Time-Wavelength Interleaving

• Time-wavelength interleaving is a set of techniques that use temporal staggering and periodic spectral separation to route and combine optical channels.
• It spans applications from integrated photonic interleavers for dense WDM to hybrid TWDM systems in optical access networks, enabling high-resolution filtering and efficient bandwidth management.
• Key implementations, such as Nyquist-filtering devices and RAC-based DWDM interleavers, demonstrate sub-GHz resolution, low crosstalk, and effective collision avoidance in scheduling.

Searching arXiv for papers directly relevant to time-wavelength interleaving and closely related interleaving architectures. Searching arXiv for papers on time-wavelength interleaving, photonic interleavers, and TWDM-related architectures. Time-wavelength interleaving denotes a family of optical and communication techniques in which temporal staggering, periodic spectral separation, or both are used to separate, combine, or schedule channels. In photonic integrated circuits, it generally refers to architectures that exploit structured delays and wavelength-selective splitting/combining so that signals are routed into different output ports in a regular, periodic spectral pattern; in optical access networks, hybrid time and wavelength division multiplexing allocates upstream traffic over time on multiple wavelengths; and in Nyquist wavelength-division multiplexing, a periodic optical filter can interleave or deinterleave odd and even spectral slices on a fixed grid (Choi et al., 22 May 2025, Bhar et al., 2017, Zhuang et al., 2015). Closely related work also uses temporal pulse interleaving, layer interlacing of passive and active photonic elements, and converter time interleaving, so the term spans several levels of abstraction rather than a single device class (Fordell et al., 19 Feb 2025, Friedman et al., 17 Jul 2025).

1. Conceptual scope and terminology

In the supplied literature, time-wavelength interleaving appears in three principal forms. First, in optical access networks, hybrid time and wavelength division multiplexing (TWDM) is the scheduling of upstream bursts over time and wavelength so that multiple optical network units can share optical line terminal resources without collision. Second, in integrated photonics, an interleaver/deinterleaver is a periodic filter that separates or combines alternate spectral slices, typically odd and even channels, on a fixed free-spectral-range grid. Third, in broader photonic architectures, “interlacing” can refer to an alternating sequence of passive and active transformation layers whose overall response is wavelength dependent.

The wavelength-domain interpretation is explicit in dense WDM and Nyquist WDM. In that setting, interleavers are the wavelength-domain counterpart of multiplexing logic: they separate or combine channels on an ITU grid with controlled phase relationships and channel spacing. The Nyquist-filtering device based on a two-ring-resonator-assisted Mach–Zehnder interferometer is described as a Nyquist-filtering interleaver/deinterleaver that can either interleave two sets of channels onto a denser Nyquist grid or deinterleave a dense super-channel into odd/even sub-bands for selective routing (Zhuang et al., 2015).

A complementary time-domain interpretation appears in orthogonal time division multiplexing. The work on Nyquist sinc sequences states that high-speed data transmission is enabled by time and wavelength division multiplexing, and that for higher data rates Nyquist orthogonal time division multiplexing (OTDM) can be combined with WDM. There, pulses are split into branches, delayed so that they land on the zero crossings of one another, and recombined to obtain an aggregate data rate of about $N$ times the single-channel rate (Dorostkar, 2019).

A common misconception is that “interleaving” always denotes literal alternation of time slots. The literature here does not support that narrow view. In the programmable silicon photonic circuit, the “interlacing” is instead the layer-by-layer alternation of passive coupled-waveguide arrays and tunable phase-shifter layers, producing programmable space-frequency linear transformations rather than classical temporal interleaving (Friedman et al., 17 Jul 2025). This suggests that the unifying concept is structured alternation and recombination across periodic degrees of freedom, not any single implementation modality.

2. Integrated photonic wavelength interleavers

Integrated photonic realizations of wavelength interleaving in the supplied literature fall into two distinct but related categories: Nyquist-filtering interleavers based on resonator-assisted interferometry, and broadband DWDM interleavers based on cascaded AMZIs or lattice filters with engineered couplers.

