Method and apparatus for spectral shaping in signal-point limited transmission systems

[0002] Not Applicable

[0003] Not Applicable

[0004] Not Applicable

[0005] The present invention relates to transmission of digital information over analog medium connected to a digital network and more particularly to the PCM modems (“Pulse Code Modulation”).

[0006] The world based on the Internet has seen tremendous growth in recent months. As more users begin browsing and downloading information from the World Wide Web, there has been a great desire to be able to increase the data rate. The desire is even greater for users accessing the Internet through an Internet service provider (ISP), since most of the users are linked up to the Net through a personal computer and a modem. Conventional analog modems, such as the V.34 modems, however, view the public switched telephone network (“PSTN”) as an analog channel, even though the signals are digitized for communications throughout most of the network. As such, various effects and impairments due to quantization impose a limitation on the data rate of the channel to about 35 Kbps. This limit has been commonly known as Shannon's Limit. (See Shannon, C. E. and W. Weaver, The Mathematical Theory of Communication,University of Illinois Press, 1949).

[0007] There has been a recent development of a high-speed communications technology based on the PCM modems, where data rate of at least 56 Kbps is said to be actually attainable. The PCM modem technology is based on the simple realization that the public switched telephone networks (PSTN) is increasingly a digital network and not an analog network. Also, more and more central site modems are connected to the PSTN through digital connections, i.e. T1 in the U.S. and E1 in Europe, without utilizing a CODEC (coder/decoder). The conventional modem, however, still interprets this digital stream as the representation of the modem's analog signal. With the PCM modems, a much higher data rate can be achieved without the complicated task of re-wiring the user's site or modifying the telephone network.

[0008] Note that by “central site modems,” it is referred to those modems installed at an ISP, or at a corporation to allow many simultaneous connections for remote LAN access. Also note that a CODEC is a device which sits between the digital portion of the network and the analog local loop for converting between analog and digital.

[0009] The recent 56 Kbps technology seeks to address an impaired section of the communications path of the PSTN digital network, where the impairment is due to the hybrid and the copper wire connection between the telephone central office and the user's home, usually referred to as the analog local loop.

[0010] Since recently, much has been described about the PCM modems and how they can and should facilitate downstream data communication at a much higher rate than present paradigm. For example, the PCM modem has been the subject of a recent Telecommunications Industry Association (TIA) Technical Committee TR-30 Standards meeting on Oct. 16-17, 1996. The submitted technical contributions include Mr. Guozhu Long's DC Suppresser for56K Modems,Mr. David C. Rife's 56 Kbps Channels,Mr. Veda Krishnan's V.pcm Modem Standard,Mrs. Vedat Eyuboglu's PCM Modems: A Technical Overview,Mr. Richard Stuart's Proposal for a High Speed Network Access Modem,and Mr. Vladimir Parizhsky's U.S. Robotics' x2 Technology: Technical Brief.These contributions are hereby incorporated by reference.

[0011] Also, there have been recent publications on the overall data communication system based on the PCM modem. The first one is a 1995 presentation disclosed by Pierre A. Humblet and Markos G. Troulis at Institute Eurecom, entitled The Information Driveway,1995, which purports to explain the basic concepts on the high speed modem. The second one is a PCT Patent Publication, dated Jun. 13, 1996, International Publication Number WO/9618261, by Brent Townshend, which discloses a High Speed Communications Systems for Analog Subscriber Connections. This Publication, on pages 17-19, discloses an overall high speed system based on the PCM modems, which also implements DC null elimination on the transmitter side. These are also hereby incorporated by reference, since they provide a fair reference to the basics of the high speed PCM modems and their environment.

[0012] Additionally, U.S. patent issued to Ender Ayanoglu of AT&T, U.S. Pat. No. 5,528,625 dated Jun. 18, 1996, entitled High Speed Quantization-Level-Sampling Modem with Equalization Arrangement, discloses a QLS modem for high-speed data communication. Another U.S. patent also issued to Ender Ayanoglu of AT&T, U.S. Pat. No. 5,394,437, dated Feb. 28, 1995, entitled High-Speed Modem Synchronized to a Remote CODEC, discloses a high-speed modem for data transmission over an analog medium in tandem with a digital network. These references are also hereby incorporated by reference.

