I Introduction
In an effort to save power and cost in wideband wireless transceiver systems, lowresolution (especially onebit) analogtodigital converters (ADCs) [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13] and digitaltoanalog converters (DACs) [14, 15, 16] are being considered in transmitter and receiver chains, especially in systems involving many such chains. The nonlinearity introduced by coarse quantization becomes a limiting factor in the achievable throughput of such a wireless system. Channel capacity is one measure of this throughput.
There is a rich literature on the subject of capacity with coarse quantization. The capacity of a system with onebit ADCs at the receiver is analyzed in [1, 2, 3, 4, 6, 7, 5, 8, 9]
with various assumptions about the channel, the channel information, and communication schemes. Communication techniques including channel estimation and signal detection for a multipleinput multipleoutput (MIMO) system with onebit ADCs at the receiver are studied in
[10, 11, 12, 13]. A communication system with onebit DACs at the transmitter is studied in [14, 15, 16].While many of the efforts consider lowresolution quantization effects at the transmitter or receiver, a few consider lowresolution quantization on both, including [17] and [18]. A linear minimummeansquarederror precoder design is proposed for a downlink massive MIMO scenario to mitigate the quantization distortion in [17] and the performance analysis of a system with a small number of onebit transceivers is studied in [18]. We focus on a model where onebit quantization is considered at both the transmitter and receiver:
(1) 
where and are the number of transmitters and receivers, and are the transmitted and received signals, is the channel matrix known to the receiver, is the additive Gaussian noise with and is independent of and , is the expected received SNR at each receive antenna. The function provides the sign of the input as its output. The channel is modeled as realvalued since only the inphase (I) information is used and the quadrature (Q) phase is ignored at the receiver. A Rayleigh channel is assumed, with each element to be independent Gaussian . This assumption appears to hold for nonline of sight (NLOS) channels in many frequency bands [19], and also appears in the analysis in [6, 7, 8, 9, 10, 11, 14].
Our contribution is a large and analysis, where the ratio is constant, of the capacity for the model (1). Analytical expressions are derived that can be used to explain the behavior of the system in various limiting regimes of operation in and .
Ii Capacity for a large number of transmitters and receivers
The capacity of the channel in (1) as a function of , , and is
(2) 
where we have normalized by , and where is the input distribution independent of , and is the mutual information notation. When with a ratio , the capacity is defined as
(3) 
We can readily see that because each transmitter can transmit at most one bit of information, and each receiver can decode at most one bit of information. Therefore, for all and .
Limiting capacities such as (3) are difficult to compute in closed form, but the “replica method” [20] can be brought to bear on the problem. Some details of how to apply the method are presented in Section IV. For now, we present the result:
(4) 
where is the capacity of a single transceiver with SNR , which is defined as
(5) 
where is the binary entropy function, and are the solutions of
(6) 
(7) 
(8) 
Equation (4) gives the capacity for any SNR and , and some limiting situations are readily analyzed, including: (i) high SNR, ; (ii) low SNR, ; (iii) many more receivers than transmitters, ; (iv) many more transmitters than receivers, . These are now presented, with only limited proofs.
Iia High SNR ()
When , the system is effectively becoming “noisefree”, and we might expect , but as we show this does not happen for all . For , , and (4) becomes
(9) 
where is defined in (5), (7) can be simplified as
(10) 
The expression (9) is not especially intuitive, but it is not difficult to solve. We show some numerical examples in Section III. It turns out that in this case, solving (9
) is essentially equivalent to solving for the “quenched entropy” for Gibbs learning of the Ising perceptron (Section 7.2 in
[21]).IiB Low SNR ()
IiC ()
IiD ()
Iii numerical evaluation of capacity
We first compare in (2) with in (4) for and , to show how the large and limit (4) can be used to approximate the exact capacity. Figure 1 displays that the approximation is accurate for small for a wide range of SNR (from 0 dB to 30 dB) when . When is larger than 0.7, is not big enough and a larger is required to get a valid approximation. We can see that can saturate at 1 when with SNR smaller than 30 dB, and an SNR higher than 30 dB is required to achieve the maximum for , but cannot achieve the maximum when . We will show later that is required to achieve the maximum.
