Cubic Tensor: Theory & Applications

• Cubic tensor is a third-order multidimensional array characterized by rank decompositions and symmetry properties that facilitate diverse mathematical and computational applications.
• In optimization, cubic tensors are integral to cubic-quartic regularization schemes that enhance convergence and stability in high-order algorithmic frameworks.
• They underpin models in physics and materials science, capturing phenomena in elasticity, magneto-optics, and phase-field theories with precise symmetry-based parameterizations.

A cubic tensor is a tensorial object of order three, with components indexed by three indices (e.g., $T_{ijk}$), and arises as a core algebraic and structural object in multilinear algebra, optimization, physics, materials science, and invariant theory. The term “cubic tensor” appears in varied technical contexts—including as a rank-3 array, as a structural motif in higher-order derivatives, within the representation theory of cubic forms, as cubic corrections or regularizations in numerical optimization, and as fundamental building blocks in magneto-optical, elasticity, and phase field theories.

1. Algebraic Structure and Rank Properties of Cubic Tensors

A cubic tensor $\mathcal{A}$ of size $n_1 \times n_2 \times n_3$ is a third-order multidimensional array $\mathcal{A}_{i_1 i_2 i_3}$ over a field $\mathbb{F}$, typically $\mathbb{R}$ or $\mathbb{C}$, with the possibility of symmetry under permutations of indices (symmetric tensors) or with no symmetries (generic tensors). The canonical form of a pure (rank-1) cubic tensor is the outer product $a \otimes b \otimes c$ with components $(a_i b_j c_k)$. The tensor rank (CP rank) is the minimal number $r$ so that $$\mathcal{A} = \sum_{t=1}^r a^{(t)} \otimes b^{(t)} \otimes c^{(t)}$$. For symmetric tensors, the symmetric/Waring rank is defined via $v \otimes v \otimes v$ representations, and the relationships between generic, symmetric, and border ranks for cubic tensors are well understood in low rank regimes. Specifically, for cubic surfaces ($4\times4\times4$ symmetric tensors), the symmetric rank and tensor rank coincide, as do the (symmetric) border ranks in sub-generic cases, with algebraic criteria given by discriminant and Hessian conditions (Seigal, 2018).

Table: Ranks for Cubic Tensors (Symmetric Case, $4 \times 4 \times 4$) | Notion | Definition | Equality Cases | |-------------------|------------------------------------------------------|---------------------------------------| | Rank | Min $r$ with sum of rank-1 outer products | Always equals symmetric rank for $n=4$| | Symmetric rank | Min $r$ sum $v_i \otimes v_i \otimes v_i$ | All cubic surfaces | | Border rank | Zariski closure of rank $\leq r$ tensors | Same as symmetric border rank $(r \le 5)$ | | Symmetric border rank | Zariski closure of symmetric rank $\le r$ | Same as border rank in these cases |

This equality (Comon's property) is proven for all cubic surfaces; the discriminant and Hessian discriminant, both classical invariants of the cubic form, fully characterize generic rank behavior and allow an explicit discriminant test for low rank (Seigal, 2018).

2. Cubic Tensors in Optimization and Regularization

Third-order tensors naturally encode cubic corrections in high-order methods for unconstrained optimization, as in the cubic-quartic regularization (CQR) paradigm. In CQR, local models of the form

$$m(s) = g^T s + \tfrac{1}{2} s^T H s + \tfrac{1}{6} T[s,s,s] + \tfrac{\sigma}{4} \|s\|^4$$

are minimized, where $g$ is the gradient, $H$ the Hessian, $T$ the cubic tensor (third derivative of $f$), and $\sigma > 0$ enforces regularization. The CQR algorithm achieves optimal global evaluation complexity $\mathcal{O}(\epsilon^{-3/2})$ for reaching approximate first-order criticality, outperforming or matching state-of-the-art ARC and quartic-regularization frameworks, especially under ill-conditioning and in the presence of large tensor components. Solving the CQR subproblem reduces to a nonlinear secular equation involving both the spectrum of $H$ and the effective cubic tensor contraction, with a robust global convergence guarantee (Zhu et al., 2023).

Special cases, such as univariate or separable cubic tensors (where $T$ is diagonal across modes), allow exact or dimensionwise reductions, obtaining the global minimizer in closed form or with drastically reduced computation. For tensors with small $\|T\|$, the convergence rate improves further, demonstrating the nuanced role of the cubic tensor structure in the global landscape and step-size control of tensor-regularized methods.

3. Cubic Tensors in Invariant Theory and Computational Complexity

Cubic tensors are central to understanding the rings of invariants under group actions, particularly in the representation theory of $$\mathrm{SL}(V) \times \mathrm{SL}(W) \times \mathrm{SL}(Z)$$ acting on $(V \otimes W \otimes Z)^{\oplus 9}$. Derksen–Makam proved that any generating set of invariants for such cubic tensor spaces requires polynomials of exponentially large degree: in $(U \otimes V \otimes W)^{\oplus 9}$, the minimal upper bound degree is at least $4^n-1$ for $n = \dim U/3$. The Grosshans principle is leveraged to transfer degree bounds from group actions to torus invariants, and to show that even for “simple” cubic tensor representations, the null cone and invariant ring possess extreme algebraic complexity. These exponential lower bounds manifest a central computational barrier for geometric complexity theory (GCT) and derandomization of polynomial identity testing (PIT) with higher-order tensor actions (Derksen et al., 2019).

