SASSMag S WON WN OOM re 3 : es SESS ; . > = ~ a a £ MOLALLA ALLLRT ET aes ~ <3 . e, we SN Ss ay St * SS Sa SS Sh SS S SOSA = : cmon Se RRASAAw ~ Th Pe Dea eer gre’ ks yo Cen ~ Pe Vee ee a Le - a Ly . \ a ~ ‘ ‘ } - t , - ; ui? c d 4 q i “PRINTERS, _ 7 205-213 Hast nee fa NEW YORK. Fah . E 2 awe ! PORTAL awn 2 py 1 8 Fe F PLL 8 - ee HRP CORLISS? ART. 2—The two general problems in approximate computa- tions, - - - - - 3—Limits of approximate numbers, ~ - 4—Definition of absolute error, - - - Definition of relative error, ~ - 5—Prop. 1. Determination of superior limit of relative error of result from given limit of absolute error, PROB. 2. Determination of superior limit of absolute error of result from given limit of relative error, PROB. 3. Determination of allowable limit of relative error from assigned limit of absolute error, - PROB. 4. Determination of allowable limit of absolute error from assigned limit of relative error, - 6é—Relative error independent of position of decimal point, 7—Order of units affected by a given relative error, 8—Rejection of redundant figures of a result, - 9—PROB. 5.—To determine how many exact figures of a result must be computed so that its relative error shall not exceed an assigned limit, - - 1I—PROB. 6. Necessary approximation of quantities for __ addition, the approximation of the result being as- Signed, pe = = = = 13—Approximation of the sum of approximate quantities, 14—PROB. 7. Necessary approximation of quantities for subtraction, the dusters of the result being assigned, - - - - 15—Approximation of the Aifference of aes ae te quan- tities, 16—PROB. 8. To form a product by peadecd multiplica- tion, exact within a unit of the zth order of decimals, 19—PROB. 9. Necessary number of decimals in factors for abridged multiplication, the product being required to the zth order of decimals, 21—Extension of the rule for Prob. 9, where the limit of error is a unit of higher order than decimals, - PAGE O ONDA DA wm & WWwWNN —_ IT 13 17 18 19 20 24 29 iv CONTENTS. ART. PAGE, 22—PROB. 10. Absolute approximation of the product of two approximate factors, - - - 23—Absolute approximation of the product of any number of approximate factors, - - ~ 24—PROB. 11. To perform an abridged multiplication so that the error of the process shall be less than that due to the errors of the factors, - - Mode of indicating limits of error of OEE: quan- tities in practice, - - - 25—PROB. 12. Relative approximation of the product of any number of approximate factors, - - 26—Absolute approximation of the square of an approxi- mate number, - - - - 27—Absolute approximation of the cube of an approximate number, ma = - ~ - 28— Relative approximation of the square and cube and ath power of an approximate number, - 29—PROB. 13. Order of units of the pene significant figure of a quotient, - 30—PROB. 14. To form a aenent by abridged division; exact within a unit of the zth order, - 32—PrRoOB. 15. Absolute approximation of the quenee. of an approximate dividend and divisor, 33—PROB. 16. To perform an abridged division so that the error of the process shall be less than that due to the errors of the quantities, - - < 34—Determination of the error of an appro quoucn by inspection, - 35—PROB. 17. Relative approximation of the quotient of an approximate dividend and divisor, - a 37—Approximate computation of the fraction eo 38—-PROB. 18. Extraction of sels root by abndeed method, - - 39—Square root of unity plus or minus a Sencil fraction, 40—PROB. 19. Absolute approximation of squee root of an approximate number, ~ - 41—Square root generally to be found with as many exact figures as there are in the number, - ~ 42—PROB. 20. Relative approximation of square root of an approximate number, ~ - - 43—Error of square root found by inspection, ~ 44—PROB. 21. Extraction of cube root by abridged method, 45—Cube root of unity plus or minus a small fraction, 31 32 32 33 36 39 40 4I 43 Aa 49 5° 52 G3 55 58 63 64 66 . 67 68 71 75 CONTENTS. ART. 46—PROB. 22. Absolute approximation cf cube root of an approximate number, - - - 47—Cube root generally to be found with as many exact figures as there are in the number, - - 48—PROB. 23. Relative approximation of cube root or #th root of an approximate number, - - 49—Error of cube root found by inspection, - 50—-Logarithmic computations, ~ ~ - 51— Natural trigonometric functions, - - 52— PROB. 24. Computation of the value of any complex monomial whose factors may be taken as accurately as we please, - - - = 53— Computation of complex polynomials, - 54—Computation of complex expressions containing factors known with only a limited degree of approximation, v PAGE. 76 86 92 94 neg PRINCIPLES” OF APPROXIMATE COMPUTATIONS. INTRODUCTORY. 1. It is frequently necessary to perform arithmet- ica] operations upon quantities whose numerical val- ues cannot be taken with absolute accuracy, either because these values are incommensurable with unity or because they are the results of measurements made in Astronomy, Physics, Chemistry, Engineering, etc., with instruments capable of giving only a limited de- gree of precision. When the ordinary rules of arith- metic are applied to a computation involving such approximate quantities the operation is often very long and tedious, and when it is finished the operator may be in doubt as to the degree of accuracy of his final result; that is, as to the amount of error to which this result is liable from the errors of the quan- tities employed. The object of the following pages is to present some simple rules for conducting com- putations involving approximate quantities, in such a manner as to require the fewest figures and to show at once the degree of accuracy of the result. The methods are partly suggested, as also a few of the “2 APPROXIMATE COMPUTATIONS. examples, by the small French treatises of Martin, Babinet and Housel, etc., but many changes have been made in the statement and demonstration of the rules, and much not found in those treatises has been added. 2. In practice we commonly either know before beginning a computation what degree of accuracy will be sufficient in the result, or else we wish to com- pute a result with all the accuracy possible. Our computations accordingly come under one or the other of the following general problems: First. A set of quantities being given, whose values may be taken as accurately as we please, to computea result with any required accuracy. Second. , we should have to find four exact figures of the root, viz. : 1.414; but if the square root of 0.0019 were required with the same limit of relative error, we should need only three exact significant figures, viz.: 0.0436. It is obvious that if the limit of absolute error of a required result were placed at a unit, instead of half a unit, of the lowest order to be retained, the number of figures necessary in the result could be determined by Rule 5, if the word a/f, wherever it occurs in the rule, were struck out. For example, if it be asked how many figures it is necessary to retain in V 19, so that if it be simply known that the absolute error is less than a wzz¢t of the lowest order retained, the rela- tive error shall be less than <5, we see that three figures, viz. : 4.35, are not enough, but that four, viz. : 4.358, are sufficient. 10. It is worth noticing, that if exact quantities are added to or subtracted from approximate quanti- ties the absolute error of the result will be the same as if only the approximate quantities were taken; and also that if an approximate quantity be multiplied or divided by an exact number the relative error will re- main unchanged. Thus if z be taken equal to 3.1415, the sum of this plus or minus any exact number will have the same absolute error; and the product or quotient of w= 3.1415 by any exact number will have the same relative error; as is evident from Art. 6. ADDITION. 11. Prosenm 6. A set of quantities being proposed, whose values may be taken as accurately as we please, it is required to determine how far each must be computed, so that the absolute error of their sum shall not exceed half a unit of the zth order. RULE 6. Jf there are not more than ten quantities to be added, compute each so that tts absolute error shall not exceed half a unit of the order next below the nth. If there are more than ten quantities to be added, indicate by the plus or minus sign the direction of the error of each quantity when computed as just stated, or else compute in cach still another figure. The rule needs but little explanation. If there are not more than ten errors in either direction, each not greater than, say, 0.005, or half a hundredth, the sum of all these errors will not exceed 0.05, or half a tenth. If we wish to indicate the direction of the error of an approximate quantity, we place a plus sign after it when its error would have to be added to give the true result, and a minus sign in the opposite case. EXAMPLE I. Find the sum of the square roots of prmerreerDe rss 22 0 8 O., 75 S410, Ly. b2, 413,04; 108, 17, and 18, so that the absolute error of the result shall be less than 0.0005. Taking each with four exact decimals, we have 14 APPROXIMATE COMPUTATIONS. 1.4142 -- 1.7321 — 22301 — 2.4495 2.64538 2.8284 3.1623 3.3166 3.4641 3.6056 3-7417 3.8730 Aolost 4.2426 42.8351 There being in the above example not more than eight errors in one direction, each less than 0.00005, the error of the sum cannot exceed 0.0004, which is ‘less than the assigned limit 0.0005. Observe, also, that in this particular example, by finding this smaller limit, viz.: 0.0004, since the last figure of the sum happens to be 0.0001, we are able to reject this and still to say that 42.835 is correct within 0.0005, or that this is the answer with five exact figures. But if we take the sum of the same numbers, omitting the last one, we find it to be 38.5925. We aré therensre evidently unable, without knowing the direction of the error, to reject the last figure of this result, and still have the answer contain five exact figures. In general when we wish to retain all the exact figures of an approximate result, and no more, we must find between what limits the true result lies, and then retain only those figures which would be kept for numbers at either limit. For example, in the case just cited, where the sum of thirteen numbers is Wed Peerage 4 ++ | ADDITION. 1s 38.5925, eight of them are taken too large and five too small; hence the true result must lie between 38.5921 and 38.5928. We can therefore take 38.59 as the result with four exact figures, but by our defi- nition neither 38.592 nor 38.593 as the result with five exact figures, for the computation does not shaw which of these results is the more accurate. EXAMPLE 2. es the series 3 3° 72 seeunraee retaining only eight decimals in re result, and make the error of the result less than a unit of the eighth decimal place. Observing the remark near the end of Art. 8, and taking nine terms, each exact to the ninth decimal place, we have, 0.666666667 24691358 1646091 130642 11290 1026 96 2 I 0.693147180 Answer, 0.69314718. [ee J++4+4 | EXAMPLE 3. Compute the expression V44+ Vo0.08 + Vo0.06 + 1.7435013...., with three exact decimals in the result. Taking four exact decimals in each term the sum will be found to be 8.9044. But the possible error of this result is greater than 0.0001. Hence the last decimal cannot be rejected without determining still 16 APPROXIMATE COMPUTATIONS. another one. (Art. 8, at end). Let the next decimal be determined. 