How does compound capacity apply to a millimeter-wave link with phase-noise or beam-pointing uncertainty?

A 71 to 86 GHz E-band backhaul radio rarely knows its exact channel: residual local-oscillator phase noise, antenna-misalignment loss of 1 to 3 dB, and atmospheric variation each shift the effective SNR within a bounded interval rather than a known distribution. Modeling the operating point as a compound set, for example an SNR known only to lie between 12 and 18 dB, lets the designer pick a modulation and coding scheme rate that the worst point in the set still decodes, which is exactly the compound-capacity philosophy. The achievable robust rate is set by the lowest-SNR member, log2(1 + gamma_min) where gamma_min is the linear SNR at the 12 dB lower bound (about 15.8), giving roughly 4.07 bit per channel use, so adaptive coding and modulation widens the set membership as link conditions are measured. This is why fade-margin design and conservative MCS selection in fixed wireless mirror compound-channel reasoning even when the term is not used explicitly.

Digital Communications

Compound Capacity

Q: How does compound capacity differ from ergodic and outage capacity?

All three describe communication under channel uncertainty but model it differently. Ergodic capacity assumes the channel varies fast enough that a codeword sees many independent realizations, so it averages the mutual information over the fading distribution. Outage capacity assumes a single fixed realization per codeword and tolerates a small probability p that decoding fails, reporting the largest rate supportable on all but the worst p fraction of channels. Compound capacity is the strictest of the three: the channel is fixed for the whole block, it is drawn from a known set S, and zero error probability is demanded uniformly over every channel in S. Compound capacity equals the maximin mutual information, max over input distribution P of the minimum over s in S of I(X;Y|s). Because nature is allowed to pick the worst channel after the input distribution is fixed, the compound rate is generally lower than the average ergodic rate but does not depend on assuming any prior over the set.

Q: Why is the compound capacity a maximin rather than a max of an average?

In the compound setting the transmitter does not know which channel from the set S is active and is given no statistical prior over S, so it cannot average. The only meaningful guarantee is a worst-case one: a code must decode reliably no matter which member of S nature selected. The transmitter therefore optimizes a single input distribution P(x) that performs well against the least favorable channel, which is a maximin problem, max over P of min over s of I(X;Y|s). A single universal codebook with a maximum-likelihood or minimum-distance decoder achieves this rate without knowing s, which is the core result of compound channel coding theorems by Blackwell, Breiman, Thomasian, and Wolfowitz. The maximin value is upper bounded by the minimax value, and for many convex sets the two coincide by a saddle-point argument.

/KOM-pownd kuh-PASS-ih-tee/

When a channel's true law is fixed for the entire transmission but unknown to the transmitter and known only to belong to a candidate set, the largest rate that still guarantees reliable decoding is the compound capacity. It is computed as a maximin of the mutual information: the encoder commits to one input distribution, then nature is free to choose the least favorable channel in the set. Unlike ergodic capacity, no averaging over a fading prior is allowed, so the worst-case member of the set dictates the achievable throughput. A single universal codebook attains this rate without the transmitter ever learning which channel is active, which is why it underpins robust modulation-and-coding choices for links operating over an uncertain but bounded SNR window.

Category: Digital Communications

Optimization: max_P min_s I(X;Y|s)

Error criterion: Uniform over set S

The Worst-Case Channel Model Behind Compound Capacity

A compound channel formalizes the common situation where a designer knows the family of channels a link might experience but not which member is in force during any given transmission. Formally, a set S of conditional distributions W_s(y|x) is fixed and made known to both encoder and decoder, but the index s is selected once, held constant for the whole codeword, and never revealed. A rate is achievable only if the same codebook decodes with vanishing error for every s in S simultaneously. This is fundamentally stricter than fading models that permit averaging, because there is no statistical prior over s to lean on and no opportunity to interleave across many independent channel draws.

