Swapping Algebraic Twins Without Breaking Math — An Introduction to Galois Theory

Galois theory reframed as a universal theory of what can be solved by breaking a problem into sequential steps — and what can't.

The central trick of Galois theory is asking: can you swap the solutions of a polynomial equation without any algebraic test noticing the difference? Take x² + 1 = 0. It has two roots, i and −i. From the perspective of the rationals, these are perfect twins — no polynomial with rational coefficients can tell them apart. The map that swaps them (complex conjugation) preserves every algebraic relationship. The collection of all such legal swaps forms the Galois group, and its structure encodes everything about how "rigid" or "floppy" the equation's solution space is.

The paper's main payoff is connecting this to solvability. A polynomial is solvable by radicals — meaning you can write its roots using +, −, ×, ÷, and nth roots — if and only if its Galois group can be peeled apart layer by layer, each layer being abelian (commutative). The quadratic, cubic, and quartic formulas exist because their symmetry groups decompose this way. The quintic fails because its Galois group contains A₅, the alternating group on five elements, which is a monolithic block with no internal factoring structure. You can't unwind it into sequential steps, so no sequential radical formula exists. This isn't a gap in human cleverness — it's a structural impossibility.

The AIT framing adds a clean lens: the Galois group measures the leftover ambiguity after a maximally compressed algebraic description has fixed everything it can. Specifying "adjoin a root of x² + 1 = 0" pins down a relation but not which root gets called i. That one bit of unresolved freedom is exactly the two-element Galois group. Symmetry, in this reading, is what remains underdetermined after maximum compression.

The paper then argues this same dichotomy — compositional/solvable vs. monolithic/irreducible — runs through differential equations (Lie's theory) and deep learning (Poggio's compositionality results). Deep networks succeed when the target function decomposes into a hierarchy of simpler functions; they fail when the structure is monolithic. The table the paper provides is genuinely clarifying: in all three domains, "solvable" means "decomposable into sequential layers," and "unsolvable" means "the symmetry is a rigid block that resists layering."

The synthesis is a staircase model of understanding: an agent composes existing tools until it hits a monolithic barrier, then must invent a new primitive (like the Bring radical, or the Airy function) that matches the problem's symmetry, then re-enters compositional search with the expanded toolkit. The cycle repeats. This is a genuinely interesting framing — not just a metaphor, but a claim that the same mathematical condition (group solvability) governs when sequential, compositional methods suffice and when genuinely new operations must be discovered.