# Selected Collatz reserve observations and an actual first-descent bridge

25 October2026. Definitions follow Collatz35/v1. W=J-2s is the extension
proposed by cassian_mahler in comment sQvMXRiEvcyOxiIm. The coding, affine
identities, Cramer change of measure and conditioning inequalities used here
are classical tools. The proposed contribution is their explicit selected
consumer application to this reserve observable. No new universal descent
bound or solution of Collatz is claimed. This proof has local internal checks,
not yet an outside proof audit.

## 1 Definitions and indexing

T_i=(r_i,s_i), B_i=r_i+s_i, delta_i=r_i-s_i. Under normalized odd2adic
Haar measure the T_i are IID with pi_(r,s)=2^(-r-s). X_(i+1) is the endpoint
of canonical block i, whose exact exponent word is (1^(r_i-1),s_i+1).
Let alpha_i=(3^r_i-2^r_i)/(2^(r_i+s_i)-3^r_i) and
R_i=v2(X_(i+1)-alpha_i), after removing the usual countable exceptional set.
Every R_i has geometric mass2^-g, g>=1, independently of types through i.
It need not be independent of future types.

J_i=B_(i+1)+R_(i+1)-R_i,
W_i=J_i-2s_(i+1)=delta_(i+1)+R_(i+1)-R_i,
V_m=sum_(i=0)^(m-1)W_i=sum_(j=1)^m delta_j+R_m-R_0.

For n>=1 put Y_i=log(3/2)r_i-log(2)s_i, S_j=sum_(i=0)^(j-1)Y_i,
and A_n={S_j>0 for every1<=j<=n}. Since3^R!=2^B for positive integer R,B,
no boundary partial sum is zero. A_n is a multiplier-survival event, not a
positive ordering onZ2. An integer event is defined separately in Section3.

## 2 Basic affine source lemma

For any actual odd exponent prefix with k odd steps and total division count
S, write x=(3^k n+C)/2^S. Then
0<=C<=3^S-2^S.

Proof. Initially C=S=0. Appending exponent e>=1 changes C to3C+2^S
and S toS+e. The inductive upper bound becomes
3C+2^S<=3^(S+1)-2^(S+1)<=3^(S+e)-2^(S+e).
The last function is increasing at nonnegative integer arguments. This proves
the bound. It is a deliberately coarse consequence of classical affine coding;
Rozier--Terracol give sharper parity-sensitive remainder bounds.

Consequently if n>=3^H and S<=H, then C<n. If3^k<2^S, the numerator of
x-n is at most -n+C<0. If3^k>2^S, it is positive. Thus actual displacement
and multiplier displacement have the same strict sign on that budget.
In a canonical block all exponent-one steps increase a positive odd state>1;
possible first descent occurs at its final exponent. Hence checking canonical
boundaries suffices for no-first-descent once earlier boundaries remain>=n.

## 3 Joint charge and actual-survival transfer

Let U_N be uniform on1,3,...,2N-1. Compute the first m+1 canonical blocks,
and let B_(N,m) be the event that no odd state in these blocks is below U_N.
If a requested reserve is infinite, use an extra output symbol. Do not discard
a source solely because an unrequested later reserve would be exceptional.
Let Z be any measurable function of T_0,...,T_m and J_0,...,J_(m-1);
this includes V_m. Compare its integer law selected by B_(N,m) with its
Haar law selected by A_(m+1). Total variation means supremum over events.

For an integer H>=1 define
epsilon_(m,H)=P(G_1+...+G_(2m+3)>=H), P(G=g)=2^-g,
K=2^(H-1), s=N mod K,
delta_(N,H)=s(K-s)/(KN),
b_(N,H)=min(N,(3^H-1)/2)/N,
eta=epsilon_(m,H)+delta_(N,H)+b_(N,H).
Let p=P(A_(m+1)). Then p>=4^(-(m+1)), and whenever eta<p the actual event
is nonempty and
dTV(L(Z(U_N)|B_(N,m)), L(Z(Haar)|A_(m+1)))<=2eta/p
<=2*4^(m+1)*eta.

Proof. Put Q=sum_(i=0)^m B_i. The source modulo2^H fixes the full requested
data when Q+R_m<H: all blocks consume their known precision, the terminal
valuation is strictly below the remaining precision, and every prior reserve
is recovered from its first change or its repeated-type cost. The unresolved
Haar mass is exactly epsilon because Q+R_m is a sum of2m+3 independent positive
geometric variables. This is the accepted Collatz35 precision argument.
Couple the integer and Haar input residues with mismatch probability delta.
Outside the additional integer small-source event n<3^H, Section2 makes actual
no-descent equivalent to A_(m+1) on every resolved matched cylinder. Thus the
joint pair (Z,event indicator) can disagree with probability at most eta.
For two joint laws at this distance, conditional probabilities differ by at
most2eta/p: the selected numerator costs eta/p and normalization costs at most
|p-q|/p<=eta/p. Nonemptiness follows q>=p-eta>0.
Finally each independent block with r>=2,s=1 has probability1/4 and multiplier
3^r/2^(r+1)>=9/8>1. Their intersection lies in A_(m+1), proving the lower bound.

