , _ X
c-{-(o — c) c^' — io — c)-' c c' ^° ''-• '
and the law of errors is then
— --^if^Ao)-o)
- (I) - ^-M-
§ 28. With respect to functions of two or more observed quantities we may also,
in case of repetitions, speak of laws of errors, only we must define more closely what we
are to understand by repetitions. For then another consideration comes in, which was out
of the question in the simpler case. It is still necessary for the idea of the law of errors
of /'(o, o') that we should have, for each of the observed quantities o and o', a series of
statements which severally may be looked upon as repetitions:
0,, Oj, Om
o'n o\, o'„.
But here this is not sufficient. Now it makes a difference if, among the special
circumstances by o and o', there are or are not such as are common to observations of the
different series. We want a technical expression for this. Here it is not appropriate only
to speak of observations which are, respectively, dependent on one another or independent;
we are led to mistake the partial dependence of observations for the functional dependence
of exact quantities. I shall propose to designate these particular interdependences of
repetitions of different observations by the word "bond", which presumably cannot cause
any misunderstanding.
Among the repetitions of a single observation, no other bonds must be found than
such as equally bind all the repetitions together, and consequently belong to the pecularities
of the method. But while, for instance, several pieces cast in the same mould may be
fair repetitions of one another, and likewise one dimension measured once on each piece,
two or more dimensions measured on the same piece must generally be supposed to be
bound together. And thus there may easily exist bonds which, by community in a cir-
cumstance, as here the particularities in the several castings, bind some or all the repe-
titions of a series each to its repetition of another observation; and if observations thus
connected are to enter into the same calculation, we must generally take these bonds into
account. This, as a rule, can only be done by proposing a theory or hypothesis as to the
38
mathematical dependence between the observed objects and their common circumstance,
and whether the number which expresses this is known from observation or quite unknown,
the right treatment falls under those methods of adjustment which will be mentioned
later on.
It is then in a few special cases only that we can determine laws of errors for
functions of two or more observed quantities, in ways analogous to what holds good of a
single observation and its functions.
If the observations o, o', o" . . ., which are to enter into the calculation of
f{o, o\ o", . . .), are repeated in such a way that, in general, o,, o't, o", ... of the j'th
repetition are connected by a common circumstance, the same for each i, but otherwise
without any other bonds, we can for each i compute a value of the function y, =
f{Oi, Oi, Oi , . . .), and laws of errors can be determined for this, in just the same way as
for separately. To do so we need no knowledge at all of the special nature of the bonds.
§ 29. If, on the contrary, there is no bond at all between the repetitions of the
observations o, o', o", ... — and this is the principal case to which we must try to reduce
the others — then we must, in order to represent all the equally valid values oi y =
f{o, o\ o", . . .), herein combine every observed value for o with every one for o', for o",
etc., and all such values of y must be treated analogously to the simple repetitions of one
single observed quantity. But while it may here easily become too great a task to com-
pute y for each of the numerous combinations, we shall in this case be able to compute
y's law of errors by means of the laws of errors for o, o', o" ...
Concerning this a number of propositions might be laid down; but one of them
is of special importance and will be almost sufficient for us in what follows, viz., that
which teaches us to determine the law of errors for the sum of the observed quantities
and o'.
If the law of errors is given in the form of relative frequencies or probabilities,
f{o) for and l.
ea.
+^>:
(35)
When the errors of observation are sufficiently small, we shall also here generally
be able to give the most different functions a linear form. In consequence of this, the
propositions (34) and (35) acquire an almost universal importance, and afford nearly the
whole necessary foundation for the theory of the laws of errors of functions.
Example 1. Determine the square of the mean error for differences of the n^"^
order of equidistant tabular values, between which there is no bond, the square of the
mean error for every value being = l^.
40
A^iJ*) = >?2(o,— 403+602— 4o,+o„) = 70^2
;rj.) = 1.1.10.14 4n-2
/^i^; 1 2 3 4 w "*•
Example 2. By the observation of a meridional transit we observe two quantities,
viz. the time, t, when a star is covered behind a thread, and the distance, f, from the
meridian at that instant. But as it may be assumed that the time and the distance are
not connected by a bond, and as the speed of the star is constant and proportional to the
known value sin p {p = polar distance), we always state the observation by the one quan-
tity, the time when the very meridian is passed, which we compute by the formula =
t -\- f cosec p.
The mean error is
Example 3. A scale is constructed by making marks on it at regular intervals,
in such a way that the square of the mean error on each interval is = /j-
To measure the distance between two objects, we determine the distance of each
object from the nearest mark, the square of the mean error of this observation being = a',.
How great is the mean error in a measurement, by which there are n intervals between
the marks we use?
X, (length) = «>?,-!- 2/; .
Example 4. Two points are supposed to be determined by bond-free and equally
good (^2 = 1) measurements of their rectangular co-ordinates. The errors being small in
proportion to the distance, how great is the mean error in the distance J?
Ki^) = 2.
Example 5. Under the same suppositions, what is the mean error in the inclina-
tion to the a;-axis?
Example 6. Having three points in a plane determined in the same manner by
their rectangular co-ordinates (a;,,yj, (xj,?/,), {x^,^/^), find the mean error of the angle
at the point (d7,,yj)
J^ + Jl + J]
^An — ji J, --.
Ji, J^, J;i being the sides of the triangle; Jj opposite to (^,,2/i)-
41
Examples 7 and 8. Find the mean errors in determinations of the areas of a
trianglff and a plane quadrangle.
X, (triangle) = \ (d\ + 2/^ + Al); k, (quadrangle) = I ( J' + A'\.
§ 30. Non-linear functions of more than one argument present very great difficulties.
Even for integral rational functions no general expression for the law of errors can be found.
Nevertheless, even in this case it is possible to indicate a method for computing the half-
invariants of the function by means of those of the arguments. To do so it seems indis-
pensable to transform the laws of errors into the form of systems of suras of powers. If
= /"(o, o', ...0"") be integral and rational, both it and its powers 0'' can be written as
sums of terms of the standard form Eko'^ • o'^ . . . oC")'', and for every such term the sum
resulting from the combination of all repetitions is ksa • .s'j . . . s^P (including the cases
where a or b ox d may be = 0), s^P being the sum of all «'•> powers of the repetitions of
oW. Thus if Sr indicates the sum of the /•'•> powers of the function 0, we get
Sr = SkSa • s't . . . sP.
Of course, this operation is only practicable in the very simplest cases.
Example 1. Determine the mean value and mean deviation of the product oo' =
of two observations without bonds. Here S^ = s„s\ and generally Sr = s^s'^, consequently
the mean value Jlf, =n^n\ and
M, already takes the cumbersome form
Example 2. f]xpress exactly by the half-invariants of the co-ordinates the mean
value and the mean deviation of the square of the distance r^ =x''-\-y^, if x and y are
observed without bonds. Here
^oir'') = So(^)«o(«/)
«2 ('•' ) = ^4 i^) «o (y) + ^s^ {x) s^ (y) + So {x) s, [y)
and
//2 in = /U W + 4//3 i^)f^i (^) + 2 (/i, {x)y + 4fi, (x) ((,,{x))^ +
+ fi, (y) + ^iMAy)l^Ay) + 2(/., (ii)r + 4/., (y) {fz, (y)Y.
§ 31. The most important application of proposition (35) is certainly the deter-
mination of the law of errors of the mean value itself. The mean value
/'I = — (o, +02+...0„)
known half-invariants >ij, k^, ..
. ^r • • , we get according to (35)
^^ir^^)
= ~{^^+--- + ^i) = ^^
^iifij)
= i,(A, + ... + ;,) = l^,
and in general
= m'-^.Xr.
42
is, we know, a linear function of the observed values, and we may treat the law of errors
for f^^ according to the said proposition, not only where we look upon o,, ... On as per-
fectly unconnected, but also where we assume that they result from repetitions made
according to the same method. For, just like such repetitions, o,, ... o„ must not have
any other circumstances in common as connecting bonds than such as bind them all and
characterise the method.
As the law of presumptive errors of o, is just the same as for Og ...Om, with the
(37)
While, consequently, the presumptive mean of a mean value for m repetitions is
the presumptive mean itself, the mean error on the mean value u, is reduced to -r= of
Vm
the mean error on the single observation. When the number m is large, the formation
of mean values consequently reduces the uncertainty considerably; the reduction, however,
is proportionally greater with small than with large numbers. While already 4 repetitions
bring down the uncertainty to half of the original, 100 repetitions are necessary in order
to add one significant figure, and a million to add 3 figures to those due to the single
observation.
The higher half-invariants of //, are reduced still more. If the k^, k^, etc., of
the single observation are so large that the law of errors cannot be called typical, no very
great numbers of m will be necessary to realise the conditions ^ai/ii) = =^4(/ii) with
an approximation that is sufficient in practice. It ought to be observed that this reduction
is not only absolute, but it holds good also in relation to the corresponding power of the
mean error |/>ijj(;Uj)2; for (37) gives
r r r
>ir(/i,):(-i,(/i,))^= m "^.(^,:>?f),
which, for instance when tn = 4, shows that the deviation of ^3 from the typical form
which appears by means of only 4 repetitions, is halved; that of /l^ is divided by 4, that
of yij is divided by 8, etc. This shows clearly the reason why we attach great importance
to the typical form for the law of errors and make arrangements to abide by it in practice.
For it appears now that we possess in the formation of mean values a means of. making
the laws of errors typical, even where they were not so originally. Therefore the standard
rule for all practical observations is this: Take care not to neglect any opportunities of
43
repeating observations and parts of observations, so that you can directly form the mean
values which should be substituted for the observed results; and this is to be done espe-
cially in the case of observations of a novel character, or with peculiarities which lead us
to doubt whether the law of errors will be typical.
This remarkable property is peculiar, however, not to the mean only, but also,
though with less certainty, to any linear function of several observations, provided only
the coefficient of any single term is not so great relatively to the corresponding deviation
from the typical form that it throws all the other terms into the shade. From (35) it is
seen that, if the laws of errors of all the observations o, o', ... o<'") are typical, the law
of errors for any of their linear functions will be typical too. And if the laws of errors
are not typical, then that of the linear function will deviate relatively less than any of the
observations o, o', ... Om-
To avoid unnecessary complication we represent two terms of the linear function
simply by o and o'. The deviation from the typical zero, which appears in the r*-^ half-
invariants (>• > 2), measured by the corresponding power of the mean error, will be less
for = o-{-o' than for the most discrepant of the terms o .and o'.
The inequation
Ar = Ar
says only that, if the laws of errors for and 0' deviate unequally from the typical form,
it is the law of errors for that deviates most. But this involves
&^ (!)'
or more briefly
T > B\
where T is positive, r > 2.
When we introduce a positive quantity U, so that
r*" = f/' > R\
it is evident that (U -\- ly > (5 + 1)'', and it is easily demonstrated that (T -\- If >
{U-\-ir. ^
Remembering that a; -f j;-» > 2, if x>0, we get by the binomial formula
(c/r-f U-yf > U-\-U-' + 2'-2>{U^+ C/"V.
Consequently
(T + ir XU -{- ir =S (K + 1)'
or
6»
44
and
^ ^ (>ir + ^'r V _ (ir(0)r^
K ^ ^, + r,v (A.(O)r'
but this is the proposition we have asserted, for the extension to any number of terms
causes no difficulty.
But if it thus becomes a general law that the law of errors of linear functions
must more or less approach the typical form, the same must hold good also of all mode-
rately complex observations, such as those whose errors arise from a considerable number
of sources. The expression "source of errors" is employed to indicate circumstances which
undeniably influence the result, but which we have been obliged to pass over as unessential.
If we imagined these circumstances transferred to the class of essential circumstances, and
substantiated by subordinate observations, that which is now counted an observation would
occur as a function, into which the subordinate observations enter as independent variables;
and as we may assume, in the case of good observations, that the influence of each single
source of errors is small, this function may be regarded as linear. The approximation to
typical form which its law of errors would thus show, if we knew the laws of errors of
the sources of error, cannot be lost, simply because we, by passing them over as unessen-
tial, must consider the sources of error in the compound observation as unknown. More-
over, we may take it for granted that, in systematically arranged observations, every such
source of error as might dominate the rest will be the object of special investigation and,
if necessary, will be included among the essential circumstances or removed by corrective
calculations. The result then is that great deviations from the typical form of the law of
errors are rare in practice.
§ 32. It is of interest, of course, also to acquire knowledge of the laws of errors
for the determinations of /j^ and the higher half-invariants as functions of a given number
of repeated observations.
Here the method indicated in § 30 must be applied. But though the symmetry
of these functions and the identity of the laws of presumptive errors for Oj, o.^, . . . Om
afford very essential simplifications, still that method is too difficult. Not even for /x^ have
I discovered the general law of errors. In my "Almindelig lagttagelseslcere" , Kobenhavn
1889, I have published tables up to the eighth degree of products of the sums of powers
Sp s^ . . ., expressed by suras of terms of the form o\o'\o"*; these are here directly appli-
cable. In W. Fiedler: "Elemente der neueren Geometrie und der Algebra der bindren
Formen", Leipzig 1862, tables up to the 10* degree will be found. Their use is more
difficult, because they require the preliminary transformation of the Sp to the coefficients
Up of the rational equations § 21. There are such tables also in the Algebra by Meyer
Hirsch, and Cayley has given others in the Philosophical Transactions 1857 (Vol. 147,
45
p. 489). 1 have computed the four principal half-invariants of n^ :
mk^in.^) = (m — ^h
■m^/l.,{fi.,) = (w — l)H, + 2m {m — 1) X]
m'^l.ifx.,} = (m - l]H^ + r2m {m - 1)H,L, + 4»« (m - 1) (m - 2)^J +
+ 8ot2(w< — l)yi5 r (38)
m';},(/i,J = {m — l)^;^ + 24m (m — l)«/l6>i., + 32m (m - l)'^ (w — 2)^5/3 +
+ 8m (m — 1) (4m2 - 9m + Q)r, + 144m2 (m — l^XJ] +
+ 96m2 (m — 1) (>w — ^Ul^-z 4' 48m« (m — 1) A^
Here m is the number of repetitions.
Of /ia and pt^ only the mean values and the mean errors have been found:
tn'n,{/xs) = (m— l)(m — 2)^3,
mnAf^;) -■= {ni-\y'(}n-2)U,+9m{m-l){m-2y^(XJ,+r,)+ \ (39)
+ 6m '^ (m — 1) (m — 2)^^ ;
and
mUi (/i J = (m - 1) (m2 — 6m + 6)X, — 6m (m — 1) AJ
m';._, (^ J = (m — 1)2 (w2 — 6m + 6)Ha +
+ 8m (m — 1) (m''' — 6m + 6) (2m'' — Ibin + 15)/le-<2 +
+ 18m(m — l)(m — 2)(m — 4)(m''-6m + 6)/i5/i3+ I (40)
+ 2m (m — 1) (17m* — 204m3 + 852m« — 1404m -j- 828)>i; -|
+ 24m2 (m — 1) (3m« — 38m« + 150m — 138)^4 1', +
+ 144m''' (m — 1) (m — 2) (m — 4) (m — 5) /I J>1., +
4- 24m»(m — l)(m2— 6m + 24)/i*.
