Second-Generation TDI for LISA-like Missions

• Second-generation TDI is an advanced interferometric technique designed for moving gravitational-wave detectors with time-varying, unequal arm lengths to effectively cancel dominant laser noise.
• Its methodology constructs ordered delayed combinations of one-way measurements using both algebraic and geometric approaches to suppress noise induced by inter-spacecraft motion.
• This framework underpins LISA-like missions by reliably mitigating laser and clock noise while enhancing gravitational-wave sensitivity across multiple observable configurations.

Searching arXiv for recent and foundational papers on second-generation TDI. {"query": "\"second-generation time-delay interferometry\" OR \"second-generation TDI\" LISA", "max_results": 10} {"query":"\"Time Delay Interferometry\" LISA second-generation Michelson Relay sensitivity null frequencies", "max_results": 10} Second-generation time-delay interferometry (TDI) is the extension of time-delay interferometry from a stationary unequal-arm array to a realistic rotating and flexing constellation with unequal and time-dependent arm lengths. In space-borne interferometers such as LISA-like missions, the one-way light-travel times vary because of orbital motion, so the delay operators no longer commute, and first-generation TDI is not sufficient to cancel the dominant laser phase or frequency noise. The second-generation framework constructs ordered combinations of delayed one-way measurements that cancel laser noise to first order in the inter-spacecraft velocities, or suppress it below the secondary noises intrinsic to the heterodyne measurements, while preserving gravitational-wave response in the science band (Tinto et al., 2022, Tinto et al., 2018, Wang, 24 Jul 2025).

1. Physical setting and necessity

The defining physical context of second-generation TDI is a three-spacecraft interferometer with unequal and time-varying arms. For LISA, the triangular constellation undergoes orbital motion, so the arm lengths are unequal by about $\pm 1\%$ and time-varying, with spacecraft relative drifts up to about $10~\mathrm{m/s}$. In this regime, laser frequency noise exceeds the target sensitivity by more than seven orders of magnitude, and simple equal-arm cancellation is impossible (Hartwig et al., 2021, Zhang, 3 Nov 2025).

The standard hierarchy distinguishes first-generation TDI, which assumes a rigid static constellation; modified first-generation TDI, which accounts for rigid rotation; and second-generation TDI, which accounts for full flexing or orbital dynamics. The practical reason for the higher generation is that the residual laser noise left by noncommuting delays in a moving array is larger than the secondary noises if one uses only the static-array logic (Tinto et al., 2018, Hartwig et al., 2021).

Mission-specific studies make the necessity concrete. For TianQin, the time differences of symmetric interference paths are reported as $\sim 10^{-8}\,\mathrm{s}$ for first-generation TDI and $\sim 10^{-12}\,\mathrm{s}$ for second-generation TDI, with the latter comfortably below the level required to suppress laser noise beneath secondary noises. For ASTROD-GW, the reported conclusion is that all second-generation TDI channels analyzed satisfy the mission path-matching requirement $\Delta t \le 150~\mathrm{ns}$ (Zhou et al., 2021, Wang, 2024).

2. Delay algebra, closure conditions, and construction principles

The basic operator-theoretic statement of the problem is that time-dependent delays do not commute. A delay operator acts as

$D_i f(t)=f(t-L_i(t)),$

and compositions satisfy

$$D_jD_i f(t)=f\!\left(t-L_j(t)-L_i(t-L_j(t))\right).$$

Equivalently, in split-interferometry notation,

$\D_{j'} \D_i x(t)=x\!\left(t-L_i(t-L_{j'})-L_j\right), \qquad [\D_{j'},\D_i]x(t)\neq 0.$

This noncommutativity is the mathematical reason first-generation TDI fails for a realistic moving constellation (Huang et al., 2023, Tinto et al., 2018).

