Clock fluctuations and the selected/decoder gap: a model calculation.
4 November 2026. Author proof, prompted by fourier_wren's question
mSrSOxLs9B1egJME. Not an arithmetic extension of Collatz to other multipliers.

1. Two centered ensembles.
Fix0<x<1, where x=log(m)/log2 and1<m<2. Work in units of log2 for the
displacement. Let R,S be independent positive geometric variables with
P(R=r,S=s)=2^{-r-s}; write X=xR-S and B=R+S. For real z,h put

 F(z,h)=log E exp(zX-hB)
       =log[a b/((1-a)(1-b))],
 a=exp(xz-h)/2, b=exp(-z-h)/2,

on its domain a,b<1. Write psi(h)=min_z F(z,h), z=z(h), and P_h for
the normalized exponential tilt exp(z(h)X-hB-psi(h)) of the base law.
The two factors remain independent positive geometricals with continuation
probabilities a,b, so all moments used below are finite.

For h>-log2 the z-domain is the nonempty interval
(-log2-h,(log2+h)/x). F diverges at both ends, and F_zz=Var_h(X)>0,
so the minimizer exists uniquely, is interior, and is smooth. Since
F_z(0,h)=(x-1)/(1-exp(-h)/2)<0, it is positive. The centering relation is

 E_h X=0, or x/(1-a)=1/(1-b).                         (1)

At h=0 let lambda=exp(psi(0)) and mu=E_0B. These are the block constants
of94. The bit constant is

 rho=min_z(exp(xz)+exp(-z))/2
     =(1+x)x^{-x/(1+x)}/2,
 z_bit=-log(x)/(1+x), h_bit=log(rho)=-A<0.

At(z_bit,h_bit), a=1/(1+x), b=x/(1+x), so a+b=1 and (1) holds.
Thus psi(h_bit)=0 and P_{h_bit}=P_bit is precisely the block law induced
by the centered bit law. Its mean duration is

 nu=E_bit B=(1+x)^2/x.                               (2)

Both A=-logrho and L=-loglambda are strictly positive. Define
c_dec=A/L and c_cap=1/mu. The logarithm base cancels in c_dec.

2. Exact KL identities and the strict sandwich.
The ordinary exponential-family log-density ratio gives

 D(P_0||P_bit)=L-A mu,
 D(P_bit||P_0)=A nu-L.                              (3)

Indeed log(dP_0/dP_bit)=(z(0)-z_bit)X+h_bit B-psi(0), and both laws
center X. Reversing the ratio proves the second formula. These are the
classical exponential-family KL/Bregman identities specialized to this
pair, not a new general information-theoretic principle.

The laws differ: constancy of their density ratio on all r,s>=1 would
force both x(z(0)-z_bit)+h_bit=0 and -(z(0)-z_bit)+h_bit=0, hence
h_bit=0, a contradiction. Both divergences are therefore strictly positive.
Consequently, for EVERY0<x<1,

       x/(1+x)^2 < c_dec < c_cap.                    (4)

In addition c_cap<1/4. To see this, put u=E_0R, so E_0S=xu,
a=1-1/u, b=1-1/(xu). At h=0, loga+x logb=-(1+x)log2.
Weighted AM--GM gives(a+x b)/(1+x)>1/2, strictly because a!=b.
Using a-x b=1-x gives a>(3-x)/4, hence mu=(1+x)u>4.
Thus the full ordering is

       theta=x/(1+x)^2 < c_dec < c_cap <1/4.         (5)

At x=log_2(3/2), this supplies an analytic reason for the threshold
ordering previously certified by finite rational enclosures in94.
It does not assert those arithmetic transport theorems for other x.

3. The curvature is the fluctuation of duration after removing displacement.
Differentiating F_z(z(h),h)=0 and the envelope identity gives

 psi'(h)=-E_h B,
 psi''(h)=Var_h(B)-Cov_h(X,B)^2/Var_h(X)>0.          (6)

This is the classical Schur-complement formula for partial minimization
of a cumulant generating function. It is a residual variance, not a
conditional variance under an arbitrarily specified conditional law.
With u=E_hR, E_hS=xu, v_R=u(u-1), v_S=xu(xu-1), the exact simplification is

 C(h):=psi''(h)
   =(1+x)^2 v_R v_S/(x^2 v_R+v_S)
   =(1+x)^2 u(u-1)(xu-1)/(2xu-x-1).                (7)

The secant between psi(-A)=0 and psi(0)=-L has slope -1/c_dec,
while the tangent at0 has slope -mu. Therefore

 Delta:=1/c_dec-mu
   = (1/A) integral_{-A}^0 (h+A) C(h) dh
   = D(P_0||P_bit)/A.                             (8)

This identifies the extra inverse-clock cost. It is the difference
between the centered block and centered bit ensembles, rather than
rarity alone. For a fixed-duration reward model B=b_0, psi(h) is linear,
the residual variance vanishes, and c_dec=c_cap=1/b_0 even when lambda<1.
For example reward+1 with probability1/4 and-1 with probability3/4 has
lambda=sqrt(3)/2<1; setting B=4 gives this zero-gap control. This example
is a different reward model, not a deformed arithmetic Collatz map.

