Abstract
Analog signal processors have attracted a tremendous amount of attention recently, as they potentially offer much faster operation and lower power consumption than their digital versions. Yet, they are not preferable for large scale applications due to the considerable observational errors caused by their excessive sensitivity to environmental and structural variations. Here, we demonstrate both theoretically and experimentally the unique relevance of topological insulators for alleviating the unreliability of analog signal processors. In particular, we achieve an important signal processing task, namely resolution of linear differential equations, in an analog system that is protected by topology against large levels of disorder and geometrical perturbations. We believe that our strategy opens up large perspectives for a new generation of robust alloptical analog signal processors, which can now not only perform ultrafast, highthroughput, and power efficient signal processing tasks, but also compete with their digital counterparts in terms of reliability and flexibility.
Introduction
For a few decades, digital signal processors (DSPs) have widely replaced analog electronic and mechanical computers for carrying out computational tasks. Such processors not only offer highspeed operation, but also ensure reliability and flexibility of the processing^{1}, a property which has established DSPs as perfect candidates for realizing largescale computational systems.
While providing highspeed and reliable operation, DSPs also suffer from several fundamental drawbacks such as highpower consumption, costly analog to digital converters, and drastic performance degradation at high frequencies^{2}. Considering these limitations, it is neither reasonable nor affordable to use DSPs for performing specific, simple computational tasks such as differentiation or integration, equation solving, matrix inversion, edge detection and image processing. Therefore, the old idea of allanalog computing and signal processing has been recently revived, driven by the development of costefficient nanofabrication techniques and promising related advances in ultrafast optics. In their pioneer work, Silva et al.^{3} theoretically demonstrated the possibility of carrying out simple computational tasks such as convolution, differentiation and integration, making use of optical waves as they propagate through engineered metamaterial layers. By going beyond the aforementioned restrictions of DSPs, such a wavebased computational scheme then inspired numerous exciting applications including analog computing^{4,5,6,7,8,9,10,11,12,13,14,15}, signal processing, equation solving^{16, 17}, optical image processing^{18,19,20,21}, optical memories^{22}, and photonic neural networks^{23}. Not only are such types of analog signal processors (ASPs) real time and ultrafast due to their wavebased nature, but also they offer low power consumption and highthroughput operation as they are free of analog/digital conversion steps. In addition, they allow the unique possibility of carrying out different computational tasks in parallel, thereby significantly reducing the total processing time^{24,25,26,27,28,29,30}.
Despite their advantageous properties, ASPs still suffer from one important limitation compared to DSPs, which severely hinders their applicability for largescale applications: while repeating the same operation always gives rise to the same result when using DSPs (which is enabled by available errorfinding algorithms and protocols in digital systems), analog signal processing is often accompanied with considerable observational error caused by the extreme sensitivity of ASPs to changes in environmental and structural parameters^{31}.
Here, we demonstrate the possibility of drastically enhancing the reliability of ASPs by leveraging the unique immunity of topological insulators^{32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47} against imperfections, a muchsought feature which has established itself as a new paradigm for realizing a large variety of reliable devices such as lasers^{48}, modulators^{49}, and lenses^{50}. More specifically, we demonstrate a topological wavebased analog system that can solve linear differential equations of arbitrary order in timedomain at the speed of the wave and prove its strong immunity against geometrical flaws and environmental changes. Our findings provide exciting perspectives for a new generation of ultrafast, highthroughput, and highly reliable analog computing systems and signal processors.
Results
Topological analog equation solver
To illustrate the core idea of our proposal, let us consider an important signal processing task, namely the resolution of a firstorder linear differential equation with a nonzero arbitrary forcing g(t) (below, we will generalize the concept to differential equations of higher order). The goal is to design a twoport wave system that, regardless of the specific form of g(t) sent at the input, outputs a signal f(t) that would be the solution of a differential equation f′(t)+αf(t) = βg(t), where α and β are real coefficients. To this aim, one could consider a conventional resonator created by, for example, a onedimensional photonic band gap material with a defect in the middle (Fig. 1a). According to coupledmode theory, the transmission coefficient through such resonator near its resonance frequency f_{0} is given by^{51}
where we have used the time harmonic convention exp(j2πft), A is an arbitrary constant, and Q is the quality factor of the resonance. Now, if we consider as input signal the source term g(t) modulated at the carrier frequency f_{0}, i.e. \(\tilde g\left( t \right) = g\left( t \right){\mathrm{cos}}\left( {2\pi f_0t} \right)\), its relationship with the output \(\tilde f\left( t \right) = f\left( t \right){\mathrm{cos}}(2\pi f_0t)\) can be obtained via inverse Fourier transform of the transfer function (TF) H(f) of Eq. 1, leading to the desired firstorder differential equation f′(t)+αf(t) = βg(t), with α = πf_{0}/Q and β = 2πA. This analysis illustrates the possibility to realize an analog equation solver by engineering a resonator and measuring the envelope of its response to the known input signal g(t) (modulated at the resonance frequency). From this example, the main advantages of analog computation over digital one are evident: there is no need for converting the input signal g(t) to a digital stream (and vice versa), and the computation is being carried out in real time with no limitation at large values of f_{0}. However, this simplicity comes with a severe drawback: when adding geometrical imperfections to the system (Fig. 1b), the transfer function can be significantly disturbed by the creation of disorderinduced modes, shifting its spectrum and introducing new resonating peaks. This leads to an output signal which has nothing to do with the correct solution.
