Ferkans — Interactive Telecom Tutor

Downlink Multi-User MIMO: The Dual Problem

The broadcast channel (BC) models the downlink: a base station with $N_t$ antennas transmits independent messages to $K$ single-antenna users. While the MAC capacity region has been known since the 1970s, the BC capacity region remained open until the early 2000s. The breakthrough came from a beautiful duality: the BC capacity region under a sum power constraint equals the capacity region of a dual MAC with the same channels but reversed roles — a result that transforms the hard encoding problem into the well-understood decoding problem.

Definition:
MISO Broadcast Channel

The $K$ -user MISO broadcast channel consists of a base station with $N_t$ antennas transmitting to $K$ single-antenna users. The received signal at user $k$ is:

$y_k = \mathbf{h}_k^H \mathbf{x} + n_k, \quad k = 1, \ldots, K$

where $\mathbf{h}_k \in \mathbb{C}^{N_t}$ is the channel to user $k$ , $\mathbf{x} \in \mathbb{C}^{N_t}$ is the transmitted signal satisfying the sum power constraint $\mathbb{E}[\|\mathbf{x}\|^2] \leq P$ , and $n_k \sim \mathcal{CN}(0, \sigma^2)$ .

The base station encodes $K$ independent messages $W_1, \ldots, W_K$ into $\mathbf{x}$ . Each user $k$ decodes only its own message $W_k$ from $y_k$ .

Unlike the MAC, each BC receiver sees only a scalar observation and cannot cooperate with other receivers. This asymmetry makes the BC fundamentally harder to analyse: superposition coding (as in the degraded BC) is no longer optimal in general, and dirty paper coding is required.

Theorem: BC-MAC Duality

The capacity region of the $K$ -user MISO broadcast channel with sum power constraint $P$ equals the union over all individual MAC power allocations of the dual MAC capacity region:

$\mathcal{C}_{\text{BC}}(P) = \bigcup_{\substack{P_1, \ldots, P_K \geq 0 \\ P_1 + \cdots + P_K = P}} \mathcal{C}_{\text{MAC}}(P_1, \ldots, P_K)$

where the dual MAC has $K$ single-antenna transmitters (with powers $P_1, \ldots, P_K$ ) and an $N_t$ -antenna receiver with the same channel vectors $\mathbf{h}_1, \ldots, \mathbf{h}_K$ .

Furthermore, the BC capacity region is achieved by dirty paper coding (DPC): encoding each user's message while pre-cancelling the interference from previously encoded users.

Duality swaps transmitter and receiver roles while preserving the channel vectors. The MAC receiver performs SIC (cancelling already- decoded users); the BC transmitter performs DPC (pre-cancelling known interference). Both operations "remove" inter-user interference one user at a time, and the information-theoretic achievable regions coincide under the same sum power. This is not merely an analogy — it is an exact equivalence of achievable rate regions.

Proof

Achievability via DPC

Order users as $1, 2, \ldots, K$ . The base station encodes user $K$ 's message first, then user $K-1$ 's message while pre-cancelling user $K$ 's codeword using DPC, and so on. By Costa's theorem, pre-cancelling known interference incurs no rate loss. User $k$ 's achievable rate is:

$R_k = \log_2\!\left(1 + \frac{P_k'|\mathbf{h}_k^H\mathbf{w}_k|^2} {\sigma^2 + \sum_{j < k}P_j'|\mathbf{h}_k^H\mathbf{w}_j|^2}\right)$

where $P_k'$ and $\mathbf{w}_k$ are the power and beamforming direction for user $k$ , and users $j > k$ have been pre-cancelled by DPC.

Converse via dual MAC

For any encoding scheme achieving $(R_1, \ldots, R_K)$ on the BC, construct a dual MAC scheme achieving the same rates. The transmit covariance matrices in the BC are transformed to individual user powers in the MAC via a power transformation that preserves the sum power.

Formally, if the BC transmits $\mathbf{x} = \sum_k \mathbf{x}_k$ with covariances $\mathbf{Q}_k$ , the dual MAC powers satisfy:

$P_k^{\text{MAC}} = \text{tr}(\mathbf{Q}_k)$

The MAC sum constraint $\sum_k P_k^{\text{MAC}} = \sum_k \text{tr}(\mathbf{Q}_k) = P$ matches the BC sum power constraint.

