LISA–Taiji Network Overview

• The LISA–Taiji network is a coordinated pair of heliocentric, triangular interferometers designed to detect millihertz gravitational waves with complementary orbital configurations.
• Its dual-constellation setup improves angular resolution by orders of magnitude for massive black hole binaries and increases detection rates for stellar binaries and EMRIs.
• Cross-correlation between the detectors enables measurement of stochastic gravitational-wave backgrounds, including chirality and non-tensor polarization features.

The LISA–Taiji network is the prospective joint operation of the space-borne gravitational-wave observatories LISA and Taiji in the millihertz band, typically understood as two triangular, heliocentric interferometers with arm lengths $2.5\times 10^6$ km and $3\times 10^6$ km, respectively, separated along Earth-like orbits and analyzed as a coherent detector network rather than as independent missions. In the literature, this network is important for two distinct reasons: it converts two LISA-class missions into a long-baseline interferometric array for resolved sources such as massive black hole binaries, stellar binary black holes, Galactic binaries, and extreme mass-ratio inspirals, and it enables cross-correlation measurements of stochastic gravitational-wave backgrounds that are inaccessible to a single planar detector, including isotropic chirality and non-tensor polarization content (2002.03603, Orlando et al., 2020, Chen et al., 2022).

1. Architecture and orbital realizations

The baseline LISA–Taiji concept places LISA in a heliocentric orbit trailing Earth by about $20^\circ$ and Taiji in a similar orbit leading Earth by about $20^\circ$, yielding an orbital separation of about $40^\circ$ and a physical baseline of order $0.7\,\mathrm{AU}$ or $\sim 1\times 10^8$ km. Both constellations are equilateral triangles with detector planes inclined by $60^\circ$ to the ecliptic and undergoing annual cartwheel motion. In this configuration, the network captures most of the localization benefit of much larger separations: for a $10^5 M_\odot$ equal-mass massive black hole binary, the angular resolution improves by about two orders of magnitude as the configuration angle increases from $3^\circ$ to $3\times 10^6$0, while the gain from $3\times 10^6$1 to $3\times 10^6$2 is only about $3\times 10^6$3 order of magnitude (Ruan et al., 2019, 2002.03603).

A second line of work formalized three alternative Taiji deployments relative to a fixed LISA orbit. These configurations differ in whether the Taiji triangle is leading or co-located, and whether its inclination is $3\times 10^6$4 or $3\times 10^6$5, producing markedly different plane-plane angles and overlap properties (Wang et al., 2021).

Network	Taiji placement	Distinctive geometry
LISA–TAIJIp	Leading Earth by $3\times 10^6$6, inclination $3\times 10^6$7	Plane angle with LISA $3\times 10^6$8
LISA–TAIJIm	Leading Earth by $3\times 10^6$9, inclination $20^\circ$0	Plane angle with LISA $20^\circ$1
LISA–TAIJIc	Co-located and coplanar with LISA	Nearly identical antenna patterns

These alternatives matter because different science cases favor different geometries. For massive binary localization and polarization determination, the LISA–TAIJIm network performs best because of its more complementary antenna pattern, whereas LISA–TAIJIc gives the largest low-frequency cross-correlation response for stochastic-background work (Wang et al., 2021). For EMRI studies, by contrast, the distinction between TAIJIp and TAIJIm is found to be negligible: the existence of a second heliocentric constellation matters more than whether the relative plane angle is $20^\circ$2 or $20^\circ$3 (Zhang et al., 8 Jun 2026).

2. Response formalism and network observables

The network is analyzed in time-delay interferometry variables rather than raw inter-spacecraft measurements. Across the literature, this has included first-generation Michelson-derived A, E, T channels, TDI 1.5 AET variables, and PD4L second-generation TDI, depending on the source class and fidelity of the detector model (Wang et al., 2021, Orlando et al., 2020, Zhang et al., 8 Jun 2026). For compact binaries, the network signal-to-noise ratio is the quadrature sum of detector contributions,

$20^\circ$4

so a dual-constellation network combines independent SNR and independent angular response (Chen et al., 2022).

