Skip to content
Shadow Theory

Section 4 9 October 2026

Complete-law preparation with every archive retained

Reading position 5 of 14

4 Complete-law preparation with every archive retained

Let XX be arbitrary genuinely retained external information, on a standard Borel space with its actual law μX\mu_X. During the present protocol XX is fixed. Include in the entrance vector the writer U∈(0,1)U\in(0,1), all nn archive inputs V1,…,VnV_1,\ldots,V_n, and any untouched future positional sources WW. Conditional on X=xX=x, suppose their complete joint law has a density fxf_x relative to product Lebesgue measure λ\lambda. The archive inputs may be arbitrarily correlated with one another, the writer and XX. All variations below are variations of this full conditional density and are averaged using μX\mu_X. In particular they are not variations of the writer marginal. For equal-mass finite measures, use TV⁡(ν,ν′)=12∣ν−ν′∣(Q)\pTV(\nu,\nu')=\frac12|\nu-\nu'|(\mathcal Q); this also covers subprobabilities of equal mass.

Write Vi(f)=∫Var⁡i(f)V_i(f)=\int\pVar_i(f) for the directional BV seminorm, with spectators and XX integrated. The same definitions apply to a restricted subdensity of total mass m≤1m\leq1. Only the explicitly active directions are varied; jumps in a fixed old archive direction are not included.

Lemma 4.1 (Grid and sectional concentration)

Let a nonnegative subdensity hh have integrated mass mm and active two-coordinate variation V=V1+V2V=V_1+V_2. Let PNP_N average on the NN by NN grid at fixed spectators. Then

∥h−PNh∥1≤V/(2N),Vi(PNh)≤Vi(h). \|h-P_Nh\|_1\leq V/(2N),\qquad V_i(P_Nh)\leq V_i(h). (4.1)

On each active unit-square fibre, with mass mzm_z and variations V1,z,V2,zV_{1,z},V_{2,z},

∥hz∥2≤(mz+V1,z)(mz+V2,z)≤mz+(V1,z+V2,z)/2. \|h_z\|_2\leq\sqrt{(m_z+V_{1,z})(m_z+V_{2,z})} \leq m_z+(V_{1,z}+V_{2,z})/2. (4.2)

If KK preserves area on each such fibre, fixes all spectators, and has active displacement ≤δ\leq\delta off a set of sectional area ≤β\leq\beta, then

TV⁡(K#(hλ),hλ)≤V2N+(m+V/2)β+4Nδ. \pTV(K_\#(h\lambda),h\lambda) \leq \frac{V}{2N}+(m+V/2)\sqrt{\beta+4N\delta}. (4.3)

If instead h≤Ch\leq C, the alternative bound is

TV⁡(K#(hλ),hλ)≤V/(2N)+2CNδ+Cβ/2. \pTV(K_\#(h\lambda),h\lambda) \leq V/(2N)+2CN\delta+C\beta/2. (4.4)
Proof

For an interval II of length ll and its mean hIh_I,

∫I∣h−hI∣≤l−1∫I×I∣h(x)−h(y)∣ dx dy≤(l/2)∣Dh∣(I). \int_I|h-h_I|\leq l^{-1}\int_{I\times I}|h(x)-h(y)|\,dx\,dy \leq (l/2)|Dh|(I).

The last inequality follows by integrating the derivative between xx and yy: its weight at t∈It\in I is 2(t−inf⁡I)(sup⁡I−t)/l≤l/22(t-\inf I)(\sup I-t)/l\leq l/2. Successive coordinate averaging contracts L1L^1 and the unused directional variations, proving the first bound in (4.1). For adjacent cell means the difference is the average of h(t+1/N)−h(t)h(t+1/N)-h(t). Integrating its derivative along that segment and summing the cells gives overlap weight at most one. This proves the variation contraction, first for smooth functions and then for BV by approximation.

For a BV square slice the one-dimensional representatives give, almost everywhere,

h(x,y)≤A(y):=∫01h(t,y)dt+Var⁡xh(⋅,y),h(x,y)≤B(x):=∫01h(x,t)dt+Var⁡yh(x,⋅). h(x,y)\leq A(y):=\int_0^1h(t,y)dt+\pVar_x h(\cdot,y), \quad h(x,y)\leq B(x):=\int_0^1h(x,t)dt+\pVar_y h(x,\cdot).

The common representatives exist by BV slicing. Thus h2≤A(y)B(x)h^2\leq A(y)B(x); Fubini proves the first inequality in (4.2), and the arithmetic–geometric mean proves the second.

Put g=PNhg=P_Nh. Its value can change under KK only on the bad set or when an active coordinate crosses a grid line. This exceptional set EE has sectional area at most s=β+4Nδs=\beta+4N\delta. Since KK preserves sectional area, so does its image. Cauchy bounds each of ∫Eg\int_Eg and ∫Eg∘K\int_Eg\circ K by s∥g∥2\sqrt{s}\|g\|_2. Their sum is divided by two in total variation. The two approximation errors together are ∥h−g∥1≤V/(2N)\|h-g\|_1\leq V/(2N), giving (4.3) after integrating the section coefficients. If 0≤g≤C0\leq g\leq C, then ∣g−g∘K∣≤C|g-g\circ K|\leq C on EE, yielding (4.4) instead. The argument prices both the exceptional set and its image.

