In an open wireless medium, security is traditionally handled at the upper layers of the network stack using cryptographic encryption. However, Physical Layer Security (PLS) offers an alternative paradigm: exploiting the physical characteristics of the radio channel—such as noise, fading, and spatial degrees of freedom—to ensure that an eavesdropper ($E$) cannot extract any information from the transmitted signal, regardless of their computational power.
The fundamental metric in PLS is the secrecy rate, which quantifies the rate at which confidential information can be reliably sent to a legitimate receiver ($Y$) while remaining perfectly secret from an eavesdropper. Mathematically, the secrecy rate $R_s$ is expressed as the difference between the capacity of the legitimate channel and the capacity of the eavesdropper channel:
$$R_s = \left[ I(X; Y) - I(X; E) \right]^+$$
While conceptually elegant, maximizing this rate presents a formidable mathematical challenge. In standard wireless communications, maximizing mutual information $I(X; Y)$ under linear channel constraints is a convex optimization problem, easily solved via classic techniques like waterfilling (as seen in cognitive radio settings analyzed by Rajiv Soundararajan). However, subtracting the eavesdropper's mutual information $I(X; E)$ completely destroys this convexity.
This article explores why secrecy rate optimization is inherently non-convex and details the algorithmic frameworks developed to navigate its challenging mathematical landscapes.
The Root of Non-Convexity: The Difference-of-Concave Structure
To understand why maximizing the secrecy rate is difficult, consider a multi-antenna transmitter optimizing its transmit covariance matrix $\mathbf{Q}$ (which dictates how power is allocated across different spatial directions).
When the channels are Gaussian, both $I(X; Y)$ and $I(X; E)$ are concave functions with respect to $\mathbf{Q}$. In isolation, maximizing either function is straightforward. However, the objective function for the secrecy rate takes the form:
$$f(\mathbf{Q}) = g(\mathbf{Q}) - h(\mathbf{Q})$$
where both $g$ and $h$ are concave. The subtraction of a concave function is equivalent to adding a convex function. Optimizing a sum of a concave and a convex function results in a Difference-of-Convex (DC) programming problem.
Standard Capacity Optimization:
Matrix Q ───► [ Concave Function g(Q) ] ───► Single Global Maximum (Easy)
Secrecy Capacity Optimization:
Matrix Q ───► [ Concave g(Q) ] - [ Concave h(Q) ] ───► Multiple Local Optima (Hard)
As a result, the optimization landscape becomes plagued by multiple local maxima, saddle points, and non-differentiable boundaries. As Ghadamali Bagherikaram et al. and Ruoheng Liu et al. demonstrate, identifying the exact boundary of the secrecy capacity region for multi-antenna broadcast channels requires overcoming these non-convex matrix hurdles. The problem becomes even more severe when dealing with multiple eavesdroppers, where the objective shifts to a non-convex max-min problem to guarantee protection against the worst-case scenario.
Algorithmic Paradigms for Overcoming Non-Convexity
To solve these problems without resorting to exhaustive, computationally prohibitive grid searches, researchers have developed several mathematical workarounds that transform or approximate the non-convex landscape into a series of convex sub-problems.
1. Semidefinite Relaxation (SDR)
When designing optimal beamforming vectors $\mathbf{w}$ for Multi-Input Single-Output (MISO) channels, the transmit covariance matrix is defined as $\mathbf{W} = \mathbf{w}\mathbf{w}^H$. This introduces a strict constraint: the matrix $\mathbf{W}$ must be rank-one. The rank-one constraint is inherently non-convex.
To bypass this, Wing-Kin Ma et al. utilized Semidefinite Relaxation (SDR). By dropping the rank-one constraint entirely, the problem is transformed into a standard convex Semidefinite Program (SDP) that can be solved efficiently using interior-point methods. Crucially, the authors proved that for many MISO secrecy formulations, the optimal solution to the relaxed problem naturally yields a rank-one matrix, meaning the relaxation is tight and transmit beamforming is globally optimal. This framework was later extended by Wing-Kin Ma et al. to solve sum secrecy rate maximization in complex full-duplex two-way relay networks.
2. Successive Convex Approximation (SCA) and DC Programming
When the non-convexity cannot be relaxed away, Successive Convex Approximation (SCA) is used to iteratively linearize the problematic convex term.