12. It is clear that if any other limit of absolute error be assigned for a sum than that stated in Prob- lem 6, the allowable limit of absolute error for each of the quantities to be added may be found by divid- ing the assigned limit of absolute error of the sum by the number of quantities to be added. EXAMPLE 4. Compute the expression 352.7856... + 7¥2+ V7 + 25.00082... + 0.000074..., with a relative error in the result not to exceed zadas. We see that the sum will be greater than 375. Hence, by Rule 4, we are allowed to make an absolute error in the sum equal to s%$30, and as there are five numbers to be added we are evidently at liberty to make an absolute error in each of them equal to xX sthta = socos, Which 1s ‘greater thansoue Hence, if the thousandths place in each number be found within a unit of that order, the result will satisfy the conditions. Taking the first four numbers with that approximation, and neglecting the fifth, we have 352.785 1.414 2.645 | 25.000 381.844 with an absolute error less than 0.005, and a relative error, therefore, less than sou%aw < sod0~- EXAMPLE 5. Calculate the sum of the reciprocals. of-the numbers 3/°7;-0; 11; 13,774,. 17-5100 22, 23, 26, 27, 28, and 29, so that the relative error of the result shall be less than sggov. ADDITION. 17 13. It is evident that if it be required to find the sum or several quantities which can each be obtained with only a limited degree of approximation, if we do not know the direction of the errors of any of them, we must take as the limit of the absolute error of the sum, the sum of the possible absolute errors of the quantities. If the absolute error of one of the quanti- ties greatly exceeds those of the others we should commonly not take the trouble to find the exact sum of the errors; but we may take some approximate value of it that we can see would exceed it. For example, in adding the numbers 76.3, 18.71, and 528.345, supposed approximate within half a unit of the lowest order in each, we should be satisfied with saying that the sum could be found exact within 0.06. EXAMPLE 6. Add the following numbers, supposed approximate within 2 units of the lowest order in each, and assign a limit to the absolute and relative error of the sum: 35.278, 26.435, 18.7346, 21.0064, 3.2176, 0.2142, 0.00125, Z SUBERACIION: 14. PRoBLEM 7. Two quantities being proposed whose values may be taken as accurately as we please, it is required to determine how far each must be computed so that the absolute error of their difference shall not exceed an assigned limit. RULE 7. Compute each quantity so that tts absolute error shall not exceed half the limit of error assigned Jor the difference. The rule needs no demonstration. And it is also evident that if the errors of the two quantities were in the same direction, these errors would not need to be made smaller than the limit assigned for the error of the difference. Thus, if a difference of two numbers be required within a unit of a given order, it will be sufficient to compute each to that order of units when- ever it can be seen that their errors will be in the same direction. EXAMPLE 7. Compute the expression V7 — V5 so that the absolute error of the result shall not ex- ceed 0.001. The answer may be either 0.409 or 0.410. EXAMPLE 8. Compute the expression V95 — V5, with a relative error less than z5dos. By calculating the first two figures of each root we find the difference will be greater than 7. Hence we are at liberty to make an absolute error in the result equal to 0.0007. But not knowing beforehand the SUBTRACTION. 19 amount of the errors that would result by taking both numbers too small or both too large, with three deci- mals each, nor the direction of the errors if each were to be taken to the nearest 3d decimal, we have to determine in each the 4th decimal, giving 9.7468 — TPN GAN ee 7-5107 Since the errors of the two quantities happen now to be in the same direction, the absolute error of the result cannot exceed 0.00005. Then the true differ- ence lies between 7.5106 and 7.5108. Either of these limits would give for the answer with four exact figures, 7.511, the absolute error being then less than half a thousandth, and the relative error, therefore, (Art. 6) less than tsdou < roo0d aS required. EXAMPLE 9. Compute the expression “V3 — V2, with a relative error less than zoos. Answer, 0.02804. 415. It hardly needs to be pointed out that if two numbers can be obtained, each with only a limited degree of approximation, the limit of the absolute error of the difference of the numbers will have to be | taken equal to the sum of the limits of the errors of the numbers, unless it is known that their errors are in the same direction; in this case the larger of the two limits of error may be taken as the limit of error of the difference. For example, if the numbers 3.725 and 1.834 are each approximate within half a thou- sandth, but the direction of the errors not known, there will be an uncertainty of a thousandth in the difference, 1.891. If, however, we have the numbers 3.725 and 1.8342 each known to be foo small by not more than half a unit of its lowest order, the error of the difference 1.8908 cannot exceed half a thousandth. MULCTIREVCAT ON: 16. Prosiem 8. Two factors being proposed, whose values may be taken as accurately as we please, it is required to form their product so that its absolute error shall not exceed a unit of the zth order of decimals. RULE 8. Jake either factor for the multiplier, and write wt with its figures in reverse order under the multiplicand, and in such a position that the original simple untts figure of the multiplier shall come under the (n+ 1)th decimal of the multiplicand. Begin each partial product with the product of the multiply- wing figure into the figure of the multiplicand directly over tt, rejecting the figures of the multiplicand to the right of this, but correcting this product, tf necessary. by adding to tt the number of units of the same orde nearest to what would be added tf the rest of the multi- plicand were used, and place the partial products with thetr right hand figures in a vertical line. The sum of the partial products will have n+ 1 decimals, the last of which must generally be regarded as doubtful. EXAMPLE 10. Compute the product 763-05403698956 X 25.4463057845 with an absolute error less than 0.0001. MULTIPLICATION. 21 Operation: BES a eee 548750 3 644.52 erscor on] fies 381 5 27018 20m 22 101 BOs 22516 4578 32 228 92 3 82 oe 6 1941 6.906 34. This method of abridged multiplication is ascribed to Oughtred. The explanation of it is very simple. In the above example the original units figure 5 of the multiplier stands under the fifth decimal figure of the multiplicand; hence, considering first the second par- tial product its right hand figure will be of the fifth order of decimals. But the original tens figure of the multiplier stands under the sixth decimal of the multiplicand, and the first deczmal of the multiplier under the fourth decimal of the multiplicand, so that the partial products made with these figures of the multiplier will also begin at the right with the fifth decimal place; and so for all the partial products, since the position value of the figures of the reversed multiplier diminishes towards the left, in the same ra- tio as that of the figures of the multiplicand increases. The position of the decimal point in the result is easy to fix. It may in any case be determined by the last sentence of Rule 8. If the multiplier in any example has no entire figures, supply the place of simple units with a zero, 22 APPRIXIMATE COMPUTA TIONS. Let us consider the possible error of a result found by Rule 8. If the rule with regard to correcting the partial products to allow for the part of the multipli- ~ cand each time rejected is carefully followed, it is evi- dent that the error of each partial product cannot exceed half a unit of its lowest order; and since the right hand figure of each partial product is of the same order as that of the final result, the error of this result, whenever the number of partial errors does not exceed twenty, cannot be greater than ten units of the lowest order obtained; that is, it will not ex- ceed ove unit of the next higher order, which is the limit assigned in the statement of the problem. ‘,. In the example worked above, the last partial prod- uct Comes from using the figure 8 of the multiplier ; and if the next ficure of the “multiplier at the left had been large enough to give more than half a unit of the lowest order in the result, we should have put in another partial product equal to one unit of that or- der. But since the number of partial products does not exceed ten, the total error of the result cannot exceed five units of its lowest order, and we may therefore reject the last figure 4 of the result, and the answer will still be correct within 0.0001, the assigned limit of error. 1%. It is easy in practice to indicate the direction of the errors of the partial products by the plus and minus signs, and thus, when the work is done, to often determine a much smaller limit of error than that assigned in advance. EXAMPLE Ioa. Make the product of the same fac- tors as in Example Io, with the same limit of error, but taking the former multiplier for the multiplicand. MULTIPLICATION, 23 25.4463 057845 6598963 0450.367 178 I241 405 — 15 2677 835 7633 892 127232 103179 76 15 Be 19416.90 636 +++ | By thus marking the direction of the observe that five of the partial products are and three are too small; hence the error of the cannot be more than 23 units of the lowest order ob- tained; which is only a quarter of the limit of error assigned in advance. But without determining a still smaller limit of error, we could not decide from either this operation or that of example 9, or both of them, whether, if we wished to reject the last decimal of the result, the one before it should be left a 3 or changed to 4. 18. EXAMPLE II. Multiply 0.995....by 9.95.... so that the absolute error of the result shall not ex- ceed 0.01. By the rule for arrangement we have 00h he pe tb RO 8955 89 6 50 9.90 I But let us examine the possible error of this result, 24 . APPROXIMATE COMPUTATIONS, supposing the 5 in each factor to be liable to an e: ror of half a unit of its own order. The first partial product would evidently be liable to an error of 44 units of the order of the last place in the product; the second partial product would be liable to an error of nearly a unit of the same order, since, besides the error of the multiplicand, we make an additional error in rejecting a figure of this partial product; and, making a similar observation, the last partial product might be wrong by about 5 units of the same order. It is therefore possible that we have exceeded the as- signed limit of error. In fact, if the true values of the numbers were 0.9945 and 9.945, their exact prod- uct would be 9.8903025, a result differing from the approximate one found above, by 0.0106975, or a trifle more than the assigned limit of error. It is thus seen that in the application of Rule 8 it will not in all cases be enough to know within a half unit of its own order the figure of the multiplicand standing over the right hand figure of the reversed multiplier, and that of the multiplier standing under the highest figure of the multiplicand. But a single additional figure in each factor will be amply sufficient. For if the multiplicand has one exact figure to the right of the multiplier, and the multiplier one to the left of the multiplicand, we may always reduce the errors of the first and last partial products each to a single unit of the lowest order obtained in the prod- uct; and since the assigned limit of error is 10 of these lowest units there is still room for 16 additional partial products with errors all in the same direction, each less than half of one of these units. 19. PRosLeM 9. The product of two factors being required within a unit of the zth order of decimals, it is required to state a rule for determining the number of decimai places to be com- puted in each factor. - MULTIPLICATION. 25 RULE 9. Either factor being regarded as the multt- plicand, compute one more decimal place in the other factor than the whole number of significant figures in the multiplicand above the nth order of decimals in- clustve.