Reference-Star Differential Imaging

• RDI is a high-contrast imaging method that uses external reference star images to model and subtract the stellar PSF, enhancing detection of faint companions and disks.
• It employs PCA/KLIP algorithms with large, carefully selected PSF libraries to mitigate self-subtraction effects common in angular differential imaging.
• Recent advances integrate synthetic references and hybrid calibration techniques to improve performance, especially under challenging conditions and limited field rotation.

Searching arXiv for recent and foundational papers on reference-star differential imaging to ground the article in the current literature. Reference-star Differential Imaging (RDI) is a high-contrast post-processing strategy in which the stellar point-spread function (PSF) of a science target is modeled and subtracted using images of other stars rather than the science target itself. In direct imaging, its principal motivation is the removal of quasi-static speckles and coronagraphic residuals without embedding the astrophysical companion or circumstellar structure in the subtraction basis. Relative to angular differential imaging (ADI), RDI avoids the fundamental self-subtraction problem because the subtraction basis does not contain the target signal, and is therefore especially relevant at small angular separations, for limited parallactic-angle rotation, and for extended or nearly symmetric circumstellar emission such as disks (Ruane et al., 2019).

1. Conceptual basis and relation to other differential-imaging strategies

RDI models the stellar halo and speckle field from separate reference-star observations. In the Keck/NIRC2 vortex-coronagraph formulation, a single frame is written as

$X = I + A,$

where $X$ is the observed frame, $I$ is the stellar PSF/speckle field after the coronagraph, and $A$ is the planet PSF. The stellar PSF is reconstructed as

$$I = \sum_{k=1}^{K} \left< X, Z^{(k)} \right> Z^{(k)},$$

where $Z^{(k)}$ are PCA basis images derived from the reference set $\{R^{(k)}\}$, and $K$ is the number of PCA modes used (Ruane et al., 2019).

The distinction from ADI is operational and statistical. ADI uses the target star’s own time series as the PSF library while the field rotates; RDI uses external references. ADI therefore depends on sky rotation to separate astrophysical signal from the PSF structure, whereas RDI depends on the quality of the reference library. For direct-imaging point sources, this makes RDI especially useful for companions at very small angular separations and for observations with limited parallactic-angle rotation; for disks, it is advantageous because it avoids the self-subtraction that plagues ADI and preserves extended morphology more faithfully (Ruane et al., 2019).

RDI is often discussed alongside simultaneous spectral differential imaging (SSDI), binary differential imaging (BDI), and hybrid approaches. In the GPIES context, RDI appears as “Reference Star Differential Imaging” (RSDI), intended to supplement ADI and SSDI when the usual $\sim 1.5$ to $3\,\lambda/D$ motion threshold at small separations leaves too few target-sequence reference frames (Gerard et al., 2016). In MagAO/Clio observations of wide binaries, BDI is explicitly described as an application of RDI and ADI to binary systems, where each star serves as the PSF reference for the other (Pearce et al., 2022). More recently, joint angular-reference differential imaging (ARDI) has treated ADI and RDI as complementary rather than competing strategies by constructing a low-rank model from both science and reference frames (Juillard et al., 2024).

A common misconception is that RDI automatically outperforms ADI. The literature does not support that as a universal rule. Several studies instead identify a regime dependence: RDI is typically favored at small separations or for morphologically extended sources, whereas ADI can catch up or outperform RDI at larger separations or when the science sequence has unusually favorable field rotation (Xuan et al., 2018).

2. Core algorithms, basis construction, and subtraction formalisms

The dominant algorithmic backbone of modern RDI is PCA/KLIP. Keck/NIRC2 reductions use PCA through VIP; JWST/NIRCam reductions for HD 19467 B use Karhunen–Loève Image Projection (KLIP) via pyKLIP; SPHERE/IRDIS archival RDI likewise uses PCA-based reconstructions from selected reference subsets (Ruane et al., 2019). In the simplest low-rank notation adopted in ARDI/IPCA work, a data matrix is decomposed as

$X$0

with $X$1 the spatial principal components, and the estimated speckle cube is

$X$2

after retaining the first $X$3 components (Juillard et al., 2024).