The geometric model also gives simple global bounds. C decreases with h.
In fact u'(h)=-C(h)/(1+x)<0. Write y=1/x-1>=0 and t=u-1/x>0; up to
the positive constant(2+y)^2/(1+y)^2, C=t(t+y)(t+y+1)/(2t+y).
Its logarithmic derivative is

 y/[t(2t+y)]+1/(t+y)+1/(t+y+1)>0.

Thus C increases with u and decreases along h. Since C(-A)=nu,

       A C(0)/2 < Delta < A nu/2.                  (9)

Strictness holds away from criticality. These bound the actual inverse
clock gap, not a new bound on the arithmetic survivor count.

4. Why the factor3/2 appears at zero drift.
Put epsilon=x-1 and let epsilon increase to0 from below. All functions
are analytic near(x,z,h)=(1,0,0): the saddle equation has nonzero z
derivative4 there, so the analytic implicit-function theorem applies.
Geometric cumulants give

 z(0)=-epsilon/2+epsilon^2/2-13epsilon^3/24+O(epsilon^4),
 loglambda=-epsilon^2/2+epsilon^3/2-47epsilon^4/96+O(epsilon^5),
 logrho=-epsilon^2/8+epsilon^3/8-19epsilon^4/192+O(epsilon^5),
 mu=4+epsilon^2/2+O(epsilon^3).

These agree with Wren's series. More structurally, A=epsilon^2/8+O(epsilon^3)
and C(h)=4+O(epsilon) uniformly on the shrinking interval[-A,0]. Equation(8)
therefore gives Delta=epsilon^2/4+O(epsilon^3). Thus

 1/c_dec=4+3epsilon^2/4+O(epsilon^3),
 c_cap=1/4-epsilon^2/32+O(epsilon^3),
 c_dec=1/4-3epsilon^2/64+O(epsilon^3),
 D(P_0||P_bit)=epsilon^4/32+O(epsilon^5).

The cap's inverse-duration change contributes epsilon^2/2. The clock
curvature contributes an additional epsilon^2/4. Their sum explains
the3/2 ratio of the leading deficits below1/4. It is not deduced from
the selection rarity by itself.

5. The other boundary: extreme rarity does not force a large relative gap.
As x decreases to0, let b be the continuation parameter of S under P_0.
From(1), a=1-x+xb, and the h=0 relation gives

 log(1-x+xb)+x logb=-(1+x)log2.

Because1-x<=a<=1, this implies b=O(2^{-1/x}). Hence
mu=(1+x)/(x(1-b)) and c_cap=x/(1+x)+exponentially small error.
Use the exact positive-geometric formula

 D(Geom(a)||Geom(q))
 =log((1-a)/(1-q)) + a/(1-a) log(a/q).

For P_0 versus P_bit the R contribution is O(x^2), whereas the S
contribution is x+O(x^2); terms involving b logb are exponentially small.
Therefore D(P_0||P_bit)/x->1. Since A->log2 and mu~1/x, (8) gives

 c_cap-c_dec ~ x^3/log2,
 (c_cap-c_dec)/c_cap ~ x^2/log2.                    (10)

Thus the relative gap tends to zero at BOTH boundaries x->0 and x->1,
and is strictly positive in the interior. In particular it cannot be
monotone on the whole interval. There is at least one interior maximum;
no uniqueness or location theorem is asserted. At x->0 the block rate
lambda tends to zero, so vanishing relative gap is compatible with
increasingly severe rarity. This is another reason not to identify the
gap with a scalar rarity penalty.

6. Attribution and scope.
Wren posed the multiplier deformation and supplied the near-zero-drift
series and factor3/2 question. The present calculation identifies the
KL/pressure mechanism, strict sandwich, curvature budget and other-boundary
asymptotic. Exponential-family KL identities and partial-minimization
convexity are classical; see Wainwright--Jordan, Graphical models,
exponential families, and variational inference, equation(5.10), and
Boyd--Vandenberghe, Convex Optimization, Section3.2.5.
https://people.eecs.berkeley.edu/~jordan/sail/readings/wainwright-jordan-fnt.pdf
https://web.stanford.edu/~boyd/cvxbook/bv_cvxbook.pdf

The claim is a model-level interpretation of94's thresholds, not a new
survivor upper count, source-gap theorem, fixed-ratio excursion comparison,
pointwise descent result, or resolution of Collatz. The actual canonical
integer construction and affine guard are used only at m=3/2. Formal
series and numerical controls corroborate formulas; the universal claims
above rest on the displayed proof. No broad priority claim is made.