What we propose is instead to form the same transfer function, but out of resonant transmission through a topological edge mode, whose existence is guaranteed by the nontrivial topologies of the surrounding bulk insulators. Figure 1c depicts such a solution based on two insulating lattices with supposedly different topologies, inspired by the Su−Schrieffer−Heeger (SSH) scheme^{32}. Like the previous case, when the input signal g(t) (modulated at f_{0}) is applied to such system, the output signal envelope f(t) is equal to the solution of the desired differential equation. However, the topological equation solver can be immune to disorder, since the presence of a single midgap interface mode can be guaranteed by bulkedge correspondence^{52} (Fig. 1d).
To test this idea on a realistic system, we designed a topological firstorder linear differential equation solver for airborne audible acoustic signals (see Supplementary Note 1 for an electromagnetic equivalent). The topological ASP system is based on sonic topological insulators inspired by the SSH scheme, obtained from solid cylinders placed in a pipe of square crosssection, as in Fig. 1c, d (see Methods for geometrical details and topological invariant calculations). The topological interface is designed to provide a resonant mode at f_{0} = 2254 Hz, with A = 1 and Q = 0.5f_{0}, aiming at solving the differential equation f′(t)+2πf(t) = 2πg(t). The transfer function H(f) of the system, calculated by threedimensional fullwave finiteelement calculations in the frequency domain, is compared to the target transfer function in Fig. 2a (green and dashed curves in the middle inset), revealing their perfect agreement. Now, consider an input signal \(\tilde g\left( t \right)\) with an arbitrarily chosen time envelope g(t) to be injected into the waveguide (Fig. 2a, left). The corresponding output signal \(\tilde f\left( t \right)\) is then calculated by convoluting \(\tilde g\left( t \right)\) with the impulse response of the system, obtained from H(f) (we have also verified our results by direct simulations in the time domain, see Supplementary Note 2). Comparing the envelope of the resulting output signal f(t) (blue line) to the exact solution of the intended differential equation (dashed line) reveals that the topological ASP system is indeed solving the equation as sound propagates through the system.
Next, we add some disorder to our equation solver by randomly shifting the position of the cylinders (average position shift is 18% of lattice period in any direction) and repeat the same procedure in the bottom panel of Fig. 2a. We notice that, despite the relatively large level of disorder, the transfer function H(f) has been left almost unaffected. Hence, the corresponding output signal f(t) still corresponds to the solution of the desired differential equation, confirming the high robustness of the proposed equation solver. To demonstrate that this property is indeed linked to the topological nature of the system, we repeat this analysis for a topologically trivial equation solver, which is based on a resonance induced by defecttunneling through a Bragg band gap. As confirmed in Fig. 2b, such resonating system is also capable of solving the firstorder differential equation. The transfer function H(f), and the output signal \(\tilde g\left( t \right)\) is however severely affected when imparting similar imperfections to the sample (position shifts have the same magnitude as that of topological case). This clearly affirms the superiority of topological ASPs over trivial ones. It should be pointed out that the choice of the input signal envelope g(t) is arbitrary here and any other temporal form can be considered for g(t) (see Supplementary Note 3).
Symmetry protection of the topological equation solver
Since onedimensional topological phases are symmetry protected, these numerical results raise an important question: what is the underlying symmetry of the proposed system that protects its edge modes? In regular tightbinding SSH chains, made of evanescently coupled identical resonators with detuned hoppings K and J (Fig. 3a, top panel), the midgap edge mode occurring at the topological boundary is protected by chiral symmetry, and a transfer function based on tunneling through this edge mode is robust to disorder in the hoppings, as long as they are weak enough not to close the band gap. However, transmission is not robust to even small levels of onsite disorder, which breaks chiral symmetry. This is exemplified in Fig. 3a–c. Figure 3a shows the midgap spectral transmission resonance associated with a perfectly ordered sample. Figure 3b shows the transfer function immunity to disorder in the couplings. Finally, Fig. 3c shows the large sensitivity of the transmission peak to arbitrarily small disorder in the resonance frequencies, which breaks the chiral symmetry. Our multiple scattering system, albeit not based on evanescent coupling, behaves similarly. The transmission peak of the ordered sample (Fig. 3d) survives disorder shifts that do not close the band gap (Fig. 3e), but not disorder in the obstacle radii (Fig. 3f). These numerical results, obtained from fullwave finiteelements simulations, are fully consistent with our topological theory, which defines topological invariants on each bulk band using the unit cell transfer matrix M_{cell}, that maps the Brillouin circle to a subspace of SU(1,1) matrices (see Methods). Remarkably, the topological invariants can only be defined under the symmetry \(M_{{\mathrm{cell}}}^2 = 1\), which holds for position disorder, but not for radii disorder, as we prove in Methods.
