AU621536B2 - Partial response channel signaling systems - Google Patents

Description

C NN COMN WEA A LTRTLH OF AUSTRALIA PATENT ACT 1952 COMPLETE SPECIFICATION 621 '~36

(ORIGINAL)

FOR OFFICE USE CLASS INT. CLASS Application Number: Lodged: Comtplete Specificatio' Lodged: Accepted: Published: Priority: Related Art-: te C C NAME OF APPLICANT: CODEX CORPORATION ADDRESS OF APPLICANT: 20 Cabot Boulevard, Mansfield, Massachusetts 02048, UNITED STATES OF AMERICA.

NAME(S) OF INVENTOR(S) David FORNEY5;)Vedat M. EYUBOGLU ADDRESS FOR SERVICE: DAVIES COLLISON, Patent Attorneys 1 Little Collins Street, Melbourne, 3000.

COMPLETE SPECIFICATION FOR THE INVENTION ENTITLED.- "PARTIAL RESPONSE CHANNEL SIGNALING SYSTEMS" The following statement Is a full description of this invention, including the best method of performing it known to us

-I-

1 1A- 1242V Partial Response Channel Signaling Systems Background of the Invention This invention relates to modulation coding and partial response systems.

In modulation coding, symbols are encoded as signals drawn from a constellation in such a way that only certain sequences of signals are possible.

In recent years, a number of kinds of trellis-type modulation codes have been developed and applied in modems) to realize coding gains of 3 to 6 dB over high-signal-to-noise-ratio, band-limited channels such as voice grade telephone channels.

Early trellis codes were due to Ungerboeck (Cjsaka et al., U.S. Patent No. 3,877,768; Ungerboeck, "Channel Coding with Multilevel/Phase Signals," IEEE e* Transactions on Information Theory, Vol. IT-28, pp.

55-67, January, 1982). Ungerboeck's codes for sending n bits per symbol are based on 4-subset or 8-subset partitions of one-dimensional (PAM) or two-dimensional (QAM) 2n+ -point signal constellations, combined with a rate-1/2 or rate-2/3 linear binary convolutional code that determines a sequence of subsets. A further set of "uncoded" bits then determines which signal points within the specified subsets are actually sent. The 1 partition and the code are designed to guarantee a U l certain minimum squared distance d 2 between min permissible sequences of signal points. Even after l giving effect to the power cost of,,an expanded signal constellation (a factor of four (6 dB) in one dimension, or a factor of two (3 dB) in two dimensions), the increase in minimum squared distance yields a coding 4~l 7 2

SI:

V t I' i gain that ranges from about a factor of two (3 dB) for simple codes up to a factor of four (6 dB) for the most complicated codes, for values of n that may be as large as desired.

Gallager Patent iication 577,34, filed February 6, 1984, discussed in Forney et al., "Efficient Modulation for Band-Limited Channels," IEEE J. Select. Areas Commun., Vol. SAC-2, pp. 632-647, 1984) devised a multidimensional trellis code based on a 16-subset partition of a four-dimensional signal constellation, combined with a rate-3/4 convolutional code. The four-dimensional subset is determined by selecting a pair of two-dimensional subsets, and the points of the four-dimensional signal constellation are made up of pairs of points from a two-dimensional signal constellation. With only an 8-state code, a d 2 of four times the uncoded minimum sequence distance can be obtained, while the loss due to expanding the signal constellation can be reduced to about a factor of 21/2 (1.5 dB), yielding a net coding gain of the order of dB. A similar code was designed by Calderbank and Sloane ("Four-dimensional Modulation With An Eight-State Trellis Code", AT&T Tech. Vol. 64, pp. 1005-1018, 1985; U.S. Patent No. 4,581,601).

I

Wei Patent ppli 7 3..

filed April 25, 1985) devised a number of multidimensional codes based on partitions of constellations in four, eight, and sixteen dimensions, combined with rate-(n-l)/n convolutional codes. His multidimensional constellationg again consist of sequences of points from two-dimensional constituent constellations. The codes are designed to minimize two-dimensional constellation expansion, to obtain vmrm ~aLI^- r r 3performance (coding gain) versus code complexity over a broad range, and for other advantages such as transparency to phase rotations. Calderbank and Sloane, "New Trellis Codes", IEEE Trans. Inf. Theory, to appear March, 1987; "An Eight-dimensional Trellis Code," Proc.

IEEE, Vol. 74, pp. 757-759, 1986 have also devised a variety of multidimensional trellis codes, generally with similar performance versus complexity, more constellation expansion, but in some cases fewer states.

All of the above codes are designed for channels in which the principal impairment (apart from phase rotations) is noise, and in particular for channels with no intersymbol interference. The implicit assumption is that any intersymbol interference 15 introduced by the actual channel will be reduced to a negligible level by transmit and receive filters; or, more specifically, by an adaptive linear equalizer in the receiver. Such a system is known to work well if the actual channel does not have severe attenuation within the transmission bandwidth, but in the case of severe attenuation ("nulls" or "near nulls") the noise power may be strongly amplified in the equalizer ("noise enhancement").

A well-known technique for avoiding such "noise enhancement" is to design the signaling system for controlled intersymbol interference rather than no 0* 0: intersymbol interference. The best-known schemes of this type are called "partial response" signaling schemes (Forney, "Maximum Likelihood Sequence Estimation of Digital Sequences in the Presence of Intersymbol Interference," IEEE Trans. Inform. Theory, Vol. IT-18, pp. 363-378, 1972).

4 In a typical (one-dimensional) partial response scheme, the desired output yk at the receiver is designed to be the difference of two successive inputs x k yk xk xk-l, rather than yk x k In sampled-data notation using the delay operaror D, this means that the desired output sequence y(D) equals rather than this is thus called a partial response system. Because the spectrum of a discrete-time channel with impulse response 1-D has a null at zero frequency the combination of the transmit and receive filters with the actual channel likewise must have a DC null to achieve this desired response. On a channel which has a null or a near null at DC, a receive equalizer designed for a 1-D desired 15 response will introduce less noise enhancement than one designed to produce a perfect (no intersymbol interference) response.

O:Uh Partial response signaling is also used to achieve other objectives, such as reducing sensitivity to channel impairments near the band edge, easing filtering requirements, allowing for pilot tones at the a6 band edge, or reducing adjacent-channel interference in frequency-division multiplexed systems.

Other types of partial response systems include a l+D system which has a null at the Nyquist band edge, and a 1-D 2 system which has nulls at both DC and the Nyquist band edge. A quadrature (two-dimensional) partial response system (QPRS) can be modeled as having a two-dimensional complex input; the (complex) response 1+D results in a QPRS-.ystem which has nulls at both the upper and lower band edges in a carrier-modulated (QAM) bandpass system. All of these partial response systems are closely related to one another, and schemes for one 111 i

S

are easily adapted to another, so one can design a system for the 1-D response, say, and easily extend it to the others.

Calderbank, Lee, and Mazo ("Baseband Trellis Codes with A Spectral Null at Zero"; submitted to IEE Trans. Inf. Theory) have proposed a scheme to construct trellis-coded sequences that have spectral nulls, particularly at DC, a problem that is related to the design of partial response systems, even though its objectives are in general somewhat different.

Calderbank et al. have adapted known multidimensional trellis codes with multidimensional signal .O"O constellations to produce signal sequences with spectral nulls by the following technique. The multidimensional signal constellation has twice as many signal points as are necessary for the non-partial-response case, and is D* divided into two equal size disjoint subsets, one of multidimensional signal points whose sum of coordinates is less than or equal to zero, the other whose sum is greater than or equal to zero. A "running digital sum" (RDS) of coordinates, initially set to zero, is adjusted for each selected multidimensional signal point by the sum of its coordinates. If the current RDS is nonnegative, then the current signal point is chosen from the signal subset whose coordinate sums are less than or equal to zero; if the RDS is negative, then the current signal point is chosen from the other subset.

In this way the RDS is kept bounded in a narrow range Snear zero, which is known to force the signal sequence to have a spectral,,null at DC. At the same time, however, the signal points are otherwise chosen from the subsets in the same way as they would have been in a non-partial-response system: the expanded

I

cr -6multidimensional constellation is divided into a certain number of subsets with favourable distance properties, and a rate-(n-1)/n convolutional code determines a sequence of the subsets such that the minimum squared distance between sequences is guaranteed to be at least d 2 min. The coding gain is reduced by the constellation doubling (by a factor of 2 or 1.5 dB, in four dimensions, or by a factor of or 0.75 dB in eight), but otherwise similar performance is achieved as in the nonpartial response case with similar code complexity.