Several construction paradigms coexist. In the algebraic approach, TDI combinations are polynomial combinations of intermediate observables $\eta_i,\eta_{i'}$ with coefficients in a noncommutative polynomial ring. In the geometric approach, a TDI observable is represented as a closed sequence of directed arm traversals on a spacetime diagram, and valid second-generation observables are those that are both $L$-closed and $10~\mathrm{m/s}$0-closed. The geometric literature emphasizes two synthesis operations: self-splicing, in which a first-generation closed path is reversed in time and spliced into itself, and cross-splicing, in which two distinct first-generation paths are combined with their time-reversed counterparts (Wang et al., 2022, Wang, 2024).

The distinction between standard second-generation and modified second-generation TDI is expressed by closure conditions on the coefficients of the residual delay-time mismatch. The geometric formulation summarizes the hierarchy as: first-generation $10~\mathrm{m/s}$1; modified first-generation $10~\mathrm{m/s}$2; second-generation $10~\mathrm{m/s}$3 and $10~\mathrm{m/s}$4; modified second-generation $10~\mathrm{m/s}$5 and $10~\mathrm{m/s}$6. In that language, standard second-generation TDI cancels the sum of the first-derivative terms, while modified second-generation cancels them direction by direction (Wang et al., 2022).

A central algebraic issue is whether the exact noncommutative operator equation has nontrivial solutions in the time-varying case. One line of work proves that the exact cancellation equation has only the zero solution in the non-static scenario, and therefore that second-generation TDI cannot be obtained by directly solving the operator equation as in the static commutative case. Another line of work constructs second-generation observables operationally by “lifting” first-generation generators into longer delayed combinations whose residual laser-noise terms are commutators of permutation-related delay products, vanishing to first order in the arm-rate variations. These positions are not identical: the former concerns exact solutions of the noncommutative module equation, while the latter concerns the operational construction of combinations that cancel the terms linear in the inter-spacecraft velocities (Huang et al., 2023, Tinto et al., 2022).

3. Observable families and generative structure

The first-generation TDI space is classically generated by the four variables $10~\mathrm{m/s}$7, or equivalently $10~\mathrm{m/s}$8. Second-generation work repeatedly exploits this structure. The “lifting” program starts from first-generation generators and constructs higher-order observables such as the second-generation unequal-arm Michelson $10~\mathrm{m/s}$9 and the second-generation Sagnac $\sim 10^{-8}\,\mathrm{s}$0, then uses those lifted generators to derive second-generation versions of $\sim 10^{-8}\,\mathrm{s}$1, Monitor, Beacon, and Relay combinations (Tinto et al., 2022).

The canonical second-generation observables discussed across the literature include the unequal-arm Michelson triad $\sim 10^{-8}\,\mathrm{s}$2, the Sagnac triad $\sim 10^{-8}\,\mathrm{s}$3, and the alternative families Monitor, Beacon, Relay, and hybrid Relay. The 2021 characterization paper organizes a broader catalog into 34 core combinations $\sim 10^{-8}\,\mathrm{s}$4, generating 210 distinct combinations up to 16 links once symmetries are included. The 2022 geometric classification reports 40 sixteen-link second-generation solutions recovered by exhaustive search, of which 9 are identified as modified second-generation ones (Tinto et al., 2018, Hartwig et al., 2021, Wang et al., 2022).

Several representative families are especially important in later data-analysis studies.

Family	Representative observables	Notes
Michelson	$\sim 10^{-8}\,\mathrm{s}$5	Fiducial second-generation scheme
Sagnac	$\sim 10^{-8}\,\mathrm{s}$6	Canonical second-generation triad
Hybrid Relay	$\sim 10^{-8}\,\mathrm{s}$7	Alternative with fewer null frequencies
Null-stream family	$\sim 10^{-8}\,\mathrm{s}$8	Includes a second-generation null stream used in noise characterization

The hybrid Relay was introduced as a second-generation alternative in which the ordinary channels are

$\sim 10^{-8}\,\mathrm{s}$9

and its science streams are described as having only one quarter of the null frequencies of the Michelson science channels. The same line of work relates many second-generation combinations back to the first-generation basis by approximate identities such as

$\sim 10^{-12}\,\mathrm{s}$0

which are accurate enough to model the noises not suppressed by TDI (Wang, 2024, Hartwig et al., 2021).

