1. Introduction
Due to the recent technological advancements in sensing technology and the internet of things, there is a growing demand for low-energy communications solutions. Indeed, since many of the sensors have only limited batteries due to environmental (in the case of energy harvesting) or replenishment limitations, these solutions need to be economical in terms of the utilized energy. Moreover, since each sensor may serve several parties, with each experiencing different conditions, these solutions need to be robust with respect to the noise level.
This problem may be conveniently modeled as the classical setup of conveying
k independent and identically distributed (i.i.d.) Gaussian source samples with minimum mean square error (MMSE) distortion over a continuous-time additive white Gaussian noise (AWGN) channel under a channel input total energy constraint
, where
E is the allowed transmit energy per source sample and unconstrained transmit bandwidth; see
Figure 1.
For the encapsulated source–coding problem for a large
k, the optimal tradeoff between the compression rate
R and the (per-sample) MMSE distortion
D [
1] (Chapter 13.2) for a memoryless Gaussian source with variance
is dictated by the rate–distortion function [
1] (Chapter 13.3):
where
is the signal-to-distortion ratio (SDR).
For the encapsulated channel-coding problem, since the bandwidth is unconstrained (i.e., grows to infinity), and the allowed energy of the channel input is constrained by
E, the maximal achievable total reliable rate (in nats) of the entire transmission—the total capacity—is given by [
1] (Chapter 9.3)
when the power spectral density of the noise (the
noise level)
N is known to the transmitter (and the receiver), and where
is the energy-to-noise ratio. We note that, in our setting, the transmit energy
E is fixed regardless of the transmission duration and bandwidth, in contrast to the power-limited setting, in which the energy
grows linearly with the transmission time for a fixed power
P. To emphasize this, following [
2,
3] and others, we make use of
to distinguish it from the more common signal-to-noise ratio (SNR), which is defined in the fixed-power scenario as
.
Returning to the overall problem of conveying
k i.i.d. source samples of a Gaussian source over a continuous-time AWGN channel subject to an energy constraint (and unconstrained bandwidth), in the limit of a large-source blocklength
k, the optimal achievable mean square error distortion per source sample is dictated by the celebrated source–channel separation principle [
1] (Th. 10.4.1), [
4] (Chapter 3.9):
, which upon substituting (
1) and (
2), amounts to
For non-Gaussian continuous memoryless sources, the optimal distortion is bounded as [
1] (Prob. 10.8, Th. 10.4.1), [
4] (Prob. 3.18, Chapter 3.9)
where the lower bound stems from Shannon’s lower bound [
5], the upper bound holds since a Gaussian source is the “least compressable” source with a given variance under a quadratic distortion measure, and
denotes the differential entropy of a sample of the i.i.d. source
x [
1] (Chapter 8), [
4] (Chapter 2.2).
While the optimal performance is known when the transmitter (and the receiver) knows the noise level and
, determining it becomes much more challenging when the noise level is unknown at the transmitter. Indeed, when the transmitter is oblivious of the true noise level, achieving (
3) for all noise levels simultaneously is impossible [
6,
7]. Instead, one wishes to achieve graceful degradation of the distortion with the noise level, namely, a scheme that would work well for a continuum of all possible noise levels without knowing the true noise level at the transmitter. Since the distortion improves exponentially with the ENR (
3)
when the noise level is known, in the absence of knowledge of the noise level at the transmitter, one might hope to attain an exponential distortion decay profile with the ENR of the form
for some
, or, equivalently,
for some
and some finite per-sample energy
E. Köken and Tuncel [
7] proved that, unfortunately, this is impossible, namely, no
(equivalently
and
) exist for which (5a,b) is achievable simultaneously for all
(equivalently, for all
). Consequently, distortion profiles that deteriorate faster with the noise level need to be sought.