Several papers make explicit that performance is often controlled less by the nominal PCA truncation than by reference selection. In the Keck vortex-coronagraph study, RDI is described as less sensitive to the exact choice of $X$4 than ADI, with the dominant performance driver being which reference frames are selected (Ruane et al., 2019). The Super-RDI framework for a $X$5-frame, multi-year Keck/NIRC2 library likewise treats metric-based frame ranking and injection-recovery optimization as co-equal to the KLIP/PCA subtraction step (Sanghi et al., 2024).

A separate line of development concerns signal-preserving modifications of KLIP for extended structures. “Constrained reference star differential imaging” rewrites the target image as

$X$6

with $X$7 the stellar signal and $X$8 the circumstellar signal, and notes that conventional RDI subtraction,

$X$9

leads to attenuated circumstellar signal

$I$0

Its remedy is to estimate the PSF model from $I$1, where $I$2 is an independent estimate of the circumstellar emission: $I$3 This formulation explicitly targets oversubtraction in disk imaging (Lawson et al., 2022).

Karhunen–Loève Data Imputation (DIKL) modifies standard KLIP in a different way by partitioning each image into an anchor matrix $I$4, intended to contain only speckles, and a boat matrix $I$5, which may contain astrophysical signal. Using boolean selection matrices,

$I$6

DIKL computes the eigendecomposition on the anchor covariance

$I$7

forms the KL basis $I$8, transfers that basis to the boat region as

$I$9

and then subtracts the boat-region speckles using coefficients derived only from the anchor target: $A$0 The stated purpose is to reduce overfitting of astrophysical structure while remaining a KLIP-derived analytical method (Ren, 2023).

3. Reference-library construction and frame-selection metrics

Reference-library quality is a recurrent limiting factor in RDI. The Keck/NIRC2 vortex-coronagraph work compared three frame-ranking metrics computed relative to the temporal median of the science frames over a speckle-dominated region within $A$1–$A$2: mean squared error (MSE), Pearson correlation coefficient (PCC), and structural similarity index metric (SSIM). The definitions provided are

$A$3

$A$4

and

$A$5

For point sources, SSIM or MSE performed better than PCC, with SSIM slightly best; pre-selecting the best-matching reference frames improved detection significance by up to a factor of 3 (Ruane et al., 2019).

The Super-RDI study expands this logic to five metrics: MSE, PCC, SSIM, flux logarithmic standard deviation indicator (FLSI), and contrast logarithmic standard deviation indicator (CLSI), with

$A$6

$A$7

and

$A$8

For a 50-target optimization subset, 38/50 targets preferred MSE, and injection-recovery tests found the best-performing region to be MSE with 1000–3000 reference frames and fewer than 500 PCs, with a local optimum near 1348 frames and 380 PCs (Sanghi et al., 2024).

Large archival libraries have enabled more systematic selection studies. On SPHERE/IRDIS H23 data, a master reference library built from five years of archival observations retained about $A$9 images per band from 725 observations after calibration and quality control. Images were aligned by minimizing $X$16 and reference selection for subtraction used a robust-scaled MSE, $X$17 after scaling by $$I = \sum_{k=1}^{K} \left< X, Z^{(k)} \right> Z^{(k)},$$0 (Xie et al., 2022).

