THEORY OF COMMUNICATION* By D. GABOR, Dr, Ing., Associate Met..ber.t (The paper was first received 25th November, 1944, and in revised form 24th Seprember, 1945) PREFACE “The purpose of thete three studies is an inquiry into the essence ofthe “information” conveyed by channels of communication, and the application of the results of this inquiry to the practical probiem of optimum utilization of frequency bands. a Fart 1, a new method of analysing signals is presented in which time and frequency play symmetccal parts, and which contains “time analysis” and “Yrequency analysis” a3 {Special eases. It shown thatthe information conveyed by a frequency band in a given {tmesnterval ean be analysed in various ways into the same number of elementary "quanta of information,” each quantum conveying one numerical datum. ‘a Part 2, this method is applied (0 the apalysis Of hearing seosations. It is shown con the basis of existing experimental material that in the band between 60 and 1 000 c/s {he homan ear can discriminate very nearly every second datum of information, aod {hat this efiency of nearly 30% is independent ofthe duration of the signals ina remark ably vide interval, ‘Ths fat, which cannot be explained by any mechanism fa the toner fat, suggests « new phenomenon in nerve conduction. At frequencies above 1 000 c/s the cffcieney of discrimination falls off sharply, proving that sound reproductions which sare far from ffl may be perceived bythe eat as perfect, and that “condensed” methods ‘of transmission and reproduction with improved waveband economy are possible in principle. Tn Part 3, suggestion are discussed for compressed transmission and reproduction of speech of ausio, and the Hest experimental results obtained with one of these methods ae described Part 1, THE ANALYSIS OF INFORMATION ‘SUMMARY Hitherto communication theory was bated on two alternative ‘methods of signal analysis, One is the description of the signal as a function of time: the other is Fourier analysis, Both are idealizaions, 26 the first method operates with sharply defined instants of. time, the second with infinite wave-trains of rigorously defined frequencies. But our everyday experiences—especially our auditory sensations — insist on a description in terms of bork time and frequency. Io the ‘resent paper this point of view is developed in quantitative language Signals are represented in two dimensions, with time and frequency as co-ordinates, » Such two-dimensional representations can be called “information diagrams,” as areas in them are proportional to the umber of independent dats Which they can convey. This is 2 con- sequence of the fact that the frequency of a sigml which is not of infinite duration caa be defined only with a certain inaccuracy, which is inversely proportional to the durstiog, and vice versa. This “uncertainty Felation” suggests a new method of description, iter= mediate between the (wo extremes of-time analysis and spectral analysis. There are certain “elementary signals” which occupy the sshallest possible area in the information diagram. ‘They are hazinonic oscillations modulated by a “probability pulse.” Each elementary simnal can be considered a5 conveying exietly one datum, or one “quantum of information.” Any signal can be expanded in terms of these by a process which includes time anaiyss and Fourier analysis as extreme cases. ‘These new methods of analysis, which ivolve-Zoriof the inathe matical apparatus of quantum theory, are illustrated by application to some problems of transmission theory, ich a8 difect generation Of single sidebands, signals transmitted i minimum time through limited frequency chaatels, frequency modulation and time-division multiplex elephony, - + Radi Section pa 7 Hii om sa abun Go, ts, Rear Lona ( INTRODUCTION “The purpose of this study is to present a method, with some new features, for the analysis of information and its transmission by speech, telegraphy, telephony, radio or television. » While this first part deals mainly with the fundamentals, it will be followed by applications to practical problems, in partituar £0 the problem of the best utilization of frequency The proce tht the rasbasion ofa ean information per unit time requifes #Gertain, minima Wave- band width dawned gradually, upon ‘Sean eBbcers uring the third decade of this’enitudy- Silay! 88 she rine ple of conservation of eneray emerged from the slowly Ua conviction ofthe impossibility of a porperuim mobile, ‘mental principle of communication engiseering ah ‘efutation of iogenious attempts to break the af ye unformu- ated law, When in 1922 John Carson'-t disproveit/the claim that frequency modulation could economize some of the band- width required by amplitude-modulation methods, he added that all such schemes “are believed to invohe fundamental fallacy." This cooviction was soon cast into @ more solid shape when, in 1926, Nyquist}? and Kopfmller'3 independently 1.~ The cisoidal function i& to be ‘considered as a limiting case, #6 the theory is correctly applicable only'to signals of finite duration, and with frequency spectra which ‘do not extend to infinity, & condition which is fulfled by all rea, pbyical signals. ‘These definitions and rules enable us to formulate the un~ ‘certainty relation quantitatively. Let us consider a finite sigaal, such as is shown, for example, in Fig. 1.6. Let us first fx the mean epoch andthe mean frequency of the signal, by means of equations (1:12) and (1.13) or (18). . These, however, do ‘ot count as data, a5 in @ continuous transmission there will ‘be some signal strength at any iostant,-and at aay frequency. ‘We consider f and j'as references, not as data. ‘The fist two ata will be therefore determined by the mean-square values of epoch and frequency, i.e. __ [resae Pe 1 Cf. Cuspats and Form: Referens 1, 9.28 (1.18) Gui) GABOR: THEORY OF COMMUNICATION pe Bb u fegeee at Saat Spraar "aan ar ~ Oa He 4 120) “The second of these has been ft transiated into “ime fan- guage,” as explained, and transformed by partial integration to Dut ils extentialy positive character into evidence. Tt may be noted that /2 and 2, and in general all mean values ‘of even order, remain unaltered if the real signal (4) or its associate, o(f) is substituted in the place of yf) = 90 + jo. Heence in the following we could again use the real instead of the complex signal, but qb will be retained in order to simplify ‘some of the analytical expressions and to emphasize the similarity with the formalas of quantum mechanics. ‘We now define what will be called “the effective duration” At and the “effective frequency width” Af of a signal by the following equations Ate [mei . 12) * AfmpnG=fRt . «(122 In words, the effective duration is defined as (27) times the ana deviation of the signal fom the mean epoch 1, and tbe effective frequency width similarly as +/(2n) times the rms. deviation from f- The choice of the numerical factor V(2m) will be justified later. ‘Using the ents RP Fen -ur Ar and Af can be expressed by means of (1.19) and (1.20). The expressions are greatly simplified if the origin of the time scale is shifted to 7, and the origin of the frequency scale to /. Both transformations are effected by introducing a new time scale 23) and a new signal amplitude He) = Henri 24), Expressing ¢.and yp by the new quantities 7 and ¥, it is found that, apart from a numerical factor 2m, (AO? and (A/)? assume the same form as equations (1.19) and (1.20) for # and /2, Multiplying the two equations we obtain. ci ave a pyenpdr| 1 # cRear + 025) ape = rt But, by a mathematical identity, a form of the “Schwarz inequality” due to Weyl and Pauli,t the expression in brackets is always larger than unity for any function ¥ for which the integrals exist, We obtain, therefore, the uncertainty relation inthe rigorous form ° MAL >E os at, 029 ‘This is the mathematical identity which is at the root of the fundamental principle of communication. We see: that the rams. duration of a signal, and its r1n.s. frequency-width define a minimuin-area in the information diagram. How large we assume this minimum area depends oa the convention for the numerical factor. By choosing it as (Qn) = 2:506 we, have made the number of elementary areas in any large rectangular fen Hs The Tacaey a cchaies"(Mebyen, Landen, seh Ni ees FLD CE RRC Ce a ces Ms Roped or beeGAROR: THEORY OF COMMUNICATION ‘region of the information diagram equal to the number of independent data which that repion can transmit, according to the result obtained in Section 1. ‘Relation (1.26) is symmetrical in time and frequency. and it suggests that a new representation of signals might be found in which 1 and f played interchangeable parts. Moreover, it suggests that it might be possible to give a more concrete inter- pretation to the information diagram by dividing i up into “ells” of size: one half, and associating each cell with an “elementary signal” which transmitted. exactly one datum. of information. This programme will be carried out in the next Section, (9) THE ELEMENTARY SIGNAL ‘The mathematical developments up to this point have run rather closely on the lines of quantum mechanics. In fact our results could have been formally obtained by replacing a co- ‘ordinate x by ¢, the momentum p by f, and Planck's constant A by unity. But now the ways part, as questions arise in the theory of information which are rather different from those which quantum theory sets out to answer. ‘The first problem arises directly from the inequality (1.26). What's the shape of the signal for which the product A‘AS actually assumes the smallest possible value, ic. for which the inequality turns into an equality? ‘The derivation of this signal form is contained in Appendix 9.3; only the result will be given here, which is very simple, The signal which occupies the minimum area Ath f = 4 is the modu- lation product of a harmonic oscillation of any frequency with a pulse of the form of a probability function. x complex form YO = en cis On fet + $) 2n ‘a toy fe and ¢ are constants, which can be interpreted as the “sharpness”: of the pulse, the epoch of its peak, and the fre- ‘quency-and phase constant of the modulating oscillation. The ‘constant «cis connected with Ar and AF by the relations m1 1 ae /Q)z A= Sony As might be expected from the symmetrical form of the con- dition from which it has been derived, the spectrum is of the same analytical fort GP) =e YEW cis [— anf —h) + 6] -- (1.28) ‘The envelopes of Goth the signal and its spectrum, or theit absolute values, have the shape of probability curves, as illus trated in Fig. 6. Their sharpness are reciprocal. ‘Because’ of its self-reciprocal character, the probability signal has. always played an important part in the theory of Fourier transforms. | In three recent papers, Roberts and Sinimonds have called attention to some of its analytical advantages! 1.12, (43 But its minimum property does not appear to have been recog- ‘ized, WC is this property which makes the modulated proba- bility pulse the natural basis on which to build up an analysis of signals in which both time and frequency are recognized as references, Tt may ‘be proposed, therefore, to call a pulse according 10 equation (1,27) an elementary signal. In the information diagraim it may be represented by a rectangle with sides Ar and Af, and area one-half, centring on the point (jf), It will be ‘shown below that ‘any. signal can be expanded into elementary signals in-such a way that theit representative coct- angles cover the whole time-frequency area, as indicated in Fig. 47.. Their amplitudes can be indicated by a number written-into the rectangle, or by shading, Each of these areas, with its associated datum, represents, as it were, one elementary al Fig. 1.6.—Envelope of the elementary signal. { L LY ; ‘ a4 aN ‘eae Fig. 1.7.—Representation of signal by logons. quantum of information, and it is proposed to call it a logon. ‘Expansion into elementary signals is a process of which Fourier analysis and time description are special cases. The first is ‘obtained at a = 0, in which case the elementary signal becomes a sine wave of infinite length; the second at «—» oo, when it to a “delta function.” Tt will be convenient to explain the expansion into elementary signals in two steps. The first step leads to elementary areas of size unity, with two associated data, but it is simpler and ‘more symmetrical than the second step, which takes us to the limit of sub-division. This first step corresponds to division of the information area by a network of lines with distances A¢ and 1/Ar respectively, as illustrated in Fig. 1.8.