SUMMARY OF THE INVENTION In accordance with the present invention there is provided apparatus for generating a sequence of digital signals Xk and/or a sequence of digital signals k 1, 2, such that the sequence of y, signals is a partial-response-coded sequence derived from the sequence of x k signals, said signals yk being a sequence S15 in a given modulation code, said apparatus comprising C0 et a coset selector for generating coset representatives c, in accordance with said given modulation code; and an encoder for selecting J said signals yk, J 1, (Yk, yk+, Yk+- 1 to be 1 congruent to a sequence of J coset representatives ck (modulo AN), AN being an N- 20 dimensional lattice, N being a positive integer, said J signals being chosen from a corresponding one of a plurality of NJ-dimensional constellations, said choice being based on a previous at least one of said plurality of NJ-dimensional constellations comprising both a point with a positive sum of coordinates and another point with a negative sum of coordinates, said encoder being arranged so that said signals x. have finite variance S,.

In preferred embodiments, either the x, or y, sequence may be delivered as an output; L 1; yk xk the code may be a trellis code or a lattice code; M may be 2 or 4 or a multiple of 4 or 2 2i; J may be 1 or the same as the number of dimensions in the modulation code; k' k 1; J is 1 and each constellation is a one-dimensional range of values centred on pXk-1, 0 r P 1, preferably P 0; there are a finite set of two non-disjoint) J-dimensional constellations; y and 7 0 911017,kxlspe.003,12040.re,6 Xtt< -7- -7x k may be real valued or complex valued.

Embodiments of the invention extend to decoding apparatus for decoding a sequence z k Yk nk, k 1, 2, into a decoded sequence yk, where the sequence of signals yk is such that the sequence is from a given modulation code of the type generated by the abovementioned apparatus; the running digital sum x, yk Yk-i Yk-2 has finite variance S; the signals yk fall in a predetermined permissible range dependant on xk, k; and the sequence n k represents noise.

A A/ A range violation monitor reconstructs the estimated running digital sum x, k

A

yk-1 compares the decoded sequence yk with a predetermined permissible

A

range based on the estimated running digital sum x k k' k, and generates an indication whenever the k is outside the permissible range.

Embodiments of the invention also provide for decoding apparatus for decoding a sequence zk yk nk, k 1, 2, where the sequence of signals y is such that the sequence is from a given modulation code, the code being capable of being generated by encoder, of the abovementioned variety, with a finite number Q of states; yv Xk Xk

L

L an integer, where the sequence X k has finite variance and the sequence n, represents noise, comprising a modified maximum S: 20 likelihood sequence estimator adapted to find MQ partial decoded sequences, up to Ssome time K, one sequence for each combination of the finite number Q of states and each of a finite number M of integer-spaced values modulo M, such that each sequence is in the code up to the time K; corresponds to the encoder being in a given state at the time K; corresponds to a value of xk at the time K that is congruent to a given one of the values, modulo M.

Embodiments of the invention are capable of adapting known modulation codes, particularly trellis codes, for use in partial response systems to achieve the same kinds of advantages that trellis codes have in non-partial response systems notably, substantial coding gains for arbitrarily large numbers n of bits/symbol with reasonable decoding complexity. Further embodiments enable the design of trellis codes for partial response systems in such a way as to achieve both a 911017,klspe.003,12040.re,7 i -8relatively low input signal power S x and a relatively low output power Sy, and permits smoothly trading off these two quantities against each other. Furthermore, higher-dimensional trellis codes can be adapted for use in partial response systems which are inherently lower-dimensional.

Other advantages and features will become apparent from the following description of the preferred embodiments, and from the claims.

1 i 911017.kxlsp.003,12040.e,8 I I9 Description of the Preferred Embodiments We first briefly describe the drawings.

Drawings Figure 1 is a block diagram of a 1-D partial response channel.

Figure 2 is a block diagram of an encoder for an 8-state Ungerboeck code.

Figure 3 is a signal constellation for the Ungerboeck code partitioned into 8 subsets.

Figure 4 is a block diagram of an equivalent encoder for the Ungerboeck code.

Figure 5 is a block diagram of a generalized N-dimensional trellis encoder.

Figure 6 is a block diagram of a modified Figure 5, based on coset representatives.

Figure 7 is a block diagram of an equivalent S'one-dimensional encoder.

Figure 8 is a block diagram of a generalized N-dimensional trellis encoder.

Figure 9 is a block diagram of a generalized *.encoder with coset precoding.

Figure 10 is a block diagram combining Figures 8 and 9.

S'2 Figure 11 is a block diagram of an encoder with RDS feedback and coset precoding.

Figures 12, 13, 14 are alternative embodiments S'of Figure 11.

Figures 15, 16, 17 are block diagrams of three equivalent filtering arrangements.

A T 3,0 Figure 18 is a block diagram of a generalized decoder.

Figures 19, 20, 21 are block diagrams of alternative encoders.

\O

Figure 22 is an alternative signal constellation.

Figure 23 is a block diagram of an encoder for use with the constellation of Figure 22.

Figures 24, 25, 26 are block diagrams of three equivalent encoders.

Figure 27 is a block diagram of an alternative decoder.

Figure 28 is a schematic diagram of an expanded signal constellation.

Figure 29 is a rhombus for use with the constellation of Figure 28.

Figure 30 is a block diagram of a two-dimensional RDS feedback encoder.

15 Figure 31 is a diagram of the dimensions of a o* rhombus for use with the constellation of Figure 28.

:Figure 32 shows a pair of constellations.

Figure 33 shows two disjoint constellations.

Structure and Operation Referring to Figure 1, the invention includes a technique for generating signal sequences to be used as inputs for a partial response channel 10, for example a one-dimensional (real) 1-D partial response baseband system with a null at DC. (Later we shall indicate briefly how to modify such a design for other types of partial response systems.) Each output signal zk of such a system is given by f; Zk k k' where the nk sequence represents noise, and the 30 yk sequence is a partial-response-coded (PRC) sequence defined by Yk Xk Xk-l' where the xk sequence is the sequence of -k2channel inputs. Because Xk Xk-1 k the xk sequence can be recovered from the PRC sequence by forming a running digital sum of the Ykvalues (given an initial value for the xk sequence); thus w'e call the xk sequence the RDS sequence. The sample variances of the RDS sequence x(D) and the PRC sequence y(D) will be denoted as S~ and S respectively.

The discrete-time partial response l-D (represented by block 12) is a composite of the responses of a chain of transmit filters, an actual channel, receive filters, equalizers, samplers, etc., designed in a conventional way to achieve a composite partial response 1-D with the noise power P (of the 15 noise sequence being small relative to the PRC power S .We shall thus want to send a relatively large number n of bits per channel input. A detector *(not shown) operates on the noisy PRC sequence z(D) to estimate x(D) (or equivalently since there is a one-to-one relationship between them). If the detector is a maximum likelihood sequence estimator, then, to 2.*'.first order, the objective is to maximize the minimum *squared distance dmin between permissible PRC *.:sequences y(D).

in some applications, the design constraint will simply be to minimize the sample variance S of the RDS (input) sequence. In others, the constraint :will be on S .In still other applications, there will be an effective power constraint somewhere in the middle of the composite filter chain, so that it will be desirable to keep both S and S ysmall, and in fact to provide for a smooth design tradeoff between them.

A related problem is the design of sequences with spectral nulls, a null at zero frequency There the objective may be to design sequences y(D) that can represent n bits per sample, that have a spectral null, that have as small a sample variance S as possible, but that also have a large minimum squared distance d 2 i between possible y(D) sequences. A common auxiliary objective is to keep the variation of the running digital sum (RDS) of the y(D) sequence limited as well, for systems reasons. Because the running digital sum sequence x(D) is, and its sample variance S x is a measure of its I variation, the present invention may also be applicable to the design of sequences with spectral nulls.

15 A number of design principles are useful in achieving our objectives. The first principle is to S' design the input (RDS) sequence x(D) so that the output (PRC) sequence taken N values at a time, is a sequence of N-dimensional signal points belonging to subsets of an N-dimensional constellation determined by a known N-dimensional trellis code. Then the minimum S2 squared distance dmi n between PRC sequences will be at least the din guaranteed by the trellis code.

S" Furthermore, a maximum likelihood sequence estimator for the trellis code can be easily adapted for use with this system, and while perhaps not optimum, it will achieve 2 the same effective din for essentially the same Sdecoding complexity as for the same trellis code in a Snon-partial response system.

