DRAFT 1

An Information-Theoretic Framework for Receiver

Quantization in Communication
Jing Zhou, Member, IEEE, Shuqin Pang, and Wenyi Zhang, Senior Member, IEEE

Abstract—We investigate information-theoretic limits and de- four times as the resolution increases by one bit, while every
sign of communication under receiver quantization. Unlike most doubling of the sampling rate leads to a one-bit loss of
existing studies that focus on low-resolution quantization, this resolution [2], [3]. The impact of ADC on the performance has
work is more focused on the impact of weak nonlinear distortion
due to resolution reduction from high to low. We consider a received increasing attention along with the recent evolution
standard transceiver architecture, which includes independent of wireless communications, in which several challenges are
and identically distributed (i.i.d.) complex Gaussian codebook faced, such as the increasing processing speed due to the uti-
at the transmitter, and a symmetric quantizer cascaded with a lization of larger bandwidth in mmWave and higher frequen-
nearest neighbor decoder at the receiver. Employing the general- cies, the increasing scale of hardware due to the use of massive
ized mutual information (GMI), an achievable rate under general
quantization rules is obtained in an analytical form, which multiple-input-multiple-output (MIMO), and the critical need
shows that the rate loss due to quantization is log (1 + γSNR), for low cost energy-efficient devices in emerging scenarios,
where SNR is the signal-to-noise ratio at the receiver front-end, e.g., massive machine-type communications (mMTC).
and γ is determined by thresholds and levels of the quantizer. A majority of the studies on ADC at communication re-
Based on this result, the performance under uniform receiver ceivers have focused on the performance and design under
quantization is analyzed comprehensively. We show that the
front-end gain control, which determines the loading factor low-resolution output quantization, and one-bit quantization
(normalized one-sided quantization range) of quantization, has has been of particular interest due to its negligible power
an increasing impact on performance as the resolution decreases. dissipation and simplicity of implementation, even without
In particular, we prove that the unique loading factor that requiring automatic gain control (AGC). In such studies the
minimizes the mean square error (MSE) of the uniform quantizer end-to-end channel is highly nonlinear, typically incurring a
also maximizes the GMI, and the corresponding irreducible
rate loss is given by log (1 + mmse · SNR), where mmse is the substantial performance loss, and necessitating a rethinking of
minimum MSE normalized by the variance of quantizer input, the transceiver design.
and is equal to the minimum of γ. A geometrical interpretation On the other side, the transceiver architecture used in
for the optimal uniform quantization at the receiver is further present wireless systems is built without considering the effect
established. Moreover, by asymptotic analysis, we characterize
of output quantization. It is thus necessary to ask, under
the impact of biased gain control, including how small rate
losses decay to zero and achievable rate approximations under such conventional transceiver architecture, how much is the
large bias. From asymptotic expressions of the optimal loading loss caused by output quantization with moderate to high
factor and mmse, approximations and several “per-bit rules” for resolution? In other words, if a small loss in achievable rate
performance are also provided. Finally we discuss more types is acceptable, how fine need the quantization be? Analytical
of receiver quantization and show that the consistency between
results on these problems appear to be lacking. Moreover,
achievable rate maximization and MSE minimization does not
hold in general. limited resolution of quantization leads to new problems, e.g.,
sensitivity of performance to the error of gain control, residual
Index Terms—Achievable rate, analog-to-digital converter,
Gaussian channel, generalized mutual information, nearest neigh- interference in multiuser systems, and so on. These largely
bor decoding rule, mean square error, MMSE, transceiver design, unexplored problems prompt us to revisit the topic of receiver
uniform quantization. quantization in communication in this work.

I. I NTRODUCTION
A. Related Work on Quantization at Communication Receivers
HE analog-to-digital conversion (ADC), including sam-
T pling and quantization, is essential for any digital receiv-
er. The power dissipation of state-of-the-art ADCs increases
We begin from the impact of output quantization in the
(discrete-time) additive white Gaussian noise (AWGN) chan-
nel, which is a benchmark model in communication theory.
This work was supported in part by the National Natural Science Foundation The performance gain of using more output quantization
of China through Grant 62231022 and in part by Henan Key Laboratory
of Visible Light Communications through Grant HKLVLC2023-B03. The levels in coded transmission (i.e., soft-decision decoding)
material in this paper will be presented in part at the IEEE International was observed very early in [4] via an information theoretic
Symposium on Information Theory (ISIT), Ann Arbor, MI, USA, June 2025 approach. In the classic textbook of Wozencraft and Jacobs [5,
[1]. (Corresponding author: Wenyi Zhang.)
Jing Zhou is with the Department of Computer Science and Engineering, Chap. 6.2], cutoff rate analysis showed that, for equiprobable
Shaoxing University, Shaoxing 312000, China (e-mail: [email protected]). uniformly spaced pulse amplitude modulation (PAM) input,
Shuqin Pang and Wenyi Zhang are with the Department of Electronic the degradation due to output quantization is approximately 2
Engineering and Information Science, University of Science and Technology
of China, Hefei 230027, China (e-mail: [email protected], weny- dB when the alphabet size equals to the number of quantization
[email protected]). levels, and the degradation vanishes when the quantization
DRAFT 2

becomes increasingly fine. Particularly, in the low-signal-to- channel under Gaussian input and nearest neighbor decoding
noise-ratio (low-SNR) limit, hard-decision decoding (one-bit [30]; see [39] for more discussions. However, GMI analyses
quantization that observes the sign of output) leads to a power also show that the AQNM-based estimation is not accurate in
loss of π/2 (approximately 2 dB) [4], [5]; see also [6, Chap. general. For example, it may overestimate the achievable rate
2.11 and 3.4].1 For K-level output quantization, it was proved in multiantenna systems [40].
that a discrete input of at most K + 1 mass points suffices to
achieve the constrained capacity [8]. B. Related Work in Quantization Theory
In vector (MIMO) channels, the quantization loss in achiev-
able rate can be very small at low-to-moderate SNR even if The rich theory of quantization was surveyed comprehen-
1~3-bit output quantization is used. This fact was shown in sively in [41], in which two well established asymptotic
information theoretic studies [9]–[11] by numerical examples theories were emphasized. The first is Shannon’s information
for full-rank channels with multiplexing gains 2~4. Although theoretic approach (rate distortion theory [42]), which places
the constrained capacity is still unknown even in the one- quantization in the framework of lossy source coding and
bit case, recent information theoretic studies have provided focuses on the high-dimension regime, thereby shedding light
various results on MIMO systems with coarse output quanti- on vector quantization. The second is the asymptotic quanti-
zation [12]–[19], which show that proper transceiver design is zation theory, which sheds light on quantizer design in the
critical in realizing efficient communication in such systems. high-resolution regime. The asymptotic quantization theory
In particular, for one-bit output quantization, it was shown that is more relevant to receiver quantization in communication
the high-SNR capacity grows linearly with the rank of channel which typically does not employ coding or vector quantization.
[13], while the low-SNR asymptotic power loss of π/2 still Although not applicable directly to receiver quantization, some
exists in the case of vector channel [12]. classical results in asymptotic quantization theory, especially
In light of these positive theoretical results, performance and those for uniform scalar quantization, are reviewed here for
design of wireless systems with low-resolution quantization comparison.
have been extensively studied in recent years; see, e.g., [20]– A basic result known in [43] and [44] (rigorously proved in
[26]. The most common approach therein, however, is not [45]) states that, for a high-resolution uniform quantizer with
information-theoretic (since exact evaluation of mutual infor- step size ℓ, the mean square error (MSE) can be approximated
mation can be difficult). Instead, achievable rate estimation by ℓ2 /12. This yields the “6-dB-per-bit rule” that each addi-
based on the additive quantization noise model (AQNM) has tional bit in resolution reduces the MSE by 6.02 dB. The rule
been widely used, which comes from Bussgang-like decom- reflects the impact of step size that causes granular distortion;
position [27]–[29]. Results in [12]–[14], [20] suggested that but it ignores overload distortion due to finite quantization
such estimation approximates the mutual information well at range. The interplay between these two types of distortion
low SNR, but becomes inaccurate at high SNR. determines how the optimal quantization range scales with
Although mutual information is a fundamental performance the resolution. The scaling law has been characterized in
measure, for communications under transceiver nonlinearity the seminal work [46] for several types of input densities.
it has limited operational meaning, in the sense that the Take the Gaussian source as example. In [46] it has been
decoder that achieves the predicted rate can be too complex shown that, for the optimal 2K-level uniform quantization
2
to implement, while that rate is not necessarily achievable that
p minimizes the MSE, 1) the loading factor scales like
by a standard transceiver architecture designed without con- 2 ln(2K) (cf. the conventional “four-sigma” rule of thumb
sidering nonlinearity because the decoder is typically mis- [44], [47], [48]), and 2) the granular distortion dominates and
matched to the nonlinear channel. In [30], [31], a more the overload distortion is asymptotically negligible. Further
meaningful performance measure that takes decoding rule into properties of the uniform quantization have been analyzed
account, namely the generalized mutual information (GMI) in [49]–[51]. For quantization at communication receivers,
[32], has been adopted, yielding analytical expressions of we need parallel results to characterize the optimal loading
the achievable rate under output quantization and nearest factor, which is essential for the design of AGC. We note
neighbor decoding rule. Under a given (possibly mismatched) that, although the importance of AGC design in the presence
decoder, the GMI determines the highest rate below which of output quantization has been recognized for a long time [6],
the average probability of error, averaged over a given i.i.d. [29], it was only investigated by numerical results in several
codebook ensemble, converges to zero as the block length works; see, e.g., [9], [10], [22], [52], [53].
N grows without bound, and it is thus a lower bound on Similar to the AQNM, there is also an additive noise model
mismatch capacity [32], [33]. The GMI has been applied for source quantization [41], [44], [47], which approximates
in various scenarios for performance evaluation, including the quantization error as an independent white noise term
the bit-interleaved coded modulation (BICM) [34], fading added to the quantizer input (but does not include a scaling
channels [35]–[37], and nonlinear fiber-optic channels [38]. In factor like that in the AQNM), though the “noise” is in fact a
fact, the rate estimation based on the AQNM (or Bussgang- deterministic function of the input. The quantization error can
like decomposition) is consistent with the GMI for scalar be white (uncorrelated between samples) when the source is
i.i.d., and it is approximately uncorrelated with the input when
1 Interestingly, the loss can be fully recovered if we replace the hard-decision
decoder (a.k.a. sign quantizer) by a carefully designed asymmetric one-bit 2 The loading factor of a quantizer is the one-sided width of its quantization
quantizer (a.k.a. threshold quantizer) and employ asymmetric input [7]. range normalized by the standard deviation of the input [47].
DRAFT 3

the resolution is sufficiently high. So the model may give a decoding, the GMI has a simple expression which can be
useful approximation under certain circumstances (see, e.g., evaluated by the correlation between the channel input and
[54]). output. Our asymptotic analyses also rely on methods and
results in asymptotic (high-resolution) quantization theory.
C. Our Work A notable tool originating from numerical analysis is the
Euler-Maclaurin summation formula [60], which was initially
1) Problem: In this paper, we consider communications introduced to source quantization theory in [49].
in the presence of output quantization, and focus on the 3) Summary of Contribution: We provide information-
effect of resolution reduction on the performance of a s- theoretic results for the transceiver architecture including i.i.d.
tandard transceiver architecture designed without considering complex Gaussian codebook and nearest neighbor decoding,
that effect. Rather than considering only low-resolution output in the presence of complex Gaussian noise and symmetric
quantization (1~3 bits) as in most existing studies, we con- receiver quantization, especially uniform quantization. Our
sider the entire region of resolution, especially the transition main contributions, including exact expressions, asymptotic
from high resolution (typically 8~12 bits or more so that formulas, and numerical results, are summarized as follows.
performance loss is negligible) to low resolution. Since in the
• In Sec. III, for the considered transceiver architecture, we
transition only weak to moderate nonlinearity is introduced,
show that the GMI under a given SNR can be expressed
it is natural to keep the transceiver architecture unchanged
by
rather than rebuild it. The considered standard transceiver
architecture includes independent and identically distributed IGMI = C − log(1 + γSNR), (1)
(i.i.d.) complex Gaussian codebook at the transmitter, and a
uniform output quantizer cascaded with a (weighted) nearest where C is the channel capacity when the resolution of
neighbor decoder at the receiver, where the loading factor of quantization is unlimited, and the parameter γ, which
the quantizer can be adjusted by gain control. See Sec. II does not depend on the SNR, is determined by thresholds
for details. We choose such an architecture in view of the and levels of the quantizer in an analytical form.
following facts. • In Sec. IV, for uniform quantization (with equispaced
thresholds and mid-rise levels), we show that optimizing
• When the impact of quantization is negligible, this archi-
the loading factor by gain control before quantization
tecture is capacity-achieving in several important chan-
(thereby minimizing γ in (1)) is increasingly important
nel models, such as the AWGN channel and the flat-
as the resolution decreases, thus imposing a critical
fading channel with Gaussian noise and channel state
challenge to the AGC. Interestingly, the problems of
information at the receiver [55]. It is also a very robust
MSE minimization and achievable rate maximization are
architecture in general noisy channels [56]. This architec-
proven to be consistent, in the sense that there is a unique
ture has been adopted in various performance evaluation
loading factor L = L∗ satisfying
problems, e.g., [35], [36], [56] for linear channels with
fading and [30], [38], [57] for nonlinear channels. L∗ = arg max IGMI (L) = arg min mse(L), (2)
L L
• The nearest neighbor decoding (minimum Euclidean dis-
tance decoding) can be implemented efficiently and has where IGMI (L) is the achievable rate as a function of the
been widely employed as a standard decoding rule in loading factor. This fact, combined with existing result,
communication systems. The uniform quantizer is also implies
√ that the optimal loading factor L∗ scales like
a standard component of practical receivers as well as 2 b ln 2 as the resolution b (in bits) increases. We further
a common assumption in performance analysis [9]–[11], prove that the minimum of γ is exactly the minimum
[14], [15], [25], [53], [58]. mean square error (MMSE) normalized by the variance of
• The achievable rate of complex Gaussian input approx- quantizer input (denoted by mmse), so that the irreducible
imates that of regular high-order modulation schemes loss in achievable rate due to uniform quantization is
such as quadrature amplitude modulation (QAM). In determined by
the AWGN channel, the high-SNR gap between their
C − IGMI (L∗ ) = log (1 + mmse · SNR) . (3)
achievable rates is 1.53 dB, which can be further reduced
by constellation shaping [59]. For uniform quantization with the optimal loading factor,
2) Method: The achievable rate results in this paper are we establish a geometrical interpretation of how the
derived based on the GMI, which, as discussed in Sec. I- MMSE and the additive Gaussian noise jointly determine
A, is a convenient performance measure for the problem the achievable rate.
of information transmission under transceiver nonlinearity. • The uniform receiver quantization is further studied by
Moreover, as a performance measure under a mismatched asymptotic analysis. In Sec. V, the impact of biased gain
decoder (since the nearest neighbor decoding rule becomes control is characterized by asymptotic behaviors and ap-
suboptimal in the presence of receiver quantization), the GMI proximations of IGMI (L). Approximations of IGMI (L) for
possesses optimality in the sense that it is the maximum overload region (L < L∗ ) and underload region (L > L∗ )
achievable rate of the i.i.d. random code ensemble, thereby of the loading factor are proposed, respectively. Specifi-
indicating the performance of a “typical” codebook [32], cally, in the high-resolution regime, we characterize the
[35]. Specifically, for Gaussian codebook and nearest neighbor loading loss in achievable rate, showing that i) the loading
DRAFT 4