* This rule follows from the arrangement of the fac- tors by Rule 8. According to that Rule the units figure of the multiplier will stand under the (z + 1)th decimal of the multiplicand. Rule 9g will then give us one figure of the multiplier to the left of the high- est significant figure of the multiplicand, as required by the last section. Thus if it were required to multi- ply 0.007425625 by 99.284376 with an absolute error less than a unit of the 4th decimal place, the arrange- ment by Rule 8 would be 0.0074 25625 67 3482.99 from which it is evident that the three left hand figures of the multiplier will not be used in forming the prod- uct. Striking them off, the number of deczmads re- maining in the multiplier, viz.: three, will exceed by one the number of significant figures in the multipli- cand above its 4th decimal place inclusive, viz.: two. The number of necessary figures in either factor will not be altered by making the former multiplier the multiplicand. For, making this change we should have 99.2843 76 52652 4700.0 In this arrangement the two left hand figures of the * Whenever we speak of significant figures in this way we of course include any zeros that come in below the highest significant figure. Thus the number o.0010702 would be regarded as having five significant figures. 26 APPROXIMATE COMPUTATIONS. new multiplier may be struck off, and the number of decimals left in it, viz.: seven, will again exceed by one the number of significant figures in the new multiplicand above its 4th decimal place inclusive, Wide sio EXAMPLE 12. Determine by Rule 9 how many decimal places would have to be calculated in V¥9057 and ¥V0.0093 so that their product by Rule 8 should not be wrong by more than a unit of the 5th decimal place. Regarding V¥9057 as the multiplicand, it will have seven figures above the 5th decimal place inclusive. Hence the number of decimal places to calculate in V0.0093 will be eight. Regarding vo0.0093 as the multiplicand it will have four significant figures above the 5th decimal place inclusive. Hence the number of decimal places to calculate in V9057 will be five. Let the figures be obtained and the product formed by Rule 8. 20. Rule 9 is framed to cover safely all cases. But in a great majority of examples, viz.: where the sum of the highest significant figures of the two fac- tors, plus one-half the number of partial products, does not exceed 18, we may do with one less exact figure ; ’ and if the sum of the highest significant figures of the two factors, plus one-half the number of partial prod- ucts, is less than 8, we shall even then generally get the result with an error less than fa/fa unit of the wth order. It is easy in any special example to re- cognize the possible error committed. EXAMPLE 13. Compute the expression V0.0003 x v1000 with a relative error less than yodu0- MULTIPLICATION. 27 The square root of 0.0003 is more than 0.017, and that of 1000 is more than 30. Hence the product will be more than o.5. By Rule 4 we are at liberty then to make an absolute error in the result equal to zO-eee = 0.00005, that is, half a unit of the ath decima place. The sum of the highest significant figures of the factors, I and 3, is so small that we shall probably be safe if we work as if the allowable error were a whole unit of the 4th decimal place, and take also one less decimal in each factor than Rule 9 would require. (The reason why the error of a result is less when the sum of the highest significant figures of the factors is small, is evidently that the sum of the errors of the first and last partial products will then also be small). Since vV0.0003 will have three significant figures above the 4th.decimal place inclusive, we must then know three exact decimals in V1000; and we may determine the requisite number of decimals in V0.0003 in a similar way, or by simply considering that we want the same number of significant figures in it as in ¥1000, which, from what has just been found, must be five. We then want the 6th decimal place in V0.0003. The multiplication will be as follows: 0.54 774 28 APPROXIMATE COMPUTATIONS. In taking each factor in this example to the nearest unit of the lowest order retained, each is too large. If we suppose the multiplicand too large by half a unit of its lowest order, the first partial product would also be too large by half a unit of its lowest order, since the multiplying figure is 1. And if we suppose the multiplier too large by half a unit of its lowest cr- der, the last partial product would be too large by about 14 units of the same order as before, since the highest figure of the multiplicand is 3. The errors of the second and third partial products cannot in- crease the error of the result to more than 3 units of its lowest order; hence the true result cannot be less than 0.54771. On the other hand, if the errors of the factors are as near zero as possible the sum of the three partial products which could then be too small will be less than a unit of the lowest order of the re- sult. Hence the true result cannot be greater than 0.54775. We may then, if we please, reject the last decimal obtained in the result, and the answer, 0.5477 will have four exact figures, and a relative error less than szgo0 < rodeos, as required. EXAMPLE 14. Compute the expression 100 7 V2, with a relative error in the result not exceeding zoq00- The product ES esas 400. ‘Then the allowable absolute error is 7é$89 = 0.04. If we work then as if the limit of error were ove unit of the second deci- mal place, instead of four, we may safely determine the requisite number of decimals by taking one less than Rule 9 would give. Regarding 100 z7 as a sin- gle factor, it will have five figures including the 2d decimal place, and we therefore need five decimals in v2. Computing these figures, and taking the value of 100 w to correspond we have MULTIPLICATION. 29 314.159 + + 124 14.1 314159 + 125664 — 3142 — an RY [eae Coca Seats 444.28 8 If the 6th decimal figure of the multiplier be sup- posed equal to 5, there would be another partial product not exceeding 2 of the lowest units of the product. As for the partial products written, only two are too small. Hence the result cannot be too small by more than 3 of its lowest units. And it cannot be too large by more than 2 of these units. Therefore the true value of the result lies between 444.291 and 444.286. Hence the result to the near- est hundredth is 444.29, with a relative error less than poo eae ao ee See 440000 =< 100000° 21. It is easy to see that if the highest significant figure of either factor is in the zth decimal place, then by Rule g there would be 2 decimals to calculate in the other factor. If the highest significant figure of one factor is # places below the wth there will then be 2 — p decimals to be calculated in the other factor; and the extension of Rule 9g to cases in which the as- signed limit of error is a unit of higher order than decimals is therefore easy. EXAMPLE 15. How many figures must be taken in the factors 9843.768 and 947.84321, so that the error - of the product shall be less than a million ? The highest figure of the: first factor stands in the third place below millions, hence the required num- 30 APPROXIMATE COMPUTATIONS. ber of decimals in the other factor is 2 — 3 = —1; that is, we may neglect the units figure. ‘The highest figure of the second factor is in the fourth place be- low millions, hence the number of decimals required in the first factor is 2 — 4 = —2; that is, we may neglect the tens. We have then 98 .059 88 a 93 Since the original units figure of the multiplier would thus come under the place of hundred thou- sands in the multiplicand, the right hand figure of the product obtained is hundred thousands; hence the product is 9300000, with an error less than a million. EXAMPLE 16. How many figures must by this ex- tension of Rule 9 be computed in each of the factors, V98734216 and vV0.0093, so that the error of the product by Rule 8 shall be less than a unit of tens place. The simple units in the value of the first factor may be disregarded, and we want four decimal places in the value of the second factor. Let the factors be found with this approximation, and their product formed by Rule 8. EXAMPLE 17. How many figures must be com- puted in V758425 and vV0.00009, so that the error of their product by Rule 8 shall be less than 100? In Article 52 will be found a general rule for deter- mining the necessary approximation of each factor where an expression contains more than two factors, either as multipliers or divisors or both. MULTIPLICATION, 31 22. PROBLEM Io. Two factors being given, each to a cer- tain degree of approximation, it is required to assign a superior limit to the absolute error of their product. RULE 10. Multiply a superior limit of each factor by the absolute error of the other factor. The sum of the two products thus obtained may be taken for the limit of the absolute error of the product of the fac- Lors. Demonstration. Suppose there be given the two approximate factors, a’ and 6’, whose absolute errors are a and f, we are to determine a limit of the error of the product, a’0’. In the most unfavorable case the errors will be in the same direction; if then a and 6 are the true values of the factors, suppose a’ =a +a, of =b+ 6. Multiplying these two equations, mem- ber by member, the product will be all!’ =ab+ apt bat af, and the absolute error of the product will be a’! — ab = apt bart af. (7) Now since a and £ are usually very small compared with a and @, the product af will be very small com- pared with af and da, so that we have for the abso- lute error of the product very nearly a’l’ — ab= af + ba. (8) In finding the limit of error by Rule 10 we shall evi- dently get a larger limit than would be given by the right hand member of equation (8), since instead of a and 6 we take by the rule quantities known to be larger than them. Hence the rule is a safe one, not- withstanding the neglect of the very small quantity afi. EXAMPLE 18. Determine the limit of absolute er- ror in the product of the factor 784.2817, supposed to have an absolute error not exceeding 0.0004, by 32 APPROXIMATE COMPUTATIONS. the factor 3.483, supposed to have an absolute error not exceeding 0.006, We may evidently assume as a safe limit, 800 X 0.006 + 3.5 X 0.0004 = 4.8 + 0.0014 The product may then be made with an error not ex- ceeding 5 simple units. 23. If we have three approximate factors, a’ = ata, W=6+, and ¢=c+y, their product will be a’b'c' = abe + aby + ach + bea, if we neglect terms containing each more than one error as a factor. The absolute error of the product will then be, very nearly, a’b'c' — abc = aby + acB + bea, (9) and if in practice we substitute in the right hand member of this equation superior limits of the ap- proximate quantities, we may evidently take for the limit of absolute error of the product of three approx- imate factors the sum of the products obtained by multiplying the error of each factor by the product of superior limits of the other two factors. The meth- od may evidently be extended to any number of fac- tors. Hence, we may take for the limit of the abso- lute error of the product of any number of approxt- mate factors the sum of the products obtained by mul- tiplying the absolute error of each factor by the con- tinued product of superior limits of all the other fac- tors. If any of the products ady, etc., would be of a much lower order than the highest, such may evi- dently be disregarded. 24. The formula, a/b‘ — ab =a + ba, assumes that the product of the approximate factors, a’d’, is to be exactly formed. If the product is made by MULTIPLICATION. 