Antti Kupiainen (2023), Nature Physics 19, 1539–1541 (2023), Comment — DOI-10.1038/s41567-023-02233-x
Key takeaways
From the reading, NOTE-821.
- From axioms to path integrals. Constructive QFT tries to define the Euclidean path integral ⟨φ(x₁)…φ(x_n)⟩ = ∫φ(x₁)…φ(x_n)e^{−S(φ)/ħ}Dφ non-perturbatively. Lattice and box cutoffs turn this into statistical mechanics of continuous spins.
- Dimension decides. Scaling sends the φ⁴ coupling to (ℓ^{4−d}λ). It is suppressed at short distances below four dimensions and at long distances above. d < 4 needs only mass renormalization. d = 4 needs infinitely many counterterms.
- One RG step is easy; iterating it is the work. The block-spin integral e^{−H_{Lℓ}(ϕ′)} = ∫e^{−H_ℓ(ϕ)}δ(ϕ′ − Cϕ)Dϕ has both cutoffs, so it is non-critical. Its perturbation series diverges like n!. A polymer expansion converges where the field is small. Large fields are improbable and are handled separately.
- What it proved. Massless 4D lattice φ⁴ has Gaussian decay ⟨ϕ(x)ϕ(y)⟩ ∼ Z|x − y|^{−2} (Gawędzki and Kupiainen 1985; Feldman et al. 1987). Continuum 4D φ⁴ is Gaussian for any bare parameters (Aizenman and Duminil-Copin 2021). The 2D Gross–Neveu model has been constructed. Hall-conductivity universality has been proved.
- Where it stops. “A more nonperturbative formalism is still missing.” Numerical RG shares the problem. The author sees tensor-network RG, which keeps actions strictly local, as more promising.
Standing in the record
Filed from the survey of 2026-09-25 of work the anthology set aside as out of scope (tier C), and read in full on 2026-10-14 in NOTE-821, from the green copy the seed could not reach. LIT-881 and LIT-873 cite it as an unread neighbour.
It does not address deep learning, as the seed guessed. It is held as the record’s orientation to rigorous RG and as context for the renormalization-and-learning accounts THEORY-194, THEORY-197 and THEORY-202. What rigorous RG controls is the iterated flow of the whole effective Hamiltonian, with its irrelevant remainder and its large fields. The exactness of a single coarse-graining step, which those accounts examine, is the part Kupiainen treats as unproblematic.
No THEORY is filed from it. Its one result fit to file, 4D φ⁴ triviality, should come from Aizenman and Duminil-Copin’s paper, which the record does not hold.