where n = 1, ..., N , q(·) denotes the quantizer which intro-

duces nonlinear distortion, g ∈ R+ is a gain-control factor,
and
VnR = Re(hXn (m) + Zn ), (5)
VnI = Im(hXn (m) + Zn ) (6)
are real part and imaginary part of the received signal, re-
spectively, where the noise Zn ∼ NC (0, σ 2 ) is independent of
Xn . In this model, we let the real and imaginary parts of the
received signal be quantized by the same rule with the same
gain-control factor.
Fig. 1. Transceiver architecture. Note (channel discretization): In practical systems, the
described transceiver architecture also includes Nyquist-type
pulse shaping at the transmitter and matched filtering com-
loss due to overload distortion decays exponentially as bined with symbol-rate sampling at the receiver. It has a
the loading factor increases, and ii) the loading loss discrete-time memoryless model at symbol-level as (4), if the
due to granular distortion decays quadratically as the channel does not introduce memory. Note that oversampling
step size decreases. In Sec. VI, focusing on the optimal (more accurately, sampling faster than the symbol rate) may
uniform quantization, we provide a new approximation improve achievable rate in the presence of output quantization,
of L∗ , an approximation of IGMI (L∗ ), and several per-bit especially in the case of one-bit quantization (see [61], [62]
rules for performance metrics such as saturation rates and and [30], [63]–[65]). However, such performance improve-
irreducible rate loss; ment relies on several conditions, including significant nonlin-
• In Sec. VII, for general quantization rules, we illustrate earity (leading to frequency dispersion that can be utilized by
that the consistency between achievable rate maximiza- oversampling), non-standard transceiver architecture (like new
tion and MSE minimization does not necessarily hold. waveform design at the transmitter [64], [65]), and sufficiently
We also discuss the possible gain of introducing more high SNR. Receivers based on oversampling are beyond the
types of quantization for communication receiver. scope of this paper since our focus is the effect of low-to-
Notation: We write f (k) = o(g(k)) to denote the asymptotic moderate nonlinearity on a standard transceiver architecture.
relationship limk→∞ fg(k)
(k)
= 0. Specifically, we use ok (1) to 1) Nearest Neighbor Decoding Rule: The decoder selects a
denote a function f (k) satisfying limk→∞ f (k) = 0, and message according to the (scaled) nearest neighbor decoding
may omit the subscript if there is no danger of confusion. rule [35], [56] as
We use φ(t) to denote the function (2π)−1/2 exp(−t2 /2), N
X
which is the probability density function (PDF) of the standard 2
m̂ = arg min |Yn − aXn (m)| . (7)
m∈M
R ∞ and use Q(u) to denote the Q-function,
normal distribution, n=1
i.e., Q(u) := u φ(t)dt. We further define φ(∞) = 0 and That is, it selects a message corresponding to the codeword (s-
Q(∞) = 0. The complex conjugate of A is denoted by A. caled by a parameter a) with the minimum Euclidean distance
The Euclidean norm of A is denoted by kAk. We use X ⊥ Y to the received vector [Y1 , ..., YN ].3
to indicate the independence of two random variables X and 2) Symmetric Quantizer and Uniform Quantizer: Let
Y. the quantizer in (4) be symmetric with 2K representation
points (levels) {±y1 , ..., ±yK } and normalized thresholds
II. P RELIMINARIES {0, ±l1, ..., ±lK−1 }, where yk ≥ 0 and lk ≥ 0 for k =
1, ..., K. Then its resolution (bit-width) is b = log2 2K bits,
A. A Standard Transceiver Architecture which typically satisfies b ∈ Z+ . Let V ∈ R be the input to
The transceiver architecture we consider is shown in Fig. p deviation σv . For both quantizers
be quantized with standard
1. For a code rate R bits/channel use (c.u.), a message in (4) we have σv = (|h|2 σx2 + σ 2 )/2. Then for the input
is selected uniformly randomly from the index set M = V , the output of the quantizer is
{1, 2, ..., ⌈2N R⌉}. If a message m is selected, then the en- q(gV ) = yk · sgn(V ), if lk−1 σv ≤ g|V | < lk σv , (8)
coder maps it to a codeword [X1 (m), ..., XN (m)] of block
length N , which is generated according to a product complex where the thresholds satisfy l0 = 0 < l1 < ... < lK−1 <
Gaussian distribution NC (0, σx2 IN ). During transmission, each lK = ∞. Apparently, the quantizer output is a nonlinear
transmitted symbol is scaled by a channel gain h ∈ C which function of its input, and the thus introduced nonlinearity
remains constant over the transmission duration of a codeword. degrades performance. In the presence of gain control, we
The scaled symbols are corrupted by i.i.d. complex Gaussian may turn our attention to an equivalent quantization rule for
noise at the receiver front-end before quantization. Then the a normalized input V /σv with adjustable thresholds {ℓk =
channel output after quantization is given by 3 Some more general forms of nearest neighbor decoding rule can be found


in, e.g., [35], [36], [56]. A more recent work including a detailed review and
Yn = q g · VnR + j · q g · VnI , (4) some generalizations is [39].
DRAFT 5

lk /g, k = 1, ..., K − 1}, where g can be adjusted to optimize To achieve the maximum GMI given in (14), the scaling factor
the performance. A special case is the uniform quantizer, in (7) should be set as
which has equispaced thresholds  
E XY
ℓk = kℓ, k = 1, ..., K − 1, (9) a = α := . (16)
σx2
and mid-rise levels Besides the proof in [30] based on direct evaluation and
 
1 optimization of the dual expression of the GMI, here we
yk = k − ℓ, k = 1, ..., K, (10)
2 provide a sketch of an alternative proof.
Proof Sketch: It has been noted in [35] and [66] that, for an
where ℓ is the step size. Thus, we define its quantization range
or support as [−Kℓ, Kℓ]. Then the loading factor or support   noise channel Y = S +U (i.e., the noise
additive uncorrelated
U satisfies E SU = 0, but is not necessarily independent of
limit of the uniform quantizer is L = Kℓ.
the input S), if S ∼ NC (0, E |S|2 ), then
 !
B. An Achievable Rate Formula from Generalized Mutual E |S|2
Information I(X; Y ) ≥ log 1 + . (17)
E [|U |2 ]
Following the notation of [33], consider a memoryless
channel X → Y with general alphabets X and Y, input prob- In [56], under Gaussian codebook and nearest neighbor de-
ability distribution PX (x), transition probability PY |X (y|x), coding rule, the achievability of the RHS of (17) and a
and decoding metric d(X, Y ). The decoder selects a message random coding converse for it are established by a geometric
according to argument,4 where the Gaussian codebook can either be an i.i.d.
Gaussian codebook or an equienergy one. In the former case
N
X the rate (17) is the GMI. For the scalar channel X → Y ,
m̂ = arg min d(Xn (m), Yn ). (11)
m∈M X ∼ NC (0, σx2 ), by a Bussgang-like decomposition, we can
n=1
always write Y = αX + D, where   α is given in (16),
As a lower bound on mismatch capacity, the GMI can be given and D = Y − αX satisfies E XD = 0. That is, we
by its dual expression as [32], [33] eliminate the correlation between the scaled input αX and
" #
e−sd(X,Y ) the corresponding distortion D by a carefully chosen scaling
IGMI = sup E log  −sd(X ′ ,Y )  , (12) factor. It is straightforward to show that (14) can be obtained
s≥0 E e |Y
by (17) when S = αX and U = D. 
where (X, Y, X ′ ) ∼ PX (x)PY |X (y|x)PX (x′ ). The GMI gives Let σy be the standard deviation of Y . We note that when
the maximum rate below which the probability of decoding E[Y ] = 0 (e.g., when the quantizer is symmetric, the channel
error, averaged over the i.i.d. random codebook emsemble, output in (4) has zero mean) we have ∆ = |ρXY |2 , where
converges to zero as the coding block length grows without
cov(X, Y )
bound. For communications with transceiver nonlinearity, if ρXY = (18)
the input distribution PX (x) is complex Gaussian with vari- σx σy
ance σx2 and the decoding rule (11) is specified by (7), then is the Pearson correlation coefficient between X and Y . Also
from (12) we obtain [30] note that the equality in (17) holds if and only if X ⊥ D and
 D ∼ NC (0, σ 2 ) (e.g., when there is no quantization in (4)),
 
IGMI = sup log(1 + s|a|2 σx2 ) − sE |Y − aX|2 so that the channel X → Y reduces to the AWGN channel,
s≥0
   the nearest neighbor decoding rule is optimal, and the GMI
sE |Y |2 equals
+ . (13)  to the  mutual information (also the channel capacity)
1 + s|a|2 σx2 σ2
log 1 + σx2 .
Maximizing the GMI by optimizing the scaling factor a yields
the following result [30, Appendix C], which provides a
III. ACHIEVABLE R ATE U NDER R ECEIVER
general approach for achievable rate analysis in the presence
Q UANTIZATION : E XACT AND A SYMPTOTIC R ESULTS
of transceiver nonlinearity with known transition probability.
Proposition 1 [30]: For a memoryless SISO channel X → Based on Proposition 1, we establish the following result
Y with transition probability pY |X (y|x) and nearest neighbor which provides an analytical expression for the achievable rate
decoding rule (7), where X, Y ∈ C and Var(X) = σx2 , the of the transceiver architecture considered in this paper.
maximum GMI under i.i.d. complex Gaussian codebook is Theorem 2: For the channel (4) where q(·) is the sym-
given by metric quantizer described in Sec. II-A2, the achievable rate
1 under i.i.d. complex Gaussian codebook and nearest neighbor
IGMI = log , (14) decoding rule (7) is
1−∆
where IGMI = log(1 + SNR) − log(1 + γSNR), (19)
 2
E XY
∆= 2 . (15) 4 The proof in [56] was intended for independent noise. However, as noted
σx E [|Y |2 ] in [35], it goes through verbatim for uncorrelated noise.
DRAFT 6

where SNR = |h|2 σx2 /σ 2 is the SNR at the receiver front-end, quantization; see, e.g., [31], which showed that nonsubtrac-
and γ is a parameter determined by the quantizer as tive Gaussian dithering improves GMI under certain circum-
stances.5 However, in our setting, nonsubtractive Gaussian
A2 dithering is always harmful, because it is equivalent to re-
γ =1− , (20)
B ducing the SNR, while the GMI is a monotonically increasing
in which function of the SNR since its derivative (in nats) satisfies
dIGMI 1−γ
√ X K = > 0. (27)
A = 2π yk (φ(ℓk−1 ) − φ(ℓk )) , (21) dSNR (1 + SNR)(1 + γSNR)
k=1 From Theorem 2, we immediately obtain the following
corollary, which shows that γ dominates the asymptotic
and
behavior of performance, especially the low-SNR slope of
K
X achievable rate, and the high-SNR saturation rate.
B=π yk2 (Q(ℓk−1 ) − Q(ℓk )) . (22) Corollary 3: The achievable rate IGMI given in (19) has the
k=1 following properties.
• The parameter γ satisfies
Proof: Applying Proposition 1, it is sufficient to show
that 2
0<γ ≤1− , (28)
π
SNR A2
∆= , (23) where the equality holds when q(·) is a one-bit quantizer,
1 + SNR B
corresponding to an achievable rate
which can be obtained by showing that 1-bit 1 + SNR
IGMI = log , (29)
  1 + π−2
E XY = 2E [Re(X) · Re(Y ) + j · Im(X) · Re(Y )] (24a) π SNR
π
2Re(h)σx 2 2
2j · Im(h)σx which converges to log2 π−2 = 1.4604 bits/c.u. at high
=p A− p A (24b) SNR (see also footnote 6). As the quantization becomes
π(|h|2 σx2 + σ 2 ) π(|h|2 σx2 + σ 2 ) increasingly fine, we have γ → 0 (from above), SNRe →
2hσx2 SNR (from below), and
=p A (24c)
π(|h|2 σx2 + σ 2 ) SNR2 2
C − IGMI = SNR · γ − γ + o(γ 2 ) nats/c.u. (30)
and 2
    4 • High- and low-SNR asymptotics: As SNR → ∞, we have
E |Y |2 = 2E Re(Y )2 = B. (25) SNRe → 1−γγ and
π  
1 1−γ 1 1
The identities (24c) and (25) can be obtained by lengthy but IGMI = log − +o . (31)
straightforward evaluations of expectations (cf. [30, Appendix γ γ SNR SNR
D]). So the saturation rate is given by
Remark: In [30, Sec. V], a parallel result of Theorem 2 for 1
real-valued channel with symmetric output quantization was I¯GMI = log . (32)
γ
given by deriving its effective SNR, while Theorem 2 shows
As SNR → 0, we have
that the GMI in the complex-valued case has exactly the same
expression except that the pre-log factor is doubled. 1 − γ2
IGMI = (1 − γ)SNR − SNR2
We note that, if the impact of quantization is omitted, 2

then the model (4) reduces to a linear Gaussian channel + o SNR2 nats/c.u. (33)
Yn = hXn (m)+Zn , and the achievable rate of the considered
An alternative expression of (19) is given in terms of
transceiver architecture is log(1 + SNR), which is also the
effective SNR (or signal to noise-and-distortion ratio) as
channel capacity C when the power of the input is σx2 . The
expression of IGMI in Theorem 2 shows explicitly that the rate IGMI = log (1 + SNRe ) , (34)
loss due to quantization is
where
C − IGMI = log (1 + γSNR) , (26) (1 − γ)|h|2 σx2 (1 − γ)
SNRe = 2 2 2
= SNR. (35)
γ|h| σx + σ γSNR + 1
which has the same expression as C except for an SNR
From this expression, we see that the effect of quantization is
reduction of at least 4.4 dB (see (28) given later). For a given
to transfer a fraction γ of the power to the denominator of the
SNR, the reduction is determined by the parameter γ, which
reflects the impact of nonlinearity and does not depend on the 5 In nonsubtractive dithering, one adds a dither signal W independent of
d
SNR. the quantizer input V before quantization, yielding an output q(V + Wd ).
Subtractive dithering has an additional step that subtracts the dither signal
Remark: Dithering, an intentional randomization technique after quantization and finally obtains q(V + Wd ) − Wd . See [41] for more
in source quantization [41], may also be beneficial in receiver discussion.
DRAFT 7

effective SNR. At low SNR, the power loss can be evaluated From (40), the normalized MSE is minimized only if
by SNR/SNRe , which asymptotically converges to 1/(1−γ) ≤ √ dA dB
π/2, where equality holds when one-bit quantization is used.6 2π = . (43)
dg dg
From Theorem 2, it is simple to check that, for one-bit
quantization, the gain control factor g has no impact on The different expressions of mse and γ show that, in gen-
the achievable rate. In general, the achievable rate has the eral, the gain control factor that minimizes the MSE does
following property. not necessarily maximizes the achievable rate. However, for
Proposition 4: The achievable rate IGMI given in (19), as a uniform quantization, the next section will show that the two
function of the gain control factor g, satisfies optimization problems are consistent; i.e., there is a unique
gain control factor (and correspondingly, a unique loading
1-bit
lim IGMI (g) = lim IGMI (g) = IGMI . (36) factor) that solves both problems simultaneously.
g→0 g→∞

