Designing from physical bounds instead of from available parts — what the machine becomes when it is built by a replicator, computes with light, and remembers holographically.
The other papers ask "what can we build?" That question silently imports the parts bin, and the parts bin encodes a decade of other people's compromises. This paper asks the inverse: what would we build if the only constraints were the ones nature enforces?
The method keeps the exercise honest. For each subsystem: state the physical bound, compute where it sits numerically, place the design against it. A design at a bound is finished — it cannot be improved without new physics. A design far from its bound is an engineering problem, and the distance says how much is left. And a goal on the wrong side of a bound is not a hard problem: it is a different machine, and saying so early is worth more than optimism.
C = B log₂(1+SNR).
Abbe/diffraction ceilings independent addressable features near λ/2 —
which is why storage density lands near one bit per λ³.
Baryon conservation forbids making matter from energy at any practical
budget, so the machine is a rearranger; this is doctrine, not aspiration
(§1).
The second law requires the entropy of any ordering step to leave in
some channel.Everything downstream — file size, data rate, storage volume, whether the transporter is possible at all — falls out of this one number, and it is not a single number but a ladder. Take a 250 g steel object, roughly a mug: about 30 cm³, about 2.7 × 10²⁴ atoms.
| description level | what it records | size | at 1 TB/cm³ holographic | at 1 Tbit/s |
|---|---|---|---|---|
| Functional | "300 mL cylinder, 4 mm wall, stoneware" — parameters, not geometry | ~200 B | invisible | instant |
Chord / .pattern | resonance model: complex poles + port-vectors, ~100 B per chord (compiler §4) | ~1 MB at 10⁴ chords | 10⁻⁶ cm³ | 8 µs |
| Voxel + material, 100 µm | 3 × 10⁷ voxels × ~20 bit (material ID + local state) | ~75 MB | 7 × 10⁻⁵ cm³ | 0.6 ms |
| Voxel + material, 10 µm | 3 × 10¹⁰ voxels — the fine-finish limit | ~75 GB | 0.075 cm³ | 0.6 s |
| Atomic microstate | every atom's species and place — the transporter's description | ~3 × 10²⁴ B | 3 × 10⁶ m³ — a cube 150 m on a side | ~860 000 years |
The design point is therefore megabytes to tens of gigabytes per object — a video file. Every intuition that a matter compiler must involve astronomical data is an intuition about the red bar, which is not what a replicator does — a point Star Trek's own engineering documentation makes by putting replicators at molecular resolution and transporters at quantum resolution.
Three rates, and the interesting result is that the one everybody thinks of is the smallest.
| loop | derivation | rate |
|---|---|---|
| Placement | 3 × 10⁷ voxels for a 250 g object in an hour | ~10⁴ voxel/s — trivial |
| Field update (drive) | the boundary must be restated far faster than the field settles: ~10⁶ states/s in the air band, ~10⁷ in the melt band, at ~18 channels × 32 bit (amplitude + phase) ≈ 576 bit/state | 0.6 → 6 Gbit/s |
| Sense (listen) | the firehose: 12 acoustic ports at 200 MSa/s × 12 bit ≈ 29 Gbit/s; the EM listen loop at 12 ports × 10 GSa/s × 8 bit | ~1 Tbit/s |
So the machine is not bandwidth-limited by building. It is bandwidth-limited by listening — and the listening is what makes the build correct, because every verb in this architecture is closed-loop (§6.2). Worse, the loop has a deadline: a correction is only useful inside the field's coherence time, which is microseconds in the acoustic band and nanoseconds in the EM band.
Terabits per second in, decision in nanoseconds. That is not a processor workload. No amount of clock speed makes a von Neumann machine consume a Tbit/s sensor stream and answer within a wave period — the data would spend its whole budget crossing a memory bus. The control loop is therefore the first-principles argument for computing in the physical layer, and it arrives without anyone having to like the idea aesthetically.
The Star Trek answer to this is the isolinear optical computer. The first-principles answer arrives at the same place twice, by different roads.
What the control loop needs is not branchy logic but a large linear transform applied to a wide sensor vector at wave speed: the compile step of the field compiler is a holographic transform, the scan step a correlation against a dictionary — exactly what optics does natively and for free (a lens performs a Fourier transform in the time light takes to cross it; a hologram, a correlation in the same). The project's compute lane already owns this substrate: the twelve-port toroidal calculator of the optical supercomputer paper, evaluated against the four-axis framework in Optimizing Compute.
This machine is supposed to be buildable by a machine like itself. That constraint rules on the computer, and the ruling is sharp:
Transistor logic is chemistry; optical computing is geometry. A modern processor needs doped junctions, atomic-layer films and lithography at a small fraction of the wavelength used to print it — a supply chain a replicator cannot close. An optical computer is waveguides, resonators, gratings and index contrast: shape. A machine whose entire purpose is to place matter to a pattern can make shape. So the endgame machine's brain is optical not because light is glamorous, but because it is the only brain the machine can build for itself.