For the case of
finite bandwidth expansion/compression
B (and finite power), by superimposing digital successive refinements [
8] with a geometric power allocation, Santhi and Vardy [
9,
10] and Bhattad and Narayanan [
11] showed that, in our terms, the distortion improves like
for an arbitrarily small
, for
large ENR values. This suggests that by taking the bandwidth to be large enough, a polynomial decay with an
of any finite degree, however large, is achievable, starting from a large enough ENR. In our setting of interest, in which the bandwidth is unconstrained, this means, in turn, that there exists a finite energy
E for which a
polynomially decaying distortion profile in
N
is attainable for any predetermined power
, however large yet finite, and for any predetermined constant
of our choice, with large enough finite per-sample energy
; for the particular choice of
for any constant
of our choice, this is equivalent to
Mittal and Phamdo [
12] constructed a different scheme, that works above a certain minimum (not necessarily large) design ENR by sending the digital successive refinements incrementally over non-overlapping frequency bands, and sending the quantization error r of the last digital refinement over the last frequency band.
The scheme of Mittal and Phamdo was subsequently improved by Reznic et al. [
6] (see also [
13,
14], [
15] (Chapter 11.1)) by replacing the successive refinement layers with lattice-based Wyner–Ziv coding [
16,
17], [
4] (Chapter 11.3) which, in contrast to the digital layers of the scheme of Mittal and Phamdo, enjoys an improvement in each of the layers with the ENR.
Kokën and Tuncel [
7] adopted the scheme of Mittal and Phamdo in the infinite-bandwidth (and infinite-blocklength) setting. Baniasadi and Tuncel [
18] (see also [
19]) further improved this scheme by allowing the sending of the resulting analog errors of all the digital successive refinements. For the case of a distortion profile that improves quadratically with the ENR (
in (6a,b)), upper and lower bounds were established by Köken and Tuncel [
7] and Baniasadi and Tuncel [
18] (see also [
19]) for the minimum required energy to attain such a profile for all ENR values. For a predetermined value of our choice
and a Gaussian source, a quadratic distortion profile (
6a) with a predefined constant
(and
) is achievable with a minimal per-sample transmit energy
E that is bounded as
A staircase profile was treated by Baniasadi [
20] (see also [
19]).
However, albeit much progress has been made in determining the minimal required energy to attain polynomially decaying distortion profiles (
6a), with particular emphasis put on the quadratically decaying distortion profile corresponding to the choice
, the upper and lower bounds in (
7) remain far apart. Moreover, no (low-delay) schemes for a single-source sample
with graceful degradation of the distortion with the ENR have been proposed.
In this work, we adapt the modulo-lattice modulation (MLM) scheme of Reznic et al. [
6] with multiple layers to the infinite-bandwidth setting, and interpret previously decoded layers that are designed for lower ENRs as side information that is known to the receiver but not to the transmitter, which allows, in turn, to apply Wyner–Ziv coding techniques [
13], [
15] (Chapter 11). By utilizing linear modulation for all the layers, we show that this scheme improves the upper (achievability) bound in (
7). We then replace the analog modulation in (some of) the layers with analog pulse-position modulation (PPM) which was shown to work well for known ENR in [
21]. We show that this scheme requires less energy to attain the same quadratic distortion profile compared to the linear-layer-only MLM scheme. Finally, we demonstrate numerically that a low-delay variant of the scheme, which encodes a single-source sample
and uses simple one-dimensional lattices, attains good universal performance with respect to the noise level.
We note that our analytic results rely on the well-established existence of good multi-dimensional lattice codes [
15] (to be precisely defined in
Section 3), which are used as a building block, along with their known theoretical guarantees. Therefore, our proposed schemes should be understood with this point in mind. That said, for a suboptimal lattice with poorer analytical guarantees, one can similarly calculate the (suboptimal) achievable performance of the scheme. Since lattices work well even in one dimension, we demonstrate the strength of the proposed technique explicitly for this practical scenario using a simple one-dimensional lattice, which amounts to a uniform grid.
The rest of the paper is organized as follows. We introduce the notation that is used in this work in
Section 1.1, and formulate the problem setup in
Section 2. We provide the necessary background of MLM and analog PPM—the two major building blocks that are used in this work—in
Section 3 and
Section 4, respectively. We then construct universal schemes with respect to the noise level in
Section 5; simulation results of our analysis for good multi-dimensional lattices and of the empirical performance of single-dimensional lattices are provided in
Section 6. Finally, we conclude the paper with
Section 7 and
Section 8 by discussing future research directions and possible improvements.