[0013] FIG. 1 depicts a conceptual diagram of the high-speed communication path using the PCM modem technology. An ISP, or “central site”, 100 is digitally connected to a telephone network 130 through its transmitter 110 and receiver 120. The network 130 is connected to a local loop 150 through a central office line card 140. The line card typically has a Pulse Code Modulation (“PCM”) CODEC implemented within. The local loop 150 is connected to the user's PC at the user's site through the user's modem 160. As can be appreciated by those skilled in the art, the connection between the ISP modem's transmitter 110 to the telephone network 130 is a digital connection with a typical data rate of about 64 Kbps. Since the parameters of the telephone network 130 and line card 140 are dictated and set by telephone company's specification and operation, the central site transmitter 110 will need to transmit the digital data in certain way to fully exploit its digital connection to the network. However, dealing with the central site transmitter in this new paradigm has its obstacles.

[0014] In this type of data communication systems, such as the transmitter 110 in the central site, the transmit signal points are determined by physical constraints, and cannot be made part of the overall transmitter design. An example is when a signal is transmitted from within the digital part of the telephone network to the residential customer. That system will be the focus of the following description, although the methods described hereinafter will in many cases be applicable to other systems.

[0015] The transmitter 110 sends 64 Kbps of data into the network 130, which eventually gets translated to an analog signal in a digital-to-analog converter in a central office line-card CODEC 140. To send the maximum of 64 Kbps of data, all transmitted bits must be determined by the incoming data. Any modification of the outgoing sequence of bits resulting in correlation between the bits represents a redundancy in the signal, and will thus result in a lower data rate. Therefore, if the analog signal is to be controlled in any way, the data rate must be lowered. The main task confronting the designers, then, is to perform the desired control with the least amount of redundancy added.

[0016] As will be described in the present application, methods and apparatus for controlling the spectrum of the line codec's output signal are disclosed. In particular, methods of minimizing the energy of the signal at unwanted frequencies are described. These methods can be readily extended to other parts of the spectrum by those skilled in the art based on the teaching of the present invention.

[0017] Impairments Presented by the Communication Channel

[0018] There are other difficulties from the channel which affect the implementation of the PCM modem and the information transmission. The communication channel from the line card PCM CODEC to the user's modem can be characterized by 4 primary functional units: the codec's anti-aliasing filter, the line card circuit, the subscriber loop and the modem's input circuit.

[0019] First, the CODEC anti-aliasing filter is mostly flat, starting to attenuate the signal at about 3.5 kHz and increasing to 15-20 dB at 4 kHz.

[0020] Second, the line card circuit is in some ways the most challenging impairment. It is difficult to characterize, since there are many types in use, and it has been designed to transmit a voice signal rather than a data signal. The hybrid transformer circuit includes a null at 0 Hz, and may induce high levels of nonlinear distortion, especially on signals with low-frequency energy.

[0021] Third, the subscriber loop itself is generally a linear impairment, creating inter- symbol interference (ISI) and picking up largely Gaussian noise. Non-Gaussian noise sources are near-end crosstalk (“NEXT”), far-end crosstalk (“FEXT”) and single frequency interference (SFI). The prime source of SFI is the electricity distribution network, adding a 50 or 60 Hz tone to the signal.

[0022] Finally, the modem input circuit consists of a line interface, including a hybrid transformer, and an analog-to-digital converter (ADC). The hybrid introduces a null in the spectrum at 0 Hz, and may also induce nonlinearities. However, since the signal level at the modem end is lower than at the central office, and the transformer can be part of the receiver design, nonlinear distortion in the modem input circuit will generally be minimal compared to that added by the central office line card. While the ADC will of course add quantization noise, which can be made negligible by design, many ADC's will add Gaussian-type noise, often with a “1/f” spectral characteristic, i.e. with highest energy at the low-frequency end of the spectrum. Also, a DC offset can be expected in the ADC, requiring some compensation by the receiver.

[0023] Given these impairments, it might seem difficult to take advantage of the low-end of the spectrum. The two transformer hybrids introduce nulls at 0 Hz, SFI is mostly around 50-60 Hz, and nonlinear distortion is increased if energy is transmitted at low frequencies (and will extend into higher frequencies). ADC noise also tends to be most prevalent at low frequencies.

[0024] Thus, it is desirable to simplify the receiver design by having the transmitter avoid sending energy at low frequencies. Without the low frequency energy, the receiver may elect to simply implement a high-pass filter to filter out noise there without the need to reconstruct the signal component removed.