The Nyquist-filtering device reported in "Full-C-band, sub-GHz-resolution Nyquist-filtering (de)interleaver in photonic integrated circuit" is an on-chip optical Nyquist-filtering interleaver/deinterleaver implemented as a two-ring-resonator-assisted Mach–Zehnder interferometer (2RAMZI) in high-index-contrast Si$_3$N$_4$/SiO$_2$ waveguides (Zhuang et al., 2015). The circuit comprises a 2 × 2 asymmetric Mach–Zehnder interferometer, two ring resonators coupled to the MZI arms, and tunable couplers and heaters. Its key geometrical relation is

$$\Delta f_{\mathrm{FSR}} = \frac{1}{\Delta T} = \frac{c}{\Delta L\, n_g},$$

where $\Delta T$ is the optical delay associated with the MZI arm-length difference $\Delta L$. Reported performance includes sub-GHz resolution, a near-rectangular passband, 12.5 GHz 3-dB bandwidth, 13.5 GHz 25-dB bandwidth, 8% transition band, 25 dB passband-stopband extinction, full C-band coverage, and more than 160 effective FSRs of 25 GHz across a bandwidth over 4 THz. The device is further positioned as a building block for Nyquist super-channel generation and for enhanced WSS-based add/drop operation, with < 0.1 dB OSNR penalty relative to digitally filtered reference signals in transmitter experiments and < 0.5 dB OSNR penalty for continue/drop cases in the ROADM-related demonstration (Zhuang et al., 2015).

The broadband DWDM interleaver reported in "Rapid adiabatic couplers with arbitrary split ratios for broadband DWDM interleaver application" uses rapid adiabatic couplers (RACs) as the key building blocks of a cascaded AMZI/lattice-filter architecture (Choi et al., 22 May 2025). The stated motivation is that conventional directional couplers and MMI couplers are compact but strongly wavelength dependent, whereas conventional adiabatic couplers are broadband and fabrication tolerant but often relatively long. RACs address this tradeoff by combining translational offset and waveguide width control. The coupler geometry is parameterized by

$$w_{s,1} = w_{bus} + \frac{\Delta w_s}{2}, \qquad w_{s,2} = w_{bus} - \frac{\Delta w_s}{2},$$

and

$$w_{e,1} = 500\,\text{nm} + \frac{\Delta w_e}{2}, \qquad w_{e,2} = 500\,\text{nm} - \frac{\Delta w_e}{2},$$

with $\Delta w_e$ as the main knob for setting arbitrary split ratio. Fabricated RACs of four different target split ratios show power splitting within $_3$0 of the design target over a 160 nm wavelength range. Using these RACs, the authors implement an 8-channel DWDM interleaver exhibiting < $_3$1 dB crosstalk for the center 8 channels with flat-top passbands, $_3$2 dB flat-top bandwidth exceeding 30 GHz for each channel, and < $_3$3 dB crosstalk over more than 40 channels with 100 GHz spacing in the 1530–1570 nm range. A partial second-order configuration further shows < $_3$4 dB crosstalk across more than 60 channels with 200 GHz spacing over roughly 1525–1640 nm, except for one marginal channel (Choi et al., 22 May 2025).

These two implementations embody different design philosophies. The 2RAMZI is explicitly an ARMA-style photonic filter with a compact delay structure, whereas the RAC-based DWDM interleaver emphasizes broadband, wavelength-insensitive splitter ratios in cascaded interferometric stages. A plausible implication is that both approaches address the same systems-level requirement: preserving the phase-delay and interference conditions channel after channel across a wide optical band.

3. Temporal pulse interleaving for timing-transfer stabilization

A distinct but closely related use of interleaving appears in fibre-optic time and frequency transfer, where temporal interleaving is used to suppress polarization-dependent timing noise rather than to route wavelength channels. "Closed-Loop Polarization Mode Dispersion Mitigation for Fibre-Optic Time and Frequency Transfer" introduces a polarization-switching pulse interleaver in which orthogonally polarized pulses are alternated from pulse to pulse and averaged in the detection chain (Fordell et al., 19 Feb 2025).