1.1. Notation
,
,
, and
denote the sets of the natural, integer, real, and non-negative real numbers, respectively. With some abuse of notation, we denote tuples (column vectors) by
for
, and their Euclidean norms by
, where
denotes the transpose operation; distinguishing the former notation from the power operation applied to a scalar value will be clear from the context. The i-th element of the vector
is denoted by
or by
, where we will use both terms throughout the paper. All logarithms are to the natural base, and all rates are measured in nats. The differential entropy of a continuous random variable with probability density function
f is defined by
and is measured in nats. The expectation of a random variable (RV)
x is denoted by
. We denote by
the modulo-
L operation for
, and by
the modulo-
operation [
15] (Chapter 2.3) for a lattice
[
15] (Chapter 2).
denotes the floor operation. We denote by
the
k-dimensional identity matrix. We denote sets of vectors by capital italic letters, where
stands for a set of
c vectors, each of length
b. All the logarithms in this work are to the natural base and all the rates are measured in nats.
2. Problem Statement
In this section, we formulate the JSCC setting that will be treated in this work, depicted in
Figure 1.
Source. The source sequence to be conveyed, , comprises k i.i.d. samples of a standard Gaussian source, namely, it has mean zero and variance .
Transmitter. Maps the source sequence
to a continuous input waveform
that is subject to an energy constraint. (The introduction of negative time instants yields a non-causal scheme. This scheme can be made causal by introducing a delay of size
. We use a symmetric transmission time around zero for convenience):
where
E denotes the per-symbol transmit-energy.
where
P is the transmit-power and
T is the transmission duration.
Channel. is transmitted over a continuous-time additive white Gaussian noise (AWGN) channel:
where
n is a continuous-time AWGN with two-sided spectral density
, and
r is the channel output signal;
N is referred to as the noise level.
Receiver. Receives the channel output signal r, and constructs an estimate of .
Distortion. The average quadratic distortion between
and
is defined as
where
denotes the Euclidean norm, and the corresponding signal-to-distortion ratio (SDR) by
since we assumed
. For non-i.i.d. samples, the variance
should be replaced by the effective variance
which clearly reduces to the (regular) variance in the case of i.i.d. zero-mean samples.
Regime. We concentrate on the energy-limited regime, viz. the channel input is not subject to a bandwidth constraint, but rather to an energy constraint
E per
source symbol (
8). As explained in the Introduction, the per-source-symbol capacity of the channel (
9) is equal to [
1] (Chapter 9.3)
where
, and the capacity is measured in nats; note that the available bandwidth is unconstrained (i.e., infinite).
Since the available bandwidth is unlimited, the receiver can learn the white noise level within any accuracy. Hence, we may assume that the receiver has exact knowledge of the channel conditions. The transmitter is oblivious of the noise level, and needs to accommodate for a continuum of noise levels. Specifically, we will require the distortion to satisfy (6a,b). Throughout most of this work we will concentrate on the setting of infinite blocklength (
). We will also conduct a simulation study for the scalar-source setting (
) in
Section 6.
3. Background: Modulo-Lattice Modulation
The overall scheme, to be introduced and analyzed in
Section 5, comprises two major components (in addition to components in the form of interleaving and “Gaussiaization” that are needed for analysis purposes) as depicted in Figure 3:
A component that assumes an additive noise vector channel of the same dimension as the source input k with unknown noise level, and constructs a layered hybrid digital–analog universal solution with respect to this noise level, where each layer accommodates a different noise level, where an estimator constructed from all the layers that were designed for larger noise levels acts as SI that is known at the receiver;
A component that modulates a single analog-source sample over a continuous-time AWGN channel which transforms the channel effectively into a one-dimensional additive channel (one channel use of a discrete-time channel), is designed for a certain noise level but attains graceful improvement if the noise level happens to be better.
Therefore, in this section, we provide the necessary background about the first component. This is a succinct background about lattices and modulo-lattice modulation which is needed to understand the machinery that is used in the proposed solutions in
Section 5, along with known performance guarantees that are relevant to this work and are needed for the analysis of the performance guarantees that are claimed in this work. Readers who are less familiar with lattices, lattice coding, and MLM are referred to the well-regarded book of Zamir on this subject [
15]. Background about the second component is provided in
Section 4.