* ‘The elementary arcas have suffixes 7 Representation of signal by a matrix of complex dbo fe . amplitudes. pcg it tor of vs einen (aS ESE it = DST rapes436 GABOR: THEORY OF COMMUNICATION jn the time direction, and & in the frequency direction. The centre lines (horizontally) may be at fy =r At, assuming. for convenience that we measure time from the “zero"th of these lines. The expansion is given by the following formula a ab? wo Y » RxD — OT RHE laf) (1.29) ‘The matrix of the complex coefficients c,, represents the signal in a symmetrical way, as itis easy to see that if the expansion ‘exists we arrive—apart from a constant factor—at the same coefficients if we expand ¢(f) instead of ¥(). ‘As the elementary signals in (1.29) are not orthogonal, the coefficients cy, are best obtained by successive approximations. In the first Spproximation we consider each horizontal strip ‘with suffic m by itself, and expand the funetion (0) as if the other strips did not existyrin: the-inerval (t, — 4A) 10(¢, + 440, ‘by putting * Moe Ge = Se gy cis Omir In this formula the exponential function, which is independent of A, has been brought over io the lef. We have now a kknowa function on the lef, and a Fourier series on the right, which by knowa methods gives immediately the first approxima tion for the coefficients c,,. This represents y(!) correctly io the intervals for which the Series are valid, but not outside them. YE the first approximations are added’ up with summation indices n, there will be a certain error duc to their overlap. ‘A second approximation can be obtained by subtracting this error from y() 8 eqn. (1.29) and repeating the Procedure. It ‘can be Expected to converge rapidly, as the exponential factor decays s0 fast that only neighbouring, strips # influence each other perceptibly. ‘This expansion gives ultimately one complex number ¢ye for ‘every two elementary areas of size one-half, The real and imaginary parts can be inierpreted as giving the amplitudes of the following two real elementary signals 30 fg oP Ae P28 ample — ty) + (130) where a2 = da/(AN2. These can be called the “cosine-type"" and “sinetype’” clementary signals, They are illustrated in Fig. 19, Wecan use them to obtain a real expansion, allocating ‘Se pe Fig. 1.9.—Real parts of elementary signal. fone datum to every cell of one-half area, But it may be noted that this will have to be necessarily a more special and less sym ‘metrical expansion than the previous one, as the transform of a cosine-type elementary signal, far example, will not in general be fof the same type, As always in communication theory desription by complex numbers is founally simpler than real data, ‘We now divide up the information plane as showa in Fig. 1 te | be fos | be | a [be ay | Aa | eee | Bae ov | te | aw | & te [be | on | oe : fa | be | ae | de Fact —— Fig, 1.10.—Expansion of arbitrary signi in cosine-type and sine-t into cells of size one-half, measuring’Ar in the time, and 4 jn the frequency, direction, Starting from the line of 2 Frequency, we allocate to these areas in every strip alternatel: cosine-type and a sine-ype elementary signal. Fvidently * ‘must start with a cosine signal at f~ 0, as the sine-type sig: ‘would be zero. This leads us to the following expansion of 1 real signal s(0):-— 0 = She Gee "e [oyu 08 2mk(r— nADIAL + bygsin Inthe + NCE nADIAN] CL: In order to find the coefficients dye and by we can carry ¢ the same process of approximation‘as explained in connect ‘with expansion (1.30), but with a difference, At the first st wwe arrive at an equation of a form SA) = a ye 008 KX > fgg Sin (K + HD with the abbreviations x= 2n(0 — mAD/Ar, and f(x) = s exp dae — nANFAD?. But the ttigonometric series on t right is not a Fourier series. {tis of a somewhat unusual ty in which the sine terms have frequencies mid.way between t cosine terms. It wil be necessary to show briefly that this seri can be used also for the representation of arbitrary functior FFirst we separate the cven and odd parts on both sides of 1 equation, by puting YF) +4- 9] = Dt yy 008 kee HAW —F(-9) = 2h by sin e+ Dx ‘The first is a Fourier series, but not the second. We have see hhowever, in Section 3, how all the frequencies contained in Function can be raised by a constant amount by means of process which involves calculating the function in quadrate ‘with it. Applying this operation to both sides of the ta ‘equation we can add + to k + 4, and obtain the ordinary Fouri sine series, which enables the coefficients to be calculated. ‘The expansion into logons is, in general, a rather inconvenie,GABOR: THEORY OF COMMUNICATION process, as the elementary signals are not orthogonal. If only approximate results are required, it may be permitted (o neglect the effect of their interference. This becomes. plausible if we consider that an elementary signal has 76°8% of its eneray inside the band Ar or Af, and only 11°6% on cither side, ‘Approximately correct physical analysis could be carried out by ‘means of a bank of resonators with resonaace curves of probs bility shape. It can be shown that if the energy collected by a resonator tuned to fis taken as 100%, the resonators on the Tight and left of it, tuned to /+ Af and f— AY, would collect ‘only 0:65% each, Roberts and Simmonds! 12,1: have sven consideration to the problem of realizing circuits with sespoases of probability shape. “Though the overlapping of the elementary signals may be of small practical consequence, it raises a question of considerable ‘theoretical interest. The principle of causality requires that any ‘quantity at an epoch f can depend only on data belonging to epochs earlier than 1. But we have seen that we could not carry out the expansion into elementary signats exactly without taking into consideration also the “overlap of the future.”” In fact, strict causality exists only in tho “time language”; as soon. 5 We use frequency as an additional reference the sort of wn- ‘certainty occurs whieh in modera Physics has often been called the “breakdown of causaity."” But rigorous time-analysis is ‘possible only with idea! axcillographs, not with any real physicat instrument; hence strict causality never applies in practice. A limitation of this concept ought not to cause difficulties to clectrical engineers who are used to the Fourier integral, ic, to an entirely non-causal method of description, (© SIGNALS TRANSMITTED IN MINIMUM TIME ‘The-dementary signals which have been discussed in the last ‘Section assure the best utilization of the information area in the sense that they possess the smallest product of effective duration by effective frequency width. It follows that, if we prescribe the effective width Af of a frequency channel, the signal trans- mitted through it in minimum time will have an envelope YO = exp — OnVANHG — Dt and, apart from a cisoidal factor, a Fourier transform. 32 : omen HEEL... 03m Bot the problem which most frequently arises in practice is somewhat dierent. Not the effective spectral width is. pre- teribed, but the total width; ie. a frequency band f ~ fis sven, outside which the spectral amplitude most be 2er0. What is the sigual shape which can be transmitted through this channel in the shortest effective time, and what is its effective duration? ‘Mathematically the problem can be reduced to finding the _spectrum if) of a signal which makes a 2 allo . + a) ‘minimum, with the condition that 4K/) is 2er0 outside the jae f — fy But this is equivalent to the condition that (f) vanishes at the limits f, and fa. Otherwise, if Y/) had a finite value at the limits but Vanished outside, the discontinuity at the limits would make the mumerator of equation (1.34) divergent. (This is the converse of the well-known fact that a signal with an abrupt break contains frequencies up to infinity, ich dec only. hyperbola, not fast enough to make /2 nite) a‘ ar ‘The problem is one of the calculus of variations, an is solved in Appendix 9.4, where it is shown that the signals transmitted jn minimuen time must be amoog the Solutions of a aifferential equation Ste ago i ass where A is an undetermined constant. But the possible values. GFA ace defied by the abnary condition that i) must vanish A the lous of the waveband.? ‘Hence al adsisible solutions are ofthe form l= $A = sind + 639 where & is an integer, We can call this the kth characteristic function of transmission through an ideal band-pass filter. Its effective duration is aM OAR aan and its effective frequency width af=G-Agl(E- ga) - 438 The shortest duration Ar belongs to & = [, i.e. to the funda- ‘mental characteristic function, which is ustrated in Fig. 1.11, UX A be Se OT ig. 1,11.~Specirum fsigal which can be traramitted a minimum ‘imme through an ideal band-pass iter, and the sigoal ite. ‘The product AtAf is also smallest for k = 1; its value is 0-571. ‘Though this s not much more than the absolute miaimum 0-5, the transmission channel is poorly willed, es the effective frequency width is only 0-456 of (f,—;). Practite has found. a way to overcome this difficulty’ by means of asymmetric, vestigial or single-sideband transmission. In these methods the spectrum is cut off at or near the centre more or less abruptly. ‘This produces a “splash,” a spreading out of the signal in time, but this effect is compensated in the reception,.when the other sideband is reconstituted and added to the received signal. ‘The advantages of a signa! of sine shape, as shown in Fig. [.1, hhave already been noticed, #s it were, empirically: by Wheeler and Loughreat in thei thorough study of evsion images. AB jin television the signals transmitted represent light intensities, Le. energies, our definitions must be applied here with a modi- fication. Either the square root of the light intensity must be substituted for y, oF the squeer-root of the Fourier transform + Proten of i dros main and geil sg, ee eee arate ae438 . of the signal for g. ‘The practical diference between these two possible definitions becomes very small in minimum problems. If we adopt the second, we obtain the same “cosine-squared” law for the optimum spectral distribution of energy which Wheeler and Loughren have considered as the “most attractive compromise."* Fig, 1.11 shows also the signal s(0) which is wansmitted in shinigum time by @ band-pass filter. ft can be seen that it differs in shape vesy litle indeed from is spectrum. It'may be noted that the total time interval in which the signal is appre- ciably diferent from zer0 is 21Uj,— It can be seen from Fig. 1.11, thai the optimum signal utilizes the edges of the waveband—in single-sideband television, the upper edge—tather poorly. But this is made even worse in television by thé convention of making the electromagnetic amplitudes proportional to the light intensities, so that the electromagnetic energy spectrum in the optimum case has the shape of a cost curve. This means that the higher frequencies will be easily drowned by atmospherics. Conditions can be improved by “compression-expansion”” methods, in which, for ‘example, the square root of the light intensity is transmitted, and squared in the receiver: OF COMMUNICATION PROBLEMS BY ‘OF THE INFORMATION DIAGRAM As the foregoing explanations might appear somewhat abstract, it appears appropriate to return to the information Siagram and to demonstrate its usefulness by means of a few examples. : . Let us take frequency modulation as a first example. Fig. 1.12 (@ DISCUSSION ‘MEANS Fig. 1.12.—Thvee representations of frequency modulation. contains thee different iustrations ofthe same slowly modulated cartier: the time representation, tht spectrum and its picture in the information diagram, It‘can be seen that the third illos- ‘ration corresponds very closely to our familiar idea of a variable frequency. The only departure from the naive expectation that its pictorial representation would *» an. undulating curve is that the curve has to be thick and bi gg s. But it appears preferable fot to show the blurring, not on. ccause it is difficult to draw, ‘but also because it’ might give rv to the idea that the picture could be replaced by a definite density distribution, Instead we have represesited it ey logons of area one-half. The shape ‘of the.rectangles, ic. the ratio AAS, is entirely arbitrary and GABOR: THEORY OF COMMUNICATION depends on the conventions of the analysis, If A‘ is taken equal to the damping time of, say, 2 funk of reeds, the picture sives an approximate description of the response of the instru- ment. It gives also a rough picture of our aural impression of a siren. How this rough picture can be perfected will be shown in Part 2 A second example is time-division multiplex telephony. a problem which almost forces on us the simultaneous considera tion of time and frequency, “Bennett '3 has discussed it very thoroughly by an irreproachable method, but, as is often the case with results obtained by Fourier analysis, the’ physical origin of the results remains somewhat obscure. An attempt will now be made to give them a simple interpretation, Jin time-division multiplex telephony, synchronized switches at both ends of a line connect the line in cyclic alternation to a nomber N of channels. Let f, be the switchiog frequency, jie, the mamber of contacts made per. second. “What is the ‘optimum switching frequency if N conversations, each occupying, a frequency band w are to be transmitted without loss of informa- tion and without crosstalk—i.e. mutual interference between channels-—and what is the total frequency-band requirement W? ‘The information diagram is shown in Fig. 1.13. The fre- 1 eT ee WLLL, vi CA" g = a 7 Fig. 1.13-—tnforpatio gran f tine division listen telephony pot ‘quency band Wii sub-divided in the time direction into rectangles of a duration Wf, ie. f, rectangles per sec. If these arc to ‘transmit independent data they cannot transmit less than one datum at a time, But one datum, or logon, at a time is also the optimum, as otherwise the recsivers. would have to dis- criminate between two or more data in the short time of contact, and distribute them somehow over the long waiting