For pole-on disk reductions with SPHERE/IRDIS, reference-library selection has also been formulated as a metadata problem. One study built H23 and K12 master libraries of 67,386 frames from 769 sequences and 54,773 frames from 746 sequences, respectively, then for each science frame preselected 10,000 frames using one metric, ranked them by frame-to-frame PCC within $$I = \sum_{k=1}^{K} \left< X, Z^{(k)} \right> Z^{(k)},$$1 to $$I = \sum_{k=1}^{K} \left< X, Z^{(k)} \right> Z^{(k)},$$2, and retained the 1000 most correlated frames. The tested parameters were seeing, coherence time $$I = \sum_{k=1}^{K} \left< X, Z^{(k)} \right> Z^{(k)},$$3, 30 m wind speed, 200 mbar wind speed and direction, elevation, epoch (MJD), DIT, Gaia $$I = \sum_{k=1}^{K} \left< X, Z^{(k)} \right> Z^{(k)},$$4, 2MASS $$I = \sum_{k=1}^{K} \left< X, Z^{(k)} \right> Z^{(k)},$$5, and spectral type. PCC-only libraries achieved the best mean contrast overall, while mixed libraries gave the most consistent results and best overall disk S/N (Stasevic et al., 3 Sep 2025).

The HST/WFC3 archival PSF approach for PDS 70 b illustrates a different selection problem: not only similarity but also reference quality. From $$I = \sum_{k=1}^{K} \left< X, Z^{(k)} \right> Z^{(k)},$$6 million WFC3/UVIS PSFs in the STScI archive, 112 F656N reference frames were identified for a subarray-matched query. These were classified into 45 Quality 1, 14 Quality 2, and 53 Quality 3 PSFs, and then filtered using a morphology criterion based on second-order image moments,

$$I = \sum_{k=1}^{K} \left< X, Z^{(k)} \right> Z^{(k)},$$7

and a peak-significance criterion $$I = \sum_{k=1}^{K} \left< X, Z^{(k)} \right> Z^{(k)},$$8. Morphology-significance criteria alone yielded a true positive fraction of 84% and a false positive fraction of 13%; adding the QFIT criterion $$I = \sum_{k=1}^{K} \left< X, Z^{(k)} \right> Z^{(k)},$$9 lowered the true positive fraction to 67% but reduced the false positive fraction to 6% (Sanghi et al., 2021).

4. Observing implementations across facilities

RDI has been implemented in markedly different operational modes. At Keck II, the NIRC2 vortex-coronagraph examples were taken in $Z^{(k)}$0 band at about $Z^{(k)}$1, with angular resolution about $Z^{(k)}$2 and a plate scale of $Z^{(k)}$3pixel, in vertical-angle mode so the pupil stayed fixed and the field rotated. The same data could therefore be reduced either with ADI or with RDI (Ruane et al., 2019).

At VLT/SPHERE, “star-hopping RDI” introduced an observing sequence in which the telescope alternates rapidly between the science star and a matched reference star. In the HR 8799 experiment, the observing recipe was about 10 minutes on the science target followed by about 5 minutes on the reference star, with a $Z^{(k)}$4 minute gap per hop. The method required a reference star within $Z^{(k)}$5 mag and roughly $Z^{(k)}$6–$Z^{(k)}$7 on the sky; for HR 8799 the chosen reference was HD 218381, $Z^{(k)}$8 away, 0.52 mag fainter in $Z^{(k)}$9, and 0.75 mag brighter in $\{R^{(k)}\}$0 (Wahhaj et al., 2021).

A different SPHERE implementation combines focal-plane wavefront control with RDI. In broadband H-band observations of HR 4796 with the four-quadrant phase mask, a dark hole was first created on-sky using pair-wise probing and electric field conjugation, then broadband images of HR 4796 and a nearby reference star were acquired with the same deformable-mirror shape. The dark-hole region covered 183 mas to 625 mas and remained stable enough between target and reference to support RDI, with the stability metric

$\{R^{(k)}\}$1

yielding $\{R^{(k)}\}$2 for the dark-hole case, corresponding to about 80% to 86% stable residual intensity (Galicher et al., 2024).