Further I know only that
m*X,{/i,) = (m— l)(m — 2)((m2-12m+12)>lf,— 60m/i2yi3}, (41)
m^/ii (//g) = (m — 1) (m« — 30^=^ + 150m'' — 240m + 120)/i6 —
— 30m (m — 1) (7m2 — 36m + 36)/ii /}., —
— 60m(m — 1) (m — 2) (3m — 8) ^J —
— 60m''(m — l)(m — 6)/l^ (42)
mn^ili-,) = (»«— l)(m — 2)(m*— 60m3+420m''— 720m + 360)^, —
— 630m (m — 1) (>« — 2) (jw^ — 8m + ^)k^X.> —
— 210m(m — l)(m — 2)(7m«— 48>«-|-60)/i^yi3 —
— 1260m2(m — l)(m — 2)(m — 10)yi3/i^ (43)
MU,{fts) = (m — l)(Hi« — 126)»6 + ]806w«— 8400w'' + 16800;«2-15120w + 5040)/( —
— 56w (w — 1) (31w« -^540/w'' -f 2340w'^ — 3600w + 1800)-i6yij —
— 1680w(m — 1)(ot — 2)(3ot=* — 40/»2-fl20>« — OGMs/ia —
— 70m(OT — l)(49w* — 720)»'' + 3168m«— 5400>M + 3240)/i5 —
— 840?w2 (m— 1) (7w« — 150^2 + 576w — 540)/i, /ij —
— 10080»«''' (m — 1) (m — 2) (/«'•' — 18?w + 40)yi5;.., —
— 840wi«(7« — l)(m''— 30^4-90);?;. (44)
Some ^I's of products of the /x,^, /i^, and /i, present in general the same charac-
teristics as the above formulae. The most prominent of these characteristics are:
1) It is easily explained that X^ is only to be found in the equation ^i(/«i) = ^i;
indeed no other half-invariant than the mean value can depend on the zero of the obser-
vations. In my computations this characteristic property has afforded a system of multiple
checks of the correctness of the above results.
2) All mean Xi(/ir) are functions of the 0* degree with regard to m, all squares
of mean errors ^^(/Jtr) are of the (—1)^' degree, and generally each ^(/ir) is a function
of the (1 — s)ti> degree, in perfect accordance with the law of large numbers.
3) The factor m — 1 appears universally as a necessary factor of ^(/ir), if only
r>l. If r is an odd number, even the factor m — 2 appears, and, likewise, if r is an
even number, this factor is constantly found in every term that is multiplied by one or
more Xs with odd indices. No obliquity of the law of errors can occur unless at least three
repetitions are under consideration.
4) Many particulars indicate these functions as compounds of factorials
(m — 1 ) (w — 2) ... [ni — r) and powers of m.
If, supposing the presumed law of errors to be typical, we put X^ = ^4 ^... = 0,
then some further inductions can be made. In this case the law of errors of fi^ may be
kUh) T I ^l(M T2 I / 2>l r\l=^ 1'+°°
e \1 ^'\1 +-=h--f^j2 =U(o)e»^do. (45)
\ / — 00
As to the squares of mean errors of fir we get under the same supposition:
^
(/^.) =
^^.
L,
(/^.) =
ir.
X,
(/««) =
m ^1
^
(//J =
24 ,.
m ^1 '
indicating that generally
(46)
47
This proposition is of very great interest. If we have a number ni of repetitions
at our disposal for the computation of a law of actual errors, tlien it will be seen that
the relative mean errors of ^/j, //j , pis ■■■fir are by no means uniform, but increase with
the index r. If m is large enough to give us /i, precisely and fi^ fairly well, then /i 3 and
ju, can be only approximately indicated; and the higher half-invariants are only to be
guessed, if the repetitions are not counted by thousands or millions.
As all numerical coefficients in yij (nr) increase with r, almost in the same degree
as the coefficients 1, 2, 6, and 24 of ^^, we must presume that the law of increasing
uncertainty of the half-invariants has a general character.
We have hitherto been justified in speaking of the principal half-invariants as the
complete collection of the fir's or ;ir's with the lowest indices, considering a complete series
of the first m half-invariants to be necessary to an unambiguous determination of a law of
errors for m repetitions.
We now accept that principle as a system of relative rank of the half-invariants
with increasing uncertainty and consequently with a decreasing importance of the half-
invariants with higher indices.
We need scarcely say that there are some special exceptions to this rule. For
instance if /i^ = — ^' , as in alternative experiments with equal chances for and against
(pitch and toss), then ).^[ii^) is reduced to = 3 X\, which is only of the (—2)'"' order.
§ 33. Now we can undertake to solve the main problem of the theory of obser-
vations, the transition from laws of actual errors to those of presumptive errors. Indeed
this problem is not a mathematical one, but it is eminently practical. To reason from the
actual state of a finite number of observations to the law governing infinitely numerous
presumed repetitions is an evident trespass; and it is a mere attempt at prophecy to
predict, by means of a law of presumptive errors, the results of future observations.
The struggle for life, however, compels us to consult the oracles. But the modern
oracles must be scientific; particularly when they are asked about numbers and quantities,
mathematical science does not renounce its right of criticism. We claim that confusion
of ideas and every ambiguous use of words must be carefully avoided; and the necessary
act of will must be restrained to the acceptation of fixed principles, which must agree
with the law of large numbers.
It is hardly possible to propose more satisfactory principles than the following:
The mean value of all available repetitions can he taken directly, without any
change, as an approximation to the presumptive mean.
If only one observation without repetition is known, it must itself, consequently,
be considered an approximation to the presumptive mean value.
The solitary value of any symmetrical and univocal function of repeated observations
48
must in the same way, as an isolated observation, be considered tlie presumptive mean of
this function, for instance /ir = ^, (fir)-
Thus, from the equations 37—41, we get by m repetitions:
;, = /ii
. m
,2
^ ^ {w — l)(w — 2)^''
m"
{f^^+j^^mf'')
[m — 1) (w, _ 2) (»j2 — \2m + 12)
(/^'^ +,^^1/^2/^3);
(47)
as to X^, P.7, /?8 it is preferable to use the equations 42 — 44 themselves, putting only
Inversely, if the presumptive law of errors is iinown in this way, or by adoption
of any theory or hypothesis, we predict the future observations, or functions of observations,
principalli/ by computing their presumptive mean values. These predictions however, though
univocal, are never to be considered as exact values, but only as the first and most impor-
tant terms of laws of errors.
If necessary, we complete our predictions with the mean errors and higher half-
invariants, computed for the predicted functions of observations by the presumed law of
errors, which itself belongs to the single observations. These supplements may often be
useful, nay necessary, for the correct interpretation of the prediction. The ancient oracles
did not release the questioner from thinking and from responsibility, nor do the modern
ones; yet there is a difference in the manner. If the crossing of a desert is calculated to
last 20 days, with a mean error of one day, then you would be very unwise, to be sure,
if you provided for exactly 20 days; by so doing you incur as great a probability of dying
as of living. Even with provisions for 21 days the journey is evidently dangerous. But
if you can carry with you provisions for 23—25 days, the undertaking may be reasonable.
Your life must be at stake to make you set out with provisions for only 17 daysor less.
In addition to the uncertainty provided against by the presumptive law of error,
the prediction may be vitiated by the uncertainty of the data of the presumptive law itself.
When this law has resulted from purely theoretical speculation, it is always impossible to
calculate its uncertainty. It may be quite exact, or partially or absolutely false, we are
left to choose between its admission and its rejection, as long as no trial of the prediction
by repeated observations has given us a corresponding law of actual errors, by which it
can be improved on.
49
If the law of presumptive errors has been computed by means of a law of actual
errors, we can, according to (37), employ the values ^j, /{g, ... and the number m of
actual observations for the determination of ^ri^i)- In this case the complete half-invari-
ants of a predicted single observation are given analogously to the law of errors of the sum
of two bondless observations by
Xr + Xrifli)-
Though we can in the same way compute the uncertainties of k^, X.^, and X^, it
is far more difficult, or rather impossible, to make use of these results for the improvement
of general predictions.
Of the higher half-invariants we can very seldom, if ever, get so much as a rough
estimate by the method of laws of actual errors. The same reasons that cause this
difficulty, render it a matter of less importance to obtain any precise determination.
Therefore the general rule of the formation of good laws of presumptive errors must be:
1. In determining X^ and X,^, rely almost entirely upon the actual observed values.
2. As to the half-invariants with high indices, say from X^ upwards, rely as
exclusively upon theoretical considerations.
3. Employ the indications obtainable by actual observed values for the intermediate
half-invariants as far as possible when you have the choice between the theories in (2).
From what is said above of the properties of the typical law of errors, it is evident
that no other theory can fairly rival it in the multiplicity and importance of its applications.
It is not only constantly applied when X^, X^, and ^5 are proved to be very small, but it
is used almost universally as long as the deviations are not very conspicuous. In these
cases also great efforts will be made to reduce the observations to the typical form by
modifying the methods or by substituting means of many observed values instead of the
non-typical single observations. The preference for the typical observations is intensified
by the difficulty of establishing an entirely correct method of adjustment (see the following
chapters) of observations which are not typical.
In those particular cases where X^ or X^ or ^^5 cannot be regarded as small, the
theoretical considerations (proposition 2 above) as to /^ g ^^^ ^^^ higher half-invariants ought
not to result in putting the latter = 0. As shown in " Videnskabernes Selskabs Oversigter" ,
1899, p. 140, such laws of errors correspond to divergent series or imply the existence of
imaginary observations. The coefficients kr of the functional law of errors (equation (6))
7
50
have this merit in preference to the half-invariants, that no term implies the existence
of any other.
This series
* (X) = k, •' + br") o]. (54)
This possibility is of some importance for the treatment of those cases in which
the single observations are bound. They must be treated then just like results, and we
must try to represent them as functions of the circumstances which they have in common,
and which must be given instead of them as original observations. This may be difficult
to do, but as a principle it must be possible, and functions of bound observations must
therefore always have laws of errors as well as others; only, in general, it is not possible
to compute these laws of errors correctly simply by means of the laws of errors of the
54
observations only, just as we cannot, in general, compute the law of errors for aE -\-bR"
by means of the laws of errors for E and E".
In example 5, § 29, we found the mean error in the determination of a direction R
between two points, which were given by bond-free and equally good (/.^{x) == X^iy) == 1)
2
measurements of their rectangular co-ordinates, viz.: Xi{Ii) = -^^, and then, in example 6,
we determined the angle V in a triangle whose points were determined in the same way.
It seems an obvious conclusion then that, as V = R' — R'\ we must have k^iV)
= X^_{R')-\-k^(R") = -^ +7p2- ^"* *'^i^ i^ "^*' correct ; the solution is ^ j( F) = ~-^,2- Tj ~ .
where J, J', and J" are the sides of the triangle. The cause of this is, of course, tliat
the co-ordinates of the angular point enter into both directions and bind R' and R" together.
But it is remarkable then that, when F is a right angle, the solutions are identical.
With equally good unbound observations, Oq, Oj, o^, and O3, we get
^2(02 — 2o, + 0o) = 6-ij(o)
^2(03—202+0,) = 6^2(0),
but
''2(03-302+30,-0,) = 20Xi{o),
although 03— 3o2+3o,— Ofl = (03— 2oj + o,) — (Oj — 2o, + 0o), according to which we
should expect to find
i?2(03— 3o2+3o, — Oo) = ;2(08— 203+0,) + /2(o2—2o, + 0o) = 12^2(0).
But if, on the other hand, we combine the two functions
R' = O0+60, — 402 and R" = 2oi+3o2 — O3,
where ki(R') = b?>Xi{p) and ;2(-K") = 14>i2(o), and from this compute ij for any function
aR'-\-bR", then, curiously enough, we get as the correct result X^iaR' -\-bR") =
(53a2 + 14J2)^3(o) = aU^(R') -\- bn,{R").
Gauss's general prohibition against regarding results of computations — especially
those of mean errors — from the same observations as analogous to unbound observations,
has long hampered the development of the theory of observations.
To Oppermann and, somewhat later, to Helmert is due the honour of having
discovered that the prohibition is not absolute, but that wide exceptions enable us to
simplify our calculations. . We must therefore study thoroughly the conditions on which
actually existing bonds may be harmless.
Let Oi,...o„ be mutually unbound observations with known laws of errors, ^1(0,),
,Jg(oj), of typical form. Let two general, linear functions of them be
[po] = p,o^ +...+;>„o„
[qo] = g,o, +... + ^„o„.
I (56)
55
For these then we know the laws of errors
X,{qo-\ = [qk,{o)l L,[ i] = [c'd'X]
[dbi]-~[abk]-[daJ\:[aaX] = [d'b'^, [dcX]-[acX]-[daX]:[aaX] = [d'c'i], [ddX]-[adX]-[daX]:[aaX] = [d'd'X\
[c'c'X]-[b'c'^ . [db'X\ : [b'b'A] = [c"c"Xl [c'd'X]-[b'd'X] • [<^b'X\ : [b'b'^ = [d'd"X]
[d'c>X] - [6'c'yi] '[d'b'X] : [b'b'Ji] = [d"c"X\, [d'd'X\-[b'd'^.[d'b'X] : [b'b'X] = [d"d"X]
[d"d"X] - [c"d"X] • [d"d'X] : [c'V'i] = [d"'d"'X]
As will be seen, there is a check by means of double computation for each of
the sums of the products properly so called. The sums of the squares are of special
importance as they are the squares of the mean errors of the transformed functions,
X^[ao] = [aaA], X^lb'o] = [b'b'X], i^[(f'o] = [c"c"X], and k^[d"'o] = [d"'d"'X].
Example. Five equally good, unbound observations Oj, Oj, O3, o^,and O5 represent
values of a table with equidistant arguments. The function tabulated is known to be an
integral algebraic one, not exceeding the 3''<' degree. The transformation into free functions
is to be carried out, in such a way that the higher diiferences are selected before the lower
ones. (Because JS certainly, J* etc., possibly, represent equations of condition). With symbols
for the diiferences, and with /(^(oi) = 1, we have then:
Function
VJo
0,
i-J'o,
FJ303-i#O3
"3 — 36^ "3
Fj03-4J^03+fFJ»03-iJ^03
Coefficients
Ooi +0oj -flOg +O0, +0oj
0-110
1-210
0-1 3-3 1
1 _4 6-4 1
3
35
3
"35
2
'35
2
T
12
35
12
35
i
5
1
10
17
35
"Y
17
3 6
12
35
3
T
_1
7
12
36
1
36
_1
7
A
3
"35
3
"35
12
35
2
7
Sums of the Products
Factors
1 -1
-1 2
-2
3
3
-6 -
6
-10
3
36
-2 3
6-
-10 -
-20
1
3 -6-
-10
20
35
-*
6-10-
-20
35
70
is selected
35 7
2 I
"~7 T
-1
-f
-1
5
f
is selected
1 7
36
1
I
is selected
are free
are both selected.
63
The complete set of free observations and the squares of their mean errors
are thus:
(0)= o, + J^o,-^y*o, = UOi + 02+0, + 0, + 0,), /,(0) = 1
(1) = Vjo,-iJ^o,+l(Vd^o,-U'o,) = ,V(-2o^-o, + o, + 2o,), >i,(l) = Jq
(2) = J^o.^^J'o, = i(2o,~o^-2o,-~o, + 2o,), ;.,(2) = f
(3)= VJ^o.-U^o, = H-Oi + 2o2-2o,+05), ^,(3) = |
(4)= J^03 = Oi-4o,+6o3-4o, + o, , ^.,(4) = 70
Through this and the preceding chapter we have got a basis which will generally
be sufficient for computations with observations and, in a wider sense, for computations
with numerical values which are not given in exact form, but only by their laws of errors.
We can, in the first place, compute the law of errors for a given, linear function of reci-
procally free observations whose laws of presumptive errors we know. By this we can
solve all problems in which there is not given a greater number of observations, and other
more or less exact data, than of the reciprocally independent unknown values of the
problem. When we, in such cases, by the means of the exact mathematics, have expressed
each of the unknown numbers as a function of the given observations, and when we have
succeeded in bringing these functions into a linear form, then we_ can, by (35), compute
the laws of errors for each of the unknown numbers.