Rate region equivalence

The rate expressions in both channels involve the same SINR values after applying the uplink-downlink SINR duality (Vishwanath, Goldsmith, Jafar, 2003). Specifically, for every DPC encoding order in the BC, there exists a corresponding SIC decoding order in the dual MAC achieving identical rates, and vice versa.

The union over all power splits $(P_1, \ldots, P_K)$ with $\sum P_k = P$ in the MAC covers all possible DPC covariance matrices in the BC, establishing $\mathcal{C}_{\text{BC}} = \bigcup \mathcal{C}_{\text{MAC}}$ . $\blacksquare$

, ,

Definition:
Dirty Paper Coding (DPC) for the Broadcast Channel

Dirty paper coding is a non-linear encoding technique for the broadcast channel that pre-cancels known interference at the transmitter. When encoding user $k$ 's message, the signals intended for users $k+1, \ldots, K$ (which have already been encoded) are known and treated as "dirt on the paper."

By Costa's theorem, if interference $\mathbf{s}$ is known non-causally at the transmitter:

$C_{\text{DPC}} = \log_2\!\left(1 + \frac{P}{\sigma^2}\right) = C_{\text{no interference}}$

The capacity is the same as if the interference did not exist. Applied sequentially for all $K$ users, DPC achieves the boundary of the BC capacity region.

DPC is capacity-achieving but non-constructive: it requires encoding over long block lengths with lattice-based or trellis-based codes. Its practical complexity is prohibitive for real-time systems, motivating the linear precoding and iterative methods studied in Sections 17.3 and 17.4. Nevertheless, DPC serves as the gold-standard benchmark against which all practical multi-user MIMO schemes are compared.

,

BC Capacity Region with DPC

Visualise the broadcast channel capacity region achieved by DPC for a 2-user MISO system. Compare the DPC region with the time-division (TDMA) and zero-forcing inner bounds. Adjust the number of BS antennas and total power to observe how the gap between DPC and linear schemes changes.

Parameters

N_t

(BS antennas)4

Total power

P

(dB)15

Example: 2-User MISO BC Rate Region

A base station with $N_t = 2$ antennas serves two users with channels $\mathbf{h}_1 = [1, 0]^T$ and $\mathbf{h}_2 = [0, 1]^T$ . Total power is $P = 20$ (linear), $\sigma^2 = 1$ .

(a) Find the maximum sum rate with DPC. (b) Find the maximum sum rate with ZF precoding. (c) Explain why they are equal in this case.

Solution

DPC sum rate

The channels are orthogonal. With DPC, optimal power allocation splits power across users. Since $|\mathbf{h}_1^H\mathbf{w}_1|^2 = 1$ and $|\mathbf{h}_2^H\mathbf{w}_2|^2 = 1$ with the natural beamforming directions $\mathbf{w}_1 = [1,0]^T$ , $\mathbf{w}_2 = [0,1]^T$ :

$R_{\text{sum}}^{\text{DPC}} = \max_{P_1+P_2=20} \left[\log_2(1+P_1) + \log_2(1+P_2)\right]$

By the concavity of $\log$ , equal power $P_1 = P_2 = 10$ is optimal:

$R_{\text{sum}}^{\text{DPC}} = 2\log_2(11) \approx 6.92\;\text{bits/s/Hz}$

ZF sum rate

With orthogonal channels, ZF precoding uses the same beamforming directions with no power penalty:

$\mathbf{W}_{\text{ZF}} = \mathbf{H}^{H}(\mathbf{H}\mathbf{H}^{H})^{-1} = \begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix}$

The ZF precoder has unit-norm columns, so $R_{\text{sum}}^{\text{ZF}} = 2\log_2(1+10) = 2\log_2(11) \approx 6.92$ .

Equality explanation

When channels are orthogonal, there is no inter-user interference to cancel. ZF precoding eliminates all interference at no cost (the pseudo-inverse is the identity), so it matches DPC. For non-orthogonal channels, the ZF power penalty causes a gap to DPC. $\blacksquare$

Quick Check

Why is dirty paper coding impractical for real-time wireless systems despite being capacity-achieving?

