Ferkans — Interactive Telecom Tutor

How Far Is Linear Precoding from Optimal?

The MIMO broadcast channel (BC) capacity — the theoretical limit — is achieved by dirty-paper coding (DPC), a nonlinear technique that pre-cancels interference at the transmitter. Linear precoding is suboptimal in general. The natural question is: how large is the gap? The answer turns out to be surprisingly small in many regimes, which is precisely why linear precoding dominates practice. This section quantifies the gap and explains the underlying theory.

Definition:
MIMO Broadcast Channel

The MIMO broadcast channel (BC) consists of a transmitter with $N_t$ antennas sending independent messages to $K$ single-antenna receivers. User $k$ observes

$y_k = \mathbf{h}_k^H \mathbf{x} + w_k, \qquad w_k \sim \mathcal{CN}(0, \sigma^2)$

subject to $\mathbb{E}[\|\mathbf{x}\|^2] \leq P_t$ . The capacity region $\mathcal{C}_{\text{BC}}$ is the set of all achievable rate tuples $(R_1, \ldots, R_{K})$ .

Definition:
Dirty-Paper Coding (DPC)

Dirty-paper coding exploits the fact that the transmitter knows all users' messages and can therefore pre-cancel interference. If user $k$ is encoded after users $1, \ldots, k-1$ (in some encoding order $\pi$ ), DPC encodes user $k$ 's signal as if the interference from users $1, \ldots, k-1$ were known non-causally at the encoder.

By Costa's theorem (1983), this pre-cancellation incurs no rate loss: the capacity of a channel with known interference is the same as without interference. Applying DPC with optimal encoding order and power allocation achieves every point on the boundary of $\mathcal{C}_{\text{BC}}$ .

DPC is the transmit-side dual of successive interference cancellation (SIC) in the MAC. Just as SIC achieves the MAC capacity by decoding and subtracting users one by one at the receiver, DPC achieves the BC capacity by encoding and pre-subtracting users one by one at the transmitter.

,

Theorem: MIMO BC Capacity Region

The capacity region of the MIMO BC with per-user channels $\mathbf{h}_1, \ldots, \mathbf{h}_{K}$ and sum power constraint $P_t$ is

$\mathcal{C}_{\text{BC}} = \bigcup_{\substack{\mathbf{S}_k \succeq 0 \\ \sum_k \text{tr}(\mathbf{S}_k) \leq P_t}} \left\{ (R_1, \ldots, R_{K}) : R_k \leq \log_2\!\left(1 + \mathbf{h}_k^H \mathbf{S}_k \left(\sum_{j > k} \mathbf{h}_k \mathbf{h}_k^H \mathbf{S}_j + \sigma^2\right)^{-1} \mathbf{h}_k\right) \right\}$

for some encoding order $\pi$ , where $\mathbf{S}_k \succeq 0$ are the per-user transmit covariance matrices.

The sum capacity is

$C_{\text{sum}}^{\text{BC}} = \max_{\substack{\mathbf{S}_k \succeq 0 \\ \sum_k \text{tr}(\mathbf{S}_k) \leq P_t}} \sum_{k=1}^{K} R_k^{\text{DPC}}.$

DPC turns the BC into a set of interference-free channels by successively encoding each user's signal knowing all previously encoded interference. The achievable rate for each user depends on its channel, its allocated covariance $\mathbf{S}_k$ , and the residual interference from users encoded later (which DPC cannot cancel). Optimising over the encoding order and power allocation yields the full capacity region.

Proof

MAC-BC duality

The key insight is the duality between the MAC and BC capacity regions. The MIMO MAC with the same channel matrix (transposed) has a known capacity region achieved by SIC. Vishwanath, Jindal, and Goldsmith (2003) showed that the BC capacity region equals the union over all noise covariances of the dual MAC capacity region.

DPC achievability

Costa's dirty-paper coding result (1983) shows that each user encoded in the DPC order achieves the same rate as if the previously encoded users' interference were absent. Combined with the MAC-BC duality, this yields the stated capacity region.

