✗ Mechanism failed 2026

Adversarial Subspace Residual Localizer

Implementation & benchmark of arXiv:2609.03264 — Certified Spherical MUSIC for 3D Localization under Adversarial Subspace Perturbations

Usefulness7/10

Difficulty5/10

Novelty7/10

Architecture Stability Accuracy Sample efficiency Embedding Graph nn Training Linear algebra Spectral theory Geometry Harmonic analysis

Source paper: Certified Spherical MUSIC for 3D Localization under Adversarial Subspace Perturbations arXiv:2609.03264 ⓘ · analyzed Sep 4, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Replace a fragile learned similarity score for continuous object coordinates with a projection residual against a learned signal subspace. Candidate coordinates are represented by normalized Fourier or positional feature vectors, and the score is the fraction of feature energy outside the estimated subspace. The score remains useful even when the estimated subspace is adversarially rotated, because the perturbation is controlled directly by a sine-theta distance rather than by assumptions about random noise.

Formulas

$$\widetilde q(\bm z)=1-\|P_{\widetilde{\mathcal U}}\varphi_{\bm z}\|_2^2,$$

$$\|P_{\widetilde{\mathcal U}}-P_{\mathcal U}\|_2\leq\varepsilon_{\rm sub},\qquad |\widetilde q(\bm z)-q(\bm z)|\leq\varepsilon_{\rm sub},$$

$$\varphi_{\bm z}(\bm\omega)=e^{\mathrm i\kappa\bm\omega\cdot\bm z},\qquad u_m^\infty(\bm\omega)=\sum_{j=1}^{s}\alpha_j^{(m)}e^{\mathrm i\kappa\bm\omega\cdot\bm x_j}.$$

$$\kappa\delta_X\gtrsim s^{2/3},$$

Mathematical statement

The paper defines the spherical MUSIC objective $\widetilde q(\bm z)=1-\|P_{\widetilde{\mathcal U}}\varphi_{\bm z}\|_2^2$, where $\varphi_{\bm z}$ is the normalized steering or feature vector for coordinate $\bm z$, $\mathcal U$ is the ideal $s$-dimensional signal subspace, $\widetilde{\mathcal U}$ is its estimated version, and $P_{\widetilde{\mathcal U}}$ is the orthogonal projector onto the estimate. The perturbation size is measured by the sine-theta distance, operationalized as $\|P_{\widetilde{\mathcal U}}-P_{\mathcal U}\|_2\leq\varepsilon_{\rm sub}$. Since $\|\varphi_{\bm z}\|_2=1$, the score perturbation obeys $|\widetilde q(\bm z)-q(\bm z)|\leq\varepsilon_{\rm sub}$, where $q(\bm z)=1-\|P_{\mathcal U}\varphi_{\bm z}\|_2^2$. The paper proves that, under separation and frame-conditioning hypotheses, the perturbed objective has one strongly convex well near each true point and a uniform value gap outside those wells.

Implementation notes

Integrate this as a continuous-coordinate localization head on top of a neural encoder. Let the encoder map an input into a matrix $Y\in\mathbb C^{M\times B}$, where $M$ is the number of sampled spatial or angular channels and $B$ is the number of examples or views. Compute the rank-$s$ left singular vectors $U_s$ of $Y$, optionally stopping gradients through the SVD during the first experiment, and define $P=U_sU_s^*$. Use a coordinate feature map with entries $\varphi_z[m]=\exp(i\kappa\omega_m^\top z)/\sqrt M$, or replace it by a learned normalized Fourier-feature map. For every candidate coordinate, compute $q(z)=1-\|U_s^*\varphi_z\|_2^2$. Train the encoder with a ranking loss forcing ground-truth coordinates to have lower $q$ than negatives, and add augmentation that randomly rotates the estimated subspace. In pseudocode: Y=encoder(x); U=top_s_svd(Y); phi=exp(1j*kappa*Omega@z)/sqrt(M); q=1-squared_norm(U.conj().T@phi); loss=hinge(q(z_pos)+margin-q(z_neg)). Estimate subspace uncertainty from two independently augmented batches using $\|UU^*-VV^*\|_2$, approximated by power iteration; the paper supplies the projector-residual construction and perturbation interpretation, while the margin and rank are hyperparameters. Start with a small synthetic 2D or 3D point-source dataset, a four-layer MLP encoder, and 64 Fourier channels. Compare against cosine similarity and an attention coordinate head under feature corruption and adversarial subspace rotation. Success means lower localization error and graceful degradation proportional to the measured projector distance at equal encoder size.

Verification

Mechanism failed

Stage 1 · Toy mechanism gate: Failed ✗

Stage 2 · Mechanism transferred to benchmark: Not tested

Stage 2 · Practical benchmark result: Not run

Methodology: Toy-system gate first; the benchmark stage runs only after a pass. How verification works

Stage 1 — Mechanism check agent confidence 9/10

Built a Fourier-feature spherical MUSIC residual localizer and numerically verified the core stability claim: for all tested adversarial rotations, the score perturbation stayed below the measured projector distance. However, the toy localization comparison showed no advantage over the cosine control at epsilon=0.20, and the residual method was worse at epsilon=0.50, so the promised localization win was not observed.

Agent confidence: 9/10

Baseline: Cosine-style control: mean grid localization error 0.00667 at projector distance 0.20; 0.00667 at 0.50.
Idea: Residual objective: mean grid localization error 0.00667 at projector distance 0.20; 0.01079 at 0.50. Maximum score perturbations for distances 0.02, 0.05, 0.10, 0.20, 0.35 were 0.00426, 0.00880, 0.02071, 0.06770, 0.11789, respectively, all within the bound.

Limitations: This was a small synthetic, non-neural experiment with a simple max-single-direction cosine control, no trained encoder, no ranking-loss training, no feature corruption/noise study, and a coarse coordinate grid. The stability check used finite random steering vectors rather than a uniform exhaustive search.

How to run: python3 experiment.py > results.json

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Artifacts

Implementation overview ⬇ Download all as ZIP 3 files · code, reports and structured results

🐍 experiment.py 4.4 KB View Raw
📄 report.md 1.4 KB View
📊 results.json 3.0 KB View Raw