Machine learning stores knowledge extensionally. Every item in a collection gets its own code, its own latent vector, its own row in a codebook. For X objects, you need X entries. The question this paper asks is whether a finite collection can instead be stored intensionally: as the decoded orbit of one compact law that returns exactly to its start. Not learned to return, not approximately returning, but algebraically guaranteed to close.
The construction is simple to state. For X objects, encode item i as the i-th phase of a fixed rotation in a learned latent space. Decode all phases with a shared nonlinear network. The rotation is a block-diagonal bank of plane rotations at integer harmonic frequencies of the fundamental 2pi/X, a real discrete Fourier operator. Because the frequencies are integers, R^X equals the identity exactly. No training is required for closure. The orbit is an address generator and a clock; the decoder is the renderer that assigns content to each address.
Why Exact Periodity Is the Mechanism, Not a Detail
The paper isolates exact periodicity as the operative constraint by comparing three operators while holding the decoder fixed. A general learned operator (Koopman-style) is erratic and eventually diverges at larger X. A learned unitary operator (norm-preserving but not periodic) is stable but degrades around the loop because its cycle does not close and phase error accumulates. The exactly periodic Fourier operator is flat and best.
On synthetic 16x16 images at X=64, the general operator scores 19.0 dB, the unitary operator 59.9 dB, and the periodic operator 68.3 dB. On real CIFAR-100 images at X=128, the general operator overflows entirely (no finite reconstruction obtained), the unitary operator collapses from 76.4 to 46.5 dB, and the periodic operator remains at 72.6 dB. At X=16 the general operator wins by 13 dB: with few phases, freedom fits better and the constraint buys nothing. The constraint pays as the orbit grows. Learning the frequencies does not help: integer harmonics already form a complete basis. The operative ingredient is exact periodicity itself.
This is the paper's sharpest result. The difference between the unitary and periodic operators is exactly the difference between "preserves norms" and "returns to start." Both are linear, both are norm-preserving, but only the periodic one closes the orbit. The failure mode of the unitary operator is not explosion or vanishing gradients. It is accumulated phase drift: each step introduces a tiny error that compounds around the loop, and by the time the orbit should close, the decoded images have drifted far from the originals.
Capacity Is the Decoder's Budget
Once closure is guaranteed by the operator, the remaining limitation is representational. The latent requires on the order of X/2 rotating planes to resolve X phases, and the decoder must have enough capacity to render the target objects. The paper accounts for this carefully.
Dense decoders carry a structural overhead of roughly 2.25x per crisp image relative to the raw image bytes. For a 32x32 grayscale image, the break-even set size always lies above the crisp ceiling. At every size tested, the construction can be crisp or compressed, but not both with dense decoders. Extending X to the full 50,000-image CIFAR corpus makes the trade sharper: compression crosses one only beyond 4,000 images, by which point reconstruction has fallen below 20 dB.
A small convolutional decoder changes the picture. On 128 images from a single CIFAR-100 class (objects sharing a low-dimensional visual manifold), a 39K-parameter convolutional decoder reaches 32.2 dB, above the 30 dB crisp line, while occupying 1.66x fewer bytes than the raw pixels. Crisp and compressed simultaneously, a regime no dense decoder reached. On 128 maximally diverse images, the same decoder falls just short at 29.3 dB. The crossing requires both shared kernels and shared structure in the represented collection. An orbit representation stores finite objects below raw-instance cost only when the collection contains structure that a shared decoder can exploit.
The Generative Law Is Free
The codebook control is the paper's cleanest accounting exercise. Replace the rotating base vector with an independently learned latent for every image: a free codebook with 16,000 latent parameters against the cycle's 128. Evaluate over the same decoders and image sets.
The free latents never win. They tie at the smaller decoder and lose modestly at the larger decoder on both shared-manifold and diverse sets, while requiring a latent store 128 times larger. At the larger decoder, the cycle leads by more than a decibel in both settings (gaps of minus 1.16 and minus 1.33 dB, with the cycle better). The structural constraint costs nothing in reconstruction on these settings. It buys parameter economy and three properties the independent codebook cannot provide by itself: a total ordering, a one-step transition from every item to the next, and exact closure.
This is the paper's strongest empirical claim. The cycle is not trading reconstruction quality for algebraic properties. It is getting those properties for free while using 128x less latent storage. The reason is straightforward: the rotation constrains the latent to lie on a known manifold (the orbit), and the decoder only needs to learn the content-to-phase mapping, not the phase-to-phase transitions. The codebook must learn everything independently for every item.
