Ferkans — Interactive Telecom Tutor

When the Problem Becomes Tractable

The general interference channel capacity is unknown, but there are important special cases where we can determine the capacity region exactly. The key insight is that when the interference is strong enough, each receiver can decode the interfering message completely — and once decoded, the interference can be subtracted, leaving a clean channel. This transforms the IC into a pair of effective MACs, and we know how to handle MACs.

We start with the extreme case (very strong interference) and work our way down to strong interference, building intuition for why the problem becomes harder as interference weakens.

Definition:
Very Strong Interference

The two-user Gaussian IC is said to have very strong interference if: $a^2 P_2 \geq P_1 + \sigma^2 \quad \text{and} \quad b^2 P_1 \geq P_2 + \sigma^2$

Equivalently, $\text{INR}_{12} \geq \text{SNR}_{1} + 1$ and $\text{INR}_{21} \geq \text{SNR}_{2} + 1$ .

In this regime, the interference is so strong that each receiver gets a better copy of the interfering message than the intended receiver does of its own.

The terminology "very strong" distinguishes this from the "strong" interference regime, which has a weaker condition and a different capacity argument.

Theorem: Capacity Region Under Very Strong Interference

Under very strong interference, the capacity region of the two-user Gaussian IC is: $R_{1} \leq \frac{1}{2}\log(1 + \text{SNR}_{1})$ $R_{2} \leq \frac{1}{2}\log(1 + \text{SNR}_{2})$

That is, both users achieve their interference-free capacity simultaneously — as if the other user were not present.

When interference is very strong, Rx 1 can decode Tx 2's message from the interference signal $aX_2$ (which it receives with SNR $a^2 P_2 / \sigma^2 \geq \text{SNR}_{1} + 1$ ), subtract it perfectly, and then decode its own message from the clean residual $Y_1 - a\hat{X}_2 = X_1 + Z_1$ with rate $\frac{1}{2}\log(1 + \text{SNR}_{1})$ . Symmetrically for Rx 2. The interference is so loud that it is easier to understand than to ignore.

Proof

Achievability

Each transmitter uses an independent Gaussian codebook with full power. Each receiver first decodes the interfering message (achievable because $I(X_2; Y_1) = \frac{1}{2}\log(1 + a^2 P_2 / (P_1 + \sigma^2)) \geq R_{2}$ when $a^2 P_2 \geq P_1 + \sigma^2$ ), subtracts it, then decodes its own message.

Converse

The single-user capacity $\frac{1}{2}\log(1 + P_k / \sigma^2)$ is an obvious upper bound (the rate cannot exceed what a user would achieve with no interference). Since the achievability matches this bound, the capacity region is established.

Definition:
Strong Interference

The two-user Gaussian IC has strong interference if: $a^2 \geq 1 \quad \text{and} \quad b^2 \geq 1$

Equivalently, $\text{INR}_{12} \geq \text{SNR}_{1} \cdot (P_2/P_1)$ and similarly for the other link. In the symmetric case ( $P_1 = P_2$ , $a = b$ ), this simplifies to $|a| \geq 1$ , i.e., the cross-link is at least as strong as the direct link.

Theorem: Capacity Region Under Strong Interference

Under strong interference ( $a^2 \geq 1$ , $b^2 \geq 1$ ), the capacity region of the two-user Gaussian IC equals the intersection of two MAC regions: $R_{1} \leq \frac{1}{2}\log(1 + P_1/\sigma^2)$ $R_{2} \leq \frac{1}{2}\log(1 + P_2/\sigma^2)$ $R_{1} + R_{2} \leq \frac{1}{2}\log(1 + P_1/\sigma^2 + a^2 P_2/\sigma^2)$ $R_{1} + R_{2} \leq \frac{1}{2}\log(1 + b^2 P_1/\sigma^2 + P_2/\sigma^2)$

Each receiver decodes both messages jointly, treating its observation as a MAC.