Such a solution of a problem may be looked upon as a transformation, by which
n observed or in other ways given values are transformed into « functions, each corre-
sponding to its particular value among the independent, unknown values of the problem.
It lies often near thus to look upon the solution of a problem as a transformation, when
the solution of the problem is not the end but only the means of determining other un-
known quantities, perhaps many other, which are all explicit functions of the independent
unknowns of the problem. Thus, for instance, we compute the 6 elements of the orbit of
a planet by the rectascensions and declinations corresponding to 3 times, not precisely as
our end, but in order thereby to be able to compute ephemerides of the future places of
the planet. But while the validity of this view is absolute in exact mathematics, it
is only limited when we want to determine the presumptive laws of errors Of sought
functions by the given laws of errors for the observations. Only the mean values, sought
as well as given, can be treated just as exact quantities, and with these the general linear
transformation of n given into u sought numbers, with altogether n^ arbitrary constants,
remains valid, as also the employment of the found mean numbers as independent variables
in the mean value of the explicit functions.
If we want also correctly to determine the mean errors, we may employ no other
transformation than that into free functions. And if, to some extent, we may choose the
64
independent unknowns of the problem as we please, we may often succeed in carrying
through the treatment of a problem by transformation into free functions; for an unknown
number may be chosen quite arbitrarily in all its n coefficients, and each of the following
unknowns looses, as a function of the observations, only an arbitrary coefficient in com-
parison to the preceding one; even the w"" unknown can still get an arbitrary factor.
Altogether are \n[n-\-l) of the n"- coefficients of these transformations arbitrary.
But if the problem does not admit of any solution through a transformation into
free functions, the mean errors for the several unknowns, no matter how many there
may be, can be computed only in such a way that each of the sought numbers are directly
expressed as a linear function of the observations. The same holds good also when the
laws of errors of the observations are not typical, and we are to examine how it is with
X-i and the higher half-invariants in the laws of errors of the sought functions.
Still greater importance, nay a privileged position as the only legitimate proceeding,
gets the transformation into a complete set of free functions in the over-determined problems,
which are rejected as self-contradictory in exact mathematics. When we have a collection
of observations whose number is greater than the number of the independent unknowns
of the problem, then the question will be to determine laws of actual errors from the
standpoint of the observations. We must mediate between the observations that contradict
one another, in order to determine their mean numbers, and the discrepancies themselves
must be employed to determine their mean deviations, etc. But as we have not to do with
repetitions, the discrepancies conceal themselves behind the changes of the circumstances
and require transformations for their detection. All the functions of the observations
which, as the problem is over-determined, have theoretically necessary values , as , for
instance, the sum of the angles of a plane triangle, must be selected for special use.
Besides, those of the unknowns of the problem, to the determination of which the theory
does not contribute, must come forth by the transformation by which the problem is to
be solved.
As we shall see in the following chapters on Adjustment, it becomes of essential
moment here that we transform into a system of free functions. The transformation begins
with mutually free observations, and must not itself introduce any bond, because the trans-
formed functions in various ways must come forth as observations which determine laws
of actual errors.
X. ADJUSTMENT.
§ 4.3. Pursuing the plan indicated in § 5 we now proceed to treat the determina-
tion of laws of errors in some of the cases of observations made under varying or different
65
essential circumstances. But here we must be content with very small results. The
general problem will hardly ever be solved. The necessary equations must be taken from
the totality of the hypotheses or theories which express all the terms of each law of error
— say their half-invariants — as functions of the varying or wholly different circumstances
of the observations. Without great regret, however, the multiplicity of these theoretical
equations can be reduced considerably, if we suppose all the laws of errors to be exclusively
of the typical form. •
For each observation we need then only two theoretical equations, one representing
its presumptive mean value ii(o,), the other the square of its mean error X^(Oi), as func-
tions of the essential circumstances. But the theoretical equations will generally contain
other unknown quantities, the arbitrary constants of the theory, and these must be elimi-
nated or determined together with the laws of errors. The complexity is still great enough
to require a further reduction.
We must, preliminarily at all events, suppose the mean errors to be given directly
by theory, or at least their mutual ratios, the weights. If not, the problems require a
solution by the indirect proceeding. Hypothetical assumptions concerning the X,^ {Oi) are
used in the first approximation and checked and corrected by special operations which, as
far as possible, we shall try to expose beside the several solutions, using for brevity the
word "criticism" for these and other operations connected with them.
But even if we confine our theoretical equations to the presumptive means ^1(0,)
and the arbitrary unknown quantities of the theory, the solutions will only be possible if
we further suppose the theoretical equations to be linear or reducible to this form.
Moreover, it will generally be necessary to regard as exactly given many quantities really
found by observation, on the supposition only that the corresponding mean errors will be
small enough to render such irregularity inoffensive.
In the solution of such problems we must rely on the found propositions about
functions of observations with exactly given coefficients. In the theoretical equations of
each problem sets of such functions will present themselves, some functions appearing as
given, others as required. The observations, as independent variables of these functions,
are, now the given observed values 0,, now the presumptive means ^1(0,); the latter are,
for instance, among the unknown quantities required for the exact satisfaction of the
theoretical equations.
What is said here provisionally about the problems that will be treated in the
following, can be illustrated by the simplest case (discussed above) of n repetitions of the
same observation, resulting in the observed values Oi, ... o„. If we here write the theo-
retical equations without introducing any unnecessary unknown quantities, they will show
the forms == .^i(o,) — X^iot) or, generally, = /i,[a(o, — oj)]. But these equations are
9
66
evidently not sufficient for ttie determination of any ^i(Oi), which they only give if another
>i,(Oifc) is found beforehand. The sought common mean cannot be formed by the introduc-
tion of the observed values into any function [a{0i — 0t)], these erroneous values of the
functions being useful only to check X.^ (Oj) by our criticism. But we must remember what
we ijnow about free functions: that the whole system of these functions [a(o, — oj.)] is only
a partial system, with w — 1 differences Oi — ot as representatives. The only w* functions
which can be free of this partial system, must evidently be proportional to the sum
0, -\- . . . -\- o„, and by this we find the sought determination by
^l{Oi) = ^(Oi + ••• + "„),
the presumptive mean being equal to the actual mean of the observed values.
If we thus consider a general series of unbound observations, o i , ... o„, it is of
the greatest importance to notice first that two sorts of special cases may occur, in which
our problem may be solved immediately. It may be that the theoretical equations concern-
ing the observations leave some of the observations, for instance o^, quite untouched; it
may be also that the theory fully determines certain others of the observations, for
instance o„.
In the former case, that is when none of all the theories in any way concern the
observation o,, it is evident that the observed value o, must be approved unconditionally.
Even though this observation does not represent any mean value found by repetitions, but
stands quite isolated, it must be accepted as the mean A,(Oi) in its law of presumptive
errors, and the corresponding square of the mean error ?..^{0i) must then be taken,
unchanged, from the assumed investigations of the method of observation.
If, in the latter case, o„ is an observation which directly concerns a quantity that
can be determined theoretically (for instance the sum of the angles of a rectilinear triangle),
then it is, as such, quite superfluous as long as the theory is maintained, and then it
must in all further computations be replaced by the theoretically given value; and in
the same way X„ (o„) must be replaced by zero, as the square of the mean error on the
errorless theoretical value.
The only possible meaning of such superfluous observations must be to test the
correctness of the theory for approbation or rejection (a third result is impossible when
we are dealing with any real theory or hypothesis), or to be used in the criticism.
In such a test it must be assumed that the theoretical value corresponding to o„,
which we will call u„, is identical with the mean value in the law of presumptive errors
for o„, consequently, that M„ = /ii(o„), and the condition of an affirmative result must be
obtained from the square of the deviation, (o„ — m„)2 in comparison with X^iOn). The
67
equation (o„ — m„)'^ = /ig (o„) need not be exactly satisfied, but the approximation must
at any rate be so close that we may expect to find ^2(^«) coming out as the mean of
numerous observed values of (o„ — u„)^. Compare § 34.
§ 44. If then all the observations o^ ... o„ fall under one or the other of these two
cases, the matter is simple enough. But generally the observations o, will be connected
by theoretical equations of condition which, separately, are insufficient for the determination
of the single ones. Then the question is whether we can transform the series of observations
in such a way that a clear separation between the two opposite relations to the theory
can be made, so that some of the transformed functions of the observations, which must
be mutually free in order to be treated as unbound observations, become quite independent
of the theory, while the rest are entirely dependent on it. This can be done, and the
computation with observations in consequence of these principles, is what we mean by
the word "adjustment".
For as every theory can be fully expressed by a certain number, n — m, of theoretical
equations which give the exact values of the same number of mutually independent linear
functions, and as we are able, as we have seen, from every observation or linear function
of the observations, in one single way, to separate a function which is free and independent
of these just named theoretically given functions, and which must thus enter into another
system, represented by m functions, this system must include all those functions of the ob-
servations which are independent of the theory and cannot be determined by it. flach of the
thus mutually separated systems can be imagined to be represented, the theoretical system by
n — m, the non -theoretical or empirical system by m mutually free functions, which together
represent all observations and all linear functions of the same, and which may be looked
upon as a complete, transformed system of free functions, consequently as unbound obser-
vations. The two systems can be separated in a single way only, although the represen-
tation of each partial system, by free functions, can occur in many ways.
It is the idea of the adjustment, by means of this transformation, to give the
theory its due and the observations theirs, in such a way that every function of the theo-
retical system, and particularly the n — m free representatives of the same, are exchanged,
each with its theoretically given value, which, pursuant to the theory, is free of error. On
the other hand, every function of the empiric system and, particularly, its m free representa-
tives remain unchanged as the observations determine them. Every general function of
the n observations \do\ and, particularly, the observations themselves are during the adjust-
ment split into two univocally determined addenda: the theoretical function \d'o\ which
should have a fixed value D\ and the non-theoretical one \d"d]. The former \d'o\ is by
the adjustment changed into D' and made errorless, the latter is not changed at all. The
result of the adjustment, D'-\-\d"d\, is called the adjusted value of the function, and may
68
be indicated as [du], the adjusted values of the observations themselves being written
M,...M„. The forms of the functions are not broken, as the distributive principle f{x-[-t/)
= f{^)-\-f{y) liolds good of every homogeneous linear function.
The determination of the adjusted values is analogous to the formation of the
mean values of laws of errors by repetitions. For theoretically determined functions the
adjusted value is the mean value on the very law of presumptive errors; for the functions
that are free of the whole theory, we have the extreme opposite limiting case, mean values
represented by an isolated, single observation. In general the adjusted values \du\ are ana-
logous to actual mean values by a more or less numerous series of repetitions. For while
X^(\do\) = X^\d'o} + X^[d"o\, we have X^idu] == X^{D')^ Li[d"u] = k.[d"o], consequently
smaller than k^ {do]. The ratio -fj^ is analogous to the number of the repetitions or
the weight of the mean value.
§ 45. By "criticism" we mean the trial of the — hypothetical or theoretical —
suppositions, which have been made in the adjustment, with respect to the mean errors of
the observations; new determinations of the mean errors, analogous to the determinations
by the square of the mean deviations, n^, will, eventually also fall under this. The basis of
the criticism must be taken from a comparison of the observed and the adjusted values,
for instance the differences [do] — [du]. According to the principle of § 34 we must expect
the square of such a difference, on an average, to agree with the square of the correspon-
ding mean error, X^ ([do] — [du]) , but as [do] — [du] = [d'o] — D, and X,i[d'o] =
X-i [do] — X^ [du], we get
X^ ([do] — [du]) = X^ [do] — X, [dti] , (68)
which, by way of parenthesis, shows that the observed and the adjusted values of the same
function or observation cannot in general be mutually free. We ought then to have
([do]-[du] )^ _
X^[do]-X^[du] ^^^>
on the average; and for a sum of terms of this form we must expect the mean to approach
the number of the terms, nota bene, if there are no bonds between the functions [rfo]— [rfj<];
but in general such bonds will be present, produced by the adjustment or by the selec-
tion of the functions.
It is no help if we select the original and unbound observations themselves, and
consequently form sums such as
(o — u)^
{X,(o)-XAu)y
for after the adjustment and its change of the mean errors, u^ ... u„ are not generally
free functions such as Oi...o„. Only one single choice is immediately safe, viz., to stick
to the system of the mutually free functions which, in the adjustment, have themselves
69
represented the observations: the n — m theoreticallly given functions and the m which
the adjustment determines by the observations. Only of these we know that they are free
both before and after the adjustment. And as the differences of the last-mentioned m
functions identically vanish, the criticism must be based upon the n — m terms corresponding
to the theoretically free functions [ao\ = A,. .. [6'o] = B' of the series
{[ao]-AY
••• +
(\b'o\ - B'Y
([ao\-AY
+
([b'o-\-Fl_
(70)
„2 [ao] — >lj [aw] I ••• ' >lj [6'o] — /ij [6'm] [aak^] i ••• ' [ft'i'yij]
the sum of which must be expected to be = w — m.
Of course we must not expect this equation to be strictly satisfied; according to
the second equation (46) the square of the mean error on 1, as the expected value of each
term of the series, ought to be put down = 2 ; for the whole series, consequently, we can
put down the expected value as n — m^'\/2(n— m).
But now we can make use of the proposition (66) concerning the free functions.
It offers us the advantage that we can base the criticism on the deviations of the several
observations from their adjusted values, the latter, we know, being such a special set of
values as may be compared to the observations like v^ ...v„ loc. cit. ; u^ . . . «„ are only
distinguished from v^ . . .v„ by giving the functions which are free of the theory the same
values as the observations. We have consequently
([g o] -Ay
{[b'o]-B')^
(O — M)'^
L ^-A")
= n — m^Vn — m. (71)
tioned
[b'b'X^-]
If we compare the sum on the right side in this expression with the above men-
, which we dare not approve on account of the bonds produced by
(0 — m)'
, by the diminution of the denomi-
^2 (o) — k^ (M)
the adjustment, then there is no decided contradiction between putting down
at the smaller value n — m only, while
U^ (o) — Xi (u)
nators, can get the value n; only we can get no certainty for it.
The ratios between the corresponding terms in these two sums of squares, conse^
quently
I, (o) — X^ (M)
1
^2 (m)
k,{0)
, we call "scales", viz. scales for measuring the influence
of the adjustment on the single observation
1 ^2 [du]
More generally we call
the scale for the function [do].
(72)
/?2 [do]
If the scale for a function- or observation has its greatest possible value, viz. 1,
A^[du] = 0. The theory has then entirely decided the result of the adjustment. But if
the scale sinks to its lowest limit = 0, we get just the reverse k2[du] = /{^[do], i. e. the
theory has had no influence at all; the whole determination is based on the accidental
70
value of the observation, and for observations in this case we get
jo — u)^
Even though the scale has a finite, but very small value it will be inadmissible to de-
pend on the value of such a term becoming = 1. We understand now, therefore, the
(0 - M)
superiority of the sum of the squares
{o — M)2
1^2 {0)
AA«)
L, {0)
m to the sum of the squares
n as a bearer of the summary criticism.
We may also very well, on principle, sharpen the demand for adjustment on the
(0 — m)21
must
part of the criticism, so that not only the whole sum of the squares
approach the value n — m, but also partial sums, extracted from the same, or even its
several terms, must approach certain values. Only, they are not to be added up as
numbers of units, but must be sums of the scales of the corresponding terms,
we may trust to the sum of the squares , ;■ /-— -^
ciously applied, may be considered as fully justified.
(o — M)2
hip)
So much
that this principle, when judi-
The sum of the squares
h [0]
possesses an interesting property which all
other authors have used as the basis of the adjustment, under the name of "the method of
the least squares". The above sum of the squares gets by the adjustment the least possible
value that - '
X, {0)
can get for values v^ . . . i?„ which satisfy the conditions of the theory.