Space-based RDI has taken both classical and synthetic forms. In JWST/NIRCam observations of HD 19467 B, the intended reference star was not acquired because of a guidestar failure, so no real reference-star data were available. RDI was nonetheless applied using synthetic PSFs built from WebbPSF and webbpsf_ext with contemporaneous optical path difference maps from the JWST wavefront sensing team; this “SynRDI” was folded into the KLIP basis used by pyKLIP, with 15 KL modes chosen empirically (Greenbaum et al., 2023). On HST/WFC3, RDI for PDS 70 b used only archival PSFs and did not depend on roll-angle scheduling, precisely because the PSF basis came from the STScI library rather than a contemporaneous calibrator (Sanghi et al., 2021).

Simulations for SCALES medium-spectral-resolution exoplanet spectroscopy studied yet another observational geometry: a planet-centered field of view of $\{R^{(k)}\}$3 arcsec in M band, covering $\{R^{(k)}\}$4–$\{R^{(k)}\}$5 at $\{R^{(k)}\}$6. Here the host star remained off-center but still generated quasi-static speckles, and RDI was modeled as direct subtraction of an idealized reference-star PSF under identical observing conditions (Desai et al., 2023).

5. Quantitative performance for companions at small angular separations

The clearest empirical case for RDI as a small-separation method comes from the Keck/NIRC2 vortex-coronagraph study. For HIP 79124 C at 192 mas and $\{R^{(k)}\}$7, the best RDI reduction gave $\{R^{(k)}\}$8, compared with the best ADI reduction at $\{R^{(k)}\}$9 using $K$0 and PA rotation $K$1. For HIP 78233 B at 141 mas and $K$2, RDI gave $K$3, compared with the best ADI result $K$4 using $K$5 and PA rotation $K$6; HIP 78233 B was the first imaging detection in that work. The reported headline result was an increase in detection significance by up to a factor of 5 relative to ADI (Ruane et al., 2019).

The broader Keck/NIRC2 performance study over 359 observations of 304 unique targets found a complementary statistical result: for the typical parallactic-angle rotation of the dataset, about $K$7, RDI provided gains over ADI for angular separations smaller than $K$8. The ADI-vs-RDI crossover was parameterized by a critical PA rotation following

$K$9

with $\sim 1.5$0, and a $\sim 1.5$1-slope fit corresponding to

$\sim 1.5$2

This implies that, for that PCA-based ADI setup, ADI begins to outperform RDI when the companion motion reaches about half a FWHM on the detector (Xuan et al., 2018).

Super-RDI extended the small-separation regime of RDI on Keck/NIRC2 by enlarging the PSF archive. Using a $\sim 1.5$3-frame multi-year library, and for typical rotation of $\sim 1.5$4, Super-RDI performed better than a widely used implementation of ADI at separations $\sim 1.5$5, gaining an average of 0.25 mag in contrast at $\sim 1.5$6 and 0.4 mag in contrast at $\sim 1.5$7. The authors interpreted this as an extension in separation space over earlier same-night, $\sim 1.5$8-frame RDI libraries (Sanghi et al., 2024).

On SPHERE/IRDIS archival H23 data, RDI outperformed ADI at small angular separations under median observing conditions and delivered an average gain of $\sim 1.5$9 mag over ADI at $3\,\lambda/D$0. The same study found that increasing the master library from 60 observations to 725 observations yielded about a 1 mag gain at $3\,\lambda/D$1 and about 0.5–0.8 mag improvements from $3\,\lambda/D$2 to $3\,\lambda/D$3, with performance saturating around 3000–5000 reference images (Xie et al., 2022).