33 abridged multiplication the errors of the process must be allowed for. But it is always easy to reduce the error resulting from the process of abridged multipli- cation to very much less than that due to the errors of approximate factors. PROBLEM 11. To determine the necessary arrangement of approximate factors for abridged multiplication, so that the er- ror due to the abridged process shall be less than that due to the error of either factor. RULE 11. When the absolute error of the factor having the greater relative error equals or exceeds § of its lowest units, take either factor for the multt- plier, reverse the order of its figures, and place tt un- der the multiplicand as far to the right as possible without having any of the significant figures of the multiplier to the right of the multiplicand or any sig- nificant figures of the multiplicand to the left of: the multiplier. When the error of the factor having the greater relative error ts between ¥ and § of its lowest untts, put the reversed multiplier one place further to the right. The reason for the rule will be clear from an exam- ple or two. We shall hereafter indicate the limit of absolute error of approximate factors by writing it in a special style of figure, following the factor, with the plus or minus sign, this limit of error being under- stood to be so many units of the lowest order retained in the factor. In reversing the multiplier its limit of error will come at the left. Where simply the plus . or minus sign follows a number the error of the num- ber will be understood to be not greater than half a unit of the lowest order retained inthe number. ‘The approximate factors of Example 18 will then be writ- ten, 784.2817 + 4 and 3.483 + 6. If we form their 3 34 APPROXIMATE COMPUTATIONS. product by arranging them according to Rule II, we shall have 297-21 Oier. SO By this operation it is evident that the errors of the first and second partial products due to the error of the multiplicand amount to about 47 units of the low- est order in the result, while the errors due to the neglected parts of the last four partial products do not exceed 2 of these units. And it is evident that if the right hand figure of the reversed multiplier had been the smallest possible, viz.: 1, and the error of the mul- tiplicand as large as 5 of its lowest units, the error of the result due to that of the multiplicand would have been at least 5 of the lowest units of the result; which would allow for Jo partial products besides the first, with errors all in the same direction, before the sum of the errors of the abridged process would equal that due to the error of the multiplicand. The result would not be essentially different if the former multi- plicand were made the multiplier. For by Rule 11 we should have VS4.201 7 gy 6 +3843 23523 + 3137 + 6257 vt- a 4h WW Co 24 MULTIPLICATION. 35 and here the possible error of the last partial product is nearly all due to that of the same factor which caused the greatest part of the error in the former operation. But let us look at one more example. Take the factors 1124.267543 + 2, and 8425.7987 + 2. If we were to make their product by arranging them like this, SAS JOO7 icy Brea eAs7O2 12 Tl it is evident that the amount of the error of their prod- uct due to the errors of the factors would be only about 2 of the lowest units obtained, whereas the errors of the product arising from the abridged process would be those due to the parts rejected from eight partial products; the limit of the errors arising from the abridged process could not then be placed in ad- vance at less than 4 of the lowest units of the prod- uct. To reduce the errors due to the abridged process of multiplication to less than those due to the errors of the factors, we need therefore, in accordance with Rule 11, to move the multiplier one place to the right. And this will be amply sufficient. For if we arrange the factors thus, BAOn 700 7 Bash 70 242 11 it is evident that the error of the product due to those of the factors will be about 24 of the lowest units of the product, while that due to the neglected parts of partial products will be less than 5 of the same units. In this example the highest significant figure of the multiplier is the smallest possible, which reduces the part of the error of the product due to that of the 36 APPROXIMATE COMPUTATIONS. multiplicand to a relatively small amount. And it is clear that if the limit of absolute error of the factor having the greater relative error were but half a unit of its lowest order we might add a zero to the factor, and call the error 5 units of the lowest order then re- tained, which would bring the case under the first part of Rule 11, already ioseand: If the relative errors of two factors are the same, or nearly the same, either factor may for the purposes of the rule be as- sumed as having the greater relative error. The rule is sufficient in all cases in which not more than ten partial products have errors in one direction. EXAMPLE 19. Multiply 3.14278 + by 0.00742 +, and let the error from the abridged process be small in comparison with the limit of error of the product due to that of either factor. Annexing a zero to each factor as just suggested, and then arranging by Rule 11, we have 3.14278 0 =e 5 5 + 0 24700.0 21999 + 1257 + 630-— On-biz6 G.0233 10) try The error due to the abridged process is less than a tenth of the possible error of the pages from that of the multiplier. EXAMPLE 20. Multiply 4.725 + 3 byo.1478 + 7 and assign a limit of error to the result. 25. PROBLEM 12. Two or more factors being given, each to a certain degree of approximation, it is required to assign a superior limit to the relative error of their product. MULTIPLICATION. : 37 RULE 12. Jake for the limit of the relative error of the product, the sum of the superior limits of the relative errors of the factors. Demonstration. We have for two factors the very nearly exact relation, equation (8), a’! —ab= apt ba. Dividing both members by ad we obtain TH ore MIPS ab eed: The first member of this equation is by definition the relative error of the product a’d’, while the second member is the sum of the relative errors of the factors a’ and 0’; and since by the rule we take instead of these latter errors superior limits of them, it is evident that Rule 12 will give a safe limit of the error of a product of two factors. It follows then that we may take for the limit of the relative error of the product of any number of approximate factors, the sum of superior limits of the relative errors of the factors. For if the product of two factors be regarded as a single factor, the limit of its relative error plus that of a third factor may be taken for the limit of the relative error of the product of the first product by the third factor, and so on. Rule 12 assumes, like Rule 10, that the product of the factors is to be exactly formed. If made by the abridged process the new errors introduced must be allowed for. But it has been shown that the additional absolute error introduced by the abridged process may easily be made much less than that due to the errors of the factors; hence the additional relative error due to the abridged process may also be cor- respondingly reduced. By placing the multiplier one aes a 38 APPROXIMATE COMPUTATIONS. or two places farther to the right than required by Rule 11 the error from the abridged process will be quite insignificant compared with that due to the errors of the factors. By indicating the limit of error of each partial product we may also if we please determine the limit of error of the actual result in- dependently of any rule. EXAMPLE 21. Determine by Rule 12 the limit of relative error in the product (65.432 + 2) (6.21242 )2).(1.5632 eae) and then compute the product. The sum of the limits of the relative errors may evidently be taken at 2 2 2 ES Sone sn a SS 60000 60000 15000 60000 5000 Arranging the first two factors by Rule 11 we have OS Ase iets pee wiles ees) herd (8 202150210 4 120 12006 Ae ee 6543 + Se LS hoes PA CGA Pa wad 406.489 8 + 256 Call this result 406.490 + 26, and make the next product as follows: 406.490 + 26 BMS ot US Stes 406.490 + 26 203, 2Ai5 see ag 24.329 0) 2b aie 1219 — Sul shea 635.424 + 124 MULTIPLICA TION. 39 The limit of absolute error of the result, as indicated by the limits of error of the partial products, is 124 of the lowest units of the result; hence the relative error of this result cannot exceed ¢3h44s5, which is a trifle less than the limit determined above, viz.: s5\ss. EXAMPLE 22. Determine by Rules 10 and 12 the absolute and relative errors to be expected in the product of the following approximate factors, and then form the product and determine its error by in- spection, as illustrated in example 21: (756.32 + 3) (25.41 = 7) (0.3248 + 2). 26. It follows from Rule 10 or equation (8) that the limit of the absolute error of the square of an ap- proximate number may be taken equal to twice a su- perior limit of the number, multiplied by the absolute error of the number. For if in equation (8) we make bo’ =a' and 6 =a we have a’? —a = 2aa. (10) And it follows from this that if the limit of absolute error of an approximate number be a unit of the th order counting down from the highest significant fig- ure inclusive, the absolute error of the square of the number will not exceed a unit of the (7 — 1)th order, counting down in the square also from its highest sig- nificant figure inclusive. For, supposing, as we may for the present purpose, the number a’ to be entire, and to contain z figures and be correct within a sim- ple unit, we shall have a= 1; anda’? will have either 2m or 2~—1 figures. But if a’ has only 2% —1 figures, the highest significant figure of a’ must be less than 4, in which case 2aqa, or the absolute error of the square, will have only z figures. Hence, if 2aa be placed under a’? so that the lowest units of 40 APPROXIMATE COMPUTATIONS. the two come in the same vertical line, there will be z—1 figures of a’? at the left of 2aa. Thus, sup- pose a! = 31572 andaw=1. We have a’? = 996791184 24a = 63144 Here, z being five, the error of the square is less than a unit of the fourth order, counting from the left in the value of a’”, If the first figure of a’ had been 5 or greater, 22a would have had one more figure, but so also would a@’”, hence the same statement would hold; and the truth of our proposition is evi- dent. And since, if @ remain constant, 2aa@ will be proportional to a, we may also state the principle that if an approximate number be exact within “alfa unit of the zth order from its highest significant figure, the square of the number will be exact within half a unit of the (z — 1)th order from its highest sig- nificant figure. It is evident that the principles of this paragraph assume that the square of the approx- imate number is to be precisely formed. If it be made by abridged multiplication, care must be taken that the errors due to the abridged process are made insignificant in comparison with 2@q@; and this is in practice always easy. The principle stated at the head of this article, or equation (10), will generally give a somewhat smaller limit of error than the other principles of this article, though it is convenient to remember that we may expect in the square of an approximate number as many exact figures less one, as there are in the number itself. EXAMPLE 23. Compute (0.0080715 + 7)? by abridged multiplication, and determine the limit of error of the result. 2%@. From the approximate value of the absolute error of the product of three factors, equation (9), MULTIPLICATION. 41 Art. 23, by making the factors and their errors equal we obtain a’ 3— g? = 3a°a (11) and if in practical examples we take in the right hand member of this equation, instead of a, a superior limit of a’, we may evidently assume as the limit of absolute error of the cube of an approximate number, the absolute error of the number, multiplied by three times the square of the superior limit of the number. It is hardly necessary to say that if the cube is form- ed by abridged multiplication the new error intro- duced must be added, or else made insignificant. EXAMPLE 24. Determine the limit of absolute er- ror of (3.456 + 2)° and (883.4 + 4)*, and perform the multiplications. For the absolute error of the first of these cubes we may take 3a7a < 3(3.5) (3.5) (0.002) < 0.075. 