When both cross-links are stronger than the direct links, each receiver can decode the interfering message alongside its own. The sum-rate constraints arise because each receiver effectively sees a MAC with both signals. The capacity region is the intersection of the two MAC regions (one at each receiver).

Proof

Achievability via joint decoding

Each receiver uses joint typicality decoding for both messages. At Rx 1, the effective MAC has inputs $(X_1, X_2)$ and output $Y_1 = X_1 + aX_2 + Z_1$ . The MAC rate constraints give: $R_{1} < I(X_1; Y_1 | X_2)$ , $R_{2} < I(X_2; Y_1 | X_1)$ , $R_{1} + R_{2} < I(X_1, X_2; Y_1)$ . Similarly at Rx 2. The achievable region is the intersection.

Converse

The converse uses Fano's inequality at each receiver. The key step is showing that under strong interference ( $a^2 \geq 1$ ), the constraint at Rx 1 is at least as restrictive as any genie-aided bound: $nR_{2} \leq I(X_2^n; Y_1^n | X_1^n) + n\epsilon_n \leq n \cdot \frac{1}{2}\log(1 + a^2 P_2/\sigma^2) + n\epsilon_n$ The strong interference condition ensures Rx 1 can decode $M_2$ reliably.

,

Example: Symmetric Strong Interference Channel

Consider the symmetric Gaussian IC with $P_1 = P_2 = P = 10$ , $\sigma^2 = 1$ , and $a = b = 2$ (strong interference: $a^2 = 4 > 1$ ). Find the capacity region and the maximum sum rate.

Solution

Individual rate bounds

$R_{k} \leq \frac{1}{2}\log(1 + 10) = \frac{1}{2}\log(11) \approx 1.73 \text{ bits}$ $

Sum-rate bounds

From Rx 1's MAC: $R_{1} + R_{2} \leq \frac{1}{2}\log(1 + 10 + 4 \cdot 10) = \frac{1}{2}\log(51) \approx 2.84$ . By symmetry ( $a = b$ ), the constraint at Rx 2 is identical.

Capacity region

The capacity region is the rectangle $R_{k} \leq 1.73$ intersected with the sum-rate constraint $R_{1} + R_{2} \leq 2.84$ . Since $2 \times 1.73 = 3.46 > 2.84$ , the sum-rate constraint is active. Maximum sum rate: $C_{\text{sum}} \approx 2.84$ bits.

Comparison with no interference

Without interference, each user gets $\frac{1}{2}\log(11) \approx 1.73$ bits, for a sum of 3.46. The interference costs about 0.62 bits in sum rate, even though each receiver can decode both messages.

Historical Note: The Strong Interference Capacity Result

1975-1987

The capacity region of the IC under strong interference was established by Sato (1981) and Costa and El Gamal (1987), building on earlier work by Carleial (1975) who showed that very strong interference allows interference-free performance. The strong interference result is one of the few complete capacity characterizations for the IC, and it relies on the elegant insight that decoding interference is beneficial when the cross-link is strong — a counterintuitive strategy that turns a nuisance into useful information.

Definition:
Weak Interference

The two-user Gaussian IC has weak interference if: $a^2 < 1 \quad \text{and} \quad b^2 < 1$

In this regime, the cross-links are weaker than the direct links. Decoding the interfering message is generally not beneficial because the interference is received too weakly. The natural strategy is to treat interference as noise (TIN): each receiver ignores the structure of the interference and treats it as additional Gaussian noise.

Theorem: Treating Interference as Noise (TIN) Rate Region

When both users treat interference as noise, the achievable rate region is: $R_{1} \leq \frac{1}{2}\log\left(1 + \frac{P_1}{a^2 P_2 + \sigma^2}\right)$ $R_{2} \leq \frac{1}{2}\log\left(1 + \frac{P_2}{b^2 P_1 + \sigma^2}\right)$

This is a rectangle — no sum-rate constraint beyond the individual bounds.

