Discrete Fourier Transform Interferometer

• Discrete Fourier transform interferometers are systems that encode optical, radio, or quantum signals using fixed delay samples processed via DFT, mapping them to spectral or spatial domains.
• They are implemented in integrated photonics, radio astronomy, and spectroscopy, offering a robust alternative to traditional tunable interferometric designs with enhanced symmetry and reduced sensitivity to fabrication variations.
• The architecture leverages analytical decompositions with fixed DFT mixing layers and phase masks to achieve scalable, efficient interferometric measurements and FFT beamforming with reduced component complexity.

Searching arXiv for relevant papers on discrete Fourier transform interferometers and closely related interferometric architectures. A discrete Fourier transform interferometer denotes a class of interferometric systems in which the measured optical, radio, or quantum signal is encoded through a discrete set of path delays, mixing operations, or spatial-frequency samples that are naturally interpreted through the discrete Fourier transform (DFT) or closely allied Fourier-transform formalisms. Across integrated photonics, radio astronomy, spectroscopy, and beamforming, the term covers several distinct but related constructions: universal multiport interferometers built from DFT mixing layers and phase masks; on-chip spectrometers that sample interferograms at discrete optical path differences; FFT-beamforming arrays that synthesize sky beams from regularly gridded antennas; and differential or nonlinear interferometers whose outputs are processed by Fourier-transform analysis. In each case, the operational principle is the same at an abstract level: interferometric measurements are arranged so that a finite set of samples in delay, mode, or aperture space can be mapped to spectral, spatial, or unitary information by a DFT-like transformation (Girouard et al., 27 Aug 2025, Kita et al., 2018, Masui et al., 2017).

1. Definition and conceptual scope

In photonic circuit theory, a DFT interferometer is a universal multiport interferometer architecture in which fixed multichannel mixing layers implementing the discrete Fourier transform are interleaved with reconfigurable diagonal unitary layers, or phase masks. A constructive result shows that any $N\times N$ unitary $U$ can be decomposed as

$U = D^{(0)} F D^{(1)} F D^{(2)} \cdots F D^{(L)},$

with diagonal unitary matrices $D^{(k)}$ and DFT matrix $F$, yielding an analytical design route for universal interferometers based on discrete Fourier transform couplers rather than solely on Mach-Zehnder interferometer meshes (Girouard et al., 27 Aug 2025).

In integrated spectroscopy, the phrase is also used in a measurement-theoretic sense. The digital Fourier transform spectrometer realizes a reconfigurable Mach-Zehnder interferometer whose switch states generate a discrete set of optical path length differences. Each setting produces one interferogram sample, and the complete set forms a digitized interferometric measurement that is reconstructed numerically as $y = Ax$, where $A$ is a calibration matrix and $x$ is the unknown spectrum (Kita et al., 2018). The on-chip interrogator based on Fourier-transform spectroscopy adopts a related but parallel architecture: an array of Mach-Zehnder interferometers with different optical path differences samples the Fourier transform of the sensor-array spectrum at discrete spatial frequencies (Peternella et al., 2018).

In radio interferometry, DFT interferometry appears as FFT beamforming. When identical antennas are arranged on a regular lattice, spatial Fourier-transform formed beams can be synthesized with computational cost $\mathcal{O}(n\log n)$ rather than $\mathcal{O}(n^2)$, provided the array geometry supports the requisite redundancy structure (Masui et al., 2017). In astronomical theory more broadly, the Fourier transform is presented as only one member of a larger family of linear canonical transforms applicable to interferometric arrays, though the DFT remains the canonical discrete implementation for distant-source imaging on regular grids (Lacki, 2015).

These usages differ in physical realization, but they share a common architecture: interferometric encoding is rendered discrete, and the inverse problem is formulated in the Fourier domain or in a directly equivalent linear-algebraic representation.

2. Unitary photonic implementations with DFT mixing layers

The most explicit use of the term in integrated photonics concerns universal multiport interferometers composed of DFT matrices and phase masks. The central result is an analytical decomposition of arbitrary unitary matrices into a sequence of $U$0 DFT mixing layers and $U$1 phase masks (Girouard et al., 27 Aug 2025). The DFT matrix is given by

$U$2

and the constructive formula culminates in an explicit sequence summarized as

$U$3

The decomposition is positioned against standard Mach-Zehnder interferometer networks such as the schemes of Reck et al. and Clements et al., which realize arbitrary unitaries through cascades of two-mode interferometers and phase shifters. Those schemes are well established, but the DFT/phase-mask architecture is distinguished by its use of fixed multichannel mixing layers with higher symmetry and by an analytical synthesis procedure that avoids optimization-based design (Girouard et al., 27 Aug 2025).