Experimental demonstration
Based on these findings we have built a prototype of the topological equation solver (Fig. 4a, top signal path). We first perform a frequencydomain measurement to obtain the transfer function of the system, H(f), by exciting the waveguide with pseudorandom noise and recording the transmitted pressure with a microphone. The graph in the middle inset represents the magnitude of the measured transfer function (green curve) compared to what we get from the numerical simulations (gray curve). As observed, the transfer function has a peak near the resonance frequency f_{0} of the topological edge mode, corresponding to the resonance parameters A = 0.87and Q = 0.03f_{0}, or differential equation parameters α = 2.7π, β = 10π/3. We next switch to a direct timedomain experiment and inject the same arbitrary input signal \(\tilde g\left( t \right)\) as in Fig. 2 into the waveguide (see Supplementary Note 4 for measurements with other types of input signals). Comparing the measured transmitted pressure \(\tilde f\left( t \right)\) (blue line) with the exact solution of the corresponding differential equation (dashed line) confirms the proper functioning of the equation solver. To probe its stability, we then randomly move the cylindrical scatterers and repeat the same procedure (Fig. 4a, bottom signal path). Remarkably, the topological ASP is still perfectly functional despite these large shifts. This exceptional property is strikingly highlighted when we compare the measured output signal from the topological equation solver with that measured at the output of its trivial counterpart in the presence of disorder of similar strength (Fig. 4b). Very different from the topological processor, the signal coming out of the trivial processor is completely distorted, which clearly validates the superior robustness of topological ASP systems. Note that a large range of values for the parameter α and β can be targeted using various means, like changing the lattice periodicity, number of unit cells, and adding losses (see Supplementary Note 5).
To further extend the reach of the approach, we investigate the possibility of solving linear differential equations of higher order. Suppose, for instance, that we want to solve the secondorder differential equation f″(t) + 6πf'′(t) + 8π^{2}f(t) = 4π^{2}g(t), which corresponds to the transfer function H(f) = 1/(2 + 3j(f − f_{0}) − (f − f_{0})^{2}). Using partial fraction decomposition, one can then write H(f) = H_{1}(f) − H_{2}(f), with H_{1}(f) = 1/(1 + j(f − f_{0})) and H_{2}(f) = 1/(2 + j(f − f_{0})). It follows that in order to solve the desired secondorder differential equation, we can realize two (firstorder) equation solvers with the transfer functions H_{1}(f) and H_{2}(f) and subtract their output signals (see Fig. 5a). This is accomplished in an analog way in Fig. 5b, where H_{1}(f) and H_{2}(f) are realized using two different topological firstorder systems with tailored dissipation losses. The analog subtraction operation is realized with an acoustic ratrace coupler. Fullwave simulations involving the full geometry with the twopipes and the ratrace coupler confirm that H(f) = H_{1}(f) − H_{2}(f) is properly implemented. Hence, when an input signal, \(\tilde g\left( t \right)\), say for example with a Gaussian envelope, is applied to the system, the envelope f(t) of the output signal follows the exact solution of the target differential equation. This is confirmed by direct finite difference time domain (FDTD) simulations.
The experimental demonstration of topological secondorder differential equation solving is provided in Fig. 5c. We designed two firstorder differential equation solvers connected to each other via our 3Dprinted acoustic ratrace coupler. The two firstorder ODE solvers are tuned to solve the secondorder ODE by adjusting the level of transmission losses using sound absorbing melamine foam. We then simultaneously excited both waveguides with the input signal \(\tilde g\left( t \right)\), and measured the output \(\tilde f(t)\). As seen in the figure, excellent agreement exists between the measured output signal envelope f(t) (solid blue line) and the expected exact solution of the corresponding secondorder differential equation (dashed line). This technique can easily be extended to the resolution of differential equations of arbitrary order (see Supplementary Note 6). It is worthy to mention further that, as an alternative route to what we proposed here, one can realize higher order and more complex transfer functions by cascading two or more SSH chains, allowing their topological edge modes to couple to each other, as we demonstrate in Supplementary Note 7.
Discussion
The robustness of the proposed topological ASPs constitutes a key step towards a new generation of allacoustic or alloptical ASPs, which not only can be much faster and simpler than DSPs, but also can compete with them in terms of reliability and flexibility. Our proposed strategy also opens new perspectives for further explorations that merge the field of topological insulators with linear system theory. For instance, while we focused here on timedomain signal processing (solving differential equations in real time), performing topological operation in other domains such as space or frequency (i.e. topological filters) seems equally promising and straightforward from our findings (an example is given in Supplementary Note 8). Another interesting direction would be to translate the concept to optics, where general design tools such as optical circuits and metatronics^{53} can enable an easy implementation of the topological concept proposed here, and facilitate the design and optimization of topological ASPs, so that they perform more complex signal processing tasks, or so that they become more robust to certain types of defects instead of others (immunity engineering). Alternatively, one may think of generalizing the concept to twodimensional topological systems protected by timereversal symmetry, which offer immunity against a broader range of defects. Finally, we envision that the coupling between topological ASPs systems and nonlinearities may lead to exciting venues for largescale neural network systems with robust processing capabilities.
