Source-linked AI summary
A Note on the Secrecy Capacity of the Multi-antenna Wiretap Channel
Tie Liu, Shlomo Shamai
TL;DR
The paper addresses the difficult nonconvex evaluation of secrecy capacity for multi-antenna wiretap channels. It develops a channel-enhancement characterization using an extremal entropy inequality and physical transmission intuition, proving a capacity formula for general vector Gaussian wiretap channels under stated matrix conditions. The resulting characterization is established through degraded-channel constructions and theorem proofs, while its scope includes power-covariance constraints and retains nonconvex Gaussian optimization issues.
Problem
Evaluating the multi-antenna secrecy-capacity expression is a functional, nonconvex problem, and single-antenna methods do not generally extend to nondegraded channels.
Method
The note uses channel enhancement and an extremal entropy inequality from multi-antenna broadcast channels, with vector Gaussian channel constructions under a power-covariance constraint.
Results
Theorem 2 states that the secrecy capacity of a general vector Gaussian wiretap channel can be written as a characterized expression, with the proof completed through matching inequalities.
Takeaways & Limitations
The alternative characterization is directly built on physical intuition regarding the optimal transmission strategy in the multi-antenna wiretap setting.
Takeaways & Limitations
The characterization focuses on vector Gaussian wiretap channels with a power-covariance constraint and includes an unresolved nonconvex covariance-optimization difficulty.
Abstract
from arXiv · showhide
Recently, the secrecy capacity of the multi-antenna wiretap channel was characterized by Khisti and Wornell [1] using a Sato-like argument. This note presents an alternative characterization using a channel enhancement argument. This characterization relies on an extremal entropy inequality recently proved in the context of multi-antenna broadcast channels, and is directly built on the physical intuition regarding to the optimal transmission strategy in this communication scenario.
1 Introduction
The multi-antenna wiretap secrecy-capacity problem reduces to evaluating a nonconvex auxiliary-variable expression, motivating indirect bounds and an alternative channel-enhancement characterization. The alternative uses an extremal entropy inequality and reflects the physical intuition behind optimal transmission.
- The channel has n_t transmit antennas and n_r and n_e antennas at the legitimate receiver and eavesdropper, respectively.
- Secrecy capacity is the maximum reliably decodable rate that remains unavailable to the eavesdropper.
- The capacity reduces to evaluating max P(U,X) [I(U; Y_r) − I(U; Y_e)] for the specific continuous-alphabet channel under a power constraint.
- Evaluating this expression is a functional, nonconvex optimization problem, and single-antenna entropy-power methods do not extend generally to nondegraded multi-antenna channels.
- Khisti and Wornell upper-bounded the secrecy objective through a genie-aided joint-output channel and optimized over the worst cooperation between legitimate receiver and eavesdropper.
- Their Gaussian saddle-point characterization matches the original optimization, establishing equal secrecy capacity when the legitimate user accesses both received signals.
- This note instead uses an extremal entropy inequality from multi-antenna broadcast channels and builds the characterization on physical intuition about optimal transmission.
2 Capacity Characterization via a Channel Enhancement Argument
The note characterizes secrecy capacity through channel enhancement: construct a degraded enhanced channel with the same secrecy capacity, then extend the result from degraded to general vector Gaussian wiretap channels.
- Channel model: The vector Gaussian wiretap channel is analyzed under a positive-semidefinite power-covariance constraint, which includes total power constraints as a special case.The constraint uses a positive definite matrix S and the ordering ⪯ denotes positive-semidefinite ordering.
- Degraded channel: Theorem 1 characterizes degraded-channel secrecy capacity when an admissible positive-semidefinite K∗_x satisfies the stated matrix condition, with Gaussian U = X optimal.The proof uses a Gaussian extremal entropy inequality to establish the result.
- Limitation: The general characterization of K∗_x relies on standard KKT conditions, which provide only a necessary condition rather than the sufficient condition available in the degraded case.The associated matrix optimization remains nonconvex, making covariance optimality difficult to certify directly.
- General channel: Theorem 2 extends the characterization to the general vector Gaussian wiretap channel using a channel enhancement argument.The enhanced channel is constructed from an optimal covariance K∗_x and an auxiliary legitimate-receiver noise covariance ˜K_r.
- General channel: The enhanced channel is degraded because ˜K_r ⪯ K_e, and Theorem 1 supplies its secrecy capacity for the reverse-capacity bound.Reducing the legitimate receiver’s noise can only increase secrecy capacity, completing the comparison with the original channel.
- Physical intuition: The construction is motivated by transmitting only along directions where the legitimate receiver is stronger, leaving the eavesdropper’s effective channel degraded relative to the legitimate receiver.Unlike broadcast-channel enhancement, improving the eavesdropper’s channel can reduce secrecy capacity, so the enhancement must preserve security carefully.
A Proof of Inequality (17)
The proof establishes inequality (17) by comparing an arbitrary input X with an optimal Gaussian input X* through a Gaussian interpolation. De Bruijn’s identity, Fisher-information inequalities, and monotonicity show that g(X) cannot exceed g(X*).
- Proof strategy: The proof targets g(X) ≤ g(X*) for every input satisfying E[XX^t] ⪯ S.The comparison is designed to address Gaussianity and covariance optimization together.
- Proof strategy: For λ ∈ [0, 1], the proof introduces an interpolating random variable formed from X and an independent Gaussian X*G.This interpolation connects the arbitrary input to the Gaussian candidate.
- Entropy and Fisher information: De Bruijn’s identity converts the derivative along the interpolation into expressions involving the Fisher information matrix J(X).The proof then invokes the vector Fisher information inequality for independent random vectors.
- Monotonicity: The Cramér–Rao inequality and covariance bound establish that g(Xλ) is monotonically nondecreasing in λ.The covariance condition used is Cov(X) ⪯ E[XX^t] ⪯ S.
- Conclusion: Evaluating the monotonicity relation at the endpoints yields inequality (17), completing the proof.The argument concludes directly after establishing the endpoint comparison.