For parameter estimation, the standard construction is a network Fisher matrix,

$20^\circ$5

with covariance $20^\circ$6 and sky-localization error derived from the angular sub-block (Chen et al., 2022). This framework underlies forecasts for massive black hole binaries, stellar binary black holes, and EMRIs.

For stochastic backgrounds, the crucial formal distinction is between self-correlations within one planar constellation and cross-correlations between LISA and Taiji. The SGWB is written in terms of Stokes parameters $20^\circ$7 and $20^\circ$8,

$20^\circ$9

with chirality parameter $20^\circ$0. The cross spectrum between channels $20^\circ$1 is

$20^\circ$2

For a single planar detector, $20^\circ$3 in the isotropic limit, so self-correlations are blind to net circular polarization; for LISA–Taiji cross-correlations, $20^\circ$4, which turns chirality into an observable (Orlando et al., 2020). This is one of the defining methodological differences between a single LISA-like mission and a LISA–Taiji network.

3. Compact-binary astronomy

For massive black hole binaries, the network was first quantified as a precision-localization instrument. For an equal-mass binary at redshift $20^\circ$5 with total intrinsic mass $20^\circ$6, a one-year LISA–Taiji overlap improves the event localization region by about four orders of magnitude relative to an individual detector; in a representative time-to-merger calculation, Taiji alone reaches $20^\circ$7 and $20^\circ$8, whereas the network reaches $20^\circ$9 and $40^\circ$0 (Ruan et al., 2019). Under a uniform comoving galaxy density $40^\circ$1, this localization volume is small enough to permit unique host-galaxy identification for substantial low-redshift subsets of $40^\circ$2, $40^\circ$3, and $40^\circ$4 binaries (Ruan et al., 2019).

For stellar binary black holes, the network changes the expected sample size from marginal to statistically useful. Using LIGO/Virgo O3a merger-rate and mass-distribution constraints, a 4-year mission yields median detection counts $40^\circ$5 for the network, compared with $40^\circ$6 for LISA alone and $40^\circ$7 for Taiji alone; at the more conservative threshold $40^\circ$8, the corresponding medians are $40^\circ$9, $0.7\,\mathrm{AU}$0, and $0.7\,\mathrm{AU}$1. Extending the overlap to 10 years raises the network counts to $0.7\,\mathrm{AU}$2 for $0.7\,\mathrm{AU}$3 and $0.7\,\mathrm{AU}$4 for $0.7\,\mathrm{AU}$5, with a non-negligible subset merging during the mission and thus enabling multiband observations with ground-based detectors (Chen et al., 2022). For detected sBBHs, the network typically yields luminosity-distance errors in the range $0.7\,\mathrm{AU}$6–$0.7\,\mathrm{AU}$7 and sky localization in the range $0.7\,\mathrm{AU}$8–$0.7\,\mathrm{AU}$9, with 10-year merging systems reaching $\sim 1\times 10^8$0–$\sim 1\times 10^8$1 (Chen et al., 2022).

For extreme mass-ratio inspirals, recent Fisher studies using fully relativistic FEW waveforms and realistic time-domain TDI responses show that a one-month LISA–Taiji observation already approaches the information content of a one-year LISA-only observation. For a representative EMRI with $\sim 1\times 10^8$2, $\sim 1\times 10^8$3, $\sim 1\times 10^8$4, $\sim 1\times 10^8$5, and $\sim 1\times 10^8$6 Gpc, the median SNR is $\sim 1\times 10^8$7 for one-month LISA, $\sim 1\times 10^8$8 for the LTp network, $\sim 1\times 10^8$9 for LTm, and $60^\circ$0 for one-year LISA. Sky localization improves from $60^\circ$1 for one-month LISA to $60^\circ$2 for LTp and $60^\circ$3 for LTm (Zhang et al., 8 Jun 2026).