An illustrative embodiment of the present invention is based on a known 8-state 2-dimensional trellis code similar to that of Ungerboeck as described in the article cited above, which uses a 128-point two-dimensional constellation to send 6 bits per :i 1-1-4 (two-dimensional) signal. (This is also similar to the code used in CCITT Recommendation V.33 for a 14.4 kbps data modem.) Figure 2 shows the encoder 20 for this code. For each six-bit symbol 21 delivered from a data source 23, 2 of the 6 input bits to encoder 20 enter a rate-2/3, 8-state convolutional encoder 22. The 3 output bits of this encoder are used in a subset selector 24 to select one of 8 subsets of a 128-point signal constellation, illustrated in Figure 3; there are 16 points in each subset (points in the eight subsets are labelled A through H respectively). The remaining 4 "uncoded bits" 26 (Figure 2) are used in a signal point selector 28 to select from the chosen subset the (two-dimensional) signal point to be transmitted. The code achieves a gain in dmi n of a factor of 5 (7 dB) over an uncoded system, but loses about 3 dB in using a 128-point rather than a 64-point constellation, so the net coding gain is about 4 dB.

One-dimensional form of the 2-dimensional Ungerboeck code The sequence of symbols xk sent over the channel is one-dimensional in a 1-D baseband partial response system. It is helpful (though not essential), therefore, to transform known trellis codes into one-dimensional form. There are two aspects to this transformation: first, to characterize the two-dimensional subsets as compositions of constituent one-dimensional subsets, and second, to characterize the finite two-dimensional constellation as a composition of constituent one-dimensional constellations. We now show how this decomposition is done for the illustrative two-dimensional Ungerboeck code, and then indicate how it may be done in the general case of an N-dimensional trellis code.

RA 4 V J^ (Lf.

The first step is to notice that each of the eight two-dimensional subsets A, B, can be viewed as the union of two smaller two-dimensional subsets, say A and B and B, etc., where each of the 16 smaller subsets can be characterized as follows. Let the possible values of each coordinate of a signal point be partitioned into four classes a, b, c, d; then each of the smaller two-dimensional subsets consists of the points whose two coordinates are in a specified pair of classes. A convenient mathematical expression for this decomposition arises if we scale Figure 3 so that signal points are one unit apart in each dimension (and the coordinates of each point are half-integers); then the classes a, b, c, d are equivalence classes (modulo 4), and each of the 16 sets A Al, B 0 are the points whose two coordinates are congruent to a given 'pair y) modulo 4, where x and y may each take on one of the four values b, c, e.g. These four values are called (one-dimensional) 'coset representatives'. The points of the constellation of Figure 3 have been labeled with Os and Is to show one possible arrangement of the 16 subsets. For example the G O point 29 has coordinates x 5/2, y 9/2, and its coset representatives are 9/2) modulo 4 or 1/2).

We may now modify Figure 2 as follows.

j tReferring to Figure 4, the three output bits of the encoder 22 plus one of the uncoded bits 30 are used as inputs to subset selector 32, which selects one of 16 3Q subsets based on the four input bits, with uncoded bit selecting between A o and A 1 or B o and B., etc., according to which of the original 8 subsets is selected by the three convolutionally encoded bits produced by encoder 22. In effect, encoder 22 and bit 7 '^OIT represent an 8-state rate-3/4 encoder, with the output selecting one of 16 subsets, although the set of possible signal point sequences has not changed. Next, designate each of these 16 smaller subsets by a pair of one-dimensional coset representatives 34, one for each coordinate, where each coset representative ck may take on one of four values. The pair of coset representatives is denoted (clk, c 2 k).

An aspect of the invention is the observation that all of the good codes cited above those of Ungerboeck, Gallager, Wei, and Calderbank and Sloane can be transformed in the same way. That is, any of S these N-dimensional trellis codes can be generated by an

N

encoder that selects one of 4 subsets, where the C t 15 subsets are specified by N 4-valued one-dimensional Scoset representatives, corresponding to congruence classes of each coordinate (modulo In some cases it is only necessary to use 2 N subsets specified by N 2-valued one-dimensional coset representatives corresponding to congruence classes of each coordinate (modulo for Ungerboeck's 4-state 2D code, Gallager's 8-state 4D code (and the similar code of Calderbank and Sloane), Wei's 16-state 4D code and 64-state 8D code, etc. Also, we have observed that many good lattice codes can be transformed in this way; e.g., the SchlAfli lattice D and the Gosset lattice E 8 can be represented by sequences of 4 or 8 two-valued Sone-dimensional coset representatives (modulo the Barnes-Wall lattices A16 and A32 and the Leech lattice A 24 can be represented by four-valued one-dimensional coset representatives (modulo 4).

A general form for all of these codes is shown in Figure 5. The encoder is N-dimensional and operates once for every N signals to be sent over the channel.

In each operation, p bits enter a binary encoder C 33 and are encoded into p r coded bits. These coded bits select (in selector 35) one of 2 p subsets of an N-dimensional signal constellation (the subsets corresponding to the 2 p r cosets of a sublattice A' of an N-dimensional lattice A, the constellation being a finite set of 2 n r points of a translate of the lattice A, such that each subset contains 2 n p points). A further n p uncoded bits selects (in .o V, selector 37) a signal point from the selected subset.

SThus the code transmits n bits for every N-dimensional n+r 15 symbol, using a constellation of 2 r N-dimensional r signal pointsN The encoder C and the lattice partition A/A' ensure a certain minimum squared distance 2 dmin between any two signal point sequences that belong to a possible subset sequence.

The observation above (about the transformability of all good codes) is the result of the '4 •mathematical observation that for all of the good trellis and lattice codes cited, the lattice 4

Z

N of t N-tuples of integer multiples of 4 is a sublattice of the lattice A' (and in some cases 2Z is). Then, for some integer q, A' is the union of 2 q cosets of 0 4 zN in The practical effect of this observation is that, provided that n q p, we can take the p r coded bits plus q uncoded bits into a subset selector that selects one of 2 q p r cosets of 4Z N in A, and further that these cosets can be identified by a sequence of N 4-valued one-dimensional coset representatives (clk C 2 cNk), where the Cjk represent integer-spaced equivalence classes I z^ J 3

A

Jr.. 1. ^CUU-U~ C ill \'1 (modulo Thus already Figure 5 can be modified as shown in Figure 6. In this modification, we assume that the 2n+r-point signal constellation divides evenly into 2 q p r subsets, each containing the same number of signal points 2 n-q-p).

The illustrative Ungerboeck code embodiment is 2 2 an example in which N 2, A Z 2 A' 2RZ 2 p 2, p r 3, q 1, and n 6.

The second step is to decompose the constellation into constituent one-dimensional constellations. For the constellation of Figure 3, each coordinate can take on one of 12 values which can be c. grouped as 8 'inner points' ±3/2, and 4 'outer points' a06 15 as suggested by boundary 31 in Figure 3.

Sae There are 2 inner points and 1 outer point in each of the four one-dimensional equivalence classes the class whose coset representative is +1/2 contains the two inner points +1/2 and and the outer point 9/2, because these three points are congruent to +1/2 (modulo Given a coset representative, therefore, it is only necessary to specify whether a point is an inner point 9 or an outer point and, if an inner point, which of the o, two inner points it is. This can be done with two bits, say blk inner or outer) and b 2 k which inner) (or with one three-valued parameter ak).

We may say that the pair (blk, b 2 k) is a range identifying parameter a k which takes on one of Sthree values, indicating the following three ranges from 0 to 4 (inner point, positive); from -4 to 0 (inner point, negative); from -6 to -4 and from 4 to 6 (outer point).

,A

A

To e ^n7 r- -1

IS

1+9 The fact that each range spans a portion of the real line of total width 4 that contains exactly one point congruent to any real number (modulo 4) means that the range-identifying parameter ak plus the coset representative ck specify a unique signal point, for any value of ck.

The signal point selector 36 of Figure 4 can then be decomposed as follows. Referring to Figure 7, three uncoded bits 40 enter a range-identifying parameter selection element 42 for each pair of coordinates. One uncoded bit determines whether any outer point is sent. If so, a second bit then determines which coordinate will contain the outer point, and the third bit selects which inner point in 1 15 the other coordinate. If not, both coordinates are inner points, and the second and third bits select which inner point in each coordinate. Thus, in sum, element 42 maps the three uncoded input bits 40 into two pairs of output bits 44 al (bll b 12 and a 2 2Q (b 21 b 22 with each pair of bits used to determine one coordinate in conjunction with the corresponding coset representative, c, or c 2 generated by coset S' representative pair selector 46. Thus the whole encoder I has been reduced to a form in which each coordinate x k (48) is selected (in a coordinate selector 50) by 4 bits, two representing ck and two representing ak (blk, b 2 k).