[0025] In the following disclosure of the present invention, methods will be disclosed for controlling the sequence of bits transmitted to the line card CODEC to eliminate as much energy as possible at low frequencies. For illustration purposes, the following disclosure will assume a redundancy of 4 Kbps, where 4 Kbps out of the 64 Kbps are generated based on the other bits. Generally, a higher redundancy may improve the performance, and the methods disclosed herein can readily be modified for a different level of redundancy.

[0026] A method and apparatus for performing spectral shaping transmitted samples with a set of predetermined frequency characteristics and a predetermined set of allowable transmitted signal levels are disclosed. The method first calculates, for each of the transmitted samples, a running measure of unwanted components upto the current sample. It then computes, for each block of the transmitted samples, an objective function based on the running measure previously calculated. It then selects, for each block of the transmitted samples, at least one redundant sample to be added or at least one transmitted sample to be modified, at structured or randomized location within the block of samples to optimize the objective function. The location of the dependent sample can further be fixed, scrambled (pseudo random) or randomized. The method of computing can be any one of the following: a Running Filter Sum, a Running Fourier Transform (“RFT”) or Fast FT (“RFFT”), or an RDS, for each transmitted sample.

[0027] Additional objects, advantages and features of the present invention will be disclosed in the following description, wherein:

[0028] FIG. 1 depicts a conceptual diagram of the high-speed communication path using the PCM modem technology.

[0029] FIG. 2 illustrates a block diagram of the environment where the spectral shaping methodology in accordance with one aspect of the present invention is applicable.

[0030] FIG. 3 illustrates a generalized process flow of the lossless spectral shaping methodology in accordance with the present invention.

[0031] FIG. 4 illustrates the signal spectrum with simple redundant-sample method, where the null is at 0 Hz.

[0032] FIG. 5 illustrates the signal spectrum with simple redundant-sign method, where the null is at 0 Hz.

[0033] FIG. 6 illustrates the signal spectrum with simple redundant-sample method, where the null is now at 60 Hz.

[0034] FIG. 7 illustrates the signal spectrum with simple redundant-sign method, where the null is at 60 Hz.

[0035] FIGS. 8 and 9 illustrate the signal spectrum with multiple frequency redundant-sample and multiple frequency redundant-sign methods, respectively, with nulls at 0 Hz and 60 Hz.

[0036] FIGS. 10 and 11 illustrate the signal spectrums with redundant-sign block code method, where block size is 4 and nulls at 0, 60 Hz and 0, 30, 60 Hz, respectively.

[0037] FIGS. 12 and 13 illustrate the signal spectrum for the block size of 8, inserting nulls at 0, 60 Hz and 0, 6, 12, 18, . . . , 60 Hz, respectively.

[0038] FIG. 14 illustrates the signal spectrum of the sequence constructed by the redundant-sign method for a block length of N=4 with a redundancy of {fraction (1/16)}.

[0039] FIG. 15 illustrates the signal spectrum of the sequence constructed by the redundant-sample method.

[0040] Methods and apparatus for spectral shaping in signal-point limited transmission systems are disclosed. In the following description, various aspects of the present invention are presented in terms of process flows, functional block diagrams and even mathematical expressions, which are the ways those skilled in the art communicate with each other in conveying their ideas and designs.

[0041] FIG. 2 illustrates a block diagram showing one environment where the spectral shaping in accordance with one aspect of the present invention is applicable. Source samples 200 from a data source, such as the ISP's server, are transmitted by a transmitter 210 (of an ISP modem) to generate transmitted samples 220. The transmitted samples 220 are transmitted, through a channel 240, to the local loop and eventually the user 250.

[0042] Those skilled in the art will recognize that the channel 240 is PCM-based with a limited set, i.e. 255 in μ-law or A-law, of signal levels, since it is specified by the telephone networks. Those skilled in the art will also appreciate that the transmitter 210 must be one using lossless coding such that no information is lost between source samples 200 and transmitted samples 220. To avoid sending the low-frequency energy to the channel 240, the low-frequency component must be eliminated (230) as the source samples 200 are coded by the transmitter 210 based on a set of characteristics of the frequency spectrum. In other words, the running digital sum (“RDS”) of any block of source samples must be minimized, which means that the low frequency energy is partially offset or entirely eliminated. Also, there must be redundancy added to ensure a lossless coding by the transmitter.