time between ‘two contacts, Hence, ifno information is to be lost, the number of contacts per second must be equal to the data of N con- Yersations each of width we. f,= 2NW. This ako Benet’ result, ‘We now consider the condition of crosstalk. "This is the exact ‘counterpart-of the problem of miniyaum transmission time in @ fixed-frequency channel, considered in the last Section, except that time and frequency are interchanged. Thus we can say at ‘once that the optimum signal form will be the sine shape of Fig. 1.1, and the frequency requirement will be very nearly 2f. ‘The chatacteistic rectangle A’AS of this signal is shown in every switching pericd, with the dimensions as obtained in the last Section. The total frequency band requirement becomes W = 2f, = 4Nw. This can be at once halved by single sideband transmission, ie. transmitting only one-half of W. But even this does not represent the limit of economy, as the signal is symmetrical not only in frequency, but also in time. In the case of the example treated in the previous: Section this was cf no use, as the epoch of the signal was unknown, Bu in time division multiplex the epoch of each signal is accurately known; hence it must be possible to halve the waveband once more and reduce W to the minimum requirement = Nw. An ingenious,GABOR: THEORY OF COMMUNICATION though rather complicated, method of achievig this, by means Of special filters associated with the receiving channels, has been described by Bennest.15 (© REFERENCES (1.1) Carsow, J. R.: “Notes on the Theory of Modulation,” Proceedings of the Institute of Radio Engineers, 1922, 10, p. 51. (1.2) Nequist, Hi: “Certain Factors affecting Telegraph Speed,” Bell System Technical Journal, 1924, 3, p. 324. (1.3) Kierwdiuer, K.: ‘Transient Phenomena’ in Wave Filters,” Elektrische Nackrichten-Technik, 1924, 1, wt. (4) Hagiuzy, RV. Le: “Transmission of Information,” Bell Systema Technical Journal, 1928, 7, p. 535. (5)-Gray, F., Horton, J. W., and Maries, C. R.: “The Production and ‘Utilization of Television Signals,” ibid, 1927, 6, p. 560. (1.6) Luscutay, F.: “Modern Communication Systems, LEE, 1932, 71, D. T16. (L.7) Caverntix, G. A., and Foster, R. M. “Fourier Integrals for Practical “Applications,” Bell Telephone System ‘Monograph B 584, 1931. (1.8) Carson, J. R., and Fry, T. C.: “*Variabledrequency Electric “Theory,” Bell System Technical Journal, 1937, 16, p. 513. (1.9) Stowakr, G. W.: “Problems Suggested by an Uncertainty Principle in Acoustis,” Journal of the Accustical Society of America, 1931, 2, p. 325. (1.10) Gotpwark; PB. C,, and Hespricxs, P. S,: “Synthet Reverberation,” Proceedings of the Institute of Radio Engineers, 1939, 27, p. 747. (U1) Rosenrs, F. F.,.and Sumonos, J. C.: "Some Properties of a Special” Type of Electrical Pulse,” Philosophica! Magazine, (Vil), 1943, 34, p. 822. (1.12) Rowers, F. F., and Soawons, J.-C.: “Further Properties of Recurrent Exponential’ and Probability Wave- Forms,” ibid (VID), 1944, 35 P. 459. (1.13) Rowents, F. F., and Simmons, J. C.: “The Physical Realizability of Electrical Networks having Prescribed Characteristics," ibid, (VM), 1944, 38, p. 778, (1.14) Waster, H. A., and’ Loucunen, A. V.: “The Fine ‘Structure of Television Images,” Proceedings of the Institute of Radio Engineers, 1938, 26, p. 540. (115) Bennerr, W.-R.: “Timedivision’ Multiplex Systems,” Ball System Technical Journal, 1941, 20, p. 199. (9) APPENDICES. (9.1) Analysis in Terms of Other than Simple Periodic Funetions ‘The discussion in Section 1 suggests a question: Why are we doing our analysis in terms of sine waves, and why do we limit ‘our communication channel, by fixed frequencies? Why aot ‘choose other orthogonal functions? In fact we could have taken, for example, the orthogonalized Bessel functions VOL) as the basis of expansion. J, is a Bessel function of fixed but arbiteary order m5 r, is the tN root of J,(x)=0; & is the expan- sion index. These functions are ortNogonal in the interval O.< £7. The factors r/r have the dimension of a frequency. We could now think of limiting the transmission channel by two “Bessel frequencies,” say jay and iq. Here the first di ence arises, The number of spectral lines between these Li will be the number of the roois of J,(x) = 0 between the limits vit and pit. But this number {S not proportional to 7. Journal 439 Hence.a Bessel channel, or a channel based on any function ‘other thaa simple harmonic functions, would not transmit che same amount of information in equal time intervals Tn principle it would be possible to construct circuits which transmitted without distortion any member of a selected set of orthogonal fuactions, But only harmonic functions satisty linear” differential equations in which time does not figure ‘explicily; hence these are the only ones which can be trans- mitted by circuits built up of constant elements. Every other system requires variable circuit components, and as there will be a distinguished epoch of time it will also require some sort of synchronization between transmitter and receiver. In come petition with fixed-waveband systems any such method will have the disadvantage that wider wavebands will be required to avoid interference with other transmissions. Though this disadvantage as jn the case of frequency modulation—might te outweighed by other advantage, investigation of such systems is outside the scope of the present study, which is mainly devoted to the problem of wavehand economy. (9.2) Mechanical Generation of Associated Signals, and the Problem of Direct Production of Single Sidebands In order to gain 2 more vivid picture of signals in quadratuie than the mathematical explanations of Section 3 can convey, it may be useful to discuss 2 method of generating them meckan- ically. Ttis obvious from equations (1.7) and (1.8) that, i order to generate the signal o(0) associated with a given signa), it is necessary 10 know not only the past but also the fare, Though formally the whole future is involved, the “relevant future" in transmission problems is usutily only a fraction of a second. This means that we cea produce o() with sufficient accuracy if wwe convert, say, O-1 sec of the future into the past; in other wards, if we delay the transmission of s() by about this interval. Fig. 114 shows a device which might accomplish this. Fig. 1.14. Device for mechanical generation ofa signal in quadrature ‘with a given signal. ‘The light of a lamp, the intensity of which is modulated by the signal s(), is thrown through a slit on a transparent rotating drum, coated with phosphorescent powder. The drum therefore carries a record of the signal with it, which decays slowly. After turning through a certain angle the record passes a sit, and here the light is picked up by 9 photocell, which transmits 5(-) with a delay corresponding to the angle.* "On the inside of the drum ‘two hyperbolically-shaped apertures are arranged at both sides Of the slit opposite to the frst photocell. The light from the two hyperbolic windows is collected by two photocels, which are connected in opposition. By comparing this arrangement + Asormewtat sinter device (for snoterpurpowe) ba mended by Gok 20) ASPEN oea0 ‘with equation (1.7) itis easy to see that the difference of the £0 photocell currents will be proportional to the function in quadra- tare with s(0). ‘The complex signal has been discussed at some length as it helps one to understand certain problems of communication engineering. One of these is the problem of single-sideband transmission. It is wel known that it is not possible to produce 4 single sideband directly. . The methed employed is to produce both sidebands and to suppress one. Equation (1-7) explains the reason. Direct singlesideband production involves knowledge of the future. The conventional modulation methods always add and subtract frequencies simultaneously. With mechanisms Tike the one shown in Fig. 1.14 it becomes possible to add or subtract them, This means forming the following expression RKO exp jen] = (0) 008 w.t — off) sin et wheré «, is the angular carrier frequency. By substituting & hharmoni¢ oscillation for s(0 isis easy to verify that c, has been added to every frequency present in the signal. Direct produc- tion of single sidebands involves, therefore, the following opera- tions: Modulate the signal with the carrier wave, and subtract from the product, the modulation product of the signal in quadrature with the carrier wave in quadrature, Tt is not, of course, suggested that this might become a practical method; the intestion was merely to throw some light on the root of @ well-known impossitility, (9.3) The Schwarz Inequality and Elementary Signals ‘The inequality pervane sageerven( fe a) 039 js valid for any rea! or complex function YY" which is continuous and differentiable and vanishes at the integration limits. The following is a modification of a proof given by H, Weyl. If ay by are two sets of m real or complex wumbers, a theorem ue to H. A. Schwarz states that [eis +--+ ab,2 < jay +... + aya%) Oot +b. C40) 1 a’s and b's are all real numbers, this can’ be interpreted as ‘expressing the fact that the cosine of the angle of two vectors with components a . . . a, and by . « . 6, in an n-dimensional Euclidian space is smalier than Unity: "This can be easily under- stood, as in a Euclidian space of any number of dimensions a ‘two-dimensional plane can be made to pass through any two vectors issuing from the origin; hence the angle between them has the same significance as in plane geometry. Equation(1.40) is a generalization of this for “Hermitian” space, in which the components oro-ordinates ofthe vetor are themselves complet By @ passage to the limit she sums in (1.40) may be replaced by integrals, so that Eayb, > [Fe and similarly for the other two sums, The real variable 7 now takes the place of the summation index. ‘The Schwarz inequality now becomes [fe dre ecrit + (Lrtetarlt Pale +See e . (1.4 ‘The second part of this inequality states the fact that the sum of the absolute squares of two conjugate complex numbers is never less than half the square of their surns. ‘We now put a fae ‘Substitution in (1.43) gives : acarrennieetén > [[(vSE + v2 )nee] a5 ‘The right-hand side can be transformed by partial integration into + (44) Ses oP nie [edogepar= -Jever 0.46) where it has been assumed that ‘Y vanishes at the integration limits. Substituting this in (1.45) we obtain the inequality (1.39), In ordor to obtain the elementary signals we must investigate when this inequality changes into an equality. From the ‘seometrical interpretation of Schwar2's inequality (1.40), it can be concluded at orice that the equality sign will obtain if, and only if, the two vectors g, b have the same direction, ie. by = Ca, In Hermitian space the direction is not changed by multiplica- tion by a complex number, hence C:need not be real. ‘This condition can be applied also to the inequality (1.29), but with a difference. (1.39) will become an equation only it both the conditions (1.41) and (1.42) become equalities; ie. if the following two equations are fulfilled f= Cy and ft = C’g* aa where C and C’ are real or complex constants. But these £0 equations are compatible if, and only if, Cac... (1.48) jn which ease the two equations (1.47) become identical. On substituting f and g from (1.44) they give the two equivalent ‘equations Ce ae 1.49) From either of these we can eliminate Y or its conjugate 7" and are led to the second-order differentia\ equation ECB) -comy . (150) Multiplying both sides by (@'¥/dr)/x, this becomes integrable and gives: But the constant is zero, as at infinity both and d'V¢r must vanish. We thus obtain the first-order equation. ay aw F = CC*Y? + const. asi ECORYS. sayGABOR: THEORY OF COMMUNICATION with the solution (apart from a constant factor) Weepadicee. 