The proposition (66) concerning the free functions shows that the condition of this minimum
is that [c"o^^ = [c"m], . . . \d"'o'] = \d"'u\ for all the free functions which are determined by
the observations, consequently just by putting for each v the corresponding adjusted
value u.
§ 46. The carrying out of adjustments depends of course to a high degree on the
form in which the theory is given. The theoretical equations will generally include some
observations and, beside these, some unknown quantities, elements, in smaller number than
those of the equations, which we just want to determine through the adjustment. This
general form, however, is unpractical, and may also easily be transformed through the
usual mathematical processes of elimination. We always go back to one or the other of
two extreme forms which it is easy to handle: either, we assume that all the elements
are eliminated, so that the theory is given as above assumed by n — m linear equations
of condition with theoretically given coefficients and values, adjustment by correlates; or,
we manage to get an equation for each observation, consequently no equations of condition
between several observations. This is easily attained by making the number of the elements
as large {^m) as may be necessary: we may for instance give some values of observations
the name of elements. This sort of adjustment is called adjustment by elements. We
71
shall discuss these two forms in the following chapters XI and XII, first the adjustment by
correlates whose rules it is easiest to deduce. In practice we prefer adjustment by correlates
when m is nearly as large as n, adjustment by elements when m is small.
XL ADJUSTMENT BY COREELATES.
§ 47. We suppose we have ascertained that the whole theory is expressed in the
equations [au] ^ A, ... [cu'] = C, where the adjusted values u of the n observations
are the only unknown quantities; we prefer in doubtful cases to have too many equations
rather than too few, and occasionally a supernumerary equation to check the computation.
The first thing the adjustment by correlates then requires is that the functions \ao\ ... [coj,
corresponding to these equations, are made free of one another by the schedule in § 42.
Let [ao]., . .. \c"o\ indicate the n — m mutually free functions which we have got
by this operation, and let us, beside these, imagine the system of free functions completed
by m other arbitrarily selected functions , [d"'o], . . . [^"o] , representatives of the empiric
functions; the adjustment is then principally made by introducing the theoretical values
into this system of free functions. It is finally accomplished by transforming back from
the free modified functions to the adjusted observations. For this inverse transformation,
according to (62), the n equations are:
and according to (35) (compare also (63))
hioi) = { ^^^ • >i, [ao] + + ^jJ^ x,[g''o^]
«.^ I , c"' I dr I ^ gf \,,, ^
= {
(74)
As the adjustment influences only the w — w* first terms of each of these equations,
we have, because [au] = A, . . . [c"u] = C", and Xil^u] = ... = l^ \c"u\ = 0,
{n r" il'" Q" 1
, - ; ^ ^ + . . . + ■■„,■, , C" + Y:J^,'r~^ \d"'o\ + . . . + [g'of.X^io,) (75)
and
72
Consequently
and
, , J [ao\-A , , „\c"o]-C"\
^.(o.)-^.(«.) = ^:Nl[aai;] + "- + [o-?Q7j) = ^^(''•-"■)- (78)
Thus for the computation of all the diiferences between the observed and adjusted values
of the several observations and the squares of their mean errors, and thereby indirectly for
the whole adjustment, we need but use the values and the mean errors of the several
observations, the coefficients in the theoretically given functions, and the two values of
each of these, namely, the theoretical value, and the value which the observations would
give them.
The factors in the expression for o,- — «<,■,
_ [ m]~A _ [c>'o]-C"
^"~" [aaX.y ^'" - [c"c"A,\ '
which are common to all the observations, are called correlates, and have given the method
its name. The adjusted, improved values of the observations are computed in the easiest
way by the formula
M. = 0. - X, (Oi){aiKa + . . . + c7K,n}. (79)
By writing the equation (78)
and summing up for all values of i from 1 to n, we demonstrate the proposition concerning
the sum of the scales discussed in the preceding chapter, viz.
^ _;,(«)
X,(o)
§ 48. It deserves to be noticed that all these equations are homogeneous with
respect to the symbol /ig. Therefore it makes no change at all in the results of the
adjustment or the computation of the scales, if our assumed knowledge of the mean errors
in the several observations has failed by a wrong estimate of the unity of the mean errors,
if only the proportionality is preserved; we can adjust correctly if we know only the
relative weights of the observations. The homogeneousness is not broken till we reach the
equations of the criticism :
73
{[ao]-A)^
+
([c"o]-C")^
_ [o-iif
L h{o) J
= \{aKa + . . . + c"K,..)HM^ = n-m ± \/2{n-m)
(82)
It follows that criticism in this form, the "summary criticism", can only be used to try
the correctness of the hypothetical unity of the mean errors, or to determine this if it
has originally been quite unknown. The special criticism, on the other hand, can, where
the series of observations is divided into groups, give fuller information through the sums
of squares
2^^^ =^(l-f;4), (83)
taken for each group. We may, for instance, test or determine the unities of the mean
errors for one group by means of observations of angles, for another by measurements of
distances, etc.
The criticism has also other means at its disposal. Thus the differences (0 — u)
ought to be small, particularly those whose mean errors have been small, and they ought
to change their signs in such a way that approximately
k^ (o.)
(84)
for natural or accidentally selected groups, especially for such series of observations as are
nearly repetitions, the essential circumstances having varied very little.
If, ultimately, the observations can be arranged systematically, either according to
essential circumstances or to such as are considered inessential, we must expect frequent
and irregular changes of the signs of — n. If not, we are to suspect the observations of
systematical errors, the theory proving to be insufficient.
§ 49. It will not be superfluous to present in the form of a schedule of the
adjustment by correlates what has been said here, also as to the working out of the free
functions. We suppose then that, among 4 unbound observations o,, 0^, O3, and o^, with
the squares on their mean errors -^2(01)1 '<2(''2)i ^ii'^s)^ ^^^ ^A"^*)^ there exist relations
which can be expressed by the three theoretical equations
\au\ = a,M, -|- a^ii^ -\- a^u^ -\- a^u^ = A
\hu\ = 6j«, + 62M2 + ^^■i'"'3 + ^4**4 "= -S
\cu'\ = r-iM, +C.^M2 +C3?/a +C.^U^ == C.
The schedule is then as follows:
10
74
The given
A B
0, /},(«, ) a,
0, ^AOt) «4
C
[ao] [60] [co]
[aaX\ [abX\ [acX\
[baX] [bbJi] [bcX]
[caX\ [cbX] [ccX\
/? =
Con-elates K,
[baX\ [caX\
[aa^] ' ' [aaX]
[ao] — A
[aaX]
Free functions
C
c'
c'
c'
ir
K
K
K
\h'o\ [c'o]
[b'b'k] ib'c'X]
[C'6'yi] [C'c'/i]
\_c'b'X\
" {b'b'X\
[6'o]-g
ib'b'k]
C"
[c"o]
[c"c">i]
Adjusted values
Scales
0,-M, M, ;,(o,-M,) >!,(m,) 1-;,{m,)
^2(0,)
Kc" ==
[c"o]-C"
[cVX]
= 3 as proof.
Criticism
{o,-uy:X,{o,)
Sum for proof
and summary criticism
The free functions are computed by means of:
C = C-rA
c'i = d — yai
B' = B-^A
b'i = bi—^Ui
\b'o] = [H-/?M
[c'6'yi] = \c.bX\-p\caX]
Ydo\ = [c6] — Y \a6\
[c'c'K] = {ccX\-r\ca^
C"
e'{
[c"o]
C'-r'B
c'i-r'b'i
[c'o]-r'[6'o]
[c"c"X] = [c'c:x\-rVh'x\
By the adjustment properly so called we compute
Oi — Ui == {aiKa^b'iKt, -\-c'iKci<)Xi{o,)
I n% i/n c"^ \
and for the summary criticism
KliaaX^-] + K\, [b'b'k^-\ + ifV- [c'-c'-yi^] =
LM J
= 3±l/6.
In order to get a check we ought further to compute [r«^] = A, \hu\ = B, and
[cm] = C, with the values we have found for m,, m^, m^, and u^. Moreover it is useful to
add a superfluous theoretical equation, for instance [(a-{-b-{-c)u\ = A-\-B-\-C, through the
75
computation of the free functions, which is correct only if such a superfluity leads to
identical results.
§ 50. It is a deficiency in the adjustment by correlates that it cannot well be
employed as an intermediate link in a computation that goes beyond it. The method is
good as far as the determination of the adjusted values of the several observations and
the criticism on the same, but no farther. We are often in want of the adjusted values
with determinations of the mean errors of certain functions of the observations; in order
to solve such problems the adjustment by correlates must be made in a modified form.
The simplest course is, I think, immediately after drawing up the theoretical equations of
condition to annex the whole series of the functions that are to be examined, for instance
[rfo], . . . [eo], and include them in the computation of the free functions. In doing so we
must take care not to mix up the theoretically and the empirically determined functions,
so that the order of the operation must unconditionally give the precedence to the
theoretical functions; the others are not made free till the treatment of these is quite
finished. The functions [d"'o], . . . | e^o] , which are separated from these — it is scarcely
necessary to mention it — remain unchanged by the adjustment both in value and in
mean error. And at last the adjusted functions [du], . . .[eu], by retrograde transformation,"
are determined as linear functions of A, B\ C", [d"'o], . . . [e^o].
Example 1. In a plane triangle each angle has been measured several times, all
measurements being made according to the same method, bondfree and with the same
(unknown) mean error:
for angle A has been found 70° 0' 5" as the mean number of 6 measurements
1) » jd t) » ti 50° 3 » » » " » 10 *'
C <> » .1 60° 0' 2" .. » » » » 15
The adjusted values for the angles are then 70°, 50°, and 60°, the mean error for single
measurement = V/SOO == 17"3, the scales 0-5, 0-3, and 0-2.
Example 2. (Comp. example § 42.) Five equidistant tabular values, 12, 19, 29,
41, 55, have been obtained by taking approximate round values from an exact table, from
which reason their mean errors are all = \/-^. The adjustment is performed under the
successive hypotheses that the table belongs to a function of the 3"", 2°^, and 1"' degree,
and the hypothesis of the second degree is varied by the special hypothesis that the 2°''
dilierence is exactly = 2, in the following schedule marked (or). The same schedule may
be used for all four modifications of the problem, so that in the sums to the right in the
schedule, the first term corresponds to the first modification only, and the sum of the
two first terms to the second modification:
10*
76
^2(0)
12
19
29
41
55
1
12
1
1 2
1
12
1
12
1
1 2
J*
VJ'
J2
(F#)'
(J7=(^T
(or
2)
(or 2)
1
_JL
2
1
-4
-1
1
1
-^
6
3
—2
2
-4
-3
1
-1
1
T
1
1
1
5
2
T
1
2
1
16
7
70
35
20
12
12
12
35
12
20
12
10
12
6
2 4
20
1 2
10
1 2
6
12
1
42
/? =
_ 1
- 2 '
r=-f
r' = o
^K.
1
70
^K.=
= -
"B' T2
i(',„=8(orl)
^( 1+ 7+160 (or +20)),
i(_4-14- 80 (or -10)),
J^( 6+ 0-160 (or -20)),
J^(_4+14_ 80 (or -10)),
T^( 1- 7+160 (or +20)),
/2(0— «)
_( 14_ 7+20)
^(16+28+ 5)
^(36+ 0+20)
_.(16+28+ 5)
_( 1+ 7+20)
For the summary criticism
(O-M)
I ^2(0)
6^ 42 7680 (or 120)
35 + 35 + 35
The hypothesis of the third degree, J^==0, where the values of 70m, and their
differences are:
839 1334 2024 2874 3849
495 690 850 975
195 160 125
-35 -35,
agrees too well with the observations, and must be suspected of being underadjusted, for
the sum of the squares of the summary criticism is only
^, where we might expect l + |/2.
The hypothesis of the second degree, J* = 0, VJ^ ^ 0, gives for 70m, and
differences:
832 1348 2024 2860 3856
516 676 836 996
160 160 160.
The adjustment is here good, the sum of the squares is
II , and we might expect 2+1/4.
The hypothesis of the first degree, A* = 0, VJ'^ .= 0, J*" = 0, gives for the adjusted
values and their differences:
9-6 20-4 31-2 42-0 52-8
10-8 10-8 10-8 10-8.
77
The deviations are evidently too large {o — u is +2'4, — 1'4, —2-2, —1-0, +2-2)
to be due to the use of round numbers; the sum of the squares is also
220-8 instead of 3 ±1/6,
consequently, no doubt, an over-adjustment.
The special adjustment of the second degree, J^=0, T^' = 0, and J* = 2, gives
for M, and its differences:
11-6 19-4 29-2 41-0 54-8
7-8 9-8 11-8 13-8
The deviations o - m = 0-4, -0-4, -0-2, 0-0, +0-2
nowhere reach a, and may consequently be due to the use of round numbers; the sum of
the squares _
4-8 instead of 3 ±1/6
also agrees very well. Indeed, a constant subtraction of 0*04 from m, would lead to
(3-4)2, (4.4)2^ (5.4)2^ (6-4)2, ^nd (7-4)^, from which the example is taken.
Example 3. Between 4 points on a straight line the 6 distances
"si
are measured with equal exactness without bonds. By adjustment we find for instance
we notice that every scale = |. It is recommended actually to work the example by a
millimeter scale, which is displaced after the measurement of each distance in order to
avoid bonds.
XII. ADJUSTMENT BY ELEMENTS.
§ 51. Though every problem in adjustment may be solved in both ways, by
correlates as well as by elements, the difficulty in so doing is often very different. The
most frequent cases, where the number of equations of condition is large, are best suited
for adjustment by elements, and this is therefore employed far oftener than adjustment
by correlates.
The adjustment by elements requires the theory in such a form that each observa-
tion is represented by one equation which expresses the mean value /i , (o) explicitely as
linear functions of unknown values, the ''elements", x, y, . . . z:
78
^liOi) = P,x-\-q^y-\- ... -{-r^z
(85)
where the p, q, ... /• are theoretically given. All observations are supposed to be
unbound.
The problem is then first to determine the adjusted values of these elements
X, tj, ... z, after which each of these equations (85) , which we call '^equations for the
observations", gives the adjusted value u of the observation.
Constantly assuming that ^^(0) is known for each observation, we can from the
system (85) deduce the following normal equations:
pkx{0)
. h{o)
=
PP
U2i0)\
x +
Pi 1
Mo)
y+-
• +
pr
Mo)
z ==
po
[Mo)\
qX^io)
I h(o) \
=
\ IP 1
[LAo)\
x +
q 1 , we make ourselves guilty of an over-
leave out terms for which
adjustment.
Example 1. The five-place logarithms in a table are looked upon as mutually
unbound observations for which the mean error is constantly l/J^ of the fifth decimal
place. The "observations", log 795, log 796, log 797, log 798, log 799, log 800, log 801,
log 802, log 803, log 804, and log 805, are to be adjusted as an integral function of the
second degree
log (800 -^t) = x'-{- y't -\- zt^.
In order to reckon with small integral numbers, we subtract before the adjustment
2-90309 + 0-00054 t, both from the observations and from the formulae. Taking 0-00001 as
our unity, we have then the equations for the observations:
— 2 = x — by + 2bz
— 2 = x — 4y-{-16z
— 1 = x — 3«/4- 9z
— 1 = X — 2y -{- iz
= a; — 1«/+ '^z
= a;
= x-{-\y+ \z
= x-^2y-\- 4z
1 = a; + 3y+ 9z
1 =: X -^ iy -\- I6z
1 = a: + 5y + 252.
From this we get
: 156, and the normal equations :
— 36 = 132a; + 0y+ 1320^
420 = 0a;+1320y+ Oz
— 540= 1320a; + 0«/ + 23496s.
