A companion to Persistence Structure of Bandwidth-Limited Observation
The “observer” in this paper is a small neural network. A recurrent model (one that processes sequences step by step, carrying forward what it’s learned so far) that watches data one episode at a time, like someone looking through a narrow window and trying to figure out what’s happening outside. It sees a short clip of some physical signal, updates its internal state, then sees the next clip. Its only job is to predict what the next observation will look like.
The experiment asks a single question: does remembering previous episodes help? For some phenomena, memory is useless. A black hole shadow looks the same every time you look. One clip is enough. For others (a gravitational wave chirp, the 11-year sunspot cycle) the signal evolves on timescales longer than any single clip. You have to accumulate observations before the structure becomes visible. The number we measure is the persistence advantage: how much better the observer performs when it remembers versus when it doesn’t. Positive means memory helps. Zero means each observation is already complete.
Everything about the observer (its architecture, its loss function, how it’s trained, how it’s evaluated) is held exactly the same across every experiment. The only thing that changes is the data it watches. So any difference in the persistence advantage is a property of the domain, not the network. To make sure, we tested four architecturally different observers: a GRU (sequential, with a forget gate), an LSTM, a Vision Transformer, and a full-attention Transformer (no sequential bottleneck, growing context window). All four agree. The network is a proxy in the same way a thermometer is a proxy. You don’t study thermometers to understand temperature, but if you build a good one and point it at different things, the readings tell you about the things, not the thermometer.
Thirteen real-data domains from independent instruments. LIGO gravitational wave strain. Event Horizon Telescope interferometric visibilities. CMB power spectra. Type Ia supernova distances. Solar wind magnetometer data. Sunspot counts. Quasar variability. Fast radio bursts. Neutrino events. The S2 star orbiting Sagittarius A*. Three additional gravitational wave detections. All public data. All run through the same pipeline.
1. There is a single axis organizing physical observation
When you embed 24 physical domains by how they respond to different observation strategies, the resulting space has one dominant direction. It explains more than half the variance. Bootstrap resampling shifts it by less than 0.01 degrees.
At one end: electromagnetic domains. Stellar lightcurves, radar Doppler, radio waterfalls. At the other end: gravitational wave strain.
This axis was hypothesized, not imposed. The conjecture that observation domains might organize along a single temporal axis motivated the experiment — but the pipeline assumed nothing about it. It emerged from the data. We defined it using the EM-gravity direction, but removing the gravitational wave domain entirely changes the axis by less than 0.01 degrees. The remaining 23 domains define essentially the same line.
The axis measures one thing: how much temporal accumulation a bandwidth-limited observer needs before the domain’s structure becomes visible.
2. The law has a closed form
The persistence advantage of a domain at position x on this axis is:
P = σ · C · τ · κ(x) · (1 − xδ)
Five parameters, each with independent empirical support:
- x: position on the temporal axis (0 = electromagnetic pole, 1 = gravitational pole)
- κ(x): a measured quadratic describing how channel-dependence increases along the axis (R² = 0.85, p < 10⁻⁸)
- C: channel-match quality, the maximum temporal autocorrelation of the observed signal, computable before training
- σ: a scale factor calibrated from GW150914
- δ: a carrier flag, 1 when electromagnetic radiation observes gravity, 0 otherwise
- τ: a provisional depth modifier for within-interface variation
3. Gravity is a statistical outlier in channel-dependence
The gravitational wave domain’s observability depends on which channels are available 8× more than any electromagnetic domain. Grubbs test: z = 6.77, p < 10⁻⁸. No other domain exceeds Δ = 2.3.
Remove temporal-structure-resolving channels (STFT, wavelets, recurrence plots) and the gravitational wave outlier vanishes. Gravity becomes indistinguishable from other domains. The outlier is channel-contingent: it exists only when the observer has the right instruments.
4. The boundary zero is exact and indestructible
When electromagnetic radiation observes a black hole (δ = 1, x = 1), persistence goes to zero. Measured: EHT Sgr A*, P = −6 × 10⁻⁶. EHT M87 (held out from calibration), P = −4.1 × 10⁻⁶.
We tried to break this zero with 14 different observation operators on M87: amplitude-only, phase-only, scrambled phases, scrambled amplitudes, permuted scans, restricted UV coverage, per-night splits. 140 total runs. Every one gives P ≈ 0. The zero survives the destruction of geometric coherence, phase structure, amplitude ordering, and temporal sequencing.
The mechanism is observational sufficiency: the black hole shadow is static at the episode scale. Each observation already contains the full signal. There is nothing to accumulate.