Methods
Bloch eigenproblem
The bulk crystal is onedimensional with lattice constant a and two obstacles per unit cell. We model it and define its topology using the transfer matrix M_{cell} of a unit cell. We start by defining the two scattering matrices S_{1} and S_{2}, as the farfield scattering matrices of each obstacle when being alone in the monomode waveguide. These matrices relate the outgoing complex signals on the left (L) and right (R) sides of the scatterers b_{L} and b_{R} to the incident ones, a_{L} and a_{R}:
Note that for now we do not make the assumption that the two matrices are equal: for instance, the cylinders could have different crosssections, or be shifted with respect to each other, etc. These matrices also usually depend on the angular frequency ω. Assuming conservation of energy during the scattering process, they must be unitary. We can therefore parametrize them very generally as
where the frequencydependent angles θ_{1,2}, α_{1,2}, ϕ_{1,2}, and Φ_{1,2} are unique once we fix the reference plane, here at the central position of the scatterers. Assuming reciprocity (S_{21} = S_{12}), we must have 2α_{1,2} − Φ_{1,2} = π, which restricts us to three parameters per scattering matrix, allowing to write:
One can then derive the associated transfer matrices M_{1} and M_{2}, defined as
and obtains
If the two scatterers are separated by a distance d in a unit cell of lattice constant a, the total transfer matrix of the unit cell M_{cell} is the product:
with
where \(L = d,\frac{{a  d}}{2},\) and c is the phase velocity. One obtains, after taking the matrix product,
with
We use the notation z* to denote the complex conjugate of z. Noting ψ〉 = [a, b]^{T}, with a and b being the forward and backward complex field amplitudes at the entrance of the unit cell, the application of Bloch theorem yields the following eigenvalue problem,
which we call the Bloch eigenproblem of the crystal. Note the nontrivial dependence of M_{cell}(ω) on ω. The most straightforward use of the above equation is the following way: for all values of ω, one can diagonalize M_{cell}(ω), and get two opposite values ±k_{B}(ω) of the Bloch wavenumber in the first Brillouin zone, and resolve the band structure. Note that M_{cell} is not unitary and is nonHermitian, meaning that in general, the values ±k_{B}(ω) are complex, allowing in principle for an infinite number of bands and bandgaps. Note further the difference with the standard tightbinding SSH model, which leads to a Hermitian eigenvalue problem that maps the Brillouin circle into the space of SU(2) matrices, and a clear topological classification of chiral symmetric systems via the winding number. Here, consistent with timereversal symmetry^{54,} \(M_{{\mathrm{cell}}}\left( \omega \right) \in {\mathrm{{SU}}}(1,1)\), a group of nonHermitian matrices^{55}. SU(1,1) Hamiltonians are found, for instance, in PTsymmetric extensions of the SSH tightbinding model^{56} where nonHermiticity of the Hamiltonian originates from the absence of energy conservation. Here, M_{cell} is not a Hamiltonian, in the sense that its eigenvalues are not related to ω, but to k_{B}, and the pseudo antiHermiticity of M_{cell} (\(\sigma _{\mathrm{z}}{\it{M}}_{{\mathrm{cell}}}^{\mathrm{\dagger }}\sigma _{\mathrm{z}} =  {\it{M}}_{{\mathrm{cell}}}\)) is related to timereversal symmetry. In Supplementary Fig. 11 we represent the band structure obtained from the transfer matrix approach, and compare it with the one obtained directly from fullwave simulations of the unit cell subjected to periodic boundary conditions (FEM method). To solve the transfer matrix eigenvalue problem, the parameters θ_{1,2}, α_{1,2}, and Φ_{1,2}, which depend on frequency, were extracted from FEM scattering simulations of a single obstacle in a waveguide. The distance between the two scatterers is taken to be \(d = \frac{a}{2}  e_{\mathrm{p}}\), with e_{p} = 2.8 cm (“trivial” case) and a = 23 cm. The rod diameter is 3.5 cm and the width of the waveguide is 7 cm. The agreement between the two approaches validates the accuracy of the multiple scattering model, in particular the underlying assumption of no nearfield interactions between the obstacles in the crystal.
Properties of the unit cell transfer matrix
To define the topology of the system in the next section, we first need to establish a few key properties of the unit cell transfer matrix. We start with general properties, before moving to more specific properties on a band or at degenerate points of the band structure.
As a direct consequence of timereversal symmetry^{54}, the transfer matrix of the system M_{cell} belongs to the group SU(1,1) of matrices of the form
which is parametrized using the Pauli matrices as
Its eigenvalues, given by \(\lambda _ \pm = \alpha _{\mathrm{R}} \pm i\sqrt {\alpha _{\mathrm{I}}^2  \beta _{\mathrm{R}}^2  \beta _{\mathrm{I}}^2}\) are real when \(\alpha _{\mathrm{I}}^2 < \left \beta \right^2\), and complex otherwise. These eigenvalues are degenerate under the condition \(\alpha _{\mathrm{I}}^2  \beta _{\mathrm{R}}^2  \beta _{\mathrm{I}}^2 = 0\), i.e. when the parameters β_{R}, β_{I} and α_{I} belong to a double cone in the (β_{R}, β_{I}, α_{I}) space. This cone is represented in the bottom panels of Fig. 6. At the tip of the cone, one has β_{R} = β_{I} = α_{I} = 0, meaning that M_{cell} reduces to M_{cell} = α_{R}σ_{0}.
