The Sufficient Observer

Two papers extract maximum information from minimal measurement — one using a supermassive black hole, one using a single pixel — and both demonstrate that observation power comes from theory, not from sensor resolution.

Domcke et al. (arXiv: 2604.01290) use broadband electromagnetic observations of M87’s supermassive black hole to constrain high-frequency gravitational waves from 10^10 to 10^27 Hz. The mechanism: gravitational waves passing through a magnetic field convert to photons via the inverse Gertsenshtein effect. M87 provides both the magnetic field (from its accretion disk and jet) and the electromagnetic background against which any excess signal would appear. The black hole isn’t a purpose-built detector — it’s a natural laboratory with the right properties. The constraint comes not from detecting gravitational waves directly but from not detecting an electromagnetic excess that their conversion would produce.

Chen et al. (arXiv: 2604.01801) achieve scene classification without full hyperspectral imaging by using compressive phasor encoding with a single pixel detector. Instead of capturing a complete hyperspectral dataset and then classifying, they encode wavelength information as phases and compress the entire spectral dimension into a few measurements. The result requires two orders of magnitude less data than conventional hyperspectral imaging. The single pixel doesn’t see less — it sees differently, extracting exactly the information needed for classification without acquiring the information that isn’t.

The structural claim: the power of an observation is determined by the theory behind it, not the resolution of the sensor. M87’s broadband spectrum constrains gravitational waves across 17 orders of magnitude in frequency — not because anyone built a detector that sensitive, but because the theoretical prediction of what graviton-photon conversion would look like is specific enough to constrain. A single pixel classifies scenes — not because one pixel is enough to image anything, but because phasor encoding extracts the spectral features that distinguish classes without acquiring the spatial features that don’t.

This inverts the common intuition that better science requires better instruments. Sometimes it does. But often the breakthrough comes from realizing that existing data already contains the answer, if you know what to look for. M87 has been observed electromagnetically for decades. The gravitational wave constraint was always there — it just required the theoretical framework to extract it. Hyperspectral scenes contain classification information in their spectra. The single-pixel compressive measurement just extracts that information without the overhead of full spatial imaging.

The mathematical structure is similar in both cases: a transform that projects high-dimensional data onto a low-dimensional space that preserves the information relevant to the question while discarding information irrelevant to it. For M87, the projection is from the full gravitational wave spectrum to the electromagnetic excess it would produce. For the single pixel, the projection is from the full hyperspectral data cube to the phasor-encoded spectral features.

Chen et al.‘s two-orders-of-magnitude data reduction is quantitatively precise about what’s lost and what’s kept. The spatial information is lost — the single pixel can’t tell you where things are in the scene. The spectral classification information is kept — the pixel can tell you what the scene contains. The sensor resolution didn’t decrease; the measurement strategy became more targeted.

Domcke et al.’s approach is even more striking: the “sensor” is a black hole 55 million light-years away, and the “measurement” is data that already existed. The gravitational wave constraint is free — it comes from reanalyzing existing observations through a new theoretical lens. No new instrument was built. No new observation was made. The information was already present, waiting for the theory that could extract it.

The implications for any field that collects data: you probably already have the answer to questions you haven’t thought to ask. The constraint isn’t data volume — it’s theoretical specificity. The question isn’t “do we have enough data?” but “do we have the right theory to read the data we already have?”