Recall that the objective is $g(\mathbf{Q}) - h(\mathbf{Q})$, where $h(\mathbf{Q})$ is the concave term representing the eavesdropper's rate. Because any concave function lies below its first-order Taylor expansion, we can approximate $h(\mathbf{Q})$ around a chosen point $\mathbf{Q}^{(k)}$ with a linear function:
$$h(\mathbf{Q}) \approx h(\mathbf{Q}^{(k)}) + \nabla h(\mathbf{Q}^{(k)})^T (\mathbf{Q} - \mathbf{Q}^{(k)})$$
Since a linear function is technically both convex and concave, subtracting it from the concave function $g(\mathbf{Q})$ results in a purely concave local approximation. Gan Zheng et al. applied this exact approach to optimize full-duplex jamming parameters, while Pu Zhao et al. leveraged it to design robust beamforming for multi-user MISO networks under imperfect channel conditions.
3. Quasi-Concavity and Bisection Searches
In specific network configurations, the non-convex objective function exhibits a property known as quasi-concavity. A quasi-concave function is one whose upper-level sets are convex.
Lan Zhang et al. discovered that while multi-antenna secrecy objectives are non-convex, they display quasi-concavity when framed relative to interference temperature constraints. This structural insight allows researchers to use a simple bisection search method. By fixing the target secrecy rate, the global optimization problem collapses into a sequence of simple convex feasibility sub-problems, guaranteeing convergence to the global optimum.
Complicating the Landscape: Hardware and Signal Constraints
The non-convexity of physical layer security becomes even more pronounced when moving from idealized mathematical models to practical wireless deployments.
Discrete Signal Constellations
Most theoretical work assumes that transmitted signals follow a continuous Gaussian distribution. However, real-world systems use discrete constellations like QAM or PSK.
Haohao Qin et al. demonstrated that when discrete inputs are introduced, the secrecy rate function becomes highly non-concave and fluctuates wildly, exhibiting numerous local maxima even in simple single-antenna setups. This occurs because the mutual information of a discrete constellation saturates at high power levels, meaning that increasing transmit power can inadvertently help the eavesdropper more than the legitimate user. For channels plagued by intersymbol interference (ISI), Aria Nouri developed specialized Expectation-Maximization (EM) style iterative algorithms to navigate these highly unstable, non-concave landscapes.
Gaussian Input Secrecy Rate: Smooth, predictable non-convexity (DC structure).
Discrete Input (QAM) Secrecy Rate: Wildly fluctuating landscape with multiple sharp local peaks due to capacity saturation.
Multi-Variable Hardware Optimization (RIS and SWIPT)
Modern wireless paradigms introduce additional optimization variables that are tightly coupled with the transmit beamformers, compounding the non-convexity:
- Reconfigurable Intelligent Surfaces (RIS): When using an IRS to reflect signals securely, engineers must optimize both the base station beamformers and the IRS phase shifts. The phase shifts introduce a non-convex unit-modulus constraint ($|\theta_n| = 1$). Ming-Min Zhao et al. and Feng Shu et al. decoupled these variables using two-timescale and alternating optimization frameworks to maximize secrecy rates despite the lack of closed-form expressions.
- Simultaneous Wireless Information and Power Transfer (SWIPT): In SWIPT networks, the system must deliver a minimum amount of harvested energy to nodes while keeping data secure. As shown by WeiQiang Xu et al. and Mingxiong Zhao et al., these energy harvesting requirements introduce non-convex quadratic constraints that require block-wise penalty functions to resolve.
Summary
Non-convex secrecy optimization works because the underlying physics of wireless propagation contains hidden mathematical structures. Whether through the tightness of semidefinite relaxations, the iterative linearization of the eavesdropper's channel via DC programming, or exploiting quasi-concavity, wireless engineers can systematically transform intractable security equations into highly efficient, deployable algorithms.
Key Foundational Work
On the Relationship Between the Multi-antenna Secrecy Communications and Cognitive Radio Communications (2009) — contributed 3 concept elements to this direction.
The Secrecy Capacity Region of the Degraded Vector Gaussian Broadcast Channel (2009) — contributed 1 concept elements to this direction.
Cooperative Jamming for Wireless Physical Layer Security (2009) — contributed 1 concept elements to this direction.