On a band, the matrix M_{cell} has a special form. Indeed, the Bloch eigenproblem implies that \(\alpha _{\mathrm{R}} \pm i\sqrt {\alpha _I^2  \left \beta \right^2} = e^{i\,k_{\mathrm{B}}a}\), from which follows that
and
implying \(\alpha _{\mathrm{I}}^2 + \alpha _{\mathrm{R}}^2 = 1 + \left \beta \right^2\), which is equivalent to \(\alpha _{\mathrm{I}}^2 = {\mathrm{sin}}^2(k_{\mathrm{B}}a) + \left \beta \right^2\), or
On a band, we therefore have
As a result, a band describes a onetoone mapping from the Brillouin circle onto a closed path \(\mathcal{C}\) in the subspace of SU(1,1) matrices M_{cell}(k_{B}) with the above form. From the Bloch eigenvalue problem \(M_{{\mathrm{cell}}}\left( \omega \right)\left \psi \right\rangle = e^{i\,k_{\mathrm{B}}a}\left \psi \right\rangle\), one deduces that on a band, M_{cell}(ω) has complex eigenvalues, meaning that \(\alpha _{\mathrm{I}}^2 > \left \beta \right^2\), i.e. the path \(\mathcal{C}\) must be inside the cone, either in the upper region α_{I} > β, or the lower one α_{I} < −β. In addition, the path \(\mathcal{C}\) can only touch the cone whenever the eigenvalues of M_{cell}, namely \(e^{i\,k_{\mathrm{B}}a}\), are degenerate. This is necessarily the case at the edges of the Brillouin zone \(\left( {k_{\mathrm{B}} = \pm \frac{\pi }{a}} \right)\), and at its center k_{B} = 0. In between, \(\mathcal{C}\) cannot touch the cone, since two distinct eigenvalues \(e^{ \pm i\,k_{\mathrm{B}}a}\) must be found, by virtue of timereversal symmetry. Finally, the path \(\mathcal{C}\) is not a loop, but a simple line, since M_{cell} is a simple function of ω, and therefore is the same for two opposite values of k_{B} on a band: it starts on the cone at \(k_{\mathrm{B}} =  \frac{\pi }{a}\) and lands on it again at k_{B} = 0, before following the reverse path between k_{B} = 0 and \(k_{\mathrm{B}} = \frac{\pi }{a}\). Figure 6a represents an example of \(\mathcal{C}\) contour for the third band of the crystal (supposedly topologically “trivial” case, with e_{p} = 2.8 cm), and Fig. 6c represents the same contour for e_{p} = −2.8 cm, corresponding to the dual system, which is supposedly topological (the topological properties will be proven in the next section). Figure 6b represents the case e_{p} = 0 cm that closes the bandgaps. As expected, in all cases the contour starts and end on the cone.
To study the conditions under which two consecutive frequency bands can touch, it is convenient to recast the Bloch eigenproblem into the equivalent form:
and think of it as follows: for each k_{B} in the first Brillouin zone, finding the bands means finding the values of ω for which the matrix \(e^{  i\,k_{\mathrm{B}}a}M_{{\mathrm{cell}}}\) has at least one eigenvalue equal to one, with the corresponding eigenvector being the Bloch eigenvector on that particular band. This can happen for infinitely many values of ω. If both eigenvalues of \(e^{  i\,k_{\mathrm{B}}a}M_{{\mathrm{cell}}}\) at a given frequency are equal to one, the band structure is doubly degenerate, which is therefore the maximum frequency degeneracy allowed by the system. Since the general form of the eigenvalues of \(e^{  i\,k_{\mathrm{B}}a}M_{{\mathrm{cell}}}\) on a band are \(\upsilon _ \pm = e^{  i\,k_{\mathrm{B}}a}\left( {\alpha _{\mathrm{R}} \pm i\sqrt {\alpha _{\mathrm{I}}^2  \left \beta \right^2} } \right) = e^{  i\,k_{\mathrm{B}}a}e^{ \pm i\,k_{\mathrm{B}}a}\), the second eigenvalue \(e^{  2i\,k_{\mathrm{B}}a}\) can only become equal to unity at the Brillouin zone edges \(\left( {k_{\mathrm{B}} = \pm \frac{\pi }{a}} \right)\), or at k_{B} = 0. As a consequence, bandgaps can only close at the center or edge of the Brillouin zone, i.e. when the contour \(\mathcal{C}\) touches the cone.
Assuming the first case, i.e. a degeneracy at \(k_{\mathrm{B}} = \pm \frac{\pi }{a}\), one has \(e^{  i\,k_{\mathrm{B}}a} =  1\). We obtain, at the particular frequency of the degeneracy,
and this matrix can only be equal to identity if \(\left \beta \right = 0\). The second case of degeneracy at k_{B} = 0 leads to the same conclusion \((\left \beta \right = 0)\). This means that when two bands touch, the contour \(\mathcal{C}\) is reaching the tip of the cone, as confirmed by Fig. 6b.
Topology of the bands
As seen in previous sections, each band defines a mapping between the Brillouin circle and a subspace of SU(1,1) matrices. We now define a topological invariant for each band, i.e. an integer quantity that is invariant upon continuous transformations of the band structure. This means that this number can only change when the band undergoes a discontinuous transformation, i.e. touches another one, or equivalently when the contour \(\mathcal{C}\) touches the tip of the cone.
Like in the standard tightbinding SSH model, we need an extra symmetry, akin to chiral symmetry, to be able to define topological invariants on each band. Here we need to require that the scattering matrices S_{1} and S_{2} are equal, taking θ_{1} = θ_{2} =θ, α_{1,2} = α_{1,2} = α and φ_{1} = φ_{2} = φ. With this extra condition, the quantity \(\beta = M_{21}\left( {\omega (k_{\mathrm{B}})} \right)\) in Eq. 14, that parametrizes the matrix M_{cell} on a band, becomes
where the quantities α, θ and φ that parametrize the S matrix of a single obstacle generally depend on ω(k_{B}). We then assume the case of nonresonant scatterers, meaning that cos θ does not vanish on the band, and the variation of α and θ on the band are negligible. Because M_{cell} always has two complexconjugate unimodular eigenvalues, ω(k_{B}) is necessarily monotonous between −π/a and 0. Let us focus our attention to the quantity \(\cos \left( {\alpha + \frac{{\omega \left( {k_{\mathrm{B}}} \right)d}}{c}} \right)\), which can potentially make the complex number β(k_{B}) vanish at some particular point of the Brillouin zone. When k_{B} goes from −π/a to 0, the angle \(\gamma = \alpha + \frac{{\omega \left( {k_{\mathrm{B}}} \right)d}}{c}\) moves monotonically between two real values, say γ_{min} and γ_{max}, defining a continuous monotonous mapping between \(\left[ {  \frac{\pi }{a},0} \right]\) to [γ_{min}, γ_{max}]. Now, two situations can arise:

(1)
The segment [γ_{min}, γ_{max}] does not contain π/2 (modulo π), in which case \(\cos \left( {\alpha + \frac{{\omega \left( {k_{\mathrm{B}}} \right)d}}{c}} \right)\) never vanishes as k_{B} go from −π/a to 0. This means that β never vanishes on the band.