[0043] The conventional way of making RDS bounded is to use an Alternate Marking Input (“AMI”) method, which generates additional, but opposite, voltage levels for the samples to bring the RDS to within a symbol size. The AMI, however, is unacceptable for PCM channels, since there is a fixed set of signal points, or voltage levels for the PCM channel as dictated by the telephone networks.

[0044] Reference is to FIG. 3, where a generalized process flow of the spectral shaping methodology in accordance with one aspect of the present invention is shown. First, a block of source samples 300 is applied to the transmitter for coding based on a set of predetermined characteristics of the transmit frequency spectrum. For each sample, a predetermined component is measured (310).

[0045] Broadly, the measure of the unwanted signal components of the transmitted samples upto the current sample can be accomplished using a Running Filter Sum (RFS). If the corresponding filter coefficients are selected to be a Fourier basis function, then the measure becomes Running Discrete Fourier Transform (RDFT) or Running Fast Fourier Transform (RFFT). It is more convenient to use RDFT or RFFT to form a set of frequency null. If only D.C. is of concern, then the measure reduces to RDS.

[0046] For every block of a predetemined number of transmitted samples, an objective function is computed (320) based on the measure already obtained. Note that the objective functions may be computed using any one of the following well-known techniques:

[0047] a) L-2 Norm of the measurement;

[0048] b) L-1 Norm of the measurement;

[0049] c) L-α Norm (for 1<α<2);

[0050] d) a weighted version of L-2 Norm;

[0051] e) a weighted version of L-1 Norm;

[0052] f) a weighted version of L-α Norm (for 1<α<2).

[0053] Finally, for every block of source samples, at least one redundant sample is added or at least one source sample is modified (330), e.g. the sign of the sample is modified, such that it will optimize the objective function. When the objective function is optimized, the unwanted signal component is minimized, e.g. the energy of the unwanted signal component is minimized.

[0054] It should be noted that the source samples are not intended to be limited to real values. However, for illustration purposes in the disclosure, only real numbers are considered.

[0055] It should be apparent to those skilled in the art that when applied to the energy at 0 Hz, the methodology will minimize the unwanted DC component in the transmitted samples. As can be appreciated by those skilled in the art, the methodology will provide arbitrary spectral-shaping, or coding, of the transmitted samples, whatever the characteristics of the transmit spectrum require.

[0056] The following description will first disclose the present invention in the context of spectral shaping in a simple DC null, followed by spectral shaping in multiple frequency nulls and then eventually to spectral pre-coding based on a predetermined spectrum shape. For the purposes of illustration in the following description, transmitted samples can be classified as source samples and dependent samples, where the source samples are unmodified and the dependent samples are either redundant samples or modified source samples, e.g. sign-modified, or both. In all measures described herein, the location of the dependent sample can be anywhere in the block, structurally (including pseudo-randomized) or randomly. If randomly, the location index must be transparent to the receiver, i.e. encoded into the transmitter sequence. For simplicity of the description below, we assume the locations of the dependent samples remain the same for all the blocks. Those skilled in the art can readily adapt the present invention for those scrambled locations based on the teaching of the present invention.

[0057] A. A Simple DC Null.

[0058] In accordance with the present invention, the effect of the DC nulls of the transformer hybrids at the PCM CODEC can be avoided by creating a DC-null in the transmit spectrum.

[0059] 1) Redundant Sample

[0060] This can be done by letting every L^th, e.g. L=16, sample counteract the DC introduced by previous samples.

[0061] Given the output signal x_n, let us define a DC measurement c_n, $c_n=\sum _{m=0}^n{x}_m.$

[0062] This is commonly known as the RDS approach. If we, when n is a multiple of L, let x_n=Q(31 c_n−1), where Q(.) quantizes its input to the closest value in the PCM digital-to-analog converter, the DC component of the signal will be greatly attenuated, i.e.: $x_n=\{\begin{array}{cc}M\left(d_n\right)& \mathrm{if}n\ne \mathrm{Lr}\\ Q\left(-c_{n-1}\right)& \mathrm{if}n=\mathrm{Lr}\end{array},r\text{is a non-negative integer,}$

[0063] where M is a suitable mapping function from data to a subset of the PCM quantization values. FIG. 4 illustrates the signal spectrum with simple redundant-sample method, where the null is at 0 Hz.

