Photonic Lanterns: Efficient Mode Converters

• Photonic lanterns are tapered waveguide devices that convert multimode optical fields into multiple single-mode outputs while preserving optical étendue.
• They enable low-loss interfacing between multimode and single-mode photonic systems, serving key roles in spectroscopy, wavefront sensing, and coherent imaging.
• Design and fabrication emphasize mode-number matching and adiabatic tapering to achieve controlled supermode evolution and efficient, broadband performance.

Searching arXiv for recent and foundational papers on photonic lanterns to ground the article in published work. Photonic lanterns (PLs) are guided-wave devices that adiabatically convert a multimode (MM) input field into multiple single-mode (SM) outputs, or equivalently transform an array of SM channels into a common MM port. In the ideal adiabatic limit, the MM end supports $N$ orthogonal modes and the transition maps these $N$ input modes one-to-one into $N$ SM outputs, enabling low-loss interfacing between multimode optics and single-mode photonic systems (Sengupta et al., 25 Jun 2026). Their significance derives from this combination of étendue preservation, compatibility with fiber and integrated photonics, and the ability to deliver stable SM outputs while retaining information about the incident complex field. In astronomy and related photonic systems, PLs have accordingly become important for diffraction-limited spectrometry, focal-plane wavefront sensing, coherent imaging, and multimode transport with reduced focal-ratio degradation (Lin et al., 2021, Lin et al., 2022, Kim et al., 2024, Benoît et al., 2020).

1. Definition, modal basis, and adiabatic conversion

A PL is a tapered waveguide device that converts a few- or multimode optical field at its input into a set of single-mode outputs (Kim et al., 2024). In the common fiber embodiment, multiple SM fibers or cores are brought together, fused, and tapered so that their coupled supermodes evolve into the guided modes of a single larger multimode core; in the reverse direction, the MM field is decoded into multiple SM channels (Leon-Saval et al., 2015, Birks et al., 2015). The transition is reciprocal and low-loss if the number of guided states is conserved and the taper is sufficiently gradual.

The central design condition is mode-number matching. For a weakly guiding step-index entrance with core radius $a$, numerical aperture $\mathrm{NA}$, and wavelength $\lambda$, the normalized frequency is

$V(\lambda) = \frac{2\pi a\,\mathrm{NA}}{\lambda},$

and the approximate number of guided modes is

$M(\lambda) = \frac{V(\lambda)^2}{2}$

in the conventions used by several of the cited studies (Lin et al., 2021, Sengupta et al., 25 Jun 2026). Near-lossless operation requires

$N \ge M(\lambda),$

across the operating band, because $M$ increases toward shorter wavelengths (Lin et al., 2021). Earlier review formulations express the same design principle as preservation of guided degrees of freedom and optical étendue, with mismatch imposing an irreducible insertion-loss floor (Leon-Saval et al., 2015, Birks et al., 2015).

Adiabaticity is the second core principle. In coupled-mode form, suppression of unwanted modal transfer requires

$N$0

so the taper must evolve slowly relative to the spacing of local propagation constants (Benoît et al., 2020, Kim et al., 2023). This condition underlies low insertion loss, low modal crosstalk, and preservation of a deterministic modal mapping. When it is violated, power couples into radiation modes or unintended guided modes, degrading throughput and calibration fidelity.

A useful physical distinction is between standard, mode-selective, and hybrid PLs. Standard PLs generally mix MM entrance modes across outputs; mode-selective PLs are engineered so that a given MM mode maps predominantly to a designated SM channel; hybrid devices isolate one selected mode, often the fundamental LP$N$1, while allowing the remaining modal subspace to mix (Lin et al., 2023, Morsy et al., 15 Apr 2026). This taxonomy matters because the desired behavior differs between applications: spectroscopy often benefits from stable SM outputs and high total throughput, whereas some wavefront-sensing or nulling architectures require controlled modal selectivity.

