Universal computation in smooth forced Navier–Stokes flows

Source and verification

This article explains family 376 of the October 6, 2026 OpenAI collection. Research conclusions are attributed to its manuscripts. The review covers statements, introductions, and selected proof mechanisms; it is not an independent verification of the complete proofs.

Overview

Family 376 constructs externally forced viscous flows whose trajectories or velocity fields encode finite instructions. Its nine manuscripts study different geometric realizations, clocks, and observation rules. The central question is whether a fixed observation occurs eventually; the constructed solutions themselves remain smooth.

🏷️ Preliminaries: forcing and material trajectories

The incompressible Navier–Stokes equations are

$$\begin{array}{c}{\partial }_{t}u+(u\cdot \nabla )u-\nu \Delta u+\nabla p=f,\\ \nabla \cdot u=0,\end{array}$$

with fixed viscosity $\nu >0$. A material particle follows $X^{\prime }(t)=u(t,X(t))$. Smooth divergence-free velocity produces an invertible, volume-preserving flow on every finite interval.

The inverse-design observation is elementary: choose a smooth divergence-free velocity starting from zero, set $p=0$, and define

$$f={\partial }_{t}u+(u\cdot \nabla )u-\nu \Delta u.$$

Then the equation holds identically. This alone says nothing about computability, support, time decay, periodicity, or a particle’s observation behavior. Those additional requirements are the content of the construction. A divergence-free velocity also does not automatically make this residual force divergence-free; the solenoidal-force variants require extra structure.

🏷️ The family claim and its quantifiers

Common computational conclusion of family 376

At a fixed positive computable viscosity, a terminating procedure takes a finite machine and input and constructs a finite effective description of a smooth force and a smooth solution starting from rest. A designated particle enters a fixed open detector if and only if the machine halts. The manuscripts give several versions with different domains and force properties. (OpenAI, 2026c; OpenAI, 2026a; OpenAI, 2026h; OpenAI, 2026d; OpenAI, 2026f; OpenAI, 2026b; OpenAI, 2026g; OpenAI, 2026e)

The compiler must finish without first determining whether the machine halts. It specifies the finite instruction table geometrically; it does not run the computation and replay its outcome in the forcing.

A companion instead detects halting through the velocity field: a vertical component exceeds a threshold in a fixed strip on the torus, or an integrated vertical signal exceeds a threshold on a cylinder. (OpenAI, 2026i) These are different observations from following one labeled particle.

🏷️ Why memory and volume compensation are necessary

A machine instruction can erase information, but a smooth flow map is invertible. The constructions therefore retain history in additional coded coordinates. Distinct predecessor states remain distinguishable after an update.

Positional encodings turn finite prefix changes into affine maps on rectangles. The motion must implement an instruction on an entire rectangle or solid box, not just on the countable set of states visited by one run. Source rectangles are separately disjoint and target rectangles are separately disjoint, though a source may overlap another target.

If a planar instruction has scale factors $a,b>0$, the three-dimensional diagonal map with factors

$$a,b,(ab{)}^{-1}$$

has determinant one. This gives the simplest explanation of normal compensation: planar compression can be balanced by expansion in the third direction. Realizing that map by a localized smooth flow still requires routing, separation, and control of the intermediate motion.

In particular, correctness at integer instruction times is insufficient. A nonhalting transition might accidentally cross the detector between those times. The manuscripts include path bounds and observation guards to exclude such false detections.

🏷️ How the nine manuscripts fit together

Manuscript theme	Distinct contribution
Finite instructions and shear flows	Reciprocal rectangle maps, history recording, and excursion bounds. (OpenAI, 2026c)
Fixed particle test	Ordinary instructions with longer time steps controlling increasingly fine spatial gates. (OpenAI, 2026a)
Eventually stationary forcing	A spatial clock produces a force constant in time after startup. (OpenAI, 2026h)
Geometric programs	Solenoidal mean-zero periodic forcing, coding sheets, and normal compensation. (OpenAI, 2026d)
Prefix instructions	Full-domain injective history processors for area-changing symbolic updates. (OpenAI, 2026f)
Rapidly vanishing forcing	All mixed derivatives decay faster than every inverse power, with a shrinking detection signal. (OpenAI, 2026b)
Velocity-field detection	Advection–diffusion turns computation into pointwise or integrated velocity observations. (OpenAI, 2026i)
Scalar potentials and slow clocks	Compact spatial support and time-decay estimates through an unbounded logarithmic clock. (OpenAI, 2026g)
Solid-box transport	Determinant-one maps on neighborhoods of full boxes, using storage and obstacle detours. (OpenAI, 2026e)

The different force properties belong to different constructions. Eventual stationarity, periodicity, and rapid decay should not be combined into a single theorem about one force.

🏷️ Clocks and solution identification

A time change explains one decay mechanism. If $V(s,x)$ is a prescribed velocity and $g(t)$ is increasing, then $u(t,x)=g^{\prime }(t)V(g(t),x)$ follows the same trajectories under the new clock. Taking $g(t)=\log (1+t)$ slows the velocity while allowing arbitrarily many instruction times, since $g(t)\to \infty$. Derivative and force estimates require differentiating the full residual, not just this first factor.