Each receiver sees its desired signal plus interference-plus-noise with effective variance $a^2 P_2 + \sigma^2$ (or $b^2 P_1 + \sigma^2$ ). The TIN rate is simply the AWGN capacity with this effective noise. No joint decoding, no rate splitting — just standard single-user decoding with degraded noise.

Proof

Achievability

User $k$ generates a Gaussian codebook with power $P_k$ . Receiver 1 treats $aX_2 + Z_1$ as $\mathcal{N}(0, a^2 P_2 + \sigma^2)$ noise (worst case by the entropy maximization property of Gaussians). The achievable rate is: $R_{1} = \frac{1}{2}\log\left(1 + \frac{P_1}{a^2 P_2 + \sigma^2}\right) = \frac{1}{2}\log\left(\frac{1 + \text{SNR}_{1} + \text{INR}_{12}}{1 + \text{INR}_{12}}\right)$

Remark on optimality

TIN is generally not capacity-achieving. It ignores the structure of the interference (which is a codeword, not Gaussian noise). However, we will see in Section 17.5 that TIN is sum-rate optimal under certain conditions on the SNR and INR values.

IC Capacity Region by Interference Regime

Compare the achievable rate regions under different strategies (TIN, joint decoding, Han-Kobayashi) for the symmetric two-user Gaussian IC as the interference strength varies.

Parameters

SNR (dB)15

a^2

(interference strength)0.5

Cross-channel gain squared: $a^2 < 1$ is weak, $a^2 > 1$ is strong

Common Mistake: Interference Is Not Gaussian

Mistake:

Assuming that treating interference as noise is information-theoretically optimal because the interference is "approximately Gaussian" by the central limit theorem.

Correction:

While the interference is the sum of many codeword symbols, it has structure (it is a codeword from a known codebook, not i.i.d. Gaussian). Strategies that exploit this structure — such as decoding the interference (strong regime) or rate-splitting (Han-Kobayashi) — can strictly outperform TIN. The Gaussian assumption in TIN is a lower bound on the interference's harmfulness, not an exact characterization.

Quick Check

In the strong interference regime ( $a^2 \geq 1$ , $b^2 \geq 1$ ), each receiver decodes both messages. Why does this help (rather than waste decoding resources)?

Decoding the interfering message allows it to be subtracted, leaving a cleaner signal for the desired message

The receiver gets two independent data streams, doubling the rate

The strong interference condition means the desired signal is also strong

Correction:

Decoding the interfering message allows it to be subtracted, leaving a cleaner signal for the desired message

Once the interfering message is decoded, it can be perfectly subtracted from the received signal. The strong interference condition ensures this decoding is reliable. The net effect is as if the interference were not present for the desired message.

Key Takeaway

The interference channel has distinct capacity behaviors depending on the interference strength. Under very strong interference, both users achieve interference-free rates. Under strong interference, joint decoding at each receiver is optimal and the capacity region is the intersection of two MAC regions. Under weak interference, the capacity is unknown in general, though treating interference as noise provides a simple achievable rate — the question is how much we lose by ignoring interference structure.

Capacity in Special Cases

When the Problem Becomes Tractable

Definition: Very Strong Interference

Theorem: Capacity Region Under Very Strong Interference

Achievability

Converse

Definition: Strong Interference

Theorem: Capacity Region Under Strong Interference

Achievability via joint decoding

Converse

Example: Symmetric Strong Interference Channel

Individual rate bounds

Sum-rate bounds

Capacity region

Comparison with no interference

Historical Note: The Strong Interference Capacity Result

Definition: Weak Interference

Theorem: Treating Interference as Noise (TIN) Rate Region

Achievability

Remark on optimality

IC Capacity Region by Interference Regime

Parameters

Common Mistake: Interference Is Not Gaussian

Quick Check

Key Takeaway

Definition:
Very Strong Interference

Definition:
Strong Interference

Definition:
Weak Interference