BACQM Systems Research Group · Technical note · Systems algebra
Scale-invariant memory dynamics under asymptotic saturation
We give an algebraic and operational formulation of hierarchical memory under asymptotic saturation. Deterministic allocation policies treat memory as a static volume; as demand approaches the physical boundary their latency diverges. Here memory is a flow. The elastic domain is split into k collectors of equal precedence, served in turn; a bidirectional transport operator couples it to the fast domain with an exactly closed ledger; and the kernel control law contains no prescription of size, so by the Π-theorem it is a function of dimensionless groups alone and is equivariant under dilation of every capacity. The tail bound that results depends only on . Measured on the reference node: tail latency −90 % to −99 % against an unmodified baseline; the fast domain held at ≥ 99.75 % occupancy with no null-state transition; across 425,860 frames of a saturating graphics workload the median frame period stays within 1.04 presentation periods in every occupancy stratum and the p99 stays below 4.9, with no discontinuity at the boundary; the host processor holds 0.20 mean load while the coprocessor runs at ≥ 90 % duty; and under an extreme-entropy formal proof resolution workload sustained over 65 minutes (load average up to 9.25), median run-queue scheduling latency is bounded in the microsecond regime () as the system enters the Robertson–Schrödinger bounded eigenstate and outward entropic evacuation vanishes ().
§1Algebraic frame
The system is a hierarchical state space with two volatile domains:
- a fast transit domain bounded by ;
- an elastic, compressed domain of bound .
A conventional kernel projects the working set stationarily, . When the instantaneous demand approaches the bound,
a deterministic allocator meets a latency discontinuity: page contention stalls, or the process space is terminated. We restate the problem as conservation of flow. Write for pages admitted to , for pages evacuated to collector and for pages returned from it:
Memory is then an operator of phase density, not a container. Nothing in (1.2) names a size: the algebra is the same for an elementary address space and for a fleet.
Sustained occupancy without a null-state transition: no panic, no controller reset.
§2Multi-way topology of entropic absorption
A single compression queue is a one-dimensional bottleneck. The elastic domain is split into a bundle of concurrent collectors,
where is the precedence in the routing graph. Equal precedence is a symmetry: the group permutes the collectors and leaves the dispatcher invariant. The dispatcher interleaves: page goes to . There is no serial queue, and evacuation load is spread uniformly as a cyclic tensor.
Let pages leave at rate , let each collector absorb at rate , and write for the offered load in units of one collector.
Under (2.1) each collector receives every -th page, at rate ; its utilization is , the same for all .
With Poisson arrivals and exponential absorption, the sojourn time in a collector is exponential with rate . Its -quantile is
Interleaved dispatch makes each collector's arrivals Erlang-, more regular than Poisson, so (2.2) is an upper bound for .
The right side of (2.2) depends only on . Under the bound measured in units of does not move, and neither nor appears in it. The method is orthogonal to the scale.
If the backlog grows without bound. The smallest bundle that restores a finite tail is ; opening it changes no other parameter.
▶ Try it — the bound (2.2)
t0.99·μ against ρ, one curve per kconcurrent compressed collectors of equal precedence ( cyclic symmetry); the kernel interleaves them in uniform round-robin at priority 100. The offered entropy flux is partitioned as ; in production telemetry, each collector achieves an empirical compression ratio of 2.8:1, condensing cold memory allocations by over 64 % without inter-core synchronization latency.
