Ferkans — Interactive Telecom Tutor

From Discrete to Gaussian MIMO

We now turn to the broadcast channel that matters most for modern wireless systems: the Gaussian MIMO broadcast channel. This is the downlink of a multi-antenna base station serving multiple single- or multi-antenna users.

The MIMO BC is not degraded in general — two users with different spatial channel matrices $\mathbf{H}_{1}$ and $\mathbf{H}_{2}$ cannot be ordered by a degradation relation unless their channels happen to have a special structure. This means superposition coding alone is insufficient. The remarkable result, due to Weingarten, Steinberg, and Shamai (2006), is that dirty-paper coding (DPC) achieves the entire capacity region of the Gaussian MIMO BC.

The point is that DPC transforms the MIMO BC into an effective degraded channel by pre-canceling the interference caused by each user's signal to the subsequently encoded users — exactly Costa's result from Chapter 12, applied sequentially.

Definition:
The $K$ -User Gaussian MIMO Broadcast Channel

The transmitter has $n_t$ antennas and serves $K$ receivers, each with $n_{r,k}$ antennas. The channel to user $k$ is: $\mathbf{y}_k = \mathbf{H}_{k} \mathbf{x} + \mathbf{z}_k, \quad k = 1, \ldots, K$ where:

$\mathbf{x} \in \mathbb{C}^{n_t}$ is the transmitted signal with power constraint $\mathbb{E}[\|\mathbf{x}\|^2] \leq P$
$\mathbf{H}_{k} \in \mathbb{C}^{n_{r,k} \times n_t}$ is the channel matrix to user $k$
$\mathbf{z}_k \sim \mathcal{CN}(\mathbf{0}, \sigma^2_{k} \mathbf{I})$ is i.i.d. Gaussian noise

The encoder sends independent messages $M_1, \ldots, M_K$ at rates $R_{1}, \ldots, R_{K}$ , and each receiver $k$ decodes only $M_k$ .

MIMO broadcast channel

The downlink of a multi-antenna transmitter serving multiple receivers simultaneously. Each receiver sees the transmitted signal through its own spatial channel matrix, making the channel non-degraded in general.

Definition:
DPC Encoding for the MIMO BC

Fix an encoding order $\pi = (\pi(1), \ldots, \pi(K))$ , a permutation of $\{1, \ldots, K\}$ . The DPC encoder processes users in the order $\pi(1), \pi(2), \ldots, \pi(K)$ :

User $\pi(1)$ : Encode $M_{\pi(1)}$ using a Gaussian codebook with covariance $\mathbf{K}_{\pi(1)}$ , producing $\mathbf{x}_{\pi(1)}$ .
User $\pi(k)$ , $k \geq 2$ : The signals $\mathbf{x}_{\pi(1)}, \ldots, \mathbf{x}_{\pi(k-1)}$ are known non-causally at the encoder (since the encoder generated them). Treat $\mathbf{s}_k = \sum_{j=1}^{k-1} \mathbf{H}_{\pi(k)} \mathbf{x}_{\pi(j)}$ as "interference known at the encoder" and apply Costa's dirty-paper coding to pre-cancel it. Encode $M_{\pi(k)}$ with covariance $\mathbf{K}_{\pi(k)}$ .
Transmit: $\mathbf{x} = \sum_{k=1}^K \mathbf{x}_{\pi(k)}$ with $\sum_{k=1}^K \text{tr}(\mathbf{K}_{\pi(k)}) \leq P$ .

Theorem: DPC Achievable Rate Region

For a fixed encoding order $\pi$ and covariance matrices $\mathbf{K}_1, \ldots, \mathbf{K}_K$ with $\sum_k \text{tr}(\mathbf{K}_k) \leq P$ , DPC achieves the rate tuple: $R_{\pi(k)} = \log \det \left( \mathbf{I} + \mathbf{H}_{\pi(k)} \left( \sum_{j=k}^K \mathbf{K}_{\pi(j)} \right) \mathbf{H}_{\pi(k)}^{H} \left( \sigma^2_{\pi(k)} \mathbf{I} + \mathbf{H}_{\pi(k)} \left( \sum_{j=k+1}^K \mathbf{K}_{\pi(j)} \right) \mathbf{H}_{\pi(k)}^{H} \right)^{-1} \right)$

The full DPC region $\mathcal{R}_{\text{DPC}}$ is the convex hull over all encoding orders $\pi$ and all valid covariance allocations $\{\mathbf{K}_k\}$ .

