Mixed Delay Constraints on a Fading C-RAN Uplink

A cloud radio access network (C-RAN) is considered where the first hop from the user equipments (UEs) to the basestations (BSs) is modeled by the fading Wyner soft-handoff model. The focus is on mixed-delay constraints where a set of messages (so called “slow” messages) are jointly decoded in the cloud unit (CU), whereas the remaining messages (called “fast” messages) have to be decoded immediately at the BSs. This paper presents inner and outer bounds on the capacity region for such a setup. Moreover, the multiplexing gain region is characterized exactly. The presented results show that for small fronthaul capacity it is beneficial to send both “fast” and “slow” messages. However, when the rate of “fast” messages is already large, then increasing it further, deteriorates the sum-rate of the system. In this regime, the stringent decoding delay on the “fast” messages penalizes the overall performance. Our results indicate that this penalty is larger at moderate SNR than at high SNR and it is also larger for random time-varying fading coefficients than for static ones.


I. INTRODUCTION
The fifth generation (5G) of wireless cellular network has to accommodate different types of data traffics with different latency constraints. Delay-tolerant traffics allow for higher transmission rate by exploiting cooperation possibilities as in cloud radio access networks (C-RAN) [1]- [5]. In C-RAN, base stations (BSs) are connected to a cloud unit (CU) via finiterate fronthaul links and data are typically decoded jointly at the CU to alleviate the effect of interference [1]. Delay-sensitive traffics are however not compatible with this new technology because they have to be decoded immediately at the BSs, and consequently cannot profit from joint processing.
In this paper, we consider transmission over a C-RAN with mixed delay constraints, i.e., where each mobile user can simultaneously send a delay-sensitive and a delay-tolerant stream. Throughout this paper, we call delay-sensitive traffic "fast" messages and delay-tolerant traffic "slow" messages. Mixed delay constraints in C-RANs have previously been studied in [6] where different decoding techniques are compared. In [6] UEs close to the BSs send only "fast" messages and it is assumed that these communications do not interfer. UEs located further away send "slow" messages and their interference pattern is modeled by Wyner's symmetric network [11], [12] with static channel coefficients.
In this work, each UE sends a pair of independent "fast" and "slow" messages. Communications between UEs and BSs are modeled by a Wyner's soft-handoff network with random time-varying channel coefficients. We present coding schemes for this setup and derive inner and outer bounds on the capacity region. Furthermore, we characterize its exact first-order high-SNR asymptotics, i.e., the multiplexing gain region. This result allows us to conclude that for moderate fronthaul capacities, the maximum "slow" multiplexing gain remains unchanged over a large regime of small and moderate "fast" multiplexing gains. The sum-multiplexing gain is thus improved if some of the messages can directly be decoded at the BSs. In contrast, for large fronthaul capacities or large "fast" multiplexing gains, this sum-multiplexing gain deteriorates by ∆ if one further increases the "fast" multiplexing gain by ∆.
At moderate SNR the conclusion based on our inner bound are slightly different: If the "fast" rate is small or moderate, then the achievable sum-rate decreases by ∆ when the rate of "fast" messages increases by ∆, and if the "fast" rate is large it decreases with a factor γ times ∆. The penalty factor γ is approximately 1 for static channel coefficients and typically higher for random coefficients. The stringent delay constraint on "fast" messages thus seems to be more harmful at moderate SNR and for time-varying channel conditions than at high SNR or for static channels.
Previous studies on mixed-delay constraints for cellular networks with cooperation [7]- [10] presented similar conclusions as our high-SNR results: For small or moderate "fast" rates the overall performance is not degraded by the stringent delay constraints. For large "fast" rates 1 bit of "fast" rate comes at the expense of 2 bits of "slow" rate.
Note that in this work, we use simple point-to-point compression techniques. Employing more sophisticated compression techniques [4] is left for our future work.