(2)
The segment [γ_{min}, γ_{max}] contains π/2 (modulo π), in which case β vanishes at least once on the band.
Since β = 0 means that the contour \(\mathcal{C}\) crosses the cone axis, we can therefore define a topological invariant η in the following way: We can count the number of times η that \(\mathcal{C}\) crosses the cone axis as k_{B} goes from −π/a to 0. This integer number changes each time γ_{max} or γ_{min} equals π/2 (modulo π), i.e. when β is zero either at the edge or center of the Brillouin zone, i.e. when a band gap closes. Figure 6 shows how the contour \(\mathcal{C}\) evolves for the third band of our system, when one goes from the trivial regime (panel a, \(\mathcal{C}\) does not cross the cone axis, η = 0) to the topological one (panel c, \(\mathcal{C}\) crosses the cone axis, η = 1). At the topological phase transition, the contour \(\mathcal{C}\) touches the tip of the cone, which closes the band gap, and the number η is not defined.
Symmetry protection
The definition of the topological invariant η as the number of times the contour \(\mathcal{C}\) crosses the cone axis between −π/a to 0 is based on two underlying symmetries, and both must be fulfilled:

1.
(1) Timereversal symmetry, which guarantees that M_{cell} belongs to SU(1,1)^{55}.

1.
(2) Equality of S_{1} and S_{2} (the farfield individual scattering matrices of both obstacles must be identical), or equivalently:
Obviously, horizontal position disorder does not change the individual scattering parameters of the object. In addition, vertical position disorder does not change it either, as demonstrated in Supplementary Fig. 12 (the only difference in the scattering spectrum are very sharp Fano interferences occurring from coupling to a acoustic bound state in the continuum, but they are far from the frequency range of interest). As a consequence, position disorder does not break \(M_{{\mathrm{cell}}}^2 = 1\). However, changing the diameter of one rod definitely changes its scattering matrix. What happens in the case of rods with different radii is that the real and imaginary part of the quantity
are never simultaneously zero, which implies that the contour \(\mathcal{C}\) can avoid crossing the cone axis by simply going around it. This is analogous to a SSH chain without chiral symmetry, where some properly chosen chiralitybreaking defects at an interface can change the winding number without closing the band gap. These results explain the outcome of the fullwave simulations presented in Fig. 3 of the main text.
Numerical methods
Fullwave simulations are all performed using Comsol Multiphysics (Acoustic and RF modules). Dispersion curves are obtained by considering a single unit cell of the lattice arrays, applying Floquet boundary condition to the lateral sides of the unit cell, and performing eigenfrequency simulations for all of the Floquet−Bloch wavenumbers.
In order to obtain the frequency spectra of the ODE solvers, we excite the system with an incident planewave with unit amplitude and measure the amount of pressure at the transmission side of the waveguide.
In order to crossvalidate our experimental measurements, we performed numerical finiteelement simulations including a viscothermal loss of 1.15 dB/m to achieve a transfer function X(ω), for example, between the injected and transmitted sound waves. We then obtained the transfer function of the loudspeaker Y(ω) by exciting the empty waveguide and measuring the associated sound pressure level at the transmission side. The transfer function Z(ω), between the voltage applied to the loudspeaker and the transmitted pressure, was then readily obtained as \(Z\left( \omega \right) = X(\omega )/Y(\omega )\).
In our FDTD simulations, we excite the waveguide from one end with the desired modulated input signal, and record the temporal evolution of the pressure field (with a time step subject to Courant–Friedrichs–Lewy (CFL) condition for ensuring stability) received at a point on the other side of the waveguide.
Experimental methods
As mentioned in the main text, an acrylic square tube is used to implement the acoustic waveguide. Nylon 6 continuous cast cylinders were then manually inserted into the waveguide to form the SSHtype array. Supplementary Fig. 13a represents the experimental setup used to achieve the transfer function of the system. The setup contains a loudspeaker, a Data Physics Quattro signal analyzer connected to a computer (not shown in the figure) controlling it, one ICP microphone measuring the transmitted sound pressure level, and a homemade anechoic termination (not shown in the figure). To obtain the transfer function of the sample, we drive the loudspeaker with a burst noise voltage (which is set as the reference signal in the setup), and measure the pressure level with respect to the reference channel using the ICP microphone. Supplementary Fig. 13b shows the experimental setup used to create an input signal (voltage) with an arbitrary time profile \(\tilde g(t)\), and to measure the temporal evolution of the output signal \(\tilde f(t)\). The setup consists of a Speedgoat Performance RealTime Target Machine with IO131 interface controlled by xPC target environment of MATLAB/Simulink, a loudspeaker, a power amplifier, a homemade acoustic termination (not shown in the figure), and an ICP microphone measuring the transmitted pressure.