Each user $\pi(k)$ sees the signals of users $\pi(k+1), \ldots, \pi(K)$ as interference (since those are encoded later and their interference is not pre-canceled), but the signals of $\pi(1), \ldots, \pi(k-1)$ are pre-canceled via DPC. The rate expression is a generalized $\log\det(\mathbf{I} + \text{SNR})$ formula with the effective noise including only the non-canceled interference.

Proof

Costa's theorem per user

For user $\pi(k)$ , the interference $\sum_{j=1}^{k-1} \mathbf{H}_{\pi(k)} \mathbf{x}_{\pi(j)}$ is known non-causally at the encoder. By Costa's theorem (Chapter 12), this interference can be pre-canceled at no rate cost, as if it were not present. The effective channel for user $\pi(k)$ is therefore: $\mathbf{y}_{\pi(k)}^{\text{eff}} = \mathbf{H}_{\pi(k)} \mathbf{x}_{\pi(k)} + \underbrace{\sum_{j=k+1}^K \mathbf{H}_{\pi(k)} \mathbf{x}_{\pi(j)}}_{\text{uncanceled interference}} + \mathbf{z}_{\pi(k)}$

Gaussian codebook achieves the rate

With Gaussian $\mathbf{x}_{\pi(k)} \sim \mathcal{CN}(\mathbf{0}, \mathbf{K}_{\pi(k)})$ , the achievable rate is the mutual information of a Gaussian MIMO channel with input covariance $\mathbf{K}_{\pi(k)}$ and noise-plus-interference covariance $\sigma^2_{\pi(k)} \mathbf{I} + \mathbf{H}_{\pi(k)} (\sum_{j>k} \mathbf{K}_{\pi(j)}) \mathbf{H}_{\pi(k)}^{H}$ . The standard $\log\det$ formula applies.

Power constraint

Since $\mathbf{x} = \sum_k \mathbf{x}_{\pi(k)}$ with independent components, $\mathbb{E}[\|\mathbf{x}\|^2] = \sum_k \text{tr}(\mathbf{K}_{\pi(k)}) \leq P$ .

,

Theorem: DPC Achieves the MIMO BC Capacity Region

The capacity region of the $K$ -user Gaussian MIMO broadcast channel equals the DPC region: $C_{\text{BC}} = \mathcal{R}_{\text{DPC}}$

That is, dirty-paper coding is optimal — no other coding scheme can achieve rates outside the DPC region.

The converse is the hard part. The proof, due to Weingarten, Steinberg, and Shamai (2006), uses a channel enhancement argument: enhance the noise at each receiver to create a degraded BC that has the same capacity region as the original, then apply the known converse for degraded BCs. The enhancement is possible because the Gaussian distribution is the worst-case noise (entropy-power inequality argument).

Proof

Achievability (DPC)

Follows from the DPC construction above and Costa's theorem. Each user sees an effective point-to-point channel with Gaussian codebook achieving the $\log\det$ rate.

Converse via channel enhancement

The key idea: for each user $k$ , enhance (reduce) the noise covariance from $\sigma^2_{k} \mathbf{I}$ to some $\mathbf{N}_k \preceq \sigma^2_{k} \mathbf{I}$ such that the resulting channel becomes physically degraded: $\mathbf{N}_1 \preceq \mathbf{N}_2 \preceq \cdots \preceq \mathbf{N}_K$ (after appropriate basis transformations). This enhancement can only increase the capacity region, so any outer bound for the enhanced (degraded) channel is also an outer bound for the original channel.

Degraded channel converse

For the enhanced degraded MIMO BC, superposition coding is optimal (Bergmans' converse extended to the MIMO case). The optimal input distribution is Gaussian (entropy maximization under covariance constraint). The resulting capacity region matches the DPC region of the original channel.

Enhancement preserves capacity

The crucial step: the enhancement is chosen so that the capacity region of the enhanced channel equals that of the original channel. This requires showing that the optimal covariance matrices $\{\mathbf{K}_k\}$ for the original channel remain optimal under enhancement — which follows from the KKT conditions of the optimization.