A k-dimensional lattice is a discrete regular array in the Euclidean space that is closed under reflection and real addition.
Definition 1 (Lattice [
15] (Def. 2.1.1))
. A non-degenerate k-dimensional lattice Λ is defined by a set of k linearly independent basis (column) vectors , and define the generator matrix G:The lattice Λ is composed of all integral combinations of the basis vectors:In particular, the origin belongs to the lattice: . Figure 2 provides examples of one- and two-dimensional lattices. A lattice induces a quantization and a partition of the space into cells, with each cell comprising all points that are closest to a specific lattice (quantization) point. These cells are referred to as
Voronoi cells.
Definition 2 (Nearest-neighbor quantizer and Voronoi cell [
15] (Chapter 2.2))
. The nearest-neighbor quantizer induced by a k-dimensional lattice Λ is defined asThe Voronoi cell is the set of all points that are quantized to : is referred to as the fundamental Voronoi cell. The breaking of ties in (16) is carried out in a systematic manner so that the induced Voronoi cells are congruent. In particular, for , where the first sum is the Minkowski sum of and the singleton ;
for , where ;
.
We next define the modulo-lattice operation with respect to the fundamental Voronoi cell.
Definition 3 (Modulo-lattice [
15] (Chapter 2.3))
. For a k-dimensional lattice Λ with a fundamental Voronoi cell , the modulo-lattice operation (with respect to ), applied to , is defined asnamely, the outcome equals the (unique) point that satisfies for some . We now define the volume, the second moment, and the normalized second moment of a lattice.
Definition 4 (Volume and second moment [
15] (Chapter 2 and 3))
. The volume of a k-dimensional lattice Λ with fundamental Voronoi cell is defined as the volume of :The second moment of Λ is defined as the second moment per dimension of a random variable that is uniformly distributed over :The normalized second moment of Λ is defined asTo attain a good MMSE using lattice quantization, should be as close as possible to the normalized second moment of a k-dimensional ball which, in the limit of , converges to . Since the effective source in intermediate layers (this will become clear in the following section) that we would like to transmit and the effective channel noise which is induced by the analog modulations over the continuous-time channel are not Gaussian in general (even after “Gaussianization” which would make them only approximately so), we would need to consider more general source and channel noise vectors that satisfy the following definition of semi-norm ergodicity (SNE).
Definition 5 (SNE [
22] (Def. 2))
. A sequence in k of random vectors of length k with a limit norm . (The original definition of [22] (Def. 2) requires for all . We use here a more relaxed definition which will prove more convenient in the following section):is SNE if for any , however small, there exists a large enough , such that for all We are now ready to describe the
k-dimensional JSCC setting and the MLM technique with side information (SI) for this setting. In the overall solution of
Section 5, the analog modulations over the continuous-time channel that will be described in
Section 4 will translate the channel into an effective
k-dimensional additive SNE noise channel (compare also the subfigures of Figure 4 in
Section 5. Over this effective channel, MLM with SI will be employed, where we will treat previous source estimators as effective side information (SI) known to the receiver but not to the transmitter [
13] and [
15] (Chapter 11).
Source. Consider a source sequence (equivalently, vector)
of length
k,
where
is an SI sequence which is known to the receiver but not to the transmitter, and
is the “unknown part” (at the receiver) with per-element variance
and is SNE (as a sequence in
k).
Transmitter. Maps
to a channel input,
, that is subject to a power constraint
Channel. The channel is an additive noise channel:
where
is an SNE noise vector that is uncorrelated with
and has effective variance
The SNR is defined as
; we use here the more common SNR notion in lieu of the ENR notion to emphasize that the channel and the source vectors (equivalently, sequences) in this section are of the same dimension
k, in contrast to the continuous-time channel of
Section 2.
Receiver. Receives , in addition to the SI , and generates an estimate of the source .
The following MLM-based scheme will be employed in the following.