Figure 2 displays in (4) for and varying from 0.1 to 10 with step 0.1. We can see that increases linearly with and when and are small, and the rate of increase slows down dramatically as and grow and nears saturation at . When is small but is large, saturates at (its upper bound). When is small but is large, we can see that increases with and reaches its maximum value 1. We show that for any , when .
Contours of constant for and are shown in Figure 3. We can observe that there is generally a sharp tradeoff between and , and that operating near the knee in the curve is generally desirable for a given capacity since both and are small.
Furthermore, the contours are dense when and start becoming sparse when , thus showing that has started to “saturate” at 0.8 and improves only slowly with further increases in either or .
The contours allow us to explore optimal operating points. For example, given a cost function where for some constant , we find an approximately optimal point to achieve is and . Attempts to make smaller will require significant increase in , and attempts to make smaller will require significant increase in .
Figure 4 shows the accuracy of the approximations of at high and low SNR. Plotted are examples when SNR is large (10 dB to 30 dB) and SNR is low (10 dB, ) of the actual capacity (4) and the corresponding approximations (9) and (12). Of particular interest is the approximately linear growth in (9) with until it reaches the saturation point when . This hard limit value of 1.24 receive antennas for every transmit antenna is perhaps surprising.
The curves for low SNR show that at is close to the low SNR approximation in (12) with a simple second order expression when . In general, when , we need for an accurate low SNR approximation according to (11).
Figure 5 presents a comparison of (4) with the large , small approximations in (14) and (16). We obtain excellent agreement for even the modest values and over a wide range of SNR. Moreover, according to (13), when , we have for any , and thus . This differs from the high SNR case, where when even as .
Iiia Tradeoff between and for fixed
We are interested in characterizing the contours in Figure 3 analytically, and we use the large approximation for in (14). Since in (14) is just a function of , to achieve some target capacity , we solve for numerically, and denote the result as . With , (13) then provides the relationship between and . To simplify the relationship, (13) can be further approximated as
(17) 
with good accuracy when . The relationship between and is then
(18) 
Iv Replica Analysis
The replica method, a tool used in statistical mechanics and applied to the theory of spin glass [20], has been applied in many communication system contexts [22, 23, 24, 25, 26]
[27, 21], errorcorrecting codes [28], and image restoration [29]. A mathematically rigorous justification of the replica method is elusive, but the success of the method maintains its popularity. We apply the method to solve for a closedform answer to (3). We omit many details, and present only the primary steps.Because the channel is unknown to the transmitter, according to [18], the optimal input distribution is , and then (2) becomes
where is the standard definition of entropy.
Then, (3) becomes
(20) 
The replica method is used to compute the limit, and the processes are similar to that used in [22, 23, 24, 25, 26].
We start with the identity
which holds for . The idea of the replica method is to compute as a function of by treating as a positive integer, and then assume the result to be valid for .
We assume the limit of and can commute, then
(21) 
where .
Now, we regard as a positive integer and we have
where is the th replica of (), and
are i.i.d. uniform distributed in
.Based on (1), we further have
(22) 
where is the th element of , is the wellknown Qfunction, and is the th row of ,
(23) 
with .
Let and . Then , where is the covariance matrix with elements for and . Then only depends on :
(24) 
and (22) becomes
where
(25) 
We can consider as the distribution of a random symmetric matrix , and we have
Similarly to [24], we apply Varadhan’s theorem and GartnerEllis theorem[30] and obtain
(26) 
where is an matrix with as its elements, and is defined as
(27) 
Based on the distribution of in (25), we further have
(28) 
where are independent uniform distributed in .
and that achieve the optimal value described in (26) are called the saddle point[21]. The saddle point either stays on the boundary ( or ) or satisfies , i.e.
(29) 
Here, we further assume that permutations among the replicas with index will not change the saddle point. This assumption is called the “replica symmetry” (RS) assumption. At the saddle point, we let
(30) 
which are called RS saddle points.