2... asm Of the two signs we can retain only the negative one, as other- wise the signal would not vanish at infinity. Puttiog 3{C] = a# we obtain the envelope of the elementary signal. ‘The signal y itself results from this by multiplying by cis 2nf/(t — 7) and is discussed in Section 5. Tt will be useful to sketch briefly the difference between the analysis based on elementary signals and the method of wave ‘mechanics. Tn the foregoing we have answered the question: ‘What functions 'Y' make the product A fA assume its soallest possible value, ie, one-half? The question posed by wave ‘mechanics is more general: What functions makes AfAt @ minimum, while fulfilling the condition of vanishing at infinity? This is a problem of the calculus of variations, which leads, instead of to eqn, (1.50), to a more general equation, called the “wave equation of the harmonic oscillator”: ey i Gr tA~ any =0 where A and a-are real constants, This equation, which con- tains (1.50) as a special case, has solutions which are finite every- where and vanish at infinity only if E A=aQn +) where n is a positive integer. These “proper” or “characteristic” solutions of the wave equation are (apart from a constant factor) o mentee pate Y=. = ‘They are known as orthogonal Hermite functions* and form the basis of wave mechanical analysis of the problem of the linear oscillator. They share with the probability function— ‘which can be considered as the Hermite function of zero order— the property that their Fourier transforms are of identical type. ‘The product AAC for the nth Hermite function is ANAS = 42a +1) ‘That is to say that the Hermite functions occupy in the informa- tion diagram areas of size $, 3,4... Because of their ortho- sonality Hermite funcons readied themgelves to the expan. Son of arbtary signals, hence ther importance. in wave Tetianic But they ae es sale forthe analy of cone ‘tinuously emitted signals, as they presuppose a distinguished ‘epoch of time t= 0, and they do not permit the sub-division of the information area into non-overlapping elementary cells.t sua st ara BERS Waa en tty aL acne i Re ee coer arr SAO Saket a (9.4) Signals Transmitted in Minimum Time theough a Given Frequency Channel It will be convenient to use “frequency language," ic, to ‘express the signal by its Fourier transform 4(f).. The problem is to make the effective duration Af of a signal minimum, with the condition that 4(f}= 0 outside an interval f, —fy, Thus 1 “# at Gay fit ral au ‘must be @ minimum, where Mo= [eer This is equivalent to making the numerator in (1.54) a mini- ‘mum with the auxiliary condition Mg = constant, and this in ‘tra can be formulated by Lagrange’s method in the form [G 4 +Ags)y=0 where Ais an undetermined multiplier. The variation of the first term is ae GSP SEB AD - [ifn he] Gee See om Bat at the limits ¢ must vanish, as it is zero outside the interval nee tom i ya aes (1.54) would not converge. Hence we bave here 34 = &6¢ = 0, ‘the first term vanishes. The variation of the second term (1.55) in (1.55) is i NG 6 + 8 asn ‘The condition (1.38) thus pves S[GE + Ae)as + Ge as)pe|e—0 sn and this can be identically fuliled for arbitrary variations 8 if, and only if, &e we ‘This is the differential equation which has to be'satisfied by the signal tanamitted in minimum time, Its solution is discussed in Section 6, $Ageo 6... G59)a GABOR: THEORY OF COMMUNICATION Part 2. THE ANALYSIS OF HEARING SUMMARY ‘The metbods developed in Part 1 are applied to the analysis of hearing sensations, in particular to experiments by Shower and Bid alph, and by Burek, Kotowski and Lichte on the discrimination of frequency and time by the human ear. I is shown that experiments ‘of widely different character lead to well-defined threshold “areas of discrimination” io the information diagram. At the best, in the interval 60-1 000 cs the human ear can discriminate very nearly evry second datum of information; i. the ear is almost as perfect as any instrumeat can be which is not responsive to ‘Over the whole ‘auditory range the efficiency is much less than 50% as the diserimina- tion falls off sharply at higher frequencies. ‘The threshold area of discrimination appears to be independent of the duration of the signals between about 20 and 250 millisec. This remarkably wide interval cannot be explained by any mechanism in the inner ear, but may be explained by a new hypothetical effect in nerve conduction, i, the mutual influence of adjacent nerve fibres, () ANALYSIS OF HEARING In relation to the ear, two rather-distinct questions will have to be answered. The first is: How many logons must be trans- mitted per second for intelligible speech? ‘The second is the corresponding question for the réproduction of speech or music which the ear cannot distinguish from the original. A precise answer to the first question will not be attempted, but some important data must be mentioned. Ordinarily it i ‘assumed that the full range between about 100 and 3000 c/s is necessary for satisfactory spetch transmission. But Dudley Homer's ingenious speech-analysing and synthetizing machine, the Vocoder?! has achieved the transmission of intellisible speech by means of 11 channels of 25 /s each, 275 c/s in all. ‘This means a condensation, or compression, ratio of about 10, ‘Another datum is an estimate by Kuptinier® of the product ‘of time-Interval by frequency-width required for the transmission ofa single letter in telephony, and in the best system of telegraphy, as used in submarine cables. The ratio is about 40. ‘This suggests that the Vocoder has probably almost reached the admissible Limit of condensation. ‘The transmission which the ear would consider as indistin- {guishable from the original presents a more exactly defined and intrinsically simpler problem, as none of the higher functions of intelligence come into play which make distorted speech intelligible, G. W. Stewart in 1931 was the first to ask whether the limit of aural sensation is not given by an uncertainty relation, which he wrote in the form ArAf = 1, without, how ever, defining Ar and Af precisely. He found the experimental ‘material insufficient to decide the question, though he concluded that there was some evidence of agreement. New experi- mental results, which have become available since Stewart's note, and a more precise formulation of the question, will allow us to give a more definite answer. In Seetion § of Part 1, methods were described for the expaa- sion of an arbitrary signal into elementary signals, allocated to cells of a lattice. Fig. 2.1 is an example of a somewhat different method of analysis, in which the elementary areas have fixed shape but no fixed position, and aze shifted so as to give a good representation with a ‘minimum number of elementary signals. We now go a step further, and adjust not only the position but also the shape of the elementary areas to the signal, in such a way that it will be approximately represented by 2 ‘minimum number of logons. This may be called “black-and- white” representation, and it is suggested that—withia certain limits itis rather close to our subjective interpretation of aural + voted by Loreen, Refereace 16 Fig. 2.1.—Sine wave of finite length. Response of a bak of eaonators Arent veapose oft seosations. Fig. 2.1 illustrates this, If a sine wave of fir ‘duration strikes a series of resonators, say a bank of reeds, w a time-constant which is a fraction of the duration, theitrespot will be approximately as shown by (o). But, as the ear har hears the two noises or “clicks” at the beginning and end the tone, is sensations can be better described by Fig. 2.1( We shall find Jater more evidence for what may be called + justable time-constant” of the ear. It appears that, ‘general, the ear tends to simplify its sensations in a similar w to the eye, and the analogy becomes very evident in the tv dimensional representation. Tt will be shown below that there is good cviderce for wi may be called a “threshold information sensitivity” of the € ie. a certain minimum area in the information diagram, wh must be exceeded ifthe ear is to appreciate more than o datum, The usefuess of this concept depends on how far) threshold value will be independent of the shape of the ar We must therefore test it by analysing experiments with tc signals of different duration. Tt.has been known for a long time (Mach, 1871) that a vi short, sinusoidal oscillation will be perceived as a noise, 1 beyond a certain minimum duration as a tone of ascertaina pitch. The most recent and most accurate experiments on t subject have been carried out by Barck, Kotowski a Lichte22.23° They found that both at 500 and 1 000c/s 1 ‘minimum duration after which the pitch could be correc ascertained was about 10 millisec for the best observers. Tr second series of experiments they doubled the intensity of + tone after a certain time, and measured the minimum durati necessary for Fearing the step. For shorter intervals the step tone couid not be distinguished from one which started W double intensity. ‘These two series of tests enable us to estimate the thresh area for very short durations. Fig. 22 explains the method a frequency of SOO-c/s. After 10 millisec the signal was j recognizable as a tone. But unless it lasted for at le 21 millisec, the ear was not ready to register a second date independent of and distinguishable fom the first. We clude, therefore, that the threshold area is determined by frequency width of the first signal and the duration of second, It is not necessary to approximate the chopped s ‘waves by elementary signals, as the ratio of the durations wo remain the same. This was 2-1 for 500 cfs, and 3-0 for 1 000: We conclude that in these regions it takes 2°1 and 3 element areas respectively to convey more than one datum to the earGABOR: THEORY he Wig. 2.2.—Experiments of Burck, Kotowskt and Lichts, Let us now consider another series of tests, the experiments ‘of Shower and Biddulph on the pitch sensitivity of the ear. Tn these tests the frequency of @ note was Varied almost sinu- soidally between a lower and an upper limit. Tae actual faw of variation was not exactly sinusoidal, as the top of the wave was flattened and rather difficult to analyse in an exact manner. In the following approximate analysis we will replace it by sinusoidal frequency modulation with a total swing 8f, equal to the maximum swing in the experiments. By this we are likely OF COMMUNICATION 443 Tt is well known? that the spectrum of a frequency-modu- Jated wave with mean frequency fo, total swing 5/ and modula tion frequency J, can be expressed by the following series ar is (209 sin 2g is (2nft + Pp sin Hf) SUB Af) eis 2a fa + hm)» 2.1) is the Bessel function of ath order, The amplitudes of the tines, spaced by the repetition frequency, are therefore pro- portional’ to J,@//2f,). Their absolute values are shown at ‘the bottom of Fig. 2.3 for four tests of Shower and Biddulph, On the other hand, the absoiute amplitudes of the side lines in the spectrum of the (wo altemating sequences of elementary signals are given by the following formulae ay cosh pry? Sym erp — (2) OPS (2) haf ‘The upper formula is valid for even, the lower for odd, orders 2. With the help of equations (2.1) and 2.2) che available constants cz and f, have been fitted 50 as to represent exactly the ratio of the first two side fines to the central one. The result is shown in Fig. 2.3, in which the elementary signals are reptesented by 2 Fig. 2.3.—Expatiments of Shower and Biddulph. ‘Toetraguesey-moduletedigmisace replaced by wo alemaling see of Saneatay sina Which Produc ery cel the sme spectrum. to commit an error ia the sense of overrating the car sensitivity, but this will give ve a safe basis for estimating the chances of deceiving the ear. The modulation frequency in Shower and Biddulph’s experiments was 2c/s, and the sensation level was kept constant at 40 db above the threshold of audibility. ‘Their results for the aviaimum variation 8f at which the trill could just be distinguished from a steady tone are as follows:— fo 62:5 125 250 500 1000 2000 400) 8000 es 2M : 0-043 0-025 0:012 0-005 0-003 0-0023 0-00225 0:0097 «fs as 27 31 29 25 30 46 9:0 295 ofs It will be seen that 8 remains almost constant up to 1000 ¢/s; from about | 000 ¢/s it is the catio 8 which is nearly constant. We now replace the signals ised in these experiments by two Periodic sequences of ‘elementary signals with frequencies fa % Hy Staggered in telation to onc another, o that pulses with higher and lower frequency alternate at intervals of 0-25 sec. {In order to approximate the actual signal as well as possible, ‘we must use the available constants f, and a: (the ‘sharpness? of the elementary signals) $0 a5 to produce nearly the same spectrum. their rectangles of area one-half, The agreement of the spectra ‘even for higher orders 1s very good up to 2.000 cfs, but less satis factory at 4000 and 8.000 o/s. But it would be useless to try better appreximations, for example by adding one or two further sequences of elementary signals, More accurate information ‘could be obtained only from experiments based on elementary signals. “It may be hoped thet such tests will be undertaken, ‘especially as Roberts and Simmonds have suggested easy methods for producing such signals. For a first orientation the results derived from the tests of Shower and Biddulph appear quite satisfactory. Tt can be seen {from Fig. 2.3 how rectangles can be constructed in the informa- tion diagram which mark the limit at which the ear can just begin to appreciate a second datum. In this case the meaning of the threshold is thet the irill.can just be distinguished from a steady tone, . Measured in units of elementary areas of one- half, their values are as follows:— Frequency... 62:5-1909 2000" 4.000 8000 ¢/s (Threshold areayO'S 2-34 2-88. 3:92 G9. ‘The reciprocals of these figures can be considered as per- formance figures of the ear as compered with an ideal instrument. In fact, the performance figure of an ideal instrument would be unity, ag it would begin to appreciate a second datum as soon. as the minimum information area of one-half was exceeded byoo ‘an amount, however small, The performance figure derived from the experiments of Shower and Biddulph between 62 and 8000/8 is shown in Fig, 2.4, The diagram also contains two GABOR: THEORY OF COMMUNICATION See ee eee eee eee e ot Fig: 2.4.