The underlying observation is that, for a linear medium without polarization-dependent loss, the average timing of two orthogonal pulses is polarization independent. The paper writes the pulse timing as

$_3$5

with optical field represented by the Jones vector

$_3$6

and propagation through a linear medium described by

$_3$7

where $_3$8 contains polarization-independent loss and dispersion, and $_3$9 is a unitary matrix containing polarization-dependent effects. In the practical implementation, the interleaver uses a delay of half the pulse period, doubling the repetition rate from 250 MHz to 500 MHz by interleaving pulses of alternating orthogonal polarization. Because only two consecutive pulses are needed for averaging, the scheme is described as compatible with very high bandwidth PMD cancellation (Fordell et al., 19 Feb 2025).

The experimental setup uses an optically phase-locked frequency comb source, a polarization-switching pulse interleaver, a 30-km dispersion-compensated fibre link plus an additional 3-km DCF module, active phase stabilization in a closed loop, a balanced optical cross-correlator (BOC) for in-loop phase detection, and a photodiode-based out-of-loop phase meter. The comb pulses are filtered and amplified to about 6 nm bandwidth at 1560 nm, with pulse duration about 620 fs. A Faraday rotator at the remote end rotates the return polarization by $_4$0, making the round-trip used for locking polarization independent (Fordell et al., 19 Feb 2025).

The headline result is quantitative. The full link has a total differential group delay of about 300 fs, with the DCF module alone contributing about 100 fs. Without interleaving, and with the polarization controller paddles moved continuously in random patterns, the out-of-loop timing drift is about 300 fs peak-to-peak. With alternating orthogonal polarizations, the delay variation is reduced to below 20 fs peak-to-peak. The paper emphasizes that this is a closed-loop demonstration, showing that the method functions in a practical timing-stabilization configuration rather than only in open loop (Fordell et al., 19 Feb 2025).

The cancellation is not perfect. The remaining drift is attributed to polarization-dependent loss (PDL) and AM-to-PM conversion in the photodiodes. For direct detection, the limiting case is stated as $_4$1. In the experiment, $_4$2 and $_4$3, giving an estimated worst-case phase drift of about 9 fs due to PDL; adding a 5 fs peak-to-peak background phase variation yields an expected worst-case total near 14 fs peak-to-peak. Residual AM-to-PM conversion is analyzed using

$_4$4

These observations motivate a double-balanced phase detector, balanced both in power and in polarization. The paper’s design conclusion is that such a receiver could remove the dominant residual error channel left after PMD averaging, and that the method could potentially yield sub-femtosecond-level, long-term time transfer in long-haul fibre links utilizing standard single-mode fibres (Fordell et al., 19 Feb 2025).

4. Programmable interlacing in the space-frequency domain

The 2025 programmable silicon photonic work broadens the notion of interleaving by replacing temporal staggering with an alternating architecture of fixed dispersive mixers and tunable phase masks. "Programmable Space-Frequency Linear Transformations in Photonic Interlacing Architectures" experimentally demonstrates linear space-frequency transformations using a $_4$5-port programmable silicon photonic circuit with an alternating architecture (Friedman et al., 17 Jul 2025).

The device consists of passive layers that are identical coupled waveguide arrays acting as fixed unitary mixers $_4$6, interleaved with active diagonal phase-shifter layers $_4$7. The overall transformation is written as

$_4$8

with

$_4$9

For the fabricated circuit, $_2$0 and $_2$1, giving 20 tunable phase parameters in principle; the actual chip had only 17 of 20 microheaters functional, yet the device still performed well. The passive-layer dynamics are modeled by coupled-mode theory,

$_2$2

with $_2$3 a tridiagonal tight-binding Hamiltonian describing nearest-neighbor coupling (Friedman et al., 17 Jul 2025).

The architectural claim is not that a single coupled-waveguide array is strongly wavelength selective. Rather, the key point is that multiple such arrays in series create a richer effective spectral response: the passive layers are weakly dispersive, but the interlaced cascade of many layers plus tunable phase shifters produces a wavelength-dependent transformation with programmable matrix elements. The implementation uses a silicon-on-insulator platform with 500 nm × 220 nm single-mode waveguides embedded in silica cladding, operating in the telecom C-band around 1550 nm. The chip includes six waveguide arrays and five phase layers, optical packaging via V-groove fiber arrays with 8° polish angle and 250 μm pitch, and thermo-optic microheaters capable of at least $_2$4 phase shift (Friedman et al., 17 Jul 2025).