Finally, an energy comparison identifies the constructed flow within its stated smooth solution class. If another solution has the same force and initial data, their difference obeys an energy inequality controlled by the constructed velocity’s gradient. Grönwall’s inequality gives equality when the required bounds and integrations hold.

This is a construction of special forced solutions and a reduction of an eventual observation problem. It is not a theorem of global regularity for arbitrary Navier–Stokes data, nor a claim of blowup, nor a fixed-resolution robust detector. The rapidly decaying construction in particular uses a shrinking signal.

🏷️ Key steps and proof outlines

A reversible recorder supplies separated rule domains

A smooth flow is invertible at every finite time, so an irreversible instruction cannot simply erase its input. The constructions record enough information to recognize the incoming rule from its output, including on complete symbolic cylinders with arbitrary tails. This is stronger than checking reversibility along the initialized computation.

The fixed-particle manuscript uses an explicit two-pulse alternative. With $J$ branches, put $\Lambda =J+1$ and record branch $j_n$ by adding $j_n{\Lambda }^{-n-1}$ to a history coordinate. Before this addition, ${\Lambda }^n{z}_n$ is an integer. Afterward,

$${\Lambda }^n{z}_{n+1}=\frac{j_n}{\Lambda }(\mathrm{mod}1).$$

A smooth periodic gate near $j_n/\Lambda$ selects the planar motion. The recording pulse freezes the planar coordinates; the motion pulse freezes the history coordinate. The force formula includes every branch, while the particle’s position selects the active one.

Differentiating this selector $k$ times costs ${\Lambda }^{nk}$. Stage lengths $S_n=2^{n^2}$ yield

$$\Vert {\partial }_t^h{D}_x^{\alpha }U{\Vert }_{\infty }\le C_{h,\alpha }{S}_n^{-1-h}{\Lambda }^{n{\alpha }_z}.$$

Its squared time integral is bounded by a convergent Gaussian series. Increasingly fine memory and bounded derivatives coexist because physical stages become very long. No force evaluation needs to predict the executed branch sequence. (OpenAI, 2026a)

Exact shear paths control the detector between checkpoints

For reciprocal scaling $(r,s)\mapsto (\lambda r,s/\lambda )$, define $X(a)(r,s)=(r+as,s)$ and $Y(a)(r,s)=(r,s+ar)$. The chronological sequence

$$Y(-\lambda ),X({\lambda }^{-1}-1),Y(1),X(\lambda -1)$$

has exactly the required product. If initial and final coordinate offsets are at most $h$, every intermediate coordinate is at most $2h$. For example, $|({\lambda }^{-1}-1)s|\le \max \{|s|,|s|/\lambda \}\le h$. Partial shears trace segments between consecutive endpoints, so this also controls every intermediate time.

Lift each rectangle to a private height, execute four shears, translate, and lower: seven stages. Height masks isolate branches during horizontal motion. Source and target masks are used at different times, so the two families may overlap. The excursion estimate proves that trajectories stay inside cutoff plateaus and avoid the detector during nonhalting transitions.

Each stage is a transverse shear with zero self-advection. Disjoint time supports imply $(U\cdot \nabla )U=0$ for the complete template, so $f=U_t-\nu \Delta U$ is already solenoidal. The geometric-program variants retain complete-path guards while changing loading, history placement, parking, and observation conventions. (OpenAI, 2026c; OpenAI, 2026d)

Prefix edits require compensation in the third direction

A prefix replacement can scale the two planar coordinates by independent factors ${\lambda }_1,{\lambda }_2$. For a solid box, choose ${\lambda }_3=({\lambda }_1{\lambda }_2{)}^{-1}$. Then $D(s)=\mathrm{diag}({\lambda }_1^s,{\lambda }_2^s,{\lambda }_3^s)$ has determinant one throughout, and every half-width stays between its endpoint values. Private heights separate full boxes during deformation; the output thickness generally differs from the input thickness. (OpenAI, 2026f)

For a moving box $c(t)+D(t)K$, put $L=\dot{D}{D}^{-1}$, so $\mathrm{tr}L=0$. The localized vector potential

$$\chi (x)\left\{\frac{1}{2}\dot{c}\times (x-c)+\frac{1}{3}[L(x-c)]\times (x-c)\right\}$$

has curl equal to $\dot{c}+L(x-c)$ wherever $\chi =1$. This realizes the affine motion on a neighborhood, not just on the central code. The box-transport manuscript ensures separation by evacuation and parking: remove every obstructing source before delivering any target. ODE uniqueness identifies the prescribed affine paths with the localized flow. (OpenAI, 2026e)

For an invariant coding plane, another compensation suffices. Given a planar field $v(Y)$, choose $\eta (z)=z-z_0$ near $z_0$ and set

$$W(Y,z)=({\eta }^{\prime }(z)v(Y),-\eta (z){\mathrm{div}}_{Y}v(Y)).$$

The two divergence terms cancel; on $z=z_0$, the horizontal component is $v$ and the normal component vanishes. Thus arbitrary planar area changes are compatible with three-dimensional incompressibility. The scalar-potentials manuscript uses this alongside compact scalar-potential realizations. (OpenAI, 2026g)

Clocks must cover every finite logical time

For a template $W(s,x)$, put $u(t,x)=g^{\prime }(t)W(g(t),x)$. ODE uniqueness gives $X_u(t,a)=X_W(g(t),a)$. The event equivalence requires $g$ to be onto $[0,\infty )$.

