An Efficient Memory Gradient Method for Extreme M-Eigenvalues of Elastic type Tensors

M-eigenvalues of fourth order hierarchically symmetric tensors play a significant role in nonlinear elastic material analysis and quantum entanglement problems. This paper focuses on computing extreme M-eigenvalues for such tensors. To achieve this, we first reformulate the M-eigenvalue problem as a sequence of unconstrained optimization problems by introducing a shift parameter. Subsequently, we develop a memory gradient method specifically designed to approximate these extreme M-eigenvalues. Under this framework, we establish the global convergence of the proposed method. Finally, comprehensive numerical experiments demonstrate the efficacy and stability of our approach.

💡 Research Summary

The paper addresses the computation of extreme M‑eigenvalues of fourth‑order hierarchically symmetric tensors, which arise in nonlinear elasticity and quantum entanglement. The authors first transform the M‑eigenvalue problem into a sequence of unconstrained optimization problems. By introducing a shift parameter t, they obtain two objective functions: the original quartic form
 f(x,y)=¼‖x‖⁴‖y‖⁴−½ A·x y x y,
and its shifted counterpart
 fₜ(x,y)=f(x,y)−t²‖x‖²‖y‖².
When the tensor possesses a positive M‑eigenvalue, the global minimum of f is –¼λ², attained at a non‑zero critical point (x,y*) with λ*=(‖x*‖²‖y*‖²). If all M‑eigenvalues are non‑positive, the shifted function guarantees a strictly negative global minimum for sufficiently large t, and any non‑zero critical point of fₜ corresponds to an M‑eigenvalue λ=(‖x‖²‖y‖²)−t.

Algorithm 1 implements an adaptive scheme: it first attempts to solve the original problem; if the norm of the obtained vectors falls below a tolerance, the algorithm switches to the shifted problem, increasing t geometrically until a non‑trivial critical point is found. The core solver for both sub‑problems is a newly proposed Memory Gradient Method (MGM).

In MGM, the variables are concatenated as z=(xᵀ,yᵀ)ᵀ and the objective Φ(z) denotes either f or fₜ. The iteration follows
 z_{k+1}=z_k+α_k d_k,
with step size α_k determined by a standard line‑search (e.g., Armijo). The search direction incorporates both the current gradient and a weighted average of the previous N directions:
 d_k=−γ_k g(z_k)+ (1/N)∑{i=1}^N β{k,i} d_{k−i},
where g(z)=∇Φ(z). The coefficients β_{k,i} are defined as β_{k,i}=‖g(z_k)‖² ϕ_{k,i}†, with ϕ_{k,i} chosen to satisfy inequalities that guarantee a sufficient descent condition. Specifically, the authors require
 ϕ_{k,1}>max{g(z_k)ᵀ d_{k−1}/γ_k,0},
 ϕ_{k,i}≥max{g(z_k)ᵀ d_{k−i}/γ_k,0} for i≥2.
These choices ensure that g(z_k)ᵀ d_k<0 for all k (Theorem 3.1) and, under mild bounds on γ_k and ϕ_{k,i}, that a stronger condition g(z_k)ᵀ d_k ≤ –c‖g(z_k)‖² holds (Theorem 3.2). Consequently, combined with a line‑search satisfying Wolfe or Armijo rules, the sequence {z_k} converges globally to a stationary point of Φ, which corresponds to an M‑eigenpair of the original tensor.

The convergence analysis proceeds by showing that Φ is bounded below, that the descent direction yields a sufficient decrease, and that the step sizes remain bounded away from zero. The memory term prevents stagnation often observed in pure steepest‑descent methods, especially in high‑dimensional settings where the landscape is highly non‑convex.

Extensive numerical experiments validate the theoretical claims. The authors test tensors of sizes (3,3), (5,5), (10,10), and larger random instances up to dimensions exceeding 100. They compare MGM against the alternating shifted power method, the BF‑GS based nonlinear power method, and a basic steepest‑descent scheme. Results show that MGM reduces the number of iterations by 30–60 % and cuts CPU time accordingly. Moreover, MGM exhibits robust performance when the shift parameter is adaptively increased, successfully handling cases with only non‑positive M‑eigenvalues. Memory usage remains modest because only a fixed number N of past directions are stored, making the method scalable to large problems.

In summary, the paper contributes a novel reformulation of the M‑eigenvalue problem, a memory‑enhanced gradient algorithm with provable global convergence, and compelling empirical evidence of its efficiency and stability. The work opens avenues for extending the approach to higher‑order tensors (order 6 or more), to tensors lacking hierarchical symmetry, and to parallel or GPU implementations for real‑time material‑property simulations.