2. Transfer matrices, principal modes, and coherent metrology

At a fixed wavelength, an $N$2-port PL is naturally described by an $N$3 complex transfer matrix $N$4 that maps modal coefficients at the MM end into complex field amplitudes at the SM outputs (Sengupta et al., 25 Jun 2026). More generally, when the number of SM inputs and MM modes differ, the encoding map can be written

$N$5

with $N$6 (Dobias et al., 23 Apr 2026). This matrix contains the full amplitude and phase information needed for beam forming, imaging, or modal reconstruction.

A particularly useful characterization is through principal modes. In the formulation used for laboratory characterization, one excites each SM port in the backward direction and measures the resulting complex field at the MM end; the recovered field $N$7 is the principal mode associated with port $N$8 (Sengupta et al., 25 Jun 2026). In an adiabatic lantern, if the input field matches $N$9, nearly all output power is delivered to the corresponding SM port. This makes principal modes both physically intuitive and operationally important, because they furnish a direct basis for coupling analysis, calibration, and wavefront sensing.

The complex field can be recovered by off-axis digital holography. If the unknown signal field $N$0 interferes with a tilted reference field $N$1, the recorded intensity is

$N$2

By placing the reference slightly off-axis, the cross-terms are shifted to sidebands in Fourier space; Fourier filtering then reconstructs the signal field: $N$3 where $N$4 selects one off-axis sideband (Sengupta et al., 25 Jun 2026). This approach was used to recover all 133 principal modes of a seven-by-nineteen PL array at 632.8 nm in a recent laboratory study (Sengupta et al., 25 Jun 2026).

The wavelength dependence of $N$5 has emerged as a central issue. Direct measurements of a 19-port non-mode-selective lantern across 1525–1575 nm showed rapid spectral evolution of the complex-valued transfer matrix, with the dominant mechanism identified as differential modal phase accumulation in the MM section (Dobias et al., 23 Apr 2026). For mode $N$6,

$N$7

and

$N$8

so increasing the MM section length $N$9 accelerates spectral decorrelation (Dobias et al., 23 Apr 2026). This provides a direct design lever: longer MM sections increase wavelength sensitivity, whereas shorter sections and reduced modal dispersion favor broadband stability.

This suggests that PL calibration is intrinsically a full-field, wavelength-dependent problem rather than a static port-to-port throughput problem. A plausible implication is that applications relying on coherent inference or broad spectral coverage must measure or model $a$0 explicitly, not merely the integrated port powers.

3. Physical implementations and fabrication strategies

PLs have been realized through several fabrication routes. A major division is between fused-bundle or multicore-fiber devices and three-dimensional integrated waveguide devices (Leon-Saval et al., 2015). In fused-bundle implementations, discrete SM fibers are packed within a low-index capillary and thermally tapered until the individual cores merge into a common MM region (Davenport et al., 2021, Vievard et al., 2024). In multicore-fiber implementations, many SM cores share a common cladding before tapering, which improves scalability to large port counts (Birks et al., 2015). Ultrafast laser inscription offers an integrated-photonic route with arbitrary three-dimensional geometries, albeit with different propagation-loss trade-offs (Leon-Saval et al., 2015).

Packing topology matters. For close-packed equal-diameter SMFs inside a circular capillary, circle-packing theory gives practical design targets for capillary inner diameter, fill fraction, and ring structure (Davenport et al., 2021). In fabricated 19- and 37-core fused bundles, packing densities of 95% and 99% of theoretically achievable values were reported, with mean adjacent core separations of 1.03 and 1.08 fiber diameters, respectively (Davenport et al., 2021). These ordered, concentric-ring packings help produce a more circular and uniform fused MM core and improve the correspondence between SM-array supermodes and MM modal structure.

