Research note Interactive · ~12 min read 2026.08Nako Sung · 성낙호
programmable photonic variational computing

What if causality
could compute the minimum?

Encode action as physical delay. Let propagation perform addition. Keep only the first wavefront that reaches each state. The result is a machine where an enormous space of histories can collapse as fast as light moves through the fabric.

RepresentationAction → time ArithmeticPropagation → add ComparisonCausality → min MemoryWinner → predecessor

The interesting part is not that light is fast.

It is that the ordering of arrival itself can become an arithmetic primitive. If the first arrival at a state is guaranteed to dominate every later history, the hardware can prune locally without ever enumerating the complete trajectories.

Implicit histories16³⁰ ≈ 10³⁶
Active frontier16 states
Critical depth30 stages

See causality choose

Three histories race toward the same state. Only the first is allowed to survive.

ABC A
A 92ps · B 128ps · C 171ps
01 / Scope before spectacle

A claim small enough to be wrong

The first machine is not a universal solver for every stationary-action problem. It targets a narrower class that is clean enough to falsify: discretized additive action over a Markov-complete state.

A[P] = Σₖ a(xₖ → xₖ₊₁)
τₖ = τbase + α · aₖ
earliest arrival = argminₚ A[P]

If a local action increment is negative, add the same per-layer baseline delay to all alternatives. Every candidate crosses the same number of layers, so the baseline shifts absolute arrival time without changing the winner.

Least action, not arbitrary stationary action. This machine returns a global minimum of the discretized objective it is programmed with. A physical Lagrangian trajectory can be a stationary point rather than a minimum; that broader class requires a different construction.

Known mathematics

Fermat, Bellman optimality and the eikonal equation are mathematically connected. That connection itself is not the novelty.

Known hardware idea

Race logic already encodes values as arrival time and uses temporal primitives for graph and dynamic-programming workloads.

Proposed contribution

Compile local action into programmable photonic delay, preserve cumulative action in time, and prune with a first-arrival photonic state element.

02 / The machine in one line

Delay + race + latch

01 / encodeLocal action
aᵢⱼ
02 / propagateTrue time delay
τᵢⱼ = αaᵢⱼ
03 / compareFirst arrival
minᵢ(Tᵢ+τᵢⱼ)
04 / rememberLatch winner
pred[j] = i*

The critical device is not a passive beam combiner. A passive junction preserves late pulses. We need an active first-arrival cell: the first valid pulse fires the output after a fixed latency, records the winning input, and suppresses later arrivals until reset.

Tₖ₊₁(j) = g + minᵢ [ Tₖ(i) + τₖ(i,j) ]
predₖ₊₁(j) = argminᵢ [ Tₖ(i) + τₖ(i,j) ]

Timestamp must survive

The cell may add a fixed gate latency g, but must not re-clock the winner to a new global time origin. Physical time is the accumulated action.

State must be complete

If future cost depends on velocity or momentum, position alone is insufficient. A node may need to represent phase-space state such as (q,p).

03 / The reason exponential histories can disappear

Why early pruning does not change the optimum

Suppose two different histories reach exactly the same Markov state j. History A arrives first:

Tᴬ(j) < Tᴮ(j)

Every future continuation from that complete state is identical for the two histories. For any future action F(j):

Tᴬ(j) + F(j) < Tᴮ(j) + F(j)

The late history can never recover. The earliest arrival is therefore a sufficient statistic for every history that collapses into the same state.

Exponential paths histories can collapse into a polynomial frontier — but only when the state representation makes Bellman’s principle valid.
04 / Interactive digital twin

Make it fail before building it

The simulator builds random layered action graphs. “Oracle” is exact dynamic programming. “Photonic” propagates physical arrival times, prunes every state to its earliest pulse, and optionally adds edge and gate timing noise.

Experiment A — first-arrival equivalence

Random continuous edge actions; the ideal model almost surely has a unique winner.

LIVE JS TWIN
6
10
500 ps
0.5 ps
Virtual trajectories
Oracle action
Photonic regret
Exact path
Oracle path Photonic first-arrival path overlap appears as one path
What the zero-noise result means. It does not show a speedup. It verifies a more basic architectural invariant: first-arrival pruning is mathematically equivalent to Bellman pruning for this workload class.