—Performance figure of the ear. BALL —Bunk, Kotowskdand Lice, 5. and B-—Shower and Riddalph. points derived from the experiments of -Burck, Kotowski and Lichte, which ft in as well as cam be expected, It is very remark- able that up to about 1 000 ¢/s the performance figure is almost 30%, which is the ideal for an instrument like the ear which cannot distinguish the phase of oscillations, i.e. rejects one-half of the data. At higher frequencies, however, the efficiency is ‘much less. ‘The good fit of the figures obtained from the experiments of Birek, Ketowski and Lichte, which were carried out with dura- tions of 10-20 millisec, with those of Shower and Biddulph, in which the threshold area reasured 250 railisec in the time direc- tion, indicates. two facts. One is that, at least wp to about 1000 ¢fs, and for durations at least in the limits 20°250 mifscc, the threshold information afea is a characteristic of the ear. Evidently the performance figure must go to zero both for ‘extremely short and for extremely long elementary signals, but within these wide and very important limits it appears to have an almost constant value. ‘The other fact which arises from the first is that the ear appears to have a time-constent adjustable at least betwen 20 and 250 ‘ilisec, and that the eat adjusts to the content of the informa- tion which it receives. But there can be litte daubt that, what- fever resonators there are in the ear, they are very strongly damped, and that their decay time is of the ordet of 20 millisee for rather less. “This is borne out by the experiments of Wesel and Lane on the amplitudes of the oscillations of the basilar membrane in the inner ear.* A pare tone excites such a broad region to oscillations that R. S. Hunt? who has recently made a thorough investigation of Wegel and Lane's data, infers from them a decay by U bel in only 2 cycles, ie. in only 2 milisec at 1000 c/s! ‘Though this estimate might be too low, there can bbe no doubt that the decay time of the ear resonators cannot substantially exceed 10 millisec, and it is impossible to imagine that they would keep on vibrating for as much as a quarter of a second. Hence, even if the duration of a pure tone is con- siderably’ prolonged beyond the 10 millisee approximately required for pitch perception, the ear resonators will stil display the same broad distribution of amplitude, This is illustrated in Fig. 2.5. In order to explain the high pitch sensitivity of the cat, as shown, fos example, by the experiments of Shower and Biddulph, ie ix sherefore necessary to assume a second mechanism which locates the centre of the resonance region with a precision increasing with the duration of the stimulus. ts effect is indie cated in Fig. 2.5. The second mechanism acts as if there were Reteresce 26, Al Haney Furrcues: wisn ee ‘Speech #88 Hearing” (Macelan, Fig. 2.5.—The two mechanisms of pitch determination, fa second resonance curve, of a non-mechanical nature, wh after about 10 millisec detaches itself from the mechani resonance curve and continues to contract until, after abc 250 millisec, it covers only a few cycles per second, ‘Both mechanisms are essential for our hearing. ‘The first itself would probably enable us to understand speech, but o the second makes it possible to appreciate music. One mig be tempted 10 locate this second function in the brain, t ‘mechanisms of nerve conduction can te imagined which mig achieve the same effect. Perhaps the simplest assumption ‘that the conduction of stimuli in adjacent nerve fibres is to 801 ‘extent unstable, so that in an adjacent pair the more strony stimulated fibre will gradually suppress the conduction in less excited neighbour. The available evidence would Justify the suggestion that this is the actual mechanism; intention js only to show that what manifests itself ast ‘adjustable time-constant"” of the ear is not necessarily cons of some higher function of intelligence, In the light of these resutts we can now approach the questi of a conidensed transmission. which entirely deceives the © ‘The performance figure as shown in Fig. 2.4 appears to indice that considerable economy might be possible, especially in § range of higher frequencies. This is brought into evidence ev more clearly in Fig. 2.6, which contains the integrals over fi Fig. 2.6—Usilization of information acea, quency of the performance figures for the ear and for an ide insfrument, Between zero and 8000c/s, for instance, tl ‘maximum number of data which the ear ein appreciate is on about one-quarter of the data which can be transmitted in band of 8 000 c/s. -Tt is even likely that further investigatiot sight substantially reduce this figure. It may be remember that the experiments on which Fig. 2.6 is based have all bet carried out with sharp or tather angular waveforms; it ismGABOR: THEORY OF COMMUNICATION unlikely that the threshold was essentially determined by logons outside the area considered in our analysis. But it must also be remembered that the “adjustable time-constant™™ makes it very difficult to deceive the ear entirely. It will be shown in Part 3 that methods are possible which could deceive any non-ideal instrument with fixed time-constant. But the ear has the 1 ‘markable property that it can submit the material presented to it not only to one test, but, as it were, to several. Ultimately only direct tests can decide whether any such scherne will work satisfactorily (@ REFERENCES 2) Hower, Dupury: “Re-making Speech," Joaraal of the ‘Acoustical Society of America, 1939, Vi, p. 169. (2.2) Borex, W., Korowsx:, P., and Licute, H.: “Develop- Part 3, ‘SUMMARY cis suggested that it may be possible to transmit speech and mosic in much rarrower wavebands than was hitherto thought necessary, not by clipping the ends of the waveband, but by condensing the information. “Two possibilities of more economical transmission 2re discussed. Both have in common that the origina) waveband is compressed in transrnission and re-expanded to the original width in reception. In the fist or “kinematical” method a temporary oF permanent record is scanned by moving slits. or their equivalents ‘hich replace one another in continuous succession before a "window. Mathematical analysis is simplest if the transmission of the window is graded according to a probability function. A simple harmonic ‘oscillation is reproduced as a group of spectral lines with frequencies which have an approximately constant ratio to the origiaal frequency, ‘The averags departure from the law of proportional conversion inverse ratio fo the time interval in which the record passes before the window. Experiments carried out with simple apparatus indicate that speech can be compressed into a frequency band of 800 or even ‘500 c/s without losing much of its intelligibility. There are various ing frequency compression in telephony by means lecically, without mechanical motion. tially in using non-siausoidal cacriers, such a8 repeated probability ‘pulses, and local oscillators producing waves of the same type. It is ‘shown that one variety of the electrical method is mathematically ‘equivalent to the kinematical method of frequency conversion. (y aNTRopUCTION High‘idelity reproduction of speech or music’ by current methods requires a waveband’ of about 8 000e/s. It has been shown in Part 1 that this band-width is sufficient for the trans- mission of 16000 exact and independent nurrerical data per second. This high figure naturally suggests the question whether all of really needed for the human car to create an iMlusion of perfection. In Part 2 it was.shown that, even in the frequency range in which itis most sensitiv, the human ear ‘can appreciate only one datum in two at the Best, and not more than one in four as an average over the whole af. range. More- over, it must be taken into consideration that, im the experiments which gave these limits: of aural discrimination, attention: was fixed on a-very simple phenomenon, Tt appears highly probable ‘that for complex sound patterns che discriminating power of the cear is very much Jess, This evidence suggests that methods of transmitting and reproducing sound may be found which are ‘much more economical than those used at preseat, in which the ‘original signal shape is carefully conserved through all the links fof transmission or reproduction. In an econornieal method the information content must be condensed to a minimum before as ‘ment of Pitch Sensations,” Elektische Nachrichten. Technik, 1935, 12, p. 326 @3) Borex, W., Korowses, P., and Licire, H.: “Audibilty of Delays,” ibid, 1935, 12, p, 355. @4) SHowen, E! G., and Binputmd, R.: “Differential Pitch Sensitivity of the Est," Journal of the Acoustical Society of Amertea, 1931, 3, . 275 @.5) Bioci, A.:"*Modulation Theory," Jounal LEE, 1944, 91, Part ffl, p. 31. 2.6) Weaet, RL, and Lave, C. £.: “Anditory Masking and the Dynamics of the Inner Ear,” Physical Review, 1924, 23, p. 266, 2.7) Hox, R. S.: “Damping and Selectivity of the Inner Ear," ourral of the Acoustical Society of America, 1942, 14, p. 50. FREQUENCY. COMPRESSION. AND EXPANSION transmistion or before recording, and the reconstniction need ‘not take place before some siage in the receiver or reproducer, ‘There is no need for the signal to be intelligible at any inter- mediate stage, Economieal methods must therefore comprise some stage of “condensing” or “coding” and some stage of See oe sGenoding” z Dudley Homer's ingenious Vocoder;©! which transmits in- teligibte speech through 11 channels of osly 25¢/s each, i a wellknown eximple of such « system, Tt operates with a Inethod. of spectral analysis and synthesis, The spectrum of speech is roughly analysed into 10 bands of 250 js eacb, and the aggregate intensity in each band is transmitted throvgh a separate channel of 25c/s. ‘The transmitted intensity is used for modulatiog a buzzer at the reosiving end, which roughly reproduces the original spectrum. The eleventh channel is used for tansmiting the “pitch,” which is, broadly speaking, the frequency of the vocal cords, The Vocader in its present form bas probably very nearly reached the limit of tolerable com> pression. Ta this Part new methods will be discussed in which the coding of the message consists essentially in compression, ic. in a pro- portional reduction of the original frequencies, and the decoding in expansion to the original range. It is evident that neither compression nor expansion can be exact if economy is to be fected. If, for instance, al frequencies were exactly halved, ‘would mean that-it would take twice the time for trans- ritting the same message and there would be no saving, Com- pression and expansion—in general, “conversion”—of fe- ‘Quencies must be rather understood in an approximate seas: ‘Thece will be unavoidable departures from the simple lineal law, and hence there will be some unavoidable distortion. But it appears that these ean be kept within tolerable limits while Bill effecting appreciable waveband economy. on) “Two corapression-expansion systems will be described. The first which operates with mechanicelly eoving parts, will be called the, “kinematical” method,* while the second does not require Mechanical motion and will be called the “eletréal” ‘method. "So far, experiments have boen carried out only wit the kinematical method, and for this reason it will occupy most of this Part, (@) THE KINEMATICAL METHOD OF FREQUENCY ‘CONVERSION, It will be convenient to explain this method by means of a particular example before generalizing the underlying principte. Assume that the message to be condensed or expanded is recorded ‘+ mci Patet Apyeaton No. 46/44.46 as a sound track on a film. . For simplicity, assume that the coriginal signal is a simple harmonic oscillation, that is to say a frequency f—to be called the “original frequency””~is produ if the record moves with standard speed w past a stationary slit Tmaginte now that the slit itself is moving with some speed u, so that its speed relative to the flm is yu. The photocell behind the film now collects fluctuations of light of frequency @.) ‘This means that all fequencies in the record are converted in a constant ratio (v — w/v. There is evidently no gain, as.it would take'the moving slit »/(o— 1) times Jonger to explore 8 certain length of the film than if it were stationary. But ‘Jet us now imagine that the film moves across « fixed window, so that the moving slit is effective only during the time in which it traverses the window. In order to get a continuous record let-a second slit appear at or before the instant at which the first slit moves, ‘out of the window, after which a third sit would appear, and 30 on. The device is still not practicable, as evidenlly every sit would produce a loud crack’ at the instant at which it ap- peared before the window and when it left it. But now sssumme that the window has continuously graded transmission, ull in the middle and fading out at both sides to total opacity. In this arrangement the slits are faded in and out gradually, so that abrupt cracks can be avoided. Ths is the prototype of a kinematical frequency convertor, schematically ilustcated in Fig. 3.1, which will be investigated below. Though Fig. 3.1.—Frequency convertor with sound film. the nomenclature will be taken from this special example, the mathematical theory can be transferred bodily to any other realization of the same principle. In Fig. 3.1 the film is supposed to move ia close contact with the slotted dram, but at different speed. A photocell collects the sum of the light transmitted by the iodividual slits and by the window. To obtain its response we must first write down the conttibution of one slit and sum over the slits. All slits will be assumed to have negligible width. For simplicity let tus measure all distances x from the middle of the window and all Uimes ¢ from the instant in which a slit, to be called the “zero”-th slit, passes varough x= 0, The other slits will be Alistinguished by suffixes A, which increese in the direction in ‘which the film is moving. Their position at the time ¢ will be called 24, - The nomenclature is explained in Fig, 3.2. Lt v be the speed of the film, while the velocity of the slits will be called waa ee 62 ‘The reason for this notation is that ean. (3.1) now simplifies to Sym fe ive. & has the meaning of a frequency-conversion ratio. GABOR: THEORY: OF COMMUNICATION eo SES ce fe eceeees a eee ee ee cee ce cH bes eet ~ oe . Fig, 3.2.