---
title: Specular Beamforming Overview
url: https://www.emergentmind.com/topics/specular-beamforming
type: topic
---

# Specular Beamforming Overview

Searching arXiv for the cited papers to ground the article in current records.
Specular beamforming denotes beamforming procedures that explicitly encode ray-consistent or mirror-law-consistent propagation rather than treating the measured field as an undifferentiated superposition. In the cited literature, the term covers two closely related but technically distinct constructions. In an acoustic waveguide, the receive-aperture field is decomposed into a small number of stable components, or “eigenbeams,” formed by narrow bundles of rays; these components support both a generalized matched-field processor and a transmit “acoustic searchlight” that illuminates one path at a time [2008.06116]. In ultrasound imaging, specular receive beamforming is formulated from Snell’s law for planar reflectors, and later extended to refraction-corrected reconstruction of cortical bone interfaces, where the objective is to enhance interface visibility while rejecting diffuse scattering and speckle [2107.08069], [2507.08497].

## 1. Stable components and the eigenbeam viewpoint

In the waveguide formulation, a point source at \((r_0,z_0)\) excites on a vertical line array at \(r=0\) a total complex field
\[
u(z)=\sum_{n=1}^N u_n(z),
\]
where each \(u_n(z)\) is the contribution of the \(n\)th eigenbeam, defined as the bundle of rays launched in a narrow angular window \(\chi_0\in[\chi'_{0,n},\chi''_{0,n}]\) [2008.06116]. In the geometric-optics approximation,
\[
u_n(z)\approx\int_{\chi'_{0,n}}^{\chi''_{0,n}} A(\chi_0)\,\exp\!\bigl[i k\,S(\chi_0)\bigr]\,d\chi_0,
\qquad
k=\frac{2\pi f}{c_0},
\]
with \(S(\chi_0)\) the eikonal of the ray launched at \(\chi_0\) and \(A(\chi_0)\) the ray amplitude at \(r=0\).

A component is called stable when the ray-bundle width \(\Delta\chi_n=\chi''_{0,n}-\chi'_{0,n}\) is so small that, under a weak range-dependent perturbation \(\delta c(r,z)\), all rays in the bundle acquire almost the same additional phase, \(k\,\delta S(\chi_0)\approx\phi_n\). In that case the measured field takes the form
\[
v(z)=\sum_{n=1}^N \gamma_n\,u_n(z),
\qquad
\gamma_n=e^{\,i\phi_n}.
\]
The requirement that \(N\) remain small, typically \(2\)–\(5\), is central: it ensures that each beam is narrow and that the representation isolates a few physically meaningful propagation branches.

The same paper gives a more systematic extraction based on the coherent-state expansion. In phase-space coordinates \(\mu=(P,Z)\), with \(P=\nu(r,z)\sin\chi\), the Gaussian beam
\[
Y_\mu(z)=\frac{1}{\sqrt{\Delta_z}}
\exp\!\Bigl[i k\,P\,(z-Z)-\tfrac{\pi}{2\Delta_z^2}(z-Z)^2\Bigr]
\]
is used to define fuzzy segments \(\sigma_n\) of the ray line in \((P,Z)\) space. One then obtains
\[
u_n(z)=\lambda^{-1}\!\int_{\sigma_n}\!a_\mu\,Y_\mu(z)\,d\mu,
\qquad
a_\mu=\int u(z')\,Y_\mu^*(z')\,dz',
\qquad
\lambda=\frac{2\pi}{k},
\]
or, equivalently, a projection operator \(Q_n\) such that \(u_n=Q_nu\) in discrete form. This formulation makes the specular structure explicit at the level of subspace decomposition rather than at the level of the full field.

## 2. Generalized matched-field processing on the stable-component subspace

Traditional matched-field processing localizes a source by maximizing
\[
K_0(\theta)=\frac{|u(\theta)^H v|^2}{\|u(\theta)\|^2\,\|v\|^2}
=\frac{v^H P_0 v}{v^H v},
\qquad
P_0=\frac{u\,u^H}{u^H u}.
\]
In a mismodeled environment, however, the peak of \(K_0\) can shift far from the true \(\theta\). The stable-component approach replaces direct comparison of the full calculated and measured fields by comparison of their stable-component subspaces [2008.06116].

