Space-Based Gravitational-Wave Astronomy

Space-based gravitational-wave astronomy begins with phase or frequency comparisons between freely falling spacecraft and ends with joint inference over detector noise, overlapping sources, and astrophysical populations. Every step matters: light-travel time makes the response frequency dependent; unequal, changing arms require ordered time-delay interferometry; unresolved binaries contribute structured foreground power; and source claims inherit waveform, mission, prior, and selection assumptions. This chapter follows that complete chain and marks which conclusions are established measurement theory and which remain configuration-dependent forecasts.

Helpful background. Probability and Statistical Inference supplies conditional probability, covariance, priors, and selection. The gravitational-wave likelihood supplies covariance-weighted inference. Gravitational Radiation and Waveforms supplies source generation and propagation, while Black-Hole Astrophysics supplies the massive-binary, EMRI, and stellar-binary populations used in forecasts.

From optical links to astrophysical inference

The ESA 2024 reference design uses a triangular three-spacecraft constellation with six directed laser links and nominal arm length $L=2.5\times10^9\ \mathrm{m}$ ESA 2024, §§2–4. Its finite light-travel time introduces the transfer frequency

$$f_*=\frac{c}{2\pi L}\simeq19.1\ \mathrm{mHz}.$$

Below, near, and above this scale, the detector cannot be represented by one frequency-independent antenna pattern with equal accuracy. A one-way fractional-frequency observable $y_{ij}(t)$ depends on the emitting and receiving events, link direction, spacecraft motion, gravitational-wave projection, laser noise, clock convention, and secondary noises.

The measurement chain is

$$\{y_{ij}\} \longrightarrow \{X,Y,Z\}_{\mathrm{TDI}} \longrightarrow \{A,E,T\}\ \text{or another declared basis} \longrightarrow \mathcal L(d|\vartheta,\lambda) \longrightarrow \text{source and population statements}.$$

The first arrow combines delayed links to suppress laser frequency noise. Because flexing-arm delays do not commute, the delay order and timestamp convention are part of the observable. The second arrow is a basis transformation, not an automatic guarantee of statistically independent channels. The covariance must be derived or estimated for the actual noise model. The final arrows require a global fit when sources and foregrounds overlap.

Route through the chapter

Read the pages in order; each supplies an input needed by the next.

1. Space Constellation and Link Response derives the endpoint response of one directed link, its finite-arm transfer function, and its dependence on sky direction and spacecraft motion.
2. Time-Delay Interferometry constructs delayed combinations that cancel laser noise, then shows why flexing arms and noncommuting delays require second-generation observables.
3. Space-Detector Noise and Confusion Foregrounds builds multi-channel covariance and a joint instrument–foreground likelihood, with an executable Whittle-likelihood control.
4. Massive Black-Hole Binary Signals connects inspiral, merger, ringdown, advance warning, and localization to declared waveform and mission assumptions.
5. Extreme-Mass-Ratio Inspiral Inference develops the multiharmonic frequency map and explains the conditional meaning of spacetime and environmental inference.
6. Multiband Gravitational-Wave Sources propagates a common-source posterior between space and ground bands without double counting priors or omitting joint selection.

The sequence deliberately places source forecasts after response, TDI, and covariance. A waveform’s intrinsic strain amplitude is not a TDI signal-to-noise ratio, and a nominal sensitivity curve is not a global-fit selection function.

The common likelihood

Let $\alpha$ label declared TDI channels and collect instrument, foreground, calibration, and response quantities into nuisance parameters $\lambda$. A common model is

$$d_\alpha(t) =\sum_s \mathcal R_\alpha(\lambda)h_s(\vartheta_s;t) +n_\alpha(t;\lambda).$$

For approximately independent positive-frequency bins, let $\widetilde{\boldsymbol r}_k$ collect the channel residuals and $\mathbf C_k(\lambda)$ their complex-bin covariance. The multichannel Whittle likelihood is

$$\log\mathcal L =-\sum_k\left[ \ln\det\mathbf C_k +\widetilde{\boldsymbol r}_k^{\,\dagger} \mathbf C_k^{-1} \widetilde{\boldsymbol r}_k \right] +\text{constant}.$$

The covariance convention absorbs the Fourier-bin and one-sided-PSD normalization; consistency is checked by simulated-noise moments and an analytic control. The determinant and quadratic terms are both summed over the same bins, and the determinant matters when $\lambda$ changes the covariance. Treating an unresolved foreground as fixed Gaussian noise can be a useful approximation, but it must be tested against nonstationarity, anisotropy, subtraction residuals, and resolved-source overlap.

Littenberg and Cornish demonstrate a prototype global analysis with multiple simulated source classes Littenberg and Cornish 2023, §§II–VI. The result establishes feasibility for its dataset and model space, not that a particular decomposition will be unique in flight data.