The same work states that previous DFT-based decompositions required expensive numerical optimization to determine phase-mask parameters for a given $U$4, whereas the proposed construction is fully analytical and constructive. The phase-mask/DFT count is reported to be reduced by 66% over previous analytical methods, from $U$5 phase masks to $U$6 (Girouard et al., 27 Aug 2025). The paper also notes a theoretical lower bound of $U$7 phase masks for universal architectures and states that the proposed analytical method comes closer to this bound than prior analytic constructions.

A further stated property is generality for even dimension $U$8, with odd $U$9 embeddable into even dimension by padding (Girouard et al., 27 Aug 2025). The explicit sequence is described as matching the computational complexity of standard Clements/Bell unitary decompositions.

This formulation places the DFT interferometer within the theory of programmable linear optics: the DFT layers act as fixed global mode mixers, while the diagonal phases supply the reconfigurability needed for universality.

3. Architectural rationale, robustness, and fabrication considerations

The DFT-interleaved architecture is motivated not only by analytical decomposability but also by physical robustness. The reported advantages include symmetry and path-independence: highly symmetric DFT mixing layers ensure that all paths through the device experience equal loss, in contrast to Mach-Zehnder meshes where unbalanced paths see higher losses (Girouard et al., 27 Aug 2025). This equal-loss property is significant in photonic hardware because loss nonuniformity directly perturbs implemented linear transformations.

The same source states that perturbations or noise in the mixing layer implementing the DFT do not affect universality, referencing recent works on tolerance to mixing-layer perturbations (Girouard et al., 27 Aug 2025). A direct implication is that the most fabrication-complex components need not be tuned with the same precision demanded by architectures whose universality depends on large numbers of accurately cascaded tunable beam splitters. The paper further states that redundant layers used in some Mach-Zehnder error-correction strategies are unnecessary in this setting.

Physical implementation is described in terms of multimode interference couplers and phase shifters. The DFT mixing layers can be realized with multimode interference couplers routinely fabricated with high uniformity on silicon photonics platforms, while the phase masks are standard single-mode phase shifters (Girouard et al., 27 Aug 2025). Reduced fabrication complexity is attributed to the fact that MMIs and phase shifters are less sensitive to fabrication variations than the precise splitting ratios required by MZIs.

This suggests a specific interpretation of the discrete Fourier transform interferometer in integrated optics: it is not merely a mathematical reformulation of unitary synthesis, but an architectural proposal that trades local tunable interferometric elements for global fixed mixers and diagonal control planes. A plausible implication is improved yield and scalability for large programmable photonic circuits, although the cited work formulates this primarily in terms of robustness, symmetry, loss tolerance, and analytical tractability rather than system-level deployment metrics (Girouard et al., 27 Aug 2025).

4. Spectroscopic DFT interferometers on chip

A second major lineage of DFT interferometers arises in integrated Fourier-transform spectroscopy. In the digital Fourier transform spectrometer, a reconfigurable Mach-Zehnder interferometer is digitized by cascaded binary optical switches in both arms. Differential path lengths are encoded in powers of two,

$U = D^{(0)} F D^{(1)} F D^{(2)} \cdots F D^{(L)},$0

and the switch permutations permit

$U = D^{(0)} F D^{(1)} F D^{(2)} \cdots F D^{(L)},$1

unique optical path length differences between the arms (Kita et al., 2018). Each switch state samples the interferogram at a discrete delay, so the instrument implements a digitized variant of FTIR.

The corresponding measurement model is

$U = D^{(0)} F D^{(1)} F D^{(2)} \cdots F D^{(L)},$2

where $U = D^{(0)} F D^{(1)} F D^{(2)} \cdots F D^{(L)},$3 is the vector of measured interferogram values, $U = D^{(0)} F D^{(1)} F D^{(2)} \cdots F D^{(L)},$4 is the calibration matrix, and $U = D^{(0)} F D^{(1)} F D^{(2)} \cdots F D^{(L)},$5 is the unknown spectrum (Kita et al., 2018). The spectral resolution under the Rayleigh criterion is

$U = D^{(0)} F D^{(1)} F D^{(2)} \cdots F D^{(L)},$6

Because both channel count and resolution scale exponentially with the number of switch stages, the architecture is presented as highly scalable. The same source attributes to it the multiplex advantage, qualitatively expressed as

$U = D^{(0)} F D^{(1)} F D^{(2)} \cdots F D^{(L)},$7

and emphasizes that only a single photodetector is required (Kita et al., 2018).

