Frequency-Invariant Beamforming Overview

• Frequency-invariant beamforming is a method that designs sensor arrays to maintain a consistent spatial response across a broad frequency range.
• Techniques such as polynomial filter design, modal matching, and convex optimization are employed to achieve stable main-lobe and sidelobe structures.
• Applications include robot audition, spatial audio capture, and wireless communications, where preserving consistent directivity is critical.

Frequency-invariant beamforming refers to the design and synthesis of sensor array or transducer array beamformers whose spatial response—the beampattern—remains effectively constant over a specified frequency range. This property is critical in applications where wideband signals are encountered or where consistent directivity over frequency supports robust source separation, dereverberation, spatial audio capture, robot audition, or wireless communications.

1. Fundamental Principles of Frequency-Invariant Beamforming

The core objective in frequency-invariant (FI) beamforming is to produce a spatial filter whose main-lobe direction, width, and side-lobe structure do not vary with frequency within the design band. This is in contrast to classical time-delay or phase-shift beamformers whose beampatterns contract, shift, or develop distortion as the frequency changes, due to wavelength scaling of spatial harmonics.

Frequency invariance is typically realized by replacing frequency-independent delay-based weights with frequency-dependent filters or by imposing optimization constraints across frequency samples. The technical challenge is to enforce the invariance constraint while achieving desired robustness (e.g., distortionless response, white noise gain bounds) and satisfying physical or implementation constraints such as geometry, causality, or filter length.

2. Canonical FI Beamformer Architectures

2.1 Robust Least-Squares FI Polynomial Beamforming

In the HRTF-based robust least-squares frequency-invariant polynomial (RLSFIP) beamforming paradigm (Barfuss et al., 2016), a microphone array is equipped with frequency-dependent filter-and-sum branches, and the final output is a polynomial (in a normalized steering parameter $D$) combination of the branch outputs.

Let $\{W_{n,p}(\omega)\}$ denote the frequency-domain weights for microphone $n$ and polynomial order $p$. The total response for a generic steering value is

$$Y_D(\omega) = \sum_{n=0}^{N-1}\sum_{p=0}^{P} D^p \, W_{n,p}(\omega) X_n(\omega)$$

where $X_n(\omega)$ is the $n$th sensor input. By fitting the polynomial beamformer’s array output at a discrete set of prototype look directions (PLDs) jointly across a frequency grid via convex optimization—with constraints enforcing distortionless response and a lower bound on the white-noise gain (WNG)—the designer obtains frequency-invariant beampatterns. Runtime steering is achieved by simply changing the scalar parameter $D$.

A notable feature of this architecture is the direct incorporation of head-related transfer functions (HRTFs) into the steering vectors, capturing the effect of complex scattering geometries (e.g., a robotic head) and improving speech recognition performance compared to free-field models.

2.2 Differential and Modal-Matched Approaches

Small-scale arrays, as found in differential microphone arrays (DMAs), can achieve frequency-invariant responses by differencing closely spaced sensors. For a first-order DMA, differencing two spatially sampled signals and normalizing achieves a $\cos\theta$ directivity pattern that is independent of frequency in the low-frequency (diffraction) regime. General higher-order and arbitrary beampatterns can be synthesized via circular-harmonic modal matching, representing the desired FI beampattern as a truncated Fourier (circular-harmonic) series and matching it to the array’s spatial response. This approach generalizes to planar arrays of first-order or higher-order directional elements and accommodates arbitrary sensor geometries and orientations (Miotello et al., 17 Aug 2025).

2.3 FIR-Weighted and Spherical Harmonic Domain FI Designs

FI beamformers can also be realized by decomposing spatial filtering into modal (spherical harmonics) components and applying frequency-dependent weightings constructed to produce frequency-flat main beams and controlled side lobes. In the time-domain nearfield FI approach, the beamformer is derived as

$${\mathsf W}_{u,v}(\omega)\;=\;\alpha_{u,v}\;\frac{1}{i\omega\tau_s}\;\left(-\,\frac{\tau_s}{\tau_{\hat d}}\right) e^{-i\omega(\tau_{\hat d}-\tau_s)}\frac{1}{h_u(\omega\tau_{\hat d})} Y_{u,v}(\theta_{\hat d},\varphi_{\hat d})$$

enforcing frequency-invariant gain at the desired focus point and flexibility in array configuration (Ma et al., 2021).