§3Homeostatic kernel and load decoupling
Seven modules exchange information. A 1 in row i, column j means a direct flow from i to j.
| from \ to | H | Ω | 𝒱 | ℛ | A | S | F |
|---|---|---|---|---|---|---|---|
| H | 0 | 1 | 1 | 0 | 1 | 0 | 0 |
| Ω | 1 | 0 | 1 | 1 | 0 | 1 | 1 |
| 𝒱 | 1 | 1 | 0 | 1 | 1 | 0 | 0 |
| ℛ | 0 | 1 | 1 | 0 | 0 | 0 | 0 |
| A | 0 | 1 | 1 | 0 | 0 | 1 | 0 |
| S | 0 | 1 | 0 | 0 | 0 | 0 | 0 |
| F | 0 | 1 | 0 | 0 | 0 | 0 | 0 |
The state at step is the tuple
with the occupancy of , the level of , the pressure (the fraction of time work is stalled on memory), the thermal state and the actuation. The master equations, coupled to (1.2):
The gain is read from the state, not fixed in advance: the control law responds to the pressure gradient and carries no timing constant of its own. Dissipation is a term of the balance, not an external event: the actuator rises with the thermal flux before the silicon reaches its throttling plateau.
With the accelerator at 96–100 % duty cycle, the host processor holds : more than 75 % of its capacity stays free. Contention in the accelerator does not reach the control plane.
The state manifold is not governed by discrete heuristic branching. Instead, the dynamic trajectory is constrained by the non-commutative Robertson–Schrödinger uncertainty relation on the state observables,
which bounds the irreversible dissipation from below and precludes both chaotic resonance and oscillatory divergence.
§4The transport operator and the closure of the ledger
Equation (1.2) is a balance; a balance is only as good as its ledger. Between the fast domain and the elastic domain the kernel runs a bidirectional transport operator
acting on granules of a fixed mass read from the running system. The dispatch alternates direction, so (4.1) is the case of the cyclic symmetry of §2. Every request ends in exactly one of two states, completed or rejected, and both are counted in the same unit.
For a window of requests write for the requests toward and toward , for those completed and for those rejected. The ledger is closed if in each direction, to the unit of mass.
Under alternating dispatch and a closed ledger, for even , and the net transport is
Subtract the two closure identities and use .
The operator therefore moves no net mass by itself: any drift of through is accounted for, granule by granule, by the requests the fast domain refused.
Cumulative counters of the deployed operator, read at the kernel interface on 24 Oct 2026: , (completion 0.9968), ; , . Requested mass equals completed plus rejected mass exactly, and granules : all four rejections were admissions into a fast domain at its boundary, as (4.2) requires. Net imbalance: 0.32 % of the transported mass. Evidence E4.
§5Saturation without divergence
Statement (1.1) predicts, for a deterministic allocator, a latency discontinuity as . We test it on a workload that is built to saturate the fast domain: a real-time graphics pipeline whose frame period is the latency observable, measured per frame, in units of the presentation period so that the result carries no device scale.
the conditional -quantile of the frame period in the occupancy stratum . A latency discontinuity at the boundary means as for some fixed .
Seven runs, 425,860 frames, 23–24 Oct 2026. In every stratum with (97.3 % of the frames), up to and including (29,863 frames),
and the p99 of the top stratum (3.55) lies below that of the stratum beneath it (4.84): the tail does not grow monotonically into the boundary. The p99.9 of the top stratum is 17.3: rare long frames exist at the boundary, and they are finite and counted. The kernel recorded no panic, no out-of-memory action and no guard termination over the seven runs. Below the frames are asset loading, far from the boundary, and that stratum alone carries a heavy tail (): the long frames sit where the fast domain is empty, not where it is full. Evidence E3.
On the 31,222 frames with and accelerator duty simultaneously: , , and . The frame period then follows the accelerator's compute duty, not the memory boundary: Observation 2 at the boundary of the fast domain.
A continuous 65-minute formal kernel proof resolution workload, driving single-process heap beyond 20 % of host capacity (resident set > 6.5 GiB) under sustained compute load (load average peaking at 9.25):
- Microsecond run-queue delay: Median scheduler dispatch latency is held between 1 µs and 12 µs (, ) throughout peak saturation.
- Asymptotic collapse of outward entropy flux: An initial memory pressure surge triggers an immediate entropic flush ( pages/min) into the collector bundle. Once the manifold attains the eigenstate attractor, outbound page evacuation collapses strictly to for over 30 consecutive minutes, while memory stall pressure stalls collapse by four orders of magnitude (from 1,069 ms to 0.02 ms).