Proof: See Appendix A. IV. O PTIMAL U NIFORM Q UANTIZATION AT THE R ECEIVER

This result can be interpreted intuitively as follows. Note
A. Achievable Rate Evaluation under Uniform Quantization
that as g → ∞ we have

Proposition 6: If q(·) is the uniform quantizer described in
Pr min |VnR |, |VnI | ≥ ℓK−1 σv → 1, (37) Sec. II-A2, then
and as g → 0 we have √ K−1
X ℓ

A= 2π ℓ · φ(kℓ) − , (44)
Pr max |VnR |, |VnI | < ℓ1 σv → 1. (38) 2
k=0

Thus, in both limits, the effective resolution of the quantizer and

1-bit K−1
reduces to one, so that the achievable rate converges to IGMI . X 1
Remark: We note that IGMI 1-bit
is not the infimum of IGMI (g) B=π 2kℓ2 Q(kℓ) + πℓ2 . (45)
8
in general. An example will be given in Sec. IV-D, Fig. 17(f). k=0

Intuitively, a smaller MSE is preferred in quantization. How- Proof: See Appendix C.

ever, for communications the main performance metric is the Specifically, for a uniform quantizer with a given resolution,
achievable rate. Thus the relationship of MSE minimization (44) and (45) show that γ is determined solely by the step
and achievable rate maximization is of interest. size ℓ or equivalently by the loading factor L, which can
Due to symmetry, we may define the normalized MSE of be optimized by adjusting the gain-control factor g in (4)
the quantizer q(·) in the channel (4) in terms of the in-phase according to the channel gain h.
component of Yn as Remark: The quantities A and B are important in our
" 2 # analysis.
VR PK−1 For uniform−k2 ℓ2
quantization, we note that the summation
k=0 ℓ · exp in A is exactly the left Riemann sum of
R
mse := E q(g · V ) − , (39) 2
σv −t2
exp 2 over [0, Kℓ] with a regular partition, and similarly,
PK−1
where we omit the index of V R and Y since we assume i.i.d. the summation π k=0 2kℓ2 Q(kℓ) in B is exactly the left
input, and we normalize the input since we use normalized Riemann sum of 2πtQ(t) over [0, Kℓ] with a regular partition.
levels of quantization. The following result shows that the Therefore, for a fixed loading factor L we have the following
normalized MSE can be expressed by A and B, and is lower high-resolution limits:
!
bounded by γ. √ X L  kL 
K−1
L
Proposition 5: lim A = 2π lim φ − (46a)
K→∞ K→∞ K K 2K
2 √  k=0
Z L
mse = 1 − 2πA − B ≥ γ, (40) −t2
π = exp dt (46b)
0 2
where equality holds if and only if  
√ 1
r = 2π − Q(L) , (46c)
A 2 2
= . (41)
B π
  !
. K−1
X L kL kL L2
Proof: See Appendix B. lim B = 2π lim Q + (47a)
K→∞ K→∞ K K K 16K 2
We define the normalized MMSE as k=0
Z L
mmse := min mse. (42) = 2π tQ(t)dt (47b)
g 0
Z ∞
π
6 The channel capacity under one-bit receiver quantization, given by 1 − = − 2π tQ(t)dt. (47c)
H2 (Q(SNR)) [6], [8], has the same low-SNR asymptotic behavior. Note that 2 L
the GMI is derived under Gaussian input and nearest neighbor decoder, and the In Fig. 2 and Fig. 3 we show numerical evaluations of IGMI
capacity is achieved by antipodal signaling (the decoding rule is unrestricted).
The difference becomes evident at high-SNR, where the capacity converges in Theorem 2, which has been maximized over L > 0 for each
∗
to 2 bits/c.u., while the GMI converges to 1.4604 bits/c.u. resolution and is then denoted by IGMI . The corresponding
DRAFT 8

14
Clearly, decreasing L reduces the granular distortion but
increases the overload distortion. When L < L∗ the overload
12 distortion increases quickly, thereby causing the waterfall of
achievable rate.
10 The numerical results in Fig. 4 and Fig. 5 reveal the
increasing importance of gain control (realized by an AGC
8 module in practical systems) as the resolution decreases: 1)
Under high-resolution output quantization, we only require
6
a rough estimate of the channel gain to guarantee that the
loading factor to be no less than a predefined threshold, say 4
4
(from the four-sigma rule of thumb [44], [47]), so that we can
stay away from the waterfall; 2) Under low-resolution output
2
quantization, such a simple strategy may increase rate loss
0
considerably, but perfect gain control needs accurate channel
-10 0 10 20 30 40 estimation which is also challenging in this case.

Fig. 2. Achievable rate with uniform output quantization and optimal loading B. Consistency Between Rate Maximization and MSE Mini-
factor. mization and a GMI-MMSE Formula
14 From source quantization theory [46], [50] we know that,
for Gaussian input, the MSE of the uniform quantizer is a
12
strictly convex function of the step size (or loading factor) and
the minimum is located at the unique solution of dMSE dℓ = 0,
10
denoted by ℓ∗ . In fact, that solution is also the unique step
8
size that maximize the GMI; that is, for uniform quantization
at the receiver there is a consistency between achievable rate
6 maximization and MSE minimization. This consistency further
leads to a GMI-MMSE formula which connects the maximum
4 achievable rate and the MMSE of quantization in a simple and
closed form. We first give the following lemma, and then use
2 it to establish the aforementioned findings in Theorem 8.
Lemma 7: For uniform quantization at the receiver, for
0
2 4 6 8 10 12 b > 1 (i.e., except for one-bit quantization), we have
dIGMI dmse
=0⇔ = 0, (51)
Fig. 3. Convergence of achievable rate to channel capacity. dℓ dℓ
and both of them are equivalent to
r
A 2
unique loading factor is denoted by L∗ . Its uniqueness will be = . (52)
B π
proved later. In Fig. 4 and Fig. 5, we show how the achievable
rate IGMI varies with the loading factor L (some more numeri- Proof: See Appendix D.
cal results and details therein will be interpreted in subsequent Theorem 8: For the channel (4) where q(·) is the uniform
sections). As the resolution increases, the increasing of L∗ is quantizer described in Sec. II-A2 with resolution b > 1,
clear. A “waterfall” near L = 0 can be observed in all figures, the input employs i.i.d. complex Gaussian codebook, and the
implying that an underestimate of the optimal loading factor receiver employs nearest neighbor decoding rule (7), we have
always causes serious rate loss. To interpret this phenomenon, the following properties.
∗
we write the normalized MSE of uniform quantization as the • The loading factor L that maximizes the GMI (19) is
sum of overload distortion and granular distortion as unique, and it is also the unique loading factor that
minimizes the MSE (39), i.e.,
mse = mseo + mseg , (48)
arg max IGMI (L) = arg min mse(L). (53)
L L
where
Z ∞ • The minimum of γ as a function of L, namely γ(L∗ ), is
2
mseo := (t − L + ℓ/2) φ(t)dt, (49) exactly the normalized MMSE (42) and satisfies
L
2
and 0 < mmse = γ(L∗ ) ≤ γ ≤ 1 − , (54)
π
K Z
X kℓ    2
1 and the maximum GMI can be written as
mseg := t− k− ℓ φ(t)dt. (50)
(k−1)ℓ 2 ∗
IGMI = log(1 + SNR) − log (1 + mmse · SNR) . (55)
k=1
DRAFT 9

5.0143 4.4982 3.9376

14 14 14

12 12 12

10 10 10

8 8 8

6 6 6

4 4 4

2 2 2

0 0 0
0 4 8 12 16 20 24 28 32 0 4 8 12 16 20 24 28 32 0 4 8 12 16 20 24 28 32

(a) b = 12 bits. Simple gain control strategy (b) b = 10 bits. Simple gain control strategy (c) b = 8 bits. Four-sigma rule of thumb is conve-
is always enough. is usually enough. nient.

3.3300 2.6816 1.9914

14 14 14

12 12 12

10 10 10

8 8 8

6 6 6

4 4 4

2 2 2

0 0 0
0 4 8 12 16 20 24 28 32 0 4 8 12 16 20 24 28 32 0 4 8 12 16 20 24 28 32

(d) b = 6 bits. Four-sigma rule of thumb (e) b = 4 bits. Four-sigma rule of thumb (f) b = 2 bits. Four-sigma rule of thumb causes
causes small loss at high SNR. causes considerable loss. significant loss.

Fig. 4. Impact of loading factor on achievable rate: Fixed resolution, varying SNR. Vertical lines and corresponding values show the optimal loading factors.

0.14 1 3.5

0.9
0.12 3
0.8

0.1 0.7 2.5

0.6
0.08 2
0.5
0.06 1.5
0.4

0.04 0.3 1

0.2
0.02 0.5
0.1

0 0 0
0 4 8 12 16 20 24 28 32 0 4 8 12 16 20 24 28 32 0 4 8 12 16 20 24 28 32

(a) SNR = −10dB. (b) SNR = 0dB. (c) SNR = 10dB ≈ SNRq (2).

7 10 14

9
6 12
8

5 7 10

6
4 8
5
3 6
4

2 3 4

2
1 2
1

0 0 0
0 4 8 12 16 20 24 28 32 0 4 8 12 16 20 24 28 32 0 4 8 12 16 20 24 28 32

(d) SNR = 20dB ≈ SNRq (4). (e) SNR = 30dB ≈ SNRq (6). (f) SNR = 40dB ≈ SNRq (8).

Fig. 5. Impact of loading factor on achievable rate: Fixed SNR, varying resolution.
DRAFT 10

Proof: First, in [46, Sec. V-A], it has been proved that Theorem 8 enables us to utilize known results on uniform
the MSE is minimized if and only if the loading factor is set quantization for minimizing the MSE. The following result
to be the unique solution of dmse ∗
dL = 0, denoted by L . The follows immediately from the equivalence (53) and existing
∗
uniqueness of L is confirmed by showing that the second results in [46], [67]. In particular, it is shown that the√optimal
derivative of the MSE is strictly positive. Lemma 7 implies loading factor L∗ grows with the resolution b like 2 b ln 2.
that the loading factor that satisfies dIdL
GMI
= dI
Kdℓ = 0 is also
GMI
Corollary 10: In the channel (4) under i.i.d. complex
unique. Noting that the achievable rate IGMI is a continuous Gaussian codebook and nearest neighbor decoding rule (7),
and differentiable function of L, from Proposition 4 we can the optimal step size ℓ∗ that maximizes the GMI (19) satisfies
infer that L∗ also maximizes IGMI (L) (it is not difficult to
lim Kℓ∗ = ∞, lim ℓ∗ = 0. (61)
exclude the other possible case that L∗ minimizes IGMI (L)). K→∞ K→∞
Then the first part of Theorem 8 is proved. The optimal loading factor L∗ = Kℓ∗ grows monotonically
According to Lemma 7, if L = L∗ then (52) holds. with K and satisfies
Combining this fact with (40) and (20), we have
L∗
π lim p = 1. (62)
γ(L∗ ) = 1 − B(ℓ∗ ) = mmse, (56) K→∞ 2 ln(2K)
2
thereby completing the proof of the second part of Theorem The corresponding MMSE consisting of the granular distortion
8. mmseg and the overload distortion mmseo satisfies
From the proof of Theorem 8, we can infer that IGMI (L) > mmse mmseg
1-bit
IGMI holds for 0 < L < ∞. Combining this with Proposition lim ∗ 2 = lim ∗ 2 = 1; (63)
K→∞ ℓ /12 K→∞ ℓ /12
4 yields the following corollary.
Corollary 9: For uniform receiver quantization, the achiev- that is, the granular distortion dominates the MMSE as the
resolution increases:
able rate given in (19), as a function of the loading factor L,
mmseo
satisfies lim = 0. (64)
K→∞ mmseg
1-bit
inf IGMI (L) = lim IGMI (L) = lim IGMI (L) = IGMI , (57)
L L→0 L→∞ For (61) and (62), we also provide new proofs of them,
1-bit
where IGMI is given in (29). respectively, by asymptotic results on the achievable rate; see
In Theorem 8 we exclude the case b = 1. In fact, for Sec. VI-A.
symmetric one-bit quantization, we have A = ℓ/2, B = πℓ2 /8, Remark: In quantization theory, the “SNR” for a quantizer is
and γ ≡ 1 − 2/π for 0 < ℓ < ∞, so that gain control is often defined as the ratio between the variance of the quantizer
unnecessary. But when input and the MSE. For the optimal uniform quantizer we
denote
4
ℓ= √ = 1.5958, (58) 1
2π SNRq = . (65)
mmse
the normalized MSE achieves its minimum as mmse1-bit = Theorem 8 implies that the maximum saturation rate can be
1 − 2/π, corresponding to the upper bound in (54). expressed as
Remark: From Theorem 2, the rate loss due to uniform 1
quantization with a loading factor L is given by log(1 + I¯GMI
∗
= log = log SNRq , (66)
mmse
γ(L)SNR). For uniform quantization, Theorem 8 characterizes
the minimum of rate loss when L = L∗ . Thus we should and the corresponding effective SNR is given by
distinguish two parts of the total rate loss as follows. sup SNRe = lim SNRe = SNRq − 1. (67)
7 SNR>0 SNR→∞
• Irreducible loss: the unavoidable part for given resolu-
tion and SNR, given by For finite SNR we have
∗ 1 − mmse (1 − mmse)Ex
C − IGMI = log(1 + mmse · SNR). (59) SNRe = SNR = . (68)
mmse · SNR + 1 mmse · Ex + σ 2
Numerical results in Fig. 2 and Fig. 3 consider only
In Table I,8 we show numerical results of optimal uniform
irreducible loss.
quantization under our transceiver architecture (an approxima-
• Loading loss: the remaining part, given by
tion of mmse given by ln(2K)/3K 2 , and an approximation
∗ 1 + γ(L)SNR of I¯GMI , denoted by IˆGMI , will be introduced in Sec. VI-B).
IGMI − IGMI (L) = log , (60)
1 + mmse · SNR As the SNR decreases, numerical results in Fig. 5 show
which is due to suboptimal loading factor and can be that the supremum and infimum of IGMI have a decreasing
reduced by improving the accuracy of gain control. Nu- ratio which converges to π/2, coinciding with the “2 dB loss”
merical results in Fig. 4 and Fig. 5 show the importance result for hard-decision decoding. Thus, a major part of the
of reducing loading loss. low-SNR capacity can always be utilized. At moderate-to-high
SNR (Figs. (5(c)-5(f))), we may roughly separate different
7 We note that the loss (59) is irreducible in the sense of GMI, and it is not
necessarily irreducible in general since the GMI is only a lower bound on the 8 Some numerical results therein have been given in [46, Table II], in which
mismatch capacity of the channel (4). an error occurred in actual SNR computation for Gaussian source, b = 8.
DRAFT 11