Table: Invariant Degree Bounds for Cubic Tensor Spaces | Action | Space | Min. Generator Degree | |---------------------------------------------|-----------------------------------|----------------------------| | $\mathrm{SL}(V)$ on $\operatorname{Sym}^3(V)^{\oplus 4}$ | $V$ of $\dim=3n$ | $3(4^n - 1)$ | | $$\mathrm{SL}(U)\times\mathrm{SL}(V)\times\mathrm{SL}(W)$$ | $(U \otimes V \otimes W)^{\oplus 9}$ | $4^n - 1$ |

These results underline the intrinsic algebraic hardness and absence of compact generating sets for the null cone in cubic tensor settings, with implications for computational invariant theory and orbit-closure algorithms.

4. Cubic Tensors in Material Science and Physics

Fourth-order and higher tensors in cubic symmetry classes precisely capture physical properties in elasticity, magneto-optics, and free energy expansions:

• Elasticity Tensors: The elasticity tensor $C_{ijkl}$ for cubic crystals, under appropriate basis, reduces to three independent constants $(C_{11}, C_{12}, C_{44})$, with Zener's ratio $A = 2 C_{44}/(C_{11} - C_{12})$ quantifying anisotropy. To estimate the "distance to cubic symmetry" of a general elasticity tensor, analytical upper-bound methods using harmonic decomposition and quadratic minimization problems yield sharp orientation-sensitive projections, crucial for robust materials modeling (Desmorat et al., 2022, Tasnádi et al., 2011).
• Magneto-optic and Conductivity Tensors: The permittivity and conductivity tensors in cubic crystals organize via cubic-in-magnetization expansions: $$\varepsilon_{ij}(M) = \varepsilon_{ij}^{(0)} + K_{ijk} M_k + G_{ijkl} M_k M_l$$, with cubic symmetry dictating that only a few principal spectra are needed to describe all quadratic effects, both in optical and transport regimes. Independent quadratic tensor parameters are extracted via directional scans and sample rotations, confirming group-theoretic predictions (Hamrlová et al., 2012, Hamrlová et al., 2016).

In phase-field modeling, the inclusion of sixth-rank “gradient-only” cubic tensors $M_{ijklmn}$ in the free energy expansion enables the capture of true cubic and hexagonal anisotropies, extending beyond the capabilities of fourth-rank models and facilitating precise description of crystal facet selections and energy penalization for interfacial energy variations (Nani et al., 2014).

5. Cubic Tensors in Tensor Completion, Statistics, and Algorithmics

Cubic tensor completion arises when determining a rank-1 (CP) decomposition of a partially observed third-order tensor. The rank-1 completion problem is equivalent to a structured matrix completion constrained by $2 \times 2$ minors vanishing across frontal slices, permitting technological advances by exploiting equivalence with rank-1 matrix recovery. When strongly completable, an explicit iterative traversal algorithm solves the problem exactly and efficiently, while in underdetermined cases, convex nuclear norm relaxation or Lasserre’s moment/SOS hierarchies yield completion or certify infeasibility in finite steps (Zhou et al., 2024).

Low-rank and sparse estimation from "cubic sketchings" is foundational in modern statistics. Efficient two-stage algorithms combine tensor moment estimators with thresholded gradient methods, attaining non-asymptotic optimal rates for exact and stable recovery of sparse low-rank cubic tensors, with key applications in high-dimensional multi-way interaction models for regression and learning (Hao et al., 2018).

6. Cubic Powers and Spectral Limits of Tensors

A distinct non-linear algebraic operation, the cubic power $T^{[3]}$ of a third-order tensor $T$ is defined componentwise by $$(T^{[3]})_{ijk} = \sum_{p,q,r} T_{i q r} T_{p j r} T_{p q k}$$. Iterating the cubic power yields a sequence $T \to T^{[3]} \to T^{[6]} \to \dots$ whose growth is captured by the Gelfand limit

$\rho(T) = \lim_{m \to \infty} \|T^{[3m]}\|^{1/m},$

which parallels the classical spectral radius for matrices. The Gelfand limit is uniquely defined (norm-independent), sub-multiplicative under tensor-norms (but not spectral or $\ell_\infty$ norms), and satisfies

• $\rho(T) = 0$ iff $T$ is nilpotent under cubic powers,
• $\rho(T) \in \{0,1\}$ if $T$ is idempotent. For norm $\|T\| < 1$, the sequence $T^{[3m]} \to 0$, characterizing contractivity. These constructions extend to all odd-order tensors (order $2k+1$) via appropriate multilinear contractions (Qi et al., 2019).

7. Specialized Constructions: Cubic Discriminants in Geometry

In differential geometry, certain rank-4 tensors encoding the discriminant of cubic polynomials—the “cubic discriminant tensors”—appear naturally on 8-dimensional Riemannian manifolds whose structure group reduces to ${SO(4)}_{ir}$. The discriminant tensor is realized by an invariant section in the bundle of hyper-Kähler curvature-type tensors and is algebraically equivalent (at each point) to the degree-4 discriminant of a cubic polynomial. Integrability of such tensors rigidly classifies the underlying geometry: only the quaternion–Kähler symmetric spaces $G_2/SO(4)$ and $G_{2(2)}/SO(4)$ admit non-flat, integrable cubic discriminants, establishing a deep connection between invariant cubic tensors and special holonomy structures (Hristova et al., 16 Aug 2025).

---

Cubic tensors, in their various guises, provide a unifying structural and computational language, connecting advanced optimization, algebraic geometry, statistical estimation, representation theory, nanoscale physics, and geometric analysis. Their roles span from defining the analytic landscape of nonlinear algorithms and physical property tensors to dictating the algebraic complexity of orbits and syzygies in invariants, with each domain exploiting tensor rank, symmetry, contraction, and invariance within its own technical framework.