The element y is consequently immediately free of x and z, but the latter must be made
87
free of one another, which is done by multiplying the first equation by 10 and subtracting
it from the third. The transformation into free functions then only requires $ = x -j-lOz
substituted for x, and we have :
- 3G = 132f ,
420 = 1320 y.
— 180 = 10296 ^,
consequently,
f = —0-2727,
XA^) = 1: 132 =
390
51480 = -007576
y = 0-3182 ,
/ij (y) = 1 : 1320 =
39
51480 = -000758
z = _0-0175,
/ij (2) = 1 : 10296 =
5
51480 = -000097
The mean error of y is consequently J- 0-0275, and that of z ^ 0-0099. The element x
is found by x = f — IO2; = —0-0977, to which corresponds X.i(x) = ,^2 (f) + lOO/tj (■2)
= 0-0173 = (0-1315)2. For log 800 we find thus 2-9030890 ± 0-0000013, and the
corresponding difference of the table is 54-318 -J^ 0-028.
For the sum of the squares of the deviations we have, according to (105)— (107),
'{o-u
AAo)
156 — 9-82 — 133-64 — 3-15
9-39,
which shows that the term of the second degree contributes somewhat to the goodness of
the adjustment. This sum of squares ought, according to the number of the observations
and the elements, to be 11 — 3 = 8, with a mean uncertainty of ^4.
The best formula for computing the adjusted values of the several observations
and their mean errors is tu = ^-\-yt-{-z(t^—10), which gives:
u
o—u
(o-uf
kiiu) Scale
log 795
= 2-9003688
+ -12
-0144
390 + 39
25 + 5
225
= 2490
•0484 ^419
log 796
= 2-9009136
--36
-1296
390 + 39
16 + 5
36
= 1194
•0232 -722
log 797
= 2-9014580
+ •20
-0400
390 + 39
9 + 5
1
= 746
•0145 ^826
log 798
= 2-9020019
--19
-0361
390 + 39
4 + 5
36
= 726
•0141 -831
log 799
= 2-9025457
+ -43
•1849
390 + 39
1 + 5
81
= 834
-0162 -806
log 800
= 2-9030890
+ -10
-0100
390 + 39
+ 5
100
= 890
-0173 -792
log 801
= 2-9036321
— -21
•0441
390 + 39
1 + 5
81
= 834
-0162 -806
log 802
= 2-9041747
--47
•2209
390+39
4 + 5
36
= 726
-0141 -831
log 803
= 2-9047170
+ •30
-0900
390 + 39
9 + 5
1
= 746
-0145 -826
log 804
= 2-9052590
+ -10
-0100
390 + 39
16 + 5
36
= 1194
-0232 -722
log 805
= 2-9058006
--06
-0036
390 + 39
25 + 5
225
= 2490
-0484 -419
•7836
12870
8-000
88
Both the checks agree: the sum of squares is 12 x 0-7836 = 9-40, and the sum
of the scales is 11 — 3.
It ought to be noticed that the adjustment gives very accurate results throughout
the greater part of the interval, with the exception of the beginning and the end. The
exactness, however, is not greatest in the middle, but near the 1" and the 3'* quarter.
Example 2. A finite, periodic function of one single essential circumstance, an angle F,
is supposed to be the object of observation. The theory, consequently, has the form :
o„ = Co-|-Ci cos r+s, sin F-j-Cj cos 2F+ s^sin 2F+ . . .
We assume that there are n unbound, equally exact observations for a series of values of F,
whose difference is constant and = -, for instance for F = 0, 60M20M80°, 240°, 300°.
n
Show that the normal equations are here originally free, and that they admit of an exceedingly
simple computation of each isolated term of the periodic series.
Example 3. Determine the abscissae for 4 points on a straight line whose mutual
distances are measured equally exactly, and are unbound. (Cmp. Adjustment by Correlates,
Example 3, and § 60).
Example 4. Three unbound observations must, according to theory, depend on two
elements, so that
0, = jcS Aj(o,) = 1
0, = xy, ,^5,(05,) = -1
The theory, therefore, does not give us equations of the linear form. This may be produced
in several ways, most simply by the common method of presupposing approximate values
of both elements, the known a for x and b for y, and considering the corrections f and; tj
to be the elements of the adjustment. We therefore put a; = a + f < ^"^^ 2/ = ^ + 'J-
Rejecting terms of the 2'"' degree, we get the equations of the observations:
o^—a^ = 2af
O2 — ah = 6f + ^'y
o., — h^ = 2h7j,
where the middle equation has still double weight. The normal equations are:
2a(o, — a2) + 26(02— «6) = (4a'' + 2i«)f+2a6)y
2a(o^ — ab) + 26(03 — 6'^) = 2ah f + (46« + 2a2) jy ;
^ is consequently not free of rj, but we find
, , 62(52o 2a.6os + a2o8) , , , a' + 26^
2«^ = «i + « (Svfpp- ' ''2(^) = 4(a2 + 62)2
, ., a^{b^o,—2abo^^aHs) , ,, 2a« + i«
89
For the adjusted value u^ of the middle observation we have
(a^ + fe^M, = ab^o,+{a'-{-b*)o,+a'^bo,, ^,{u,) = i- ^^J^-
If we had transformed the elements (comp. § 62) by putting
f = aC — ^0
jy ='b^-\-au,
or
y = i(i + + «o,
we should have obtained free normal equations
2(a2oi + 2a60j + 6*03) — 2(a2 + 62)2 _ i{a^J^b^)^i:
2(— a6o, + (a2_62)o2 + aJo3) = 2(a'' + b'')^u.
If we had placed absolute confidence in the adjusting principle of the sum of
squares as a minimum, a solution might have been founded on
(0, - ««)2 + 2 (02 - ab)^ + (03 - *")' == min.
The conditions of minimum are:
1 d min 5,1/ j.\ /. a
1 rfmin 7\ I / i,ow A
^ -^j^ = {o^ — ab)a->r(o.^—b')b = 0.
The solution with respect to a and b is not very difficult. We see for instance
immediately that
(o,-a^)(o,-b^) ^ {0,-abf
or
OjOg — 0^ = b^o^ — 2abo.^-{- a^o.^.
Still better is it to introduce s^ = a^-\-b'^, by which the equations become
{o^ — s^)a-\-o,J) =
o^a-\-(o3—s^)b = 0, .
consequently,
s*-sHo,-^o,)-\-o,o,-ol =
(»'-H^) =(°'-i^)+";-
If the errors in 0,, 0^, and O3 are not large, o^o.^ — 0-^ must be small; one of the
two values of s^ must then be small, the other nearly equal to Oi-^-o.^; only the latter can
be used.
12
90
Further,
we
get;
a
=
02
s«
03-
0, — ;
(
[b
r=
Os-
-s«
Ol-
-s"
a^
(03-
0,+C
'3-
-28"'
i« =
(0,-
, — 28"
In this way we avoid guessing at approximate values (for which otherwise we
should perhaps have taken a'^^o^ and 1^ = 0^). The values which we have here found
for a^ and b^, and to which may be added
— ab
are really exact; and if we substitute them in the above normal equations, we get f =
and 7j = 0.
Even when, as in this case, the theory is not linear, it is not unusual for the
sum of the squares to be a minimum. Caution, however, is necessary; particularly, it
may happen that the sum of the squares becomes a maximum for the found elements, or
for some of them.
We may also in another way make the equations of this example linear, namely,
by considering the logarithms of o^, 0^, Og as the observed quantities, and finding the
logarithms of the elements from the equations which will then be linear.
logo, = 2 log a;
logOg = log a: + logy
logog = 2 logy.
In this way we throw the difficulty over upon the squares of the mean errors. As
dz
log (z-\-dz) = log« +
z
we may approximately take
^2 (log ^) = ^ -^z i^)-
If a and b also here indicate approximate values of x and y, the weights of the
3 equations, respectively, become proportional to a*, 2a^b^, and b*. Thus we find the
normal equations
2aMogo, + 2a2ftMogo., = {4a* + 2a%^) log x ^ 2a^b^og y
2a^¥ log 02 + 2b* log 03 = 2a^b^ log x + {4b*+2a-'b^) log y ,
91
which give the simple results
21ogx = logo. - (^,3^j log-^ , ^,(logx) = ^^.(J^^.).
This solution agrees only approximately with the preceding one. It might seem
for a moment that, in this way, we might do without the supposition of approximate values
for the elements, but this is far from being the case. For the sake of the weights we
must, with the same care, demand that a and x, as also b and «/, agree, and we must
repeat the adjustment till the squares of the mean errors get the theoretically correct
values. And then it is only a necessary, but not a sufficient condition, that x — a and
y ^ b are small. Unless the exactness of the observations is also so great that the mean
errors of o,- are small in proportion to o, itself, the laws of errors of the logarithms cannot
be considered typical at the same time as those of the observations themselves.
Example 5. The co-ordinates of four points in a circle are observed with equal
mean errors and without bonds: Xj ^ 20, y. = 10; x^^ 16, y^ = 18; x^ ^ 3, ^a = 1'^!
and x^ =2, y^ = 4. In the adjustment for the co-ordinates a and b of the centre and
the radius /•, we cannot use the common form of the equations
(x-a)^ + {y-b)' = r\
because it embraces more than one observed quantity besides the elements. In order to
obtain the separation of the observations necessary for adjustment by elements, we must
add a supplementary element, or parameter, F, for each point, writing for instance
Xi = a -\- r cos F, , yt = b -\- r sin F,- .
As the equations are not linear we must work by successive corrections A a, A 6,
Ar, AF, of the elements, of which the first approximate system can be obtained by
ordinary computation from 3 points. For the theoretical corrections Axt and Ayt of the
co-ordinates we get by differentiation of the above equations
Axi = A a -{- Ar cos Vi — A Vi • r sin F,-
Ayi = A 6 + A >• • sin F,- + A F.- • r cos F,-.
These equations for the observations lead us to a system of seven normal equations.
By the "method of partial elimination" (§ 61) these are not difficult to solve, but here the
simplicity of the problem makes it possible for us immediately to discover the artifice.
We know that every transformation of equally well observed rectangular co-ordinates results
in free functions. The radial and the tangential corrections
A Xi cos Vi -\- A yt sin F,- = A w,-
and
A Xi sin Vi — A y,- cos F,- = A <<■
12*
92
can, consequently, here be taken directly . for the mean values of corrections of observed
quantities, and as only the four equations
A <,■ = A a sin F,- — A 6 cos F, — /• A Fj
contain the four corrections A F, of the parameters, they can be legitimately reserved for
the successive corrections of the elements. In this way
A «i = A a cos F, ■+ A i sin F^ + A r
with equal mean errors, ^.^in) = k^i^) = k^iiy), are the "equations for the observations"
of this adjustment, and give the three normal equations:
[A n cos F] = A a [cos^ F] + Ab [cos F sin F] + A r [cos F]
[A n sin F] = A a [cos F sin F] + A 6 [sin^ F] + A r [sin F]
[Aw] = A a [cos F] -\- Ab [sin F] + A r • 4.
In the special case under consideration, we easily see that the first, second, and
fourth point lie on the circle with r = 10, whose centre has the co-ordinates a = 10 and
6 = 10; the parameters are consequently:
Fi = 0°0'0, F^ = 53°7'8, Fg = 135°0'0, and F,= 216°52'2.
For the third point the computed co-ordinates are: X3 = 2'9290 and y3 = 17-0710,
consequently, A a;3 = +0-0710 and A^g = —0-0710, A <3 = 0, and A «3 = —0-1005;
all other differences A a;,- = and A yt = 0. The "equations for the observations" are :
1-0000 A a + 0-0000 A 6 + 1-0000 A /• = 0-0000
0-6000 A a + 0-8000 A 6 + 1-0000 A r = 0-0000
— 0-7071 A a + 0-7071 A 6 + 1-0000 A r = — 0-1005
— 0-8000 A a — 0-6000 A & + 1 -0000 A r = 0-0000.
The normal equations are:
2-5000 A o + 0-4600 A 6 + 0-0929 A r = + 0-0710
0-4600 A a + 1-5000 A 6 + 0-9071 A r = — 0-0710
E = 0-0929 A a + 0-9071 A 6 + 4-0000 A r = — 0-1005.
By elimination of A r we get
2-4978 A a + 0-4390 A 6 == + 0-0733
B = 0-4390 A a + 1-2943 A 6 = — 0-0482 ;
and by eliminating A b
A = +2-3490 A a = +0-0896.
From jK, B, and A we compute
Aa = +0-0381 , A6 = —0-0501, and Ar = —0-01465.
The checks are found by substitution of these in the several equations. The 4 equations
For cliecking
+ 0-0002
0-0000
0-0000
+ 0-0001
— 0-0001
0-0000
the sum of squares
(here = (8 — 7)/} 2) is consequently
93
for the observations give the following adjusted values of Aw,:
Ami = +0-0234, A M2 = — 0-0319, A ^3=— 0-0770, and Aw, = —00151;
(o — u)
= (0-0234)2 + (0-0319)2 + (0-0235)^ + (O-OISI)^ = 0-00235.
For this, by the equation (108), we get
0-01010 — 0-00271 — 0-00356 — 0-00147 = 0-00236
as the final check of the adjustment.
The 4 equations for A tt give us
AF, = -4-17'2, AF5, = +20'8, AF3 = -2'9, and AF, = — 21'6.
Thus, by addition of the found corrections to the approximate values,
r = 9-98535, a = 10-0381, 6 = 9-9499,
F, = 0°17'2, F, = 53°28'6, F3 = 134°57'1, and V, = 216°30'6,
we have the whole system of elements for the next approximation, if they are not the
definitive values. In both cases we must compute by them the adjusted values of the co-
ordinates, according to the exact formidw, the resulting differences, obs. — comp., are:
Point Ax Ay An At
1 —0-0232 +0-0002 —0-0232 +0-0002
2 +0-0191 +0-0257 +0-0320 0-0000
3 +0-0166 —0-0166 —0-0234 —0-0001
4 —0-0123 —0-0090 +0-0152 0-0000.
The sum of the squares, [(Aa;)2 + (A y)^] = 0-00236, agrees with the above value,
which indicates that the approximation of this first hypothesis may have been sufficient.
Indeed, the students who will try the next approximation by means of our final differences,
will, in this case, find only small corrections.
From the equations A, B, and R, which express the free elements by the original
bound elements, A a, Ab, Ar, we easily compute the equations for the inverse trans-
formation:
A a = 0-4257 • A
A6 = — 0-1444 -^ + 0-7726- B
A r = 0-0228 • A — 0-1752 • B + 0-25 • E.
By these, any function of the elements for a given parameter can be expressed as a linear
function of the free functions A, B, and R; and by yi^C^) = 2-3490 >i 5, X^(B) = l-2di?, X^,
94
and ^2 (J?) = 4^2, the mean error is easily found. Thus the squares of the mean errors of
the co-ordinates x and y are
X^(x) = {2-3490 ( 0-4257 + 0-0228 cos F)" + 1-2943 (-0-1752 cos F)« +4(0-25 cos Vf}l^
'i2(y) = {2-3490(— 0-1444 + 0-0228sinF)'' + l-2943( 0-7726 — 0-1752 sin F)" +4(0-25 sin F)''}>i2
Only the value k^ = 0-00236, found by the summary criticism, is here very
uncertain.
XIIL SPECIAL AUXILIARY METHODS.
§ 57. We have often occasion to use the method of least squares, particularly
adjustment by elements; and this sometimes requires so much work that we must try to
shorten it as much as possible, even by means which are not quite lawful. Several temp-
tations lie near enough to tempt the many who are soon tired by a somewhat lengthened
computation, but not so much by looking for subtleties and short cuts. And as, moreover,
the method was formerly considered the best solution — among other more or less good —
not the only one that was justified under the given supposition, it is no wonder that it
has come to be used in many modifications which must be regarded as unsafe or wrong.