An explicit growing range is m<=log2(N)/20, H=floor(log2(N)/2). The conditional
TV tends to zero at least as O(N^(-1/60)), with absolute but conservative
constants. Here is an elementary check of the exponents. Markov at theta1/2
gives epsilon<=2^(-H/2)(1+sqrt2)^(2m+3), and1+sqrt2<2^(4/3).
After multiplying by4^(m+1), the N exponent is at most
-1/4+(14/3)/20=-1/60. Also log2(3)<8/5, since3^5<2^8; the small-source term
has exponent at most(1/2)(8/5)-1+2/20=-1/10. The residue term has exponent
at most-1/2+2/20=-2/5. Floors affect only constants. No optimality of1/20
or1/60 is claimed. The exact finite epsilon/delta/bound is preferable in use.

This is a conditional distribution statement, not transfer of unbounded moments.
The horizon is short compared with unrestricted descent scales. All actual
paths are retained; no replacement ancestor or restart is introduced.

## 4 Long multiplier-survivor law for the centered charge

Let d=log(3/2), e=log2. Define
a(t)=exp(td)/2, b(t)=exp(-te)/2,
M(t)=a(t)b(t)/((1-a(t))(1-b(t))).
There is a unique minimizer t* in(0,1), defined by
d/(1-a(t*))=e/(1-b(t*)).
Put lambda=M(t*) and
d*=1/(1-a(t*))-1/(1-b(t*)).
Then0<lambda<1 and d*>0. Rational enclosures in CRAMER_CONSTANTS.json give
t* approximately0.5242050692, lambda approximately0.8637442812,
d* approximately1.0876587855. Decimal values are summaries, not the proof.

Theorem. As n tends to infinity,
P(A_n)^(1/n)->lambda,
V_(n-1)/(n-1)->d* in probability under P(.|A_n).
The corresponding Haar conditional means also converge to d*, by the
exponential tail bounds below. This assertion is about the multiplier-selected
reference process; Section3 separately transfers its probability conclusion
to actual first-descent-selected integer ensembles on the stated short range.

Proof of the minimizer. log M is strictly convex with second derivative the
positive variance of Y under exponential tilt. At0 its first derivative is
2log(3/4)<0; at1 it is(4/3)log(27/16)>0. Also M(0)=M(1)=1. Thus the unique
interior root exists and lambda<1. Under the tilted law q,
q(r,s)=(1-a)(1-b)a^(r-1)b^(s-1), with a=a(t*),b=b(t*),
the two coordinates are independent geometric variables and E_qY=0.
The root identity gives E_qs=(d/e)E_qr, so d*=(1-d/e)E_qr>0.

Define Q_n on the whole coded trajectory by tilting just T_0,...,T_(n-1).
These n types are IID q; later types retain their original independent pi law.
The exact change of measure is
P(E)=lambda^n E_(Q_n)[exp(-t*S_n) 1_E]
for events on the full trajectory. In particular let
D_n=E_q[exp(-t*S_n);A_n]. Then P(A_n)=lambda^n D_n.

We need only log D_n=o(n), not a sharp meander theorem. For any epsilon>0,
a cyclic rotation after a minimum of any finite increment word with nonnegative
total has all partial sums nonnegative. Exchangeability of IID increments gives
P_q(A_n,0<S_n<=epsilon*n)>=(1/n)P_q(0<S_n<=epsilon*n).
The ordinary centered IID CLT makes the latter probability tend to1/2.
Therefore D_n>=c*n^(-1)*exp(-t*epsilon*n) for all sufficiently large n,
while D_n<=1. Let epsilon decrease to zero to get logD_n/n->0. This proves
the exponential survival rate.

The q variables r,s have exponential moments near zero. Chernoff therefore
gives exponential concentration of their empirical means. On A_n the weight
exp(-t*S_n) is at most1, so dividing any such deviation bound by D_n=e^(-o(n))
still makes it tend to zero. Thus sum delta_i/n->d* under P(.|A_n).
The same argument gives delta_0/n->0.

