Acoustic Causal Tracing: Techniques and Applications

• Acoustic causal tracing is a methodological approach that infers, analyzes, and enforces causality in acoustic systems using physical sum rules and time-domain constraints.
• It employs techniques ranging from geometric ray tracing and caustic analysis to causal attention in transformer-based neural networks, ensuring robust and interpretable audio models.
• The approach spans applications such as spectral shaping of underwater resonators and privacy-preserving contact detection through controlled acoustic signaling.

Acoustic causal tracing encompasses a class of methodologies for inferring, analyzing, or enforcing the directionality, stability, or partition of causal effects within acoustic systems, both for physical modeling of wave propagation and for high-level sequence modeling of audio data. Approaches span fundamental physical sum rules constraining scattering bandwidths, geometric tracing of ray stability and caustic formation, strictly causal models for time reversal in attenuating media, transformer-based neural architectures for causal relevance in audio, and system-level tracing via acoustic signals for privacy-preserving contact detection.

1. Foundational Principles of Causal Constraints in Acoustics

Causality, as a physical and mathematical principle, mandates that acoustic system responses cannot precede their excitation. In frequency-domain scattering, this requirement enforces analytic properties on the transmission and extinction coefficients. In high-dimensional acoustic modeling (e.g., audio signal processing), causality ensures any representation, attribution, or classification is grounded only on past or present evidence, never future observations.

A key principle is the acoustic sum rule (Baldin-type) for scattering, which states that the integrated extinction cross-section, weighted by the inverse square of frequency, is “locked” to static effective mass and stiffness parameters of the scatterer. Causal attention mechanisms in neural sequence models, by analogy, restrict transformations to rely only on temporally preceding tokens to prevent leakage of future information and enforce interpretability in terms of causal acoustic attributions (Qu et al., 6 Jan 2026, Owino et al., 18 Dec 2025).

2. Universal Causal Sum Rule and Scattering Resource Tracing

The universal acoustic Baldin sum rule, rigorously derived in the scattering context for a one-dimensional subwavelength scatterer of thickness $L$ in background fluid, provides a quantitative “ledger” for the allocation of scattering resources across frequency:

$$\int_0^{\infty} \frac{\sigma_{\mathrm{ext}}(\omega)}{\omega^2}\,d\omega = \Gamma$$

where

$$\Gamma = \frac{\pi L}{2c_0} \left( \frac{K_0}{M_{\mathrm{eff}}(0)} + \frac{\rho_{\mathrm{eff}}(0)}{\rho_0} \right)$$

This relation, derived by exploiting the Kramers–Kronig relations and optical theorem, imposes a global constraint: no passive (i.e., causal and linear) acoustic scatterer can exceed the scattering bandwidth dictated by its static (monopole and dipole) parameters. Practical causal tracing consists of evaluating the partial sum

$$\gamma(\omega) = \int_0^{\omega} \frac{\sigma_{\mathrm{ext}}(\omega')}{\omega'^2} d\omega'$$

allowing an experimenter or designer to quantitatively track how “scattering resources” are distributed up to any frequency (Qu et al., 6 Jan 2026).

Experimentally, this sum rule is confirmed for both monopole and dipole underwater resonators, including complex structures with multiple elastic and Fabry–Pérot modes. The rule enables spectral shaping: by tuning the static impedance to match the background (i.e., $\Gamma_m = \Gamma_d$), one can spectrally broaden transmission loss—a principle validated with Fano-type resonator experiments.

3. Geometric Ray Tracing and Causal Caustic Analysis

Causal tracing at the geometric (ray) level addresses the stability of acoustic rays and the formation of caustics in inhomogeneous, potentially moving, layered media. The propagation of rays derived from the eikonal approximation can be framed as a geodesic flow in an effective (pseudo-)Riemannian manifold.

The location and structure of caustics correspond to conjugate points of the geodesic flow, identified by the vanishing of the Jacobi (deviation) field:

$\frac{d^2Y}{ds^2} + K(s) Y = 0$

where $K(s)$ is the sectional curvature derived from local sound speed and flow gradients. The causal tracing algorithm in layered media proceeds by integrating ray and deviation (Jacobi) equations, matching at interfaces, and locating zeros of the Jacobi field corresponding to caustic points. Near caustics, amplitude corrections are made using canonical diffraction integrals (Airy-type), ensuring that traced field intensity remains physically meaningful (Bergman, 2015).

This methodology provides principled, geometric diagnostics for ray stability, caustic localization, and their dependence on both smooth and layered environmental structure.

