Transit Centroid Vetting: Ruling Out Background Eclipsing Binary Blends

What Centroid Vetting Is and Why Blends Fool Us

A transit survey measures brightness in a photometric aperture spanning many arcseconds — 3.98″ per pixel for Kepler, a coarse 21″ for TESS. Within that aperture, the light of the intended target can be contaminated by unresolved neighbours. When a fainter background or bound eclipsing binary (EB) undergoes a deep eclipse, its light loss is diluted by the bright target, so the summed aperture flux dips by a shallow, planet-like amount.

The trap: a 1% dip that looks like a hot Jupiter can equally be a 50%-deep EB eclipse diluted 50:1. Radius alone cannot tell them apart. But spatial information can. Because the eclipsing source sits at a slightly different sky position than the target, removing its light during the eclipse pulls the aperture's flux-weighted centroid toward the target and away from the blend. Centroid vetting exploits exactly this: it asks whether the location of the dimming matches the location of the star we think we are watching.

The Mechanism: Difference Imaging and the Centroid Shift

The workhorse implementation is difference imaging, formalised in the Kepler Data Validation module. For every pixel, the pipeline averages the flux in-transit and subtracts it from the average out-of-transit flux. The resulting difference image isolates only the light that disappeared during the dip. Fitting the telescope's Pixel Response Function (PRF) to that difference image locates the transit source; fitting the PRF to the out-of-transit image locates the target. The vector between them is the centroid offset.

For a constant target diluted by an offset eclipsing object, the flux-weighted centroid follows a linear relation:

• ξ(t) ≈ (F_e/F_sys − 1) · (1 − D₀) · x_e

where D₀ is the dilution (fraction of light from the constant target), x_e is the sky offset of the eclipsing object, and F_e/F_sys is the eclipsing object's normalised flux fraction. The shift peaks at eclipse minimum, scales with both the depth and the separation, and reverses sign on either side of the offset — a distinctive, rediscoverable signature.

Key Numbers and a Worked Example

Consider a target contaminated by a background EB offset by x_e = 2″, contributing 10% of the aperture light, whose eclipse is 30% deep. The diluted system dip is δ_sys ≈ 0.10 × 0.30 = 3%. The centroid shift is roughly ξ ≈ (1 − D₀) · x_e · δ_sys ≈ 0.90 × 2″ × 0.03 ≈ 0.05″ — small, but well within reach of a PRF fit whose formal precision reaches 0.1″ or better for a bright Kepler star.

• Detection threshold: a centroid offset is flagged when it exceeds 3σ; offsets below ~0.1″ are treated as systematics and ignored regardless of sigma.
• NGTS on the ground achieves centroid precisions of 0.25–0.75 milli-pixels (a few milli-arcseconds), recovering roughly 80% of blended EBs.
• Genuine planets should show zero offset — only about 6% of real planet transits produce any detectable centroid motion, and even then only from photometric noise or crowding, not from the planet itself.

How It Is Observed and Where It Is Applied

Centroid vetting is baked into the automated pipelines of every major transit survey. In Kepler's DV report, two metrics anchor the test: the OotOffset (out-of-transit source vs transit source) and the KicOffset (transit source vs the catalogued star position). A red asterisk falling outside the 3σ blue circle on the DV summary page is the classic visual cue that the transit belongs to a different star.

• Kepler / K2: per-quarter difference images; K2 additionally needs motion-decorrelation because thruster-fire pointing drift smears the centroid.
• TESS: the enormous 21″ pixels make in-aperture blends common, so centroiding is paired with Gaia astrometry and per-pixel light-curve mapping to localise the source.
• NGTS: the first fully automated ground-based centroid-vetting algorithm, correcting for atmospheric and instrumental centroid systematics before differencing.

Crucially, a null result is not a confirmation — it only rules out resolvable offset blends. High-resolution imaging and radial velocities finish the job.

How It Differs From Its Cousins

Centroid vetting is one tool in a false-positive toolkit, and it targets a specific culprit — the spatially offset blend.

• vs. odd/even depth test: catches EBs whose primary and secondary eclipses differ in depth. Works on on-target EBs where the centroid barely moves; centroiding is blind there.
• vs. secondary-eclipse / phase-curve search: flags a hidden stellar companion by its thermal secondary. Complementary, sensitive to unresolved bound pairs.
• vs. V-shaped vs U-shaped morphology: grazing EBs mimic depth but not shape; a purely geometric discriminant.
• vs. high-contrast imaging (AO / speckle): resolves neighbours down to ~0.05″ directly, whereas centroiding infers sub-pixel offsets statistically without resolving them.
• vs. statistical validation (BLENDER, vespa): computes a false-positive probability by weighing all scenarios; centroid offset is a key input to that calculation, not a competitor.

Centroiding's unique strength is reaching below the diffraction limit — it exposes blends far tighter than any camera can split.

Significance, Famous Cases, and Open Problems

Centroid vetting rescued the statistical credibility of the Kepler catalogue: the mission's astrophysical false-positive rate is a manageable ~10% largely because difference-image centroiding culled most background EBs before follow-up. It converted a candidate list riddled with blends into a demographically usable sample underpinning modern occurrence-rate and eta-Earth estimates.

Cautionary cases remain. TESS objects of interest such as TOI-864.01 have been unmasked as hierarchical eclipsing binaries where dilution masqueraded as a planet, and analyses of systems like HIP-44302 illustrate ongoing refinement in how offsets are interpreted. Open challenges persist: TESS's coarse pixels leave many blends unresolved even after centroiding; bound companions within an arcsecond can evade the offset test entirely; and crowded fields, saturated stars, and correlated red noise inflate false alarms. Distinguishing a genuinely off-target signal from a centroiding systematic at the 0.1″ floor is still where careful vetters earn their keep.

Angular resolution and centroid sensitivity across major transit surveys.

Survey	Pixel scale	Centroid precision	Blend sensitivity
Kepler	3.98″/pixel	~0.1–0.5″ (PRF fit)	Sub-pixel offsets flagged at >3σ
TESS	21″/pixel	~few arcsec	Coarse; needs difference imaging + Gaia
NGTS (ground)	4.97″/pixel	0.25–0.75 mpix (~1–4 mas)	Recovers ~80% of blended EBs
K2	3.98″/pixel	Degraded by pointing drift	Motion-corrected centroids required