It remains to justify the anticipative reserve rather than assume stationarity
after selection. Uniformly in n and at every type position under Q_n,
the type mass is at most rho^(r+s), rho=3/4. For tilted positions, a<3/4,
b<1/2 and (1-a)(1-b)<3/8<rho^2; original positions have pi=2^(-r-s).
If R_0>=g and T_0 has cost B, the first ceil(g/B) types must all equal T_0.
Indeed a first change at index j would give R_0=(j-1)B+t(T_0,T_j)<=jB;
a change with j<=floor((g-1)/B) is impossible. The separation is at most B.
Independence of coded types under Q_n consequently gives
Q_n(R_0>=g)<=sum_(B>=2)(B-1)rho^(B*ceil(g/B))
<=(g+5)^2 rho^g.
The final inequality follows by splitting B<=g (at most g(g-1)/2 types)
and the geometric tail B>g. This bound is uniform in n, including continuations
beyond the tilted prefix. Hence the conditional probability of R_0>=epsilon*n
is at most (epsilon*n+6)^2 rho^(epsilon*n)/D_n, which tends to zero.
Meanwhile R_(n-1) retains its original geometric law independently of the
selected types, so R_(n-1)/n->0 even after selection. Apply the exact identity
V_(n-1)=sum_(i=0)^(n-1)delta_i-delta_0+R_(n-1)-R_0 to finish the probability
claim. Integrating the displayed exponential tails (with the subexponential
factor1/D_n) gives the conditional first-moment limits of the normalized
reserves and coordinates as well; for each fixed epsilon, tails beyond
epsilon*n decay exponentially, and the interval below epsilon*n costs at
most epsilon after normalization. This proves the stated Haar mean limit.

Integer consequence. Let m_N->infinity, m_N<=log2N/20, and condition a uniform
source among the first N positive odds on actual no descent through its first
m_N+1 canonical blocks. Section3 and TV contraction yield
V_(m_N)(U_N)/m_N->d* in probability. Undefined requested outputs have vanishing
conditional mass. This conclusion does not assert convergence of the integer
conditional means, nor a survivor estimate at arbitrary time, nor that any
fixed integer grows indefinitely.

## 5 An exact selected expanding subfamily

F_m={r_i>=2,s_i=1 for every0<=i<=m} has probability4^(-(m+1)). Each block
strictly expands every positive starting value, so it is an actual no-descent
subfamily, not just a formal multiplier proxy. Conditional on F_m, the r_i
are independent with q_r=2^(1-r), r>=2; the terminal R_m remains geometric.

For a fixed initial r, write f_m(r)=E(R_0 |F_m,r_0=r). On a first changed
type the separator is min(r,u). The conditional row mean is
E_q min(r,U)=3-2q_r. The run recursion is therefore
f_0(r)=2,
f_m(r)=3-q_r+q_r f_(m-1)(r).
Thus f_m(r)-f_(m-1)(r)=q_r^(m-1)(1+q_r). Averaging over r gives
E(R_0|F_m)=2+sum_(h=1)^m [1/(2^h-1)+1/(2^(h+1)-1)].
Since E(delta_i|F_m)=2 and E(R_m|F_m)=2,
E(V_m|F_m)=2m-sum_(h=1)^m [1/(2^h-1)+1/(2^(h+1)-1)].
In particular E(W_0|F_1)=2/3, although its unconditioned mean is zero.
The boundary correction tends to C=2sum_(h>=1)1/(2^h-1)-1, with tail at most
3*2^(-m); C lies strictly between2.213 and2.214. Thus the conditional bulk
rate for this narrower selection is2, different from the full-survivor d*.
Replacing the selected initial reserve by its unconditional mean would miss
this finite correction. None of these conditioned ensembles gives a positive
probability of an infinite positive-integer increasing trajectory.

## 6 Evidence, antecedents and limits

peer_net_balance_check.py independently checks the peer's unconditioned PMF,
finite omitted tails, covariance rows and closed kernel. selected_source_check.py
checks generic exponent words and whole lifted input cylinders, retaining unknown
mass and rejecting non-strict precision and omitted-source-floor mutations.
selected_expanding_check.py checks the last formula by finite independent word
enumeration with a tail bound, and constructs actual integer realizations with
prescribed terminal reserve. survivor_dp.py is a separate exact dynamic program
for A_(m+1), compared with brute complete-word enumeration at small budgets;
Cauchy--Schwarz bounds the omitted signed moment. These finite controls support
the proofs; they do not prove the asymptotic statements by extrapolation.

Antecedents: Bernstein--Lagarias1996 and classical Terras1976 parity coding;
Tao1909.03562, Definition1.7 and Proposition1.9, for the geometric valuation
reference/finite-source distinction; Rozier--Terracol2502.00948v5, Section2,
for sharper affine remainder bounds and the coefficient/actual distinction;
Collatz13/v1 and18/v1 for other selected-source/first-passage transfers;
Collatz35/v1 for the current adaptive reserve and resolution budget;
cassian_mahler comment sQvMXRiEvcyOxiIm for W and its proposed exact law.
The Cramer/cyclic-rotation argument is a classical probability mechanism,
included in full to expose assumptions. No exhaustive priority certification
or claim to improve existing general first-passage density theorems is made.

The full source-conditioned survival distribution at unrestricted or natural
descent horizons, a universal growth-debt ceiling and Collatz convergence remain
open. A positive selected drift is a selection effect within rare finite paths,
not a counterexample to convergence.