Data availability
The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.
References
 1.
Proakis, J. G. & Monolakis, D. G. Digital Signal Processing: Principles, Algorithms, and Applications (Prentice Hall, Upper Saddle River, NJ, 1996).
 2.
PallásAreny, Ramón, and John G. Webster. Analog signal processing. John Wiley & Sons, 1999.
 3.
Silva, A. et al. Performing mathematical operations with metamaterials. Science 343, 160–163 (2014).
 4.
ZangenehNejad, F. & Khavasi, A. Spatial integration by a dielectric slab and its planar graphenebased counterpart. Opt. Lett. 42, 1954–1957 (2017).
 5.
Bykov, D. A. et al. Firstorder optical spatial differentiator based on a guidedmode resonant grating. Opt. Express 26, 10997–11006 (2018).
 6.
Wu, W., Jiang, W., Yang, J., Gong, S. & Ma, Y. Multilayered analog optical differentiating device: performance analysis on structural parameters. Opt. Lett. 42, 5270–5273 (2017).
 7.
Kwon, Hoyeong et al. Nonlocal metasurfaces for optical signal processing. Phys. Rev. Lett. 121, 173004 (2018).
 8.
Zuo, S., Wei, Q., Tian, Y., Cheng, Y. & Liu, X. Acoustic analog computing system based on labyrinthine metasurfaces. Sci. Rep. 8, 10103 (2018).
 9.
Zhang, W. et al. Implementing quantum search algorithm with metamaterials. Adv. Mater. 30, 1703986 (2018).
 10.
Fang, Y., Lou, Y. & Ruan, Z. Ongrating graphene surface plasmons enabling spatial differentiation in the terahertz region. Opt. Lett. 42, 3840–3843 (2017).
 11.
Hwang, Y., Davis, T. J., Lin, J. & Yuan, X.C. Plasmonic circuit for secondorder spatial differentiation at the subwavelength scale. Opt. Express 26, 7368–7375 (2018).
 12.
Zhu, Tengfeng et al. Generalized spatial differentiation from the spin Hall effect of light and its application in image processing of edge detection. Phys. Rev. Appl. 11, 034043 (2019).
 13.
Dong, Z., Si, J., Yu, X. & Deng, X. Optical spatial differentiator based on subwavelength highcontrast gratings. Appl. Phys. Lett. 112, 181102 (2018).
 14.
Youssefi, A., ZangenehNejad, F., Abdollahramezani, S. & Khavasi, A. Analog computing by Brewster effect. Opt. Lett. 41, 3467–3470 (2016).
 15.
Pors, A., Nielsen, M. G. & Bozhevolnyi, S. I. Analog computing using reflective plasmonic metasurfaces. Nano Lett. 15, 791–797 (2015).
 16.
Yang, T. et al. Alloptical differential equation solver with constantcoefficient tunable based on a single microring resonator. Sci. Rep. 4, 5581 (2014).
 17.
Wu, Jiayang et al. Compact tunable silicon photonic differentialequation solver for general linear timeinvariant systems. Opt. express 22, 26254–26264 (2014).
 18.
Zhu, T. et al. Plasmonic computing of spatial differentiation. Nat. Commun. 8, 15391 (2017).
 19.
Guo, C., Xiao, M., Minkov, M., Shi, Y. & Fan, S. Photonic crystal slab Laplace operator for image differentiation. Optica 5, 251–256 (2018).
 20.
Roberts, Ann, Daniel, E. G. ómez & Timothy, J. Davis Optical image processing with metasurface dark modes. JOSA A 35, 1575–1584 (2018).
 21.
ZangenehNejad, F. & Fleury, R. Performing mathematical operations using highindex acoustic metamaterials. New J. Phys. 20, 073001 (2018).
 22.
Ferrera, M. et al. Onchip CMOScompatible alloptical integrator. Nat. Commun. 1, 29 (2010).
 23.
Brunner, D., Soriano, M. C., Mirasso, C. R. & Fischer, I. Parallel photonic information processing at gigabyte per second data rates using transient states. Nat. Commun. 4, 1364 (2013).
 24.
Vourkas, I., Stathis, Di & Sirakoulis, G. C. Massively parallel analog computing: Ariadne’s thread was made of memristors. IEEE Trans. Emerg. Top. Comput. 6, 145–155 (2018).
 25.
Li, Ming et al. Reconfigurable optical signal processing based on a distributed feedback semiconductor optical amplifier. Sci. Rep. 6, 19985 (2016).
 26.
Liu, Weilin et al. A fully reconfigurable photonic integrated signal processor. Nat. Photon. 10.3, 190 (2016).
 27.
Santagiustina, M., Chin, S., Primerov, N., Ursini, L. & Thévenaz, L. Alloptical signal processing using dynamic Brillouin gratings. Sci. Rep. 3, 1594 (2013).
 28.
Willner, AlanE. et al. Alloptical signal processing. J. Light Technol. 32, 660–680 (2014).
 29.
Koos, Christian et al. Alloptical highspeed signal processing with silicon–organic hybrid slot waveguides. Nat. photonics 3, 216 (2009).
 30.
Corcoran, Bill et al. Optical signal processing on a silicon chip at 640Gb/s using slowlight. Opt. express 18, 7770–7781 (2010).
 31.
Kuo, S. M., Lee, B. H. & Tian, W. RealTime Digital Signal Processing: Fundamentals, Implementations and Applications (Wiley, New York, 2013).
