Begin with the reference library in the 2012 KLIP paper
In Soummer, Pueyo and Larkin, arXiv:1207.4197v1, read Sections 2.1–2.3 and Equations 5–12 before inspecting Figures 1–2. A reference library is a set of images from which the method learns recurring spatial patterns. Its covariance eigenvectors define a basis of image patterns, ordered by how much variation they explain.
Find where the number of retained modes enters the projection. Removing more modes can reduce residual stellar structure and also remove more source light. Figure comparisons show a trade-off between detectability and characterization; a darker residual image does not specify its surviving flux calibration.
Write down whether the candidate source can enter the references. A separate reference star and an angular sequence containing the target lead to different source contamination. The same basis notation can describe both, but their responses need different calibration.
Reading: Soummer
Pueyo Section 2 changes the basis as well as the source image
Read Sections 2.1–2.5 of arXiv:1604.06097v1 in order. First identify the unperturbed stellar references. Then locate the small astrophysical signal added to them. The perturbation changes the reference covariance, its eigenvalues and its eigenimages. Subtracting only a fixed stellar projection keeps one contribution and omits the basis change.
Use Appendix E, especially Equations E18 and E20, to trace the perturbation terms. Small eigenvalue separations can amplify a change in the eigenvectors. This is a concrete reason to compare finite injection amplitudes and reference selections when the learned basis is unstable.
The public synthetic validation repeats the complete reduction after each channel injection. For one saved KL mode, fixing the reference basis differs from recomputing it by about 1.14 percent in the tested spatial measurements. The full finite-injection derivative itself agrees much better at small amplitude. Those are two different comparisons: derivative linearity and completeness of the basis response.
Reading: Pueyo
Read the spectrum examples as injection experiments
Continue through Sections 3.1–3.3 and 4.1–4.2, then Appendix F, Equations F8–F12. Track which symbol is the injected input spectrum, which is the processed output and which is the inferred spectrum. The illustrative J-band examples include synthetic planets in observed backgrounds. Their known spectra make bias measurable.
An empirical PSF injection into a frozen background asks how a processing choice responds to that added light. It does not by itself sample new photon noise or a new atmospheric realization. Our complete-acquisition tests generate independent counts before rebuilding the references and their covariance.
The independent blind-null calculation also repeats the selection: 140 candidate positions and five choices consisting of mean ADI, mean ASDI and one, two or three KL modes. In sixty-four null acquisitions none exceeds the selected synthetic source peak. The saved result therefore reports a finite binomial upper bound, not a Gaussian discovery level. Its population keeps one prescribed optical-path sequence.
Reading: Pueyo · Mawet et al.
Greco and Brandt tell you how neighboring colors fluctuate
Open arXiv:1602.00691v1 at Section 2, Equations 1–6. The covariance joins fluctuations in different spectral channels. Its off-diagonal structure concerns random deviations around a mean; the response matrix concerns how changing an input signal changes the mean. Neither matrix replaces the other.
Read Sections 3.1–3.2 and Equations 7–8 with the likelihood normalization beside you. A spectrum can contain broad correlated fluctuations, so treating every channel as independent can overstate the information in a broad molecular feature. A prior can further shape an inferred atmosphere posterior. Identify that prior before interpreting a narrow physical-parameter interval.
The sparse GPI calculation provides a useful diagnostic: its full thirty-dimensional spatial fit has a large residual even when the ten aperture sums can be solved exactly. That is a request to improve the image and noise model before passing the extracted points to an atmospheric fit.
Reading: Greco & Brandt
Samland’s extraction starts earlier than a wavelength cube
In Samland et al., arXiv:2210.06390v1, start at Section 3.4, then read Sections 4.2–4.3, 4.6, 5.2–5.3 and 6. A lenslet integral-field spectrograph lays many short spectra onto one two-dimensional detector. Each lenslet image has a finite shape, and adjacent spectral contributions can overlap.
The model fits those raw detector contributions to extract a wavelength cube. Spectral cross-talk can therefore arise before ADI or SDI begins. An already-reduced GPI cube supplies neither the original detector pixels nor the complete lenslet calibration needed to repeat that extraction.
Keep a simple drawing beside the text: raw detector, extracted wavelength cube, PSF subtraction, measured spectrum. Put each error term at the stage where it enters. SDI mixing after extraction and lenslet cross-talk during extraction then stay distinct, even when both create spectral correlations.
Reading: Samland et al. · GPI Data Pipeline 1.4.0 data formats
Electric-field control changes the photons that arrive
Haffert et al., arXiv:2303.13719v1, Sections 2.1–2.4 and Equations 1–22, introduces implicit electric field conjugation, or iEFC. Small mirror probes change coherent stellar speckles. Comparing positive and negative probes reveals their response, so a measured control matrix can guide a correction.
To first order, the difference between the two probe intensities isolates a cross-term between the stellar field and the probe field. The derivation treats an incoherent companion/background component as unchanged across the small probe pair, so that component cancels from the difference. A real off-axis source can also have command-dependent throughput, which needs its own calibration. The argument assumes the field and instrument remain sufficiently stable across the pair and the adopted response is locally linear.
The correction reduces coherent stellar leakage before the science exposure. A learned post-processing basis instead acts on images already recorded. Both need throughput and calibration checks, but they are different interventions in the optical path. The observing Lab implements post-processing; the preserved optical bench shows its pupil and focal planes.
Reading: Haffert et al.
Read the July 2026 on-sky result on its own scale
The fixed source is arXiv:2607.08146v1, On-sky dark-hole diggin’ with implicit Electric Field Conjugation on MagAO-X. Read all four pages, especially Sections 2–2.1 and 3–4 and Figures 1–3. Treat this version as a preprint. The internal calibration and later on-sky sequence test whether a measured control matrix transfers to the aligned observing system.
The apparatus has a dedicated non-common-path deformable mirror after the AO/science split. That placement matters: the controlled aberrations are not simply the same error signal already sensed by the main AO system. Inspect the source, mirror conjugates and science camera together before comparing the images.
Keep three numerical statements separate. Internal bench suppression reaches a different scale from the reported on-sky contrast improvement of roughly two to twenty. A future contrast of one hundred-thousandth near one diffraction unit is a goal in this source. Compare sensitivity after accounting for the probe averaging and science integration times in each observation.
Reading: Haffert et al. · Haffert et al.
A short set of questions for the next paper
Locate the plane where the method acts, the actual measured pixels and the reference selection. Find the source that supplies throughput and ask whether it changes the learned basis. Then locate the response matrix, the noise covariance and the position or background nuisance treatment.
Finally, check an injected or held-out source, a null selection and an explicit failure. Keep the fixed version and the figure or equation next to each conclusion. That small habit makes new high-contrast methods much easier to compare.
Reading: Pueyo · Greco & Brandt · Mawet et al.