Low-tubal-rank tensor recovery is often solved through factorized formulations, which are scalable but nonconvex. While convergence guarantees have been studied for such methods, their global optimization landscape remains less understood. This letter studies a symmetric and noiseless low-tubal-rank tensor sensing problem under the t-product algebra, covering both exact-rank and over-estimated-rank factorizations. Under a standard t-RIP condition, we prove that the factorized objective has no spurious local minima in the noiseless case. We further provide a frequency-localized certificate for the negative curvature along a Procrustes displacement, showing the certified escape-curvature scale is an error-weighted average of slice-wise spectral strengths. Simulations support the theoretical findings.