What is stable and what is forecast

Layer	Stable core	Configuration-dependent boundary
Link response	A plane wave is sampled at separated emission and reception events, producing a finite-arm transfer function	Orbit, ephemeris, clocks, calibration, gaps, and approximation order set the realized response
TDI	Delayed link combinations can cancel common laser noise in the assumed delay algebra	Flexing, noncommutation, ranging error, interpolation, clock noise, and data gaps set residual leakage
Noise inference	Cross-channel covariance and likelihood normalization define spectral inference	Instrument evolution, foreground population, overlap, subtraction, segmentation, and priors set recovered components
Massive binaries	Chirp phase, modulation, harmonics, merger, and ringdown carry complementary information	Warning, localization, host, and remnant precision depend on source, waveform, mission, and selection
EMRIs	Three fundamental frequencies generate a coherent, slowly evolving harmonic comb	Spacetime deviations, environment, yield, and precision depend on waveform completeness, response, overlap, and priors
Multiband sources	Conditionally independent likelihoods combine for one common source	Association, event yield, alert utility, and new-physics reach depend on both detector networks and joint selection

All source-yield, localization, parameter-precision, strong-field, and multiband claims in this chapter have an evidence cutoff of 2026-08-18 and are labelled as forecasts or simulation results. No LISA flight observation is presented. Mission schedules and performance should be refreshed from a superseding official configuration rather than inferred from an older forecast.

What the results establish and how to reproduce them

The chapter’s common public stress test is LISA Data Challenge 2b “Spritz”: simulated second-generation $X,Y,Z$ fractional-frequency TDI streams at 5 s cadence, based on a Keplerian constellation and including laser noise, secondary noise, gaps, artifacts, and an unresolved Galactic foreground LISA Data Challenge 2b, dataset description. It is neither flight data nor a universal noise model.

An end-to-end reproduction should pin the ESA configuration or stated alternative, orbit or ephemeris file, link geometry and timestamps, phase or fractional-frequency convention, laser and clock model, ordered delays, TDI generation, interpolation, cadence and band, observing interval and duty, challenge-file hashes, calibration, source population, waveform and response commits, foreground, covariance, priors, sampler, selection statistic, thresholds, injection measure, and seeds. Report failed searches, missed modes, posterior coverage, residual spectra, and selection efficiency.

Two controls bracket the analysis:

1. In a static unequal-arm simulation, the declared first-generation TDI polynomial must cancel a common laser process to numerical precision while retaining an injected gravitational wave.
2. In a flexing-arm simulation, the residual must show the predicted noncommuting-delay scaling; a second-generation construction must reduce it. Only then should cadence refinement, orbit error, gaps, foreground, and overlapping sources be added one at a time.

Tinto and Dhurandhar develop the TDI algebra and its generations in Tinto and Dhurandhar 2021, §§3–6. For moving-arm frequency data, the chain rule also introduces Doppler factors when differentiating delayed phases; the TDI page explains why ordinary phase-delay operators alone are insufficient. A zero laser residual in one discretization is not enough: delay order, interpolation convergence, signal transfer, and an independently generated injection all need checks.

Review the chapter

1. Symmetric channel covariance. Suppose $X,Y,Z$ have equal one-sided auto-spectral density $S$ and equal real cross-spectral density $C$. Define

$$A=\frac{2X-Y-Z}{3}, \qquad E=\frac{Z-Y}{\sqrt3}, \qquad T=\frac{X+Y+Z}{3}.$$

Find the three auto-spectral densities and explain the limitation of this diagonalization.

Solution

Linearity of covariance gives

$$S_A=S_E=\frac23(S-C), \qquad S_T=\frac13(S+2C).$$

All three cross-spectra vanish because the covariance of $X,Y,Z$ is invariant under permutations. Thus $A,E,T$ diagonalize this ideal symmetric covariance. Unequal arms, unequal link noises, time evolution, gaps, calibration error, or anisotropic residuals break the symmetry; then the actual frequency-dependent covariance must be estimated and inverted rather than assumed diagonal.

2. A spectral-amplitude control. In independent positive-frequency bins, let the model PSD be $S_k=a q_k$ with known $q_k\gt0$. The Whittle negative log likelihood, up to constants, is

$$\ell(a) =\sum_{k=1}^{N} \left[ \log(aq_k) +\frac{I_k}{aq_k} \right].$$

Derive the maximum-likelihood estimator. If the model is true and $I_k/(aq_k)$ are independent unit-mean exponential variables, find its mean and standard deviation.

Solution

Differentiating gives

$$\frac{d\ell}{da} =\frac{N}{a} -\frac{1}{a^2}\sum_{k=1}^{N}\frac{I_k}{q_k}.$$

Therefore

$$\widehat a =\frac1N\sum_{k=1}^{N}\frac{I_k}{q_k}.$$

Each term $I_k/q_k$ has mean $a$ and variance $a^2$, so

$$\mathbb E[\widehat a]=a, \qquad \operatorname{sd}(\widehat a)=\frac{a}{\sqrt N}.$$

This is a normalization and convergence control for ideal independent bins. Windowing, overlap, nonstationarity, cross-channel covariance, and fitted foreground shapes reduce the effective information and must be handled in the real likelihood.

References

• European Space Agency. LISA Definition Study Report. ESA-SCI-DIR-RP-002 (2024). Official PDF.
• LISA Data Challenge. “Challenge 2b: Spritz.” Public simulated dataset and documentation. Dataset.
• Littenberg, Tyson B., and Neil J. Cornish. “Prototype Global Analysis of LISA Data with Multiple Source Types.” Physical Review D 107 (2023): 063004. DOI. Open PDF.
• Tinto, Massimo, and Sanjeev V. Dhurandhar. “Time-Delay Interferometry.” Living Reviews in Relativity 24 (2021): 1. DOI.