II. PROBLEM SETUP
Consider the uplink communication of a multi-cell C-RAN with K UEs and K BSs. UEs and BSs are indexed by 1, . . . , K. Each BS is connected to a CU via a separate fronthaul link of capacity C (see Fig. 1). At a given time t ∈ {1, . . . , n}, the signal received at BS k is described as where X k,t and X k− . . . . . .
represent the time-t fading coefficients. We assume that the sequence of channel coefficients is i.i.d. over time and distributed according to the K-tuple distribution of a given stationary and ergodic process Each BS k has perfect channel state information (CSI) about its own channel, i.e., it observes the realizations of {(G k,t , F k,t )} for all t ∈ {1, . . . , n}. The UEs know only the statistics of the random channel coefficients and are said to have no CSI. Figure 1 shows an extract of our system model. Each UE k wishes to convey the pair of independent messages (M and has to be decoded by BS k as we explain shortly. The "slow" source message M are the rates of transmissions of the "fast" and the "slow" messages.
UE k computes its channel inputs X n k := (X k,1 , . . . , X k,n ) as a function of the pair (M for some function φ (n) k on appropriate domains so that the average block-power constraint is satisfied. Each BS k decodes the "fast" source message M (F ) k based on its own channel outputs Y n k := (Y k,1 , . . . , Y k,n ). So, it produces:M using some decoding function ψ (n) k on appropriate domains. It further produces the fronthaul message using some encoding function The CU then decodes the set of "slow" messages as by means of a decoding function b (n) .
The main focus of this paper is the achievable sumrates of "fast" and "slow" messages. Given a maximum fronthaul link capacity C and a maximum allowed power P, the pair of (average) rates (R (F ) , R (S) ) is called achievable if for each positive integer K there exists a sequence (in n) of encoding and decoding functions {φ tends to 0 as n → ∞ and the rates satisfy Note that the notation lim refers to the limit superior. Definition 1 (Capacity Region): The capacity region C(P, C) is the closure of the set of all rate pairs (R (F ) , R (S) ) that are achievable with power P and fronthaul link capacity C. We are particularly interested in the capacity in the asymptotic high-SNR regime. The pair of multiplexing gains (S (F ) , S (S) ) is called achievable with frontaul multiplexing gain µ, if there exists a sequence of rates {R (F ) (P), R (S) (P)} P>0 so that and for each P > 0 the pair (R (F ) (P), R (S) (P)) is achievable with fronthaul capacity Definition 2 (Multiplexing Gains): The closure of the set of all achievable multiplexing gains (S (F ) , S (S) ) is called multiplexing gain region and denoted S (µ).
and the random process {W β i } ∞ i=−∞ as the unique stationary process satisfying: (17) Notice that the joint process {(F i , G i , W β i )} ∞ i=−∞ is also stationary and ergodic.
For a given β ∈ [0, 1], let R(β) ⊆ R 2 be the set of all non-negative pairs (R (F ) , R (S) ) that satisfy and Proof: See Section IV. Theorem 2 (Capacity Outer Bound): Assuming log |G 0 | and log |F 0 | are integrable near 0, any rate pair (R (F ) , R (S) ) in the capacity region C(P, C) satisfies the following four constraints: Proof: Omitted due to space limitations. Corollary 1 (Multiplexing Gain Region): The multiplexing gain region S (µ) is the set of all nonnegative pairs (S (F ) , S (S) ) satisfying Proof: The converse holds by Theorem 2 and the achievability by Theorem 1. Specifically, for the achievability part, it suffices to prove that the two pairs are achievable. The multiplexing gain pair in (22) can be achieved by silencing every second UE, which decomposes the network into K/2 non-interfering point-to-point links. If µ ≥ 1, the multiplexing gain pair (S (S) = 1, S (F ) = 0) is achieved by a scheme where each BS quantizes its observed outputs to noise level and the CU decodes all the transmitted "slow" messages based on these quantized outputs. If µ < 1, then the multiplexing gain pair (S (S) = µ, S (F ) = 1/2 − µ/2) in (23) is achieved by a scheme that time-shares the schemes above over fractions of 1 − µ and µ of the time. Figure 2 illustrates the proposed inner and outer bounds on the capacity region for independent random processes {G i } and {F i }, where each F i is circularly Gaussian of variance σ 2 F and each G i is circularly Gaussian of variance σ 2 G . Numerical simulations are performed for different values of σ 2 F . The figure also presents inner and outer bounds on the capacity region assuming static channel coefficients (regions in red). As can be seen from the figure, for small values of R (F ) , the slope δR (F ) δR (S) of the inner bound is approximately −1 both for static and random channel coefficients. This means increasing the rate of "fast" messages by ∆, decreases the rate of "slow" messages by ∆ and thus the sum-rate remains constant. For large values of R (F ) and random time-varying channel coefficients, the slope of the inner bound is around −3.5 for σ 2 F = 0.2 and around −4 for σ 2 F = 0.3. In contrast, this slope is around −2.7 for static channel coefficients. Increasing an already large "fast" rate R (F ) thus penalizes the sum-rate of the system and is more pronounced under random fading.  Figure 3 shows the multiplexing gain region for different values of µ. We notice that when µ < 1, for S (F ) ≤ 1 2 − µ 2 , the multiplexing gain of "slow" messages is constant and solely limited by the fronthaul capacity. In this regime, the summultiplexing gain of the system is increased by decoding parts of the messages directly at the BSs. When µ < 1 and S (F ) > 1 2 − µ 2 , or when µ ≥ 1, the slope of the boundary of the region is −2. In these regimes the maximum sum-multiplexing gain is decreased by ∆ when the "fast" multiplexing gain increases by ∆.
Random Code Construction: For each k ∈ {1, . . . , K}, generate codebooks C u,k , C v,k , and C w,k randomly. Codebook Reveal all codebooks to all terminals. We explain the encoding and decoding operations assuming that (Recall that BS k and the CU know these realizations.) UE k: Sends BS k: Decodes its "fast" message M (F ) k based on its own channel outputs Y n k = y n k . It then looks for a uniqueî k such that where A (n) (·) refers to the jointly typical set as defined in [14] and where given G 0 = g and F 0 = f the pair (U, Y ) is a centered bivariate Gaussian vector of covariance matrix If none or more than one such indicesî k exist, BS k declares an error. Otherwise it declareŝ Subsequently it forms the difference c n k := y n k − u n k (î k ), and looks for an index k such that (ĉ n k ( k ), c n k , f n k , g n k ) ∈ A (n) (PĈ CF G ).

V. CONCLUSION
We presented inner and outer bounds on the capacity region of a C-RAN under mixed delay constraints and characterized the multiplexing gain region of this network. We obtained the following conclusions. When the fronthaul capacities are small, then the overall performance of the system can be improved if some of the data streams (the delay sensitive streams) are directly decoded at the BSs. The stringent delay constraint on these streams however becomes harmful when their rate is too large. In this regime the total sum-rate has to be decreased by a penalty factor γ times ∆ when the delaysensitive rate is increased by ∆. The penalty factor γ ≈ 1 for static channel coefficients or in the high-SNR regime, and it can be significantly larger for random channel coefficients and at moderate SNRs.
To reduce the gap between the proposed inner and outer bounds on capacity, in future works we plan to include more sophisticated multi-user compression techniques [4], [5] at the BSs and the cloud processor.