TABLE I
N UMERICAL R ESULTS OF O PTIMAL PARAMETERS AND P ERFORMANCE M ETRICS FOR U NIFORM Q UANTIZATION

b 2K L∗ ℓ∗ γ(L∗ ) = mmse ln(2K)/3K 2 SNRq (dB) I¯GMI

∗ (bits/c.u.) Iˆ
¯GMI (bits/c.u.)
1 2 1.5958 1.5958 0.3634 0.2310 4.40 1.4604 2.11
2 4 1.9914 0.9957 0.1188 0.1155 9.25 3.0728 3.11
3 8 2.3441 0.5860 3.7440×10−2 4.3322×10−2 14.27 4.7393 4.53
4 16 2.6816 0.3352 1.1543×10−2 1.4441×10−2 19.38 6.4369 6.11
5 32 3.0102 0.1881 3.4952×10−3 4.5127×10−3 24.57 8.1604 7.79
6 64 3.3300 0.1041 1.0400×10−3 1.3538×10−3 29.83 9.9091 9.53
7 128 3.6395 5.6868×10−2 3.0433×10−4 3.9486×10−4 35.17 11.6821 11.31
8 256 3.9376 3.0762×10−2 8.7686×10−5 1.1282×10−4 40.57 13.4773 13.11
9 512 4.2237 1.6499×10−2 2.4919×10−5 3.1730×10−5 46.03 15.2924 14.94
10 1024 4.4982 8.7855×10−3 6.9970×10−6 8.8138×10−6 51.55 17.1248 16.79
11 2048 4.7614 4.6498×10−3 1.9444×10−6 2.4238×10−6 57.11 18.9722 18.65
12 4096 5.0143 2.4484×10−3 5.3554×10−7 6.6104×10−7 62.71 20.8325 20.53

scenarios of receiver quantization as follows, which show that interpretation for such optimal uniform quantization can be
channel estimation and AGC design become more challenging established. For simplification we first introduce an equivalent
as the resolution decreases. p shown in Fig. 6, where we let X = hX, Y = σv Y =
model
• SNR-limited (high resolution) scenario: b > 2 log10 SNR Y · (|h|2 σx2 + σ 2 )/2, V = V R +jV I , Z = Z, and W = Y−V
+b1 bits and SNRq ≫ SNR, where b1 can be 2~3. The (we write W(V) in Fig. 6 to emphasize that W is a function
irreducible loss (59) is negligible, and significant rate of V), so that
improvement can be obtained by increasing the SNR. In
Y = X + Z + W = V + W, (70)
fact, we have
∗
d(C − IGMI ) where the quantization error W satisfies
= mmse + o(mmse2 ), (69)    
dSNR E |W|2 = MMSE = mmse · E |V|2 = 2 · mmse · σv2 . (71)
which suggests that the irreducible loss increases slowly
with the SNR. A large overestimate of the optimal Thus, the high-resolution limit of Y is V, and the high-SNR
gain control factor does not cause significant loading limit of V is X.
loss, thereby allowing the simple gain control strategy Now we are ready to illustrate the geometry of optimal uni-
described at the end of Sec. IV-A. In this scenario the form quantization at the receiver in N -dimensional Euclidean
impact of receiver quantization is not important. There is space, as shown in Fig. 7. We use boldface letters to denote
still room to reduce the resolution if accurate gain control codewords or signal vectors in the equivalent model (70), e.g.,
is possible. X = [X1 , ..., XN ]. Since i.i.d. codebook is considered, the
• Resolution-limited (low resolution) scenario: b ≤ vector X and other vectors  in Fig.
 7 are all i.i.d. random
2 log10 SNR bits and SNRq < SNR. The achievable rate vectors. Thus, we have E kXk2 = N Ex , and the empirical
is seriously limited due to large irreducible loss, while average power of the input codeword converges in probability
accurate gain control is required to avoid large loading to Ex , i.e.,
 
loss. Significant rate improvement can be obtained by 1
lim kXk2 − Ex > ǫ = 0, ∀ǫ > 0. (72)
increasing the resolution. N →∞ N
• Moderate resolution scenario (the remaining cases): The
resolution is enough to maintain a small irreducible loss, √ that the length of X (in asymptotic√sense)
Wepthus briefly say
is E [kXk2 ] = N Ex . Similarly, the length of Z is N σ 2 .
but an overestimate of the optimal gain control factor may
The geometry in Fig. 7 includes two Pythagorean relations as
cause considerable loading loss.
follows.
We note that the consistency between rate maximization and
• For the additive noise channel V = X + Z, we have
MSE minimization does not hold in general; see discussions  
E XZ = 0, implying
on nonuniform quantization in Sec. VII.      
E kVk2 = E kXk2 + E kZk2 . (73)
C. Geometry of Optimal Uniform Quantization at Receiver • For the quantization channel Y = V + W, it has been
We have shown that, in the standard transceiver architecture shown in [67] that the MMSE uniform quantization
shown in Fig. 1, if the gain control factor g is set appropriately satisfies E YW = 0; i.e., the error W is uncorrelated
so that the loading factor of the quantizer q(·) is equal to the with the output of the quantizer, yielding
optimal value L∗ , then the achievable rate attains its maximum      
given by the GMI-MMSE formula (55). In fact, a geometrical E kVk2 = E kYk2 + E kWk2 . (74)
DRAFT 12

LMMSE estimator of Y from V is ρ2V Y = 1 − mmse. The

following result provides some equivalent expressions of the
condition (52) for GMI maximization, which also interprets
the Wiener coefficients of the other two projections shown in
Fig. 7.
Proposition 11: The following four conditions are equiva-
Fig. 6. A simplified equivalent model of the transceiver architecture when lent to each other:
L = L∗ . r
A 2
= , (76)
B π
   
E XY E XV
= , (77)
E [|Y|2 ] E [|V|2 ]
   
E YX E VY
= , (78)
E [|X|2 ] E [|V|2 ]
and
ρXY = ρXV · ρV Y , (79)
where
 
E XV SNR
= = ρ2XV , (80)
E [|V|2 ] 1 + SNR
and
 
E VY
= 1 − mmse = ρ2V Y , (81)
E [|V|2 ]
respectively.
Proof: See Appendix E.
Remark: Proposition 11 shows that the optimal uniform
quantization rule should let
Fig. 7. Geometry of optimal uniform quantization at the receiver. SNR
X̂LMMSE (Y) = Y (82)
1 + SNR
p
N (Ex + σ 2 ), hold; i.e., it should let the Wiener coefficient be equal to the
√ Thus, the lengths p of V, W, and Y are one in (80). Similarly, the optimal uniform quantization rule
2
N · MMSE, and N (Ex + σ − MMSE), respectively. We
then have two right triangles in Fig. 7: the quantization should let
triangle 0VY, which determines the saturation rate I¯GMI ∗
ŶLMMSE (X) = (1 − mmse)X (83)
in (66), and the noise triangle 0XV, which determines the
channel capacity C, i.e., the limit of achievable rate under fine hold, where the Wiener coefficient is equal to the one in (81).
quantization. These two triangles jointly determine the triangle Moreover, from (83) we obtain an additive uncorrelated noise
∗ model as
0XY, and consequently determine the achievable rate IGMI .
We note that YVX and 0XY are not right  triangles.
 In fact,
2 Y = (1 − mmse)X + D, (84)
for finite SNR and
 2finite
  2  E |Y − X| is strictly
resolution,
 
smaller than E |Z| + E |W| , because the error W is where D satisfies E XD = 0 and
correlated with the noise Z. Treating the error as independent  
noise always reduces the achievable rate. E |D|2 = (1 − mmse)(mmse · Ex + σ 2 ). (85)
In Fig. 7 there are also some perpendicular lines (dashed) The channel (84) is a scaling of the channel Y = αX + D
which indicate some projections. For example, the projection in the proof of Proposition 1. We can apply Proposition 1
of X onto V is the linear minimum MSE (LMMSE) estimator directly to (84) and obtain exactly the same GMI expression in
of X from V as Theorem 8. Correspondingly, the scaling factor for the nearest
 
E XV neighbor decoder in Fig. 7 should be set as 1 − mmse.
X̂LMMSE (V) = V, (75) We next discuss how the triangles in Fig. 7 determine the
E [|V|2 ]
performance via three angles as follows.
i.e., a scaling of V with a scalar (called Wiener coefficient),
  1) θXV ∈ (0, π/2): it is determined by the SNR and satisfies
2
which
 2is  equal to ρ2XV = SNR/(1 + SNR) since E XV = SNR = cot2 θXV (86)
E |X| = SNR · σ . Similarly, the Wiener coefficient of the
DRAFT 13

and V. A SYMPTOTIC A NALYSIS : C HARACTERIZATION OF

r IGMI (L)
SNR
ρXV = cos θXV = . (87) From previous numerical results, we have noted that ad-
1 + SNR
justing the loading factor by gain control is essential in
The high-resolution limit of achievable rate (i.e., the capacity)
approaching the maximum achievable rate. To understand
can be expressed as
the impact of biased gain control, this section investigates


C = log(1 + SNR) = log csc2 θXV . (88) the achievable rate IGMI (L) as a function of the loading
 q  factor L when the resolution and the SNR are given. Our
2) θV Y ∈ 0, arccos π2 : it is determined by the resolu- basic approach is asymptotic analysis, by which we establish
tion, or equivalently by mmse, and satisfies analytical results and interpret some phenomena observed in
numerical results. In source quantization, asymptotic analysis
mmse = sin2 θV Y (89) yields high-resolution quantization theory, which provides
fairly accurate results for resolutions equal to or greater than 3
and bits [41]. Our results for receiver quantization exhibit similar
√ accuracy.
ρV Y = cos θV Y = 1 − mmse. (90)
In summary, the behavior of IGMI (L) can be characterized
The high-SNR limit of achievable rate (i.e., the saturation rate) by two regions as follows.
can be expressed as 1) Overload Region (L < L∗ ): In this region overload
1
distortion dominates the performance, but the impact of gran-
I¯GMI
∗
= log = log csc2 θV Y . (91) ular distortion is also important when L is close to L∗ . The
mmse
achievable rate behaves like a waterfall except except for
3) θXY ∈ (max(θXV , θV Y ), π/2): it is determined by θXV the case where L is very close to L∗ . For high-resolution
and θV Y as (see Proposition 11) quantization at high SNR, we have IGMI ≈ log(L3 /4φ(L)),
which increases approximately linearly with L. On the other
cos θXY = cos θXV cos θV Y , (92)
hand, as L → 0, the effective resolution of the quantizer
1-bit
so that reduces to one bit and IGMI (L) converges to IGMI given in
r (29).
SNR √
ρXY = cos θXY = 1 − mmse. (93) 2) Underload Region (L > L∗ ): In this region granular
1 + SNR distortion dominates the performance. As L increases, the rate
The maximum achievable rate can be expressed as loss increases like log(1+L2K −2 SNR/12) over a wide range.
∗

When L ≫ L∗ , the effective resolution reduces to one bit and
IGMI = log csc2 θXY , (94) IGMI (L) converges to IGMI1-bit
again.
corresponding to an effective SNR as SNR∗e = cot2 θXY . Specifically, when the resolution is sufficiently high so that
The preceding relations show that, under varying resolu- the irreducible loss is negligible, the loading loss (60) can
tion and fixed SNR (θV Y varies while the triangle 0XV is be expressed as log(1 + γ(L)SNR), which is characterized as
fixed), according to the relationship (92), the optimal loading follows.
∗
factor lets cos θXY be directly proportional to cos θV Y . For • When L → L from below (in the overload region), the
increasingly fine quantization, the limit of the triangle 0XY loading loss decays like O(4φ(L)/L3 )SNR; i.e., it decays
is the triangle 0XV, the limit of ρ2XV Y is ρ2XV V, and the exponentially with L; see Theorem 12.
∗
limit of achievable rate is the channel capacity C. Similarly, • When L → L from above (in the underload region), the
under varying SNR and fixed resolution (θXV varies while loading loss decays like L2 K −2 SNR/12; i.e., it decays
the triangle 0VY is fixed), the optimal loading factor lets quadratically with L; see Theorem 14.
cos θXY be directly proportional to cos θXV . For increasing We first establish the preceding high-resolution asymptotic
SNR, the limit of the triangle 0XY is the triangle 0VY, the results for loading loss in Sec. V-A, and then utilize them to
limit of ρ2V Y X is ρ2V Y V, and the achievable rate tends to the characterize the behavior of IGMI (L) in Sec. V-B.
saturation rate. The low-SNR slope of the achievable rate is
cos2 θV Y = 1 − mmse.
Finally we note that a suboptimal loading factor always A. Decay of Rate Loss under High-Resolution Receiver Quan-
breaks the geometry in Fig. 7. For example, if we fix the tization
loading factor and let b increase without bound, then the We first consider the impact of bias in gain control when the
granular distortion vanishes but the overload distortion keeps irreducible loss is negligible. In other words, we consider how
∗
unchanged, so that Y cannot converge to V. For a given IGMI (L) converges to IGMI ≈ C as L → L∗ . Our method is
resolution, as the SNR increases, a suboptimal loading factor to characterize asymptotic behaviors of rate loss under finite
breaks the Pythagorean relation (74).p In this case the length of loading factor and finite step size, respectively, in the high-
Y can be larger  or smaller  than N (Ex + σ 2 − MMSE), but resolution limit. Then the accuracy of the obtained asymptotic
2
the distance E kY − Vk and the angle θV Y always become formulas is evaluated numerically for finite-resolution receiver
larger, thereby reducing the achievable rate. quantization.
DRAFT 14