Scheme 1. [MLM-based JSCC with SI [
13], [
15] (Chapter 11)]
Transmitter: Transmits the signal
where
is a lattice with a fundamental Voronoi cell
and a second moment
P,
is a scalar scale factor, and
is a dither vector which is uniformly distributed over
and is independent of the source vector
; consequently,
is independent of
by the so-called crypto lemma [
15] (Chapter 4.1).
Receiver:
Receives the signal
(
27) and generates the signal
where
is the equivalent channel noise, and
is a channel scale factor.
Generates an estimate
:
where
is a source scale factor.
When
in (
30) falls within
, the modulo operation does not come into play, resulting in an effective additive noise channel from
to
. Thus, we want the probability of this “correct lattice decoding” event to be bounded from below by
for some small
. On the other hand, conditioned on the correct lattice decoding event, we want the quantization noise, which is governed by
, and consequently by the shape of
, to have a small normalized second moment to be good for MMSE estimation. The following theorem provides guarantees for the achievable distortion using this scheme and is aggregated from [
13], [
15] (Chs. 11.3, 6.4, 9.3), and [
22] (see also the exposition about correlation-unbiased estimators (CUBEs) in [
23]).
Theorem 1. The distortion (10) of Scheme 1 is bounded from above byfor , and that satisfywhere is the distortion given a lattice-decoding-error event [13] (Equation (24)) and is bounded from above byand the lattice parameters and are defined asMoreover, for any , however small, and any , there exists a sequence of lattices, , that are good for both channel coding [22] (Def. 4) and mean squared error (MSE) quantization [22] (Def. 5), viz.,respectively, and, therefore, this sequence of lattices achieves a distortion that approaches . Remark 1. By our definition of SNE sequences, for each finite k the actual variance of the unknown part and the noise variance may be higher than for every higher than their asymptotic quantities. Consequently, also the second moment of for every would be taken to be higher than its asymptotic value.
That said, as k grows to infinity, these slacks become negligible and the performance converges to that of (32), (38). The following choice of parameters is optimal in the limit of infinite blocklength,
, in the Gaussian case (
comprises i.i.d. Gaussian samples,
comprises i.i.d. Gaussian samples) [
4] (Chapter 11.3) when the SNR is known.
Corollary 1 (Optimal parameters [
13], [
15] (Chapter 11.3))
. The choice , , , , yields a distortion D that is bounded from above as in (32) withwhereMoreover, for any , however small, there exists a sequence of lattices that attains (38) and, therefore, in the limit , and above converge to and the distortion D approaches , which converges, in turn, to Consider now the setting of an SNR that is unknown at the transmitter but is known at the receiver. In this case, although the receiver knows the SNR and can, therefore, optimize
and
accordingly, the transmitter, being oblivious of the SNR, cannot optimize
for the true value of the SNR. Instead, by setting
in accordance with Corollary 1 for a preset minimal-allowable-design SNR,
, Scheme 1 achieves (
44) for
and improves, albeit sublinearly, with the SNR for
. This is detailed in the next corollary.
Corollary 2 (SNR universality)
. Assume that for some predefined . Then, the choice , , and with respect to (as it cannot depend on the true SNR), and and (may depend on the true SNR) yields a distortion D that is bounded from above, as in (32) for that is given in (39) with . Moreover, for any , however small, there exists a sequence of lattices that satisfies (38); therefore, in the limit , converges to , converges to , and the distortion D approaches , which converges, in turn, to Corollary 3 (Source power uncertainty)
. Assume now additionally that the transmitter is oblivious of the exact power of , , but knows that it is bounded from above by : . Then, the distortion is bounded according to (32), withfor the parametersMoreover, for any , however small, there exists a sequence of lattices that attains (38) and, therefore, in the limit of , converges to , converges to , and the distortion D is bounded from above in this limit by :where ϵ decays to zero with . For , the bound (48c) approaches . The following result is a simple consequence of Theorem 1 and avoids exact computation of the optimal parameters.
Corollary 4 (Suboptimal parameters)
. Assume the setting of Corollary 3 but with not necessarily uncorrelated with , and denote . Then, the distortion is bounded according to (32) withfor the parameters , .We refer to by since now may depend on .
The following property will prove useful in
Section 5 when treating non-Gaussian noise through “Gaussianization”.