All constellations commonly used with the C11 above-cited codes can be decomposed in this way. The principles are similar to those discussed in my U.S.

Patent 4,597,090 and in Forney et al., "Efficient Modulation..." cited above, where N-dimensional constellations were built up from constituent 2-dimensional constellations; a similar buildup from 'i 2-dimensional constituent constellations was used by Wei in conjunction with trellis codes in his U.S. Patent Sppl4.itio, cited above.

The general form of encoder for N-dimensional codes is shown in Figure 8. For every N coordinates, p bits 51 enter an encoder 52 and p r coded bits 54 are produced; these plus q uncoded bits 56 enter a selector 58 which selects a sequence of N coset representatives the remaining n-p-q uncoded bits 62 are transformed (in a selector 64) into a sequence of range-identifying parameters ak (66) which together with the ck determine (in a signal point selector 68) a sequence of N signal point values x (70) by a .I signal point selection function f(c k ak) which S 15 operates on a one-dimensional basis. In general, the range-identifying parameter ak determines a subset of the real line (one-dimensional constellation) of width (measure) 4 which contains exactly one element congruent to any possible c k value (modulo and the function f(c k ak) selects that element. For all codes cited, the coset representative alphabet may be taken as Sfour integer-spaced values (modulo for some codes, the coset representative alphabet may be taken as two integer-spaced values (modulo 2) (in which case the ranges are of width The size of the ak alphabet is as large as necessary to send n bits per N coordinates. The signal point sequences generated by this form of encoder are generally the same as those in L" ithe original code, and in particular, are separated by 30Q the same minimum squared distance d min as the original code.

RA4/.

o ?+TO Coset Precoding N-dirmnsional signal point sequences generated by known good trellis codes, when serialized to one-dimensional signal points, cannot in general be used as inputs to the partial response channel of Figure 1 without degradation of din (because of intersymboi interference). However, a technique which we call coset precoding allows the adaptation of these known codes to partial response systems without increase of S or 2 x degradation of dmi n The general technique is illustrated in Figure 9.

We use the same convolutional encoder 52 as used by the known trellis code, preferably in the form of Figure 8. The p r coded output bits 54, rather 15 than selecting a subset directly, are converted (as in Figure 8) in a subset selector/serializer 70 into a sequence ck of N one-dimensional coset representatives cl, .,c

N

corresponding to the subset that would be selected in a non-partial response system. These coset representatives are then 'precoded' (in a precoder S' 72) into an alternative (or 'precoded') coset representative sequence ck' where

C

k Ck-l' ck (modulo 4) (In the cases where it is possible to use modulo 2 coset representatives, this precoding can be done modulo 2.) Thus the precoded coset representative sequence 74 is a running digital sum modulo 4 (or 2) of the ordinary coset representative sequence. Precoded coset representatives ck' can then be grouped N at a time in 3Q grouper 75 to specify (in signal point selector/serializer 76) an N-dimensional subset; a signal point can then be selected (based on the uncoded bits 78) in the usual way; and the resulting signal 1 Rp 4* d g .0 -22 point can be sent out as a sequence x(D) of N one-dimensional signals xk over the partial response channel (in the same order as they were precoded).

Note that if the c k are half-integers, then the ck' alternate between two sets of 4 values, one set displaced by 1/2 from the other. This has only a minor effect; we can, for example, "dither" alternating coordinates xk by +1/4 and -1/4 so as to accommodate this periodicity. Alternatively, we may let the c k alphabet be integer-valued, 1, 2, then the ck' are always from the same alphabet, e.g., t3/2}. These offsets of ck' or ck do not affect the d 2 of the code.

min If the encoder is in the form of Figure 8, then 15 Figure 9 can be put in the form of Figure 10, where the I t isame blocks do the same things. In particular, since we have characterized the function f(ck, ak) as one that selects the unique element congruent to ck in a range identified by a k it does not matter if the precoding changes the ck' alphabet from the ck alphabet; indeed, the (modulo 4) in the precoder is unnecessary in principle, though possibly useful in practice.

With either Figure 9 or Figure 10, it can be shown that the PRC sequence Yk Xk xk-1 has elements that are congruent to ck (modulo so they Sfall in the subsets of the original trellis code, and Stherefore have at least the same d .The RDS min sequence xk has the same average energy S x as in the original trellis code if the c k alphabet is the same as the ck alphabet; even if not, approximate equality still holds. (In the illustrative embodiment, the average energy per coordinate is 10.25, with variables, and thus S 2S; the spectrum of the RDS sequence (xk} is flat (white) within its Nyquist band; the spectrum of the PRC sequence {yk} is the same as that of the partial response channel.

Even if the ck are not integers, these statements are still approximately true.

Coset precoding can be modified for other kinds r'kof partial response systems as follows. For a 1+D I(one-dimensional) partial response system, use the same system except with Ck-1 subtracted rather than added in precoder 72, so that ck ck ck 1 (modulo

L

For a I-D system, replace the delay element D by a delay element D L so that ck kL ck For a 1+D two-dimensional system, use two 1+D precoders in parallel, with pairs of outputs from the subset selector/serializer as inputs, and with the two outputs determining the real and imaginary (in-phase and quadrature) parts of the two-dimensional signal point to be transmitted.

RDS Feedback Depending on the application, it may be desirable to reduce the average energy y of the PRC sequence, at the cost of increasing the average..energy Sx of the RDS sequence. This will also tend to flatten the PRC spectrum, while raising the low-frequency content of the RDS spectrum. Justesen, "Information Rates and Power Spectra of Digital Codes", IEEE Trans. Inform. Theory, Vol. IT-28, pp. 457-472, dt ii t r a -4- 1982, has introduced the notion of a 'cutoff frequency" fbelow which the PRC spectrum is small and above which it tends toward flatness, and has shown thatf0 is approximated by f 0 (S /2S )f\~whr is the Nyquist band edge frequency.) A general method of performing this cradeoff while maintaining the dmi of the trellis code in the PRC sequences is to augment the encoder of Figures 9 or 10 as follows.

The PRC sequence may be computed from the RDS sequence; for the 1-D channel, each PRC signal is just Yk Xk xkl. Referring to Fig. 11, we may let eq.the signal point selector 80 base each x k on xk-1 (by feeding x k back through a delay element 82) as well as on the current precoded coset representative c I k and on the range-identifying parameter ak in **;such a way that large PRC values Yk (calculated in summer 84) are avoided. As long as the signals x k are 9still chosen to be congruent to the ckI (modulo 4), the signals Yk will be congruent to the ck (modulo and therefore will preserve the d mi of the trr&.lis code. (Note that although the idea is to *precompute the PRC value Ykto keep it small, what is actually fed back is the previous RDS value xk1 so that we call this RDS feedback.) k1 For the illustrative embodiment, this could work as follows. As already noted, the normal selection 9. function~ f(ck ak') of selector 80 can be characterized by saying that the 8 inner points are the 8 half-integer values lying in the range frm -4 to +4, while the 4 outer points are the 4 half-integer values lying in the range from -6 to -4 and +4 to We can vary the inner point range and outer point range as a function of xk..l as long as the inner poin~t range

FIAA

spans 8 signal points, 2 from each equivalence class, while the outer point range spans 4 signal points, 1 from each equivalence class.

A general way of doing this is to translate all ranges by a translation variable R(xk- 1 which is a function of That is, in the illuscrative embodiment, the inner point range is modified to be from -4 R(xk-1) to 4 R(xk-l) and the outer point range to be from -6 R(xk-1) to -4 R(xk-1) and from 4 R(xk to 6 R(xk_).

The function R(xk-1) should be generally increasing with Xk_ 1 so as to reduce the yk. We have been able to show that the optimum choice is R(xkl) xxk ,1 where 8 is a parameter in the range 0 8 1. When 8 0, the RDS feedback through element 82 disappears and Figure 11 reduces to coset precoding as in Figure 10. With this choice, if S O is the value of S in the ordinary case (8 then it is approximately true that

S

x S/(1 2); S 2S B); The spectrum Sx(f) of the RDS sequence is proportional to 1/(1 2Bcos e 2 where 9 f/fN; The spectrum S of the PRC sequence is 2 proportional to 2(1 cos8)/(l 23 cose 82); the "cutoff frequency" f 0 is (l-8)fN' The xk are limited to the range form k to and the Yk are limited to the range -M to M, if the range of coordinates in the orignal code is from -M/2 to M/2.