The following result shows that, the rate loss due to on- which follows from bounds for the Q function as [68]
ly overload distortion decays exponentially as the loading
factor increases. Asymptotically, the loading loss is directly φ(t) t
proportional to the overload distortion, and is also directly > Q(t) > φ(t). (102)
t 1 + t2
proportional to the SNR.
Theorem 12: In the channel (4) under i.i.d. complex Gaus- The proof is completed by combining (100) and (30).
sian codebook and nearest neighbor decoding rule (7), the
Remark: By the lower bound in (102), one can show that
rate loss due to uniform quantization with a loading factor L
φ(t)− tQ(t) < φ(t)/t2 . Thus the integral in (98) can be upper
satisfies
  bounded by
4φ2 (L)
C − IGMI = mseo + (1 + oL (1)) SNR nats/c.u. Z ∞
L2 φ(t)
(95a) dt = L−1 φ(L) − Q(L). (103)
  L t2
4φ(L)
= (1 + oL (1)) SNR nats/c.u. (95b)
L3 Using the lower bound in (102) again, we obtain
in the high-resolution limit, where
Z ∞ 4φ(L)
mseo < , (104)
mseo := 2 (t − L)2 φ(t)dt (96) L3
L

is the infimum of the normalized overload distortion mseo which implies that, as L increases, mseo converges to
given in (49). 4φ(L)L−3 from below (cf. (99)).
Proof: Combining (46c) and (47c), we obtain the high- On the other hand, the rate loss due to only granular
resolution limit of γ as a function of L as distortion decays quadratically as the step size vanishes. To
prove this we need the following lemma which is one of the
γ̄(L) := lim γ (97a) various forms of the Euler-Maclaurin summation formula [60].
K→∞
1
R∞
2 Lemma 13: For a real-valued continuously differentiable
4 − L tQ(t)dt − 21 − Q(L)
= 1
R∞ (97b) function f (t) defined on [a, b], we have
4 − L tQ(t)dt
R∞
4 L (φ(t) − tQ(t)) dt − 4Q2 (L) Z K−1
!
= R∞ . (97c) b
f (a) X f (b)
1 − 4 L tQ(t)dt f (t)dt = ℓ + f (a + kℓ) +
a 2 2
k=1
In the high-resolution limit, granular distortion vanishes and
ℓ2 ′
the normalized MSE includes only the overload distortion, − (f (b) − f ′ (a)) + o(ℓ2 ), (105)
namely mseo . From Proposition 5, (46c), and (47c), we obtain 12
another expression of the overload distortion as
Z ∞ where ℓ = b−a
K .
mseo = 4 (φ(t) − tQ(t)) dt. (98) This lemma characterizes the error of numerical integration
L using the composite trapezoidal rule with an evenly spaced
From [46, Lemma 7 and Eqn. A9], we can infer that9 (uniform) grid. Based on Lemma 13, the following result can
be obtained.
4φ(L) Theorem 14: In the channel (4) under i.i.d. complex Gaus-
mseo = (1 + oL (1)) . (99)
L3 sian codebook and nearest neighbor decoding rule (7), the rate
We thus obtain loss due to uniform quantization with a step size ℓ satisfies
mseo − 4Q2 (L)

γ̄(L) = R∞ (100a)
1 − 4 L tQ(t)dt C − IGMI = mseg + o(ℓ2 ) SNR nats/c.u. (106a)
 2 
4φ2 (L) ℓ 2
= mseo + (1 + oL (1)) (100b) = + o(ℓ ) SNR nats/c.u. (106b)
L2 12
4φ(L)
= (1 + oL (1)) , (100c)
L3 in the high-resolution limit, where
where we utilize the fact
φ(t) mseg := lim mseg (107)
Q(t) = (1 + ot (1)) , (101) K→∞
t
9 The is the supremum of the normalized granular distortion mseg
derivation in [46] begins from the original form (96). However, we
can also begin from (98) and confirm (99) directly by bounds of Q-function given in (50).
1 √1
[68], e.g., t+1/t φ(t) < Q(t) ≤ 2
φ(t). Proof: Applying the Euler-Maclaurin summation formula
3t/4+ t +8/4
DRAFT 15

in Lemma 13, for A we have

101
Z Kℓ
−t2
exp dt
0 2 100
K−1
!
1 X −k 2 ℓ2 1 −K 2 ℓ2
=ℓ + exp + exp 10-1
2 2 2 2
k=1
ℓ2 −K 2 ℓ2
+Kℓ exp + o(ℓ2 ) (108a) 10-2
12 2
K−1
X −k 2 ℓ2 ℓ 10-3
= ℓ · exp −
2 2
k=0
10-4
ℓ −K 2 ℓ2 Kℓ3 −K 2 ℓ2
+
exp + exp + o(ℓ2 ) (108b)
2 2 12 2
= A + o(ℓ2 ), (108c) 0 1 2 3 4 5 6

and for B we have

Z L Fig. 8. Approximations of small irreducible rate loss: b = 12 bits.
2tQ(t)dt
0
! 101
K−1
X
=ℓ 2kℓQ(kℓ) + KℓQ(Kℓ)
k=1 100
2
 
ℓ Kℓ −K 2 ℓ2 1
− Q(Kℓ) − √ exp − + o(ℓ2 ) (109a)
6 2π 2 2 10-1
K−1
X ℓ2 ℓ2
= 2kℓ2 Q(kℓ) + − + o(ℓ2 ) (109b)
8 24 10-2
k=0
B ℓ2
= − + o(ℓ2 ). (109c) 10-3
π 24
As K → ∞, we have L = Kℓ p = ∞. Then in the high-
resolution limit we have A = π/2 + o(ℓ2 ) and B/π = 10-4
1/2 + ℓ2 /24 + o(ℓ2 ), which yield 0 1 2 3 4 5 6

ℓ2
mse = mseg = + o(ℓ2 ) (110)
12 Fig. 9. Approximations of small irreducible rate loss: b = 6 bits.
and
π
1 + o(ℓ2 ) ℓ2
γ̄ = 1 − 1
2
ℓ2
= + o(ℓ2 ). (111)
π 2 + 24 + o(ℓ2 ) 12 In this case the achievable rate is well approximated by
L2 SNR/12K 2. When the resolution decreases, e.g., in Fig.
The proof of (106a) is completed by combining (110) and
9 where the resolution is 6 bits, the approximation (112)
(30), while the proof of (106b) is completed by combining
becomes less useful, especially at high SNR. But the ap-
(111) and (30).
proximation (113) is still satisfactory. For higher accuracy,
In Theorem 12 and Theorem 14 we assume output quan-
by considering both the overload distortion and the granular
tization with unlimited resolution. We now check whether
distortion, we propose an approximation given by
they provide useful approximations of rate loss under finite
resolutions. In Fig. 8, we show the rate loss C − IGMI due to
12 bits output quantization. According to Theorem 12, a small d
C − IGMI (L) ≈ log (1 + m se · SNR) , (114)
irreducible loss in overload region can be approximated by
C − IGMI (L) ≈ 4φ(L)L−3 SNR nats/c.u., (112) where
and according to Theorem 14, a small irreducible loss in
underload region can be approximated by 4φ(L) 4φ2 (L) L2
d
m se = 3
+ 2
+ , (115)
L L 12K 2
C − IGMI (L) ≈ L2 SNR/12K 2 nats/c.u. (113)
Clearly, these approximations successfully capture the decay in which the first two terms of the RHS come from (95a),
of rate loss when it is dominated by the overload distortion. and the last term comes from (106). As shown in Fig. 9,
When the loading factor L exceeds 4, the increasing granular the proposed formula well approximates the transition from
distortion kicks in and dominates the performance quickly. overload region to underload region.
DRAFT 16

B. Properties and Approximations of IGMI (L) 10

3.3300

We first discuss some general properties of IGMI (L). The 9

derivative of the achievable rate
8
dIGMI −SNR dγ
= (116) 7
dL γSNR + 1 dL
6
shows that, for the SNR-limited scenario (defined in Sec. IV-
B), where γSNR ≪ 1 over a wide range of L, we have 5

dIGMI dγ 4
≈ −SNR , (117)
dL dL 3

which implies that the penalty of a small bias of gain control 2

increases approximately linearly with the SNR; see Fig. 8 and 1
Fig. 9. This is consistent with Theorem 12 and Theorem 14.
0
For the resolution-limited scenario (also defined in Sec. IV-B), 0 4 8 12 16
where γSNR ≫ 1, we have
dIGMI 1 dγ Approximations of IGMI (L): b = 6 bits.
≈− , (118) Fig. 10.
dL γ dL
2.6816
7
which does not varies with the SNR, implying that the
achievable rate saturates; see high-SNR curves in Fig. 4(e) 6
and Fig. 4(f).
Apart from the aforementioned cases, in general it is not 5
easy to establish a global property for IGMI (L) from the
derivative. We next propose approximations for IGMI (L) by 4
asymptotic analysis of γ(L).
1) Overload Region: In this region it is not easy to find a 3

simple and accurate approximation for IGMI (L). In the range

2
of resolution of practical interest (e.g., b ≤ 12 bits), the loading
factor in the overload region is small, so that the asymptotics
1
(101) can be inaccurate and the asymptotic expression of γ̄(L)
in (100) is less useful. For example,
Fig. 10 and Fig. 11 0
show that C − log 1 + 4φ(L)L−3 SNR is not a satisfactory 0 4 8 12 16

approximation. Instead, using

C − log (1 + mseo SNR) . (119) Fig. 11. Approximations of IGMI (L): b = 4 bits.

may approximate the waterfall better, but it cannot be ex-

pressed in a closed form. A special case is high-resolution simpler approximation C − 12L2 K −2 SNR nats/c.u. in the
quantization at high SNR, where the waterfall in the overload high-resolution scenario; see Fig. 8 and Fig. 9. However, it
region can be approximated by the high-resolution limit of the becomes less useful as the loading loss increases, as shown
saturation rate log(1/γ) as in Fig. 10 and Fig. 11. The approximation (121) also implies
  that the saturation rate can be approximated by
1 L3
log = log (1 + oL (1)) (120a) K2
γ̄(L)
 
4φ(L) I¯GMI (L) = log(1/γ) ≈ log . (122)
12L2
L3
≈ log (120b) In Figs. 5(d)-5(f), we can observe that (122) well approximates
4φ(L)
√ IGMI in the resolution-limited scenario.
2π
= L2 /2 + 3 ln L + ln nats/c.u., (120c)
4 VI. A SYMPTOTIC A NALYSIS : C HARACTERIZATION OF L∗
∗
which does not depend on resolution or SNR; see Figs. (5(d)- AND IGMI
∗
5(f)). This section focuses on IGMI , the maximum achievable rate
2) Underload Region: It is easier to approximate IGMI (L) of the transceiver architecture described in Sec. II-A when
in this region. We utilize the asymptotic expression (111) of the receiver quantization is uniform with the optimal loading
γ̄ and let the resolution be finite with 2K levels, yielding factor L∗ . In Sec. VI-A, we consider approximations of the
  optimal loading factor L∗ , which is essential for achieving
12L2 ∗
IGMI (L) ≈ C − log 1 + SNR , (121) IGMI by accurate gain control. In Sec. VI-B, we characterize
K2 ∗
IGMI as a function of the resolution b, and further provide
which is highly accurate even when the resolution is low; some per-bit rules for different performance metrics such as
see Fig. 10 and Fig. 11. We have also considered an even saturation rates and irreducible rate loss.
DRAFT 17

A. Approximations of Optimal Loading Factor with an error term

"  
We first provide a new proof of the property (61) for the step 4 ln(2K) 3
size and loading factor of the optimal uniform quantizer. In ǫ = 2 ln 1+ 1−
2K 4 ln(2K)
[67], it was shown that the uniform quantizer that minimizes
the MSE must satisfy (61) if the input density has infinite √ !!3/2 #
1 3 4 2π
support. Here we show that (61) is a natural corollary of the · 1+ ln ln(2K) + ln (130)
2 ln(2K) 2 3
achievable rate results in Theorem 12 and Theorem 14.
New Proof of (61): We consider two possible cases other included, which satisfies limK→∞ ǫ = 0.
than (61) and exclude them respectively. First, if the high- In the following result, we adopt a new way to approximate
resolution limit of the optimal step size, namely limK→∞ ℓ∗ , is L∗ , which begins from (124) in the preceding new proof of
strictly larger than zero, then limK→∞ L∗ = ∞. However, we (62).
can infer from Theorem 14 that, for a step size bounded away Proposition 15: In the channel (4) under i.i.d. complex
from zero, if the resolution is high enough, then there must Gaussian codebook and nearest neighbor decoding rule (7),
exist a smaller step size that reduces the loss in achievable rate. the optimal loading factor L∗ that maximizes the GMI (19)
Thus limK→∞ ℓ∗ must be equal to zero. Second, on the other satisfies
side, if limK→∞ L∗ < ∞, then limK→∞ ℓ∗ = 0. However, n o
we can infer from Theorem 12 that, given a finite L, if the lim L∗ − L̂0 = 0, (131)
K→∞
resolution is high enough, then there must exist a larger L that
reduces the loss in achievable rate. Thus limK→∞ L∗ cannot where L̂0 is the unique real-valued solution of the transcen-
be finite. Therefore, we conclude that limK→∞ ℓ∗ = 0 and dental equation10
limK→∞ L∗ = ∞ hold simultaneously. 
18
The scaling law (62) was obtained in [46] (as a special case L2 + 6 ln L − ln = 4 ln(2K). (132)
of a more general result) by analyzing the derivative of the π
MSE. Here we give a simpler proof of (62) from the condition Proof: According to (61) and (123), the difference be-
(52) for achievable rate maximization. The proof utilizes the tween the numerator and the denominator of (124) vanishes
Euler-Maclaurin formula. as K tends to infinity. Therefore, L∗ can be approximated with
New Proof of (62): In the high-resolution regime, by com- vanishing error by the solution of
bining (108), (109), and (52) we obtain Z ∞
R∞ L2
−t2 ∗2 √ 2tQ(t)dt − Q(L) = , (133)
Kℓ∗ exp 2 dt + o(ℓ ) L 24K 2
R∞ 1 ∗2 2 = 2π, (123)
∗
Kℓ∗ 2tQ(t)dt − 24 ℓ + o(ℓ ) since its LHS and RHS are continuous functions of L. Ac-
yielding cording to (102) and its variation
 
Q(L∗ ) 1
lim R ∞ 1 ∗2
= 1. (124) 1− φ(t) < 2tQ(t) − φ(t) < φ(t), (134)
K→∞
L∗ 2tQ(t)dt − 24K 2 L 1 + t2
Noting that we note that, instead of (133), we can turn to the solution of
Q(L∗ ) −φ(L∗ ) equation
lim R ∞ = lim = 1, (125)
K→∞
L∗
tQ(t)dt K→∞ −L∗ Q(L∗ ) 24K 2 φ(L) = L3 , (135)
where the second equality follows from (102), we obtain
∗2
which is equivalent to (132). The uniqueness of L̂ is clear since
24K 2 exp −L2 the LHS of (132) is a smooth and monotonically increasing
lim √ = 1. (126)
K→∞ 2πL∗ 3 function of L.
The proof is completed by noting that (126) implies (62).  Remark: In fact, the equation (135) can also be derived
Although p(62) provides a simple approximation of L∗ from analysis of the MSE: Just let the derivative of the MSE
as L̂1 = 2 ln(2K), it can be refined by more elaborate approximation (115) be zero, and omit the exponential terms
techniques. In [46, Sec. V] three approximations have been therein except for the dominated one.
proposed. For Gaussian input, the first is just L̂1 , and the other In Fig. 12, we show L∗ , its approximation p L̂0 from (131)
two satisfy a stronger condition in Proposition 15, the simple approximation 2 ln(2K), and
n o the approximations L̂2 and L̂3 from [46]. Note that the case
lim L∗ − L̂i = 0, i = 2, 3, (127) b = 1 does not require gain control. We see that the new
K→∞
approximation L̂0 is accurate for all resolutions b ≥ 2. It
where
r is much better than L̂2 and close to the more complicated
32π approximation L̂3 .
L̂2 = 4 ln(2K) − 3 ln ln(2K) − ln (128)
9
10 A transcendental equation x + ln x = ln a can be solved by the Lambert
and
q W function as x = W (a). Similarly, the solution L̂ in Proposition 15 can be
given by a special case of the generalized Lambert W function, although less
L̂3 = L̂22 + ǫ, K > 1, (129) is known about this function [69].
DRAFT 18

5.5

4.5

3.5

2.5

1.5

0.5

0
1 2 3 4 5 6 7 8 9 10 11 12

Fig. 12. Numerical results, asymptotics, and approximations of L∗ .

