CALLIOPE · BCOM

This note aims to demystify Galois Theory by connecting its foundational definitions to a broader principle of computational and compositional tractability. We begin with the elementary example of ${Q}(i)$ to illustrate how algebraic symmetries arise from the indistinguishability of roots. We then proceed to the heart of the theory: the rigorous link between symmetry and solvability. We reframe the Abel-Ruffini theorem not merely as a limit on polynomials, but as a positive definition of structure: a problem is "solvable" if and only if its symmetry group is compositional (decomposable into abelian steps). When this condition fails—as with the monolithic $A_5$ group—we show that progress requires the discovery of new atomic primitives (the Nuclear Option"). Finally, we discuss this logic in the context of the continuous domain (Lie Theory) and Artificial Intelligence. Drawing on Poggio’s work on compositionality and Kolmogorov Theory, we demonstrate that learning" is the algorithmic cycle of discovering these symmetries. The resulting staircase" in the Kolmogorov Structure Function reveals that intelligence is the interplay of two modes: exploiting existing tools through composition, and inventing new cognitive cores to further compositionally compress the unsolvable".

Galois theory is secretly a theory of what makes problems solvable — and that same logic governs why deep learning works.

The paper starts with a simple, concrete example: the equation x² + 1 = 0 has two roots, i and −i. From the perspective of the rational numbers, these two roots are completely indistinguishable — no polynomial with rational coefficients can tell them apart. That indistinguishability is a symmetry: you can swap i and −i everywhere and all the arithmetic still checks out. The collection of all such valid swaps forms the Galois group of the equation. For this example it has just two elements, corresponding to one bit of unresolved ambiguity left over after the defining constraint pins down everything else.

The deeper payoff comes when you ask: which equations can be solved by a formula? The quadratic formula exists. So do cubic and quartic formulas. But no general formula exists for degree-5 polynomials — this is the Abel-Ruffini theorem. Galois reframed this not as a failure but as a structural diagnosis. A formula using nested radicals (square roots, cube roots, etc.) works precisely when the symmetry group of the equation can be peeled apart layer by layer into simple, commutative (abelian) pieces — like an onion. The paper calls this "compositionality." When the group is instead a monolithic block — specifically A₅, the symmetry group of the icosahedron — no such peeling is possible, and no radical formula can exist. The quintic isn't unsolvable because it lacks structure; it's unsolvable because its structure is too rigid to decompose.

The "nuclear option" is what happens next: invent new primitives. Hermite and Kronecker showed that elliptic modular functions, which carry exactly the icosahedral symmetry that radicals lack, can solve the quintic. The pattern generalizes: when your toolkit's symmetry doesn't match the problem's symmetry, you're stuck — until you define a new atomic operation that does match. The paper traces the same logic into differential equations via Lie theory, where solvability again means the symmetry group of the solution space decomposes into abelian layers, enabling step-by-step integration.

The final move connects all of this to AI and learning theory. Drawing on Poggio's work, the paper argues that deep networks succeed because the functions describing physical reality are compositional — they decompose into hierarchical layers of simpler functions, exactly mirroring a solvable Galois group. When a learner hits a phenomenon whose structure is monolithic relative to its current toolkit, it stalls. Progress requires discovering a new cognitive primitive — a new layer type, a new inductive bias — that compresses the previously incompressible. The Kolmogorov Structure Function (a curve trading model complexity against prediction error) descends not smoothly but in a staircase: long slopes of compositional optimization punctuated by sharp drops when a new primitive is discovered and added to the library. Intelligence, the paper concludes, is this cycle — compose, stall, discover, re-compose — running indefinitely.