—Explanation of notations. AE the spacing of two slits is 5 the position of the kth slit time rit ven by md othe 2. ‘The record will be'characteriaed by the signal s() which would produce if it were scanned, in the ordinary way by stationary slit in the-position x= 0. Hence, if the wind: ‘were fully transparent, the signal dué to the Ath slit at time would be 30 alo) = Het hte) ‘The total reproduced signa, i.e. the light sum collected by ¢ photocell, is obtained from this by multiplying by the trar mission coefficient P(2) of the window and summing over k, Jn all the: following calculations we will assume that tt transmission follows a probability law. ‘This law has unig properties. in Fourier, analysis and will immensely simplify 0 investigations: Other laws wich appear equally simple a pri, and which may have even some practical advantages—such ‘triangular or trapezoidal windows-—lead to expressions. whic are too complicated for anything but umerical- discussio “Hence we assume PO) =exp— GIN 2. . G. Nis a mumber, t6 be called the “slit number,” which chard terizes the reproduction process. It is the mmber of slits the length over which the transmission of the window falls fro unity to Ie. The total length of the window in which transmission exceeds 1% is°4:3Ns. Thus we can say broad ‘hat the total number of slits simultaneously before the windo is 4-3N, The reproduced signal, ic. the total light collected by photocell, at time ¢, is 3 ) ee — GINGA = Hh). ©. This, in combination with eqn. 3.3, is a complete descriptic of the operation of the frequency convertor. It will now 1 iustated in the special case in which 5 is @ simple harmon oscillation SM m cis Inf ‘Tre complex form will be used, with the understanding that 11 real part constitutes the physical signal, Simple harmon oscillations are suitable for the analysis, as their spectrum wi ‘consist of a few lines. But it may be mentioned that analys in terms of the elementary signals discussed in Part 1 (harmon oscillations with probability envelope) can be carried out almo equally simply, as the reproduction of an elementary sign:GABOR: THEORY OF COMMUNICATION consists also in the sum of a few elementary signals. This is carried out in Appencix 7.1, but in the text only the more familiar method of Fourier analysis will be employed, Substituting the signaf (3.7) in eqn. (3.6) and using eq. (3.3), wwe objain Sexo [0 — a api eo — At 0 ‘The meaning of this somewhat complicated. expression is ex- plained in Fig. 3.3. Each, slit, as it passes before the window, {Yo - tp Vv Fig. 3.3,—The contributions of individual slits and the resulting Hight output. ‘transforms the sine wave into an clementary signal. By adding up the contributions of the individual slits we obtain for some frequencies a very neatly faithful reproduction, i.e. an almost pure tone but of differeat frequency from the original. For ‘other frequencies we obtain strong beats. ‘A-more convenient and complete description of the frequency conversion process is obtained by Fourier analysis. It will now ‘be convenient to measure distances in time intervals, and to introduce, instead of the sit spacing s, the time interval + between the passage of two consecutive slits before a fixed point cml Ww (3.9) With this notation the Fourier transform, i. the spectrum of the signal s(#), becomes, by known rules,* SUA) = Wat exp = (aNAAF — Hf)? “e cis Imkr(f — f) = @.100) This expression allows of a simple interpretation. The second factor Deeis Imkr —h) is the sum of an infinite numbér of complex vectors of sibie Jength, with an angle of 2ar(/— jf) between io consecutive vectors. This series, though not caswergent, is summable,t and its sum is zero for ail values of f except those for which af f)~aninteser . . 2. GH) Physically this means that the spectrum consists of sharp Lines which differ from one another by multiples of 1/r. In other * The gloaletion caved out ia Appeodin 7.1, The ue fot Fourier terafarmt ay be load is eteraoe £3, ted) parenliely Gaels othe pe 0, "Sommation a fo be andesood ig the see of Cabo. Cf, Warrtatcn- ‘wheats Modern Avail 3b. 135. ia 47 ‘words, the spectrum consisis of all combination notes of the original frequency f with the repetition frequency 1/7. ‘The absolute sharpness of the spectral lines is a consequence of the assumption that the slits pass before the window at ‘mathematically exact equal intervals. In each spectral line S(f) is a “delta function,” i.e. @ sharp peak of infinite height but finite area, But as in what follows we shall always bave to deal ‘with line spectra, it is more convenient to re-iacerpret S(f) asa function which is zero except at certain discrete values of f, where it assumes Gaite values, proportional to the amplitude Of the spectral Tines. In the same sense, we write the second factor of eqn. (3.10<) somewhat more simply as Beis zai f— fy) = DBF — fk)» B12) and interpret this as a “selecting factor” which bas zero value everywhere except for those values of f which fulfi condition - 3.4. —The selection factor. (3.11), where it assumes the value unity (see Fig. 34). Thus vwe write eqn. (3.10) SUA) = exp — GHNTAS — RAPER — fo— Kir) 3.100) “The first factor is independent of the summation index 4 and represents an attenuation function of probability shape, which has its maximum at 7 = he ie. at frequencies’ which have been converted in the correct ratio x. The sharpness of this attenuation curve is reciprocal to the sharpnéss of the transmission curve of the window, measured in units of time, Thus, if the window were infinitely broad we should obtain exact conversion of all frequencies. ‘But this would have the disadvantage that short signals occurring at some definite time would be reproduced at completely in- Gefinite times (with an infinite numberof repetitions). Con versely, if the window were infinitely short the attenuation would ‘be zero and the frequencies scattered evenly over all possible values defined by eon, G.11). Thus we mect again-the fundamental uncertainty relation between frequency and time (or rather, “epoch”) which was discussed in some detail in Pare 1. It follows immediately from previously obtained results that the probability window is ideal in the sense that it produces the smallest possible product of the linked uncertainties of frequency and epoch, as defined in Part 1.* Nevertheless the probability window is not necessarily the best from @ practical point of view. “Some possible improvements will be discussed later. Equation (3.102) or (3.106),allows also a simple graphical interpretation, which is explained in Fig. 3.5 in a oumerical example, The original frequency fy és the ordinate; the repro- duced frequencies f are the abscissae. Both are conveniently ‘measured in unis 1/r, ic. a6 multiples of the repetition fre- ‘quency. All poiats (//4) which satisfy condition (3.11) lie 02 lines at 45° to the two axes, and intersect the horizontal axis at integral values of fr. Thé attenuation curve exp — (NAS ~ nf need be‘drawn only once, though in the Figure it has been done twice inorder to give a clear visual impression of the way i, acai nentins was dtd, sur om ontan ator, wens448 ag. 3.5.—Ditora of requeney compresion Which the amplicude is distributed over the (ff) plane. The spectral lines are given by the height of the attenuation curve above the points in which a line fg = constant crosses the lines Uf = far = integer, as shown in an example, ‘This Figore shows the action of the frequency convertor at ‘one glance. The correctly converted frequency f= x fo appears in the reproduction only where a line ( —f)r = an integer inter- sects the-line f= fy This condition is always fulilled. for ‘fo= 0, and for all frequencies which are multiples of er an G13 ‘This may be called the length of the “cycle of reproduction”, as, the quality of reproduction varies cyclically with this period. 1f fis an integral multiple of F the reproduction can be made almost perfect, asthe side lines can be almost entirely suppressed if the slit number NV is made sufficiently large. As can be seen in Fig. 3.6, N= 1 is sufficient to achieve this. BUt this improve- Fig. 3.6.—Diagram of frequency expansion. Sled ‘ment in the reproduction of certain tones is made at the cost of others. “If Nis large, not only the side lines but almost all tudes near the middle ‘of a cycle of reproduction will be suppressed, ie. certain notes will be missing. It is evident that ‘compromise must be struck between the purity of reproduction at the ends and at the middle of every cycle length F: ‘The effect of the slit number N on the quality of the sepro= GABOR: THEORY OF COMMUNICATION Te WT tT Fig, 3.7-—influence of sit mimibar on aualiy of reproduction. duction is shown in Fig, 3.7. Three cases are illustrated, a for an expansion ratio «= 2, and for'N = 0°25,'0-5 and 1 It may be recalled that the average number of slits before the ppart of the window in which the transmission exceeds 1% i 4:3N: Tn each case a full cycle of reproduction is shown, wi ‘ten equally-spaced original frequencies. At the left, N= 0°25, the Figure shows the effect of 16 ‘small slit numbers. The reproduction is very “noisy,” no fre ‘quency being reproduced as an approximately pure tone. Ther litle difference between the spectra of frequencies near th middle or ends of the cycles they are all of uniformly poo quality. ‘At the right, NV = 1, the Figure shows the effect of a too larg number of slits (cf, Fig. 3.6). The frequencies at the ends © the cycle are reproduced nearly ideally, as practically pure tone: but the frequencies in the middle of the cycie are almost entirel missing in the reproduction. ‘The best compromise appears to be N= 0°S, shown in th middle of the Figure. ‘The end frequencies are still reproduced a ‘almost pure tones, and the intensity falls off little towards th le of the cycle. (The intensity is obtained by squaring th amplitudes shown in the Figure and finding their sum. It falls in the middle of the eycle, 19 0°56 of the maximum.) ‘The spectr Of the intermediate tones consists mostly of only two lines; is these will be vibrating tones, vibrating with a beat frequency 0 Afr. The beats are strongest in the middle, where the tw spectral components have equal amplitudes. It might appear at first sight that, by reducing the beat fre quency below any limit, the reproduction could be made perfec to any desired degree. "But there are limits to the increase of + As Wis fixed more or less at 0-4-0-5, 7 can be increased onl by making the window longer. “The leagth of the window ma, ‘be now defined as the length of time J in which a point of th film passes through the part of the window in which the trans mission exceeds 1% is T=4-3Nyoe 43NT1~W) 2. Gul Hence, for the optimum, N= 0-5 T= 0-477 ~ x) If the time 7 is too Jong, the time resolution in the repre action willbe poor. “Determining the best compromise betwee time resolution and frequency reproduction is a matte: fo experiment. On general grounds one would expect that th ‘window length 7 must be kept below the limit at which the ea could besin to separate the contribution of the two or mor slits which are simultaneously before the window. For speec the optimum of T is probably about 100 millsec; for musiGABOR: THEORY OF COMMUNICATION probatly about 250 millisec.* With x ~ 2 this would make the beat frequency 21 ¢/s for speech, and about 8 os for music. _ It may be noted by comparing eqns. (3.13) and (3.14) that a simple reciprocity relation obtains between the cycle length F and the window length 7, of the form FT=43N . G15) With optimum choice of 1 the valoe of this is about 2. Thus for a window length of 100 millise: the optimally-reproduced frequencies are spaced by about 20./s; for T= 250 millisec by about 8¢/s. In the reproduction the spacing will be « times more. ‘The theory so far discussed was based on the assumption of @ probability window, which not only has the advantage of mathematical simplicity, but also gives the most advantageous reciprocity relation between time resolution and frequency reso- tution, “But the optimum mumber N was fouid to be only ,, about 0-5, which means that there are on the average only about two slits before the window. This might produce a slight but noticeable noise in the optimally-reproduced frequencies, in particular for f= 0 (background), Hence it may be advan- tageous to depart somewhat from the probability shape ia order to suppress the noite. Fig. 3.8 shows window transmission Dea ae Fig. 3.8.