Coupling Matrix
From Lowpass Prototype to Physical Hardware
The coupling matrix formulation, developed by Atia and Williams in the early 1970s and generalized by Cameron in the late 1990s, replaced the older element-by-element circuit synthesis with a compact linear-algebra description of the whole filter. Once a desired response (Chebyshev passband ripple plus a set of prescribed transmission zeros) is captured as a pair of characteristic polynomials, those polynomials are converted directly into a normalized coupling matrix. The matrix is normalized in the lowpass domain, so a single matrix describes the filter shape independent of center frequency and bandwidth; denormalizing it to a real design simply scales every entry by the fractional bandwidth and sets the resonant frequency.
Physically, each off-diagonal entry corresponds to a real coupling structure. In a combline or cavity filter, the main-line couplings Mi,i+1 map to iris widths or probe depths between adjacent resonators, the input and output couplings MS1 and MNL map to the tap point or coupling loop that sets the external Q, and the diagonal entries map to small frequency detunings applied with tuning screws. The power of the method is that it lets a designer reshape the response purely in matrix form, verify it instantly by computing S-parameters from the matrix, and only then commit to a physical layout. The same matrix later becomes the target for computer-aided tuning on the production floor.
The N+2 Matrix and Its Entries
In the N+2 convention the rows and columns are ordered S, 1, 2, up to N, then L. The matrix is symmetric for a reciprocal, lossless filter, so only the upper triangle carries independent information. Main-line entries sit on the first off-diagonal; cross-couplings sit further from the diagonal and are what distinguish a folded, cascaded-triplet (CT), or cascaded-quadruplet (CQ) topology. A direct source-to-load entry MSL is what enables a fully canonical filter to place all N transmission zeros at finite frequencies.
Governing Equations
[A] = [R] + Ω[U] − j[M]
S21 = −2j × [A−1]N+2,1, S11 = 1 + 2j × [A−1]1,1
Denormalization to physical couplings:
kij = FBW × Mij (inter-resonator coupling)
Qe = 1 / (FBW × MS12) (external quality factor)
Where [U] is the identity with zeros at the S and L positions, [R] holds the port terminations, Ω is the normalized lowpass frequency, [M] is the coupling matrix, FBW is fractional bandwidth, and j = √−1. Example: a 4-pole, 1% FBW filter with M12 = 0.92 gives a physical coupling k12 ≈ 0.0092.
Topology and Transmission-Zero Capability
| Topology | Cross-coupling pattern | Max finite TZs | Zero symmetry | Typical use |
|---|---|---|---|---|
| Inline (Chebyshev) | None (tridiagonal) | 0 | n/a | Simple bandpass, max stopband rolloff via order |
| Cascaded triplet (CT) | One per triplet | 1 per triplet | Asymmetric | Single-sided rejection, group-delay shaping |
| Cascaded quadruplet (CQ) | One per quad | 2 per quad | Symmetric pair | Symmetric sharp skirts, duplexer channels |
| Folded canonical | Multiple, folded | N − 2 | Mixed | Compact high-order satellite filters |
| Transversal / N+2 canonical | S-to-all, with MSL | N | Arbitrary | Reference form for synthesis and rotation |
Frequently Asked Questions
What is the difference between the N+2 and N x N coupling matrix?
The N × N matrix represents only the resonators and folds the ports into input and output external Q values. The N+2 matrix adds explicit source (S) and load (L) rows and columns, giving an (N+2) × (N+2) array. That lets the source and load couple to multiple resonators and carry a direct MSL term, so the N+2 form (Cameron's modern standard) can synthesize every realizable response, including fully canonical filters with N finite transmission zeros, without the rank limits of the N × N matrix.
How do cross-couplings create transmission zeros?
A cross-coupling links two non-adjacent resonators, creating a bypass path alongside the main inline path. At one frequency the bypass signal arrives equal in amplitude and 180° out of phase with the direct signal, so they cancel and form a finite-frequency zero in S21. The sign of the coupling (electric versus magnetic) decides which side of the passband the zero falls on, which is how asymmetric responses are built. The count of independent cross-coupling paths sets the maximum number of zeros.
How is a coupling matrix extracted from measured S-parameters?
The measured or simulated response is fit to a rational polynomial model to recover the characteristic polynomials, from which a canonical coupling matrix is computed. A sequence of Givens similarity rotations preserves the eigenvalues, and hence the response, while zeroing the entries the chosen physical topology cannot realize, yielding a matrix whose nonzero terms map onto real irises and screws. Diagnostic extraction from measured data reports how far each coupling and resonance sits from target to guide computer-aided tuning.