Using scientific machine learning for experimental bifurcation analysis of dynamic systems

The paper explicitly sets up the CBC root-finding equation e1(A1^t, V) = A1^t − A1(V) (its Eq. B.7) and assumes a unique solution V* for each target amplitude, implementing a cubic polynomial surrogate over sampled airspeeds to locate roots; however, it does not supply the mathematical conditions ensuring injectivity, continuity of the inverse, or interpolation/root-locus error bounds . It specifies the reduced UDE model M_red ẍ + B_red ẋ + S_red x + f_NL^red(x) + U(x; V) = 0 (its Eq. (13)) and introduces process noise as in Eq. (17), but again gives no rigorous existence/persistence guarantees for reproducing the full branch with a learned U(x;V) . Finally, the paper correctly attributes the low-amplitude onset bias in noisy CBC to an Ornstein–Uhlenbeck baseline (with corroborating plots and a qualitative explanation of the apparent shift of Hopf with increasing σ), but it does not derive an analytical scaling law for the shift . By contrast, the candidate solution supplies the missing hypotheses (strict monotonicity or uniqueness-on-amplitude sets, regularity and a derivative lower bound, C^1 universal approximation, and hyperbolicity/structural stability) and delivers standard, correct arguments: an O(h^{r+1}) root error for interpolation under |A1'|≥m>0, a C^1-approximation-plus-persistence proof of a UDE reproducing the branch to within ε, and a near-Hopf OU calculation yielding Var[h] ∼ const·σ^2/|μ(V)| and the leading-order bias V_H − V^-(A1^t,σ) ≍ σ^2/A1^{t2}. The only caveat is that uniform hyperbolicity must exclude neighborhoods of cycle-folds; this is easily handled by restricting to subintervals away from saddle-node points. Net: the paper’s narrative is empirically sound but incomplete; the model’s solution is correct under clearly stated assumptions.