Let \(W=[u_1,\dots,u_N]\) be the \(N_a\times N\) matrix whose columns are the calculated stable components for trial \(\theta\). With the economy-size singular-value decomposition
\[
W=\sum_{m=1}^N \alpha_m\,\xi_m\,\eta_m^H,
\]
the projector onto \(\mathrm{span}\{u_n\}\) is
\[
P=\sum_{m=1}^N \xi_m\,\xi_m^H.
\]
The generalized similarity is then
\[
K(\theta)=\frac{v^H P v}{v^H v},
\]
and source localization proceeds by maximizing \(K(\theta)\). Because \(P\) only “sees” the \(N\)-dimensional stable-component subspace, this generalized matched-field processing is far less sensitive to errors outside those \(N\) beams.

The formulation is accompanied by explicit assumptions and approximations: \(u_n^H u_m\approx0\) for \(n\neq m\), perturbation-induced phases \(\phi_n\) are random and independent, and the model \(v\approx\sum\gamma_nu_n\) is accurate at the true \(\theta\). In the model problem reported in the paper—deep-water waveguide, \(500\) Hz CW, array length \(250\) m, \(\delta c_{\rm rms}=0.25\,\)m/s, correlation scales \(\ell_r=5\) km and \(\ell_z=0.5\) km, with \(N=3\) eigenbeams—traditional similarity at the true source gave \(\overline{K_0}\approx0.3\) with large scatter over \(40\) realizations, whereas the generalized criterion gave \(\overline{K}\approx0.7\), tightly clustered near unity. In single-realization uncertainty surfaces, \(K_0(r_s,z_s)\) was fractured into many local maxima, while \(K(r_s,z_s)\) remained a single smooth peak near the true \((r_0,z_0)\).

## 3. Continuous-wave transmission, pulsed arrivals, and path-selective illumination

The same stable-component formalism yields an explicit transmit design. To emit a narrow continuous-wave beam that travels along the \(n\)th eigenbeam, the aperture excitation is the phase conjugate of the received stable component,
\[
u_{0,n}(z)=\bigl[u_n(z)\bigr]^*\,B(z),
\]
where \(B(z)\) is a smooth window taper vanishing at the array ends. In discrete form, for sensor depths \(z_i\),
\[
w_n=\bigl[u_n(z_1)^*,u_n(z_2)^*,\dots,u_n(z_{N_a})^*\bigr]^T
\odot
\bigl[B(z_1),\dots,B(z_{N_a})\bigr]^T,
\]
followed by normalization, for example \(\|w_n\|=1\) [2008.06116].

Algorithmically, the procedure is: ray tracing from the nominal focus \((r_0,z_0)\), identification of \(N\) disjoint angular intervals corresponding to eigenbeams, computation of \(\{u_n(z)\}\) either geometrically or through coherent-state projection, formation of the weight vectors \(w_n\), and continuous-wave emission through the array. The resulting field forms, to leading order, a narrow beam traveling along eigenbeam \(n\). The paper’s numerical comparison is explicit: driving the array with the phase-conjugate of the full field \(u^*(z)\) focuses at \((r_0,z_0)\) but launches three overlapping beams along all eigenrays, whereas driving with each stable component \(u_n^*(z)\) individually produces one narrow beam that follows exactly the central ray of eigenbeam \(n\).

For a pulsed source \(s(t)\), the receive data \(V(z,t)\) can be Fourier transformed to \(v(z,f)\), projected onto each \(u_n(z,f)\), and inverted back to time to form \(G_n(t)\), the pulse arriving via eigenbeam \(n\). In a perturbed environment the peak is delayed by
\[
\delta t_n=\frac{\delta S_n}{c_0},
\]
and the set \(\{\delta t_n\}\) may be used as travel-time constraints in a tomographic inverse problem to refine \(c(r,z)\) or source position. This makes specular decomposition not only a beamforming device but also an inversion primitive.