Every tested pure binary black hole merger, by contrast, shows positive persistence. Ten out of ten distinct events. All positive. The boundary between zero and positive is sharp.
5. GW170817 tests the carrier flag
GW170817 was a neutron star merger. A gravitational wave event with an electromagnetic counterpart. The law predicts that when the carrier is present in the gravitational signal, δ activates and persistence suppresses.
Measured: P ≈ 0. A gravitational wave event at zero persistence, exactly where the law says it should be.
6. The axis is not about force type. It’s about observation cost
Among domains with measurable temporal structure, the implied axis positions form a monotonic ordering:
| Domain | C | Implied x | What it is |
|---|---|---|---|
| Solar wind | 0.261 | 0.52 | fast EM |
| CMB | 0.980 | 0.63 | acoustic structure |
| S2 orbit | 0.918 | 0.83 | slow orbital dynamics |
| Sunspots | 1.000 | 0.91 | 11-year cycle |
| GW150914 | 0.987 | 1.00 | chirp signal |
Fast electromagnetic phenomena sit near x = 0. Slow electromagnetic phenomena (sunspots, the solar cycle) sit near x = 1 alongside gravitational waves. The axis measures accumulation cost relative to the observer’s bandwidth, not force type.
7. Four architectures agree
The axis was discovered with a ViT-Small embedder and confirmed with a GRU recurrent network. An LSTM ablation (50 seeds, p = 2 × 10⁻⁶) preserves it. A full-attention Transformer (no sequential bottleneck, no forget gate, growing context window) reproduces the same persistence pattern across five key domains.
The temporal axis is a property of the observation problem. Not the network.
8. The law generalizes beyond physics
Non-physical domains with high temporal autocorrelation land on the same κ(x) curve as gravitational waves. Weather (seasonal cycle), random walks (Brownian drift). Implied positions: weather at x = 0.98, random walk at x = 0.89.
Non-physical domains with low temporal autocorrelation (stock returns, seismic catalogs) scatter off the curve.
The curve is universal. The geography (EM at one pole, gravity at the other) is specific to physics.
9. The axis positions shift with bandwidth
When you double the observation window from 16 to 32 episodes, LIGO strain drops from x = 1.00 to x = 0.66. Solar wind drops from x = 0.54 to x = 0.11. The κ(x) curve holds at every tested bandwidth. What moves is where each domain sits on it.
Axis positions are not intrinsic to domains. They are ratios between the domain’s temporal structure and the observer’s bandwidth. With enough bandwidth, every domain approaches x = 0. There is no absolute frame.
10. C predicts the ordering of black hole mergers
Across nine pure binary black hole events, the channel-match quality C (maximum temporal autocorrelation of the strain signal) is the strongest predictor of persistence (Spearman ρ = 0.83, p = 0.005). Inverse peak frequency is also significant (ρ = −0.75, p = 0.020): slower mergers need less persistence, consistent with the bandwidth result.
A decomposition of C into named components shows that only temporal autocorrelation predicts persistence. Signal-to-noise ratio does not. Detector cross-correlation does not. C is literally a measure of how much the observable signal persists across episodes.
What a black hole is, on the observation manifold
A domain where every observation is already complete. The shadow doesn’t evolve between episodes. There is no temporal structure to accumulate. The persistence function has an exact zero there. Not because the observer fails, but because the domain has nothing left to teach. In general relativity this boundary is called the event horizon. On the observation manifold it is the node where P = 0, confirmed by the data to six decimal places.
What would kill this
- An electromagnetic observation of a black hole with large P.
- A pure gravitational wave event with P = 0.
- A domain at high x with low channel-dependence.
- Manifold dimensionality increasing when channels are removed.
None of these has occurred across 25 confirmed tests from 13 real-data domains.
What the paper is not claiming
The paper does not claim that gravity “is” the observation interface in a physically deep sense. It claims that persistence cost under bounded observation has reproducible structure across diverse physical domains, that this structure is organized by a single axis with EM and gravitational poles, that the structure is observer-independent, and that it shifts lawfully with bandwidth.
The model is empirical. The carrier flag δ is an imposed boundary constraint, not a deduction. The depth parameter τ is provisional. Catalog-type domains where accumulation is statistical rather than temporal (fast radio bursts, ECG) produce unphysical axis positions and mark the boundary of the law’s current form.
Whether the EM-gravity geography reflects a contingent fact about the tested domains or a more general regularity remains open.
Read the full paper