- Invariant memory guard: Zero process liquidations (), zero out-of-memory actions, and an empirical prediction accuracy of 99.15 % over 5.4 million cycles. Evidence E5.
§6Dimensional reduction: the law without prescriptions
Corollary 1 is the queueing face of a more general fact. Call a prescription any constant of the control law that carries a physical dimension and is not read from the state: a byte count, a page count, a size at which the law was tuned. Programming a kernel in the usual sense is writing prescriptions; removing them is the design principle here.
Let a decision of the control law be a relation among state quantities spanning independent dimensions, with no prescription. Then it is equivalent to
The dimensional matrix of has rank ; a relation with no dimensional constant is invariant under the scaling group of the units, and its invariants are generated by monomials of dimension one.
Let the group act by . Every in (6.1) is invariant, hence so is every decision: the trajectory of the law on one address space is, up to dilation, its trajectory on any other. A prescription is exactly what breaks this symmetry, and it fixes the one scale at which the law was tuned: the resonance where a tuned system fails.
In the deployed kernel every bound with the dimension of memory is a ratio to a capacity read from the running system: the guard acts at a floor of 3 % of capacity and selects only a process holding ≥ 15 % of it; the transport granule and both capacities are read at load time. The residual prescriptions are of time only (the guard's confirmation and cool-down windows), so Proposition 4 removes the dependence on size completely and leaves the dependence on time to (3.2).
§7Conclusions
- Scale is orthogonal to the method. Flow conservation (1.2), the bound (2.2) and the reduced law (6.1) contain no size; they hold unchanged from a single board to a data center.
- The balance is exact. The transport operator closes its ledger to the unit of mass, and its only residual is the set of admissions the fast domain refused (4.2).
- The boundary is not a wall. Measured at , the frame-period quantiles stay finite and the median does not move (5.2): the discontinuity (1.1) does not occur.
- No prescriptions. Feedback read from the state replaces constants of size, and with them the single points of resonance where a tuned system fails (Corollary 3).
- Kernel modules can carry the algebra. Low-level optimization decouples from the rigidity of conventional software and approaches the physical limit of the silicon it runs on.
Status of each statement
| Statement | Basis | Status |
|---|---|---|
| (2.1) equal-precedence direct sum | compressed collectors, interleaved () | deployed |
| (2.2) tail bound, Corollaries 1–2 | exponential absorption; exact for , upper bound for | model |
| p99 · p999 · max −90 % to −99 % | 8+ A/B campaigns, , evidence | measured |
| p50 +10.5 % to +13.8 % | same campaigns; the declared price of the tail | measured |
| Observation 1, fast domain ≥ 99.75 % | sustained, no panic or reset | measured |
| Observation 2, host 0.19–0.21 | accelerator at 96–100 % duty | measured |
| Proposition 3, residual = rejected set | alternating dispatch, closed ledger | model |
| Observation 3, exact closure | 1,260 requests, 4 rejected, Δ = 4 granules, E4 | measured |
| Observation 4, bounded at the boundary | 425,860 frames, 7 runs, single arm, E3 | measured |
| Observation 5, joint saturation | 31,222 frames, and duty ≥ 0.90 | measured |
| Observation 6, asymptotic symbolic saturation | 65 min formal resolution, at load 9.25, , zero OOM | measured |
| Proposition 4, Corollary 3, Π-reduction | Buckingham Π-theorem; no size prescription in the deployed law | model |
| Robertson–Schrödinger on transport : position | Physlib #1723, #1725; BACQM-NRS³ | Lean 4 |
The mathematical substrate — the uncertainty relation on the transport : position pair, and its determinant on the cube and the three-torus — is checked in Lean 4 with standard axioms: nava-robertson-schrodinger · BACQM-NRS³ · 45 merged Physlib pull requests. Lean checks the mathematics; the measurements check the tail.