## 4. Specular receive beamforming for planar reflectors in ultrasound

In the ultrasound comparison study, specular beamforming is defined against three established receive beamformers: delay-and-sum (DAS), filtered delay-multiply-and-sum (DMAS), and minimum-variance (MV) [2107.08069]. DAS assumes a locally homogeneous, diffuse-scattering medium and applies geometry-driven delays and static apodization. DMAS is a non-linear coherence-enhancing beamformer based on pairwise multiplication of delayed signals followed by band-pass filtering for the second harmonic. MV minimizes output power subject to unit gain in the look direction and depends strongly on the choice of subarray length \(L_s\).

Specular beamforming is instead built on Snell’s law for planar reflectors. For a pixel \(P\) and assumed reflector orientation \(\alpha_g\),
\[
y_{\rm SB}(P,\alpha_g)
=\sum_{j=1}^T s\!\Bigl(\tau_P(\alpha_j,\alpha_r)\Bigr),
\qquad
\alpha_r=\alpha_j-2\alpha_g,
\]
where \(\alpha_j\) is the \(j\)th plane-wave transmit angle, \(\alpha_r\) is the specular receive angle, and \(\tau_P(\alpha_j,\alpha_r)\) is the one-way travel time from the virtual reflection point on the reflector to the transducer elements after applying both transmit and receive delays. The method can be extended by correlating \(y_{\rm SB}\) with a pre-computed matched filter \(h(\alpha_j,\alpha_g)\), and the displayed image can be formed from \(\max_{\alpha_g} y_{\rm SB}(P,\alpha_g)\) or its matched-filter analogue.

The reported quantitative evaluation uses contrast ratio and generalized contrast-to-noise ratio. At reflector angles \(\alpha_g=20^\circ,30^\circ\), DAS yields negative or low contrast ratio (\(-0.8\) dB, \(-3.4\) dB) and \(\mathrm{gCNR}\sim0.5\); DMAS and MV improve both metrics, with contrast ratio \(8\)–\(18\) dB and \(\mathrm{gCNR}\) \(0.7\)–\(0.99\); and SB achieves the highest contrast ratio, approximately \(20\)–\(27\) dB, with \(\mathrm{gCNR}\sim0.98\)–\(0.99\). At depths \(4.3\) mm and \(17.5\) mm, DMAS degrades at \(4.3\) mm because only a few elements receive specular energy, but recovers at \(17.5\) mm as more channels contribute. DAS remains competitive when reflectors are near-normal to the array, with contrast ratio approximately \(28\)–\(31\) dB and \(\mathrm{gCNR}=1\), while SB again reaches approximately \(30\)–\(33\) dB and \(\mathrm{gCNR}=1\).

These results situate specular beamforming as an application-tailored receive model rather than a generic replacement for diffuse-medium beamforming. In this comparison it is best at purely planar specular structures, but it suppresses all non-specular components, so soft-tissue features may disappear.

## 5. Refraction-corrected specular beamforming for cortical bone

The cortical-bone extension models a two-layer geometry with a soft-tissue layer of speed \(c_1\) overlying cortical bone of speed \(c_2\), an external interface \(D_e\) approximated by \(z=a_0+a_1x+a_2x^2\), and an internal reflector \(D_i\) given by \(z=b_0+b_1x+b_2x^2\) [2507.08497]. Snell’s law is enforced at the tissue–bone boundary and the specular law is enforced at the internal interface:
\[
\sin(\alpha_t+\phi_e)/c_1=\sin(\gamma_t+\phi_e)/c_2,
\qquad
\gamma_t+\gamma_r=2\theta_l,
\]
with \(\phi_e=\arctan(2a_2x_J+a_1)\) and \(\theta_l=-\arctan(2b_2x_Q+b_1)\).