2.4 Frequency-Regularized Optimization

For loudspeaker arrays exhibiting heterogeneous transducer characteristics, frequency-invariant constant directivity is imposed as a quadratic equality (generalized Rayleigh quotient, GDI) across frequency, with frequency-regularization to ensure handoff between overlapping passbands. Two designs—maximum efficiency constant directivity (MECD) and maximum sensitivity constant directivity (MSCD)—are formulated as quadratic programs with closed-form or fast iterative solutions, maintaining flat directivity index (DI) across the operational band (Luo, 2024).

3. Frequency-Invariant Beamforming Methodologies

A diversity of direct synthesis and optimization frameworks have been advanced:

• Polynomial/FIR filter design: Weights are constrained to polynomial or FIR filter forms, tuned to enforce spatial invariance at sampled frequencies and directions (as in RLSFIP and FIR spatial filtering).
• Modal matching: Beampatterns are projected onto basis functions (e.g., circular or spherical harmonics), and modal weights are computed via pseudoinverse, ensuring the sum matches a frequency-independent target pattern (Miotello et al., 17 Aug 2025, Ma et al., 2021).
• Coordinate transformations: Spherical-to-planar mappings enable closed-form, IDFT-based design for arbitrary FI beam shapes on uniform planar arrays, replacing iterative optimization (Son, 2020).
• Null constraints and convex QP: Arrays combining omnidirectional and directional microphones permit FI beampatterns by imposing multi-constraint QPs that guarantee specified main lobe, nulls, and WNG for robustness (Zhang et al., 25 Aug 2025).
• Wideband surface-based approaches: RIS (reconfigurable intelligent surfaces) are treated via space-frequency transformation and stationary phase approximation, deriving closed-form phase-shifting laws to deliver flat beampatterns across wide bandwidths without frequency-dependent hardware (Qian et al., 2024).
• Data-driven autograd optimization: For 2D (azimuth and elevation) FI beamforming with complex array geometries, automatic differentiation and multi-frequency penalties are used to minimize beamwidth and out-of-beam power while constraining invariance metrics across frequency (Ortigoso-Narro et al., 24 Nov 2025).
• Smoothness-regularized per-tone designs: In multicarrier (FBMC/OQAM) wireless systems, per-tone SVD-based beamforming is regularized via phase-aligned smoothing or orthogonal iteration to ensure beamformer continuity and preserve system orthogonality (Qiu et al., 2018).

4. Performance, Robustness, and Implementation Aspects

Frequency-invariant beamformers are characterized by their ability to maintain:

• Consistent main-lobe width and pointing across wide frequency ranges, with beampattern variance typically within 1–2 dB (or less) in the main lobe.
• Robustness to array errors (e.g., gain/phase uncertainties, sensor directivity mismatch), as enforced by WNG or DI constraints.
• Preservation of system-level goals, such as flat ASR word error rate (ASR-WER in robot audition), constant directivity for sound reinforcement, or uniform communication-rate and resolution in wireless and radar applications.
• Computational efficiency where possible: Fast closed-form or analytic methods (IDFT, modal pseudoinverse, polynomial interpolation) replace high-dimensional iterated optimization, permitting real-time or high-throughput synthesis (Son, 2020, Qian et al., 2024).
• Design flexibility and geometry agnosticism: Modal-domain (spherical/circular harmonics) and polynomial methods fully decouple the FI constraint from sensor placement, enabling adaptation to arbitrary layouts or robot heads (Ma et al., 2021, Miotello et al., 17 Aug 2025).

Experiments demonstrate that FI beamformers can provide nearly identical beampatterns across frequency samples, avoid spatial aliasing and main-lobe distortion, and maintain distortionless responses in the look direction while suppressing energy at prescribed nulls. In practical robot audition, FI beamformers using HRTF-based steering vectors outperform free-field models by up to 15% absolute ASR-WER (Barfuss et al., 2016). Similarly, frequency-invariant loudspeaker array designs with frequency regularization realize perfectly flat DI over octaves and prevent destructive interference at crossover (Luo, 2024).