—Window shapes with zero noise for two and three slits. shapes for two and three slits which produce no noise when passing before aa even background, as the light sum is constant in any position. Though the mathematical theory of such ‘windows is very much more complicated, itis not to be expected. that they would produce essentially different results from proba- bility windows of comparable effective width, @) DISTORTIONS RESULTING FROM THE COMPRESSION- EXPANSION CYCLE ‘A full eycle of condensed transmission of the kind discussed consists in compression by a factor x < 1, folloned by expansion in the ratio ix. In general, if two conversion processes ate applied in succession to a simple harmonic oseilation of fre- Gency fy the resulting spectrum is given by sf) =k exp — {rN yr [fell — ep) + Ky + NYS — {fo + ke) PY BY — fo — Kir, — mit) - G16) ‘The derivation is given in Appendix 7.2. All data N, 1, x of the first conversion have been given a suffix 1, those of the second conversion the suffix 2. & and m are summation indexes which ‘run overall integral values. ‘The second factor is again a selection operator, which is zero. for al values off with the exception of those where Lahot kin + mig . . Gay Tis means that only those equenes wil appeae ia the ‘spectrum which correspond to combination tones of the original ade eee thc opment ce ‘ot 9, Pa “49 frequency with one or the other or both of the repetition fre- quencies Ifry and I/ry. These form, in general, x double series, Which in the particularly important practical examples to be Considered reduces 0 a simple series. Tn what follows we will consider only pairs of conversion processes which, on the averege, reconstruct the origiaal fre- quencies. The condition for this is keel. G18) ‘The ambiguity of sign expresses the fact that positive and nega- tive frequencies are equivalent. But only the plus sign will be considered, and it will be assumed, rooreover, that both x and ‘eq are Positive. Negative conversion ratios are less advanta- ‘geous, a8 fora given window length they requie higher repetition frequencies (Eqn. (3.14)], The whole compression-expansion cycle will be characterized by the compression ratio x,0.< x <1, and the expansion ratio will be assumed as I/x. ‘To simplify the discussion it will be assumed that the window length T'is the same in the transmitter and in the receiver. ‘This corresponds to optimum conditions, as it wil evidently be best to operate at both ends with the longest permissible 7, which may have different values for speech and for music. This ‘means TH 3— Nil —= Nei ole . B19) or ma NaleN, + G20) A second simplifying assumption will be tytz= p= aninteger . . . . G21) ‘This again is an assumption which is fulfiled in the most ii portant practical cases. In the interest of optimum transmission ‘the slit number will be used, in both the transmitter and the receiver, which gives the best results in simple conversion (N= 0-4 — 0°65), and if 1 is the reciprocal of an integer 4, 4.4... « the condition (3.21) will be fulfilled. Mathematically this has the advantage that the double series of frequencies in the reproduced spectrum bry + ming row becomes a simple series, with period ifr as in simple conversion. We write Kir, + mity = e+ pny = nr, 322) so that the spectral lines are now characterized by the single suffix n, which can be called the “order mumber.” As S(/) will be different from zero for integer values of n, and for these only, wwe can now omit the selection operator 8 in eqn. (3.16), on tbe Understanding that we consider only integral values of 2. qa. (.17) now becomes Safin, 6. G2) Eliminating f by means of eqn, (3.23) and introducing the assumptions (3.20) and @.21) into eqn. (3.16) we now. obtain the simplified formula Son) = Tk exp — CANAL for — 6) + KP 4 Ll) +k mR}. 624) In this sum, however, not all integral values of & are included, but only those which are compatible with the given value of the order n. Ifthere are fo valuss ke mg which satisfy the equation nm gt map all other values which sais it must be of the form * k= ky top mam? 2%450 where v is any integer. It will therefore be- convenient to introduce v as the summation index, and to make the convention that ky is the smallest positive number in the sequence of k's. In other words, let k be the residue of m divided by p, of, in the notation of the elementary theory of numbers, 1 = ke(mod p) soe . G25) ‘As a further simplification we note that Stun) is a periodic function of f, with a cycle length 7 ‘ Fo sgaa =e G29 ANT (Sy ko) =) [é@ +B+) + (B+ B+-%)] © 020 By rearranging the terias ia the exponent this can be written, finally, Shon = exp 4(28) Sy, aNg\*( fo ko me? Drew -2(28) Geter re 2). can This formula lends itself well to graphical interpretation, In. Fig. 3.9 the ordinate is again the original frequency fg, measured Fig. 3.9—Explanation of frequency-conversion diagrams. in units, and the abstisiae are the reproduced fiequencies f ‘A line at 45° through the origin represents the correct recon version law, f-= J This is the line of zero order, n= 0. Parallel to this we draw lines through all multiples of I/r, on the faxis. These are the loci of all non-zero intensities. It we Imagine the amplitude Sif) asa surface above the (fy) plane, this surface consists of a number of profiled planes, projecting above the lines n = constant. On the line n= 0 we have evidenty a maximum of S(f0) for every integral value of fF These may be called the “prin- cipal maxima." At the side lies of higher order there will also be maxima, but because of the probability function in front of the sum these will be smaller, "We can dravr lines connecting, these maxima of different orders. We obtain a set of straight ‘connecting the points where SIF + kelp ~ mxf2p= an inseger 3.29) AE the order 1 increases by one, by eqn. (3.24) kp also increases bby unity, and fof changes by <(-w . 330) GABOR: THEORY OF COMMUNICATION as shown in Fig. 39. Tt ean be shown from the geometry of Fig, 3.9 that these lines wilf iotesect the hortzontal axis at multiples of Pith = 48) 31) ‘These lines, together with the lines n= integer, form a network with intersections at ‘every maximum of the spectral function Shon). ‘Along each lige m= constant, the spectral amplitude is the same function of fF, apart from the shift (3,30) and the factor exp [— HaNahp)*a4] which varies with m but is a constant along each Tine, Thus it is suficient to compute the amplitude function once, for m= 0, where the shift is zero and the exponential factor unity. This faction is SU) =En exp ~ UaN JOE +P. G32) This, as a function of fF, is the sum of probability functions, recurring at unit distance, Tt is shown in Appendix 7.3 that it can be reduced to a recognized transcendental function of analysis, the theta function 69, Fi . 3.10 shows this function ig. 3.10.—The function Sf, 0). for two values of the parameter Naf. In the cases which are Of practical interest Naf is equal to or larger than unity, and the probability functions become so sharp that their overiap is negligible, and (.32) consists of recurring peaks of probability shape. Its now possible to construct diagrams, which may be called frequency reconversion diagrams, which show the reproduced spectrum of any pure original tone in the same way as the previous simple conversion diagrams. Fig. 3.11 is a first example of such a diagram, with «=; ie. the cycle consists ia compression to one-half, followed by expansion 10 the original range. The slit numbers are assumed as N, = Ny = 4, which was previously found to represeot the most advantageous com- promise. The diagram can be considered as three-dimensional, with the profiles of the S-function at right angles to the Ug) plane, The amplitudes are plotted in the direction fg, $0 that the spectrum corresponding fo, any original frequency /o can be immediately constructed by drawing a horizontal line and plotting the heights of the S:function at the intersections with the lines of constant order. ‘This is carried out for a full eycle of reproduction in Fig. 3.12, which may be compared with Fig. 3.7 (central figure) illustrating the result of the expansion, starting from an undistorted record. Te must be noted that + in Fig. 37 corresponds to 7 in Fig. 3.12, and as 7, = 2r, the minimum interval between two frequencies in the spectrum in Fig, 3.12 is half of that in Fig. 3.7. If this is borne ia mind, it can be scen immediately that the difference ‘between the two cases is mainly that the two side-lnes in Fig. 3.7 have now split up into two fines each (with some insignificant satellites), and the centre of gravity of these two lines follows very nearly the same course as in Fig. 3.7. But it has been showa before that with «= d, 1/7, can be made so small that the ear can hardly, if at all, distinguish between the two tones.452 fence as compared with the case «= 4 is quantitative, The ‘eat Grequency between the doublets is now about 4/7, twice as large as before. If in Fig, 3.14 the doublets are imagined as merged into one, the lines connecting them will be almost vertical, Thus we can, interpret the operation of the frequency reconvertor in a some- what different way. It acts very nearly like a musical instrument with a discrete set of frequencies, which tries to imitate speech or music as closely as possible with a limited nuraber of tones. It is well known that if a vowel is sung into an open piano with the loud pedal depressed it will echo the vowel very. clearly, The frequency roconyestor performs a similar imitation, but with the difference that its fixed frequencies are set at equal arithmetical, not geometrical, intervals, Hence the reproduction will tend 9 become more perfect at higher frequencies. At fower frequencies there must necessarily be departures from perfect reproduction. This becomes evident if itis remembered that the frequency convertor does not change the rhythm or ““time-pattern" of speech or music. Tn frequency language this means. that frequencies well below the audible range are repro- ‘duced almost with the original value, whatever the value of x. Summing up, we can say drat a frequency compressor and an expander operating in succession produce as close a reproduction of the original as is compatible with the uncertainty relation, and the limit is set almost entirely by the expansion, the errors Introduced by the compression being relatively smal. @) PROVISIONAL REPORT ON EXPERIMENTAL WORK ‘Theory can give a complete description of the operation of the frequenty convertor either in time language, or in frequency language, or in the more general representation discussed in previous communications, but it does not enable us to draw ‘conclusions on the quality of the reproduction. In order to subject the theory to a fist rough test, a 16-mm ‘sound-film projector was converted by few simple modifica- tions into a frequency convertor. Fig. 3.15 {s a photograph of the essential parts, and Fig. 3.16 is a schematic illustration of ‘the optical arrangement. "The usual single, stationary slit of the sound head was replaced by a slolted drum which rotated round an axis passing through the filament of the exciter Jamp. The drum was of 0-00S-in steel tape, and the width of the slits was also about 0-005 in. ‘The condenser lens was replaced by as large a lens as the fitting would take, with a free diameter of about 1 inch. Immediately in froot of the slotted drum a frame was arranged for the “‘window.'” In the case of films with variable-area sound tracks this was a film with graded transmission, produced by & graphic process or sprayed with an aitbrush. For. variable- density films the window was cut out of black fllm or paper to the desired shape. The window and the slits behind it were ‘imaged on the film by the same microscope objective as used in ordinary operation, which reduoed their image to about one~ quarter. Thus, allowing for optical errors, the effective slit- ‘widih was 0-0015-0-002 in, The maximum length, 7, of the ‘window which could be utilized was limited both by the diameter of the condenser lens and by the collecting system which guides the collected light to the photocell, Measured on the film it ‘was about 6mm. Sound-flm moves at the standard speed of 483 mm/sec; thus the maximum 7 was about 32 millisec. By running the flm on the “silea” setting, at about 125 mmyseo, this could be increased to about 48 millisec. The shortness of these times was a severe Jimitalion of the apparatus, ‘The improvement between 32 and 48 millisec was so marked that it appears to confirm the expectation that the optimum Tis con- siderably longer, probably 100 millise, perhaps even more. ‘The slotted drum had a slepped pulley attached to it which GABOR: THEORY OF COMMUNICATION Fie, 3.15—16mm sound. projector converted into an exper ‘mental frequency convertor. Optical arrangement in frequency convertor. could be driven at different speeds by means of a spring belt from another. stepped pulley atachel to a sprocket of the Projector. By crossing the belt the motion could be reversed. ‘The following values of x were tried:— = 0-25 033 O42 1:5 F-75 2-0 3-0 3-33 Tt became evident in the first experiments that the window length of 32 millisee was insufficient for the reproduction of music, hence the later tests were mosily restricted to the repro- duction of speech. The uneven rotation of the drum due to the elasticity of the spring belt was also much less objectionable with speect than with nmusic. Male speech remains completely intelligible with « = 1-5, ic. if the frequencies are raised by 50%, though a baritone changes into a high tenor. The intelligibility falls