Star-hopping RDI on SPHERE targeted even smaller separations in the search for a fifth planet in HR 8799. The authors reported that the contrast improvement at $3\,\lambda/D$4 could be up to 2 magnitudes over ADI, and that an injected planet at $3\,\lambda/D$5 was not recovered at all in the 1.5-hour IRDIS ADI subset, whereas the star-hopping RDI reductions recovered it clearly. The full IFS $3\,\lambda/D$6 limits were 11.2 mag at $3\,\lambda/D$7, 13.5 mag at $3\,\lambda/D$8, 14.4 mag at $3\,\lambda/D$9, and 15.0 mag at $X$00, corresponding to 6.5, 3.1, 2.3, and 1.8 $X$01, respectively, for an assumed age of 30 Myr and BT-Settl models (Wahhaj et al., 2021).

Not all companion-oriented RDI studies report gains. The GPIES RSDI/TLOCI work, using a PSF archive of 207 references and focusing on the inner 100 mas to 300 mas annulus of the December 2014 51 Eri dataset, found no apparent improvement in SNR from adding the PSF library or from explicit forward-model SNR optimization. The bootstrap SNRs were 12.55 for tar+lib opt, 12.45 for tar opt, and 12.95 for un-opt. This has been interpreted in the paper as evidence that current ADI+SSDI workflows may already be optimized for that dataset and that limited gains can be achieved by using a PSF archive (Gerard et al., 2016).

6. Disk imaging, morphology preservation, and extended-source methods

Disk imaging is the domain in which RDI’s conceptual advantage over ADI is most explicit. In the Keck/NIRC2 vortex-coronagraph study, ADI of MWC 758 was dominated by self-subtraction, whereas RDI recovered the north and south spiral arms and retained a proposed third spiral arm visible in both reductions. For 2MASS J16042165-2130284, ADI was essentially consistent with noise for all $X$02, while RDI clearly revealed the ring of scattered light. The J1604 RDI reduction used the best 88 reference frames out of 150 and 44 PCs, and the ring morphology remained stable if the number of frames or PCs was varied by about 20% (Ruane et al., 2019).

SPHERE/IRDIS archival RDI reinforced the same point at survey scale. It resolved 33 disks in total intensity—19 planet-forming disks and 14 debris disks—with 4 disks detected only with RDI and not with ADI: TW Hya, V1094 Sco, UX Tau A, and HD 169142. Two disks, DG Tau A and HD 131488, were resolved in scattered light for the first time, and three disks, V1094 Sco, UX Tau A, and SZ Cha, were detected in total intensity for the first time. The study emphasizes that the disks not detected in ADI but recovered in RDI tend to be low-inclination systems, exactly those most vulnerable to ADI self-subtraction (Xie et al., 2022).

The HD 169142 study provides a canonical reinterpretation enabled by RDI. It defines RDI simply as subtraction of the reference image of one or several stars from the target image, and uses the complete SPHERE GTO database to find the highest-correlation reference dataset, typically with correlation coefficient $X$03. In the resulting reductions, a bright inhomogeneous ring at $X$04 mas became apparent, showing that bright “blobs” seen in PCA reductions were actually part of the ring. RDI also showed a marginal inner ring at $X$05 mas. Combined with polarized-light imaging and cADI simulations, this led to the conclusion that earlier candidate companions were more plausibly disk features tracing a bright ring with Keplerian motion (Ligi et al., 2017).

Constrained-RDI methods specifically target the oversubtraction that can persist even in classical RDI for disks. PI-constrained RDI uses a polarized-intensity image $X$06 as a conservative estimate of disk morphology, with $X$07, and can further convert PI into a total-intensity estimate through a fractional-polarization model

$X$08

leading to

$X$09

For AB Aurigae, this produced an oversubtraction-free detection of the disk in total intensity and allowed a more decisive recovery of the spectral signature of AB Aurigae b (Lawson et al., 2022).