28. By dividing both members of equation (10) by a’, and those of equation (11) by a? we have the approximate equations | ae PE (12) a a ——{— = 38 (13) Jrom which tt ts evident that we may take for the limits of the relative errors of the square and cube of an approximate number respectively twice and three times the limit of the relative error of the number, a result evidently in harmony also with Rule 12. EXAMPLE 25. Apply the principle just stated, to determine the limits of the relative errors of the re- 42 APPROXIMATE COMPUTATIONS, sults of Examples 23 and 24, and from the relative errors thus found deduce also the absolute errors. For. the relative error of the first cubein Example 24 we have 3° ee , and since the cube will not a 3400 exceed 42 we may take for the limit of absolute er- 6 X 42, 3400 the Iimit determined in Example 24. ror, which is the least trifle less than 0.075, It evidently follows from Rule 12, that we may take for the limit of the relative error of the nth power of an approximate number, n times a limit of the rela- tive error of the number. If 2 were extremely large, however, we should not assume the limit of the rela- tive error of the number at its smallest possible value. mining the order of units of the highest signif e ota quotient. oe RULE 13. Observe how many places to We right or left the decimal point of the dividend would have to be moved in order that the first figure of the quotient should be simple units. If the point would have to be moved n places towards the right the first signifi- cant figure of the true quotient wrll be in the nth dec- imal place. If the point would have to be moved n places towards the left the first figure of the true guotient will be n places above simple units, that ts, the true quotient will have n+ 1 figures at the left of the decimal point. The rule may be illustrated by an example or two. In what follows we shall arrange quantities in division as we arrange them in algebra, viz.: with the divisor at the right ‘of the dividend, and the quotient below the divisor. Suppose then we have for division 0.1941690| _763.05436 os 0.0002 It is plain that if the decimal point of the dividend were moved four places to the right the first figure of the quotient would be 2 simple units. If we correct that quotient by moving the decimal point back four 44 APPROXIMATE COMPUTATIONS. places to the left, the 2 will evidently be brought into the fourth decimal place. If the dividend and divi- sor were as follows: 7428.436 0.04269 ie I 74008.8 it is clear that if the decimal point of the dividend were moved five places to the left the first figure of the quotient would be simple units. If we correct that quotient by moving its decimal point five places to the right the true quotient will evidently have six figures at the left of the decimal point, the results in each of these illustrations agreeing with the rule. 20. PROBLEM 14. A dividend and divisor being proposed, whose values may be taken as accurately as we please, it is re- quired to form their quotient so that its absolute error shall not exceed a unit of the zth order from the decimal point. RULE 14. Begin by determining the first signifi- cant figure of the quotient, and the order of its untts. Then count the number of figures that the quotient must contain, from this figure inclusive down to the nth inclusive. Beginning at the left, assume in the divisor a number of significant figures greater by one than the number just counted, and take at the left of the dividend as many significant figures as there are in the product of this assumed divisor by the first sig- nificant figure of the quotient. Subtract this product Srom the assumed dividend, and instead of annexing any figure to the remainder, reject a figure at the right of the divisor to determine the second figure of the guotient. Continue the process of division by throw- | ing off successively the figures of the divisor, until the quotient contains one figure below the nth order. In multiplying the portion of the divisor each time re- tained, by the successive quotient figures, have regard DIVISION. 45 to the rejected part of the divisor so far as to make each partial product tf possible exact within half of one of its lowest units. Whether the extra figure of the quotient can be discarded without passing the assigned limit of error must be determined by the special conditions of each example. EXAMPLE 26. Compute the expression 194 16.9063468085... 763.05 403678956... with an absolute error in the quotient less than a unit of the 6th decimal place. We see that the decimal point in the dividend would have to be moved one place to the left for the first figure of the quotient to be simple units, hence there will be two entire figures in the quotient. Therefore we need nine figures of the divisor and ten of the dividend. We take then 19416.90635|76 3.054037 15261 08074|25 4 463058 4155 82561 3815 27018 349 55543 305 22161 35 33382 30 52216 4 81166 4 57832 23334 22892 442 382 60 46 APPROXIMATE COMPUTATIONS. We mark with a dot each figure when it is rejected from the divisor. Let us consider the possible errors introduced in this example. The error of the divi- dend employed is less than half a unit of the lowest order retained in it; the error of the assumed divisor being also less than half a unit of its lower order, the error of the first partial product cannot exceed a unit of the lowest order in that, since the first quotient figure, by which we multiply, is 2. If the direction of the error of the divisor were not known, the limit of error of the second partial product would also bé a unit of its lowest order, since it would be doubtful whether we ought to carry a 3 or a 4 to its lowest figure, from the product of the rejected figure 7 of the divisor by the second figure of the quotient. As for the remaining partial products, none of their er- rors can be more than half a unit of their lowest order. But the remainder, 60, by which the 7th dec- imal of the quotient is determined, cannot be in error by more than the sum of the errors of the dividend and the partial products preceding this remainder, that is, > +1+1+3= 5. If the true remainder at the foot ought then to be 55 instead of 60, the 7th decimal of the quotient would be 7, but should be followed by other figures; but if the true remainder ought to be 65, the 7th decimal of the quotient would still be 8, though followed by other figures. There- fore the 7th decimal of the quotient actually obtain- ed cannot be wrong by more than a unit of its own order. Hence, a fortzori, the error of the quotient is less than a unit of the 6th decimal place, as required. In fact, from the limit of error now determined, we may reject the 7th decimal of the quotient, either in- creasing the previous one or not, without passing the assigned limit of error. DIVISION. 4 7 31. It is plain that if the left hand significant fig- ure of a divisor be the smallest possible, viz: 1, and it be followed by one or more zeros, then when the division has been continued until all the figures of the divisor but the 1 have been rejected, and we are ready to find the next figure of the quotient, an er- ror in the remainder to be used in determining this quotient figure, equal to 7 units of the order of the remainder, will cause an error in the quotient also equal to 7 units of the order of the quotient figure to be found. But by the application of Rule 14 this quotient figure will be of the order next below the mth. In the most unfavorable case possible, the first figure of the quotient will be the largest possible, viz: 9; for then the limit of error of the first partial prod- uct may be 45 units of its order, that of the second partial product 1 unit of the same order. If we add to these the possible error of the dividend, } a unit of the same order, we have still room for 8 more par- tial products with errors of 3 a unit each, all in the same direction, before the error of a remainder shall equal 10 units of its lowest order. Hence, if Rule 14 is followed, the error of the quotient cannot exceed IO units of the order next below the zth, or I unit of the zth order, except there should be more than 10 partial products having errors in one and the same direction, a case evidently not often likely to occur in practice. And it is clear that if the figure of the divisor following the last one to be used by the rule is known correctly, we may reduce the error of the first partial product to less than a unit of its lowest order; and then, if the number of partial products is less than 10, the error of a quotient found by the rule will in such cases be less than Aa/f a unit of the zth order; so that if we wish to reject the 48 APPROXIMATE COMPUTATIONS. extra figure of the quotient we can if necessary ad just the previous figure so that the error of the quo- tient shall still be less than a unit of the zth order. It is not necessary to state a special rule for fixing the number of decimals to be computed in a pro- posed dividend and divisor so that a quotient shall be correct within a unit of an assigned order; for the application of Rule 14 is sufficient to determine the required number. EXAMPLE 27. Compute the expression ——, with a relative error less than zodu07-. 2 The quotient will be greater than 2, hence, by Rule 4, we may make an absolute error = 0.0002. We shall be safe then if we apply Rule 14 as if to find a quotient with a limit of absolute error equal to a unit of the 4th decimal place. We need then, by Rule 14, 6 figures each in the divisor and dividend. The division will be as follows: 2 82842|> 22145 31317 28284 3033 2828 205 I41 64 mes 7 There are only 5 partial products preceding the last remainder, and the error of this cannot exceed 4 DIEISLON. 49 units of the same order, hence the error of the quo- tient must be less than 3 units of the order of its fig- ure 5. If we reject the 5, the answer will be 2.2214, with an absolute error less than 0.0001, and a rela- tive error less than s9$095, and, @ fortzorz, less than 1 T0000: 2 ; , with EXAMPLE 28. Compute the expression a relative error in the quotient less than yodo07 - EXAMPLE 29. Compute the expression 0.54674321 =a V 0.0003 with a relative error in the quotient less than ys'o0 ° 32. PROBLEM 15. A dividend and divisor being given, each to a certain degree of approximation, it is required to assign a superior limit to the absolute error of their quotient. RULE 15. Multiply a superior limit of the quotient by the relative error of the divisor, divide the absolute error of the dividend by an inferior limit of the dt- visor, and take the sum of the two results for the re- guired limit. Demonstration. It is evident that the most unfa- vorable case, or that in which the error of the quo- tient will be the greatest, is that in which the errors of the dividend and divisor are in opposite directions; for example, the dividend too large and the divisor too small. Suppose then that @ and 6 represent an exact dividend and divisor, a’ and 0 the correspond- ing approximate dividend and divisor, and a and 8 their absolute errors. In the most unfavorable case we have a ata F 38 50 APPROXIMATE COMPUTATIONS. / The absolute error of the quotient s would then be ata a. a+rba a ap © b—-f b- io — p) ae batap_a@i pia fiat ae Gee (15) But Rule 15 will evidently give a superior limit of the last member of this equation. Hence it is a safe rule. EXAMPLE 30. Determine the limit of absolute er- 4237.5 = 5 85.846 + 4 By Rule 15 we may evidently take 4500 0.004 4 O51 5 - LAs 80 80 80 160 EXAMPLE 31. Assign limits to the absolute errors of the following indicated quotients: 20eeA 2 et 0 0.47328 + 7 5 334-725. + 3, 8874 s 0432 0 ae 784.3 = 4 In the second of these expressions the error of the de- nominator is supposed to be 0, and in the last the error of the numerator is supposed to be o. 33. It is assumed in Rule 15 that the division of the approximate numbers will be exactly made, or, at least, that if the abridged process of division is em- ror of the quotient <10508 DIVISION. 