You might also like

• The Subtle Art of Not Giving a F*ck: A Counterintuitive Approach to Living a Good Life

  

  From Everand

  The Subtle Art of Not Giving a F*ck: A Counterintuitive Approach to Living a Good Life

  Mark Manson

  Rating: 4 out of 5 stars

  4/5 (5820)
• The Gifts of Imperfection: Let Go of Who You Think You're Supposed to Be and Embrace Who You Are

  

  From Everand

  The Gifts of Imperfection: Let Go of Who You Think You're Supposed to Be and Embrace Who You Are

  Brené Brown

  Rating: 4 out of 5 stars

  4/5 (1093)
• Never Split the Difference: Negotiating As If Your Life Depended On It

  

  From Everand

  Never Split the Difference: Negotiating As If Your Life Depended On It

  Chris Voss

  Rating: 4.5 out of 5 stars

  4.5/5 (852)
• • Magazines
  • Podcasts
  • Sheet music
• Principles: Life and Work

  

  From Everand

  Principles: Life and Work

  Ray Dalio

  Rating: 4 out of 5 stars

  4/5 (609)
• The Glass Castle: A Memoir

  

  From Everand

  The Glass Castle: A Memoir

  Jeannette Walls

  Rating: 4.5 out of 5 stars

  4.5/5 (1717)
• Grit: The Power of Passion and Perseverance

  

  From Everand

  Grit: The Power of Passion and Perseverance

  Angela Duckworth

  Rating: 4 out of 5 stars

  4/5 (590)
• Hidden Figures: The American Dream and the Untold Story of the Black Women Mathematicians Who Helped Win the Space Race

  

  From Everand

  Hidden Figures: The American Dream and the Untold Story of the Black Women Mathematicians Who Helped Win the Space Race

  Margot Lee Shetterly

  Rating: 4 out of 5 stars

  4/5 (898)
• Sing, Unburied, Sing: A Novel

  

  From Everand

  Sing, Unburied, Sing: A Novel

  Jesmyn Ward

  Rating: 4 out of 5 stars

  4/5 (1104)
• Shoe Dog: A Memoir by the Creator of Nike

  

  From Everand

  Shoe Dog: A Memoir by the Creator of Nike

  Phil Knight

  Rating: 4.5 out of 5 stars

  4.5/5 (540)
• The Perks of Being a Wallflower

  

  From Everand

  The Perks of Being a Wallflower

  Stephen Chbosky

  Rating: 4.5 out of 5 stars

  4.5/5 (2104)
• The Hard Thing About Hard Things: Building a Business When There Are No Easy Answers

  

  From Everand

  The Hard Thing About Hard Things: Building a Business When There Are No Easy Answers

  Ben Horowitz

  Rating: 4.5 out of 5 stars

  4.5/5 (349)
• Bad Feminist: Essays

  

  From Everand

  Bad Feminist: Essays

  Roxane Gay

  Rating: 4 out of 5 stars

  4/5 (1018)
• Elon Musk: Tesla, SpaceX, and the Quest for a Fantastic Future

  

  From Everand

  Elon Musk: Tesla, SpaceX, and the Quest for a Fantastic Future

  Ashlee Vance

  Rating: 4.5 out of 5 stars

  4.5/5 (474)
• The Outsider: A Novel

  

  From Everand

  The Outsider: A Novel

  Stephen King

  Rating: 4 out of 5 stars

  4/5 (1867)
• Her Body and Other Parties: Stories

  

  From Everand

  Her Body and Other Parties: Stories

  Carmen Maria Machado

  Rating: 4 out of 5 stars

  4/5 (822)
• The Sympathizer: A Novel (Pulitzer Prize for Fiction)

  

  From Everand

  The Sympathizer: A Novel (Pulitzer Prize for Fiction)

  Viet Thanh Nguyen

  Rating: 4.5 out of 5 stars

  4.5/5 (122)
• Angela's Ashes: A Memoir

  

  From Everand

  Angela's Ashes: A Memoir

  Frank McCourt

  Rating: 4.5 out of 5 stars

  4.5/5 (441)
• The Emperor of All Maladies: A Biography of Cancer

  

  From Everand

  The Emperor of All Maladies: A Biography of Cancer

  Siddhartha Mukherjee

  Rating: 4.5 out of 5 stars

  4.5/5 (271)
• Brooklyn: A Novel

  

  From Everand

  Brooklyn: A Novel

  Colm Tóibín

  Rating: 3.5 out of 5 stars

  3.5/5 (1947)
• The Little Book of Hygge: Danish Secrets to Happy Living

  

  From Everand

  The Little Book of Hygge: Danish Secrets to Happy Living

  Meik Wiking

  Rating: 3.5 out of 5 stars

  3.5/5 (403)
• A Man Called Ove: A Novel

  

  From Everand

  A Man Called Ove: A Novel

  Fredrik Backman

  Rating: 4.5 out of 5 stars

  4.5/5 (4771)
• The World Is Flat 3.0: A Brief History of the Twenty-first Century

  

  From Everand

  The World Is Flat 3.0: A Brief History of the Twenty-first Century

  Thomas L. Friedman

  Rating: 3.5 out of 5 stars

  3.5/5 (2259)
• Steve Jobs

  

  From Everand

  Steve Jobs

  Walter Isaacson

  Rating: 4.5 out of 5 stars

  4.5/5 (808)
• Devil in the Grove: Thurgood Marshall, the Groveland Boys, and the Dawn of a New America

  

  From Everand

  Devil in the Grove: Thurgood Marshall, the Groveland Boys, and the Dawn of a New America

  Gilbert King

  Rating: 4.5 out of 5 stars

  4.5/5 (266)
• The Art of Racing in the Rain: A Novel

  

  From Everand

  The Art of Racing in the Rain: A Novel

  Garth Stein

  Rating: 4 out of 5 stars

  4/5 (4208)
• A Tree Grows in Brooklyn

  

  From Everand

  A Tree Grows in Brooklyn

  Betty Smith

  Rating: 4.5 out of 5 stars

  4.5/5 (1929)
• The Yellow House: A Memoir (2019 National Book Award Winner)

  

  From Everand

  The Yellow House: A Memoir (2019 National Book Award Winner)

  Sarah M. Broom

  Rating: 4 out of 5 stars

  4/5 (98)
• A Heartbreaking Work Of Staggering Genius: A Memoir Based on a True Story

  

  From Everand

  A Heartbreaking Work Of Staggering Genius: A Memoir Based on a True Story

  Dave Eggers

  Rating: 3.5 out of 5 stars

  3.5/5 (231)
• Yes Please

  

  From Everand

  Yes Please

  Amy Poehler

  Rating: 4 out of 5 stars

  4/5 (1903)
• Team of Rivals: The Political Genius of Abraham Lincoln

  

  From Everand

  Team of Rivals: The Political Genius of Abraham Lincoln

  Doris Kearns Goodwin

  Rating: 4.5 out of 5 stars

  4.5/5 (234)
• The Woman in Cabin 10

  

  From Everand

  The Woman in Cabin 10

  Ruth Ware

  Rating: 3.5 out of 5 stars

  3.5/5 (2522)
• Fear: Trump in the White House

  

  From Everand

  Fear: Trump in the White House

  Bob Woodward

  Rating: 3.5 out of 5 stars

  3.5/5 (738)
• Wolf Hall: A Novel

  

  From Everand

  Wolf Hall: A Novel

  Hilary Mantel

  Rating: 4 out of 5 stars

  4/5 (3973)
• John Adams

  

  From Everand

  John Adams

  David McCullough

  Rating: 4.5 out of 5 stars

  4.5/5 (2409)
• On Fire: The (Burning) Case for a Green New Deal

  

  From Everand

  On Fire: The (Burning) Case for a Green New Deal

  Naomi Klein

  Rating: 4 out of 5 stars

  4/5 (74)
• The Light Between Oceans: A Novel

  

  From Everand

  The Light Between Oceans: A Novel

  M.L. Stedman

  Rating: 4.5 out of 5 stars

  4.5/5 (789)
• Manhattan Beach: A Novel

  

  From Everand

  Manhattan Beach: A Novel

  Jennifer Egan

  Rating: 3.5 out of 5 stars

  3.5/5 (792)
• The Constant Gardener: A Novel

  

  From Everand

  The Constant Gardener: A Novel

  John le Carré

  Rating: 3.5 out of 5 stars

  3.5/5 (104)
• The Unwinding: An Inner History of the New America

  

  From Everand

  The Unwinding: An Inner History of the New America

  George Packer

  Rating: 4 out of 5 stars

  4/5 (45)
• Rise of ISIS: A Threat We Can't Ignore

  

  From Everand

  Rise of ISIS: A Threat We Can't Ignore

  Jay Sekulow

  Rating: 3.5 out of 5 stars

  3.5/5 (137)
• Little Women

  

  From Everand

  Little Women

  Louisa May Alcott

  Rating: 4 out of 5 stars

  4/5 (105)