Lemma 1 ([
24] (Lemmas 6 and 11))
. Let be a sequence of lattices that satisfies the results in this section, and let be a dither that is uniformly distributed over the fundamental Voronoi cell of . Then, the probability density function (p.d.f.) of is bounded from above aswhere is the p.d.f. of a vector with i.i.d. Gaussian entries with zero mean and the same second moment P as , and decays to zero with k. 4. Background: Analog Modulations in the Known-ENR Regime
Following the exposition at the beginning of
Section 3 and
Figure 3, we concentrate now on the second major component that is used in this work, that of analog modulations for conveying a scalar zero-mean Gaussian source (
) over a channel with infinite bandwidth, where both the receiver and the transmitter know the channel noise level, or equivalently,
. To that end, we will review next the analog linear modulation and the analog PPM and will supplement the known results for the latter with a new robustness result for a source distribution that deviates from Gaussianity in Corollary 6.
Consider first analog linear modulation, in which the source sample
x is linearly transmitted with energy
E, (under linear transmission, the energy constraint holds only on average, and the transmitted energy is equal to the square of the specific realization of
x) using some unit-energy waveform
Note that linear modulation is the same (“universal”) regardless of the true noise level. Signal space theory [
25] (Chapter 8.1), [
26] (Chapter 2) suggests that a sufficient statistic of the transmission of (
51) over the channel (
9) is the one-dimensional projection
y of
r onto
:
where
z is a standard Gaussian noise variable. The MMSE estimator of
x from
y is linear and its distortion is equal to
and improves only linearly with the
.
Consider now analog PPM, in which the source sample is modulated by the shift of a given pulse rather than by its amplitude (which is the case for analog linear modulation):
where
is a predefined pulse with unit energy and
is a scaling parameter. In particular, the square pulse (Clearly, the bandwidth of this pulse is infinite. By taking a large enough bandwidth
W, one may approximate this pulse to an arbitrarily high precision and attain its performance within an arbitrarily small gap) is known to achieve good performance. This pulse is given by
for a parameter
, which is sometimes referred to as
effective dimensionality. Clearly,
.
The optimal receiver is the MMSE estimator
of
x given the entire output signal:
The following theorem provides an upper bound on the achievable distortion of this scheme using (suboptimal) maximum a posteriori (MAP) decoding, which is given by
where
is the (empirical) cross-correlation function between
r and
with lag (displacement)
, and
is the autocorrelation function of
with lag
.
Remark 2. Since a Gaussian source has infinite support, the required overall transmission time T is infinite. Of course, this is not possible in practice. Instead, one may limit the transmission time T to a very large—yet finite—value. This will incur a loss compared to the the bound that will be stated next; this loss can be made arbitrarily small by taking T to be large enough.
Theorem 2 ([
21] (Prop. 2))
. The distortion of the MAP decoder (57) of a standard Gaussian scalar source transmitted using analog PPM with a rectangular pulse is bounded from above bywithbounding the small- and large-error distortions, assuming . In particular, in the limit of large , and β that increases monotonically with ,whereand in the limit of . Remark 3. For a fixed β, the distortion improves quadratically with the . This behavior will prove useful in the next section, where we construct schemes for the unknown-ENR regime.
Setting
in (
60) of Theorem 2 yields the following asymptotic performance.
Corollary 5 ([
21] (Th. 2))
. The achievable distortion of a standard Gaussian scalar source transmitted over an energy-limited channel with a known ENR is bounded from above aswhere as . The following corollary, whose proof is available in
Appendix A, states that the (bound on the) distortion is continuous in the source p.d.f. around a Gaussian p.d.f. Such continuity results of the MMSE estimator in the source p.d.f. are known [
27]. Next, we prove the required continuity directly for our case of interest with an additional technical requirement on the deviation from a Gaussian p.d.f.; this result will be used in conjunction with a non-uniform variant of the Berry–Esseen theorem in
Section 5.