After what we have seen of the difference between free and bound functions, it will be
understood that the consequences of transgressions against the method of least squares
stand out much more clearly in the mean errors of the results than in their adjusted
values. And as — to some extent justly — more importance is attached to getting tolerably
correct values computed for the elements, than to getting a correct idea of the uncertainty,
the lax morals with respect to adjustments have taken the form of an assertion to the
effect that we can, within this domain, do almost as we like, without any great harm,
especially if we take care that a sum of squares, either the correct one or another, becomes a
minimum. This, of course, is wrong. In a text-book we should do more harm than good
by stating all the artifices which even experienced computers have allowed themselves to
employ, under special circumstances and in face of particularly great difficulties. Only
a few auxiliary methods will be mentioned here, which are either quite correct or nearly
so, when simple caution is observed.
§ 58. When methodic adjustment was first employed, large numbers of figures
were used in the computations (logarithms with 7 decimal places), and people often com-
plained of the great labour this caused; but it was regarded as an unavoidable evil, when
the elements were to be determined with tolerable exactness. We can very often manage,
however, to get on by means of a much simpler apparatus, if we do not seek something
95
which cannot be determined. During the adjustment properly so called, we ought to be
able to work with three figures. But this ideal presupposes that two conditions are satis-
fied: the elements we seek must be small and free of one another, or nearly so; and in
both respects it can be difficult enough to protect oneself in time by appropriate trans-
formation. Often it is only through the adjustment itself that we learn to know the
artifices which would have made the work easy. This applies particularly to the mutual
freedom of the elements. The condition of their smallness is satisfied, if we everywhere use
the same preparatory computation as is necessary when the theory is not of linear form.
By such means as are used in the exact mathematics, or by a provisional, more
or less allowable adjustment, we get, corresponding to the several observations o, . . . o„,
a set of values v^ . . .Vn which are computed by means of the values x^^ . . . z^ of the
several elements x . . .z, and which, while they satisfy all the conditions of the theory with
perfect or at any rate considerable exactness, nowhere show any great deviation from the
corresponding observed value. It is then these deviations o, — «?, and x — x^... which are
made the object of the adjustment, instead of the observations and elements themselves
with which, we know, they have mean error in common. When in a non-linear theory
the equations between the adjusted observation and the elements are of the general form
Ui = F{x, . . .z),
they are changed into
«'-^'=(f)„(^-^»)+--- + (f)„(^-^o) (109)
by means of the terms of the first degree in Taylor's series, or by some other method of
approximation. If the equations are linear
Ui = PiX + . . . + i-iZ ,
we have, without any change, for the deviations:
Ui — Vi = 'Piix — Xa)^ ... + ''•• (^ — ^o)- (110)
No special luck is necessary to find sets of values, ?;,,... a;^, ... ^n, whose devia-
tions 0, — Vi show only two significant figures ; and then computation by 3 figures is, as
far as that goes, sufficient for the needs of the adjustment.
The method certainly requires a considerable extra-work in the preparatory com-
putation, and it must not be overlooked that computations with an exactness of many
decimal places will often be necessary in this part ; especially Vi ought to be computed with
the utmost care as a function of a;„ . . . z,, , lest any uncertainty in this computation should
increase the mean errors, so that we dare not put X^ {o—v) = k^ (o).
This' additional work, however, is not quite wasted, even when the theory is linear.
The list of the deviations o,- — »,• will, by easy estimates, graphic construction, or directly
96
by the eye, with tolerable certainty lead to the discovery of gross errors in the series of
observations, slips of the pen, etc., which must not be allowed to get into the adjust-
ment. The preliminary rejection of such observations may save a whole adjustment; the
ultimate rejection, however, falls under the criticism after the adjustment.
In computing the adjusted values, particularly w,-, after the solution of the normal
equations, we ought not to rely too confidently on the transformation of the equations into
linear form or into equations of deviations for o,- — vt . Where it is possible, the actual
equations «,■ = F(x, . . . z) ought to be employed, and with the same degree of accuracy
as in the computation of »,. In this way only can we see whether the approximate system
of elements and values has been so near to the final result as to justify the rejection of
the higher terms in Taylor's series. If not, the adjustment may only be regarded as
provisional, and must be repeated until the values of «,-, got by direct computation,
agree with the values through Ui — Vi in the linear equations of adjustment.
On the whole the adjustment ought to be repeated frequently till we get a sufficient
approximation. This, for instance, is the rule where the observations represent probabilities,
for which ^^ (o,) is generally known only as functions of the unknown quantities which
the adjustment itself is to give us.
§ 59. The form of the theory, and in particular the selection of its system of
elements, is as a rule determined by purely mathematical considerations as to the
elegance of the formulae, and only exceptionally by that freedom between the elements
which is wanted for the adjustment. On the other hand it will generally be impossible
to arrange the adjustment in such a way that the free elements with which it ends, can
all be of direct, theoretical interest. A middle course, however, is always desirable, for the
reasons mentioned in the foregoing paragraph, and very frequently it is also possible, if
only the theory pays so much respect to the adjustments that it avoids setting up, in the
same system, elements between which we may expect beforehand that strong bonds will
exist. Thus, in systems of elements of the orbits of planets, the length of the nodes and
the distance of the perihelion from the node ought not both to be introduced as elements;
for a positive change in the former will, in consequence of the frequent, small angles of
inclination, nearly always entail an almost equally large negative change in the latter. If
a theory says that the observation is a linear function of a single parameter, t, the formula
ought not to be written u=p-\-qt, unless all the fs are small, some positive, and others
negative, but u = r'\-q{t — tQ), where t^ is an average of the parameters corresponding to
the observations. If we succeed, in this way, in avoiding all strongly operating bonds,
and this can be known by the coefficients of all the normal equations outside the diagonal
line becoming numerically small in comparison with the mean proportional between the
two corresponding coefficients in the diagonal line, then we have at any rate attained so
97 •
much that we need not use in the calculations for the adjustment many more decimal
places than about the 3, which will always be sufficient when the elements are originally
mutually free, and not during the adjustment are first to be transformed into freedom
with painful accuracy in the transformation operations.
If, by careful selection of the elements, we even get so far that no sum of the
products [pq]^) in numerical value exceeds about ^ of the mean proportional between the
corresponding sums of squares V[pp] [qq] , or in many cases only ^ of these amounts,
then we may consider the bonds between the elements insignificant. The normal equations
themselves may then be used to determine the law of error for the elements; we compute
provisionally a first approximation by putting all the small sums of products = 0, and in
the second approximation we correct the [j9o]'s by substituting the sums of the products
and the values of the elements as found in the first approximation. For instance:
[po] — [pq\ y^ — .... - [pr\ z, = [pp\ x^ (111)
while
= t:{M-M-...-M}. ,„s,
As the errors in these determinations are of the second order, it will not, if the o's
themselves are small deviations from a provisional computation, be necessary to make any
further approximations.
Even if the bonds between the elements, which are stated in terms of the sums
of the products, are stronger, we can sometimes get them untied without any transforma-
tion. If we can get new observations, which are just such functions of the elements that
the sums of the products will vanish if they are also taken into consideration, we will of
course put off the adjustment until, by introducing them into it, we cannot only facilitate
the computation, but also increase the theoretical value and clearness of the result. And
if we can attain freedom of the elements by rejecting from a long series of observations
some single ones, we do not hesitate to use this means; especially as such unused observa-
tions may very well be employed in the criticism. If, for instance, an arctic expedition
has made meteorological observations at some fixed station for a little more than a com-
plete year, we shall not hesitate in the adjustment, by means of periodical functions, to
leave out the overlapping observations, or to make use of the means of the double values,
giving them the weight of single observations.
') In what follows we write, for the sake of hrevity, [pq\ for [^J.
13
§ 60. Though of course the fabrication of observations is, in general, the greatest
sin which an applied science can commit, there exists, nevertheless, a rather numerous and
important class of cases, in which we both can and ought to use a method which just
depends on the fabrication of such observations as might bring about the freedom of the
theoretical elements. As a warning, however, against misuse I give it a harsh name: the
method of fabricated observations.
If, for instance, we consider the problem which has served us as an example in the
adjustment, both by correlates and by elements, viz. the determination of the abscissae for
4 points whose 6 mutual distances have been measured by equally good, bondfree observa-
tions, we can scarcely after the now given indications look at the normal equations,
"12 + 0,3 -|- Oj4 = Sxj — 1x2 — la^a — 1»4
— "12 + ''23 + "24 = —lXi-\-Sx^—lx.^—lx^
— 0,3 — O23 + O34 = — la;,— laj2+3a-3 — la?,
Oj, Oj, O3, = la;, lajj i^x■^-\-6x^,
without immediately feeling the want of a further observation :
= la;, 4- Iscg 4" 1^3 + 1^4 1
which, if we imagine it to have the same weight = 1 as each of the measurements of
distance ^,(o„) ^ Xr — a;,, will give by addition to the others, but without specifying the
value of 0,
+ 0,2 + 0,3 + 0,, = 4x,
— 0,2 + 028 +»« 4 = 4a;2
— 0,3 — 023+03, == 4a;8
— 0, , — 02, — O3, = 4a;,,
and consequently determine all 4 abscissae as mutually free and with fourfold weight.
What in this and other cases entitles us to fabricate observations is indeter-
minateness in the original problem of adjustment — here, the impossibility of determining
any of the abscissae by means of the distances between the points. When we treat
such problems in exact mathematics we get simpler, more symmetrical, and easier solu-
tions by introducing values which can only be determined arbitrarily; and so it is also in
the theory of observation. But the arbitrariness gets here a greater extent, because not
only mean values, but also mean errors must be introduced for greater convenience. And
while we can always make use of a fabricated observation in indeterminate problems for
the complete or partial liberation of the elements, we must here carefully demonstrate,
by criticism in each case, that the fabrication we have used has not changed anything
which was really determined without it.
99
In the above example, this is seen in the first place by O disappearing from all
the adjusted values for the distances Xr — Xs, and then by O's own adjusted value,
determined as the sum x^-\-X2^x^'\-x^, and leading only to the identity = 0. The
adjustment will consequently neither determine nor let it get any influence on the
other determinations. The mean errors show the same and, moreover, in such a way that
the criterion becomes independent of whether has been brought into the computation
as an indeterminate number or with an arbitrary value, for, after the adjustment as well
as before, we have for Oj X^iO) == 1. The scale for is consequently = 0, and this is
also generally a sufficient proof of our right to use the method of fabricated observations.
§ 61. The method of partial eliminations. When the number of elements is
large, it becomes a very considerable task to transform the normal equations and eliminate
the elements. The difficulty is nearly proportional to the square of that number. Long
before the elements would become so numerous that adjustment by correlates could be
indicated, a correct adjustment by elements can become practically impossible. The special
criticism is quite out of the question, the summary criticism can scarcely be suggested, and
the very elimination must be made easier at any price. If it then happens that some of
the elements enter into the expressions for some of the observations only, and not at all in
the others, then there can be no doubt that the expedient which ought first to be employed
is the partial elimination (before we form the normal equations) of such elements from the
observations concerning them. These observations will by this means be replaced by certain
functions of two observations or more, which will generally be bound; and they will be
so in a higher and more dangerous degree the fewer elements we have eliminated. By
this proceeding we may, consequently, imperil the whole ensuing adjustment, the foundation
of which, we know, is unbound or free observations as functions of its elements.
If now it must be granted that the difficulties can become so great that we cannot
insist on an absolute prohibition against illegitimate elimination, we must on the other
hand emphatically warn against every elimination which is not performed through free
functions, and much the more so, as it is quite possible, in a great many cases in which
abuses have taken place, to remain within the strictly legitimate limits of the free functions,
by the use of "the method of partial eliminations".
This is connected with the cases, in which some of the observations, for instance
Oy . . . Om, according to the theory, depend on certain elements, for instance x, . . .y, which
do not occur in the theoretical expression for any other of the observations. Our object is
then, by the formation of the normal equations to separate o^ . . . o^ as a special series of
observations. We begin by forming the partial normal equations for this, and then imme-
diately perform the elimination of x, . . . y from them, without taking into consideration
whether these equations alone would be sufficient for a determination of the other elements.
13*
100
As soon as x . . . y are eliminated, the process of elimination is suspended. The trans-
formed equations containing these elements (which now represent functions that are free of
all observations, and functions which depend only on the remaining elements z, . . . m), are
put aside till we come back to the determination ot x . . . y. The other partially transformed
normal equations, originating in the group o^ . . . Om, are on the other hand to be added,
term by term, to the normal equations for the elements z,-... u, formed out of the remain-
ing observations, before the process of elimination is continued for these elements.
That this proceeding is quite legitimate becomes evident if we imagine the
elements x . . . y transformed into the elements jc'. . . «/', which are free oi z . . . u, and then
imagine x' . . . y' inserted instead ot x . . . y in the original equations for the observations.
P^r then all the sums of products with the coefficients ot a^ . . . y' will identically become
== 0, and the sums of squares and sums of products for the separated part of the observa-
tions will, as addenda in the coefficients of the normal equations (compare (57)), come outi
immediately, with the same values as now the transformed normal equations.
As an example we may treat the following series of measurements of the position
of 3 points on a straight line. The mode of observation is as follows. We apply a millimeter
scale several times along the straight line, and then each time read oif by inspection with
the unaided eye either the places of all the points against the scale or the places of two
of them. The readings for each point are found in its separate column, and those on the
same row belong to the same position of the scale. (Considered as absolute abscissa-
observations such observations are bound by the position of the zero by every laying
down of the scale; but these bonds are evidently loosened by our taking up the position
against the scale of an arbitrarily selected fixed origin yr as an element beside the abscissae
1, x^, Xg of the three
points).
All mean
errors are supposed to be equal.
Position
Point
Eliminated free Elements
mm.
17-22 = y, +h{^i + ^2) .
of
the Scale
1
I
mm.
6-9
II
mm.
27-54
III
mm.
2
8-35
54-95
31-65
= 1/2 + h (^1 4- ^a)
^
3
7-9
54-5
31-20
= 1/3 +5(^1+^3)
4
21-16
47-2
34-18
== y4 + i («2 + ^3)
II
5
10-74
36-7
23-72
— ys +1(^2 + ^3)
II
6
4-06
30-1
17-08
= Ve +-H^2 + ^3)
t.sl'
7
31-45
51-98
78-06
53-83
= 2/7 +U^i + ^i + ^-i)
8
32-9
53-5
79-5
55-30
= ys -[-^(^1 + ^2 + ^3)
9
9-6
30-3
56-22
32-04
= 2/9 +i(a;i + a;2 + a;3)
(JQ
ri-
10
20-16
40-78
66-8
42-58
= Vio+H^i + '^i + ^a)
II
11
18-9
39-5
65-56
41.32
= yii+U^i+^i + ^a)
w
101
As the theoretical equation for the i"" observation in the s"" column has the form
and every observation, therefore, is a function of only two elements, there is every reason
to use the method of partial elimination. If we choose first to eliminate the ys, we have
consequently to form normal equations for each of the 11 rows. Where only two points
are observed these normal equations get the form
Or + O, = 2yi -\- Xr -\- Xs
Or = «/i + Xr
for three points the form of the normal equations is
Oi + Ojj + O3 == 3«/i + ^i + ^•^ + ^3
Ol = y-' + ^i
03 = !/i -\-Xs.
Of these equations those referring to the ij, have given the eliminated free elements
stated above to the right of the observations after the perpendicular.