As 0 approaches 1, Sy approaches S

O

and S approaches a flat spectrum with a sharp null at DC. Meanwhile, S x becomes large and S (f) X Xx

T.

as- -26approaches a spectrum, except that it remains finite near DC. We have been able to show that this is the best possible tradeoff between S Sy, and S0' Figures 12, 13, 14 show three equivalent ways of generating xk and/or yk based on c k a k and xk-. Figure 12 corresponds most closely to Figure 11.

In Figure 12, the feedback variable c'k_1 in coset precoder 72 is replaced by Xk_, since c'k-i Xk- (modulo and only the value of c'k (modulo 4) is used in selector 80. R(a k denotes the range identified by a k and R(xk-1) represents the range translation variable introduced by RDS feedback. Since Yk Xk Xk-1 C'k Xk-1 (modulo 4) and c k ck Xk- (modulo yk c S 15 (modulo 4).

Figures 13 and 14 are mathematically equivalent to Figure 12, in the sense that if they have the same starting value Xk_ and the same sequence of inputs (c k ak), they will produce the same sets of outputs (xk' Yk In Figure 13, yk is chosen as the unique element congruent to ck in the range R(ak) S. XkI and xk is determined from Yk as k yk Xk-l' so xk c'k ck S. Xk-l (modulo and is the unique element in the range R(ak) R(xk-1) congruent to c'k (modulo In Figure 14, an innovations variable i k is chosen as the unique element congruent to c"K ck K k xk- 1 R(xk-1) (modulo 4) in the range R(ak), and xk is determined from ik as xk i k R(xk1), so xk c k R(xk-l) cl (modulo and is the unique element in the range R(ak) R(xk-1) congruent to c'k (modulo 4).

Figure 12 combines the delay element in the precoder with the delay element necessary for RDS feedback, and V13^ is most useful if xk is the desired output and the c k are always from the same alphabet, e.g. Figure 13 eliminates the precoder altogether, and is most useful if yk is the desired output and the ck are always from the same alphabet, e.g. 1/2, Figure 14 takes the range translation variable R(xk outside of the selector, so that the ik are always chosen from the same range (the union of all the innovations equence i(D) is approximately a sequence of independent identically distributed random variables ik (ignoring the minor variations induced by the c' k congruence constraint) and this auxiliary sequence can be useful if a white (spectrally flat) sequence deterministically related to to: 15 x(D) or y(D) is desired.

Figures 15, 16, 17 illustrate three equivalent filtering arrangements for use with the and i(D) sequences of Figures 12, 13, 14. In Figure 15, the S' RDS sequence x(D) is filtered in a transmit filter HT(f) before being transmitted (as signal over r the actual channel (not shown). In Figure 16, the PRC sequence y(D) is filtered in a transmit filter H'T(f) whose response is equivalent to that of a cascade of a sampled-data filter and HT(f); since y(D) has a DC null, it does not matter that the response of is infinite at DC (particularly if HT(f) also has a DC null). In Figure 17, the innovations sequence i(D) is filtered in a transmit filter H" whose response is equivalent to that of a cascade of a 1/(1-3D) sampled-data filter and HT(f); this is equivalent to Figures 15, 16 if R(xk_) Bxk-1; otherwise, the equivalent sampled-data filter is the filter corresponding to xk i k R(xk-l), in -24 general nonlinear. Any one of these equivalent forms may be preferable depending on HT(f), R(xk1l), and the implementation technology being used.

Certain modifications of the above RDS feedback systems may be desirable in practice. For example, it may be desirable to change the form of the ranges R(ak) from those used when R(xk-1) 0. For example, in the illustrative embodiment, a simply implemented form of RDS feedback is as follows: when Xk-1 is positive, let yk be chosen as usual in the range from -4 to 4 if ak indicates an inner point, but if ak indicates an outer point, let yk be the number congruent to ck in the range from -4 to when Xk-1 is negative, use the range from 4 to 8 for outer points. Then the range of the PRC sequence yk is limited to -7 1/2 to +7 1/2, rather than -11 to 11 when there is no RDS feedback; the PRC variance S is reduced to 13.25 from 20.5, a reduction of 1.9 dB, and about 1.1 dB above S 10.25; 0 the mean of Yk is 3/2 if Xkl 1 is positive, and 3/2 if Xk_1 is negative, so that the RDS sequence tends to remain in the 25 neighborhood of zero. While it is difficult to compute Sx exactly, it follows from the facts that E[Ykxk-1 S /2 and E[yk Xk-1 E[Ixk-11] that the mean of the 'o absolute value of xk is S /3 4.42, so that the RDS sequence xk is fairly well bounded. (With no RDS feedback, the mean of the absolute value of xk is 2.75); 1: -ii the variance of the yk' given Xk-1' is S 11, about 0.3 dB higher than the Sg 10.25 possible with no RDS feedback. The minimum possible S for S 13.25, S 11, is S, 19.5, corresponding to B 0.66.

x 2 Since S x SIxi E[I|xl S must be greater than (4.42) 19.5, so with this simple method we achieve less than the optimal spectral tradeoff; every possible yk is associated with a unique pair (ck' ak). As we shall discuss in more detail below, this means that a decoder need I not keep track of an estimated running digital sum of the estimated PRC sequence, and that 15 there will be no error propagation in the St. decoder.

In summary, this simple method does not achieve the best power tradeoff between S x and Sy, but does

S

t effectively limit not only S but also the peak values y of Yk' keeps the RDS sequence xk fairly well bounded, and avoids error propagation at the receiver.

Thus these methods allow for trading off of S versus S from the unconstrained case, where the xk sequence is uncorrelated, S x has the same energy S O as is necessary to send n bits per symbol in the non-partial-response case, and S 2

S

x (ii) Salmost to the case where the yk sequence is uncorrelated, Sy S

O

and S x becomes very large.

These tradeoffs are possible for all trellis and lattice codes cited.

RA_

'T 0 L di 17 I 1 11 Ia Decoding The above methods succeed in generating PRC sequences that belong to a known good code, and 2 therefore have a d i at least as great as that of the code.

Referring to Fig. 18, a suitable detector for the noise received PRC sequence z(D) y(D) n(D) is therefore a maximum likelihood sequence estimator (Viterbi algorithm) for the known good code, adapted as follows: A first step of decoding may be, for each noisy received PRC value zk Yk nk' for each of the four classes of real numbers congruent to the four one-dimensional coset representatives cjk (modulo j 1,2,3,4, find (block 92) the closest element yjk in each class to zk and its 'metric' mjk (jk Zk) (squared distance from zk); In a code based on an N-dimensional lattice partition a second step of decoding may be, for each of the 2 p r cosets of A' in A, to find the best (lowest metric) of the 2 q cosets of 4Z N whose union is that coset of by summing the respective metrics of the constituent one-dimensional metrics mjk and comparing these sums (block 94); Decoding can then proceed in the usual manner (block 96), using as a metric for each coset of A' the best metric determined in step The decoder will ultimately produce an estimate of the sequence of cosets of which can be mapped to a sequence of estimated coset representatives c k which can be mapped to the

T

I

I

I

3o *P 4 44 4 4r .4 4 4 4 4 4**t 4r 4 *4r 4 4 I. *4 S. V 4 4 4i 4 14 corresponding yk' from which the original ak and xk can be recovered if desired (block 98). These last steps require that the decoder

A

keep track of the running digital sum xk- 1 of the estimates yk" Since the PRC sequences are in che known code, the error probability of this decoder will be at least as good as that of the known code, in the sense of 2 achieving at least the same effective dmin min' However, because the PRC sequences are actually only a subset of the known code sequences, such a decoder is not a true maximum likelihood sequence estimator for the PRC sequences. As a result, it may occasionally decode to a sequence which is not a legitimate PRC sequence.

Legitimate PRC sequences must satisfy the two following additional conditions: A legitimate finite PRC sequence y(D) must be divisible by 1-D; the sum of its coordinates must be zero; the range constraints imposed by the signal point selector must be satisfied for all y A (or equivalently x k or ik), based on the reconstructed values of the RDS Xk-,l If this decoder makes a normal decoding error, corresponding to a short period of wrong coset estimates followed by correct coset estimates, it is possible that the corresponding finite PRC error sequence will have a running digital sum other than zero. This will cause a persistent error in the decoder's estimated running 30 digital sum x.l. which may lead to occasional mapping _i A A errors back to the yk, ak' and ultimately x k even though the cosets c k are correct, for as long as the error in the RDS estimate persists.

pri 3- The decoder must therefore continually monitor (block 99) whether the range constraints in the reconstructed yk and xk are satisfied. If they are not, then it knows that its estimated RDS xk_- is incorrect; it should adjust by the minimum amount necessary for the range constraint to be satisfied, assuming that the coset sequence ck is correct. With probability 1, this will eventually result in resynchronization of the estimated RDS to the correct value, and normal decoding can resume. However, there may be a considerable period of error propagation.