For Galactic binaries, the network has been tested in fully realistic mock data containing $60^\circ$4 sources from the Radler population. A coherent LISA–Taiji analysis with the iterative GBSIEVER pipeline increases the number of confirmed binaries by $60^\circ$5 over a single detector, from 10,388 to 18,151, and the residual after subtraction approaches the ideal confusion-noise floor much more closely than in a single-detector analysis (2206.12083). This has network-wide consequences because improved subtraction of the Galactic foreground directly benefits EMRI, massive-black-hole, and stochastic-background searches.

4. Stochastic backgrounds, chirality, and non-GR polarization

The LISA–Taiji network has a special role in stochastic gravitational-wave background science because cross-correlation between separated constellations accesses observables unavailable to a single planar interferometer. For isotropic SGWBs, the central quantities are the total intensity $60^\circ$6 and the circular-polarization Stokes parameter $60^\circ$7. A single LISA-like triangle is blind to $60^\circ$8 in the isotropic limit, but LISA–Taiji cross-correlations produce a nonzero parity-sensitive response. In a Fisher forecast for a flat spectrum with 4 years of data and 75% duty cycle, a maximally chiral background can be measured clearly when $60^\circ$9; at lower amplitude, $10^5 M_\odot$0 becomes much less well constrained, with $10^5 M_\odot$1 errors on $10^5 M_\odot$2 typically 1–2 orders of magnitude larger than those on $10^5 M_\odot$3 (Orlando et al., 2020).

This isotropic-chirality problem is also where the alternative network geometries differ most sharply. LISA–TAIJIc has $10^5 M_\odot$4 up to about $10^5 M_\odot$5 mHz and is therefore the best pure cross-correlation geometry for SGWB detection, whereas LISA–TAIJIm gives the best sky localization and polarization determination for massive binaries because its antenna patterns are more complementary. Subsequent SGWB work showed that, although TAIJIp has a somewhat larger low-frequency overlap than TAIJIm, the latter is competitive with or superior to TAIJIp for several specific isotropic spectral shapes, and the two have essentially identical capability to separate stochastic components through parameter estimation (Wang et al., 2021, Wang et al., 2021).

For parity violation in particular, later analyses comparing only the p and m configurations found that the LISA–TAIJIm network has sensitivity to circular polarization approximately one order of magnitude greater than LISA–TAIJIp at lower frequencies, and correspondingly smaller Fisher errors on the polarization fraction for power-law, single-peak, and broken-power-law spectra (Chen et al., 2024). This makes the TAIJIm geometry the preferred choice if parity-violating cosmology is prioritized.

The network has also been used to study anomalous vector and scalar polarizations in isotropic SGWBs. In the baseline LISA–Taiji geometry, the symmetry of the $10^5 M_\odot$6 and $10^5 M_\odot$7 cross-correlations allows a linear combination in which the tensor contribution cancels algebraically, leaving only vector and scalar components. Under a 10-year overlap, the resulting effective sensitivities reach $10^5 M_\odot$8 and $10^5 M_\odot$9 in the mHz band (Omiya et al., 2020). This constitutes a direct network-level test of polarization content beyond GR.

Chiral early-universe models provide a concrete target for this capability. For chiral backgrounds generated by axion–dark-photon and axion–Nieh–Yan couplings, Fisher forecasts with the LISA–Taiji network find relative $3^\circ$0 errors below $3^\circ$1 and $3^\circ$2 for the normalized model parameters of the two mechanisms, and relative errors around $3^\circ$3 and $3^\circ$4 for the circular-polarization parameters, respectively (Su et al., 26 Mar 2025). In this sense, the network is not merely a detector of stochastic power; it is a spectropolarimetric instrument for new-physics backgrounds.

5. Cosmology with standard and dark sirens

The network’s compact-binary localization capability feeds directly into GW cosmology. In the bright-siren case, a forecast based on three massive-black-hole formation models—pop III, Q3d, and Q3nod—found that over 5 years the number of standard sirens with electromagnetic counterparts is approximately doubled by the network relative to Taiji alone: $3^\circ$5 vs $3^\circ$6 for pop III, $3^\circ$7 vs $3^\circ$8 for Q3d, and $3^\circ$9 vs $3\times 10^6$00 for Q3nod. Using network standard sirens alone, the relative precision on $3\times 10^6$01 can reach $3\times 10^6$02 in favorable models; when combined with Planck distance priors, the network yields $3\times 10^6$03, or about $3\times 10^6$04, for constant-$3\times 10^6$05 dark energy in the Q3nod scenario (Wang et al., 2021).