A recent example of large-scale device fabrication is the monolithic glass array used for the WaveDriver concept: seven independent 19-port PLs integrated in one component, for a total of 133 SM ports (Sengupta et al., 25 Jun 2026). Four lanterns used a uniform “A” design and three a “B” design with alternating core sizes and a depressed-well structure. The device was fabricated by fiber-bundle/tower tapering rather than ultrafast laser inscription. Common parameters included taper factor 7.14, taper length 1 cm, $a$1, $a$2, $a$3, $a$4, and outer cladding diameter $a$5 (Sengupta et al., 25 Jun 2026). The A design used 19 SM cores in a hexagonal lattice with 3.11 $a$6 core diameter and 27.4 $a$7 pitch, while the B design used alternating diameters of 4.04, 4.60, and 5.22 $a$8, 28.4 $a$9 pitch, and a depressed well (Sengupta et al., 25 Jun 2026).

Visible-light implementations pose stricter mode-count and alignment constraints because the mode count rises at shorter wavelengths. A 19-port visible PL integrated on SCExAO/Subaru used separate SM fibers bundled in a low-index capillary and tapered to form an MM input of about 25 $\mathrm{NA}$0 core and about 99 $\mathrm{NA}$1 cladding diameter, with measured transmission at 780 nm of 88.9% on average, ranging from 83% to 95% (Vievard et al., 2024). That device fed a pseudo-slit spectrograph and demonstrated the first visible-light on-sky PL spectroscopy (Vievard et al., 2024).

These examples show that fabrication is not merely a manufacturing detail. Core geometry, lattice ordering, taper profile, and index homogeneity directly determine adiabaticity, spectral behavior, polarization sensitivity, and the conditioning of downstream inverse problems.

4. Spectroscopy, throughput, and focal-ratio behavior

The principal instrumental appeal of PLs is that they can couple aberrated or time-varying focal-plane light into multiple diffraction-limited SM beams with higher total throughput than direct SMF injection while preserving the spectrographic advantages of SM delivery (Lin et al., 2021). In simulations for diffraction-limited spectrometry in the $\mathrm{NA}$2 band, coupling efficiency into PLs scaled roughly linearly with Strehl for AO-filtered Kolmogorov turbulence. At 10% Strehl and $\mathrm{NA}$3, a 3-port PL achieved approximately $\mathrm{NA}$4 SMF throughput, a 6-port approximately $\mathrm{NA}$5, and a 19-port approximately $\mathrm{NA}$6 (Lin et al., 2021).

Beam shaping with phase-induced amplitude apodization improves few-port performance, especially for SMFs and 3-port PLs. In those simulations, PIAA gave an approximately 30% relative coupling boost for an SMF and approximately 25% for a 3-port PL, whereas gains became negligible beyond about 10 modes (Lin et al., 2021). The design implication stated there is explicit: if pixel count is constrained, a 3-port PL with PIAA provides a favorable trade; if photon noise dominates and pixel budget is available, larger $\mathrm{NA}$7 provides more throughput and tip-tilt resilience (Lin et al., 2021).

Visible-light measurements on Subaru/SCExAO make the throughput advantage concrete under real atmospheric residuals. In laboratory characterization, the visible PL achieved on-axis coupling of 61.8% at 642 nm, a peak coupling across the map of 80%, and a best averaged coupling over the field of view of 51% $\mathrm{NA}$8 10% at $\mathrm{NA}$9 (Vievard et al., 2024). On sky, under median seeing and large tip/tilt residuals, the average coupling efficiency was 14.5% $\lambda$0 7.4% with a maximum of 42.8% at 680 nm (Vievard et al., 2024). The same study argued that a visible SMF would have been strongly limited by its $\lambda$1 mas acceptance at 700 nm, whereas the PL at $\lambda$2 delivered a field of view of about 80 mas (Vievard et al., 2024).

The dependence on focal ratio follows directly from acceptance angle. For the 633 nm WaveDriver lanterns, $\lambda$3 implies an acceptance half-angle in air of $\lambda$4 and an optimal injection near $\lambda$5; simulations predicted peak-throughput values of $\lambda$6 and $\lambda$7 for the two designs (Sengupta et al., 25 Jun 2026). On muirSEAL, a 19-port lantern with $\lambda$8 implied an optimum injection $\lambda$9, consistent with a measured throughput peak near $V(\lambda) = \frac{2\pi a\,\mathrm{NA}}{\lambda},$0, though the best measured injection efficiency was only 31.6%, well below the expected $V(\lambda) = \frac{2\pi a\,\mathrm{NA}}{\lambda},$1, suggesting alignment-dominated loss rather than inherent lantern loss (Sengupta et al., 27 Aug 2025).

