Iyeumaring-AION-LATCH-VI / reports /Compact_Head_Theory.md
PureOne's picture
Upload verified Iyeumaring AION-LATCH VI research dataset
792dec0 verified
|
Raw History Blame Contribute Delete
9.77 kB

A certified decoder bridge from finite scalar data to compressed neural heads

Scope

The scalar optimal filter supplies a universal regularity/retention operation, but its oracle form assumes access to a residual function. Finite Lipschitz envelopes already address that access issue. This continuation instead studies certificate-preserving compression and a data-adaptive decoder whose guarantees hold on an entire interval, rather than only on sampled queries.

The setting is deliberately restricted: one real input, one scalar output, a known piecewise-affine nonnegative guard, and bounded/Lipschitz heads. Piecewise-linear ReLU interpolation, linear programming, Lipschitz envelopes, and metric-entropy arguments are established methods. The proposed contribution is their explicit coupling to the retention contract, including exact continuous-domain constraints, separation, a compression guarantee, and an entropy obstruction. No worldwide priority or general RSI claim is established.

1. The finite-data information contract

Work on (I=[-1,1]), with protected set (K={0}), guard (g(x)=|x|), and scalar residuals (r) satisfying (r(0)=0) and (\operatorname{Lip}(r)\le1). Given consistent exact samples ((x_i,y_i)), the established envelopes are

[ \ell(x)=\max{-g(x),\max_i[y_i-|x-x_i|]},\quad u(x)=\min{g(x),\min_i[y_i+|x-x_i|]}. ]

Let (m=(\ell+u)/2), (w=(u-\ell)/2). Both envelopes themselves belong to the sample-consistent class. Therefore the exact conditional worst-case error of any fixed decoder (f) is

[ \mathcal R_S(f)=\sup_{r\text{ consistent}}|r-f|\infty =\sup{x\in I}\max{u(x)-f(x),f(x)-\ell(x)} =\sup_x[w(x)+|m(x)-f(x)|]. ]

This certificate concerns every sample-consistent 1-Lipschitz target, not an assumed target formula. It can favor a different prediction from the best prediction on a particular smooth ground truth.

2. Exact whole-interval constraints for a compressed ReLU head

Take knots (-1=z_0<\cdots<z_{M-1}=1), including every guard kink. Let (h) be the continuous piecewise-linear interpolant of nodal values (v_i), and let (f=g h). On interval ([z_i,z_{i+1}]), write

[ s_i=(v_{i+1}-v_i)/(z_{i+1}-z_i),\quad g(x)=g_i+\sigma_i(x-z_i). ]

For a desired head Lipschitz cap (H), the following linear inequalities in the nodal values are necessary and sufficient for the whole-interval contract:

[ |v_i|\le1,\quad |s_i|\le H, ]

[ |\sigma_i v_i+g_i s_i|\le1,\quad |\sigma_i v_{i+1}+g_{i+1}s_i|\le1. ]

Proof. A linear head reaches its maximum absolute value at an endpoint, and its derivative is (s_i). The output derivative is

(f'(x)=\sigma_i h(x)+g(x)s_i),

an affine function on that interval. Its maximum absolute value is attained at an endpoint, where it equals the last two expressions. Continuity and integration give the global Lipschitz bounds; necessity follows from the same local derivatives. The exact gate supplies (f=0) wherever (g=0). ∎

Thus the certificate is not an interpolation of finite Lipschitz tests. It explicitly controls all points between the knots. For a distance to finitely many protected points, include both protected locations and distance-switching points as guard breakpoints; the argument is unchanged.

The head has an exact ReLU representation

[ h(x)=v_0+s_0(x-z_0)+\sum_{j=1}^{M-2}(s_j-s_{j-1})[x-z_j]_+. ]

It needs (M-2) hinges plus an affine term. A segment-index decoder evaluates the same function in (O(\log M)) lookup time and (O(1)) local arithmetic, using (O(M)) stored coefficients. Direct dense ReLU evaluation instead costs (O(M)). The certificate uses actual segment slopes and gate derivatives; generic sums of absolute network weights can be much more conservative.

3. Data-adaptive minimax fitting with a continuous certificate

Minimize (E) over the nodal values, under the preceding contract and the semi-infinite linear inequalities

[ g(x)h(x)-\ell(x)\le E,\quad u(x)-g(x)h(x)\le E\qquad(x\in I). ]

The head is linear in its coefficients at every fixed (x), so any finite subset of these constraints is an ordinary LP. On each interval formed by the head knots, guard kinks, and envelope switches, (g h) is quadratic and the envelopes are affine. Consequently the exact continuous separation oracle requires only endpoints and any interior stationary point of each risk quadratic.

Proposition. If a finite-cut LP has optimum (E_-), and its head has oracle risk (E_+), then the true optimum within this compressed head class lies in ([E_-,E_+]). If (E_+-E_-\le\varepsilon), the stored globally feasible decoder is (\varepsilon)-optimal within that class.