4. Causal Models in Attenuating Acoustic Media and Time-Reversal Imaging

Strongly causal models of attenuated wave propagation ensure finite signal front speed and adherence to the principle that no field arises prior to excitation (i.e., $p^\mathrm{a}(x,t)=0$ for $t<0$). Key models include the KSB (fractional attenuation) and NSW (one-relaxation) models, parameterized by fractional exponents or differential relaxation times.

Time-reversal imaging in such models requires adaptation: after observing boundary data, one reconstructs the initial pressure by solving an appropriately modified (attenuation-aware) time-reversed PDE. The correction is constructed by asymptotic expansions:

$$K(\omega) = \omega + \alpha A_1(\omega) + O(\alpha^2)$$

$$\Gamma^\mathrm{a}(\omega) = \Gamma(\omega) + \alpha\,\Delta\Gamma_1(\omega) + O(\alpha^2)$$

and the imaging functional is corrected perturbatively, ensuring convergence to the true initial state as the order of expansion increases. This methodology enables causal tracing of energy and information flow in attenuating acoustic systems, guaranteeing that inversion and imaging respect causal physical constraints (Kalimeris et al., 2012).

5. Causal Tracing in Deep Audio Modeling

In neural sequence models for audio, such as the DACH-TIC architecture, causal tracing is implemented through architectural and training constraints designed to enforce and surface the causal structure of acoustic signals.

• Causal attention masking: All self-attention in transformers is strictly lower-triangular, ensuring that token $i$ can only attend to tokens $j\leq i$, reflecting a “no-future-peeking” causal constraint:

$$a_{ij} = \begin{cases} \frac{\exp(q_i^\top k_j/\sqrt{d})}{\sum_{j' \leq i} \exp(q_i^\top k_{j'}/\sqrt{d})} & \text{if } j \leq i \ 0 & \text{if } j > i \end{cases}$$

with strict lower-triangular masking in the pre-softmax scores.

• Controlled perturbation (“pseudo-intervention”) training: Forcing the model to ignore spurious correlational cues by generating perturbed acoustic counterfactuals (e.g., pitch shifts, local energy suppression) and penalizing output deviation between original and perturbed inputs.
• Multi-task supervision for token-wise causal relevance: Outputs a “salience map” that explicitly estimates token-level causal relevance, making the model’s attributions interpretable and aligned with domain priors.
• Domain-adversarial objective: A gradient reversal layer (GRL) coupled with a domain classifier promotes invariance to environmental or domain artifacts, ensuring that learned causal attributions are not confounded by spurious correlations induced by recording conditions.

Performance measures such as the Counterfactual Stability Score (CSS), Causal Fidelity Index (CFI), and ablation studies confirm that each of these causal tracing components substantively contributes to robustness, interpretability, and domain generalization (Owino et al., 18 Dec 2025).

6. Privacy-Preserving Acoustic Tracing Systems

Acoustic-based contact tracing systems, such as Acoustic-Turf, operationalize causal tracing on a system level by employing controlled, ephemeral ultrasonic signal broadcasts for encounter detection. Devices exchange randomized, per-minute ultrasonic ID packets; presence is inferred by detecting these packets within a time window, and causal contact is determined strictly by acoustic signal presence—ensuring that only truly physically proximal encounters register as “contacts.”

Detection algorithms exploit the recent acoustic presence (not future or far-past) of peers, aligning system operation with temporal and spatial causality dictated by the propagation loss and absorption of high-frequency sound in air and through obstacles (Luo et al., 2020). Privacy is assured through unlinkable ID generation and local-only matching, avoiding correlation or leakage of sensitive historical contact chains.

7. Implications and Generalization of Acoustic Causal Tracing

The evolving landscape of acoustic causal tracing reveals a unifying structure:

• In physical wave systems, sum rules and geodesic diagnostics provide quantitatively exact, causality-driven constraints linking static properties to frequency-dependent response, enabling optimal design within strict physical bounds (Qu et al., 6 Jan 2026, Bergman, 2015).
• In signal processing and machine learning, causal attention, perturbation-based augmentation, and pseudo-interventions systematically enforce causal interpretability and robustness, with direct applications in clinical, affective, and environmental audio analysis (Owino et al., 18 Dec 2025).
• At the systems level, acoustic tracing mechanisms informed by physical propagation constraints support privacy-preserving proximity inference, outperforming wireless correlational approaches and naturally limiting tracing to physically meaningful causal connections (Luo et al., 2020).

The principled integration of causal tracing—from theory to implementation—enables robust, interpretable, and physically grounded advancement across acoustic science and engineering domains.