PLs also mitigate focal-ratio degradation in multimode transport when implemented in mode-selective PL–MCF–PL links. A 7 m six-mode link built from two mode-selective PLs bracketing a six-core MCF showed significantly improved FRD relative to a conventional six-mode MMF at 1550 nm, because transfer of power from lower-order to higher-order modes was inhibited (Benoît et al., 2020). This is a distinct but related expression of the same advantage: PL architectures can preserve modal structure and output angular distribution better than unconstrained multimode transport.

5. Wavefront sensing and adaptive-optics roles

PLs have developed into a focal-plane wavefront-sensing modality because their output intensities encode the phase-sensitive overlap of the focal-plane field with the lantern’s principal modes (Lin et al., 2022). In the PLWFS formalism, if $V(\lambda) = \frac{2\pi a\,\mathrm{NA}}{\lambda},$2 is the principal mode associated with port $V(\lambda) = \frac{2\pi a\,\mathrm{NA}}{\lambda},$3, the coupling coefficient is

$V(\lambda) = \frac{2\pi a\,\mathrm{NA}}{\lambda},$4

and the measured output is

$V(\lambda) = \frac{2\pi a\,\mathrm{NA}}{\lambda},$5

For small aberrations expanded in a modal basis $V(\lambda) = \frac{2\pi a\,\mathrm{NA}}{\lambda},$6, the measurement vector obeys a linearized relation

$V(\lambda) = \frac{2\pi a\,\mathrm{NA}}{\lambda},$7

with $V(\lambda) = \frac{2\pi a\,\mathrm{NA}}{\lambda},$8 the response matrix (Lin et al., 2023).

Theoretical and numerical analyses have quantified the usable linear and nonlinear regimes. For few-port PLWFSs at 1550 nm, good linearity was reported out to about $V(\lambda) = \frac{2\pi a\,\mathrm{NA}}{\lambda},$9–$M(\lambda) = \frac{V(\lambda)^2}{2}$0 rad RMS wavefront error, depending on lantern type and basis, while degeneracy onset was around $M(\lambda) = \frac{V(\lambda)^2}{2}$1–$M(\lambda) = \frac{V(\lambda)^2}{2}$2 rad RMS for most configurations (Lin et al., 2023). Throughput rose with port count, from about $M(\lambda) = \frac{V(\lambda)^2}{2}$3 to $M(\lambda) = \frac{V(\lambda)^2}{2}$4, and hybrid lanterns with beam shaping could direct about 80% of unaberrated flux into a selective science port at the cost of reduced linear range (Lin et al., 2023). The broader implication is that PLWFS is not a single sensor but a family of sensors whose observability, conditioning, and throughput depend strongly on port count, taper geometry, symmetry breaking, and auxiliary optics.

Laboratory and on-sky demonstrations have confirmed this role. On the SEAL testbed, a 19-port PL was used in a multi-WFS single-conjugate AO architecture with a modulated pyramid WFS, the PL handling low-order aberrations and non-common-path aberrations while the pyramid handled higher-order atmospheric content (Sengupta et al., 2024). Three reconstruction strategies were compared—linear inversion, neural network, and Gerchberg–Saxton—with the neural network giving the best overall performance in simulations and a held-out test-set root-mean-squared error of $M(\lambda) = \frac{V(\lambda)^2}{2}$5 rad (Sengupta et al., 2024). The same work introduced an experimental method to identify the PL propagation matrix $M(\lambda) = \frac{V(\lambda)^2}{2}$6 from intensity-only measurements, making model-based reconstructors practical (Sengupta et al., 2024).