[0064] 2) Redundant Sign

[0065] Another method is to modify the sign of every L, e.g. L=2, sample based on c_n−1, then, $x_n=\{\begin{array}{cc}M\left(d_n\right)& \mathrm{if}n\ne \mathrm{Lr}\\ \mathrm{sgn}\left(-c_{n-1}\right)M^{\prime }\left(d_n\right)& \mathrm{if}n=\mathrm{Lr}\end{array},r\text{is a non-negative integer,}$

[0066] where M′ only selects positive quantization values. Clearly, d_nshould preferably have one less bit when n=L k.

[0067] FIG. 5 illustrates the signal spectrum with simple redundant-sign method, where the null is at 0 Hz.

[0068] As can be observed from FIGS. 4 and 5, both methods create a DC-null in the transmit spectrum without much effect on the spectrum above 100 Hz. The receiver will not have to reconstruct the signal component at DC, thus greatly simplifying the receiver circuit.

[0069] B. Creating Nulls at Individual Frequencies.

[0070] 1) Single Frequencies

[0071] The above-described methods can be generalized, or extended, to insert nulls (or rather, near-nulls) at any frequencies. Let us define c_n^(k)(now a complex number) as, $c_n^{\left(k\right)}=\sum _{m=0}^n{x}_m{e}^{-{\mathrm{j\omega }}_{k}m},$

[0072] where Ω_krepresents the frequency 𝑓_k=Ω_k/2πT with symbol period T. If we make every L^th, e.g. L=16, sample dependent, then we select the dependent samples to minimize ${\left|C_n^{\left(k\right)}\right|}^2.$

[0073] It should be noted that L is not restricted to 16 in any sense and those skilled in the art can readily apply a different value in accordance with their special applications. It can readily be shown that the best selection for x_nis then, $x_n=Q\left(-\mathrm{Re}\left[c_{n-1}^{\left(k\right)}{e}^{{\mathrm{j\omega }}_{k}n}\right]\right).$

[0074] It is then possible to have a generalization for the DC-null redundant-sample method as: $x_n=\{\begin{array}{cc}M\left(d_n\right)& \mathrm{if}n\ne \mathrm{Lk}\\ Q\left(-\mathrm{Re}\left[c_{n-1}^{\left(k\right)}{e}^{{\mathrm{j\omega }}_{k}n}\right]\right)& \mathrm{if}n=\mathrm{Lk}\end{array},e.g.L=16.$

[0075] FIG. 6 illustrates the signal spectrum with simple redundant-sample method, where the null is now at 60 Hz.

[0076] A generalization for the DC-null redundant-sign method is now, $x_n=\{\begin{array}{cc}M\left(d_n\right)& \mathrm{if}n\ne \mathrm{Lr}\\ \mathrm{sgn}\left(-\mathrm{Re}\left[c_{n-1}^{\left(k\right)}{e}^{{\mathrm{j\omega }}_{k}n}\right]\right)M^{\prime }\left(d_n\right)& \mathrm{if}n=\mathrm{Lr}\end{array},e.g.L=2.$

[0077] FIG. 7 illustrates the signal spectrum with simple redundant-sign method, where the null is at 60 Hz. As can be observed by those skilled in the art, if Ω_k=0, this becomes the same as the special case with the DC null.

[0078] 2) Multiple Frequencies

[0079] The generalized method as described above can in fact be done at more than one ii frequency. If we define a set of Ω_k, k=0,1, . . . ,K−1, and minimize the sum of the (k) squared norms of c_n^(k), it can be found that the best selection for x_nfor the redundant sample method is, $x_n=\{\begin{array}{cc}M\left(d_n\right)& \mathrm{if}n\ne \mathrm{Lr}\\ Q\left(-c_{n-1}\right)& \mathrm{if}n=\mathrm{Lr}\end{array},$ $\mathrm{where}{C}_n=\frac{1}{K}\sum _{k=0}^{K-1}\mathrm{Re}\left[c_{n-1}^{\left(k\right)}{e}^{{\mathrm{j\omega }}_{k}n}\right],\mathrm{and}\mathrm{for}\mathrm{the}\mathrm{redundant}\mathrm{sign}\mathrm{method},$ $x_n=\{\begin{array}{cc}M\left(d_n\right)& \mathrm{if}n\ne \mathrm{Lr}\\ \mathrm{sgn}\left(-c_{n-1}\right)M^{\prime }\left(d_n\right)& \mathrm{if}n=\mathrm{Lr}\end{array}.$

[0080] If the goal is to suppress a band of frequencies, then those skilled in the art can readily replace summation (Σ) with integration (∫) in calculating C_n.