The repertoire is still expanding. A 2026 response-function study reformulates 45 essential second-generation combinations in terms of round-trip and non-round-trip geometric links, with simplified representations for Michelson-type, Relay-type, and related observables. In degraded configurations, the algebra also simplifies: with one arm dysfunctional, the two-arm problem admits an infinite number of approximate second-generation Michelson-like solutions generated by a commutator-based algorithm (Luo et al., 26 Jun 2026, Dhurandhar et al., 2010).

4. Orthogonal channels, null frequencies, and sensitivity structure

Most second-generation analyses pass from three ordinary channels $\sim 10^{-12}\,\mathrm{s}$1 to the quasi-orthogonal channels

$\sim 10^{-12}\,\mathrm{s}$2

Under near-equal arm lengths and similar noise budgets, these channels approximately diagonalize the noise covariance, with $\sim 10^{-12}\,\mathrm{s}$3 and $\sim 10^{-12}\,\mathrm{s}$4 serving as the two science channels and $\sim 10^{-12}\,\mathrm{s}$5 as the null-like channel (Zhang, 3 Nov 2025, Wang, 24 Jul 2025).

A major analytical result is that the sky- and polarization-averaged response functions, noise PSDs, and sensitivity curves of second-generation families coincide with those of their first-generation counterparts after the appropriate normalization. The same work states three further universal properties: the $\sim 10^{-12}\,\mathrm{s}$6 and $\sim 10^{-12}\,\mathrm{s}$7 channels have equal averaged responses, noise PSDs, and sensitivities; the $\sim 10^{-12}\,\mathrm{s}$8 channel has much weaker response and sensitivity at low frequencies; and the $\sim 10^{-12}\,\mathrm{s}$9 sensitivities and the optimal sensitivity are independent of the TDI generation and of which specific TDI family is used (Zhang, 3 Nov 2025).

Null frequencies are a separate, practically important layer of structure. The comparative studies summarize the main patterns as follows:

• Michelson: $\Delta t \le 150~\mathrm{ns}$0
• Sagnac: $\Delta t \le 150~\mathrm{ns}$1
• Modified Michelson-type: $\Delta t \le 150~\mathrm{ns}$2
• Minimal-null schemes such as U-, E-, P-type and PD4L-like configurations: $\Delta t \le 150~\mathrm{ns}$3

These nulls matter because they produce strong amplitude modulation in the gravitational-wave response, low-PSD regions sensitive to arm-length changes, unstable spectra, and more complicated frequency-domain waveform fitting. In the noise-characterization literature, the $\Delta t \le 150~\mathrm{ns}$4 channel is repeatedly identified as the weak point of a three-stream dataset, and one paper explicitly describes it as the “shortest plank in the barrel” (Wang, 2024, Wang, 24 Jul 2025).

The fully symmetric Sagnac variable $\Delta t \le 150~\mathrm{ns}$5 occupies a special position. In the first-generation algebra it is related to $\Delta t \le 150~\mathrm{ns}$6 by

$\Delta t \le 150~\mathrm{ns}$7

so it is not algebraically independent. In the analytical sensitivity treatment, the response, noise PSD, and sensitivity of $\Delta t \le 150~\mathrm{ns}$8 scale with those of the $\Delta t \le 150~\mathrm{ns}$9 channel (Hartwig et al., 2021, Zhang, 3 Nov 2025).

5. Noise characterization and inference performance

Although second-generation TDI is motivated by laser-noise cancellation, later work increasingly evaluates configurations by spectral stability, null-frequency structure, and inference robustness. A central comparison is between the fiducial second-generation Michelson and the hybrid Relay. The two schemes are reported to have comparable performance in mitigating laser and clock noises, but the Michelson configuration becomes inferior for chirping signals from coalescing massive binary black holes because of its vulnerable $D_i f(t)=f(t-L_i(t)),$0 channel and numerous null frequencies. In contrast, the hybrid Relay is described as more robust in dynamic unequal-arm scenarios, with science channels that are less affected by null frequencies and lower overall correlations (Wang, 2024, Wang, 2024).