Recent segmented-aperture work extended PL sensing to phase discontinuities. A 19-port lantern was used to reconstruct segment piston offsets on the muirSEAL testbed, with a representative linear range of about $M(\lambda) = \frac{V(\lambda)^2}{2}$7 rad, corresponding to $M(\lambda) = \frac{V(\lambda)^2}{2}$8 nm at $M(\lambda) = \frac{V(\lambda)^2}{2}$9, and minimal cross-talk at small amplitudes (Cuevas et al., 29 Aug 2025). In simulation, closed-loop control over 50 iterations improved Strehl from 0.902 to 0.992 while reducing RMS piston error from 82.61 nm to 0.273 nm (Cuevas et al., 29 Aug 2025). This explicitly addresses a limitation of many conventional pupil-plane sensors, which show poor sensitivity to discontinuities at internal pupil boundaries.

An on-sky second-stage control demonstration on the Shane 3 m used a non-dispersed 19-port PL after first-stage AO correction, reducing PL-reconstructed residual WFE from about 3.8 nm RMS to about 1.3 nm RMS and increasing the PSF peak by about 15% in closed loop (Sengupta et al., 25 Nov 2025). The loop targeted quasi-static NCPAs, not atmospheric residuals, because the overall DM command latency was about 1 s (Sengupta et al., 25 Nov 2025). This indicates that even narrowband, non-dispersed PLs can serve as minimally invasive retrofits for second-stage focal-plane control.

6. Coherent imaging, spectroastrometry, and hybrid science architectures

PLs are not limited to photometric sensing. Because each SM output carries a stabilized complex amplitude corresponding to a modal projection of the pupil field, pairs of outputs can be interfered to measure modal coherences (Kim et al., 2024). In a 6-port coherent-imaging study, the pupil-plane principal modes $N \ge M(\lambda),$0 were interpreted as complex-valued effective apertures. If

$N \ge M(\lambda),$1

then pairwise interference in an ABCD combiner yields the mutual coherence $N \ge M(\lambda),$2 through four phase-stepped intensities: $N \ge M(\lambda),$3 Simulations showed that PL interferometric observables behave similarly to aperture masking interferometry on scales below $N \ge M(\lambda),$4, but with greater sensitivity to symmetries and the ability to break some 180° position-angle degeneracies (Kim et al., 2024).

PL-fed spectroastrometry uses a different aspect of the same modal redistribution. In a 6-port formulation, the normalized output intensities vary linearly with small tip-tilts, so the wavelength-dependent centroid can be recovered from port spectra without rotating a slit (Kim et al., 2024). The linearized retrieval is written

$N \ge M(\lambda),$5

with centroid estimate

$N \ge M(\lambda),$6

(Kim et al., 2024). Mock observations of accreting protoplanets showed recoverable line-driven astrometric signals in both linear and nonlinear regimes, with exposure times of 1–3 hr under assumed AO jitter (Kim et al., 2024). ELT-oriented studies further argued that PLs extend spectroastrometric and interferometric science to smaller angular scales and higher contrasts by combining $N \ge M(\lambda),$7 sensitivity scaling with simultaneous two-dimensional centroid sensing (Kim et al., 2023).

Hybrid science-and-control architectures are a further development. The Hybrid-Mode Selective Photonic Lantern (HMSPL) is designed so that the LP$N \ge M(\lambda),$8 mode is directed into a central SMF feeding a mid-resolution spectrograph, while surrounding SMFs route coherent starlight to a low-resolution spectrograph for wavefront sensing (Morsy et al., 15 Apr 2026). The measurement model is written

$N \ge M(\lambda),$9

for the sensing channels, with the same lantern simultaneously serving as science feed and focal-plane WFS (Morsy et al., 15 Apr 2026). In this conception, PLs become common-path interfaces that eliminate non-common-path aberrations between spectroscopy and sensing, an explicit requirement for deep-contrast, long-duration observations such as those envisaged for HWO (Morsy et al., 15 Apr 2026).

This suggests that PLs are increasingly better understood not as isolated couplers but as modal front ends for integrated photonic systems, in which throughput, sensing, interferometry, and spectroscopy are co-designed.