61 ployed, care will be taken that the errors arising from the contractions of the process shall be very small in comparison with those due to the errors of the divi- dend and divisor. PROBLEM 16. A dividend and divisor being given,each to a certain degree of approximation, it is required to state a rule for dividing by the abridged process, so that the error of the quotient due to this process shall be less than that due to the errors of the quantities. RULE 16. Begin the division as usual. Then, if the sum of the errors of the dividend and of the first partial product formed does not exceed as many units of the lowest order employed in the dividend as half the number of partial products to follow, add places to the atvidend and divisor until tt does. EXAMPLE 32. Calculate by the abridged process the quotient 4.35278 + 3.14159 +’ and let the error from the abridged process be less than that due to the limits of error of the quantities. ole «Jee 6 4.352780 + 5|3.141590 + 5 3141590 + 5]1.385534 + 5 I 211190 042477 2 2 268713 2oT Soe 17386 15708 — 1678 1571 — 107 AC ae Dje a7 2 APPROXIMATE COMPUTATIONS. By thus adding a zero to the dividend and divisor we make the limit of error of the dividend and first par- tial product, due to the limits of error of the quanti- ties, 10 units of the lowest order then in the dividend, while the only partial products which are in error from the abridged process are the last four, and the sum of the errors of these does not exceed 2 units of the same order as before, hence the error of the quotient due to the use of the abridged process is much less than that due to the limits of error of the quantities. 34. We have illustrated in the above example a method by which the limit of the absolute error of an approximate quotient may be determined by an in- spection of the work. By indicating the error of the dividend and of each partial product, it is evident that the sum of all these errors cannot exceed 14 of the lowest order of units in the dividend actually employed. Therefore, since the first figure of the divisor is 3, the error of the quotient cannot exceed 5 units of the order of the figure 4 of the quotient, determined by using the last remainder, 13. EXAMPLE 33. Perform the division of Example 30, having regard to Rule 16, and determine the limit of error by inspection. 423795) 4:48 §OAG Cee 34338 + 2/4936 £2 803 7 772, O thers a Tat 258 55 ic. o DIVISION. 53 It will be noticed that after finding what the last figure of the quotient is to be we do not write the corresponding partial product; but we merely con- sider what the quotient figure in the place just found would be if the remainder by which it is determined were diminished or increased by the possible error of that remainder. Thus in the above example, if the remainder 53 were diminished by 8, the 4th quotient figure would be 5 instead of 6; and if the remainder 53 were increased by 8, the 4th quotient figure would be 7; but without a closer determination of the limits of error of the partial products we should have to assume that the 7 might be followed by other figures. Therefore if we take 6 for the fourth figure of the quotient, we place the limit of error at 2 units of its order, so as not to understate the error, though in reality we can see that the error could not be more than a trifle over I unit of that order; which is the limit determined by Example 30, supposing the divi- sion were to be made exactly. EXAMPLE 34. Calculate the expression 7 2.74384 £7 having regard to Rule 16. EXAMPLE 35. Perform the divisions of Example 31, having regard to Rule 16, and determine the limits of error by inspection; then compare these limits with those found by Rule 15. 35. PROBLEM 17. A dividend and divisor being given, each to a certain degree of approximation, it is required to assign a superior limit to the relative error of their quotient. RULE 17. Take for the limit of the relative error of the quotient, the sum of the superior limits of the relative errors of the dividend and divisor. 54 APPROXIMATE COMPUTATIONS. Demonstration. In the exact expression, equation / (15), for the absolute error of at viZen 7 E+ Swe shall make but a very slight error if we substitute 0 for 6’. Making the substitution, and dividing by the expression for the true quotient, we have the relative a’ ‘ error of Pp very nearly, NA Rape 14 BOTS _ Bye a Takg Wan (16) b Rule 17 will give a superior limit of the right hand member of this equation. Hence it is a safe rule. 36. It is assumed in Rule 17, as in Rule 15, that the division of the approximate numbers is to be ex- actly performed, or that the additional error from the abridged process is to be very small in comparison with that due to the error of the dividend or divisor. But we have shown in Rule 16 how to make the er- ror from the abridged process less than that due to the errors of the quantities; and it is evident that we may in any case make the error of the abridged proc- ess quite insignificant by adding to the dividend and divisor one or two more places than required by Rule 16. EXAMPLE 36. Determine by Rule 17 the limit of relative error of WRA2E ocd 10.3864 £9 : and from the relative error and a limit of the quo- DIVISION. 55 tient find, by Rule 2, a limit to the absolute error. Then make the division by Rule 16, and determine the error by inspection. EXAMPLE 37. Apply the same process to the quantities in examples 30 and 31. 3%. Besides the general abridged process of divi- -sion already explained, there is a special formula which may sometimes be used with advantage when the divisor is but little greater or less than unity. Sup- pose it to be required to compute the quotient ’ I @ being any number, and 7x a small fraction. Find- ing two terms of the quotient by algebraic division, and adding the indicated quotient of the remainder, we may write a ax ia opm Qa iy Q( laws) 1 ent If we neglect the last term of this expression, and take a(1—~*) for the quotient we shall make an abso- 2 ax é: lute error equal to : The corresponding rela- tive error will be ax’ 10a eee Mulan are Dad a I+2 Now when ~# is very small, x? will be very much smaller, and in such cases we may take a(1—-+) as : a : the approximate value of eee By doing so we sub- wer stitute a multiplication for a division; and we may if we please easily determine the limit of absolute error 56 APPROXIMATE COMPUTATIONS. committed, knowing that the approximate result is too small, and its relative error equal to 2”. EXAMPLE 38. Compute by the above method the expression y-y¢77, and determine the limit of error committed. For y-a677 We substitute 2(0.9923) = 1.9846. The relative error of this being (0.0077)? < 0.00006, the absolute error cannot be greater than 2(0.00006) = 0.00012, (Rule 2); that is, the result 1.9846 is toc small by a little more than a unit of the 4th decima! place. By treating the fraction in the same way as we a treated ——_, we may also deduce itz y ax me ew naa If we take aA a(i+-z), we make an absolute er- ax ; ror equal to : and a relative error equal to —r ax’ | Broa — 2 a eee = this error being in the same direction as before, and having the same expression. EXAMPLE 39. Divide 2 by 0.9923. We substitute 2(1.0077) = 2.0154; the limit of the relative error of this result being, as in the last exam- ple, 0.00006, and the absolute error in fact < 0 00012, though if the relative error were quite up to 0.0006 DIVISION. 57 the absolute error would slightly exceed 0.00012, since the quotient is greater than 2. It is evident then that where x is a small fraction, . a : \ in place of -—— we may substitute a(1—-~), and for I++ yay We may substitute a(1+-), and the results will in each case be a little too small, the relative error being +’, and the absolute error being the quotient multiplied by x”. EXAMPLE 40. At a temperature, ¢ = 5° centi- grade, and a pressure of I atmosphere, the volume, V, of a mass of hydrogen being 1 litre, what will be its volume, ’,, at zero, under the same pressure ? We have from Physics the formula V Ke 1+74(0.0036613) yo = I e I ° 1+5(0.0036613) | 1+0.0183065. Say V, = 1X(1—0.0183065) = 0.9817 very nearly. The relative and absolute errors of this result being a little less than (0.02)? = 0.0004, we may be sure that the answer within a unit of the 4th decimal place is 0.9820. or EXAMPLE 41. Suppose the volume of the gas ata temperature of —5° to be 4 litres, what will be its volume at zero? In the formulas of example 40, make ¢= —5, and V=4 SQUARE ROOT. 38. PROBLEM 18. Any exact number being given, it is re- quired to find its square root by an abridged method, but so that the absolute error of the root shail not exceed a unit of the (2z+1)th order, counting to the right from the highest signifi- cant figure of the root inclusive. RULE 18. Employ the ordinary process of extract- ing the square root until n+ 1 significant figures of the root have been found. Form the next trial divisor as usual, and in forming the new dividend bring down only one new figure from the original number. finish the work by dividing this dividend by the trial divisor gust formed, contracting the latter one place at the right after cach quotient figure ts found, and placing the quotient figures in the root, until n addt- tional figures have been found, observing not to tn- crease the last one unless it would be followed by a gure at least as large as 7. EXAMPLE 42. Compute V10498.59325783, so that the result shall contain nine figures and be exact within a unit of the lowest order retained. We put 2x+1 =9, whence ~+1=5. We shall then find five figures by the ordinary process, and four by division. SQUARE ROOT. 69 10498.59325783 ( 102.462643 I 202 ) 0498 404 2044) 9459 81 76 20486 ) 12 8332 eet 2050 20492) 54165 40984 Lod 12295 886 820 66 61 5 After finding five figures of the root, the next trial divisor, 20492, is formed as usual; but since the next figure of the root is not to be annexed to this trial divisor before multiplying, we must evidently bring down but one new figure to the dividend. In bring- ing down this new figure, we do not increase it, what- ever it is followed by in the original number, for the contracted process tends to make the result a trifle in excess, as will be seen below. ‘The last four figures of the root are given by division. To show that the error of a square root found by Rule 18 is less than the limit assigned in the state- ment of the problem, we may proceed as follows: Since the process of extracting the square root of a number which is partly decimal does not essentially 60 APPROXIMATE COMPUTATIONS. differ from that in which the number is entire, we may confine the investigation of the principle to the ex- traction of the roots of whole numbers. These roots, however, may be entire, or they may be partly iec- imal. Suppose the square root of an entire number to be separated into two parts, the part at the right being made to contain z figures besides the decimal figures, and the part at the left at least 2-+ 1 figures. De- note the relative value of the left hand part by a, and that of the remaining part by &. (By the relative value of the left hand part we mean the value of the z+ 1 figures with z zeros added). If WV be the number whose root we are considering we shall then have, N= (at bv =a + 2ab + &. (17) If we employ the ordinary process of extracting the square root of the number J, until we have found the part a, and after forming as usual the next divi- dend bring down to this the remaining periods of the number, this dividend will evidently be equal to N—a’. Finding the value of V — a from equation (17) and denoting it by R, we have R= 2ab+ &, Ie b whence Oe ree es (18) 22. OR Let us determine a superior limit of the value of 2 this last term, ae Since & has but xz figures above a the decimal point, 0’ cannot have more than 2n figures above that point, whereas a, with its relative value, must have at least 2n + 1 places above the decimal point; and 2a@ must hence be more than twice as ee great as 0. Therefore the value of ae less than atl SQUARE ROOT. 61 half a simple unit. It is thus evident from equation (18) that if the dividend R were exactly divided by 2a, and the quotient taken for the part 4 of the square root, the result would be too large, but by less than half a untt. In employing the abridged process of division for finding the value of the quotient = the only addi- tional error introduced is that due to the rejection of some figures from a few partial products. 