Looping Video: When Phase Is Meaning
For looping video, the cyclic order is not an arbitrary address assignment but the content's own temporal structure. The construction becomes a per-clip video representation in the NeRV family, with one crucial difference: the temporal embedding is exactly periodic in the loop length by construction rather than by training.
Seven standard benchmark clips at 128x128 RGB, all represented as palindromed loops (played forward then backward so every clip closes exactly). Every run is roughly 2.5x smaller than the raw frames. The cycle leads the frame-index baseline on all seven clips, by 0.7 dB on average. On the seam (mean absolute difference between the rendered frame at the wrap point and at phase zero), the two representations differ by five orders of magnitude on every clip and seed: the cycle wraps at numerical precision (never exceeding 3.3e-7), while the frame-index representation never falls below 1.5e-2, producing a visible pop.
The seam makes the representation claim measurable. Exact closure is a property the standard embedding does not have and cannot reliably learn; the cycle has it by construction. A fair alternative is to pin the baseline's frequencies directly to integer harmonics of the loop. That harmonically pinned embedding closes its seam (at or below 8.5e-8), confirming that exact periodicity, however imposed, is the ingredient responsible for closure. But it trails the cycle by 1.9 dB on average and is unstable on one clip (Sintel, with a 6.2 dB spread from one collapsed run). Routing harmonics through a learned embedding network costs fidelity and stability relative to the fixed rotation of a single base vector, which provides the same harmonic structure with zero additional parameters.
The Bandlimited Cycle as Interpolator
Training on every other frame and evaluating the held-out frames fails badly with the full harmonic bank (10 to 18 dB). The reason is aliasing: frequencies above the training grid's Nyquist limit are unconstrained and oscillate freely between samples. Capping the bank at Nyquist repairs the failure. Held-out fidelity rises by 9 to 22 dB, reaches the model's own trained-frame ceiling on near-static content (35.5 to 36.0 vs. 35.7 to 36.2 dB), and beats the frame-index embedding on all three clips tested.
The bandlimited cycle is the Whittaker-Shannon interpolator on the circle. A loop represented at one frame rate can be rendered at another from the representation alone. This is a practical consequence: the same stored representation supports variable playback rates without retraining, something a standard frame-index embedding cannot do without aliasing artifacts.
What the Finite-Knowledge Perspective Adds
The measurements support a representation principle. A finite collection can be encoded as an orbit when three roles are separated: the group action supplies addresses, order, and exact recurrence; the decoder supplies the content associated with each address; and shared structure in the content determines whether the representation compresses or interpolates.
For unrelated objects, the orbit is a compact addressing law. It does not create semantic neighborhoods between adjacent phases, as the missing-phase experiment demonstrates: training a 32-phase cycle on only sixteen phases leaves the untrained phases as blurry class-average blobs, 40 dB below the trained ones. A denoising objective does not change this. The midpoint between unrelated images corresponds to no image, and local smoothing does not invent one. A cyclic indexing of arbitrary finite items supplies order and closure, not semantic interpolation.
For naturally cyclic collections, the orbit can also encode genuine transition structure. Looping video is the measured example. Periodic simulations, robotic gaits, biological cycles, and procedural animation are plausible extensions precisely because their finite states already possess a cyclic order and require recurrence. Taxonomies, graphs, permutation spaces, and branching workflows are not naturally one-dimensional cycles. Forcing them into one orbit would provide an enumeration but could destroy their native adjacency. A broader program would replace the single cyclic generator with generators of a finite group or semigroup acting on latent space, so that the algebra of the representation matches the combinatorial structure of the knowledge.
Scope and Honest Limits
Everything is small-scale and internally controlled: grayscale 32x32 images, 128x128 clips, decoders from tens of thousands to a few million parameters. All experiments were designed, run, and verified on Apple silicon (M1 and M4 machines) with no cloud compute. The codebook control and video study are replicated over three retraining seeds. The operator-comparison tables remain single-seed, with margins of tens of decibels.
Compression claims are against raw pixels and matched baselines, never against image or video codecs. Conventional video codecs compress these clips much further and are not the comparison. For natively non-looping content played once, the construction offers no advantage over a standard frame-index embedding. On content that does not loop, exact periodicity offers no advantage.
The claim is deliberately scoped. Finite cyclic knowledge can be stored as dynamics rather than as independent instances, with exact recurrence supplied by algebra and content supplied by a shared decoder. The economy is not in the decoder weights per se; it is in replacing X independent latent entries with a law. Whether the full representation is smaller than the raw objects depends on the decoder's ability to exploit shared structure. In that sense, the work is about compressing relations (the regularity of how items are addressed and rendered) rather than merely compressing pixels.