Corollary 6. Consider the setting of Theorem 2 for a source p.d.f. that satisfieswhere , is the standard Gaussian p.d.f., and is a symmetric absolutely continuous non-negative bounded function with unit integral , that is monotonically decreasing for (and for , by symmetry) and satisfies ; thus, there exists such that Then, the distortion of the decoder that applies the decoding rule (57) is bounded from above bywhere denotes the bound on the distortion for a standard Gaussian source of Theorem 2, and is a non-negative constant that depends on . (This is no longer the MAP decoding rule since is no longer a Gaussian p.d.f.). 5. Main Results
In this section, we construct JSCC solutions for the unknown-ENR communication problem. As already explained at the beginning of
Section 3, the proposed solution, which is depicted in
Figure 3 (cf.
Figure 1), is composed of two major components:
Aa layered MLM-based component that works well for a continuum of possible noise levels over k-dimensional additive SNE noise channels, where each layer accommodates a different noise level, with layers of lower noise levels acting as SI in the decoding of subsequent layers;
An analog modulation component that is designed for a particular ENR of the continuous-time channel but improves for high ENRs and induces a k-dimensional additive SNE noise channel for the first component.
Following the exposition in the introduction, since an exponential improvement with the ENR cannot be attained in this setting for an infinite number of noise levels let alone a continuum thereof [
7], following [
7,
18], we consider polynomially decaying profiles (6a,b).
We first show, in
Section 5.1, that replacing the successive refinement coding of [
7,
18] with MLM (Wyner–Ziv coding) with
linear layers results in better performance in the infinite-bandwidth setting (paralleling the results of the bandwidth-limited setting [
6]).
In
Section 5.2, we replace the last layer with an analog PPM one, which improves quadratically with the ENR (
in (
6b)) above the design ENR (recall Remark 3).
In principle, despite analog PPM attaining a gracious quadratic decay with the ENR (recall Remark 3) only above a predefined design ENR, since the distortion is bounded from above by the (finite) variance of the source, it attains a quadratic decay with the ENR for all , or equivalently, for all and in (6a,b).
That said, the performance of analog PPM deteriorates rapidly when the ENR is below the design ENR of the scheme, meaning that the minimum energy required to obtain (
6a) with
and a given
is large. To alleviate this, we use the above-mentioned layered MLM scheme. Furthermore, to achieve higher-order improvement with the ENR (
in (6a,b)), multiple layers in the MLM scheme need to be employed.
We now present a simplified variant of the general scheme that is considered throughout this section. This variant is also depicted in
Figure 4a. The full scheme, which incorporates interleaving for analytical purposes, is available in
Appendix B and depicted in
Figure A1.
M-Layer Transmitter:
First layer ():
Transmits each of the entries of the vector
over the channel (
9) linearly (
51):
for
, where
is a continuous unit-norm (i.e., unit-energy) waveform that is zero outside the interval
, say
of (
55),
is the allocated energy for layer 1, and
E is the total available energy of the scheme.
Other layers: For each :
Calculates the
k-dimensional tuple
where
, and
denotes the
entry of
;
,
, and
take the roles of the
, and
of Scheme 1, and are tailored for each layer
i;
is chosen to have unit second moment.
For each
, views
as a scalar-source sample, and generates a corresponding channel input,
using a scalar JSCC scheme with a predefined energy
that is designed for a predetermined
, or equivalently,
, such that
and
.
Receiver: Receives the channel output signal
r (
9), and recovers the different layers as follows.
First layer (): For each :
Recovers the MMSE estimate of given , where .
If the true noise level N satisfies , sets the final estimate of to and stops. Otherwise, determines the maximal layer index for which and continues to process the other layers.
Other layers: For each in ascending order:
For each
, uses the receiver of the scalar JSCC scheme to generate an estimate
of
from
, where
Using the effective channel output
(that takes the role of
in Scheme 1) with SI
, generates the signal
as in (
30) of Scheme 1, where
is a channel scale factor.
Constructs an estimate
of
:
as in (
31) of Scheme 1, where
is a source scale factor. The final estimate if
.