DIKL targets similar failure modes analytically. On SPHERE/IRDIS star-hopping RDI data for HD 169142, PDS 201, and HD 129590, DIKL reportedly recovered a two-ring system for HD 169142, revealed fine extended structures not seen in KLIP for PDS 201 and HD 129590, and agreed with DI-sNMF within $X$10 in disk-hosting regions while being about 3 orders of magnitude faster (Ren, 2023). ARDI with iterative PCA generalizes the same objective by combining science and reference information. Across synthetic disk-injection tests and 48 real protoplanetary-disk datasets, it often improved disk-image quality relative to ADI or RDI alone, particularly for ambiguous extended structures and for cases in which the reference library quality was suboptimal (Juillard et al., 2024).

7. Operational constraints, caveats, and emerging directions

RDI’s performance depends on PSF matching, observing conditions, and the structure of the reference library. The Keck/NIRC2 performance survey found that random-forest models could explain 70%–80% of the variance in ADI detection limits but only 30%–50% of the variance in RDI detection limits, with the most important RDI predictors including $X$11 magnitude, total integration time, PA rotation, FWHM, and RDI reference-library size. This lower predictability was explicitly linked to factors not yet modeled well, especially target-reference PSF similarity and vortex centering accuracy (Xuan et al., 2018).

Reference quality is also not reducible to a single scalar metric. The disk-focused SPHERE/IRDIS library-selection study concluded that no single physical parameter fully captures PSF similarity across all observing conditions. PCC-only libraries gave the best mean contrast, mixed libraries gave the most robust results and best disc S/N, epoch matching was especially useful for small-separation discs, and spectral type and DIT were poor standalone selectors (Stasevic et al., 3 Sep 2025). This suggests that RDI reference selection is inherently multi-factorial.

Synthetic RDI on JWST has introduced additional caveats. For HD 19467 B, the use of WebbPSF-derived references provided the diversity needed to enhance contrast in the inner working region, but the choice of 15 KL modes was made by visual inspection, and the paper explicitly called synthetic RDI somewhat heuristic in that dataset. The MASKLWB mask position was not yet well calibrated, with an approximate $X$12-offset of 70 mas, and the absence of a real reference star meant that direct comparison between synthetic and classical RDI could not yet be made (Greenbaum et al., 2023).

Reference-star properties themselves can degrade RDI. Simulations of JWST/NIRCam coronagraphy with binary reference stars found that the brightest tested companions, with relative brightness $X$13, produced the worst local sensitivity loss of 3.02 magnitudes, whereas companions at $X$14 relative brightness had almost no practical effect. Position angle altered local sensitivity loss by roughly 0.3–0.5 mag, and the authors recommended preferring the dimmest binary possible, avoiding binaries brighter than or equal to $X$15 if possible (Stephenson et al., 9 Apr 2025).

At the same time, recent work emphasizes that RDI need not remain a stand-alone alternative to ADI. ARDI treats angular and reference diversity jointly (Juillard et al., 2024); dark-hole-assisted RDI demonstrates that focal-plane wavefront control can make target-to-reference residuals sufficiently stable for post-processing gains (Galicher et al., 2024); and synthetic-reference strategies on JWST suggest that the routinely measured optical configuration can, in some regimes, substitute for a missing calibrator (Greenbaum et al., 2023). This suggests a broader shift from “RDI versus ADI” toward hybrid calibration architectures in which reference information, angular diversity, and model-based corrections are combined rather than opposed.

In summary, the literature presents RDI as a family of PSF-subtraction methods centered on external-reference calibration. Its strongest demonstrated use cases are close-in companions under limited field rotation, survey regimes with modest PA rotation and large reference libraries, and disk imaging where ADI self-subtraction distorts or suppresses extended morphology. Its principal limitations are the need for highly correlated references, the difficulty of predicting performance from metadata alone, and the persistence of oversubtraction when disk signal leaks into the PSF model. Current developments—large archival PSF libraries, constrained or imputation-based KLIP variants, synthetic references, dark-hole coupling, and ARDI—indicate that the most active frontier is not the basic definition of RDI itself, but the design of reference models that preserve astrophysical signal while remaining operationally efficient across ground- and space-based instruments.