Proof. Removing risk constraints can only lower the optimum, so (E_-) is a lower bound. The head already satisfies the exact whole-interval contract, and the oracle evaluates its risk on every continuous subinterval, so (E_+) is a feasible upper bound. ∎

The implementation adds the maximally violated risk point and repeats. No claim of a fixed iteration count or general strong-polynomial bound is needed. The observed runs terminate well below the declared tolerance.

For a numerical LP dual audit, project inequality multipliers to (\lambda\le0), then scale so the error-variable weight (A_E^T\lambda\le1). Every nodal coefficient, constant-head coefficient, and affine intercept/slope in the experiment has absolute value at most one. Thus the corrected bound

[ E_-\ge b^T\lambda-|A_\theta^T\lambda|_1 ]

is valid for the finite LP: multiply primal inequalities by (\lambda), use nonnegativity of (E), and bound the coefficient-stationarity residual by the unit coefficient box. The certificate identity is exact in real arithmetic. The implementation reports floating residuals and does not replace outward interval arithmetic or a formal LP proof checker.

4. A constructive guarantee when compressing an existing feasible head

Suppose (|h_*|\le1), (\operatorname{Lip}(h_*)\le H), and (f_*=g h_*) is 1-Lipschitz. Let the mesh width be (\Delta), include all affine guard breakpoints, assume (|g'|\le1), and write (R=\sup_I g). Let (Ih_*) interpolate the sampled head and put

[ \beta=(1+H\Delta)^{-1},\qquad \widetilde h=\beta Ih_*. ]

Theorem. The compressed decoder retains exact gate preservation, head norm at most one, head Lipschitz cap (H), and output Lipschitz cap one. Moreover,

[ |g\widetilde h-f_*|_\infty \le R,\frac{3H\Delta}{2(1+H\Delta)}. ]

Proof. Interpolation preserves the head norm and its Lipschitz cap. On a mesh interval, the output secant slope has magnitude at most one because its endpoints equal (f_). Its two derivative endpoints differ from that secant by (\pm\sigma_i s_i\Delta_i), so the interpolated output is at most ((1+H\Delta))-Lipschitz. Scaling restores the output contract. The standard Lipschitz interpolation estimate is (|Ih_-h_*|_\infty\le H\Delta/2). Therefore

(|\beta Ih_*-h_*|\le\beta H\Delta/2+(1-\beta)=3H\Delta/[2(1+H\Delta)]).

Multiply by (R). ∎

In practice one can replace the universal scale by the largest allowed scale computed from the exact segment contract. The comparison uses that stronger actual scale. Merely interpolating the head without checking the output derivative is insufficient: head regularity and output regularity are different constraints.

Another compression route retains a subset (C) of the samples. If every discarded sample ((x_i,y_i)) has a retained representative ((x_j,y_j)) with

[ |y_i-y_j|+|x_i-x_j|\le\eta, ]

then (0\le\ell_S-\ell_C\le\eta), (0\le u_C-u_S\le\eta), and the midpoint decoders differ uniformly by at most (\eta/2). Proof: each discarded lower cone is within (\eta) above its retained cone by the triangle inequality; upper cones give the opposite bound. The midpoint difference is half the difference of two numbers in ([0,\eta]). The factor one half is sharp for the midpoint construction. Applying the scalar nonlinear filter preserves this prediction difference because it is nonexpansive in its residual input, while each reconstructed midpoint retains the full regularity contract. This subset compression reduces envelope evaluation cost, whereas fitting the small head removes dependence on sample count during inference.

5. Compression cannot remove target information

Even the same head/output contract contains a large family. For (0<\epsilon\le1/4), place (m=\lfloor1/(8\epsilon)\rfloor) disjoint triangular tents on ([1/2,1]), each of base width (4\epsilon), peak height (\epsilon), and slope magnitude (1/2). Assign independent signs to the tents and call the resulting output (f). It is zero near (K), has Lipschitz constant at most (1/2), and its head (h=f/|x|), extended by zero elsewhere, satisfies

[ |h|_\infty\le2\epsilon,\qquad \operatorname{Lip}(h)\le1+4\epsilon\le2. ]

The head derivative bound follows from (|h'|\le |f'|/x+|f|/x^2) on the positive supports. Hence every one of the (2^m) sign patterns meets the pilot's (H=2) contract. Distinct patterns differ by (2\epsilon) at a peak. A deterministic decoder using fewer than (m) bits must map two patterns to the same code and cannot uniformly approximate both to error strictly below (\epsilon). Thus certified compression requires at least (\lfloor1/(8\epsilon)\rfloor) bits for this family, even if data fully identify the target. This is a concrete metric-entropy obstruction, not a new general information-theoretic principle.

The learned nine-coefficient heads therefore demonstrate compression for the disclosed small task families. They do not establish arbitrary-dimensional scalability or compression of unrestricted learning algorithms.