We prove a polynomial bound in the "true complexity" problem of Gowers and Wolf. The proof uses only repeated applications of the Cauchy-Schwarz inequality, answering negatively a question posed by Gowers and Wolf.To choose and reason about the sequence of Cauchy-Schwarz steps needed, we need to introduce several layers of formalism and theory. The highest level of abstraction in this framework concerns building what we term "arithmetic circuits" encoding computations in multilinear algebra.It is plausible this machinery could be used to generate arithmetic inequalities in greater generality, and we state some conjectures along these lines. FREDDIE MANNERS 6. A proof of Theorem 1.1.5 80 6.1. The structure of the proof 80 6.2. Some stashing-type proofs 81 6.3. The Bridge gate 83 6.4. The super AGate 84 6.5. The big AggreGate 86 6.6. StarGates and Lemma 6.1.2 88 6.7. Building a StarGate 91 6.8. Assigning a StarGate 94 7. Conjectures 99 7.1. Cases to avoid 99 7.2. A precise phrasing 100 7.3. Requiring an elementary argument 101 7.4. A special case 102 References 102 1. Note we have control of Λ Φ by some Gowers norm • U scs+1 , by the Cauchy-Schwarz complexity argument [GW10, Theorem 2.3]. Hence we are free to modify f i by small errors in the • U scs +1 -norm. 2. Apply an inverse theorem for the Gowers •GTZ12], or for quantitative bounds [GM21, Man18b]). 3. Taking steps 1 and 2 together, we may assume WLOG that f i are "U scs+1 structured functions": i.e., nilsequences (if n = 1 and p is large) or phase polynomials (if p is fixed and n is large).11 To be accurate, some of these results proved the original non-multilinear conjecture, with However, Hatami and Lovett [HL11] (in finite fields) and Altman's argument (for n = 1) prove the multilinear version discussed here.12 The original version of [GT10a] did not mention this extra hypothesis, but it was later observed by Altman that the proofs assumed it implicitly.