For $g(t)=\log (1+t)$, let $a=(1+t{)}^{-1}$. The identity

$${\partial }_t[a^{r}A(g(t),x)]=a^{r+1}({\partial }_s-r)A(g(t),x)$$

gives derivative decay at every fixed order. An integrable speed would traverse only finite logical time. The eventually stationary construction instead places the clock in a spatial coordinate and turns on a fixed autonomous processor. A scalar startup speed that becomes one makes its residual force eventually stationary; this is a separate prescription from the decaying-clock variant. (OpenAI, 2026d; OpenAI, 2026h)

The rapidly vanishing family shrinks the spatial motions themselves. In its addressed-curl construction, disjoint local supports give derivative bounds $C{b}_n{\delta }_n^{-|\alpha |}$ without a factor counting addresses. Choosing $b_n\le {\delta }_n^{n+10}$ and ${\delta }_n\le 2^{-n}$ absorbs every fixed derivative order, including the two derivatives in viscosity. This differs from slowing a fixed trajectory. The alternating and lattice variants implement related memory separation with different encodings. (OpenAI, 2026b)

A velocity detector also needs a diffusion estimate

The velocity-observation manuscript builds a scalar advection–diffusion component. In block $n$, a short injection creates a bump, and a short stirring burst performs $n$ logical periods. Let $w^0$ be the nondiffusive transported bump. Its difference from the new diffusive component has source $\Delta w^0$, so the maximum principle gives

$$\Vert w^{\mathrm{new}}-w^0{\Vert }_{\infty }\le 2{\delta }_n{D}_n<1/16,$$

where $D_n$ bounds the transported bump’s Hessian. Forward and inverse flow derivative estimates determine $D_n$ before the computation runs.

At a halting visit the reference bump has value one in the detector. On a nonhalting trajectory its whole support stays a definite distance away. Heat-kernel bounds and summable total injected mass control old bumps during later waits. Together these estimates produce the fixed threshold gap above and below $1/2$ for all times. Merely encoding a particle path would not establish this velocity-level observation. (OpenAI, 2026i)

See Also

• on finite-time blowup for forced Navier-Stokes studies a separate forced singularity claim; smoothness of the prescribed residual is crucial in both settings.
• Family 368 also requires geometric constructions to preserve injectivity on whole neighborhoods.

🏷️ Manuscripts and verification scope

• Finite Instructions and Solenoidal Shear Flows — pinned manuscript and companion source.
• A Fixed Particle Test for Computation in a Forced Viscous Flow — pinned manuscript and companion source.
• Universal Computation with Eventually Stationary Navier–Stokes Forcing — pinned manuscript and companion source.
• Geometric Programs for Solenoidal Forcing — pinned manuscript and companion source.
• Prefix Instructions and Incompressible Flows — pinned manuscript and companion source.
• Computation under Rapidly Vanishing Navier–Stokes Forcing — pinned manuscript and companion source.
• Velocity-Field Detection of Computation in Forced Navier–Stokes Flows — pinned manuscript and companion source.
• Scalar Potentials and Slow Clocks for Forced Fluid Computation — pinned manuscript and companion source.
• Incompressible Box Transport and Finite Computation — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The scope page documents selected statements from four companions: coding-sheet realization, rapidly decaying alternating-coordinate memory, velocity-field detection, and solid-box transport with a balanced recorder. Its coding-sheet statement does not cover that companion’s particle-halting detector or solid-box variant. The other selected statements carry their own support, observation, and solution-comparison hypotheses; the page is not blanket verification of all nine manuscripts. No Lean build, axiom audit, or manuscript-to-formal-statement equivalence check was performed for these notes.

OpenAI collection provides prerequisites and the reading sequence; coverage register records completed and pending families.

🐻  OpenAI 2026a. A Fixed Particle Test for Computation in a Forced Viscous Flow.

🐻  OpenAI 2026b. Computation under Rapidly Vanishing Navier–Stokes Forcing.

🐻  OpenAI 2026c. Finite Instructions and Solenoidal Shear Flows.

🐻  OpenAI 2026d. Geometric Programs for Solenoidal Forcing.

🐻  OpenAI 2026e. Incompressible Box Transport and Finite Computation.

🐻  OpenAI 2026f. Prefix Instructions and Incompressible Flows.

🐻  OpenAI 2026g. Scalar Potentials and Slow Clocks for Forced Fluid Computation.

🐻  OpenAI 2026h. Universal Computation with Eventually Stationary Navier–Stokes Forcing.

🐻  OpenAI 2026i. Velocity-Field Detection of Computation in Forced Navier–Stokes Flows.