By subtracting these equations from the corresponding other equations we get,
in the cases where there are 2 points:
Or — i{0r + 0s)= \Xr — ^X,
0, — \{0r-\-0,) = —\Xr + \Xs ,
and in cases where there are 3 points :
Oi — i(Oi + o., + 03) = |aJi — ia;.,— |a;3
02— 3(01 + 02 + ^3) = —1^1 + 1^2 — \^i
03 — M01 + O2 + O3) = —h^i — \^2 + i^-6-
By forming the sum of these differences for each column, and counting, on the
right side of the equations, ' how often each element occurs with one other or with two
others, we consequently get the ultimate normal equations:
-168'^8= 'Ix.-'ix^-lx^
— 37"71 = — -^x^^-^x^ — -g-a;^
+ 206-69 = -'{x.-'lx^ + '^x,.
The case is here simple enough to be solved by a fabricated observation. How is
its most advantageous form found, when its existence is given?
Answer: g^ = ^-f-|l + |l, weight = 23712 ,
102
after which we get the normal equations:
114
U9
— 168-98
114 *i
37-71 = ^x.
^ + 206-69 =
consequently,
mm.
Xi = — 25-38 , x^ =
From these we now compute the ys:
«/i = 32-295 — 0,
4-77 , and ^3 = + 21-24.
2/2
2/6
= 33-72 — ,
= 33-27 — .
= 25-945 — ,
= 15-485 — 0.
== 8-845 — ,
= 56-80 -0,
= 58-27 — 0,
= 35-01 - ,
y,o= 45-55 — 0,
2/1,= 44-29-0.
2/7
We need not here state the adjusted values for the several observations, nor their
differences, of which it is enough to say that their sum vanishes both for each row and
for each column; their squares, on the other hand, will be found to be:
I
II
III
Total:
-0002
-0002
•0004
1
•0001
2
1
1
2
2
2
4
6
6
12
2
2
4
9
25
4
38
1
1
2
9
36
9
54
1
1
2
1
4
9
14
Total: -0025
-0077
•0036
•0138
2' = ^0028
2' = •Olio
For the summary criticism we notice that the number of observations is 27, the
number of the elements is 3 + 11 — 1 = 13, divisor consequently = 14 (one element being
wholly engaged by the fabricated observation 0). The unit of the mean error is therefore
determined by £^ = 0-0010, and the mean error on single reading j^ 0-032, which agrees
well with what we may expect to attain by practice in estimates of tenth parts.
103
As to special criticism it is liere, wliere the weights of the eliminated free
functions are respectively 2 and 3 times the weight of the single observation, while the
weights of a;, , x^, and x^ after the adjustment become respectively
easy to compute the scales
759
H4'
-, and ^, very
1 —
X^(u)
1
1
Xi(o) ' Weight after the adjustment "
With 759 as common denominator we find for the several scales and the sums of their
most natural groups:
1
2
3
4
5
6
7
8
9
10
11
I
II
III
327
327
654
331-5
331-5
663
331-5
331-5
663
336
336
672
336
336
672
336
336
672
436
442
448
1326
436
442
448
1326
436
442
448
1326
436
442
448
1326
436
442
448
1326
3170
3545
3911
10626
3996
2' = 6630
The comparison with the sums of squares in the groups, divided by £'", shows then for
point r 2-5 instead of ^ = 4-2 ±1/8^, for point 11 7-7 instead of 4-7±V/9¥, for
point III 3-6 instead of 5-1^1/10-2, for all positions of the scale with two readings
2-8 instead of 5-'d:i::VlO-6 , and for positions with 3 readings 11-0 instead of 8-7 ^ 1/17-4.
The limit of the mean error is consequently reached only in the group of point II, where
(7-7 — 4-7)* = 9-0 < 9-4, and it is nowhere exceeded. We have a checii by summing
the scales:
i2«^= 14 = 27-11-3 + 1.
§ 62. In such cases in which the circumstances and weights of the observations
are distributed in some regular way, this will often facilitate the treatment of the normal
equations. The elimination of the elements and the transformation of the normal equations
into such whose left hand sides can be regarded as unbound observations, as they are free
104
functions of the original observations, need not always be so iirmly connected with one another
as in the ordinary method. If we, in a suitable way, take advantage of regularity in the obser-
vations, and thereby are able, to find a transformation which sets the normal equations free,
then the determination of the several elements will scarcely throw any material obstacles
in our way. But in order to find out any special transformations, we must know the
general form of the changes of the normal equations resulting from transformation of the
original elements into such as are any homogeneous linear funtions of them whatever.
If the equations for the unbound observations in terms of the original elements
have been
Oi-^ Vi^ + iiV + r^z,
the normal equations will be:
[qo] = {qp\x+{qq^y + \qr]z
\ro\ = [rp]x-j- [rg'Jy -|- [r/]0.
And if we wish to substitute new elements, f, ;j, and (^, for the old ones, we make use of
substitutions in which the original elements are represented as functions of the new ones,
therefore
2/ == /.,f + A:,, + Z,C [ (114)
The equations for the observations then have the form
Oi = (jo./ii+?<^2+»'<-^3)f+{j'.-^i+?.-^2+»'.^3)'? + (i't^+9'J2 + n^3)C- (115)
The new normal equations may be formed from these, but the form becomes very cumbrous,
the equation which specially refers to $ being
[{phi + qh^-Jrrh;) o] = [(ph^-i-qh^ +rh^)^ f + [{ph^ + qh^^^rhs) (Mi + ^a + '^a)] V +
+ [iphi + qhi + rh^) {pi, + ql., + rl^)\ C-
The computation ought not to be performed according to the expressions for the coefficients
which come out when we get rid of the round brackets under the signs of summation [ ].
But it is easy to give the rule of the computation with full clearness. The old normal
equations are first treated exactly as if they were equations for unbound observations, for
X, y, and z, respectively; expressed by the new elements, consequently by multiplication,
by columns, by A,, h^, and h^ and addition; by multiplication by k,, k^, and kg and
addition; and by multiplication by ^i, l^, and l^ and succeeding addition. Thereby, certainly,
we get the new normal equations, but still with preservation of the old elements:
105
Upki-^qk^-\-rks)o] = [{pki-{-qk^-{-rks)p]x + [{pk,+qk^^rk^)q]y + [{pk,-^qk^-\-rk^)r]3 \{\16}
[{ph + ql2-i-rl.^)o] = [{ph+ ql2-i-rls)p]x-^[{pl,-^ ql^+rl.,)q]y^[{pli-i-qL,-\-rl.,)r]z
The second part of the operation must therefore consist in the substitution of the
new elements for the original ones in the rigiit hand sides of these equations. In order
to find the coefficients of f , tj, and C' we must therefore here again multiply the sums of
the products, now by rows, by
'11 '2* '3
and add them up.
Example. It happens pretty often, for instance in investigations of scales for
linear measures, that there is symmetry between the elements, two and two, Xr and Xm^r,
so that for instance the normal equation which specially refers to -Xr, has the same coeffi-
cients, only in inverted order, as the normal equation corresponding to x^-r', of course,
irrespective of the two observed terms [po] on the left hand sides of the equations.
Already P. A. Hansen pointed out that this indicates a transformation of the elements
into the mean values Sr = k{xr-{-Xm-r) and their half diiferences dr = }^(xr — a;„,_,.). In
this case therefore the equations for the old elements by the new ones have the form
Xf ^^^ Of. (*j.
Xm—r =^ Sr Ctr ,
and the transformation of the normal equations is, consequently, performed just by forming
sums and differences of the original coefficients. If the normal equations are
[ao] = 4a; + 3?/ + 22: + 1m
[60] = 3a; + 6«/ + 40 + 2M
{co\ = 2a; + 42/ + 62 + 3m
[rfo] = la; + 2«/ + 30 + 4m ,
the procedure is as follows:
[ao\ + {do\ --= 5a; + Sy + bz + bu = lO^i^ + lO^i^
[bo\ -^{co-\ = bx + \0y + 10^ + bu = 10 ^^±^ + 20 ^^
[«o] - [f/o] = .3a; + I2/ - l2 - 3m = 6^^+2^^
{bo\ — [co] = la;+ 2y— 2z - \u = 2 "^^^+4
2
14
106
As in this example, we always succeed in separating the mean values from the halt
differences, as two mutually free systems of functions of the observations.
§ 63. The great simplification that results when the observations are mere repe-
titions, in contradistinction to the general case when there are varying circumstances in
the observations, is owing to the fact that the whole adjustment is then reduced to the
determination of the mean values and the mean errors of the observations. Before an adjust-
ment, therefore, we not only take the means of any observations, which are strictly speaking
repetitions, but we also save a good deal of work in the cases which only approximate to
repetitions, viz. those where the variations of circumstances have been small enough to allow
us to neglect their products and squares. It has not been necessary to await the systematic
development of the theory of observations to know how to act in such cases.
When astronomers have observed the place of a planet or a comet several times
in the same night, they form a mean time of observation t, a mean right ascension a,
and a mean declination , and consider a and 8 the spherical co-ordinates of the star at
the time t.
With the obvious extensions this is what is called the normal place method, the
most important device in practical adjustment. Such observations whose essential circum-
stances have "small" variations, are, before the adjustment, brought into a normal place, by
forming mean values both for the observed values themselves and for each of their essential
circumstances, and on the supposition that the law which connects the observations and
circumstances, holds good also, without any change, with respect to their mean values.
Much trouble may be spared by employing the normal place method. The question
is, whether we lose thereby in exactness, and then how much.
We shall first consider the case where the unbound observations o are linear
functions of the varying essential circumstances x, . . . . z, the equation for the observa-
tions being:
?.^{o) = a-{-bx~\- -\- de.
With the weights v we form the normal equations:
[vo] = a[v]-\-b[vx]+ ... -ird[vz] (117)
[vxo] = a [v.t] -\- b [vx^] -|- + '^ [vxz] \
(118)
[vzo] = a\yz^ + ^[^'■Z'^] + • + (^[f^*]. J
If the whole series of observations is gathered into a single normal place, 0,
corresponding to the circumstances X . . . Z, and with the weight F, we shall have:
107
V = [V]
VO = [vo]
VX = [vx\
VZ = [vz\ ,
and as
= a + bX-^....-J^dZ, (117a)
this normal place will exhaust the normal equation (117) corresponding to the constant
term, both with respect to mean value and mean error. But if we make the other normal
equations free of (117), we get, by the correct method of least squares:
lv(o — 0)(x—X)] = blvix—Xy]-^ . . . -{-d[v{x—X){z—Z)]
[v{o — 0)(z — Z)] = b[v(x—X)(z — Z)]-{- . . . +d[v(2 — Zf]
(118 a)
for the determination of the elements b . . . d, and these determinations are lost completely
if the whole series is gathered into a single normal place. Certainly, the coefficients of these
equations (118a) are small quantities of the second order, if the x—X and z — Z are
small of the first order.
If, on the other hand, we split up the series, forming for each part a normal
place, and adjusting these normal places instead of the observations according to the
method of the least squares, then the normal equation corresponding to the constant
term is still exhausted by the normal place method; and besides this determination of
a-\-bX-\- -^dZ the normal place method now also aflords a determination of the other
elements b . . . d, in such a way, however, that we suffer a loss of the weights for their
determination. This loss can become great, nay total, if the normal places are selected in
a way that does not suit the purpose; but it can be made rather insignificant by a
suitable selection of normal places in not too small a number.
Let us suppose, in order to simplify matters, that the observations have only one
variable essential circumstance x, of which their mean values are linear functions, con-
sequently
>li(o) = a-\-bx,
and that the x's are uniformly distributed within the utmost limits, x^ and a;, ; we then let each
normal place encompass an equally large part of this interval, and we shall find then, this
being the most favourable case, with n normal places, that the weight on the adjusted value of
the element b becomes 1 — 1—1 , if by a correct adjustment by elements the corresponding
weight is taken as unity. The loss is thus, at any rate, not very great. And it can be
made still smaller, if the distribution of the essential circumstance of the observations is
108
uneven, and if we can get a normal place everywhere where the observations become
particularly frequent, while empty spaces separate the normal places from each other.
The case is analogous also when the observations are still functions of a single
or a few essential circumstances, but the function is of a higher degree, or transcendental.
For it is possible also to form normal places in these cases; and we can do so not only
when the variations of the circumstances can be directly treated as infinitely small within
each normal place, which case by Taylor's theorem falls within the given rule. For if we
have at our disposal a provisional approximate formula, y = f{po), and have calculated the
deviation from this, o — y, of every observation (considering the deviations as observations
with the essential circumstances and mean errors of the original observations), then we
can use mean numbers of deviations for reciprocally adjacent circumstances as corrections
which, added to the corresponding values from the approximate formula, give the normal
values. Further, it is required here only that no normal place is made so comprehensive
that the deviations within its limits do not remain linear functions of the essential
circumstances.
Also here part of the correctness is lost, and it is difficult to say how much. The
loss is, under equal circumstances, smaller, the more normal places we form. With twice
(or three times) as many normal places as the number of the unknown elements of the
problem, it will rarely become perceptible. With due regard to the essential circumstances
and the distribution of the weights we can reduce it, using empty spaces as boundaries
between the normal places.
A suitable distribution of the normal places also depends on what function the
observations are of their essential circumstances. As to this, however, it is, as s^ rule,
sufficient to know the behaviour of the integral algebraic functions, as we generally, when
we have to do with functions which are essentially different from these, will try through
transformations of the variables to get back to them and to certain functions which
resemble them in this respect.
We need only consider the cases in which we have only one variable essential
circumstance, of which the mean value of the observation is an algebraic function of the
,-th degree. We are able then, on any supposition as to the distribution of the observations,
0, and their essential circumstances, x, and weights, v, to determine r-fl substitutive
observations, 0, together with the essential circumstances, X, and weights, V, belonging
to them, in such a way that they treated according to the method of the least squares
will give the same results as the larger number of actual observations. The conditions are:
[0V-] ^ OoVo+ ...+OrVr j
(119)
[OX'V] = XlOo Fo + . . . + TrOrVr j
109
and
[,] = F„ + . . . + Vr
[x^-'v] = x':v, + ...+xrvr.
(120)
These Sr-\-2 equations are not quite sufficient for the determination of the 3r-l-3
unknowns. We remove the difficulty in the best way by adding the equation:
The elimination of the Vs (and O's) then leads to an equation of the r+l degree, whose
roots Xq, . . . Xr are all real quantities, if the given x's have been real and the v's
positive. When the roots are found, we can compute, first V^, ... Vr and afterwards
Oq, ... Or, by means of two systems of r-f 1 linear equations with r-\-l uniinowns.