Avoidance of Error Propagation We now give a general method of avoiding error propagation at the receiver. The method works best when the signal constellation consists of all points in A within an N-cube, but is not restricted to that case.

It may be regarded as a generalization of the principles i of earlier forms of precoding (modulo M) for use with coded sequences.

The basic idea is that each possible PRC value Yk should correspond to a unique (c k ak) value, when the code can be formulated in one-dimensional form as in Figure 7; or, more generally, that each group of N Yk values should correspond not only to a unique sequence of N ck values but also to a unique set of uncoded bits, if a general N-dimensional signal point selector is used as in Figure 6. Then the inverse map

A

from decoded yk to coded and uncoded bits is Sindependent of the decoder's estimate of the running digital sup, so that the decoder need not keep track of the RDS; error propagation does not occur.

Thus in Figure 18, block 99 can be eliminated.

r. *3 I- Figure 19 shows how this may be done where the code can be formulated in one-dimensional form, as in the illustrative embodiment. From ck and a k a signal point selector selects a value s f(c a k as in Figure 8. In the illustrative embodiment, s k takes on one of 12 values, namely, the half-integral values in the range from -6 to 6. In general, s k will take on one of the values from an integer-spaced alphabet in a range of width M; we denote this range by R 0 Then, as in Figure 13, yk is selected as the unique number congruent to s k (modulo M) in the range Rg R(xkl) Xk 1 of width M, where R(xk-l) is an RDS feedback translation variable, 'and Xkl is the previous RDS signal point. The 15 current RDS xk is computed as yk Xk-l.

'tl, t Figures 20 and 21 are equivalent methods of generating xk and/or yk from the sequence s k such that yk sk (modulo analogous to Figures 12 and 14. In Figure 21, an innovations variable ik is generated which is more or less white and uniformly I distributed over the range R 0 so that its variance

S

O is approximately M 2 /12; thus SO 12 for the illustrative embodiment, a penalty of about 0.7 dB over t the value of S o 10.25 achievable with no RDS feedback. As in Figures 12, 13, 14, all three sequences Xk' yk' and ik carry the same information, and as in Figures 15, 16, 17, any can be used as the input to a filter which shapes the spectrum for transmission.

SThe penalty in the innovations variance is eliminated if the original code coordinates are uniformly distributed over a range RO; if the original constellation is bounded by an N-cube with side

R

O

_i 33 As an illustrative embodiment with a square constellation, we use the same two-dimensional 8-state Ungerboeck encoder as in Figure 2, except with the 128-point constellation of Figure 22 rather than that of Figure 3. The constellation consists of alternate points from the conventional 256-point 16 x 16 constellation; thus the coordinates have the 16 half-integral values 1/2, 3/2, 15/2}, but with the restriction that the sum of the two coordinates must be an even integer modulo 2).

The minimum squared distance between signal points is 2 thus 2, rather than 1; and the d of the code is min rather than 5. The variance of each coordinate is now 21.25 rather than 10.25, which after scaling by 2 is 15 a loss of 0.156 dB relative to the Figure 3 constellation, since the cross is more like a circle than is the square. (In lattice terminology, we are now partition RZ 2 2 ,using the 8-way lattice partition RZ2/4 rather 2 2 l than Z 2 /2RZ 2 jt '20 It will be observed that now each of the eight subsets corresponds to a unique pair of coset representatives (cl, c 2 modulo 4, such that c i

C

2 0 (modulo Therefore, the three coded bits of Figure 2 determine a pair of coset representatives directly in subset selector 24, rather than with the aid of an uncoded bit as in Figure 4. The four uncoded bits then select one of the 16 points in the selected subset. In this case, the uncoded bits may simply be taken two at a time to determine one of the four ranges to -4 to 0, 0 to 4, or 4 to 8. This is conveniently expressed by letting the two-bit range identifying parameters (al, a 2 each represent one of the four values t6}; then the coordinate selection function is simply s k f(ck, ak) c k k k k' 34~ C 4 Cr14 'Itt 54 4 I. I I4 144-S St 4 4

C~

S I I 4 St I S I I SI St I IC 4* 5 S 4

C.

55 0 f 4 4 4* ak. Note that the possible values for s k are the 16 half-integral values in the range R 0 from -8 to 8, of width M 16.

Conventional precoding may then be done modulo 16. The entire encoder is illustrated in F'igure 23.

The RDS value x k is the sum s k xk-1 (modulo 16). In this case the x k values are essentially independent identically distributed (white) random variables, and Yk =xk xkl sk (ouo1) To obtain spectral tradeoffs via RDS feedback as in Figures 12, 13, 14, let s k continue to represent the desired congruence class Of Yk (modulo 16), and let R(xkl1) be an RDS feedback variable as in Figures 12, 13, 14, ideally equal to 3 x k-l. Then Figures 24, 15 25, 26 show three equivalent methods of obtaining sequences x k and/or Yk xk xkl such that Y s(modulo 16) and Sand S have the k k x y desired tradeoff, given S 0 21.25. Here R0is the range from -8 to 8.

In this case the innovations variable i k has variance So 0 1 /12 21.33, essentially the same as the variance of each coordinate in Figure 22, so that there is no penalty beyond the 0.16 dB involved in using Fig. 22 rather than Fig. 3.

25 As already noted, the decoder need not keep track of the RflS, because, given the estimated PRC AA A sequence Yk' the ^ck ak, and ultimately the original input bit sequence are uniquely determined.

However, if the decoder does keep track of the estimated 30, RfS and thq corresponding ranges that the yk should fall into, it can detect that an error has occurred whenever the decoded Ykfalls outside the estimated

I

V-4r

A

f: 1 3-6

I

SI~

I.'

t t I 41 range. Even if not used for error correction, such range violation monitoring can yield an estimate of decoder error rate.

Augmented Decoders A true maximum likelihood sequence estimator would take into account the entire state of the encoder and channel, which in general will include the value of the RDS xk_ 1 (the channel state) as well as the state of the encoder C. Such a decoder would achieve the true d 2i of the PRC sequences, and would be free of error propagation. However, because Xk_- in general takes on a large number of values, in principle possibly an infinite number with RDS feedback, such a decoder may not be practical. In addition, to achieve the true d in may require an essentially infinite decoding delay because the code/channel combination becomes quasi-catastrophic when n is large, as we shall explain more fully below.

It may be worth considering augmenting the decoder to at least achieve the true d 2 i of the code, however. Because all finite PRC sequences are divisible by 1-D, all finite-weight error sequences must have even weight. Thus, the true d 2 is always even. In the illustrative embodiment, the true dmi n is actually 6, not A general method for achieving the true d in in such cases while only doubling the effective number of states in the decoder is as follows. Let the decoder split each state of the encoder C into two, one corresponding to an even RDS and one to an odd. During decoding, two sequences then merge into the same state only if their estimated RDS has the same value (modulo Thus, it becomes impossible for two sequences differing by an odd-weight 36 I -2 error sequence to merge, so that the effective dmin is the weight of the minimum even-weight error sequence in the original code. Further, if there is a decoding error that results in a persistent estimated RDS error, as discussed above, that error must be at least 2, so it will tend to be detected sooner.

The decoder of Figure 18 can be used, modified only as shown in Figure 27. For most codes, each of the subsets of the signal constellation (cosets of A' in A) will contain points all of which have a sum of coordinates which is either even or odd. For example, in Figure 3, four of the eight subsets contain points whose coordinate sum is 0 (modulo and 4 contain points whose coordinate sum is 1 (modulo Thus the 15 metric of each subset (coset of A' in A) can be S- determined as before in blocks 92 and 94; the maximum likelihood sequence estimator 196 is then modified to find the best sequence of cosets that is in the code, and has a running digital sum congruent to 20 zero (modulo The decoded coset sequence is mapped back to Yk and x k in block 98 as before, with adjustment of xk_l by block 99 if necessary (adjustments will now be by multiples of 2), S* There is a drawback to this technique, however, in addition to the doubling of the decoder state space.

Two sequences may differ by an odd-weight error sequence followed by.a long string of zeros (no differences).

The decoder may then follow parallel pairs of states in the decoder trellis for a very long time, without S 3Q resolving the ambiguity. This 'quasi-catastrophic' behavior can ultimately be resolved by the maximum likelihood sequence estimator only by a range violation i 31 a 4 4 a..

t 94 t (a I Ie ar 54 a 4 4c due to the differing RDS parity on the two paths. Thus, the decoding delay required to achieve the true d may be very large.

min For this reason, it will generally be preferable simply to chose an encoder C with twice the number of states, and use an unaugmented decoder for C.