Historical Note: The Resolution of the MIMO BC Capacity

2003-2006

The MIMO BC capacity problem was one of the most actively pursued questions in information theory during 2000-2006. Caire and Shamai (2003) showed that DPC achieves the sum capacity of the MISO BC (single-antenna receivers). Vishwanath, Jindal, and Goldsmith (2003) and Viswanath and Tse (2003) independently established DPC optimality for the sum rate via MAC-BC duality. The full capacity region (all rate tuples, not just sum rate) was finally established by Weingarten, Steinberg, and Shamai in 2006, using the elegant channel enhancement technique.

It is worth noting that the practical impact of this result extends far beyond the specific DPC scheme. The capacity region characterization justifies the design of practical MU-MIMO precoding schemes (zero-forcing, regularized ZF, MMSE precoding) that approximate DPC at lower complexity — see Book telecom, Chapter 17.

⚠️Engineering Note

DPC vs. Practical Linear Precoding

While DPC is information-theoretically optimal for the MIMO BC, its implementation complexity is prohibitive for real systems. Practical 5G NR base stations use linear precoding (zero-forcing, MMSE, or regularized ZF) instead.

The gap between DPC and zero-forcing precoding is at most $\log_2(K)$ bits per user for $K$ users, but in practice the gap is much smaller — typically 1-3 dB for the sum rate at moderate SNR. For massive MIMO systems ( $n_t \gg K$ ), the gap vanishes asymptotically because the users' channels become nearly orthogonal.

The practical message: DPC tells us the limit of what is possible; linear precoding gets us most of the way there with orders of magnitude less complexity.

Practical Constraints

•
DPC requires non-causal knowledge of all users' messages — impractical in real-time systems
•
Linear precoding requires only CSI, not message knowledge
•
Massive MIMO with MRT or ZF approaches DPC performance

🎓CommIT Contribution(2003)

DPC Sum-Capacity Optimality for the MISO BC

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

Caire and Shamai established that dirty-paper coding achieves the sum capacity of the MISO broadcast channel (single-antenna receivers). This was a key stepping stone toward the full MIMO BC capacity region, later proved by Weingarten, Steinberg, and Shamai (2006). The paper introduced the connection between DPC and Costa's writing-on-dirty-paper theorem in the multiuser MIMO context, and showed that the sum rate with DPC can far exceed that of linear precoding, especially when user channels are not orthogonal.

DPCMISO BCsum capacityView Paper →

Why This Matters: Connection to MU-MIMO Precoding

The MIMO BC capacity region established here is the theoretical foundation for all multi-user MIMO downlink techniques used in 4G LTE and 5G NR. The DPC encoding order corresponds to successive interference pre-cancellation, while practical systems approximate this with linear precoding (ZF, MMSE). The MAC-BC duality (Section 16.4) is used in practical beamforming design: it is easier to optimize the uplink and then transform the solution to the downlink.

See full treatment in The $K$-User MIMO Broadcast Channel

DPC Rate Region for the 2-User MISO BC

Visualize the DPC capacity region for a two-user MISO broadcast channel ( $n_t = 2$ , single-antenna receivers). Adjust the channel vectors and SNR to see how the capacity region changes with channel geometry.

Parameters

Channel 1 angle (degrees)30

Angle of channel vector $\mathbf{h}_1$ from the $x$-axis

Channel 2 angle (degrees)120

Angle of channel vector $\mathbf{h}_2$ from the $x$-axis

SNR (dB)15

Transmit SNR in dB

DPC Pre-Cancellation for the MIMO BC

Step-by-step animation of dirty-paper coding for the two-user MIMO BC. The encoder processes users sequentially, pre-canceling interference for later users via Costa's theorem. The last user sees an interference-free channel.

Common Mistake: Encoding Order Matters for DPC

Mistake:

Assuming that the DPC rate region is the same for all encoding orders.

Correction:

Different encoding orders favor different users. User $\pi(1)$ (encoded first, canceled last) sees the most interference and typically gets the lowest rate. User $\pi(K)$ (encoded last, all interference pre-canceled) sees a clean channel. The full DPC region is the convex hull over all $K!$ encoding orders, which allows time-sharing between orderings.