Based on (20), (21), (26), and (30), we can get (4), where is the solution when the saddle point is on the boundary (). The remaining expressions in (4) are obtained when the saddle point satisfies (29), and the corresponding RS saddle point is the solution of (6)(8).
Iva Extension to complex signals
The realvalued model (1) is now extended to both I and Q phase at the transmitter and receiver, and hence
(31) 
where are defined as
where are the real and imaginary parts of the channel, and are the real and imaginary parts of the transmitted and received signal, and and are additive noise. The elements in , , , and are independent Gaussian , and is the expected received SNR at each receive antenna.
Since the channel is known only to the receiver, the uniform input is optimal and the channel capacity is
When with a ratio , the capacity is defined as
(32) 
Similarly to the analysis for real signal, we have
(33) 
(34) 
Using the replica method with the RS assumption, we obtain
(35) 
and therefore
(36) 
Consequently, the IQ model capacity is twice the Ionly capacity.
V Conclusion
We have presented the capacity per transmitter in the limit where the number of singlebit transmitters and receivers is large, and where was fixed. A flat Rayleigh fading channel was considered, and we assumed the channel was only known by the receiver. We were able to derive a variety of approximations using the analytical results, and showed that the largesystem formulas are useful even for a small numbers of transmitters and receivers. We examined how saturated with either large or , and gave formulas for exploring the contours of fixed as a function of and . Further work in expanding these results to different channel models would be of great interest.
Acknowledgment
The authors are grateful for the support of the National Science Foundation, grants ECCS1731056, ECCS1509188, and CCF1403458.
References
 [1] J. Singh, O. Dabeer, and U. Madhow, “On the limits of communication with lowprecision analogtodigital conversion at the receiver,” IEEE Transactions on Communications, vol. 57, no. 12, pp. 3629–3639, 2009.
 [2] S. Krone and G. Fettweis, “Capacity of communications channels with 1bit quantization and oversampling at the receiver,” in 2012 35th IEEE Sarnoff Symposium, 2012, pp. 1–7.
 [3] J. Mo and R. W. Heath, “Capacity analysis of onebit quantized MIMO systems with transmitter channel state information,” IEEE Transactions on Signal Processing, vol. 63, no. 20, pp. 5498–5512, 2015.
 [4] ——, “High SNR capacity of millimeter wave MIMO systems with onebit quantization,” in 2014 Info. Theory and Applications Workshop, 2014, pp. 1–5.
 [5] A. Mezghani and J. A. Nossek, “Analysis of 1bit output noncoherent fading channels in the low SNR regime,” in 2009 IEEE Int. Symp. Information Theory, 2009, pp. 1080–1084.
 [6] ——, “On ultrawideband MIMO systems with 1bit quantized outputs: Performance analysis and input optimization,” in 2007 IEEE Int. Symp. on Information Theory, 2007, pp. 1286–1289.
 [7] ——, “Analysis of Rayleighfading channels with 1bit quantized output,” in 2008 IEEE Int. Symp. Information Theory, 2008, pp. 260–264.
 [8] C. Mollén, J. Choi, E. G. Larsson, and R. W. Heath, “Onebit ADCs in wideband massive MIMO systems with OFDM transmission,” in Acoustics, Speech and Signal Processing (ICASSP), 2016 IEEE International Conference on. IEEE, 2016, pp. 3386–3390.
 [9] ——, “Uplink performance of wideband massive MIMO with onebit ADCs,” IEEE Transactions on Wireless Communications, vol. 16, no. 1, pp. 87–100, 2017.
 [10] J. Choi, J. Mo, and R. W. Heath, “Near maximumlikelihood detector and channel estimator for uplink multiuser massive MIMO systems with onebit ADCs,” IEEE Transactions on Communications, vol. 64, no. 5, pp. 2005–2018, 2016.