Programming is performed in situ using a tunable laser, a 4-channel optical switch network, polarization controllers, a 4-channel detector, an electrical SMU, and a Python-based iterative controller. The wavelength-dependent training objective is

$_2$5

with the Frobenius norm. Demonstrated applications include single-wavelength permutation routing and port-selective wavelength demultiplexing/filtering. The paper reports four representative wavelength-routing configurations, including 1520 nm and 1580 nm routed to output port 1, 1530 nm to port 1 and 1570 nm to port 2, and related multiwavelength assignments, with at least 10 dB extinction ratio in the worst case (Friedman et al., 17 Jul 2025).

This work is directly relevant to time-wavelength interleaving only in the broader sense that it studies interlaced photonic architectures that act jointly in the spatial and spectral domains. It does not implement a temporal interleaver in the classical sense of alternating time slots. That distinction is important: here, interleaving is architectural and spectral, not slot-based.

5. TWDM as a scheduling problem in optical access networks

At the network level, time-wavelength interleaving appears as a media-access and collision-avoidance problem. "Constrained Receiver Scheduling in Flexible Time and Wavelength Division Multiplexed Optical" treats TWDM as the next-generation passive optical access approach approved by FSAN, emphasizing that the OLT can allocate bandwidth to ONUs on multiple wavelengths, thereby providing bandwidth flexibility, security and privacy, passivity, and excellent reach (Bhar et al., 2017).

The paper’s principal contribution is the identification of an additional collision domain in secure flexible TWDM architectures. Standard schemes such as EFT, LFT, EFT-VF, LFT-VF schedule only the OLT receiver, thereby preventing receiver collision. In the 4th and 5th architectures, however, there are two collision domains: $_2$6, the receiver collision at the OLT, and $_2$7, a group collision at the switch. Group collision occurs because multiple ONUs may desire different receivers, but if they upstream simultaneously to the same switch port only one-to-one routing is possible. ONUs mapped to the same switch port form a group $_2$8, and only one ONU from a particular group can upstream at any instant (Bhar et al., 2017).

To address both constraints, the paper proposes the MAC protocol Constrained Earliest Void Filling (CEVF). It uses Request messages $_2$9 and Grant messages $$\Delta f_{\mathrm{FSR}} = \frac{1}{\Delta T} = \frac{c}{\Delta L\, n_g},$$0, with $$\Delta f_{\mathrm{FSR}} = \frac{1}{\Delta T} = \frac{c}{\Delta L\, n_g},$$1. Scheduling is formulated in terms of receiver voids $$\Delta f_{\mathrm{FSR}} = \frac{1}{\Delta T} = \frac{c}{\Delta L\, n_g},$$2 and group voids $$\Delta f_{\mathrm{FSR}} = \frac{1}{\Delta T} = \frac{c}{\Delta L\, n_g},$$3. The earliest possible scheduling instant is

$$\Delta f_{\mathrm{FSR}} = \frac{1}{\Delta T} = \frac{c}{\Delta L\, n_g},$$4

and candidate-void intersection is defined through

$$\Delta f_{\mathrm{FSR}} = \frac{1}{\Delta T} = \frac{c}{\Delta L\, n_g},$$5

$$\Delta f_{\mathrm{FSR}} = \frac{1}{\Delta T} = \frac{c}{\Delta L\, n_g},$$6

A valid void must satisfy

$$\Delta f_{\mathrm{FSR}} = \frac{1}{\Delta T} = \frac{c}{\Delta L\, n_g},$$7

where $$\Delta f_{\mathrm{FSR}} = \frac{1}{\Delta T} = \frac{c}{\Delta L\, n_g},$$8 is the upstream transmission time and $$\Delta f_{\mathrm{FSR}} = \frac{1}{\Delta T} = \frac{c}{\Delta L\, n_g},$$9 is the receiver tuning time included in the grant duration. The selected start time is $\Delta T$0, and the grant scheduling instant is $\Delta T$1 (Bhar et al., 2017).