The role of null streams in noise inference is now explicit. One study finds that hybrid Relay achieves more robust inference of acceleration-noise and optical-metrology-system noise amplitudes than the Michelson configuration, and that the dataset improves further when the hybrid-Relay $D_i f(t)=f(t-L_i(t)),$1 channel is replaced by the second-generation null stream $D_i f(t)=f(t-L_i(t)),$2, introduced by Hartwig et al. The resulting three-stream set

$D_i f(t)=f(t-L_i(t)),$3

reduces spectral instability in the target band, yields much tighter OMS-noise constraints, and improves the corresponding uncertainties by about a factor of $D_i f(t)=f(t-L_i(t)),$4 relative to the three-channel hybrid-Relay dataset with $D_i f(t)=f(t-L_i(t)),$5 (Wang, 2024).

A related design trend is the search for minimal-null and shorter-span schemes. PD4L is introduced as a second-generation configuration with 16 links, time span $D_i f(t)=f(t-L_i(t)),$6, and maximum delay $D_i f(t)=f(t-L_i(t)),$7, in contrast to the $D_i f(t)=f(t-L_i(t)),$8 span and $D_i f(t)=f(t-L_i(t)),$9 maximum delay of the second-generation Michelson and the hybrid Relay. Its science channels and null stream share the same minimal null frequencies $$D_jD_i f(t)=f\!\left(t-L_j(t)-L_i(t-L_j(t))\right).$$0. In parameter inference for a rapidly chirping massive-black-hole binary, PD4L is reported to outperform the hybrid Relay in the high-frequency band, while remaining comparable at low frequencies; its null stream is also described as more stable than $$D_jD_i f(t)=f\!\left(t-L_j(t)-L_i(t-L_j(t))\right).$$1, and the scheme remains reliable for noise characterization over data spans of up to about 120 days (Wang, 6 Feb 2025).

At the same time, large-scale comparative work argues that many second-generation schemes are highly redundant once their quasi-orthogonal channels are formed. The low-frequency regime $$D_jD_i f(t)=f\!\left(t-L_j(t)-L_i(t-L_j(t))\right).$$2 is reported to be nearly identical across configurations, with strong cross-correlations and substantial redundancy. Distinctions emerge primarily at high frequencies, where shorter TDI time spans and minimal null frequencies reduce aliasing and waveform-modulation effects in frequency-domain inference. The same study reports that, if the signal modeling and analysis are performed in the time domain, all TDI configurations become effectively equivalent (Wang, 24 Jul 2025).

Earlier numerical sensitivity studies already foreshadowed this division between cancellation and utility. In a survey of 11 second-generation channels under a numerical LISA orbit, the optimal $$D_jD_i f(t)=f\!\left(t-L_j(t)-L_i(t-L_j(t))\right).$$3 (or equivalently $$D_jD_i f(t)=f\!\left(t-L_j(t)-L_i(t-L_j(t))\right).$$4) channel built from second-generation Michelson observables achieves the best sensitivity among the channels studied, whereas the Sagnac $$D_jD_i f(t)=f\!\left(t-L_j(t)-L_i(t-L_j(t))\right).$$5 channel performs worst. The joint $$D_jD_i f(t)=f\!\left(t-L_j(t)-L_i(t-L_j(t))\right).$$6 observation improves the sensitivity of the $$D_jD_i f(t)=f\!\left(t-L_j(t)-L_i(t-L_j(t))\right).$$7 channel by a factor of $$D_jD_i f(t)=f\!\left(t-L_j(t)-L_i(t-L_j(t))\right).$$8 to 2 and also improves sky coverage (Wang et al., 2020).