Example: DPC for the 2-User MISO BC

Consider a MISO BC with $n_t = 2$ antennas, two single-antenna users with channels $\mathbf{h}_1 = [1, 0]^T$ and $\mathbf{h}_2 = [0, 1]^T$ (orthogonal channels), and total power $P = 10$ . Noise variance $\sigma^2 = 1$ for both users. Compute the DPC sum rate and compare with zero-forcing.

Solution

Orthogonal channels simplification

With orthogonal channels, each user's signal does not interfere with the other. The optimal strategy is simply to allocate power $P_1$ to user 1 and $P_2 = P - P_1$ to user 2, with $\mathbf{K}_1 = P_1 \mathbf{e}_1 \mathbf{e}_1^H$ and $\mathbf{K}_2 = P_2 \mathbf{e}_2 \mathbf{e}_2^H$ .

DPC rates

With either encoding order: $R_{1} = \log(1 + P_1), \quad R_{2} = \log(1 + P_2)$ The DPC encoding order is irrelevant because there is no cross-interference to pre-cancel. Sum rate: $R_{1} + R_{2} = \log(1 + P_1) + \log(1 + P - P_1)$ .

Comparison with ZF

Zero-forcing with orthogonal channels is identical to DPC — each antenna serves one user with no interference. The ZF and DPC sum rates coincide: $R_{\text{sum}} = \max_{0 \leq P_1 \leq P} \left[ \log(1 + P_1) + \log(1 + P - P_1) \right]$ By concavity of $\log$ , equal power allocation $P_1 = P_2 = 5$ is optimal, giving $R_{\text{sum}} = 2\log(6) \approx 5.17$ bits/channel use.

Remark on non-orthogonal channels

When channels are not orthogonal (e.g., $\mathbf{h}_1$ and $\mathbf{h}_2$ are close in angle), DPC strictly outperforms ZF because it can pre-cancel cross-interference instead of nulling it out — ZF wastes degrees of freedom to avoid interference, while DPC uses them constructively.

Quick Check

In the DPC encoding for the MIMO BC, user $\pi(K)$ (the last user encoded) achieves the rate of a point-to-point MIMO channel. Why?

All interference from previously encoded users is pre-canceled via Costa's theorem

User $\pi(K)$ has the strongest channel

The noise at user $\pi(K)$ is zero

Correction:

All interference from previously encoded users is pre-canceled via Costa's theorem

User $\\pi(K)$ is encoded last, so the encoder knows all other codewords $\\mathbf{x}_{\\pi(1)}, \\ldots, \\mathbf{x}_{\\pi(K-1)}$ non-causally and pre-cancels their interference completely. The effective channel is interference-free.

Channel enhancement

A converse proof technique where the noise at each receiver is reduced (enhanced) to create a degraded broadcast channel whose capacity region contains that of the original channel. The enhancement is chosen so that the capacity region is preserved, allowing the use of the simpler degraded-channel converse.

The MIMO Broadcast Channel

From Discrete to Gaussian MIMO

Definition: The KKK-User Gaussian MIMO Broadcast Channel

MIMO broadcast channel

Definition: DPC Encoding for the MIMO BC

Theorem: DPC Achievable Rate Region

Costa's theorem per user

Gaussian codebook achieves the rate

Power constraint

Theorem: DPC Achieves the MIMO BC Capacity Region

Achievability (DPC)

Converse via channel enhancement

Degraded channel converse

Enhancement preserves capacity

Historical Note: The Resolution of the MIMO BC Capacity

DPC vs. Practical Linear Precoding

DPC Sum-Capacity Optimality for the MISO BC

Why This Matters: Connection to MU-MIMO Precoding

DPC Rate Region for the 2-User MISO BC

Parameters

DPC Pre-Cancellation for the MIMO BC

Common Mistake: Encoding Order Matters for DPC

Example: DPC for the 2-User MISO BC

Orthogonal channels simplification

DPC rates

Comparison with ZF

Remark on non-orthogonal channels

Quick Check

Channel enhancement

Definition:
The $K$ -User Gaussian MIMO Broadcast Channel

Definition:
DPC Encoding for the MIMO BC