∗
Remark: To approach IGMI , it is sufficient to use a simple If we replace mmse in (55) by 4b3·4lnb2 , then we obtain IˆGMI (b).
∗
linear approximation of L as According to (140), it is straightforward to show that the gap
b+4
∗
between IGMI and IˆGMI (b) is ob (1) for an arbitrary finite SNR,
L̂lin = . (136) thereby completing the proof.
3
The approximation (140) implies a 6-dB-per-bit-rule as
Fig. 12 shows that it is accurate for 2 ≤ b ≤ 7. When b ≥ 8,
its accuracy degrades, but the loading loss caused is negligible 10 log10 SNRq ≈ 6.02b − 10 log10 b + 0.34 (dB), (141)
since it grows quadratically with L; see numerical results in which has been known since [46]. Therefore, in the high-
Fig. 4. resolution regime, each additional bit in resolution reduces
the MSE by four times:
B. Maximum Achievable Rate Approximations and Per-Bit mmse(b)
Rules lim = 4. (142)
b→∞ mmse(b − 1)
∗
We have found a simple relationship between IGMI and The approximation (140) is still not accurate enough for
mmse in Theorem 8, and we have also given numerical moderate to low resolutions; see Table I. More approximations
results of mmse for b = 1, 2, ..., 12 in Table I. However, a for MMSE or SNRq can be found in [46, Sec. IV] based
∗
direct connection between IGMI and the resolution is still very on refined asymptotic formulas of L∗ and its asymptotic
useful. The following result provides such a connection via relationship with MMSE. However, using (140) is enough to
approximating mmse by b. get the simple and useful approximation of IGMI ∗
in Proposition
Proposition 16: The achievable rate given in (55) satisfies 16. In Fig. 13 we compare IGMI with its approximation IˆGMI .
∗

∗
IGMI = IˆGMI (b) + ob (1), (137) It is shown that the accuracy is acceptable when b ≥ 2.
We next introduce some per-bit rules for other performance
where metrics. The following one is obtained from Proposition 16
 
4b ln 2 immediately.
ˆ
IGMI (b) = C − log 1 + SNR . (138)
3 · 4b • A 2-bpcu-per-bit rule for saturation rate: The saturation
rate I¯GMI
∗
(b) given in (66) satisfies
Proof: From (62) and (63) we obtain
I¯GMI
∗
(b) = Iˆ¯GMI (b) + o(1), (143)
3K 2
lim mmse = 1, (139)
K→∞ ln(2K) where
which yields a high-resolution approximation of MMSE given Iˆ¯GMI (b) = 2b − log2 b + 0.11 bits/c.u., (144)
by which implies that
4b ln 2 4b ln 2

mmse = (1 + ob (1)) b
≈ . (140) lim I¯GMI
∗
(b) − I¯GMI
∗
(b − 1) = 2 bits/c.u. (145)
3·4 3 · 4b b→∞
DRAFT 19

14 101

12

100
10

10-1
6

4
10-2

0 10-3
-10 0 10 20 30 40 2 4 6 8 10 12

Fig. 13. Approximation of IGMI

∗ . Fig. 15. A 5-dB-per-bit rule for irreducible rate loss.

100
20
100
-1 18
10

16
10-2 10-2
14

10-3 12
10-4
10
-4
10
8

10 -5 6 10-6

4
10-6
2 10-8

10-7 0
2 4 6 8 10 12 2 4 6 8 10 12 14 16

Fig. 14. Normalized MMSE and saturation rate under different resolutions. Fig. 16. For irreducible rate loss, a 6-dB-per-bit rule is inaccurate.

Numerical evaluations of Iˆ ¯GMI (b) for 1 ≤ b ≤ 12 are This rule can be confirmed by numerical results in Fig. 15.
shown in Table I. In Fig. 14, it is shown that mmse decreases Although the 6-dB-per-bit rule is well-known in quantization
exponentially as the resolution increases, and correspondingly, theory and gives the correct asymptotics (as shown in (141)),
the saturation rate I¯GMI
∗
grows approximately linearly with it is less accurate unless the rate loss is extremely small, see
the resolution. The improvement per bit is 1.6-1.9 bits/c.u. Fig. 16.
in our range of interest, although the high-resolution limit is
2 bits/c.u. VII. D ISCUSSIONS ON F URTHER Q UANTIZATION RULES
The second per-bit rule can be observed from results in AT THE R ECEIVER
Table I, suggesting a roughly 5-dB-per-bit increase of SNRq :
Based on the analytical framework given in Theorem 2, a
10 log10 SNRq ≈ 5b. (146) major part of this paper has focused on the simplest receiver
quantization scheme, namely scalar uniform quantization with
Since the rate loss in (55) is determined by SNR/SNRq , (146)
mid-rise levels. This section briefly discusses nonuniform and
implies that, in our range of interest, a 5 dB increase of SNR
other types of quantization rules in several aspects.
requires an extra bit of resolution to maintain the same rate
loss. More specifically, we have the following rule.
• A 5-dB-per-bit rule for irreducible rate loss: We require A. Relationship Between MSE and Achievable Rate: Numeri-
a resolution of at least cal Examples
For the channel (4) with a uniform quantizer q(·) whose
2 log10 SNR + b0 (147)
resolution is at least two bits, we have shown that the unique
∗
bits so that the irreducible loss C − IGMI can be as small gain control factor that minimizes the MSE also maximizes the
−b0
as 10 bits/c.u., where b0 ≥ 0. GMI. When non-uniform quantization rules are allowed, one
DRAFT 20

1 1 1.4
3 3 GMI GMI
GMI 2.5
mse mse 1.2
mse 0.8 0.8
2.5 2.5
2 1
2
0.6
2 0.6
0.8
1.5
1.5 1.5
0.4 0.4 0.6
1
1 1
0.4
0.2 0.2
0.5 0.5 0.5
0.2

0 0 0 0 0 0
0 2 4 6 8 10 0 1 2 3 4 5 0 2 4 6 8 10

(a) Uniform quantizer: (b) Optimal nonuniform quantizer: (c) Nonuniform quantizer:
thresholds 0, ±0.25, ±0.5, ±0.75, thresholds 0, ±0.5006, ±1.0500, ±1.7480, thresholds 0, ±0.2, ±0.6, ±0.7,
levels ±0.125, ±0.375, ±0.625, ±0.875 levels ±0.2451, ±0.7560, ±1.3440, ±2.1520 levels ±0.1, ±0.5, ±0.7, ±0.9

3.005 0.04 3.02 0.055 3.5

GMI GMI 2
3 mse GMI 3
mse 0.0395
3 mse
2.98 0.05
2.5
0.039 1.5
2.96
2.995
2
0.0385 2.94 0.045
1
1.5
2.99
2.92
0.038
1
2.9 0.04
0.5
2.985
0.0375
2.88 0.5

2.98 0.037 2.86 0.035 0 0

2.1 2.2 2.3 2.4 2.5 2.6 1.6 1.8 2 2.2 2.4 2.6 0 5 10 15 20

(d) Part of (a): 2.1 ≤ L ≤ 2.6 (e) Optimized quantizer with equispaced thresholds: (f) Nonuniform quantizer:
thresholds 0, ±0.25, ±0.5, ±0.75, thresholds 0, ±0.2, ±0.6, ±0.7,
levels ±0.1224, ±0.3673, ±0.6122, ±0.9693 levels ±0.1, ±0.9, ±0.8, ±0.2

Fig. 17. GMI, MSE and γ under gain control: SNR = 10 dB.

may vary the thresholds {ℓk } and levels {yk } of q(·) (possibly From Fig. 17(a), we confirm that, under uniform quantiza-
under some constraints) to reduce γ so that the GMI (19) tion, the parameter γ converges to mmse when the loading
can be optimized, yielding thresholds {ℓ∗k } and levels {yk∗ }. factor approaches its optimal value L∗ (in this example it
However, when this optimized quantization rule is applied in is 2.3441); see also Fig. 17(d). Interestingly, this consistency
a given channel, to achieve that optimized GMI, we need to also occurs in Fig. 17(b), where we use the optimal 8-level
set the gain control factor g appropriately according to σv quantizer obtained numerically in [30] by Lloyd’s algorithm
(the standard deviation of the quantizer input) to guarantee [70]. But for the nonuniform quantizer given in Fig. 17(c)
that the thresholds satisfy ℓk = lk /g = ℓ∗k for 1 ≤ k < K. and the optimized quantizer with equispaced thresholds (also
Otherwise, the performance will be degraded. We next explore obtained in [30]) given in Fig. 17(e), the gain control factor
the impact of gain control on the MSE and the GMI of the that maximizes the GMI and the one that minimizes the MSE
channel (4) under nonuniform quantization rules by examples; are different. The highly nonlinear quantizer given in Fig. 17(f)
see numerical results given in Fig. 17, where we set SNR = 10 provides an example of unusual behavior of IGMI (g), which
dB. The examples are described as follows, each of which achieves its maximum when the MSE is relatively large.
satisfies 2K = 4, i.e., b = 2. Remark: In achievable rate evaluation, if we use the additive
noise model of the quantizer, i.e., treating the MSE as additive
• The uniform quantizer (with equispaced thresholds and
Gaussian noise, then we obtain an estimate as
mid-rise levels); see Fig. 17(a) and Fig. 17(d).  
• The optimal nonuniform quantizer: the thresholds and Ex
R̂ = log 1 + 2 (148)
levels are optimized to maximize the GMI; see Fig. 17(b). σ + 2 · mse · σv2
• Two nonuniform quantizers without optimization: the first  
mse
one has monotonically increasing levels, and the second = log(1 + SNR) − log 1 + SNR , (149)
1 + mse
one is highly nonlinear since its levels are no longer
monotonically increasing; see Fig. 17(c) and Fig. 17(f), which may cause underestimate or overestimate. In particular,
respectively. it overestimates the achievable rate when the optimal uniform
• An optimized quantizer with equispaced thresholds: its quantization is used, while it may significantly underestimate
thresholds and levels are optimized to maximize the the achievable rate when mse is much larger than γ, which is
GMI under the constraint that the thresholds must be possible; see the numerical results in Fig. 17 when the gain
equispaced; see Fig. 17(e). control factor (or loading factor) is far away from its optimal
DRAFT 21

value. which follows from L’Hôpital’s rule. Moreover, according to

∗
(102), the largest level yK given in (152) satisfies
B. On Improving Uniform Quantization by Post-Processing ∗ 1
(K − 1)ℓ > yK > (K − 1)ℓ + , (154)
In [39], [57], it has been shown that, the GMI (14) given (K − 1)ℓ
in Proposition 1 can be improved by introducing a post- ∗
implying that yK /ℓ converges to K − 1 as ℓ → 0. From these
processing of the channel output as Y → Ỹ , thereby extending facts we can conclude that the gain of MMSE post-processing
the channel X → Y to X → Ỹ ; The GMI is maximized by
asymptotically vanishes as the resolution increases.
letting
Ỹ = E [X|Y ] , (150) C. On Possible Gain of Futher Quantization Rules
i.e., using a conditional expectation operator as the post- For communication receivers, nonuniform quantization rules
processor, which gives an MMSE estimation of X under can be realized directly, or indirectly by combining a uniform
the observation Y . Thus we can infer that the optimized quantizer with a pre-processing of its input (e.g., companding
quantizer with equispaced thresholds considered in Sec. VII-A in pulse-code modulation (PCM)), depending on the cost of
is equivalent to a uniform quantizer combined with an MMSE implementation. There have been some works on numerical
post-processing of its output. In fact, for a quantizer with ℓ- optimization of nonuniform quantization rules with respect to
spaced thresholds, the optimal levels have been shown to be different performance measures, e.g., cutoff rate [71], mutual
[30] information [8], [72], GMI [30], and MSE [21], [23]. But
no general characterization of the performance gain has been
φ((k − 1)ℓ) − φ(kℓ)
yk∗ = , k = 1, ..., K − 1 (151) given.
Q((k − 1)ℓ) − Q(kℓ)
As shown in Fig. 17(b), for nonuniform quantizer, there
and is also a consistency between GMI maximization and MSE
∗ φ((K − 1)ℓ) minimization, but we must optimize all the thresholds and
yK = . (152)
Q((K − 1)ℓ) levels, rather than only a single factor (g or L). In [30,
Appendix E] a proof of this consistency has been given. Thus,
Such levels satisfy yk∗ = E [V |V ∈ [(k − 1)ℓ, kℓ)] , k =
∗ the maximum achievable rate can be expressed as
1, ..., K − 1 and yK = E [V |V ≥ (K − 1)ℓ], where V ∼
N (0, σv ) is the input of the quantizer; i.e., the optimal level yk∗
2 ∗
IGMI = log(1 + SNR) − log(1 + mmsenu · SNR), (155)
is the centroid of the corresponding interval. This is consistent
where mmsenu is the normalized minimum MSE of all possible
with the aforementioned result that a conditional expectation
nonuniform quantizers. This result can also be understood
operator maximizes the GMI. Comparing Fig. 17(e) and Fig.
via a similar geometrical interpretation as that given in Sec.
17(d), we see that, under the same loading factor L = kℓ,
IV-C. From (155) we can infer that, under our setting of
such a post-processing indeed improves the GMI slightly.
transceiver architecture, the maximum gain in achievable rate
Interestingly, the optimal loading factor that maximizes the
from replacing the optimal uniform quantizer by the optimal
GMI in Fig. 17(e) is smaller than the optimal loading factor in
nonuniform quantizer is determined by the reduction of the
Fig. 17(d). In other words, when post-processing is available,
normalized MMSE from mmse to mmsenu . The numerical
the optimal step size of the uniform quantizer becomes smaller,
results in Fig. 17(a) and Fig. 17(b) show that the rate gain
and a concatenation of an MMSE uniform quantizer and an
is marginal for b = 2. For higher resolutions the rate gain
MMSE post-processor is strictly worse than the optimized
grows slowly. When b = 4, according to the numerical results
quantizer with equispaced thresholds.
Since the example Fig. 17(e) shows that the gain of the in [30], the optimal non-uniform quantization improves the
MMSE post-processing is limited for low-resolution uniform saturation rate log(1/γ) by no more than 4.4%, and opti-
quantization, we may infer that the gain is also limited under mized quantization with equispaced thresholds improves the
high resolution. In fact, as the step size decreases, the centroid saturation rate by 1.5%. In fact, for nonuniform quantization
of the interval [(k − 1)ℓ, kℓ) given in (151) tends to be its with Gaussian input, the asymptotic quantization theory, in
midpoint.11 To see this, note that for an arbitrary c > ℓ/2, particular the Panter-Dite formula [73], yields a “6-dB-per-