Simulation results in OMNET++ use 64 ONUs, 1 Gbps ONU and OLT transceivers, $\Delta T$2, ONU line rates of 31.25 Mbps, 62.6 Mbps, and 125 Mbps, 1 Gb ONU buffers, and self-similar Pareto on/off traffic with shape parameters 1.2 and 1.4. At 31.25 Mbps and 62.5 Mbps, CEVF achieves higher throughput than EFT-VF on the secure flexible architectures because EFT-VF suffers packet loss due to group collisions. At $\Delta T$3, throughput for EFT-VF on Fig. 1(a) and for CEVF is approximately 99%. At 125 Mbps, CEVF throughput saturates at approximately 80% for $\Delta T$4 and 85% for $\Delta T$5 because group constraints create unused voids and because ONUs of a group effectively share only one receiver’s line rate, with $\Delta T$6 (Bhar et al., 2017).

The theoretical throughput upper bound is given as

$\Delta T$7

The same paper also reduces online search complexity through void ordering and void hopping, proving that the modified CEVF requires at most $\Delta T$8 steps to find the appropriate void and has worst-case complexity $\Delta T$9. This network-level formulation makes clear that time-wavelength interleaving is not solely a filter-design problem; it is equally a constrained scheduling problem across multiple collision domains (Bhar et al., 2017).

6. Adjacent interleaving frameworks and open extensions

Several adjacent works do not implement wavelength interleaving directly, but they clarify how interleaving behaves when parallel channels are combined, calibrated, or generalized. Their relevance is methodological rather than literal.

"A Direct Calibration Algorithm for ADC Interleaving" studies a $\Delta L$0-way time-interleaved ADC system in which multiple ADC cores sample the same analog input in a staggered pattern to increase effective sampling rate (Chan et al., 25 Nov 2025). Each core samples with interval $\Delta L$1, the interleaved time step is $\Delta L$2, and the measured sample is modeled as

$\Delta L$3

with mismatch parameters $\Delta L$4 (phase delay/timing skew), $\Delta L$5 (gain), and $\Delta L$6 (offset). Using a sinusoidal reference $\Delta L$7, the paper shows that mismatch signatures separate into specific Fourier families:

$\Delta L$8

and

$\Delta L$9

This permits direct recovery by inverse DFT, with overall complexity reported as $$w_{s,1} = w_{bus} + \frac{\Delta w_s}{2}, \qquad w_{s,2} = w_{bus} - \frac{\Delta w_s}{2},$$0. In the noiseless case, once sampling exceeds the theoretical minimum, gain and offset can be recovered to machine precision, often around $$w_{s,1} = w_{bus} + \frac{\Delta w_s}{2}, \qquad w_{s,2} = w_{bus} - \frac{\Delta w_s}{2},$$1–$$w_{s,1} = w_{bus} + \frac{\Delta w_s}{2}, \qquad w_{s,2} = w_{bus} - \frac{\Delta w_s}{2},$$2, while phase delay recovery is more sensitive. The paper explicitly states that it does not directly handle bandwidth mismatch or wavelength-domain physical effects without an analogous model (Chan et al., 25 Nov 2025).

"Analog Time Multiplexing for Digital-to-Analog Conversion" shifts the multiplexing stage of a time-interleaved sigma-delta DAC from the digital domain into the analog domain (Martínez-Heredia et al., 26 Feb 2026). Instead of one high-speed DAC operating at $$w_{s,1} = w_{bus} + \frac{\Delta w_s}{2}, \qquad w_{s,2} = w_{bus} - \frac{\Delta w_s}{2},$$3, the architecture uses $$w_{s,1} = w_{bus} + \frac{\Delta w_s}{2}, \qquad w_{s,2} = w_{bus} - \frac{\Delta w_s}{2},$$4 low-speed DACs clocked at $$w_{s,1} = w_{bus} + \frac{\Delta w_s}{2}, \qquad w_{s,2} = w_{bus} - \frac{\Delta w_s}{2},$$5 with relative phase shifts of $$w_{s,1} = w_{bus} + \frac{\Delta w_s}{2}, \qquad w_{s,2} = w_{bus} - \frac{\Delta w_s}{2},$$6, and sums their outputs in analog. The resulting stage is equivalent to filtering by a comb filter,