Experiment B — noise boundary

Sweep timing error and watch exact-path accuracy fall as physical arrival ordering becomes ambiguous.

MONTE CARLO

Model: every programmed edge delay and every node forwarding latency receives independent Gaussian timing error. Static calibration error and dynamic jitter should eventually be separated experimentally.

05 / The exponential number is real — and easy to misuse

Exponential virtual space, polynomial active frontier

If each of d decisions has b possible states, the number of complete histories is b^d. With first-arrival pruning, only one winner per state needs to continue.

Scaling explorer

Compare implicit trajectory count, physical fabric, dynamic-programming work and propagation depth.

RESOURCE ACCOUNTING
16
30
15 ps
50 ps
Implicit histories b^d
Programmable edges
Active frontier
Estimated optical latency
Do not call this an exponential complexity speedup. The same optimal-substructure that allows photonic pruning also allows classical dynamic programming. The opportunity is to spatialize the min-plus reductions so wall-clock time follows physical propagation depth — not to claim that P=NP has been bypassed.
06 / A claim the community can actually test

Advantage should mean crossover, not rhetoric

The fair comparison is not against exhaustive enumeration. It is against the best optimized electronic algorithm for the same graph, accuracy target and I/O assumptions.

TTSphotonic(accuracy, energy, I/O) < TTSbest-electronic(same accuracy)
and / or
EDPphotonic < EDPbest-electronic
QuantityPhotonic architectureFair classical baselineWhat to measure
Candidate historiesb^d implicitb^d implicitDo not use as a speed metric alone
Work representationO(d·b²) spatial delays / gatesO(d·b²) min-plus opsArea, utilization, memory traffic
Critical depth≈ O(d) propagationO(d log b) ideal parallel reductionMeasured end-to-end latency
Precisionarrival-time margin / jitterdigital numeric precisionExact path + regret
Energylaser + gates + programming + detectioncompute + SRAM/HBM + controlJoules per solved instance
07 / A build that can fail honestly

From fiber bench to integrated fabric

0 / Fiber digital-physical twin

3 states × 3–8 stages. Fixed fiber delays implement transition cost; switched delay segments implement programmable potentials. Validate action↔arrival calibration and the noise model.

1 / First-arrival cell

Characterize one active node: set/reset behavior, latency distribution, late-pulse extinction, fan-out and predecessor recording.

2 / Integrated lattice

Combine programmable true-time-delay edges with first-arrival cells. Scale width until electronic min-plus reduction becomes the meaningful adversary.

First-arrival cell: what must actually be measured

ParameterWhy it mattersInitial target
Forwarding latency gAdds directly to every stage<50 ps aspirational; characterize achievable value first
Latency jitter σgCan reorder near-tied candidatesMuch smaller than the minimum timing margin
Late-pulse extinctionDetermines whether the exponential pulse population is actually pruned>20 dB proof-of-principle target
Reset / refractory timeLimits instance throughput, not single-instance latencyMeasure separately
Winner-port memoryAllows full trajectory reconstructionOne predecessor symbol per state
The hard device problem is the latch, not the delay.Programmable photonic delay is an established direction. The distinctive challenge is an ultrafast state element that forwards one winner, suppresses later pulses, preserves temporal ordering, and resets efficiently.
08 / How we should try to kill the idea

The experiments that matter most are adversarial

Near-tie test

Sweep ΔA until αΔA approaches total timing uncertainty. The measured failure curve should match the digital twin.

State-aliasing test

Merge histories that share position but hide different velocity. If pruning fails, show that a phase-space state restores correctness.

Loss adversary

Attenuate the true winner without changing delay. It should still win until detector threshold makes it disappear.

Electronic crossover

Benchmark the same Bellman update on CPU, GPU and ideally FPGA/ASIC-like hardware with symmetric I/O accounting.

If it only beats brute force, we learned little. If it beats optimized min-plus hardware at equal accuracy with full resource accounting, we have a result.
Selected prior art

The work that defines the boundary