bit” rule12 for the maximum SNR as [41]
φ c − 2ℓ − φ c + 2ℓ
lim

10 log10 SNRnu
q = 6.02b − 4.35 + o(1), (156)
ℓ→0 Q c − ℓ − Q c + ℓ
2 2


where SNRnu nu
φ c − 2ℓ − φ c + 2ℓ q = 1/mmse . Therefore, under our setting, the
= lim R c+ 2ℓ (153a) rate gain increases logarithmically with the resolution b, and
ℓ→0
c− 2ℓ
φ(t)dt for b ≤ 10 the gain is limited (cf. (141)).
c− 2ℓ ℓ

c+ 2ℓ ℓ

According to (156), even if the best nonuniform scalar
2 φ c − 2
+ 2 φ c +

2 quantization is used, the MSE achieved in the high-resolution
= lim 1 ℓ ℓ
(153b)
2 φ c+ 2 +φ c− 2
ℓ→0
12 There is another version of the 6-dB-per-bit rule as B
eff = (SNRq −
= c, (153c) 1.76)/6.02, which has been widely used in practice to calculate the effective
number of bits (ENoB) of a quantizer [2], [3]. It is derived under assumptions
11 In [54] the asymptotic convergence of centroids to midpoints was proved that the granular distortion is uniformly distributed over [−ℓ/2, ℓ/2] and there
under uniform quantization with infinitely many levels. is no overload distortion; see [2] for more details.
DRAFT 22

regime is still 2.72 times larger than the theoretical limit from constellation shaping. But it is natural to ask whether
implied by the rate-distortion function of a Gaussian source our results are still useful in the cases of QAM, phase-shift
[42].13 If we allow coded uniform scalar quantization (i.e., keying (PSK), and other commonly used inputs, which are all
representing the quantizer output by a variable-rate lossless bounded (unlike the unbounded Gaussian input). Since in these
code), the gap can be reduced to only 1.53 dB (πe/6) worse cases Proposition 1 does not apply, the GMI evaluation may
than the theoretical limit (see [41] and references therein). rely heavily on numerical computation. We left this problem
Unfortunately, coded quantization does not help in commu- to future study. Nevertheless, the insight gained in this work
nication receivers since the bottleneck therein is the limited will be very helpful for the finite-alphabet case.
resolution of the quantizer rather than the cost of representing Finally, we list some topics that can be addressed following
its output. Applying vector quantization to communication our information-theoretic framework of receiver quantization.
receivers is a more challenging topic. Nevertheless, the gap to • The orthogonal frequency division multiplexing (OFDM),
the fundamental limit reminds us that it is possible to alleviate which typically generate Gaussian-like signal. Therein, a
the ADC resolution bottleneck in communication receivers by new effect is that the nonlinear distortion due to quan-
exploring new quantization mechanisms. tization leads to intercarrier interference. The problem
becomes more complicated when the channel introduces
VIII. C ONCLUDING R EMARKS time-dispersion.
The goal of this study is to evaluate the impact of resolution • Multiuser channels, especially the Gaussian multiple-
reduction on information-theoretic limits of communications access channel. In this case a new phenomenon due to
with receiver quantization. Leveraging the GMI as a basic receiver quantization is that, when successive interference
tool, which enables us to take the decoding rule into con- cancellation (SIC) is used to decode a user, there exists
sideration, we establish an array of exact and asymptotic residual interference from other users which may signif-
results under a standard transceiver architecture. Our results icantly reduces the achievable rate of that user.
indicate a critical issue in system design that arises as the • Multiantenna channels. We need an extended version
resolution decreases, namely, optimizing the loading factor by of the proposed analytical framework which applies for
gain control, which minimizes the loss in achievable rate by different receiver architectures (e.g., MMSE receiver,
eliminating the loading loss. The remaining irreducible loss maximal ratio combining receiver).
is an appropriate evaluation of the inherent robustness of the
considered transceiver architecture. Our results also establish A PPENDIX A
explicit connections between the MSE (affected by the gain P ROOF OF P ROPOSITION 4
control) and the rate loss. Although for general receiver
quantization rules, smaller MSE does not necessarily imply We first note that all the adjustable levels in {ℓk =
smaller rate loss, for the commonly used uniform quantizer lk /g, k = 1, ..., K − 1} tend to zero as g → ∞, and
we prove that the unique loading factor that minimizes the they tend to infinity as g → 0. For the case √of g → ∞,
MSE also maximizes the GMI (i.e., achieves zero loading from (21) and (22) we have limg→∞ A/yK = 2πφ(0) and
2
loss). For perfect gain control, we show that the irreducible limg→∞ B/yK = πQ(0). For the case of g → 0, noting that
loss is determined by the product of the normalized MMSE A and B can be expressed by
and the SNR, and provide a geometrical interpretation for this
√ K−1
X
result. Performance approximations and per-bit rules in this A = y1 + 2π φ(ℓk )(yk+1 − yk ) (157)
case are also given. For imperfect gain control, to understand k=1
its impact, we characterize the decay of small rate loss in
the high-resolution regime, and propose approximations of the and
achievable rate as a function of the loading factor which is K−1
X
π 2 2
fairly accurate for moderate resolutions. These results provide B= y1 + π Q(ℓk )(yk+1 − yk2 ), (158)
2
insight into transceiver design with nonnegligible quantization k=1
effect, especially choice of quantizer resolution and design of respectively, it is direct to check that limg→0 A/y1 = 1 and
AGC. limg→0 B/y12 = π/2. Therefore, in both cases γ tends to 1 −
A limitation of this work is that the obtained analytical 2/π and the limits of the GMI are the same, IGMI1-bit
.
results are derived under the Gaussian codebook. Like many
classical information-theoretic results given by mutual infor-
mation, for the GMI, it is also not easy to get closed-form A PPENDIX B
or analytical results without assuming Gaussian input. For P ROOF OF P ROPOSITION 5
practical systems with finite alphabet, we can expect that our We begin from the definition (39) in which the quantization
result applies well for Gaussian-like constellations that comes rule is given by (8) and is equivalent to
13 The comparison is made under the assumption that the signal at the
q(gV ) = yk · sgn(V ), if ℓk−1 ≤ v < ℓk , (159)
receiver front-end is represented by b bits per channel use before feeding
into the decoder. The MSE of the representation can be reduced by lossy
source coding (treating the received signal as a source). A scalar quantizer is where v = |V |/σv ∼ N (0, 1). Utilizing the symmetries of the
a very simple lossy source coding scheme. input distribution and the quantization rule with respect to the
DRAFT 23

K−1 K−1
!
origin, we have X X 1
2
= πℓ (2k + 1)Q(kℓ) − Q(kℓ) + (163c)
K Z
X ℓk 8
k=0 k=0
mse = 2 yk2 φ(v)dv K−1
X
ℓk−1 1
k=1 =π 2kℓ2 Q(kℓ) + πℓ2 . (163d)
K Z ℓk
X 8
k=0
−2 2yk vφ(v)dv + 1 (160a)
k=1 ℓk−1 A PPENDIX D
XK
P ROOF OF L EMMA 7
= 1−4 yk (φ(ℓk−1 ) − φ(ℓk ))
k=1 Let K > 1. According to Theorem 2, the LHS of (51) is
K
X equivalent to dγ
dℓ = 0, or
+2 yk2 (Q(ℓk−1 ) − Q(ℓk )) (160b)
dA A dB
k=1 = . (164)
2 √  dℓ 2B dℓ
= 1− 2πA − B . (160c) According to Proposition 5, the RHS of (51) is equivalent to
π
The lower bound and condition of equality follow from dA 1 dB
r !2 = √ . (165)
dℓ 2π dℓ
1 2
mse − γ = A− B ≥ 0. (161) Let
B π
K−1
X  
−k 2 ℓ2
A PPENDIX C C= ℓ · k 2 ℓ2 · exp , (166)
2
P ROOF OF P ROPOSITION 6 k=0

According to (21), we can derive (44) by straightforward which is strictly positive when K > 1. Noting that
derivation as K−1
X
dA −k 2 ℓ2 ℓ
X
K−1
ℓ

−(k − 1)2 ℓ2 −k 2 ℓ2
 ℓ· = ℓ · exp −
A= kℓ − exp − exp dℓ 2 2
k=0
2 2 2
k=1
  X 
K−1
−k 2 ℓ2

ℓ −(K − 1) ℓ2 2 − ℓ · k 2 ℓ2 · exp (167a)
+ Kℓ − exp (162a) 2
2 2 k=0
K−1  
X 
K−1
−(k − 1)2 ℓ2 −k 2 ℓ2
 X
2 2 −k 2 ℓ2
= kℓ exp − exp =A− ℓ · k ℓ · exp (167b)
2 2 2
k=0
k=1
K−1   =A−C (167c)
1 X −(k − 1)2 ℓ2 −k 2 ℓ2
− ℓ exp − exp
2 2 2 and
k=1
K−1 √
−(K − 1) 2 2
ℓ ℓ −(K − 1) ℓ 2 2
1 dB √ X 2 2π 2
+ Kℓ exp − exp (162b) √ ℓ· = 2π 2kℓ Q(kℓ) + ℓ
2 2 2 2π dℓ 8
K−1 k=0
X −k 2 ℓ2
= ℓ · exp −
ℓ
, (162c) X 
K−1
−k 2 ℓ2

2 2 2 2
k=0
− ℓ · k ℓ exp (168a)
2
k=0
and according to (22), we obtain (45) as r
2 X 
K−1
−k 2 ℓ2

X
K−1
1
2 =
π
B− ℓ · k 2 ℓ2 exp
2
(168b)
B=π k− ℓ2 (Q((k − 1)ℓ) − Q(kℓ)) r k=0
2
k=1
 2 2
= B − C, (168c)
1 π
+π K− ℓ2 Q((K − 1)ℓ) (163a)
2 it is direct to check that both (164) and (165) are equivalent
 K−1
X to (52), thereby completing the proof.
= πℓ2 k 2 (Q((k − 1)ℓ) − Q(kℓ))
k=1
K−1 A PPENDIX E
X
− k (Q((k − 1)ℓ) − Q(kℓ)) P ROOF OF P ROPOSITION 11
k=1 First, (80) and (81) can be obtained by the Pythagorean
K−1
X 
1 relations (73) and (74), respectively. We begin from (24c) and
+ (Q((k − 1)ℓ) − Q(kℓ)) (25) and rewrite them as
4
k=1
 2 r r
1    2 2   2
+π K− ℓ2 Q((K − 1)ℓ) (163b) E XY = E |X| A = E XV A (169)
2 π π
DRAFT 24