$$w_{s,1} = w_{bus} + \frac{\Delta w_s}{2}, \qquad w_{s,2} = w_{bus} - \frac{\Delta w_s}{2},$$7

which acts as a moving-average filter with zeros at $$w_{s,1} = w_{bus} + \frac{\Delta w_s}{2}, \qquad w_{s,2} = w_{bus} - \frac{\Delta w_s}{2},$$8. The jitter result is explicit:

$$w_{s,1} = w_{bus} + \frac{\Delta w_s}{2}, \qquad w_{s,2} = w_{bus} - \frac{\Delta w_s}{2},$$9

For $$w_{e,1} = 500\,\text{nm} + \frac{\Delta w_e}{2}, \qquad w_{e,2} = 500\,\text{nm} - \frac{\Delta w_e}{2},$$0, the paper reports about 12 dB improvement, matching simulations, while ideal-case SNDR remains essentially unchanged at 69.706 dB for the conventional output and 69.715 dB for the proposed output. The trade-off is comb-filter passband distortion; for $$w_{e,1} = 500\,\text{nm} + \frac{\Delta w_e}{2}, \qquad w_{e,2} = 500\,\text{nm} - \frac{\Delta w_e}{2},$$1, $$w_{e,1} = 500\,\text{nm} + \frac{\Delta w_e}{2}, \qquad w_{e,2} = 500\,\text{nm} - \frac{\Delta w_e}{2},$$2, and the paper estimates $$w_{e,1} = 500\,\text{nm} + \frac{\Delta w_e}{2}, \qquad w_{e,2} = 500\,\text{nm} - \frac{\Delta w_e}{2},$$3 for less than 3 dB amplitude distortion (Martínez-Heredia et al., 26 Feb 2026).

The most speculative extension is "Complementary Fractional Dimensional Order of Nyquist Sinc Sequences for Time Division Multiplexing" (Dorostkar, 2019). It proposes complementary Nyquist sinc sequences (CNSS) and fractional-dimensional order Nyquist sinc sequences for OTDM, beginning from cosine-series and sine-series representations of Nyquist sinc sequences:

$$w_{e,1} = 500\,\text{nm} + \frac{\Delta w_e}{2}, \qquad w_{e,2} = 500\,\text{nm} - \frac{\Delta w_e}{2},$$4

for odd $$w_{e,1} = 500\,\text{nm} + \frac{\Delta w_e}{2}, \qquad w_{e,2} = 500\,\text{nm} - \frac{\Delta w_e}{2},$$5, and

$$w_{e,1} = 500\,\text{nm} + \frac{\Delta w_e}{2}, \qquad w_{e,2} = 500\,\text{nm} - \frac{\Delta w_e}{2},$$6

as the complementary sequence for even $$w_{e,1} = 500\,\text{nm} + \frac{\Delta w_e}{2}, \qquad w_{e,2} = 500\,\text{nm} - \frac{\Delta w_e}{2},$$7. It further argues that a desired optical delay can be realized through an electrical phase shifter via

$$w_{e,1} = 500\,\text{nm} + \frac{\Delta w_e}{2}, \qquad w_{e,2} = 500\,\text{nm} - \frac{\Delta w_e}{2},$$8

The same paper states that the fractional-dimensional trajectory is not unique, that the “best set of solutions” must be selected for data transmission, that the dimensional transformation and inverse transformation are under further investigations, and that the work lacks complete simulation or experimental validation. The notion is therefore an open conceptual extension rather than an established architecture (Dorostkar, 2019).

Taken together, these adjacent works show that interleaving creates structured periodic artifacts when channels are mismatched, and that those artifacts can sometimes be canceled, filtered, or calibrated by exploiting periodicity itself. This suggests a broader methodological link between time-wavelength interleaving, converter interleaving, and programmable photonic interlacing: all depend on careful control of relative delay, phase, and channel weighting, even when the physical platforms are different.