In the dark-siren case, where no electromagnetic counterpart is identified and host galaxies are marginalized statistically, the network remains powerful because it reduces the host-galaxy count inside the 3D localization volume. One 5-year study found that the LISA–Taiji network can constrain the Hubble parameter within $3\times 10^6$06 accuracy and possibly down to $3\times 10^6$07 or better. In a representative event, the network reduced the sky area from $3\times 10^6$08 for Taiji alone to $3\times 10^6$09, and the number of candidate host galaxies from 1022 to 3 (Wang et al., 2020). The same work classified the best events as “diamond” $3\times 10^6$10, “gold” $3\times 10^6$11, “green” $3\times 10^6$12, and “blue” $3\times 10^6$13, and found that all diamond and gold network events lie at $3\times 10^6$14, where low-redshift $3\times 10^6$15 inference can be formulated in a nearly model-independent way (Wang et al., 2020).

These cosmological applications depend on astrophysical assumptions about massive-black-hole populations, counterpart rates, and galaxy catalogs, but the network contribution is geometrically robust: higher SNR, smaller $3\times 10^6$16, tighter $3\times 10^6$17, and hence fewer plausible hosts. In practice, this is why the LISA–Taiji network appears repeatedly in forecasts for standard sirens, dark sirens, and host-galaxy identification rather than only in source-detection studies.

6. Design trade-offs, limitations, and extensions

A recurring theme in the literature is that there is no single globally optimal geometry for every science case. The co-located TAIJIc option maximizes overlap-reduction functions and therefore stochastic-background cross-correlation, but it does not provide the long-baseline triangulation that drives the best massive-binary localization. Conversely, TAIJIm yields the best sky localization and polarization determination for massive binaries, and later work showed that it is also competitive with or superior to TAIJIp for several SGWB and parity-violation targets; this is why multiple authors identify TAIJIm as the most balanced overall option for a joint mission with LISA (Wang et al., 2021, Wang et al., 2021, Chen et al., 2024).

A common misconception is that a second LISA-class mission would only increase sensitivity by roughly $3\times 10^6$18. That is accurate for some amplitude-dominated parameters, but incomplete. The network also changes the inverse problem itself: it introduces new baselines, new antenna patterns, new parity-sensitive phases, and independent cross-correlations. This is why localization for $3\times 10^6$19 massive black hole binaries can improve by orders of magnitude rather than only by the factor expected from SNR scaling, and why isotropic SGWB chirality is measurable only with a network (Ruan et al., 2019, Orlando et al., 2020).

The forecast literature also shares several limitations. Fisher analyses for SGWBs assume isotropy, Gaussianity, and stationary noise, often neglect correlated environmental noise and imposing simplified spectral models; compact-binary forecasts may neglect spin, eccentricity, merger-ringdown, or confusion foregrounds depending on the source class; dark-siren studies typically assume simplified galaxy distributions or omit clustering information (Orlando et al., 2020, Chen et al., 2022, Wang et al., 2020). These are not peculiar to LISA–Taiji, but they affect how forecast numbers should be interpreted. In particular, the TAIJIc configuration is repeatedly noted as potentially vulnerable to correlated noise because co-location improves overlap and environmental commonality simultaneously (Wang et al., 2021).

The network concept is also extensible. Three-detector configurations including TianQin have already been studied for EMRIs, with the result that a one-month LISA–TAIJI–TianQin observation can achieve parameter constraints comparable to, and in some cases tighter than, a one-year LISA-only observation (Zhang et al., 8 Jun 2026). This suggests that the LISA–Taiji network is best understood not only as a specific two-mission proposal, but as the first realizable instance of a broader mHz detector-network architecture in which constellation geometry is a science driver in its own right.