4.6. To take a defi- nite case, Suppose an entire number, a’, to contain 6 figures, and its error to be a simple unit. Then, since the cube root a would be greater than 46, we have 3a” > 6000; hence, a < gq4gy < 0.0002. But the root a will have 2 figures besides the decimals ; hence its error will be less than 2 units of the 6th order from its highest figure. And if the first figure of the root here were as large as 6, we should have 3a° > 10000, and a would therefore be less than 0.0001. Jt follows, then, that when an approximate number contains n significant figures, and ts exact within a unit of its lowest order, we may find the cube root with an error not exceeding a unit of the order of its nth significant figure, except when the first significant figure of the root ts 4 or 5, and in those cases the limit of error need not exceed 2 units of the same order. In the majority of cases we may find as many exact significant figures in the cube root of CUBE ROOT. 79 an approximate number as there are in the number. (Compare Example 58, &c.) EXAMPLE 56. Determine, by Rule 22, the limit of absolute error of 4/0.000136425+7, and then find the root by the abridged process, making the error of the process, however, not more than a tenth of the limit so determined. Do the same with 4/365.046+. 48, PrRopLEM 23. Any approximate number being given, with a relative error not exceeding an assigned limit, it is required to assign a limit to the relative error of the cube root of the number. RULE 23. Take for the limit of the relative error of the cube root, one-third the limit of the relative _ error of the number. Demonstration. Dividing both members of equa- tion (22) by a, we have eh en lien o — 2. meee a 3 3 a*—a,. The factor meyer being here the relative error of the approximate number a”, the limit of the relative error of the cube root a’ may evidently be taken equal to one-third that of the number, provided the root of the approximate number be precisely found, or the new error in finding it be made insignificant. Rule 23 evidently agrees with the principles of Art. 28. And it is clear that we may extend the method to aroot of any index, avd take for the limit of the relative error of the nth root of an approxt- mate number, the nth part of the limit of the relative error of the number. EXAMPLE 57. Assign a limit to the relative error of 4/842.731-.6, and from this limit find, by Rule 2,a limit to the absolute error of the result. Also find 80 APPROXIMATE COMPUTATIONS. this latter limit by Rule 22, and compare the two an- swers. 49. Byamethod similar to that of Art. 43, we may determine the possible error ofa cube root by in- specting the possible error of the last remainder. EXAMPLE 58. Compute 4/1272.4386+7 as closely as the limit of error allows. Assuming that we can find eight figures in the root, we will find five before beginning the contracted division. 1272.4386+7 (10.836248 + 7 I 30000 272438 308 2464) 32404 | 2bort2 16 64|3499200 12726600 3243 9729|3508929 | 10526787 6 9|351866700 | 2199813000 32496 194976|352061676 | 2112370056 36.35225 0688 | 87442,944 Bibra e! <7: 16991 14090 + 2901 2818 + 83° 707 The final remainder, 83, is liable to error of 700 units of its lowest order, from the original error of the num- ber. Allowing 3 units of the same order for the re- jected parts of the partial products, and adding the 83 to its possible error, 703, the quotient of the sum by 352, the last divisor employed, will evidently be but a trifle over 2 units of the same order as the last figure 8 of the root, found by the same divisor. The error due to neglecting the part of the divisor CUBE ROOT. SI 2 3 corresponding to ao = is, by Rule 21, less thana tenth of a unit of the same order; hence the total error of the root can be but little over two units of that order. We call the limit 3, so as not to under- state it. The limit, if found by Rule 22, would be just about 2 of these units. EXAMPLE 59. Compute 4/752.43275438 as closely as the limit of error allows, and determine the limit by inspection. * 43 LOGARITHMS AND TRIGONOMETRIC FUNCTIONS. $0. Ifwe examine a table of common logarithms, we shall notice that the greatest difference between the logarithms of any two consecutive numbers, of figures each, is less than 5 units of the order of the ath decimal of the logarithm, and that the least dif- ference of any two such logarithms is about half of one of these units. For example, the logarithms of 1000 and 1001 are respectively 3.000000 and 3.000434, while the logarithms of 9998 and 9999 are 3.999913 and 3.999957. On the other hand, the greatest dif- ference between any two numbers corresponding to two logarithms that differ by a unit of the zth order of decimals, is less than 3 units of the order of the uth significant figure of the numbers, and the least difference of any two such numbers is about a quar- ter of a unit of the same order. For example, the numbers corresponding to the logarithms 3.999800 and 3.999900 are 9995.4 and 9997.7, while the num- bers corresponding to the logarithms 3.000000 and 3 OCCOTOO are 1000 and 1000.23. The limits of difference stated above do not take into account the slight inaccuracy which will some- times occur in finding by interpolation logarithms or numbers intermediate between those given directly in the tables. But from the examples it is evident TRIGONOMETRIC FUNCTIONS. 83 that in finding the logarithm of an approximate num- ber, an error in the number equal to a unit of the order of its zth significant figure will be liable to make the zth decimal of the logarithm uncertain ; and an error in a logarithm equal to a unit of the ath order of decimals will be liable to make the wth significant figure of the number found from the loga- rithm uncertain. Hence, if computations are made with the ordinary six-figure logarithms, the result cannot in general be depended on as accurate beyond the 5th significant figure, though if the data to begin with are exact the 6th significant figure of the result may not be far out of the way. But since it isclear from the examples given that the amount of uncertainty in the logarithm of an ap- proximate number consisting of z figures will vary very much according to the value of the highest sig- nificant figure of the number, it is not easy to givea general rule which will always determine the smallest obtainable limit of error in the result of a computa- tion of approximate quantities by means of loga- rithms. This limit is, however, easily determined in any special example by indicating in the work the possible error of each step of the computation, which may be readily ascertained by observing what change would occur in each logarithm taken from the table, for a change in the approximate number equal to its possible error, or what change would occur in a num- ber to be found from an approximate logarithm, for a change in this logarithm equal to its possible error. The same method is evidently applicable with the logarithms of trigonometric functions. EXAMPLE 60. Given one side of a triangle a= 3500+ 7 feet, and the adjacent angles, B = 65°30’+ 30” and C = 85°30'+30”, compute the side 6 by means of logarithms, and assign the limit of error. 84 APPROXIMATE COMPUTATIONS. We have the formula d = qsin B sin the angle 4 will be 29°+1’, and finding the logarithms, we have: log a .. (3500+7).. . 3-544068+ 725 log sin B (65°30’4307) . 9.959023+ 29 a.c.log sin A (29° + OL’). . 0.314429+ 229 log 6 (6569.3+5-9) . 3.817520+353 d=66 BIO. ae Se) We place at the right of each logarithm the change which the tables would give, if the quantity whose logarithm we are taking were changed by the amount of its possible error. In looking out the number for 4, corresponding to the sum of the logarithms, we divide the possible error of this sum, viz., 383, by the tabular. difference 66, giving the possible error of 0 a little less than six feet. Observing that EXAMPLE 61. Compute in the same way the side ¢ of the same triangle. 4. We may employ a like method in working examples by means of natural trigonometric functions. EXAMPLE 62. Find the side 4, of example 60, by means of natural functions, and assign its limit of error. We have, nat sin (65°30’+ 30”)=o-909964 7 Multiply by a reversed, 7£ 0053 2729088+ 21 454098+ 4 O+ QI 3184°86-L 776 TRIGONOMETRIC FUNCTIONS. 85 Divide by nat sin(29°+ 1’)=0-484814 26 3184:864776 | 048481426 2908864756 | 6569:346:0 27600 24241-- 16 3359 29094 2 450 290 436 14 When we reach the remainder 450, by which the units figure of the quotient is obtained, we see that the possible error of this remainder has amounted to 290 units of its lowest order. Dividing this possible error by the same portion of the divisor that was used in finding the units figure of the quotient, we Have, as the limit of error of the resujt, 6 feet, which only differs by a tenth of a foot from the limit of error when the computation was made by log- arithms, while the result itself is the same. It is thus seen that natural functions with five decimals give nearly the same precision as logarithmic func- tions with six decimals. EXAMPLE 63. Compute in the same way the side c of the same triangle, and compare with the result of example 61. COMPLEX COMPUTATIONS. 32. Having explained the elementary processes of approximate computations, we will now consider the casein which several of these processes are to be employed in a single problem. We will first take ex- amples in which the quantities proposed may be found as accurately as we please, and afterwards give a few in which some of the quantities are known with only a limited degree of accuracy. PROBLEM 24. Any complex monomial being proposed, whose value is required with an absolute error not exceeding an assigned limit, and whose factors may each be taken as ac- curately as we please, it is required to assign an allowable limit to the error of each factor, and to compute the value of the monomial. RULE 24. Make a rough calculation of a superior limut of the result, and from this determine, by Rule 3, the allowable retative error of the result. Then count the number of single factors whose values are ta be taken approximately, both in the numerator and denominator of the monomial, and add to the number of such factors the number of operations that are to be performed on them after their values are obtained divide the allowable relative error of the final result by the sum thus made, and the quotient will be an al- lowable relative error for each single factor. Com- pute each single factor until its error does not exceed this limit, and then perform the remaining operations an such a way that the new error introduced by each abridged process shall be very small compared with that due to the errors of the factors. Indicate, as the work proceeds, the limit of error of each partial result, and lastly of the final result. COMPLEX COMPUTATIONS. 87 EXAMPLE 64. Compute the value of Tee Vi0 xz — v 5.27963289.... 