Converse (Weingarten--Steinberg--Shamai)

Weingarten, Steinberg, and Shamai (2006) proved the converse: no scheme can achieve rates outside this region. Their proof uses an extremal entropy inequality to show that Gaussian inputs are optimal. $\blacksquare$

,

Definition:
MAC-BC Duality

The MAC-BC duality states that the capacity region of the MIMO BC with channel matrix $\mathbf{H}$ and sum power $P_t$ equals the capacity region of the dual MIMO MAC with channel matrix $\mathbf{H}^{H}$ and the same sum power constraint:

$\mathcal{C}_{\text{BC}}(\mathbf{H}, P_t) = \mathcal{C}_{\text{MAC}}(\mathbf{H}^{H}, P_t).$

This means every point achievable by DPC in the BC is also achievable by SIC in the dual MAC, and vice versa.

The duality is computational as well as theoretical: to find the DPC power allocation for the BC, one can solve the (simpler) MAC power allocation problem and transform the solution.

Theorem: Gap Between Linear Precoding and DPC

For i.i.d. Rayleigh fading with $K$ users and $N_t$ antennas, the sum-rate gap between DPC and RZF precoding (with optimal $\alpha$ ) satisfies:

High SNR, fixed $K$ : The gap is bounded by $K \log_2(e) \approx 1.44K$ bits/s/Hz, independent of $P_t$ and $N_t$ .
Massive MIMO ( $N_t \to \infty$ , $K$ fixed): The gap vanishes: $C_{\text{sum}}^{\text{DPC}} - R_{\text{sum}}^{\text{RZF}} \to 0$ .
High loading ( $K/N_t \to 1$ ): The gap grows logarithmically in $P_t/\sigma^2$ .

Linear precoding "wastes" degrees of freedom on interference management, while DPC can pre-cancel interference for free. When $N_t \gg K$ , there are ample degrees of freedom and the waste is negligible. When $K$ approaches $N_t$ , the degrees of freedom are scarce and DPC's ability to cancel interference without losing any becomes increasingly valuable.

Proof

Upper bound via single-user channels

DPC achieves at most $C_{\text{sum}} \leq \sum_k \log_2(1 + P_t\|\mathbf{h}_k\|^2/(K\sigma^2))$ (each user gets an interference-free channel with its share of power). RZF achieves $R_k \geq \log_2(1 + (N_t - K)P_t/(K\sigma^2))$ on average. The difference per user is bounded by $\log_2(N_t/(N_t - K))$ , which vanishes as $N_t/K \to \infty$ .

Exact large-system result

Using the deterministic equivalent from random matrix theory, the per-user rate gap is $(1 + o(1))\log_2(e)$ bits/s/Hz at high SNR, giving a total gap of $K\log_2(e)$ for the sum rate. $\blacksquare$

Linear Precoding Gap to DPC Capacity

Compare the sum rate of MRT, ZF, RZF, and the DPC sum capacity as a function of SNR. Observe that RZF closely tracks DPC at moderate loading, while the gap widens as $K/N_t$ increases.

Parameters

N_t

32

K

8

Example: Quantifying the DPC Gap

For $N_t = 64$ , $K = 8$ , $P_t/\sigma^2 = 10$ dB, compute the sum rate for MRT, ZF, RZF, and DPC. Express the gaps in bits/s/Hz and as a percentage of the DPC capacity.

Solution

DPC sum capacity

Using waterfilling over the effective channels: $C_{\text{sum}}^{\text{DPC}} \approx 36.8$ bits/s/Hz (Monte Carlo average).

Linear precoding rates

$R_{\text{sum}}^{\text{MRT}} \approx 31.2$ , $R_{\text{sum}}^{\text{ZF}} \approx 34.6$ , $R_{\text{sum}}^{\text{RZF}} \approx 35.4$ bits/s/Hz.

