Signal Injection–Recovery Test¶

An end-to-end validation protocol — instantiates Model-Guided Signal Separation

Adds a known synthetic signal into real data, runs the whole extraction pipeline, and checks how faithfully it comes back — measuring the pipeline's bias, completeness, and detection limit.

Signal Injection–Recovery Test validates a pipeline by feeding it a signal whose truth is known and measuring what comes back. Because the injected signal was constructed, comparing recovered to injected exposes the pipeline's amplitude bias, its shape distortion, and — most distinctively — the fraction of signals it recovers at all, its completeness, across the entire chain: preprocessing, separator, and thresholds together. What makes it this mechanism and not its siblings is that it tests recovery end-to-end rather than inspecting any single stage or residual. Where the whiteness check asks "is the leftover structureless?", this asks "if a real signal were there, would we get it back, and undistorted?"

Example¶

To calibrate a transit-detection pipeline that hunts for planets in stellar light curves, analysts inject thousands of synthetic transits of known depth and orbital period directly into real light curves, then run the unmodified pipeline and record which injections are recovered and how the measured depth compares to the injected one. The output is a completeness map — recovery fraction as a function of planet size and period — together with an honest detection limit (the size below which recovery collapses) and any systematic depth bias the pipeline imposes. None of it rests on a real discovery; the injected signals are fabricated on purpose, which is exactly what makes the truth known.

How it works¶

Known signals are generated to span the parameter range of interest, added into real (or realistic) data, and passed through the full pipeline blind to which are injected. Recovered parameters are compared to the injected truth to yield three things a single run of the pipeline never reveals: the recovery fraction, the bias in recovered amplitude or shape, and the signal level at which recovery falls off. Repeated across the parameter grid, it maps performance. Its distinguishing contribution is completeness — it is the only sibling that measures what the pipeline misses, not merely how it distorts what it catches.

Tuning parameters¶

Injection parameter grid — the range and density of injected signals; a broad, dense grid maps performance finely but multiplies pipeline runs.
Injection realism — injecting into real backgrounds with realistic signal shapes versus idealized ones; idealized injections flatter the pipeline.
Number of injections — more give tighter completeness and bias estimates but scale the compute linearly.
Blind vs. informed recovery — whether the recovery step knows an injection is present; blind recovery is the honest test.

When it helps, and when it misleads¶

It helps by turning "seems sensitive" into a measured detection limit and a bias correction — the one test that quantifies what a pipeline misses rather than only how it mangles what it catches. It misleads when the injections are unrealistic: an idealized signal recovered perfectly says little about real ones, and the test validates recovery only of the shapes you actually injected, so failures on un-injected morphologies stay invisible. The classic misuse is injecting only easy signals to advertise sensitivity. The discipline is to inject realistic and even adversarial cases across the true parameter range, and to pair the result with a residual whiteness check on real outputs.

How it implements the components¶

extraction_validation_plan — it is the validation plan: a defined protocol for exercising the whole pipeline against known injected truth.
calibration_reference — the injected known signals are the ground-truth reference that recovery is measured against.
information_loss_metric — recovery fraction and amplitude bias quantify exactly how much of a known signal the pipeline loses or distorts.

It builds no separator of its own — the operator components belong to the extraction mechanisms — and it does not inspect the residual for leftover structure, which is Residual Leakage and Whiteness Check's job.

Instantiates: Model-Guided Signal Separation — the end-to-end recovery validator for a whole pipeline.
Consumes: the full extraction pipeline under test, run unmodified on injected data.
Sibling mechanisms: Residual Leakage and Whiteness Check · Held-Out Sample Test · Supervised Representation Learning · Matched Filtering · Deconvolution and Inverse Filtering · Kalman or Particle Filter

Notes¶

The two validators are complementary and neither substitutes for the other. Injection–recovery measures whether a known signal returns intact — its bias and completeness — while Residual Leakage and Whiteness Check measures whether the residual is structureless — its leakage. A pipeline can pass one and fail the other, which is why a trustworthy extraction is checked from both sides.