Reconstruction is underdetermined and therefore regularized. The stated elastic-net-with-smoothing objective is

$U = D^{(0)} F D^{(1)} F D^{(2)} \cdots F D^{(L)},$8

The paper reports that this reconstruction can provide noise suppression and spectral resolution enhancement beyond the classical Rayleigh criterion, with a minimum resolvable wavelength spacing of $U = D^{(0)} F D^{(1)} F D^{(2)} \cdots F D^{(L)},$9 versus a classic device-limited Rayleigh resolution of $D^{(k)}$0 in its tests (Kita et al., 2018).

A related integrated architecture is the planar spatial heterodyne spectrometer used as an on-chip interrogator for photonic sensors. It consists of an array of Mach-Zehnder interferometers with distinct optical path differences, each using a $D^{(k)}$1 multimode interferometer combiner. The $D^{(k)}$2 MMI yields three output signals with nominal $D^{(k)}$3 relative phase shifts, enabling recovery of complex Fourier coefficients (Peternella et al., 2018). For interferometer $D^{(k)}$4,

$D^{(k)}$5

and these complex voltages are proportional to a Fourier integral of the input spectrum: $D^{(k)}$6

Spectrum reconstruction for $D^{(k)}$7 interferometers is written as

$D^{(k)}$8

with spectral resolution

$D^{(k)}$9

With 9 MZIs, the paper reports $F$0 pm (Peternella et al., 2018).

The same device also illustrates an important distinction between native Fourier resolution and parameter-estimation precision. The sensor resonance deviations are modeled as

$F$1

and solving this nonlinear system with Newton’s method allows sub-resolution tracking whose limit is set by SNR rather than by $F$2 (Peternella et al., 2018). Experimentally, the minimum modulation amplitude obtained is 400 fm, more than two orders of magnitude smaller than the FT spectrometer resolution, while about 92% of thermal induced phase drift is compensated by using one sensor as a reference (Peternella et al., 2018).

5. Differential, nonlinear, and time-domain variants

Discrete Fourier transform interferometry is closely related to, but not identical with, broader Fourier-transform interferometry. The efficient differential Fourier-transform spectrometer designed for Sunyaev-Zel'dovich effect measurements is based on a Martin-Puplett interferometer configuration rather than a digital DFT sampling architecture, yet it exemplifies the same principle of encoding spectral information into an interferogram that is analyzed through Fourier-transform methods (Schillaci et al., 2014). The instrument divides the observed sky field into two halves along the meridian, routes them to the two input ports of the MPI, and measures brightness differences between conjugate sky pixels. It operates over $F$3 (30–600 GHz), has maximum spectral resolution $F$4 (1.9 GHz), unvignetted throughput of $F$5, and a measured common-mode rejection ratio upper limit of $F$6 (Schillaci et al., 2014).

The nonlinear interferometer for mid-infrared gas spectroscopy likewise uses Fourier-transform analysis of an interferogram, but here the interferometric encoding exploits induced coherence between correlated photons. A Michelson-type nonlinear interferometer generates broadband signal-idler photon pairs by SPDC, scans delay in the mid-infrared idler arm, and detects only the near-infrared signal photons with a silicon avalanche photodiode (Lindner et al., 2020). The measured interferogram $F$7 is Fourier transformed to recover the spectrum,

$F$8

and the transmission is obtained as the ratio of sample and reference spectra. For a maximum optical delay $F$9, the spectral resolution is reported as $y = Ax$0, enabling rotational-line-resolving spectroscopy over a bandwidth of over $y = Ax$1 (Lindner et al., 2020).

A more radical variant is Fourier transform spectrometry without Fourier analysis of the interferogram. That work argues that precision is gained by fitting theoretical time-domain lineshape expressions directly to the interferogram, thereby avoiding errors associated with Fourier integrals, apodization, peak fitting with standard interpolation functions, background selection, and Stokes energy shifts (Lagos et al., 2018). The measured interferogram is modeled as

$y = Ax$2

This method does not redefine the interferometer as non-Fourier in hardware terms; rather, it contests the necessity of Fourier-domain post-processing for certain spectroscopic inference tasks.

Together, these examples show that the “DFT interferometer” label is most exact when the hardware itself enforces a discrete set of delays or mixing operations, but Fourier-transform interferometry more generally encompasses devices whose outputs are continuous interferograms later discretized numerically.