DPC requires perfect CSI at both transmitter and receiver

DPC encoding complexity grows exponentially with the number of users and requires non-linear operations over long block lengths

DPC only works for 2-user systems

DPC requires full-duplex operation at the base station

Correction:

DPC encoding complexity grows exponentially with the number of users and requires non-linear operations over long block lengths

DPC is fundamentally a non-linear encoding technique based on lattice codes or binning schemes. The encoding must operate over very long block lengths to approach capacity, and the complexity of finding the optimal lattice point grows with dimensionality, making real-time implementation infeasible for practical systems.

Common Mistake: Assuming ZF Achieves the BC Capacity

Mistake:

Believing that zero-forcing precoding achieves the broadcast channel capacity because it eliminates all inter-user interference.

Correction:

While ZF eliminates inter-user interference, it does so at a power penalty: the ZF precoder $\mathbf{W} = \mathbf{H}^{H} (\mathbf{H}\mathbf{H}^{H})^{-1}$ amplifies the transmit power by the factor $\text{tr}((\mathbf{H}\mathbf{H}^{H})^{-1})$ , which can be large for ill-conditioned channels. DPC avoids this penalty by non-linearly pre-cancelling interference without power amplification. The gap between ZF and DPC is significant at low-to-moderate SNR and when $K$ approaches $N_t$ (the system is heavily loaded). At high SNR with $K < N_t$ , the gap vanishes in the DoF sense (both achieve the same multiplexing gain), but a constant offset remains.

🎓CommIT Contribution(2003)

Broadcast Channel Capacity via Dirty Paper Coding

G. Caire, S. Shamai (Shitz) — IEEE Transactions on Information Theory

Caire and Shamai independently (alongside Vishwanath, Jindal, and Goldsmith) established that dirty paper coding achieves the capacity region of the MIMO Gaussian broadcast channel. Their approach connected the multi-antenna BC to the degraded broadcast channel through a Sato-type argument, providing a constructive encoding scheme and proving that the DPC rate region is an outer bound. This result — one of the most important in multi-user information theory — resolved a long-standing open problem and established DPC as the theoretical benchmark for all downlink multi-user MIMO schemes. The complete converse was later established by Weingarten, Steinberg, and Shamai (2006) using an enhanced entropy power inequality.

broadcast-channeldirty-paper-codingmimo-capacityView Paper →

Key Takeaway

The BC-MAC duality theorem is the single most powerful tool in multi-user MIMO theory: it converts the hard encoding problem (broadcast channel) into the well-understood decoding problem (multiple-access channel) by swapping transmitter and receiver roles under the same sum power constraint. DPC at the BC transmitter mirrors SIC at the MAC receiver — both "peel off" users one at a time.

Broadcast Channel (BC)

A multi-user channel where a single transmitter sends independent messages to multiple receivers. The MISO BC models the cellular downlink where a multi-antenna base station serves single-antenna users.

Dirty Paper Coding (DPC)

A non-linear encoding technique that pre-cancels known interference at the transmitter with no rate penalty. Based on Costa's theorem (1983), DPC achieves the capacity region of the MIMO broadcast channel.

BC-MAC Duality

The fundamental result that the broadcast channel capacity region under a sum power constraint equals the union of dual multiple-access channel capacity regions over all power splits satisfying the same sum power constraint.

Broadcast Channel (BC)

Downlink Multi-User MIMO: The Dual Problem

Definition: MISO Broadcast Channel

Theorem: BC-MAC Duality

Achievability via DPC

Converse via dual MAC

Rate region equivalence

Definition: Dirty Paper Coding (DPC) for the Broadcast Channel

BC Capacity Region with DPC

Parameters

Example: 2-User MISO BC Rate Region

DPC sum rate

ZF sum rate

Equality explanation

Quick Check

Common Mistake: Assuming ZF Achieves the BC Capacity

Broadcast Channel Capacity via Dirty Paper Coding

Key Takeaway

Broadcast Channel (BC)

Dirty Paper Coding (DPC)

BC-MAC Duality

Definition:
MISO Broadcast Channel

Definition:
Dirty Paper Coding (DPC) for the Broadcast Channel