 32.
Asbóth, J. K., Oroszlány, L. & Pályi, A. in A Short Course on Topological Insulators. Springer Lecture Notes in Physics Vol. 919 (Springer, New York, 2015).
 33.
Wang, Z., Chong, Y., Joannopoulos, J. D. & Soljačić, M. Observation of unidirectional backscatteringimmune topological electromagnetic states. Nature 461, 772–775 (2009).
 34.
He, C. et al. Acoustic topological insulator and robust oneway sound transport. Nat. Phys. 12, 1124–1129 (2016).
 35.
Wang, P., Lu, L. & Bertoldi, K. Topological phononic crystals with oneway elastic edge waves. Phys. Rev. Lett. 115, 104302 (2015).
 36.
Gorlach, M. A. et al. Farfield probing of leaky topological states in alldielectric metasurfaces. Nat. Commun. 9, 909 (2018).
 37.
Yves, S. et al. Crystalline metamaterials for topological properties at subwavelength scales. Nat. Commun. 8, 16023 (2017).
 38.
Khanikaev, A. B., Fleury, R., Mousavi, S. H. & Alù, A. Topologically robust sound propagation in an angularmomentumbiased graphenelike resonator lattice. Nat. Commun. 6, 8260 (2015).
 39.
ZangenehNejad, Farzad & Fleury, Romain Topological Fano resonances. Phys. Rev. Lett. 122, 014301 (2019).
 40.
Hafezi, M., Demler, E. A., Lukin, M. D. & Taylor, J. M. Robust optical delay lines with topological protection. Nat. Phys. 7, 907–912 (2011).
 41.
Rechtsman, M. C. et al. Photonic Floquet topological insulators. Nature 496, 196–200 (2013).
 42.
Cheng, X. et al. Robust reconfigurable electromagnetic pathways within a photonic topological insulator. Nat. Mater. 15, 542–548 (2016).
 43.
Chen, W. J. et al. Experimental realization of photonic topological insulator in a uniaxial metacrystal waveguide. Nat. Commun. 5, 5782 (2014).
 44.
Susstrunk, R. & Huber, S. D. Observation of phononic helical edge states in a mechanical topological insulator. Science 349.6243, 47–50 (2015).
 45.
Yu, SiYuan et al. Elastic pseudospin transport for integratable topological phononic circuits. Nat. Commun. 9.1, 3072 (2018).
 46.
Lumer, Y., Plotnik, Y., Rechtsman, M. C. & Segev, M. Selflocalized states in photonic topological insulators. Phys. Rev. Lett. 111, 243905 (2013).
 47.
Wang, Z., Chong, Y. D., Joannopoulos, J. D. & Soljačić, M. Reflectionfree oneway edge modes in a gyromagnetic photonic crystal. Phys. Rev. Lett. 100, 013905 (2008).
 48.
Harari, G. et al. Topological insulator laser: theory. Science 359, eaar4003 (2018).
 49.
Yu, H. et al. Topological insulator as an optical modulator for pulsed solidstate lasers. Laser Photonics Rev. 7, L77–L83 (2013).
 50.
Hassler, F., Akhmerov, A. R. & Beenakker, C. W. J. Flatlens focusing of electrons on the surface of a topological insulator. Phys. Rev. B—Condens. Matter Mater. Phys. 82, 125423 (2010).
 51.
Joannopoulos, J., Johnson, S. and Winn, J. Photonic Crystals: Molding the Flow of Light. Princeton University Press. (2008).
 52.
Qi, X. L. & Zhang, S. C. Topological insulators and superconductors. Rev. Mod. Phys. 83, 1057 (2011).
 53.
Li, Y., Liberal, I. & Engheta, N. Dispersion synthesis with multiordered metatronic filters. Opt. Express 25, 1937–1948 (2017).
 54.
Griffiths, D. J. & Steinke, C. A. Waves in locally periodic media. Am. J. Phys. 69, 137–154 (2001).
 55.
Wybourne, B. G. (1974). Classical Groups for Physicists. Princeton University Press.
 56.
Esaki, K., Sato, M., Hasebe, K. & Kohmoto, M. Edge states and topological phases in nonHermitian systems. Phys. Rev. B—Condens. Matter Mater. Phys. 84, 205128 (2011).
Acknowledgements
This work was supported by the Swiss National Science Foundation (SNSF) under Grant No. 172487. The authors would like to thank Dr Nadège Kaina for useful discussions about topological properties of multiple scattering systems.
Author information
Affiliations
Contributions
F.ZN. carried out the research work under the supervision of R.F. Both authors participated in discussing the results and writing the manuscript.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Journal peer review information: Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. Peer reviewer reports are available.
Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
ZangenehNejad, F., Fleury, R. Topological analog signal processing. Nat Commun 10, 2058 (2019). https://doi.org/10.1038/s41467019100863
Received:
Accepted:
Published:
Further reading

A single inversedesigned photonic structure that performs parallel computing
Nature Communications (2021)

Analogue Optical Spatiotemporal Differentiator
Advanced Optical Materials (2021)

Topological optical differentiator
Nature Communications (2021)

The Design of Split Ring Resonator (Srr) Based Terahertz Bandpass Filter and Comparison of Various Types of Filters.
Journal of Physics: Conference Series (2021)

Topologically Protected Exceptional Point with Local NonHermitian Modulation in an Acoustic Crystal
Physical Review Applied (2021)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.