6. Spatial DFT interferometry in radio and astronomical systems

In radio interferometry, discrete Fourier transform interferometry is instantiated through FFT beamforming. For an array of identical antennas on a regular lattice, beam synthesis can be performed with the spatial DFT, reducing cost from $y = Ax$3 to $y = Ax$4 (Masui et al., 2017). The voltage beamforming weights are

$y = Ax$5

the corresponding FFT beams are

$y = Ax$6

and the steering angles satisfy

$y = Ax$7

The formalism supports more than simple beam synthesis. The same work derives exact regridding for arbitrarily pointed tied-array beams,

$y = Ax$8

as well as windowed beamforming for sidelobe suppression and likelihood-based source localization (Masui et al., 2017). Under a simple uncorrelated-noise model, enough FFT beams preserve the information content of redundancy-stacked visibilities; the same paper also emphasizes that correlated noise from cross-talk or sky contributions makes standard Fourier-transform beamforming not strictly information preserving.

The wider theoretical setting is given by the proposal of arbitrary transform telescopes. There, classical interferometry is interpreted as deriving angular distributions from the Fourier transform of the electric field on the ground, but the Fourier transform is presented as one element of the broader class of linear canonical transforms (Lacki, 2015). The fractional Fourier transform is specifically identified as useful for sources that are close to the interferometer, because it effectively focuses the array at a finite distance rather than at infinity. The spatial Fourier transform relation is written as

$y = Ax$9

This framework clarifies why DFT interferometers dominate practical implementations: regular sampling, distant sources, and FFT algorithms make the Fourier basis computationally and physically natural.

A different integrated-optics astronomical direction is represented by astrointerferometry with discrete optics, where discrete diffraction in a two-dimensional array of coupled waveguides is used to determine the phase and amplitude of mutual correlation functions between any pair of three telescopes (Minardi et al., 2010). The output field at waveguide $A$0 is

$A$1

and time-averaged intensities form a linear system

$A$2

from which the real and imaginary parts of pairwise coherences can be recovered. The paper states that the method is not a direct discrete Fourier transform, but it is conceptually analogous in that it mixes input coherences into a finite set of output observables from which visibilities are reconstructed (Minardi et al., 2010).

7. Interpretation, boundaries of the term, and related misconceptions

The expression “discrete Fourier transform interferometer” can be misleading if treated as naming a single device type. The literature instead supports at least three technically distinct meanings.

First, in programmable photonics, it refers to a universal multiport interferometer architecture built from DFT mixing layers and phase masks (Girouard et al., 27 Aug 2025). Here the DFT is a physical mode-mixing primitive.

Second, in integrated and nonlinear spectroscopy, it refers to interferometers that acquire a discrete set of interferogram samples over path-delay states or path-delay scans, with reconstruction performed through a DFT or an equivalent calibrated inverse problem (Kita et al., 2018, Peternella et al., 2018, Lindner et al., 2020). Here the DFT acts primarily as an analysis operator on discretely sampled interferometric data.

Third, in radio astronomy, it denotes aperture arrays whose regular spatial sampling permits direct FFT beam synthesis (Masui et al., 2017). Here the DFT maps sampled electric fields or visibilities into beams on the sky.

A common misconception is that Fourier-transform interferometry necessarily requires explicit Fourier transformation of the interferogram. The time-domain fitting approach shows that, for some spectroscopic applications, physically motivated direct interferogram fitting can replace Fourier-domain processing while retaining the same interferometric measurement hardware (Lagos et al., 2018). Another misconception is that DFT-based interferometers are merely numerical reformulations of conventional architectures. In the universal multiport setting, the decomposition into DFT layers and phase masks is a hardware architecture with distinct robustness and fabrication properties, not just an alternative software synthesis method (Girouard et al., 27 Aug 2025).

The cited works also delineate the limits of DFT-centric thinking. Arbitrary transform telescopes argue that the Fourier basis is not uniquely privileged in principle, although it remains dominant because most astrophysical sources are distant and slowly varying and because FFT algorithms are efficient (Lacki, 2015). In radio beamforming, correlated noise and non-ideal array behavior mean that Fourier-transform formed beams do not strictly preserve all information contained in the full visibility matrix (Masui et al., 2017). In photonic spectroscopy, apparent resolution limits associated with finite Fourier sampling may be surpassed in parameter estimation through calibration and nonlinear inversion, as in the 400 fm sensor-modulation result achieved with a native $A$3 pm Fourier resolution (Peternella et al., 2018).

Taken together, these developments define the discrete Fourier transform interferometer not as a singular instrument but as a family of interferometric systems whose structure, sampling, or synthesis is organized around discrete Fourier-transform operations. The family spans universal linear optics, on-chip spectrum analysis, radio beamforming, and specialized differential or nonlinear interferometric sensing, with the DFT serving either as the physical mixer, the numerical inversion basis, or both.