0.4318965021.... so that the absolute error shall not exceed 0.001. As approximation for assigning the allowable relative error of the result, we may substitute i raxV2 as 13X12 0.4 0.4 limit of the final result. Hence, the allowable rela- tive error of the result may be taken at 74,5. There are 5 single factors to be taken approximately, and there are 5 operations to perform on them after their values are found ; we therefore divide the allowable relative error of the result by 10, which gives the limit of the relative error for each factor, goto. By Rule 4 we may then assume the allowable absolute error of V2 at gyhg7=0. 000025, that of W1oat a305 —0.00005, and that of 2 at z5#y9=0.000075. Taking then V10 with four decimals, = 2.1544 + and multiplying by 7=3.1416—, we have 2.1544 + I—00141.3 O40 3202 rs 21544 + 8618 — BN ESO kee irre! ~~ 6.76826 + 16 Dividing this result by 5.27963 + <4; that is, 4 is a superior 88 APPROXIMATE COMPUTATIONS, 6.76826 + 16 |5.279°3 + 527903 + 1.28196 + 4 148863 105593 — 43270 42237 — 1033 528 — 505 StI 30 + 78 Extracting the square root of the quotient, 1.28196+4 (1.132244+3 I 21) 28 ue 223) 719 OGG 2262) 5060 4524 536 452 84t 47 Taking 4/2 with five decimals=1.25992+, and multiplying, T;1 3224; tee + 29952.1 1.132240 + 39 226448 + S001 25-- aan IOI9QO + IOIQ + 23 05 1.426532 + 44 COMPLEX COMPUTATIONS, 89 Finally, dividing this result by 0.4318965+, 1.426532 + 44 | 0.431896,5+ BAe RREY f | 3.3029 + 2 130842 129569 1273 864 409 389 20 +46 We may then take for the final result, with three exact decimals, 3.303+, the absolute error being then less than 0.0003, or less than 0.001, the assigned limit. The reason for Rule 24 may be given in a few words. If a given monomial were composed of any number of approximate factors, either for multiplication or division, and if there were to be no new error made by abridged processes of computation, we might take, for the allowable limit to the relative error of each factor, the allowable relative error of the final result divided by the number of such factors. But since the abridged processes are liable to produce further errors, we must allow for them in assigning the limit of error of the factors, which Rule 24 evidently does. But since, as the operation proceeds, the error in- creases, it would not even with this allowance be safe to let the new error, produced by each abridged pro- _cess, be equal to that due to the error of the two fac- tors on which we are operating. How much less than this the error of each operation ought to be made will depend on the degree of complexity of the expression which we have to compute. For all ordi- nary cases, if nine-tenths of the error of each partial result is due to the errors of the two factors which gO APPROXIMATE COMPUTATIONS. give it, there will evidently be a sufficient margin of safety. If in the statement of a problem the limit of relative error of the result be assigned in advance, we may evidently employ the method of Rule 24, except that we are saved the necessity of making the rough cal- culation of a limit of the result to begin with. EXAMPLE 65. Compute V857 x 4/9847.27 5 1 with a relative error in the result less than zp45>. There are 3 factors, and 2 operations after they are found, hence the allowable absolute error of 857 may be taken at 35,5 = 0.00058, that of 9847.27 at ay4°57 = 0.0004, and that of mat sa3a9 = 0.00000. Take, therefore, “857 with 3 exact decimals, 9847.27 with 4, and a also with 4; make the mul- tiplication so as to have 3 decimals in the product, and the division so as to have 2 decimals in the quo- tient, and the result will be 199.73 = 7. EXAMPLE 66. Compute within 0.0001 the radius of a circle whose area shall be equal to that of a regu- lar hexagon, one side of which equals I. If R be the radius, we may easily obtain the form- ala, R =y/3¥3, We see that the result will be less than 1, hence the allowable relative error of the result is zo94y,. And, since we may regard the relative error of a sq. root as equal to half that of the number, this expression not lowable relative error of the quantity under the large radical Further, since multiplying and dividing by \ COMPLEX COMPUTATIONS, OI exact factors does not alter the relative error, we need to allow for the error of only one operation on the two approximate factors, before taking the final square root. We have then for the allowable relative error of each of these factors zghoq. It is just as well to Wilt, Wee and we may then make the abso- atta. lute error of V27 equal to gy$oa, Say 0.0001, and that of 27 equal to sp$sy, Say 0.0002. Take, therefore, 4 decimals each in V 27 and 22, make the division so as to retain 4 exact decimals in the quotient, and extract the root so that the error of it shall be mostly due to that of the quantity. (Compare Art. 41). The answer, with 4 exact decimals, is 0.9093 +. EXAMPLE 67. Compute within 0.001 the radius of a sphere, whose volume shall be equal to that of a truncated pyramid, the altitude of which is 0.752, the bases being *regular hexagons, whose sides are re- spectively 1.42 and 0.843. The formula will be ae Ve * [0-732 (1-42)’+ (1.42) (0.843) + (0.843)? 2 7 It is not difficult to see that the result will be less than 1. Then the allowable relative error of the result will be zj4,. There are 3 factors that will be taken approximately, viz., WY 3, 7, and |(1-42)’+ (1°42) (0°843)+(0°843)’], and we allow for 3 opera- tions, besides those with the exact factors, so that the relative error of each of the approximate factors may be made 75,5. The factor just written in brack- ets will be greater than 3, hence its allowable abso- lute error will be 5 = 0°0005. The multiplications and additions to obtain this factor may then as well be made exactly, but only 3 exact decimals need 92 APPROXIMATE COMPUTATIONS, be retained in the value of it when found. The answer within 0.001 is R=0.839-+. EXAMPLE 68. Find within 0.0001 the radius of a sphere inscribed in a cone whose altitude is equal to the diameter of its base, and whose volume is equal loos fe AS, The formula will be en = V9 iis / 27 3 EXAMPLE 69. ‘The volume of a gas at a tempera- ture 7°, and pressure 7™, Being ”, the volumeme temperature 0°, and pressure 760" is given by the formula. x A. 760(1 + ¢ X 0:0036650) Supposing 7. = 1056.7..... , Af = 202-0) 150.1..., determine how many more decimals would have to be given in each of these quantities, so that V could be calculated with an absolute error not exceeding 0-002. (Proceed by finding the allowable limits of error of the factors, that of the result being 0.002. ) EXAMPLE 70. Compute within one cent the amount of eighty-nine dollars and thirty-seven cents at compound interest for three years and three hun- dred and forty-seven days, at the rate of seven and one-half per cent. per annum. Denoting the amount by A, we shall have Al i= 780'37 (1-075)"(1 TELE) | 305 os. If a polynomial be proposed, whose terms may be found as accurately-as we please, and if it is to be calculated with a relative error in the result not exceeding an assigned limit, then, before we can COMP EX COMPUTATIONS, 93 ¢ ssign the allowable limit to the error of each term, we have to determine a rough inferior limit of the final result, from which the allowable absolute error of this result may be found; and from this, as in ad- ‘ dition and subtraction, the allowable absolute error of each term, and then, if necessary, the allowable relative error of each term, and so that of each factor in each term. EXAMPLE 71. Compute the length of one side of a regular pentedecagon inscribed in a circle whose radius s I, with a relative error in the result less Chet a eS ‘ 000 The formula will be i 7V 1042 in 7¥3 (V5—1). We can see by a little trial that the first of these two terms will be greater than 0.8, and that the sec- ond will be less than 0.6; hence, the result will be greater than 0.2, which is, therefore, an inferior limit of the result. We may then, by rule 4, assume the allowable absolute error of the result at ;92, = 0.0002. The absolute error of each of the two terms may then be 0.0001. The allowable errors of the approximate factors may now be assigned, and the result com- puted. EXAMPLE 72. Compute the expression : 7 G3 ™ + wae ree an fe ean ev to) + 732—V10 4/2 V7 V6 a i) with a relative error in the at less than = et. In solving such a problem as this, it is well to ar- range the work on paper in such a way that after having carried the computation of each term far enough to serve for assigning an inferior limit of the result, from which to determine the allowable errors, 94 APPROXIMATE COMPUTATIONS. the computations may be resumed again at the same points, without having to repeat any of the work. 24. If it be required to compute as accurately as possible the value of a complex expression which contains factors whose values are only known with a limited degree of approximation, it is semetimes convenient to determine in advance the degree of precision which we may expect in the result ; while, in other cases, the form of the expression is such that the most convenient way is to proceed directly with the computation, taking care that the errors in- troduced by abridged processes are made small in comparison with those due to the errors of the quan- tities, and indicating the possible error of each partial result when found, and lastly that of the final re- sult. If a monomial contains a factor whose rela- tive error is very much greater than those of the other factors, it will generally be useless to retain in these latter all the figures that might be taken, al- though to obtain the result as closely as may be it is best to work with one or two redundant figures. Or, if, in a series of terms for addition or subtraction, one of them would have a much greater absolute error than the others, then it will, of course, avail little to compute these latter to a much lower order of units than can be found in the term having the greatest absolute error. EXAMPLE 73. An iron cylinder, weighing 6 kilo- grammes, is found to be lengthened 2 millimetres by passing from the temperature of melting ice to that of boiling water. Considering the specific gravity of iron, and the coefficient of dilatation, the radius ¢ the cylinder in decimetres. yay, be computed from the formula : ; 6. 7.8 X 16.920 X 7 COMPLEX COMPUTATIONS. ape fe Assuming that the factors 7.8 and 16.920 are liable to an absolute error of half a unit of the lowest order in each, determine what degree of relative and ab- solute approximation may be expected in the value of RX, and then make the computation. The error of the result will evidently be almost entirely due to that of the factor 7.8. If the com- putation were to be made with the greatest possible precision, we might then expect the limit of relative error of the result to be about 1. 7385 = shy. We can see that the quantity under the radical will not differ much from ;4;, hence the result will be about 0.12, and the absolute error, therefore, approxi- mately 442 = 0.0004. We can probably employ the abridged modes of computation without increasing the error to more than 0.0005, which will enable us to get the result within half a unit of thousandths place. Taking mw = 3.142—, the denominator may be found=414.6+4 30, the quotient of 6 by this = 0.01447 +7z and the square root of this = 0.120345. EXAMPLE 74. Determine the limit of relative and absolute error of the value of the following expres- sion, supposing the computations were to be exactly made; then make the computations by abridged processes, having regard to Rules 11 and 16, and de- termine by inspection or otherwise the limit of error of the result as obtained. (67.34547) (6 326147) (1.7253-42) (4.2785 47) (0.627344 /) Taking the sum of. limits. of. the relative errors of all the factors, we have THbo otsubc0 t ts000 t co000 + erbo0=ie0b00- But tetas < o c a a < mw So iJ] tu —_ = o = c a 511 CSSSSE AOR