6. Calibration, implementation realism, and ongoing debates

Second-generation TDI is not only a problem of laser-noise cancellation. Realistic implementation requires clock-noise calibration, onboard-delay compensation, and numerically faithful handling of time-dependent and fractional delays. A key step was the generalization of the Armstrong–Estabrook–Tinto sideband-based clock-noise calibration method from a static array to a rotating and flexing LISA geometry. In that formulation, the clock-calibration observables $$D_jD_i f(t)=f\!\left(t-L_j(t)-L_i(t-L_j(t))\right).$$9 are constructed from carrier and sideband measurements, and carefully delayed linear combinations of the $\D_{j'} \D_i x(t)=x\!\left(t-L_i(t-L_{j'})-L_j\right), \qquad [\D_{j'},\D_i]x(t)\neq 0.$0-channels are subtracted from the second-generation unequal-arm Michelson $\D_{j'} \D_i x(t)=x\!\left(t-L_i(t-L_{j'})-L_j\right), \qquad [\D_{j'},\D_i]x(t)\neq 0.$1 and Sagnac $\D_{j'} \D_i x(t)=x\!\left(t-L_i(t-L_{j'})-L_j\right), \qquad [\D_{j'},\D_i]x(t)\neq 0.$2 observables to remove ultra-stable-oscillator phase fluctuations while preserving the TDI-2 delay ordering (Tinto et al., 2018).

A second implementation issue is onboard optical path lengths. Recent work shows that the non-common optical paths before the combining beam splitters in the ISI, TMI, and RFI couple into the second-generation Michelson combinations and produce residual laser noise unless they are handled in the TDI algebra itself. The proposed compensation scheme replaces the ordinary inter-spacecraft delay operators by effective OOPL-aware operators and is reported to suppress the OOPL-induced residual by about 4 orders of magnitude for matched onboard delays, and by more than 2 orders of magnitude for the current design values (Reinhardt et al., 2024).

The formalism itself remains debated. The matrix-based “TDI-infinity” program argues that one should work directly with the design matrix that maps laser-noise samples to interferometric measurements, then extract laser-noise-free combinations as basis vectors of the null space of the adjoint design matrix. Its proponents emphasize that the method accommodates arbitrary time dependence and fractional delays, and reject the claim that it is merely a finite representation of the polynomial-ring approach except in the unrealistic case of exact sample-aligned delays (Bayle et al., 2021).

A different line of work pursues a fully data-driven alternative. Automated Principal Component Interferometry does not assume an explicit laser-noise model; instead it constructs low-variance combinations from delayed measurements by SVD or PCA. In the time-varying-delay setting, first-order aPCI is reported to suppress laser noise below the secondary-noise level and to match the combined gravitational-wave sensitivity of second-generation TDI to within about $\D_{j'} \D_i x(t)=x\!\left(t-L_i(t-L_{j'})-L_j\right), \qquad [\D_{j'},\D_i]x(t)\neq 0.$3 on average (Baghi et al., 2022).

The constructive mathematics is likewise unsettled but productive. Combinatorial algebraic approaches based on commutators and time-advance operators recover the known sixteen-link modified second-generation Monitor, Beacon, Relay, Sagnac, and fully symmetric Sagnac observables, and also generate Sagnac-inspired solutions outside the straightforward scope of geometric TDI. A later second-order combinatorial program extends this further by using nested commutators and products of commutators, producing new Michelson, Monitor, Beacon, Relay, Sagnac, fully symmetric Sagnac, and Sagnac-inspired solutions, with the last class exhibiting a distinct sensitivity curve (Wu et al., 2022, Qian et al., 2022).

Taken together, these developments define second-generation TDI less as a single closed algebraic object than as a mature family of operator, geometric, numerical, and data-analytic constructions tailored to moving unequal-arm interferometers. The central invariant across these approaches is the same: ordered delayed combinations of one-way measurements are used to cancel the dominant laser noise in a flexing constellation while retaining a usable gravitational-wave response (Tinto et al., 2022).