If, for instance, the essential circumstances of the actual observations are contained
in the interval from — 1 to -\-l, and if the observations are so numerous and so equally
distributed that they may be looiied upon as continuous with constant mean error every-
where in this interval; if, further, the sum of the weights = 2; then the distribution of
the substitutive observations will be symmetrical around 0, and, for functions of the lowest
degrees, be
= o{
X
V
r = 3
(X =
\x =
(X =
6
•000
2-000 '
— -577,
1-000,
— -775,
-556,
— -861,
-348,
— -906,
•237,
— -932,
•171,
— •949,
•129,
+ •577
1^000 '
•000,
•889,
— ^340,
•652,
— -538,
•479,
— -661,
•361,
— •742,
•280,
+ ■775
-556 '
+ -340,
-652,
•000,
•569,
— •239,
•468,
- •406,
•382,
+ •861
•348 '
+ ^538,
•479,
+ ^239,
•468,
•000,
•418,
+
+
906
237 '
661,
361,
406,
382,
— •932
•171 '
+ -742, + ^949
•280, ^129
If, in another example, the distribution of the observations is, lilcewise, continuous,
but the weights within the element dx proportional to e^^^ consequently symmetrical with
maximum by a; = 0, then the distribution for the lowest degrees, the only ones of any
practical interest, will be
no
f X = -000
^\f= 2-000 '
jX = —1-000, +1-000
^ I r = 1-000, 1-000 '
X = -1-732, -000, +1-732
V = -333, 1-333, -333 '
/- = 3
X = -2-334, — -742, + -742, +2-334
V = -092, -908, -908, -092
_ fJt = —2-857, -1-356, -000, +1-356, +2-857
^ '^ \V = -023, -444, 1-067, -444, -023 '
^ (Z== -3-324, -1-889, --617, + -617, +1-889, +3-324
** ~ ^ i F = -005, -177, -818, -818, -177, -006 '
r = 6
jX = —3-750, —2-367, —1-154, -000, +1-154, +2-367, 3-750
\ V = -001, -062, -480, -914, -480, -062, -001
If we were able now to represent these substitutive observations as normal places,
then we should be able also, by the use of such tables in analogous cases, to prevent any
loss of exactness. It would be possible entirely to evade the application of the method of
the least squares; we had but to form such qualified normal places in just the same
number as the adjustment formula contains elements that are to be determined. This,
however, is not possible. Certainly, we can obtain normal places corresponding to the
required values of the essential circumstance, but we cannot by a simple formation of
mean numbers give them the weight which each of them ought to have, without employing
some of the observations twice, others not at all. By taking into consideration how much
the extreme normal places from this reason must lose in weight, compared to the sub-
stitutive observations, we can estimate how many per cent the loss, in the worst case,
can amount to. In the first of our examples we find the loss to be 0, for r = and
r = 1; but for r = 2 we lose 15, for r = 3 we lose 19, for r = 4 we lose 20, and
for greater values of r 21 p. c.
Example. Eighteen unbound observations, equally good, ^j(o) = j-^, correspond
to an essential circumstance whose values are distributed as the prime numbers p from
1049 to 1171. Taking (^—1105): 100 = x as the essential circumstance of the observa-
tion 0, we have:
Ill
X
X
X
— •56
— •41
— •14
-•15
+ •18
— •24
-•54
+ •50
-12
— •08
— •02
— •32
+ •33
-•21
+•24
+ •46
+ •48
+ •09
-•44
-•42
-•30
— •03
— •15
+ •48
+ •39
+ •12
+ •04
+ •12
+ •21
+ •40
+ •58
+ •66
-•24
-•18
+ •18
-•39
Dividing
these observations
into groups indicated by the horizontal lines, we get
the 6 normal pla(
3es:
X
weight
— •550
+
045
2
«
-•407
+
100
3
— •108
• —
034
5
+ •145
+
115
4
+ •470
+
255
2
+ •620
—
315
2
If we suppose the mean values of the observations to be a function of the third,
eventually second, degree of x, X^{o) = a-\-bx-\-cx^-^dx^, we have by ordinary application
of the adjustment by elements the normal equations:
6^72 == 216^00 a— \-20b ^29-98 c + \-Md
— 3-07 = -1^20 a + 29-98 &+ l-94c + 8-llf/
— 1^08= 29-98rt+ h946+ S-nc + l-21d
— l-U =
1^94 a
+
8^1U+ l-2lc
+ 2-56d.
the free
equations :
6^72 =
216-00 a
—
1^206 + 29^98 c
+ l-9id
- 3^03 =
29^976 + 2^11 c
-{-8-12d
- 1^79 =
3^80 r
+ •37c^
- ^54 =
•305 rf
a == +
•09,
a' = +^10,
b = +
•40,
6' = - -01,
c =- —
•30,
c' = —-47,
we get:
d = —1-n,
where a', b\ c' are the coefficients in the functions of second degree, obtained by pre-
supposing d = 0.
112
Now, by application of the normal places instead of the original observations, we
obtain on the same suppositions the normal equations:
6-72 = 216-00 a — 1-20 b + 29-45 c + 1-87 d
_ 2-84 = —1.20 a + 29-45 6+ l-87c+7-93a!
— -54= 29-45 a + 1-87 ft + 7-93 c+ 1-14 rf
— 1-57 = 1-87 a + 7-93 ft + 1-14 c + 2-45 d.
By the free equations:
6-72 = 216-00 0— 1-20 ft + 29-45 c+ 1-87 (^
we get:
-2-80 =
29-44 ft + 2-03 c + 7-94 d
— 1-26 =
3-77 c+ -Sid
- -76 =
-263 d.
a = + -07,
a' = + -08,
ft = -f- -69,
h' = --07,
c = - -07,
c' = --33,
d = —2-88.
A comparison between these two calculations, particularly between the leading
coefficients in the free equations, shows that the loss of weight amounts to 1 — |^, or
14 per cent. But it is only in the equation for d that the loss is so great; in the equa-
tions for ft and c, respectively, it is only two and one per cent.
Our normal places are very good if the function is only of the first or second
degree; for the function of third degree they can be admitted even though the values of
the elements a, ft, c, d have changed considerably. For functions of 4''' or higher degrees
these normal places would prove insufficient.
§ 64. That graphical adjustment is a means which can carry us through great
difficulties, we have shown already in practice by applying it to the drawing of curves of
errors. The remarkable powers of the eye and the hand must, like a deus ex machina,
help us where all other means fail.
Adjustment by drawing is restricted only by one single condition: if we are to
represent a relation between quantities by a plane curve, there must be only two quantities;
one of these, represented by the ordinate, is, or is considered to be, the observed value;
and the other, represented by the abscissa, is considered the only essential circumstance
on which the observed value depends.
Examples of graphical adjustment with two essential circumstances do occur,
however, for instance in weather-charts. In periodic phenomena polar co-ordinates are
preferred. But otherwise each observation is represented by a point whose ordinate and
113
abscissa are, respectively, the observed value and its essential circumstance; and the adjust-
ment is performed by free-hand drawing of a curve which satisfies the two conditions
of being free from irregularities and going as near as possible to the several points of
observation. The smoothness of the curve in this process plays the part of the theory,
and it is a matter of course that we succeed relatively best when the theory is unknown
or extremely intricate : when, for instance, we must confine ourselves to requiring that the
phenomenon must be continuous within the observed region, or be a single valued function.
But also such a theoretical condition as, for instance, the one that the law of dependence
must be of an integral, rational form, may be successfully represented by graphical adjust-
ment, if the operator has had practice in the drawing of parabolas of higher degrees. And
we have seen that also such functional forms as have the rapid approximation to an asymptote
which the curves of error demand, lie within the province of the graphical adjustment.
As for the approximation to the several observed points, the idea of the adjust-
ment implies that a perfect identity is not necessary; only, the curve must intersect the
ordinates so near the points as is required by the several mean errors or laws of errors.
If, after all, we know anything as to the exactness of the several observations before we
make the adjustment, this ought to be indicated visibly on the drawing-paper and used
in the graphical adjustment. We cannot pay much regard, of course, to the presupposed
typical form and other properties of the law of errors, but something may be attained,
particularly with regard to the number of similar deviations.
If we know nothing whatever as to the exactness of the several observations, or
only that they are all to be considered equally good, there can be only a single point in
our figure for each observation. In a graphical adjustment, however, we can and ought
to take care that the curve we draw has the same number of obserted points on each
side of it, not only in its whole extent, but also as far as possible for arbitrary divisions.
If we know the weights of the observations, they may be indicated on the drawing, and
observations with the weight n count w-fold.
In contradistinction to this it is worth while to remark that, with the exception
only of bonds between observations, represented by different points, it is possible to lay
down on the paper of adjustment almost all desirable information about the several laws of
errors. Around each point whose co-ordinates represent the mean values of an observation
and of its essential circumstance, a curve, the curve of mean errors, may be drawn in
such a way that a real intersection of it with any curve of adjustment indicates a devia-
tion less than the mean error resulting from the combination of the mean errors of the
observed value and that of its essential circumstance, if this is also found by observation,
while a passing over or under indicates a deviation exceeding the mean error. Evidently,
drawings furnished with such indications enable us to make very good adjustments.
15
114
If the laws of errors both for the observation and for its circunastance are typical,
then the curve of mean errors is an ellipse with the observed points in its centre.
If, further, there are no bonds between the observation and its circumstance, then
the ellipse of mean errors has its axes parallel to the ordinate and the abscissa, and their
lengths are double the respective mean errors.
If the essential circumstance of the observation, the abscissa, is jjnown to be free
of errors, the ellipse of the mean errors is reduced to the two points on the ordinate,
distant by the mean error of the observation from the central point of observation. In
special cases other means of illustrating the laws of errors may be used. If, for instance,
the mean errors as well as the mean values are continuous functions of the essential
circumstance of the observation, continuous curves for the mean errors may be drawn on
the adjustment paper.
The principal advantages of the graphical adjustment are its indication of gross
errors and its independence of a definitely formulated theory. By measuring the ordinates
of the adjusted curve we can get improved observations corresponding to as many values
of the circumstance or abscissa as we wish, and we can select them as we please within
the limits of the drawing. But these adjusted observations are strongly bound together,
and we have no indication whatever of their mean errors. Consequently, no other adjust-
ment can be based immediately upon the results of a graphical adjustment.
On the other hand, graphical adjustment can be very advantageously combined
with interpolations, both preceding and following, and we shall see later on that by this
means we can remedy its defects, particularly its limited accuracy and its tendency to
place too much confidence in the observations, and too little in the theory, i. e. to give
an under-adjustment.
By drawing we attain an exactness of only 3 or 4 significant figures, and that is
frequently insufficient. The scale of the drawing must be chosen in such a way that the
errors of observations are visible; but then the dimensions may easily become so large that
no paper can contain the drawing. In order to give the eye a full grasp of the figure,
the latter must in its whole course show only small deviations from the straight line, which
is taken as the axis of abscissae. This is a practical hint, founded upon experience. The
eye can judge of the smoothness of other curves also, but not by far so well as of that
of a straight line. And if the line forms a large angle with the axis of the abscissae,
then the exactness is lost by the flat intersections with the ordinates. Therefore, as a rule,
it is not the original observations that are marked on the paper when we make a graphical
adjustment, but only their differences from values found by a preceding interpolation.
In order to avoid an under-adjustment, we must allow | of the deviations of the
curve from the observation-points to surpass the mean errors. It is further essential that
115
the said interpolation is based on a minimum number of observed data; and after the
graphical adjustment has been made, it is safe to try another interpolation using a smaller
number of the adjusted values as the base of a new interpolation and a repeated graphical
adjustment.
If the results of a graphical adjustment are required only in the form of a table
representing the adjusted observations as a function of the circumstance as argument, this
table also ought to be based on an interpolation between relatively few measured values,
the interpolated values being checked by comparison with the corresponding measured
values. A table of exclusively measured values will show too irregular diiferences.
When we have corrected these values by measuring the ordinates in a curve of
graphical adjustment, they may be employed instead of the observations as a sort of normal
places. It has been said, however, and it deserves to be repeated, that they must not be
adjusted by means of the method of the least squares, like the normal places properly so
called. But we can very well use both sorts of normal places, in a just sufficient number,
for the computation of the unknown elements of the problem, according to the rules of
exact mathematics.
That we do not know their weights, and that there are bonds between them, will not
here injure the graphically determined normal places. The very circumstance that even distant
observations by the construction of the curve are made to influence each normal place, is an
advantage. It is not necessary here to suffer any loss of exactness, as by the other normal
places, which, as they are to be represented as mean numbers, cannot at the same time be
put in the most advantageous places and obtain the due weight. As to the rest, however, what
has been said p. 108 — 110 about the necessity of putting the substitutive observations in
the right place, holds good also, without any alteration, of the graphical normal places.
The method of the graphical adjustment enables us to execute the drawing with
absolute correctness, and it leaves us full liberty to put the normal places where we like,
consequently also in the places required for absolute correctness ; but in both these respects
it leaves everything to our tact and practice, and gives no formal help to it.
As to the criticism, the graphical adjustment gives no information about the mean
errors of its results. But, if we can state the mean error of each observation, we are able,
nevertheless, to subject the graphical adjustments to a summary criticism, according to
the rule V-'(o_m)2
I'-
m.
And with respect to the more special criticism on systematical deviations, the graphical
method even takes a very high rank. Through graphical representations of the finally
remaining deviations, o — m, particularly if we can also lay down the mean errors on the
same drawing, we get the sharpest check on the objective correctness of any adjustment.
15*
116
From this reason, and owing to tlie proportionally slight difficulties attached to it,
the graphical adjustment becomes particularly suitable where we are to lay down new
empirical laws. In such cases we have to work through, to check, and to reject series
of hypotheses as to the functional interdependency of observations and their essential
circumstances. We save much labour, and illustrate our results, if we work by graphical
adjustment.
Of course, we are not obliged to subject observations to adjustment. In the pre-
liminary stages, or as long as it is doubtful whether a greater number of essential circum-
stances ought not to be taken into consideration, it may even be the best thing to give
the observations just as they are.
But if we use the graphical form in order to illustrate such statements by the
drawing of a line which connects the several observed points, then we ought to give this
line the form of a continuous curve and not, according to a fashion which unfortunately
is widely spread, the form of a rectilinear polygon which is broken in every observed
point. Discontinuity in the curve is such a marked geometrical peculiarity that it ought,
even more than cusps, double-points, and asymptotes, to be reserved for those eases in
which the author expressly wants to give his opinion on its occurrence in reality.
XIV. THE THEORY OF PROBABILITY.
§ 65. We have already, in § 9, defined ''probability' as the limit to which — the
law of the large numbers taken for granted — the relative frequency of an event approaches,
when the number of repetitions is increasing indefinitely; or in other words, as the limit
of the ratio of the number of favourable events to the total number of trials.
The theory of probabilities treats especially of such observations whose events
cannot be naturally or immediately expressed in numbers. But there is no compulsion in
this limitation. When an observation can result in different numerical values, then for
each of these events we may very well speak of its probability, imagining as the opposite
event all the other possible ones. In this way the theory of probabilities has served as
the constant foundation of the theory of observation as a whole.
But, on the other hand, it is important to notice that the determination of the
law of errors by symmetrical functions may also be employed in the non-numerical cases
without the intervention of the notion of probability. For as we can always indicate the
mutually complementary opposite events as the "fortunate" or "unfortunate" one, or as
"Yes" and "No", we may also use the numbers and 1 as such a formal indication. If
117
then we identify 1 with the favourable "Yes"-event, with the unfavourable "No", the
sums of the numbers got in a series of repetitions will give the frequency of affirmative
events. This relation, which has been used already in some of the foregoing examples, we
must here consider more explicitly.
If repetitions of the same observation, which admits of only two alternatives, give
the result "Yes" = 1 m times, against m times "No" = 0, then the relative frequency
for the favourable event is
m + M
But if we employ the form of the symmetrical functions
for the same law of actual errors, then the sums of the powers are
Sg = m-\-n , Sj = §2 .... = Sr = m. (121)
In order to determine the half-invariants by means of this, we solve the equations
m =
m =
tn =
m • ,u 1 + 3m • j«o -f 3m • /ig + (w -j- «) /«4 i
and find then
m
Ml
fit
Fa
mn
f^i ==
{m -{- nf
mn {n — tn)
{m-\-nY
mn {n^ — 4mn + »»*)
(122)
(m-f-w)*
Compare § 23, example 2, and § 24, example 3.
All the half-invariants are integral functions of the relative frequency, which is
itself equal to fi^. The relative frequency of the opposite result is — qj- = 1 — fXi; by
interchanging m and w, none of the half-invariants of even degree are changed, and those
of odd degree (from fi.^ upwards) only change their signs.
In order to represent the connection between the laws of presumptive errors, we
need only assume, in (122), that m and n increase indefinitely, while the probability of the
event becomes p = — ^-— , and the probability of the opposite event is represented by
— ^-^ — = 1 — p = q. The half invariants are then :
m-{-n ^ ^
h = M
-^3 = mii—p)
K = P