Gaps

Scheme	Sum Rate	Gap to DPC	Gap (%)
MRT	31.2	5.6	15.2%
ZF	34.6	2.2	6.0%
RZF	35.4	1.4	3.8%

At $N_t/K = 8$ , RZF captures over 96% of the DPC capacity. The gap would narrow further with more antennas.

🎓CommIT Contribution(2018)

Massive MIMO Has Unlimited Capacity

G. Caire — IEEE Transactions on Information Theory

Caire (2018) proved a landmark result: in the massive MIMO regime with spatial correlation, the capacity grows without bound even with pilot contamination. The key insight is that users with sufficiently different spatial covariance matrices can be separated by exploiting the eigenstructure of $\mathbf{H}\mathbf{H}^{H}$ . This result, combined with the small gap between linear precoding and DPC at large $N_t/K$ , implies that simple linear precoders can approach the unlimited capacity in practice — a foundational justification for massive MIMO deployment.

massive-mimocapacitypilot-contaminationspatial-correlationView Paper →

Historical Note: Costa's Dirty-Paper Coding

1983--2006

In 1983, Max Costa published a remarkable information-theoretic result: if the transmitter knows the interference $S$ non-causally, the capacity of the channel $Y = X + S + Z$ is the same as without interference, $C = \frac{1}{2}\log_2(1 + P/\sigma^2)$ . The name "dirty-paper coding" comes from the analogy: writing on paper with dirt stains, if you know where the stains are, you can write around them without losing any information. Twenty years later, this theoretical curiosity became the key to the MIMO BC capacity, when Caire and Shamai (2003) and Vishwanath, Jindal, and Goldsmith (2003) showed that DPC achieves the capacity region.

Historical Note: The MIMO Broadcast Channel Capacity Proof

2003--2006

The MIMO BC capacity was one of the great open problems of multi-user information theory. While DPC achievability was established by 2003, the matching converse remained open until 2006, when Weingarten, Steinberg, and Shamai proved that DPC with Gaussian codebooks is optimal. Their proof used a new "enhanced channel" technique and an extremal inequality for entropy. The result settled a decade-long debate about whether nonlinear schemes beyond DPC might be needed.

Quick Check

Why is DPC not used in practical wireless systems despite achieving the BC capacity?

It requires exponentially complex encoding

It only works with perfect CSI

It achieves lower rates than linear precoding

It violates per-antenna power constraints

Correction:

It requires exponentially complex encoding

DPC requires structured binning over high-dimensional lattices or random codes. The encoding complexity grows exponentially with the number of users and the block length. No practical DPC implementation exists that approaches the theoretical gains at reasonable complexity.

Why This Matters: Linear Precoding in Practice — The Pragmatic Choice

The small gap between RZF and DPC ( $< 5\%$ at $N_t/K \geq 4$ ) explains why every 4G/5G base station uses linear precoding rather than DPC. The massive MIMO philosophy reinforces this: by deploying many antennas ( $N_t \gg K$ ), the gap becomes negligible, and the system designer gets near-optimal performance with a simple matrix multiplication. This is the fundamental engineering insight of massive MIMO.

Dirty-Paper Coding (DPC)

A nonlinear encoding technique where the transmitter pre-cancels known interference by encoding each user's signal as if the interference from previously encoded users were absent. Achieves the MIMO BC capacity region but is computationally intractable in practice.

MAC-BC Duality

The property that the capacity region of the MIMO broadcast channel equals the capacity region of the dual MIMO multiple-access channel with the transposed channel matrix and the same sum power constraint.

Key Takeaway

DPC achieves the MIMO BC capacity but is impractical. Linear precoding (RZF) captures over 95% of the DPC capacity when $N_t/K \geq 4$ , and the gap vanishes in the massive MIMO regime. This small gap is the fundamental justification for using linear precoding in all modern cellular systems: the complexity savings are enormous while the performance loss is negligible.

Connection to DPC and BC Capacity