[0081] FIGS. 8 and 9 illustrate the signal spectrum with multiple frequency redundant-sample and multiple frequency redundant-sign methods, respectively, with nulls at 0 Hz and 60 Hz.

[0082] C. Block Coding of Redundant Samples

[0083] The performance of the above-described redundant-sample method can be further enhanced by considering blocks of data and jointly optimizing several redundant samples.

[0084] If we make every L^th, e.g. L=16, samples redundant, then we can consider jointly optimizing N redundant samples in a super block of (N×16) samples. In other words, we would select x_n, x_n−16, . . . , x_{n−(N−1)*16}simultaneously to minimize the sum of squared norms of c_n^(k). It should be appreciated by those skilled in the art that the N jointly optimized redundant samples x_n, x_n−16, . . . , x_{n−(N−1)*16}, without quantization, are roots of the following linear equation: A x=b, where $A=\left\{\begin{array}{cccc}K& \sum _{k=0}^{K-1}\cos \left(16{\omega }_k\right)& \cdots & \sum _{k=0}^{K-1}\cos \left(16\left(N-1\right){\omega }_k\right)\\ \sum _{k=0}^{K-1}\cos \left(16{\omega }_k\right)& K& \cdots & ⋮\\ ⋮& \cdots & ⋰& \sum _{k=0}^{K-1}\cos \left(16{\omega }_k\right)\\ \sum _{k=0}^{K-1}\cos \left(16\left(N-1\right){\omega }_k\right)& \cdots & \sum _{k=0}^{K-1}\cos \left(16{\omega }_k\right)& K\end{array}\right\}$ $x=\left\{\begin{array}{c}{x}_n\\ x_{n-16}\\ ⋮\\ x_{n-16\left(N-1\right)}\end{array}\right\}$ $b=-\left\{\begin{array}{c}\sum _{k=0}^{K-1}\mathrm{Re}\left[c_{n-1}^{\left(k\right)}{e}^{{\mathrm{j\omega }}_{k}n}\right]\\ \sum _{k=0}^{K-1}\mathrm{Re}\left[c_{n-1}^{\left(k\right)}{e}^{{\mathrm{j\omega }}_k\left(n-16\right)}\right]\\ ⋮\\ \sum _{k=0}^{K-1}\mathrm{Re}\left[c_{n-1}^{\left(k\right)}{e}^{{\mathrm{j\omega }}_k\left(n-\left(N-1\right)16\right)}\right]\end{array}\right\}$

[0085] Here the term $c_{n-1}^{\left(k\right)}\mathrm{is}{c}_{n-1}^{\left(k\right)}=\sum _{\underset{\underset{\dots n-16\left(N-1\right)}{m\ne n-16,n-32,}}{m=0}}^{n-1}{x}_m{e}^{-{\mathrm{j\omega }}_{k}m},$

[0086] which is slightly different from before.

[0087] When each entry of x can only take value from a finite set of quantization values, the exact solution of x will be replaced by its quantized version. A simple way to get a quantized version of roots of above linear equation is independently quantizing each of the entries. A more preferred way will be jointly quantizing the whole vector by searching through a small neighborhood of x and finding the quantized candidate vector which minimizes the sum of the squared norms of c_n^(k).

[0088] It should readily be appreciated by those skilled in the art that using FFT to perform coding of redundant samples is a trivial extension of the above-described methodology. In that case, $c_n^{\left(k\right)}=\sum _{m=0}^n{x}_m{e}^{-{\mathrm{j\omega }}_{k}m}$

[0089] will be replaced by $c_p^{\left(k\right)}=\sum _{m=0}^{P-1}{x}_m{e}^{-j\frac{2\pi k}{p}m},$

[0090] where c_P^(k)corresponds to the FFT at the actual frequency of 𝑓_k=k/PT.