Storage lands on the same wavelength argument as everything else. Volumetric holography stores on the order of one bit per λ³. At λ = 500 nm that is 8 × 10¹² bit/cm³ ≈ 1 TB/cm³ BOUND. Set against §2, a full object library is small:
| at 1 TB/cm³ | objects held |
|---|---|
| 1 cm³ (a sugar cube) | ~10⁴ objects at 100 MB each |
| 10 cm³ (a chip, an "isolinear card") | ~10⁵ objects — more than a household will ever ask for |
| 1 cm³, atomic-microstate objects (§2 red bar) | 3 × 10⁻⁷ of one object |
The current art is a long way below the bound, and the useful thing about the recent work is the direction it moves: Chen, Wang, Wu, Song, Yang, Lin & Tan encode in amplitude, phase and polarisation together and decode with a convolutional network straight from intensity, removing the step-by-step reconstruction (Optica 13, 591–601, 2026) MEASURED, explicitly research-stage. That is not an attack on the λ³ bound — nothing beats it — but it is the right strategy for approaching it: use more of light's degrees of freedom per site. λ³ counts diffraction-limited sites, not bits per site, and amplitude × phase × polarisation is a larger alphabet than intensity alone.
Von Neumann's universal constructor arranges primitive parts it cannot synthesise — it works from a "sea of parts." The honest question is not whether the sea vanishes, but how far it drains.
| subsystem | self-fabricable? | why |
|---|---|---|
| structure, enclosure, chassis | yes | bulk geometry — the machine's core competence |
| phononic screens, metasurfaces, gratings | yes | patterned geometry at the design wavelength; the machine's own optical channel patterns at µm |
| single-crystal transducers | plausible | directional solidification with field-programmed microstructure is already a specified capability (spec §6.1) |
| optical compute elements | yes, if optical | waveguides and resonators are shape — §4.2 |
| holographic media | plausible | a photopolymer or doped glass volume is a bulk phase |
| conventional ICs, power semiconductors | no | doping and lithography chemistry outside the machine's repertoire — the irreducible residue |
So the closure fraction is high and not unity — the correct von Neumann answer, stated plainly: the sea of parts shrinks to a handful of chips. A machine that makes its own structure, transducers, apertures, optics and memory — and buys a controller — is the honest form of the claim, and the first version worth calling self-replicating without an asterisk.
Trek separates the replicator (molecular resolution, inanimate objects) from the transporter (quantum resolution, living beings). It is easy to read that as a plot convenience. Read against §2 it is exactly where the bound falls:
| replicator — molecular | transporter — atomic/quantum | |
|---|---|---|
| description size | 10⁷–10¹¹ bit | ~3 × 10²⁵ bit |
| store one object | 10⁻⁶–0.1 cm³ | 3 × 10⁶ m³ |
| move at 1 Tbit/s | µs–s | ~10⁶ years |
| verdict | an engineering problem | a different machine — and no-cloning forbids the quantum-exact version outright |
The gap is thirteen orders of magnitude in storage and six in time, and no plausible improvement in either closes it: you cannot engineer your way across a factor of 10¹³ by being clever about encoding, because the bound is counting distinguishable states, not bytes. This is the one place in the document where the answer is "no, and here is the number."
The replicator is reachable. The transporter is not — and the same arithmetic that says so is what tells us the replicator's file is the size of a video.
| bound | where it sits | where we sit | distance |
|---|---|---|---|
| Object description (Abbe / molecular) | ~10⁷–10¹¹ bit | .pattern chords, MB-class | at it — the format is already the compressed form |
| Storage density (1 bit/λ³) | ~1 TB/cm³ | current holographic art, research-stage | orders below; direction correct (multi-DOF encoding) |
| Sense channel (Shannon) | set by band × SNR | ~1 Tbit/s design target | near it for the aperture we have — more needs more aperture |
| Compute latency (loop deadline) | wave period: ns–µs | optical/physical layer | at it by construction — that is why the layer exists |
| Compute energy (Landauer) | ~3 zJ per erased bit | far above; but the linear optical transform is nearly reversible | far — and the honest headroom is reversibility, not clock tricks |
| Matter (baryon conservation) | rearrangement only | rearranger by design | at it — doctrine, §1 |
| Energy ledger (2nd law) | ordering exports entropy | harvest loop + entropy-routing cooling | at it in principle; efficiency is the open engineering |
| Self-closure (von Neumann) | sea of parts never empties | everything but the controller | near it — residue is a handful of chips |