□

Let Tn\mathcal T_n be the complete composition of the exact cycles and Bn\mathcal B_n the corresponding baker composition. Let RnR_n be the final ready quantile and Zn=(X,B1,…,Bn,W)Z_n=(X,B_1,\ldots,B_n,W) all retained external and positional nuisance variables. The internal memories are retained in the quantum wave; there is no additional hidden internal coordinate to marginalize in the specified ontology.

Theorem 4.2 (Quantitative preparation with retained archives)

Restrict the actual entrance law to a subdensity hx≤fxh_x\leq f_x of mass m=1−τm=1-\tau, with directional variations Vw,V1,…,VnV_w,V_1,\ldots,V_n in the writer and warmup archives. Set

Sn=2Vw+∑j=1nVj,Hn∗=n+Sn/2,En=Sn2N+Hn∗β+4Nδ. S_n=2V_w+\sum_{j=1}^n V_j,\qquad H_n^*=n+S_n/2, \qquad E_n=\frac{S_n}{2N}+H_n^*\sqrt{\beta+4N\delta}. (4.5)

Suppose the identical exact cycle has the sectional comparison bounds of Lemma 3.3. Then

TV⁡(Law⁡(Rn,Zn),U⊗Law⁡(Zn))≤τ+Vw4 2n+2En,Pr⁡(some archive miscopies its own writer sign)≤τ+Hn∗β+nEn.\begin{align}\pTV\bigl(\pLaw(R_n,Z_n),\pUnif\otimes\pLaw(Z_n)\bigr) &\leq\tau+\frac{V_w}{4\,2^n}+2E_n, \tag{4.6}\\ \Pr(\hbox{some archive miscopies its own writer sign}) &\leq\tau+H_n^*\sqrt\beta+nE_n. \tag{4.7}\end{align}

Each correctly copied archive sign persists throughout all later prescribed stages and holding intervals. If the full entrance density is capped by CC, one may instead take τ=0\tau=0 and

Encap=Sn2N+2nCNδ+nCβ/2. E_n^{\rm cap}=\frac{S_n}{2N}+2nCN\delta+nC\beta/2. (4.8)

In that capped case the all-record failure also has the simpler bound nCβnC\beta.

Proof

After jj ideal cycles, the archive values determine si=⌊2bi⌋s_i=\lfloor2b_i\rfloor and vi=2bi−siv_i=2b_i-s_i for 1≤i≤j1\leq i\leq j. Writing Hj=∑i=1j2j−isiH_j=\sum_{i=1}^j2^{j-i}s_i, the inverse writer coordinate is u0=2−j(u+Hj)u_0=2^{-j}(u+H_j). The full inverse Jacobian is one: the writer factor 2−j2^{-j} cancels the jj archive factors two. At fixed old archives, differentiation in the active writer therefore multiplies its variation by 2−j2^{-j}, whereas variation in the next unused vj+1v_{j+1} is unchanged. Any old archive-digit seam is in a fixed coordinate. Thus the sum of active variations before all nn ideal steps is at most SnS_n.

Telescope Tn−Bn\mathcal T_n-\mathcal B_n by replacing one ideal step at a time, using ideal prefixes and exact suffixes. Deterministic pushforward contracts total variation, so Lemma 4.1 bounds the total boxed endpoint discrepancy by EnE_n; using nm≤nnm\leq n gives (4.5). The same bound controls every partial-prefix error.

For the complete ideal output, fix all original source values and XX. The archives reveal the dyadic cell containing u0u_0, as well as those source values. Replacing the original density by its cell mean in u0u_0 is exactly the operation that makes the remainder uniform while keeping the ideal archive marginal. The interval inequality in the proof of Lemma 4.1 gives L1L^1 error at most 2−nVw/22^{-n}V_w/2 and hence TV error at most Vw/(4 2n)V_w/(4\,2^n). Transferring both the complete endpoint law and its archive marginal from ideal to exact costs 2En2E_n.

The removed entrance subprobability and the product of its own final nuisance marginal with U\pUnif each have mass τ\tau. Their distance is at most τ\tau. This proves (4.6) with that tail charged once, not once per cycle.

For each record, the wrong-copy event is contained in its pre-copy bad set. At the ideal boxed prefix, sectional Cauchy bounds its mass by (m+Vactive,j/2)β(m+V_{{\rm active},j}/2)\sqrt\beta. Replacing this prefix by the actual boxed prefix costs at most EnE_n. Sum over jj, add the initial tail once, and obtain (4.7). The zero archive current established above proves subsequent persistence. Under a cap, exact area preservation keeps that same cap at every actual prefix, giving nCβnC\beta directly. The capped grid lemma gives (4.8).

□

If the averaged complete quantile Fisher information is Jq=∫∑i∣∂uilog⁡fx∣2fx dλ dμXJ^{\rm q}=\int\sum_i|\partial_{u_i}\log f_x|^2 f_x\,d\lambda\,d\mu_X, score Cauchy gives

Sn≤n+4Jq,Vw≤Jwq. S_n\leq\sqrt{n+4}\sqrt{J^{\rm q}},\qquad V_w\leq\sqrt{J_w^{\rm q}}. (4.9)

This is one sufficient regularity condition. The physical-score result below is strictly broader and does not assume finite quantile Fisher.