For each transmit–receive pair \((i_t,i_r)\) and each image point \(P\), the method finds interface points \(J\) and \(K\) on \(D_e\) and a mirror point \(Q\) on \(D_i\) satisfying Snell’s and specular laws. The refraction-corrected two-way travel time is
\[
\sigma_{i_t,i_r}(P)
=
\frac{|P_t-J|+|P_r-K|}{c_1}
+
\frac{|J-P|+|K-P|}{c_2}.
\]
Delayed echoes are mapped into the specular domain through
\[
f(\beta;P)=\sum_{i_t,i_r} S(\sigma_{i_t,i_r}(P),\gamma_t,\gamma_r)
\Big|_{(\gamma_t+\gamma_r)/2=\beta}.
\]
A model-based matched filter \(h(\beta;P,b_1,b_2)\) is then computed, and the normalized cross-correlation
\[
\chi(\theta_l,b_2;P)
=
\frac{\int f(\beta;P)\,h_0(\beta;P,0,b_2)\,d\beta}
{\sqrt{\int f^2\,\int h_0^2}}
\]
yields a specularity index \(\Psi(P)=\max_{\theta_l,b_2}|\chi|\) and a best-fit local orientation \(\tilde{\theta}_l(P)\). The final image is
\[
I_{\rm SP}(P)=\Psi(P)\cdot\sum_{\beta} w(\beta;\tilde{\theta}_l(P))\,f(\beta;P),
\]
where \(w\) is a Hann window of half-width \(\eta\cdot\pi/2\).

Implementation details are explicit: a \(2.5\) MHz phased array with \(96\)–\(128\) elements and pitch approximately \(0.295\) mm on a fully programmable Vantage system recorded a \(96\times96\) synthetic aperture data set, element-by-element; DAS base images used \(f\)-number \(=0.5\); and subject-specific sound speeds were estimated by autofocus in vivo and from the head-wave in ex vivo water-coupled scans. The endosteal interface contrast metric is
\[
C_{EI}=\mu_E/\mu_I,
\qquad
C_{EI}({\rm dB})=20\log_{10}(\mu_E/\mu_I).
\]
In vivo, specular beamforming improved \(C_{EI}\) by \(1\) to \(13\) dB while maintaining the relative contrast between the outer and inner surfaces of the cortex; ex vivo on elderly femurs with porosity \(5\)–\(16\%\), the mean gain was \(1.1\)–\(7.6\) dB depending on subvolume and sample. The reported specularity maps gave \(\Psi>0.7\) at periosteum and \(\Psi>0.5\) at endosteum.

## 6. Comparative interpretation, limitations, and scope

Across these formulations, specular beamforming is not a single algorithm but a family of physics-constrained beamforming constructions. In the waveguide case, the constraint is that only a few dominant eigenbeams should be compared or excited; in ultrasound, the constraint is that the receive path should satisfy the mirror-law geometry of a planar or curved interface, possibly with refraction. This suggests a unifying view in which specular beamforming replaces full-field matching by matching on a geometrically admissible subset of propagation paths.

The practical advantages are domain-specific and explicitly delimited in the cited work. In multipath acoustics, generalized matched-field processing becomes less sensitive to inevitable inaccuracies of the environmental model, and stable-component phase conjugation provides a specular beamformer that can illuminate one path at a time [2008.06116]. In ultrasound of planar reflectors, specular beamforming achieves the highest contrast ratio and generalized contrast-to-noise ratio in the reported angulation and depth experiments, but it suppresses non-specular diffuse components and therefore may omit surrounding tissue features [2107.08069]. In cortical bone imaging, explicit modeling of Snell’s law and refraction enhances the visibility of the endosteal interface and reduces speckle from intracortical pores, yet the planar-reflector assumption underestimates curvature in highly curved anatomy and the full curved-model increases computational load by approximately \(100\times\) versus DAS [2507.08497].

Several distinctions follow directly. Specular beamforming should not be conflated with conventional DAS plus altered apodization: the cited ultrasound formulation changes the delay law itself through \(\alpha_r=\alpha_j-2\alpha_g\) and admits a matched-filter extension. It also should not be conflated with phase conjugation of the entire received field: in the waveguide results, full-field phase conjugation launches overlapping beams along all eigenrays, whereas stable-component phase conjugation launches a single narrow beam along one prescribed eigenbeam. Finally, the cited literature consistently frames specular beamforming as application-tailored. It is most effective when the dominant physics is specular reflection or stable ray-bundle propagation, and less appropriate when the objective is to preserve diffuse scattering or soft-tissue texture as primary image content.

Source: https://www.emergentmind.com/topics/specular-beamforming