Remark 4 (Interleaving)
. To guarantee independence between all the noise entries , we use interleaving in the full scheme, which is described in Appendix B in (A8) and (A11). We note that this operation is used to simplify the proof that the resulting noise vector is SNE (recall Definition 5). Remark 5 (Gaussianization)
. To use the analysis of Section 4 of analog PPM for a Gaussian source, we multiply the vectors by orthogonal matrices that effectively “Gaussianize” its entries, as shown in the full description of the scheme in Appendix B, in (A8) and (A11). In particular, this is achieved by a Walsh–Hadamard matrix by appealing to the central limit theorem; a similar choice was previously proposed by Feder and Ingber [28], and by Hadad and Erez [29], where in the latter, the columns of the Walsh–Hadamard matrix were further multiplied by i.i.d. Rademacher RVs to achieve near-independence between multiple descriptions of the same source vector (see [29,30,31] for other ensembles of orthogonal matrices that achieve a similar result). Interestingly, the multiplication by the orthogonal matrices (since Walsh–Hadamard matrices are symmetric, they further satisfy ) Gaussianizes the effective noise incurred at the outputs of the analog PPM JSCC receivers. Remark 6 (JSCC-induced channel)
. The continuous-time JSCC transmitter and receiver over the infinite-bandwidth AWGN channel induce an effective additive-noise channel of better effective SNR and source bandwidth. Over this induced channel, the MLM transmitter and receiver are then employed. This interpretation is depicted in Figure 4b, with representing the effective additive noise vectors. We next provide analytic guarantees for this scheme for linear and analog PPM layers in
Section 5.1 and
Section 5.2, respectively, in the infinite-blocklength regime. In
Section 6, we compare the analytic and empirical performance of these schemes in the infinite-blocklength regime, as well as comparing the empirical performance of these schemes for a single-source sample. The treatment of the infinite-blocklength regime pertains to the full scheme as presented in
Appendix B. The comparison for a single-source sample, uses the simplified variant of Scheme 2.
5.1. Infinite-Blocklength Setting with Linear Layers
We start with analyzing the performance of the scheme where all the
M layers are transmitted linearly and
M is large; we concentrate on the setting of an infinite-source blocklength (
) and derive an achievability bound on the minimum energy that achieves a polynomial distortion profile (6a,b). A constructive proof of the next theorem is available in
Appendix C. In particular, this proof specifies all the scheme parameters, such as the energy allocated to each layer and the minimal noise level it is designed for.
Theorem 3. Choose a decaying order , a design parameter , and a minimal noise level , however small. Then, a distortion profile (6a) with L and is achievable for all noise levels for any transmit energy E that satisfiesfor a large-enough-source blocklength k, whereIn particular, the choice achieves a quadratic decay () for any transmit energy E that satisfiesfor a large-enough-source blocklength k. We note that already this variant of the scheme offers an improvement compared to the hitherto best-known upper (achievability) bound of (
7).
The choice of the minimal noise level dictates the number of layers M that need to be employed: the lower is, the more layers M need to be employed.
Remark 7. In the proof in Appendix C, we use an exponentially decaying noise level series , which facilitates the analysis. Nevertheless, any other assignment that satisfies the profile requirement and energy constraint is valid and may lead to better performance; for further discussion, see Section 7. 5.2. Infinite-Blocklength Setting with Analog PPM Layers
In this section, we concentrate on the setting of an infinite-source blocklength () and a quadratically decaying profile ( in (6a,b)) using analog PPM.
To that end, we use a sequence of
linear JSCC layers as in
Section 5.1, with only the last layer replaced by an analog PPM one; since analog PPM improves quadratically with the ENR (recall Remark 3),
M need not go to infinity to attain a quadratically decaying profile.
Theorem 4. Choose a design parameter , and a minimal noise level , however small. Then, a quadratic profile () (6a) with is achievable for all noise levels for any transmit energy E that satisfiesfor a large-enough-source blocklength k. This theorem, whose proof is available in
Appendix D, offers a further improvement over the upper bounds in (
7) and Theorem 3 for a quadratic profile. Again, the proof of Theorem 4 in
Appendix D is constructive and details the scheme parameters, such as the energy allocated to each layer and the minimal noise level, it is designed for.
Remark 8. Replacing all layers but the first layer with analog PPM ones should yield better performance, but complicates the analysis. Moreover, a similar analysis to that of Theorem 3 for may be devised, but for would require multiple layers as the distortion of analog PPM decays only quadratically. Both of these analyses are left for future research.