[0091] D. Block Coding of Redundant Signs

[0092] The Redundant Sign method as described above can be further improved by considering blocks of data. Several signs within the block can be selected in combination to optimize. For example, if we consider a block of 4 samples, and first select positive quantization points for each sample based on the input data. Assuming a redundancy of {fraction (1/16)}, we then need to encode 2 more bits for the block, and have 2 degrees of freedom. We now want to select the set of 4 signs S_n=(s_n,s_n+1,s_n+2, s_n+3) to minimize the energy at the selected frequencies 𝑓_k, given x′_n+l=M′(d_n+l), l=0,1,2,3 and $c_{n-1}^{\left(k\right)},k=0,1,\dots ,K-1.$

[0093] In effect, we want to direct the vector X′_n=(x′_n,x′_n+1,x′_n+2,x′_n+3) in such a way as to minimize C_n. We are restricted by 2 bits, but the other 2 bits can be selected. Suppose the set of possible S_nis divided into 4 subsets A_i, i=0,1,2,3, and a subset is selected based on the input bits, then a S_nis selected out of the 4 subset members to minimize C_n. Then we would want the elements of each subset to be as different from each other as possible, in order to handle a variety of different X_nwell.

[0094] One way to spread the S_nis to construct a binary code, where each subset has members with maximum distance between them, and the subsets are cosets of each other. One example would be:

[0095] A₀=(0000,0011,1100,1111)

[0096] A₁=(0001,0010,1101,1110)

[0097] A₂=(0100,0111,1000,1011),

[0098] where the binary codes represent the sign

[0099] A₃=(0101,0110,1001,1010)

[0100] combinations where a “0” represents “+”, and a “1” represents “−”. Here, the 2 data bits would select one of the 4 subsets, and then the sign combinations in each subset would be tested to find which one gave the lowest C_n.

[0101] That sign combination would then be used to multiply the positive input samples. In the receiver, the decoding operation would identify the signs. The receiver would then find which subset the sign combination belonged to, and output the 2 bits corresponding to that subset.

[0102] As can be observed by those skilled in the art, an interesting side-benefit is that for each subset, the negative of each member of the subset is also in the subset. Thus, if the signal gets multiplied by −1 in the channel, the resulting subset sequence is unchanged. Therefore, no differential encoding/decoding is necessary, the coding system is sign-invariant.

[0103] FIGS. 10 and 11 illustrate the signal spectrums with redundant-sign block code method, where block size is 4 and nulls at 0, 60 Hz and 0, 30, 60 Hz, respectively.

[0104] As can be expected, the larger the subset, the better the performance. One example of a subset structure for a block length of 8, where one still assumes a redundancy of {fraction (1/16)}, is illustrated in Table 1 (now using hexadecimal notation).

[0105] FIGS. 12 and 13 show the signal spectrum for the block size of 8, inserting nulls at 0, 60 Hz and 0, 6, 12, 18, . . . , 60 Hz, respectively.

[0106] If more attenuation is desired, the block size can be increased to 16 or even higher, although the computational complexity will grow considerably. The redundancy can also be increased. For a redundancy of ⅛, all sign-bits are available for manipulation. This will increase the low-frequency attenuation over that shown in FIGS. 9 and 10 by more than 10 dB.

[0107] E. Spectral Precoding According to a Specified Spectral Shape

[0108] The foregoing methods can be further generalized to shape transmitted sequence spectrum with respect to a specified spectral characteristic. Let the z-transform of the desired spectral shaping function be G(z) and its inverse be defined as $F\left(z\right)=1/G\left(z\right)=\sum _{m=-\infty }^{\infty }{f}_m{z}^{-m},$

[0109] where the filter coefficient, 𝑓_m, can be complex-valued.

[0110] In practice, we might be only interested in the cases where F(z) is causal, i.e., 𝑓_m=0, for all m<0, or we may use a delayed and windowed version of F(z). In addition, we might prefer F(z) to be minimum phase. This can be done by selecting an all-pass filter A(z) such that F(z) is minimum phase and $F\left(z\right)=\frac{A\left(z\right)}{G\left(z\right)},$

[0111] where |F(e^−jΩ)|=1/|G(e^−jΩ)| and F(z) is assumed to be causal and stable. Once we have obtained the desired filter coefficients {𝑓_m}, the unwanted signal component contained in the available sequence {x_m} for 0≦m≦n at epoch n can be written as: $c_n=\sum _{m=0}^n{x}_m{f}_{n-m}.$

[0112] Its energy can be reduced or completely canceled by jointly optimizing several redundant samples which will be selected with respect to the dominant weights of 𝑓_mor jointly selecting several redundant signs. These two methods are similar to the aforementioned block coding with redundant samples and block coding with redundant signs.