 [11] Y. Li, C. Tao, L. Liu, G. SecoGranados, and A. L. Swindlehurst, “Channel estimation and uplink achievable rates in onebit massive MIMO systems,” in 2016 IEEE Sensor Array Multi. Sig. Proc. Work., 2016, pp. 1–5.
 [12] J. Mo, P. Schniter, N. G. Prelcic, and R. W. Heath, “Channel estimation in millimeter wave MIMO systems with onebit quantization,” in Signals, Systems and Computers, 2014 48th Asilomar Conference on. IEEE, 2014, pp. 957–961.
 [13] C. Studer and G. Durisi, “Quantized massive MUMIMOOFDM uplink,” IEEE Transactions on Communications, vol. 64, no. 6, pp. 2387–2399, June 2016.
 [14] A. K. Saxena, I. Fijalkow, A. Mezghani, and A. L. Swindlehurst, “Analysis of onebit quantized ZF precoding for the multiuser massive MIMO downlink,” in Signals, Systems and Computers, 2016 50th Asilomar Conference on. IEEE, 2016, pp. 758–762.
 [15] Y. Li, T. Cheng, L. Swindlehurst, A. Mezghani, and L. Liu, “Downlink achievable rate analysis in massive MIMO systems with onebit DACs,” IEEE Communications Letters, 2017.
 [16] S. Jacobsson, G. Durisi, M. Coldrey, and C. Studer, “Massive MUMIMOOFDM downlink with onebit DACs and linear precoding,” arXiv preprint arXiv:1704.04607, 2017.
 [17] O. B. Usman, H. Jedda, A. Mezghani, and J. A. Nossek, “MMSE precoder for massive MIMO using 1bit quantization,” in 2016 IEEE Int. Conf on Acoust, Speech and Signal Proc., 2016, pp. 3381–3385.
 [18] K. Gao, N. Estes, B. Hochwald, J. Chisum, and J. N. Laneman, “Powerperformance analysis of a simple onebit transceiver,” in Information Theory and Applications Workshop (ITA), 2017. IEEE, 2017, pp. 1–10.
 [19] T. S. Rappaport, R. W. Heath Jr, R. C. Daniels, and J. N. Murdock, Millimeter wave wireless communications. Pearson Education, 2014.
 [20] M. Mezard, G. Parisi, M. A. Virasoro, and D. J. Thouless, “Spin glass theory and beyond,” Physics Today, vol. 41, p. 109, 1988.
 [21] A. Engel and C. Van den Broeck, Statistical mechanics of learning. Cambridge University Press, 2001.

[22]
T. Tanaka, “Analysis of bit error probability of directsequence CDMA multiuser demodulators,”
Advances in Neural Information Processing Systems, pp. 315–321, 2001.  [23] ——, “A statisticalmechanics approach to largesystem analysis of CDMA multiuser detectors,” IEEE Transactions on Information theory, vol. 48, no. 11, pp. 2888–2910, 2002.
 [24] ——, “Statistical learning in digital wireless communications,” in International Conference on Algorithmic Learning Theory. Springer, 2004, pp. 464–478.
 [25] D. Guo and S. Verdú, “Randomly spread CDMA: Asymptotics via statistical physics,” IEEE Transactions on Information Theory, vol. 51, no. 6, pp. 1983–2010, 2005.
 [26] ——, “Multiuser detection and statistical mechanics,” in Communications, Information and Network Security. Springer, 2003, pp. 229–277.
 [27] E. Gardner, “The space of interactions in neural network models,” Journal of physics A: Mathematical and general, vol. 21, no. 1, p. 257, 1988.
 [28] A. Montanari and N. Sourlas, “The statistical mechanics of turbo codes,” The European Physical Journal BCondensed Matter and Complex Systems, vol. 18, no. 1, pp. 107–119, 2000.
 [29] H. Nishimori and K. M. Wong, “Statistical mechanics of image restoration and errorcorrecting codes,” Physical Review E, vol. 60, no. 1, p. 132, 1999.
 [30] A. Dembo and O. Zeitouni, “Large deviations techniques and applications,” Applications of Mathematics, vol. 38, 1998.
Comments
There are no comments yet.