For example, there is a 16-state 2-dimensional Ungerboeck code with din 6; even though it may have a somewhat larger error coefficient than the 8-state code with an augmented 16-state decoder, we believe that in practice it will be preferable.

It may be worth mentioning that PRC sequences drawn from the 4-state two-dimensional Ungerboeck code 2 also have a true d of 6, since that code has min dmn 4' with the only weight-4 error sequences being single coordinate errors of magnitude 2, which are also not divisible by 1-D. A 16-state decoder which keeps track of the RDS modulo 4 can achieve this d However, in this case not only is the code min' quasi-catastrophic, but also the error coefficient is large, so again it would seem that the ordinary 16-state 2D Ungerboeck code would be preferable.

Quadrature Systems As mentioned earlier, a complex (or quadrature) 25 partial response system (QPRS) may be modeled as a 1 D sampled-data filter operating on a complex-valued RDS sequence x(D) to produce a complex-valued PRC sequence y(D) (1 D) yk xk Xk-l' When used with double-sideband quadrature amplitude modulation over a bandpass channel, such a system results in nulls at both band edges, fc fN' where fc is the carrier frequency and fN 1/2T is the width of a single Nyquist band.

36 When N is even and 4Z is a sublattice of as is the case with all good codes previously mentioned, then we can adapt a known good code for use in a QPRS system by using essentially the same principles as before. A coset of 4Z can be specified by N/2 complex-valued coset representatives ck' where coset representatives take on one of 16 possible values, corresponding to 4 integer-spaced values (modulo 4) for the real and imaginary parts of ck' respectively. The general picture of Figure 8 then holds, except that coset selector 58 and range-identifying parameter selector 64 select N/2 complex-valued coset representatives ck and range-identifying parameters ak' and the signal point selector operates once per 15 quadrature signal and puts out complex-valued signals Coset precoding as in Figure 9 is done by forming the complex-valued precoded coset c'k (modulo 4) once per quadrature symbol. RDS .feedback as in Figures 11, 12, 13 is done by using a 20 function R(ak) that identifies a region of complex space of area 16 that contains exactly one element from any coset of 4Z 2 and a complex-valued translation variable R(xk ideally equal to 8xk-1. In the 2 S2 cases where 2Z or 2RZ is a sublattice of A', precoding can be done modulo 2 or 2 2i, respectively, and R(ak) can identify a region of area 4 or 8 containing exactly one element from any coset of 2Z 2 2 or 2RZ respectively.

Higher-Dimensional Systems 30 We have shown embodiments in which coordinates of N-dimensional symbols are formed on a signal-by-signal (one or two-dimensional) basis, with signal-by-signal feedback of the previous RDS value Similar kinds of performance can be obtained by r 1 I St..

I

I 4 44 1 4 1 4: S Ca II S S 'SI 39 systems which select signals on a higher-dimensional basis. In such systems, the precoded coset representatives must be grouped as in Figure 9 so as to select subsets in the appropriate dimension, signal points then selected in that dimension, and the coordinates then serialized again for transmission over the channel. If the order of cosets is maintained, then such a system retains the property that the PRC sequences are from the given code and have the specified 2 dmin In such a system it may be more natural to do (RDS) feedback on a higher-dimensional basis rather than on each signal.

N-Dimensional Codes Although representation of codes in one-dimensional form is desirable, it is not essential.

In this section we show how codes may be generated directly in N dimensions. In certain forms, the N-dimensional code is entirely equivalent to its one-dimensional counterpart. In other forms, simplified embodiments may be obtained.

Again, we shall use for illustration the 8-state 2-dimensional Ungerboeck-type code of Figure 2, with the 2-dimensional 128-point constellation of Figure 3. In this constellation, recall that each coordinate 25 takes on values from the alphabet of the 12 half-integral values in the range from -6 to 6; the two-dimensional constellation uses 128 of the 144 possible pairwise combinations of elements of this alphabet.

As a first step, we expand the signal constellation to an infinite number of values, as follows. Let the expanded constellation consist of all pairs of numbers that are congruent to some point in the original (Figure 3) constellation (modulo 12). Thus the L-o points in the expanded constellation consist of pairs of half-integral values. If we regard the original constellation as a cell bounded by a 12 x 12 square 98, then the expanded constellation consists of the infinite repetition of this cell throughout 2-space, as indicated diagrammatically in Figure 28. Note that each cell contains only 128 of the 144 possible points; there are 4 x 4 'holes' 99 in the expanded constellation.

The key property of this expanded constellation 101 is that if we place a 12 x 12 square anywhere in the plane (with sides oriented horizontally and vertically), the square will enclose exactly 128 points, one congruent to each of the points in the original constellation, An even more general statement is true: 15 if we place a rhombus 102 with horizontal width 12 and tit vertical height 12 (see Figure 29) anywhere in the plane, it too will enclose 128 points, one congruent to S* each point in the original constellation.

Referring to Figure 30, we may now implement RDS feedback on a two-dimensional basis as follows. Let xk- 1 represent the running digital sum of all yk previous to the current (two-dimensional) symbol. Let R(xkl) now denote a region of the plane corresponding to a 12 x 12 rhombus as in Figure 29, with both the shape and the location of the rhombus possibly depending on Let (Y 0 YQ,k+) denote the point in the original constellation that would be selected (in selectors 104, 105) by the three coded bits and four uncoded bits according to the unconstrained code (Figure Then (in selector 106) select (Yk' yk+l) as the unique point in the two-dimensional expanded constellation that lies within the egion R(xk1_) and is congruent to (y0,k Yk+l (modulo I v .4 4- 12); these will be the two coordinates yk. We can obtain (xk, xk+l) from x k Yk Xk-1, Xk+l Yk+l Xk as shown.

We now sho6 that this two-dimensional system can produce the same outputs as the one-dimensional RDS feedback system (modulo 12) shown earlier, with the optimal one-dimensional RDS feedback variable R(xkl) 3 Referring to Figure 31, in one dimension, given Xk-1l Yk is chosen as the unique value in the range RO 8xk_- Xk- 1 congruent to s k (modulo 12), where we now recognize that s k is congruent to yo,k Hence, one coordinate of the rhombus used in the two-dimensional system can be taken to lie in the same width-12 range. Then, given Xk_ 15 and yk, and thus also xk Yk Xkl, Yk+ is chosen as the unique value in the range R 0 (l-8)xk

R

0 (1- 3 )yk (l-8)xk_ 1 that is congruent to Sk+l y0,k+l (modulo 12). Thus, Yk+l lies in the range R 0 (l-8)xk- 1 (same as yk)' shifted by 20 -(l-B)yk' Thus, by proper choice of rhombus, we can emulate the performance of a one-dimensional (modulo 12) RDS feedback system with a two-dimensional system. It will thus have the same advantages, including avoidance of error propagation and near-optimal tradeoff between

SS

x Sy and S o and the same disadvantages, notably the increase in S O to 12 over the 10.25 otherwise ,possible.

We can choose other two-dimensional RDS a a o. 30 feedback variables (regions) to further simplify implementation, and achieve other advantages, at the cost of suboptimal power tradeoffs. For example, a system almost identical to the simplified one-dimensional system described earlier results if we 1

V

4-3 1 let R(xk_l) be the square 120 of side 12 centered at when Xk_ is positive, and the square 122 centered at when Xkl is negative. Thus we use one of the two constellations 124, 126 shown schematically in Figure 32.