[0113] The only difference is in the performance measure. To state clearly, we define the set of L dominant tap indices of 𝑓_mas . Then for each block of N samples, where suppose N>max{},we select L redundant samples x_n−l, ∀lε subject to saturation and quantization constraints of PCM signaling or select L redundant signs s_n−l, ∀lε such that |c_n|²is minimized. The search of the optimal redundant samples or signs is straightforward. Any of the constrained linear programming algorithms would be sufficient, as can be appreciated by those skilled in the art.

[0114] For illustration purposes, we consider a special case where the most dominant tap is 𝑓₀. Further, we assume 𝑓₀is real-valued without loss of generality. For this case, it is reasonable to modify the current sample x_naccording to c_n−1to minimize |c_n|². The parameter c_n−1here is referred to as a running filter sum (RFS) at time n−1. If the “redundant-sample” method is employed for each block of N samples, then the desired sequence can be constructed as $x_n=\{\begin{array}{cc}M\left(d_n\right)& \mathrm{if}n\ne \mathrm{rL}\\ Q\left(-c_{n-1}/f_0\right)& \mathrm{if}n=\mathrm{rL}\end{array}.$

[0115] If the “redundant-sign” method is used, then the desired sequence is given as $x_n=\{\begin{array}{cc}M\left(d_n\right)& \mathrm{if}n\ne \mathrm{rL}\\ \mathrm{sgn}\left(-c_{n-1}/f_0\right)M^{\prime }\left(d_n\right)& \mathrm{if}n=\mathrm{rL}\end{array},$

[0116] where M′(d_n)=|M(d_n)|. It can be shown that as long as the RFS is bounded, the constructed sequence satisfies the specified spectral requirements. Note that if we choose G(z)=1−z⁻¹, then 𝑓_m=1 for all m≧0. The RFS then reduces to RDS. If we choose G(z)=1−e^−jΩz⁻¹, then 𝑓_m=e^−jΩmfor all m≧0 and a given normalized frequency Ω. The RFS reduces to RFT.

[0117] In order to demonstrate the effectiveness of the various methods in accordance with the present invention, we consider G(z)=1−z⁻², a so-called modified duobinary signal. This class of sequences does not possess a sharp discontinuity at Nyquist frequency and as a result, it can be easily approximated by a physically realizable filter. It can be shown that the corresponding inverse filter is given as, $F\left(z\right)=\sum _{m=0}^{\infty }{z}^{-2m},$

[0118] where we select (ROC) |z|<1. Accordingly, we can use two separate RDS' to account for the even samples and the odd samples respectively, $c_n=\{\begin{array}{cc}\sum _{m=0}^{n/2}{x}_{2m}& \mathrm{if}n\mathrm{is}\mathrm{even}\\ \sum _{m=0}^{\left(n-1\right)/2}{x}_{2m+1}& \mathrm{if}n\mathrm{is}\mathrm{odd}\end{array}.$

[0119] The sequence {x_n} can be constructed by any of the aforementioned methods.

[0120] FIG. 14 shows the power spectrum of the sequence constructed by the redundant-sign method for a block length of N=4 with a redundancy of {fraction (1/16)}.

[0121] The last example considers G(z)=1−αz⁻¹where the weighting factor α or forgetting factor can be used to control the memory of the residual running digital sum at the cost of slightly lower attenuation at DC. FIG. 15 gives the power spectrum of such sequence constructed by the redundant sample method (N=8, α=0.99).

[0122] Although only a few exemplary embodiments of this invention have been described in detail above, those skilled in the art will readily appreciate that many modifications are possible in the exemplary embodiments without materially departing from the novel teachings and advantages of this invention. Accordingly, all such modifications are intended to be included within the scope of this invention as defined in the following claims. In the claims, means-plus-function clauses are intended to cover the structures described herein as performing the recited function and not only structural equivalents but also equivalent structures. Thus although a nail and a screw may not be structural equivalents in that a nail employs a cylindrical surface to secure wooden parts together, whereas a screw employs a helical surface, in the environment of fastening wooden parts, a nail and a screw may be equivalent structures.