and [17] K. Gao, J. N. Laneman, and B. Hochwald, “Capacity of multiple one-bit

   2 transceivers in a Rayleigh environment,” in Proc. IEEE Wireless Commun.
E |Y|2 = E |V|2 B, (170) Net. Conf. (WCNC), May 2018.
π [18] Y. Nam, H. Do, Y. Jeon, and N. Lee, “On the capacity of MISO channels
with one-bit ADCs and DACs,” IEEE J. Sel. Areas Commun., vol. 37,
respectively. Combining them with (80) we obtain no. 9, pp.2132–2145, Sept. 2019.
   r [19] N. I. Bernardo, J. Zhu, Y. C. Eldar, and J. Evans, “Capacity bounds for
E XY E XV πA one-bit MIMO Gaussian channels with analog combining,” IEEE Trans.
2
= 2
, (171)
E [|Y| ] E [|V| ] 2 B Commun., vol. 70, no. 11, pp. 7224–7239, Nov. 2022.
[20] A. Mezghani and J. A. Nossek, “Capacity lower bound of
thereby implying that (52) is equivalent to (78). According to MIMO channels with output quantization and correlated noise,”
(81), we have in Proc. 2012 IEEE Int. Symp. Inf. Theory, July 2012, Cam-
bridge, MA, USA, July 2012, pp. 1732-1736. Available: http-
    s://mediatum.ub.tum.de/doc/1171263/1171263.pdf (not available on IEEE
E VY = (Ex + σ 2 )(1 − mmse) = E |Y|2 . (172) Xplore.).
[21] L. Fan, S. Jin, C. K. Wen, and H. Zhang, “Uplink achievable rate for
Substituting (172) into (171) yields massive MIMO systems with low-resolution ADC,” IEEE Commun. Lett.,
   r vol. 19, no. 12, pp. 2186–2189, Dec. 2015.
E XY E VY πA [22] K. Roth and J. A. Nossek, “Achievable rate and energy efficiency of
= , (173)
E [|X|2 ] E [|V|2 ] 2 B hybrid and digital beamforming receivers with low resolution ADC,”
IEEE J. Select. Areas Commun., vol. 35, no. 9, pp. 2056-2068, Sept.
thereby implying that (52) is equivalent to (77). Combine (171) 2017.
and its equivalent form (173), we obtain [23] S. Jacobsson, G. Durisi, M. Coldrey, U. Gustavsson, and C. Studer,
“Throughput analysis of massive MIMO uplink with low-resolution
πA2 ADCs,” IEEE Trans. on Wirel. Commun., vol. 16, no. 6, pp. 4038–4051,
ρ2XY = ρ2XV ρ2V Y , (174) June 2017.
2B 2 [24] Y. Li, C. Tao, G. Seco-Granados, A. Mezghani, A. L. Swindlehurst, and
thereby implying that (52) is equivalent to (79). L. Liu, “Channel estimation and performance analysis of one-bit massive
MIMO systems,” IEEE Trans. Signal. Process., vol. 65, no. 15, pp. 4075–
4089, Aug. 2017.
R EFERENCES [25] S. Dutta, C. N. Barati, D. Ramirez, A. Dhananjay, J. F. Buckwalter, and
S. Rangan, “A case for digital beamforming at mmWave,” IEEE Trans.
[1] J. Zhou, S. Pang, and W. Zhang, “A high-resolution analysis of receiver Wirel. Commun., vol. 19, no. 2, pp. 756–769, Feb. 2020.
quantization in communication,” IEEE Int. Symp. Inf. Theory (ISIT), Ann [26] Q.-U.-A. Nadeem and A. Chaaban, “Analysis of one-bit quantized linear
Arbor, MI, USA, June 2025. precoding schemes in multi-cell massive MIMO downlink,” IEEE Trans.
[2] R. H. Walden, “Analog-to-digital converter survey and analysis,” IEEE J. Commun., vol. 72, no. 5, pp. 2577–2594, May 2024.
Sel. Areas Commun., vol. 17, no. 4, pp. 539–550, Apr. 1999. [27] J. J. Bussgang, “Crosscorrelation functions of amplitude-distorted Gaus-
[3] B. Murmann, “ADC Performance Survey 1997-2023,” [Online]. Avail- sian signals,” Technical Report No. 216, Research Laboratory of Electron-
able: https://github.com/bmurmann/ADC-survey ics, Massachusetts Institute of Technology, Cambridge, MA, Mar. 1952.
[4] F. J. Bloom, S. S. L. Chang, B. Harris, A. Hauptschein, and K. C. Morgan, [28] Ö. T. Demir and E. Björnson, “The Bussgang decomposition of nonlin-
“Improvement of binary transmission by null-zone reception,” Proc. IRE, ear systems: Basic theory and MIMO extensions [lecture notes],” IEEE
vol. 45, no. 7, pp. 963–975, July 1957. Signal Process. Mag., vol. 38, no. 1, pp. 131–136, Jan. 2021.
[5] J. M. Wozencraft and I. M. Jacobs, Principles of Communication Engi- [29] A. Lozano and S. Rangan, “Spectral vs energy efficiency in 6G: Impact
neering. New York, NY, USA: Wiley, 1965. of the receiver front-end,” IEEE BITS Inf. Theory Mag., vol. 3, no. 1, pp.
[6] A. J. Viterbi and J. K. Omura, Principles of Digital Communication and 41–53, March 2023.
Coding. New York, NY, USA: McGraw-Hill, 1979. [30] W. Zhang, “A general framework for transmission with transceiver
[7] T. Koch and A. Lapidoth, “At low SNR, asymmetric quantizers are better,” distortion and some applications,” IEEE Trans. Commun., vol. 60, no.
IEEE Trans. Inf. Theory, vol. 59, no. 9, pp. 5421–5445, Sep. 2013. 2, pp. 384–399, Feb. 2012.
[8] J. Singh, O. Dabeer, and U. Madhow, “On the limits of communication [31] N. Liang and W. Zhang, “Mixed-ADC massive MIMO,” IEEE J. Sel.
with low-precision analog-to-digital conversion at the receiver,” IEEE Areas Commun., vol. 34, no. 4, pp. 983–997, Apr. 2016.
Trans. Commun., vol. 57, no. 12, pp. 3629–3639, Dec. 2009. [32] J. Scarlett, A. G. i Fabregas, A. Somekh-Baruch, and A. Martinez,
[9] B. M. Murray and S. Reisenfeld, “Maximizing the cutoff rate in a “Information-theoretic foundations of mismatched decoding,” Found.
quantized MIMO wireless systems with AGC,” in Proc. 1st Int. Conf. Trends Commun. Inf. Theory, vol. 17, no. 2, Mar. 2020.
Wireless Broadband Ultra Wideband Commun. (AusWireless’06), Sydney, [33] A. Ganti, A. Lapidoth, and İ. E. Telatar, “Mismatched decoding revisited:
2006. General alphabets, channels with memory, and the wide-band limit,” IEEE
[10] B. M. Murray and I. B. Collings, “AGC and quantization effects in a Trans. Inf. Theory, vol. 46, no. 7, pp. 2315–2328, Nov. 2000.
zero-forcing MIMO wireless system,” in Proc. IEEE 63rd Veh. Technol. [34] A. Martinez, A. G. i Fábregas, G. Caire, and F. M. J. Willems,
Conf., 1802–1806, May 2006. “Bit-interleaved coded modulation revisited: A mismatched decoding
[11] J. A. Nossek and M. T. Ivrlač, “Capacity and coding for quantized MI- perspective,” IEEE Trans. Inf. Theory, vol. 55, no. 6, pp. 2756–2765,
MO systems,” in Proc. Intl. Conf. Wireless Commun. Mobile Computing June 2009.
(IWCMC), 2006, Vancouver, Canada, pp. 1387–1391. [35] A. Lapidoth and S. Shamai (Shitz), “Fading channels: How perfect need
[12] A. Mezghani, J. A. Nossek, and A. L. Swindlehurst, “Low SNR ’perfect side information’ be?” IEEE Trans. Inf. Theory, vol 48, no. 5,
asymptotic rates of vector channels with one-bit outputs,” IEEE Trans. pp. 1118–1134, May 2002.
Inf. Theory, vol. 66, no. 12, pp. 7615–7634, Dec. 2020. [36] H. Weingarten, Y. Steinberg, and S. Shamai(Shitz), “Gaussian codes and
[13] J. Mo and R. W. Heath, “Capacity analysis of one-bit quantized MIMO weighted nearest neighbor decoding in fading multiple-antenna channels,”
systems with transmitter channel state information,” IEEE Trans. Signal IEEE Trans. Inf. Theory, vol. 50, no. 8, pp. 1665–1686, Aug. 2004.
Process., vol. 63, no. 20, pp. 5498–5512, Oct. 2015. [37] G. Kramer, “Information rates for channels with fading, side information
[14] J. Mo, A. Alkhateeb, S. Abu-Surra, and R. W. Heath, “Hybrid archi- and adaptive codewords,” Entropy, vol. 25, no. 5, May 2023.
tectures with few-bit ADC receivers: Achievable rates and energy-rate [38] M. Secondini and E. Forestieri, “Scope and limitations of the nonlinear
tradeoffs,” IEEE Trans. Wirel. Commun., vol. 16, no. 4, pp. 2274–2287, Shannon limit (invited paper),” J. Lightw. Technol., vol. 35, no. 4, pp.
April 2017. 893-902, Feb. 2017.
[15] S. Rini, L. Barlett, E. Erkip, and Y. C. Eldar, “A general framework for [39] Y. Wang and W. Zhang, “Generalized nearest neighbor decoding,” IEEE
MIMO receivers with low-resolution quantization,” in Proc. IEEE Inf. Trans. Inf. Theory, vol. 68, no. 9, pp. 5852–5865, Sept. 2022.
Theory Workshop (ITW), Kaohsiung, Taiwan, Nov. 2017, pp. 599–603. [40] B. Li, N. Liang, and W. Zhang, “On transmission model for massive
[16] A. Khalili, S. Rini, L. Barletta, E. Erkip, and Y. C. Eldar, “On MIMO MIMO under low-resolution output quantization,” in Proc. 2017 IEEE
channel capacity with output quantization constraints,” In Proc. 2018 85th Veh. Technol. Conf. (VTC Spring), Sydney, NSW, Australia, Jun.
IEEE Int. Symp. Inf. Theory (ISIT), pp. 1355–1359. 2017.
DRAFT 25

[41] R. M. Gray and D. L. Neuhoff, “Quantization,” IEEE Trans. Inf. Theory, [69] I. Mezö and Á. Baricz, “On the generalization of the Lambert W
vol. 44, no. 6, Oct. 1998. function,” Trans. Amer. Math. Soc., vol. 369, no. 11, pp. 7917–7934,
[42] T. Berger, Rate Distortion Theory: A Mathematical Basis for Data Nov. 2017.
Compression. Englewood Cliffs, NJ: Prentice-Hall, 1971. [70] S. Lloyd, “Least squares quantization in PCM,” IEEE Trans. Inf. Theory,
[43] B. M.Oliver, J. R. Pierce, and C. E. Shannon, “The philosophy of PCM,” vol. IT-28, no. 2, pp. 129–137, Mar. 1982.
Proc. IRE, vol. 36, no. 11, pp. 1324–1331, Nov. 1948. [71] J. L. Massey, “Coding and modulation in digital communication,” in
[44] W. R. Bennett, “Spectra of quantized signals,” Bell Syst. Tech. J., 1948, Proc. Zurich Sem. Digital Commun., Vol. 2, No. 1, 1974.
vol. 27, no. 3, pp. 446–472. [72] Q. Yu and M. Médard, “The asymptotic solutions of the capacity
[45] H. Gish and J. N. Pierce, “Asymptotically efficient quantizing,” IEEE maximal quantization problem,” in Proc. 2015 IEEE 82nd Veh. Technol.
Trans. Inf. Theory, vol. IT-14, pp. 676–683, Sept. 1968. Conf. (VTC2015-Fall), 2015.
[46] D. Hui and D. L. Neuhoff, “Asymptotic analysis of optimal fixed-rate [73] P. F. Panter and W. Dite, “Quantization distortion in pulse-count modu-
uniform scalar quantization,” IEEE Trans. Inf. Theory, vol. 47, no. 3, lation with nonuniform spacing of levels,” Proc. IRE, vol. 39, pp. 44–48,
March 2001. Jan. 1951.
[47] A. Gersho and R. M. Gray, Vector Quantization and Signal Compression.
Boston, MA: Kluwer, 1992.
[48] N. S. Jayant and P. Noll, Digital Coding of Waveforms: Principles and
Applications to Speech and Video. Englewood Cliffs, NJ: Prentice-Hall,
1984.
[49] S. Na and D. L. Neuhoff, “Asymptotic MSE distortion of mismatched
uniform scalar quantization,” IEEE Trans. Inf. Theory, vol. 58, no. 5, pp.
3169–3171, May 2012.
[50] S. Na and D. L. Neuhoff, “On the convexity of the MSE distortion of
symmetric uniform scalar quantization,” IEEE Trans. Inf. Theory, vol. 64,
no. 4, pp. 2626–2638, Apr. 2018.
[51] S. Na and D. L. Neuhoff, “Monotonicity of step sizes of MSE-optimal
symmetric uniform scalar quantizers,” IEEE Trans. Inf. Theory, vol. 65,
no. 3, pp. 1782–1792, Mar. 2019.
[52] F. Sun, J. Singh, and U. Madhow, “Automatic gain control for ADC-
limited communication,” in Proc. IEEE Global Telecommun. Conf.
(GLOBECOM), Miami, FL, USA, Dec. 2010.
[53] S. Krone and G. Fettweis, “Optimal gain control for single-carrier
communications with uniform quantization at the receiver,” in Proc. 2010
Int. Conf. Acoust. Speech Sig. Process. (ICASSP’10), Dallas, USA, Mar.
2010.
[54] D. Marco and D. L. Neuhoff, “The validity of the additive noise model
for uniform scalar quantizers,” IEEE Trans. Inf. Theory, vol. 51, no. 5,
pp. 1739–1755, May 2005.
[55] D. Tse and P. Viswanath, Fundamentals of Wireless Communication,
New York: Cambridege Unviersity Press, 2005.
[56] A. Lapidoth, “Nearest neighbor decoding for additive non-Gaussian
noise channels,” IEEE Trans. Inf. Theory, vol. 42, no. 5, pp. 1520–1529,
May 1996.
[57] W. Zhang, Y. Wang, C. Shen, and N. Liang, “A regression approach
to certain information transmission problems,” IEEE J. Select. Areas
Commun., vol. 37, no. 11, pp. 2517–2531, Nov. 2019.
[58] Y. Wu, L. M. Davis, and R. Calderbank, “On the capacity of the discrete-
time channel with uniform output quantization,” in Proc. 2009 IEEE Int.
Symp. Inf. Theory (ISIT’09), Seoul, Korea, 2009, pp. 2194–2198.
[59] G. D. Forney and G. Ungerboeck, “Modulation and coding for linear
Gaussian channels,” IEEE Trans. Inf. Theory, vol. 44, no.6, pp. 2384–
2415, Oct. 1998.
[60] J. Stoer and R. Bulirsch, Introduction to Numerical Analysis. 3rd edition.
New York, NY: Springer.
[61] E. N. Gilbert, “Increased information rate by oversampling,” IEEE Trans.
Inf. Theory, vol. 39, no. 6, pp. 1973–1976, Nov. 1993.
[62] S. Shamai (Shitz), “Information rates by oversampling the sign of a
bandlimited process,” IEEE Trans. Inf. Theory, vol. 40, no. 4, pp. 1230–
1236, July 1994.
[63] T. Koch and A. Lapidoth, “Increased capacity per unit-cost by over-
sampling,” in Proc. IEEE 26th Conv. Elect. Electron. Eng. Israel, Eilat,
Israel, Nov. 2010, pp. 684–688.
[64] L. T. N. Landau, M. Dörpinghaus, and G. P. Fettweis, “1-bit quantiza-
tion and oversampling at the receiver: Sequence-based communication,”
EURASIP J. Wirel. Commun. Netw., vol. 2018, no. 1, pp. 83, Apr. 2018.
[65] R. Deng, J. Zhou, and W. Zhang, “Bandlimited communication with one-
bit quantization and oversampling: Transceiver design and performance
evaluation,” IEEE Trans. Commun., vol. 69, no. 2, pp. 845-862, Feb.
2021.
[66] B. Hassibi and B. M. Hochwald, “How much training is needed in
multiple-antenna wireless links?” IEEE Trans. Inf. Theory, vol. 49, no.4,
pp. 951–963, Apr. 2003.
[67] J. A. Bucklew and N. C. Gallagher, “Some properties of uniform step
size quantizers,” IEEE Trans. Inf. Theory, vol. IT-26, no. 5, pp. 610–613,
Sept. 1980.
[68] P. Borjesson and C.-E. Sundberg, “Simple approximations of the error
function Q (x) for communications applications,” IEEE Trans. Commun.,
vol. 27, no. 3, pp. 639–643, Mar. 1979.