The capacity of this setup, proven by Blackwell, Breiman, and Thomasian and independently by Wolfowitz, is the maximin of the mutual information. The transmitter picks a single input distribution P(x) that performs as well as possible against the least favorable channel, and nature responds by minimizing over s. A universal decoder, such as maximum mutual information or minimum-distance decoding, achieves this rate without knowing s, so robustness costs nothing beyond the worst-case rate penalty. Because the maximin is dominated by the minimax, designers often verify whether a saddle point exists; for convex, compact sets the two values coincide and the least-favorable channel is well defined.

Relationship to SNR Uncertainty

For an additive white Gaussian noise link whose received SNR is known only to lie in an interval, the set S is parameterized by signal-to-noise ratio and the worst member is simply the lowest SNR. The compound capacity then collapses to the Shannon formula evaluated at that minimum SNR. This explains why fade-margin budgeting and conservative channel capacity targets are, in effect, compound-channel reasoning: the engineer trades peak throughput for a rate that the lowest expected operating point still supports without retransmission.

Governing Relations

Compound capacity (maximin form):
C_comp = max_P(x) min_{s ∈ S} I(X; Y | s) bits per channel use

Minimax upper bound (saddle point when sets are convex):
C_comp ≤ min_{s ∈ S} max_P(x) I(X; Y | s)

AWGN interval, SNR ∈ [γ_min, γ_max]:
C_comp = log₂(1 + γ_min)

Where I(X;Y|s) is the mutual information under channel s, P(x) is the input distribution, and γ_min is the worst-case linear SNR. Example: a set spanning 12 to 18 dB has γ_min ≈ 15.8, giving C_comp ≈ 4.07 bit per channel use, versus 6.00 at 18 dB.

Capacity Notions Under Channel Uncertainty

Capacity notion	Channel behavior per block	Knowledge of state	Error criterion	Formula core
Compound	Fixed, unknown, from set S	Set known, index unknown	Zero, uniform over S	max_P min_s I
Ergodic	Varies fast over codeword	Distribution known	Zero, averaged	E[ I ]
Outage	Fixed random realization	Distribution known	Allows p outage	max R : Pr(I < R) ≤ p
Arbitrarily varying	Adversarial, can change per symbol	Set known	Zero, worst sequence	Random-code maximin
Shannon (single channel)	One fixed known channel	Fully known	Zero	max_P I(X;Y)

Common Questions

Frequently Asked Questions

How does compound capacity differ from ergodic and outage capacity?

Ergodic capacity averages mutual information over a known fading distribution, assuming each codeword sees many independent realizations. Outage capacity fixes one realization per block and tolerates a small failure probability p. Compound capacity is strictest: the channel is fixed for the whole block, drawn from a known set S, and zero error is demanded uniformly over every member. It equals max_P min_{s ∈ S} I(X;Y|s), so it is typically below the ergodic average but needs no prior over the set.

Why is the compound capacity a maximin rather than a max of an average?

The transmitter is given no statistical prior over the set S, so averaging is meaningless; the only useful guarantee is worst-case. A single input distribution P(x) is optimized against the least favorable channel, giving max_P min_s I(X;Y|s). One universal codebook with a maximum-likelihood or minimum-distance decoder achieves this without knowing s. The maximin is bounded above by the minimax, and for convex, compact sets the two coincide through a saddle-point argument.

How does compound capacity apply to a millimeter-wave link with pointing or phase-noise uncertainty?

A 71 to 86 GHz E-band backhaul radio seldom knows its exact channel: residual LO phase noise, 1 to 3 dB antenna-misalignment loss, and atmospheric variation shift the effective SNR within a bounded interval. Treating the operating point as a compound set, say SNR between 12 and 18 dB, the robust rate is fixed by the worst member, log₂(1 + γ_min). This is why conservative MCS selection and fade-margin budgeting in fixed wireless mirror compound-channel reasoning even when the term is not named.

Related Terms

← Composite Triple Beat Compound Slot →