Machine Learning Invariants of Tensors

The Big Picture

Imagine describing the shape of a crystal. No matter how you rotate or flip it, certain properties stay the same: the number of faces, the angles between them, the overall symmetry. These unchanging quantities are called invariants, and physicists have been hunting them for over a century.

The problem scales badly. A single electromagnetic field has two independent invariants. A gravitational field in ten dimensions? Dozens, tangled in hidden relationships. Sorting out which invariants are truly independent and which are disguised combinations of others has traditionally required years of mathematical heroics.

Elamaran, Ferko, and Scarlett have automated much of that heroism. Their algorithm uses random sampling, numerical linear algebra, and graph enumeration to discover the independent invariants of any tensor (a structured mathematical object, like a matrix but with potentially more dimensions). What used to be a years-long grind becomes an overnight computation.

Their test case: an antisymmetric 3-form field in six-dimensional space, a structure that shows up all over string theory. The algorithm confirmed it has exactly five independent invariants, which constrains the physical theories you can build from that field.

Key Insight: By generating random instances of a tensor and hunting for linear dependencies between candidate scalars, this algorithm can experimentally identify which invariants are truly independent, turning a hard algebraic problem into a tractable numerical one.

How It Works

The idea is straightforward: compute two quantities for thousands of random inputs; if some weighted sum always equals zero, the quantities are secretly related. The algorithm runs in three stages.

Stage 1: Graph enumeration. Every scalar quantity you can build from a tensor can be drawn as a tensor network, a diagram where nodes represent copies of the tensor and edges represent pairs of indices being summed over. The algorithm systematically enumerates all inequivalent diagrams at each order, building a complete candidate list. For a rank-3 antisymmetric tensor (one where swapping any two indices flips the sign), the count at low orders is manageable. Higher orders grow combinatorially, but the enumeration stays systematic.

Stage 2: Numerical evaluation. For each candidate scalar, the algorithm draws many random instances of the tensor, with components sampled from Gaussian distributions, and computes every invariant for every draw. This produces a large matrix: rows are random samples, columns are candidate invariants.

Stage 3: Linear algebra. The team searches for syzygies, linear relations of the form $c_1 I_1 + c_2 I_2 + \ldots = 0$ holding across all draws. Numerically, this means finding the null space of the data matrix. A nonzero null vector reveals a genuine algebraic identity: one invariant is a combination of the others. When no new independent invariants appear across several consecutive orders, the list is considered complete.

The 3-form test case. The team applied the algorithm to an antisymmetric 3-form $H_{\mu\nu\rho}$ in six dimensions. This isn’t an exotic curiosity; it appears in string theory and supergravity as the field strength of the B-field in the NS-NS (Neveu-Schwarz) sector. The algorithm builds scalars from $H$ three ways:

• Trace variables: Multiply copies of $H$ and sum over shared indices using the metric, forming combinations like $H_{\mu\nu\rho}H^{\mu\nu\rho}$ and higher-order versions.
• Hodge dual variables: Use the six-dimensional Levi-Civita symbol $\epsilon_{\mu\nu\rho\sigma\lambda\kappa}$ to construct a dual of $H$, then build invariants from that.
• Spinor variables: Rewrite the tensor in spinor indices via six-dimensional Weyl matrices, revealing invariants invisible in the original picture.

After running the algorithm across all three families, the answer was clear: only five invariants are truly independent. All others can be expressed as explicit algebraic functions of these five, with exact formulas provided in the paper.

The payoff is concrete. Any Lagrangian for a 3-form in six dimensions (the master function from which equations of motion and physical predictions derive) is a function $L(x_1, x_2, x_3, x_4, x_5)$ of exactly five variables. No more, no less. Theories that look different on the surface may turn out to be the same theory in disguise.

Why It Matters

In string theory, the independent invariants of the Riemann tensor in ten dimensions are essential for classifying higher-derivative corrections to supergravity, the so-called $\alpha’$ expansion. Working these out by hand has occupied experts for decades. Automating invariant-finding could cut through years of that effort.

The same approach applies in quantum information. The independent invariants of a multi-partite quantum state under local unitary transformations are exactly the entanglement measures. Identifying them is the mathematical backbone of classifying genuinely distinct kinds of entanglement. Point the same algorithm at a different tensor, and it could push that classification further.

Numerical linear algebra and random sampling turn out to be experimental tools for pure mathematics. Not approximations, but hypothesis generators later confirmed analytically. That’s a mode of doing math that deserves wider adoption.

Bottom Line: By treating invariant theory as a data problem, this team confirmed that a 6D antisymmetric 3-form has exactly five independent invariants and built a reusable algorithm that can answer the same question for any tensor, with applications across string theory, quantum information, and algebraic geometry.