As in the previous one-dimensional system, inner points are always chosen from the same set regardless of Xk_, but outer points are varied so as to bias Yk in a positive or negative directicn. The i0 ranges of the Yk are strictly limited from -7 1/2 to 7 1/2. In fact, this system is identical to the earlier simplified system, except that Yk+l is chosen on the basis of Xkl rather than xk. In practice, all the I measures of performance and spectrum will be very I 2. 15 similar.

i. Another variant yields a system akin to that of i the Calderbank, Lee, and Mazo type. A CLM-type system uses an expanded signal constellation with twice the ,a ordinary number of signal points, divided into two disjoint constellations, one to be used when Xkl is positive, and the other when xk_ 1 is negative. Figure 33, for example, shows a 16 x 16 square constellation divided into two disjoint constellations 110, 112 of 128 points each, such that each such constellation divides evenly into 8 subsets of 16 points each. One constellation consists of points the sum of whose coordinates is positive or zero and is used when Xk_ 1 Sis negative; the other consists of points whose coordinate sums are negative or zero and is used when Xkl 1 is positive. In two dimensions, doubling the constellation size doubles Sy and thus does not yield a favorable power tradeoff; however, in higher dimensions the penalty due to the use of two disjoint constellations is less.

r"i k L i VT

^QA

13 -4 These ideas can be generalized to N dimensions, as follows. If there is a one-dimensional formulation of the code as in Figure 8, using modulus M, then an N-cube of side M completely surrounds the N-dimensional constellation, and the resulting cell can be replicated to cover N-space without compromising the minimum squared distance between code sequences that are congruent to original code sequences modulo M. Then we may use an N-dimensional RDS feedback function R(xk-l) where for all Xk_,1 R(xk-) is a region of N-space of volume MN that contains exactly one point in each equivalence class of N-vectors modulo M, in an N-dimensional analogue of Figure Other embodiments are within the following 15 claims.

t4*o o 4 0 So o 0* 0 4 9 a 44 4

Claims (31)

1. Apparatus for generating a sequence of digital signals Xk and/or a sequence of digital signals k 1, 2, such that the sequence of Yk signals is a partial- response-coded sequence derived from the sequence of Xk signals, said signals Yk being a sequence in a given modulation code, said apparatus comprising a coset selector for generating coset representatives c k in accordance with said given modulation code; and an encoder for selecting J said signals Yk, J 1, (Yk, Yk+1, k+J-1) to be congruent to a sequence of J coset representatives c k (modulo A, being an N- dimensional lattice, N being a positive integer, said J signals being chosen from a plurality of NJ-dimensional constellations, said choice being based on a previous xk, at least one of said plurality of NJ-dimensional constellations comprising both a point with a positive sum of coordinates and another point with a negative I 15 sum of coordinates, said encoder being arranged so that said signals Xk have finite I j variance S x

2. The apparatus of claim 1 wherein the relationship between Xk signals and Yk signals is yk Xk L an integer.

3. The apparatus of claim 1, wherein said encoder comprises a generator of a sequence of alternative coset representatives ck' chosen so that the sequence of coset representatives c, is a partial-response-coded sequence derived from the sequence of ck' signals, and a selector for selecting said signals Xk to be congruent to said sequence of alternative coset representatives c k (modulo

4. The apparatus of claim 1 wherein said x k and y, sequences have variances S, and Sy and wherein the ratio of variance Sy to variance S x is selectable within a predetermined range. The apparatus of claim 3 wherein the relationship between the xk signals 920102,kxlspe.003,12040.r,44 i 1- j 45 and Yk signals is y, Xk L an integer and wherein Ck' Ck c'k-L (modulo in the case when Yk Xk Xk-, ck' C, C'k-L (modulo in the case where Yk Xk Xk-L.

6. The apparatus of claim 4 wherein the relationship between the x, signals and Yk signals is Yk x, Xk-. L an integer.

7. The apparatus of claim 6 wherein said x k sequence and/or said yk sequence is capable of representing n bits per signal, wherein said Yk signals fall within an alphabet of possible Yk signals that are spaced apart within said alphabet evenly by a spacing A, and wherein said encoder is capable of causing said sequence yk to have a variance S, less than 2S o and said sequence Xk to have a variance S, not much greater than S, 2 S o being approximately the minimum signal power S. required to represent n bits per signal with a A-spaced alphabet.

8. The apparatus of claim 6 wherein said x, and y, signal sequences are capable of representing n bits per signal, said Yk signals fall within an alphabet of V possible y, signals that are spaced evenly by an amount A, said predetermined range is controlled by a parameter P, and S. is approximately and Sy is approximately S o being approximately the minimum signal power required to represent n bits per symbol with a A-spaced alphabet in accordance V with said code.

9. The apparatus of claim 7 wherein said sequence yk is a sequence in a given modulation code. The apparatus of claim 2, wherein the sequence y, is a sequence of one- dimensional signals in a given N-dimensional modulation code, said modulation code being based on an N-dimensional constellation partitioned into subsets associated with said code, said subsets each containing N-dimensional signal points, the choice of said subset being based on coded bits and uncoded bits of said signal points, and wherein said encoder is adapted to derive from said coded and 911017lspe.003,12040.e,45 4 -r -I C 46 uncoded bits, for each said N-dimensional symbol, a set of N, M-valued one- dimensional coset representatives c, corresponding to congruence classes of each of the N coordinates (modulo each coset representative designating a subset of one-dimensional values in a one-dimensional constellation of possible coordinate values for each of said N dimensions, each said one-dimension signal in said sequence being selected from said possible coordinate values based on uncoded bits.

11. The apparatus of claim 2, 5, 6 or 7 further comprising an output at which said sequence y, is delivered. i! S12. The apparatus of claim 2, 5, 6 or 7 further comprising an output at which Ssaid sequence Xk is delivered. i

13. The apparatus of claim 2, 5, 6 or 7 wherein L is 1.

14. The apparatus of claim 2, 5, 6 or 7 wherein the relationship between the Xk S,3 signals and Yk signals is yk xk L an integer.

15. The apparatus of claim 2, 5, 6, 9 or 10 wherein said modulation code is a trellis code.

16. The apparatus of claim 2, 5, 6, 9 or 10 wherein said modulation code is a lattice code.

17. The apparatus of claim 2, 5 or 9 wherein M is 2.

18. The apparatus of claim 2, 5 or 9 wherein M is 4.

19. The apparatus of claim 2, 5 or 9 wherein M is a multiple of 4. C

20. The apparatus of claim 2 wherein J is 1. 911017)kisepe 23,1200.rc,46 47

21. The apparatus of claim 2 wherein J is the same as the number of dimensions of said modulation code.

22. The apparatus of claim 2 wherein k' is k-1.

23. The apparatus of claim 2 wherein J is 1 and each said constellation is a one-dimensional range of values centred on Xk-1, 0 1.

24. The apparatus of claim 23 wherein p 0. The apparatus of claim 2 wherein there are a finite set of said J- dimensional constellations.

26. The apparatus of claim 25 wherein there are two said J-dimensional constellations.

27. The apparatus of claim 2, 5, 6 or 7 wherein y, and xk are real valued.

28. The apparatus of claim 2, 5, 6 or 7 wherein y, and xk are complex valued.

29. The apparatus of claim 2 or 5 wherein yk and x, are complex valued and wherein M is 2+2i, i=V-1. The apparatus of claim 2 wherein at least two of said J-dimensional constellations are not disjoint.

31. A decoding apparatus for decoding a sequence z Yk nk, k 1, 2, into a decoded sequence yk, where the sequence of signals Yk is such that said sequence is from a given modulation code, generated by the encoding apparatus of claim 2; the running digital sum x, yk1 Yk-2 has finite variance S,; S(c) said signals y, fall in a predetermined permissible range dependent 911017,kIspe.003,12040.r 47 r 48 on xk, k' k; and the sequence n k represents noise; said decoding apparatus including a range violation monitor comprising: a means for reconstructing an estimated running digital sum Xk y, Yk- 1 and a means for comparing said decoded sequence yk with said predetermined permissible range based on said estimated running digital sum x, k' k, and for generating an indication when said Yk is outside said permissible range.

32. A decoding apparatus according to claim 31, in which said estimated running digital sum xk is adjusted based on said indication so that y will be inside said permissible range.

33. A decoding apparatus according to claim 32, in which said adjustment is by the minimum possible amount such that y, falls inside said permissible range.

34. A decoding apparatus for decoding a sequence xk y, nk, k 1, where sequence nk represents noise and the sequence of signals Yk is such that said sequence is from a given modulation code generated by the encoding apparatus of claim 2 with a finite number Q of states; and Y Xk Xk-L, L an integer, where said sequence xk has finite variance and the sequence nk represents noise; said decoding apparatus comprising: a means for receiving the sequence z; and a modified maximum likelihood sequence estimator responsive to the receiving means, said estimator being adapted to find MQ partial decoded sequences, up to some time K, where M, Q, and K are positive finite numbers, one such said sequence for each combination of said finite number Q of states and each of a finite number M of integer-spaced values modulo M, such that each said sequence is in said code up to said time K; 920102,bkxlspc.003,12040.r,48 I 49 corresponds to said encoder being in a given said state at said time K; and corresponds to a value of x, at said time K that is congruent to a given one of said values, modulo M. Decoding apparatus as in claim 34 wherein M is 2.

36. Decoding apparatus as in claim 31 wherein M is 4.

37. An apparatus for generating a sequence of digital signals substantially as hereinbefore described with reference to the drawings. age ttt Dated this 17th day of October, 1991 CODEX CORPORATION By its Patent Attorneys DAVIES COLLISON 911017 isPca)3,12040.re,49