7. Limitations, discrepancies, and current research directions

A persistent issue is the mismatch between simulation and experiment. The 2026 laboratory characterization of the seven-by-nineteen WaveDriver device found that measured coupling matrices varied substantially between lanterns; some reproduced the simulated upper-left correlation cluster, others did not, and the banded adjacency patterns seen in simulation were not prominent experimentally (Sengupta et al., 25 Jun 2026). Three A lanterns matched simulations closely, but one A lantern showed lower sensitivity and stronger inner-ring correlations, while B lanterns generally underperformed relative to simulation (Sengupta et al., 25 Jun 2026). The authors attributed this to unmodeled depressed-well structure, possible non-adiabatic sections, residual stress birefringence, and fabrication drift (Sengupta et al., 25 Jun 2026).

Absolute throughput and crosstalk remain incompletely characterized in some recent experimental campaigns. The WaveDriver study normalized modes to simulated peak throughput values of 0.946 for A and 0.905 for B and therefore did not directly report insertion loss or modal crosstalk (Sengupta et al., 25 Jun 2026). Similarly, some wavefront-sensing studies prioritized response matrices and dynamic range over full optical metrology (Sengupta et al., 27 Aug 2025). This does not invalidate the sensing conclusions, but it means direct system comparisons must be made carefully.

Chromaticity is another fundamental constraint. The full-field transfer matrix of non-mode-selective lanterns can decorrelate rapidly with wavelength, with a half-power correlation near $M$0 nm in one measured 19-port device (Dobias et al., 23 Apr 2026). Broad-band wavefront sensing or spectroscopy therefore requires either spectrally resolved calibration, lantern designs with intentionally reduced modal phase evolution, or architectures that exploit rather than suppress chromatic transfer-matrix structure (Dobias et al., 23 Apr 2026, Sengupta et al., 25 Jun 2026). The failure of broad H-band non-dispersed sensing in the Shane on-sky test, and its restoration with a 30 nm band centered at 1550 nm, is consistent with this broader point (Sengupta et al., 25 Nov 2025).

Other active issues include polarization dependence, thermal and mechanical drift, scaling to larger port counts, and conditioning of inverse problems. Several studies explicitly recommend tighter control of core diameter and pitch, better taper design to enforce stricter adiabaticity, annealing to reduce stress birefringence, and regular recalibration of principal modes or transfer matrices to track drift (Sengupta et al., 25 Jun 2026, Dobias et al., 23 Apr 2026). Nonlinear reconstruction, spectral dispersion, and polarization separation are repeatedly proposed as ways to increase effective degrees of freedom and extend dynamic range (Lin et al., 2023, Cuevas et al., 29 Aug 2025).

A recurring misconception is that PLs are simply efficient multimode couplers. The literature shows that this is incomplete. Standard PLs do provide efficient MM-to-SM conversion, but their practical value increasingly lies in the structured modal information preserved in the output, whether for spectroscopy, interferometry, or focal-plane control (Lin et al., 2022, Kim et al., 2024, Sengupta et al., 25 Jun 2026). Another misconception is that more ports automatically solve all sensing or throughput problems. The cited work instead shows explicit trade-offs: larger $M$1 generally improves throughput and mode coverage, but increases calibration complexity, detector footprint, and often the difficulty of maintaining uniform fabrication and well-conditioned reconstruction (Lin et al., 2021, Lin et al., 2023).

Current research is therefore converging on a set of linked objectives: physically accurate full-field modeling including measured index profiles and depressed-well structures; broadband transfer-matrix metrology; improved fabrication control for higher uniformity; hybrid lanterns that allocate science and sensing modal subspaces deliberately; and closed-loop demonstrations in realistic high-contrast systems (Sengupta et al., 25 Jun 2026, Morsy et al., 15 Apr 2026, Sengupta et al., 25 Nov 2025). The cumulative picture is that PLs now occupy a distinct position in photonics: they are modal transducers whose utility depends as much on their coherent transfer properties as on their raw throughput.