5. Extensions and Special Cases

• RIS and Spatial-Chirp Approaches: The wideband challenge in RIS-assisted systems is addressed by precomputing spatial phase laws via stationary phase, mapping the desired spectral response to the spatial domain; this obviates the need for hardware-based frequency-dependent delays (Qian et al., 2024). InFocus, for massive phased arrays, embeds a spatial FMCW chirp in the aperture phase to achieve frequency-invariant focusing and mitigate beam squint and misfocus (Myers et al., 2020).
• Nearfield and 3D Control: Time-domain modal beamformers enable frequency-invariant nearfield focusing, not only discriminating sources in angle but also in range. The modal-based approach decouples the FI design from array specifics, allowing for arbitrary reconfiguration (Ma et al., 2021).
• Arbitrary Beampatterns and Null Specification: Recent methods permit direct null placement and prescribed mainlobe control independent of analytic pattern formulas, useful for complex microphone arrangements or mixed omni/directional arrays, as in null-constraint optimization (Zhang et al., 25 Aug 2025).
• Elevation Frequency Invariance: 2D geometries such as concentric circular arrays, optimized via multi-frequency autograd, enable frequency-invariant beampatterns not just in azimuth (as with most line arrays) but also in elevation (Ortigoso-Narro et al., 24 Nov 2025).

6. Comparative Analysis and Application-Specific Considerations

The table summarizes representative architectures and their key properties:

Method/Class	Key Mechanism	Notable Properties
HRTF-based RLSFIP	Polynomial branch+LS	Runtime steering, HRTF robustness, ASR-WER
Modal-matching (DMA, DLA)	Circular/spherical harmonics LS	Arbitrary layout, FI via order selection
FIR/IDFT-based (planar)	Spherical→planar mapping + IDFT	Arbitrary beampatterns, O(N^2 log N) comp.
Null-constrained QP	QP with distortion/nul/WNG constr	Robust, steerable, prescribed nulls
RIS/SFM-chirp (InFocus)	Fourier/Stationary Phase Law	Hardware simplicity, flat wideband gains
Autograd multi-freq	PyTorch, RProp/diff constraints	Full 2D FI, fine beamwidth control
SVD + Smoothness (FBMC)	Per-tone SVD + alignment	Wireless MIMO, preserves orthogonality

The choice of methodology is dictated by the sensor/transducer hardware, spatial aperture, desired angular discrimination, frequency band, processing latency, and robustness requirements of the target application.

7. Challenges and Directions

Despite considerable progress, several ongoing challenges persist:

• Sensor directivity and mutual coupling: Real elements exhibit frequency-dependent sensitivity, which must be modeled for accurate FI performance, particularly in small or heterogeneous arrays (Miotello et al., 17 Aug 2025).
• Robustness to modeling and environmental uncertainty: Accurate FI beamformers must tolerate errors in transfer function measurement (e.g., HRTF), array perturbations, or channel nonstationarities.
• Extension to non-uniform, adaptive, or sparse arrays: Closed-form FI solutions for irregular geometries or arrays with missing elements remain a topic of active investigation.
• Hardware constraints: Implementation on platforms with quantized phase shifters, constrained FIR filter length, or limited computational resources shapes the actual achievable invariance.

A plausible implication is that future FI beamforming solutions will increasingly integrate data-driven optimization, explicit physical modeling of non-ideal sensors, and application-specific adaptation to dynamic environments, especially in real-time embedded contexts.

---

Key references include HRTF-based RLSFIP design (Barfuss et al., 2016), modal matching for DMAs (Miotello et al., 17 Aug 2025), convex null-constrained differential arrays (Zhang et al., 25 Aug 2025), spatial-chirp wideband phased arrays (Myers et al., 2020), time-domain modal nearfield FI beamforming (Ma et al., 2021), fast IDFT-based broadband design (Son, 2020), and RIS space-frequency transformation (Qian et al., 2024).