Introduction
Quantum query complexity provides a framework for understanding the power and limitations of quantum algorithms when access to the input is restricted to oracle queries. In the worst-case setting, an algorithm must succeed on every input satisfying the promise. The polynomial method [5 ] and the adversary method [2 , 18 ] are central tools for proving quantum query lower bounds. The general adversary bound characterizes bounded-error query complexity for function evaluation up to constant factors, but constructing suitable adversary matrices and bounding their norms can be technically demanding.
In the average-case setting, an algorithm’s success probability is averaged over a prescribed input distribution. A lower bound for such a distribution also gives a worst-case lower bound whenever its support consists of valid inputs. The structure of the distribution offers an additional way to analyze the information acquired through queries. Zhandry’s compressed-oracle technique [27 ] can be viewed as a quantum analogue of lazy sampling: it represents a random function by a superposition of small, orthogonal databases. Each query changes the database by a bounded amount, and successful search can be related to the amplitude of databases containing the required witnesses. A lower bound can therefore be obtained by controlling how quickly these witnesses form.
The original construction exploits the independence of function values at different inputs. Extending this approach to correlated distributions is substantially more difficult. For a random permutation, injectivity couples the values across the domain; for a random function with prescribed multiplicities, observing one value constrains the multiplicities remaining elsewhere. Compressed permutation oracles [14 , 15 ] provide a way to represent these dependencies through partial injections. They also suggest a route to collision bounds under exact-multiplicity promises, where the independence needed by the original compressed oracle is unavailable.
Collision search under such promises includes two different regimes. In the first, many equality classes can supply a target collision. A basic example is a function of pure type ( k n ) (k^n) ( k n ) , in which every image value has exactly k k k preimages. For 2 ≤ r ≤ k 2 \le r \le k 2 ≤ r ≤ k , every class contains an r r r -collision. This is the abundant-collision regime familiar from the collision problem and from the equal-image case of set equality [11 , 9 ] . In the second regime, only a few classes contain a target collision, while smaller collisions may be plentiful. This is the situation underlying r r r -distinctness. The type ( k n , k + 1 ) (k^n,k+1) ( k n , k + 1 ) isolates the difficulty: there is one ( k + 1 ) (k+1) ( k + 1 ) -collision to find, surrounded by n n n classes of size k k k . We study these two collision-search regimes within a single exact-type formulation.
We consider functions f : [ N ] → Z q f : [N] \to\mathbb{Z}_q f : [ N ] → Z q of exact type
λ = ( 1 n 1 , … , K n K ) , N = ∑ d = 1 K d n d . \lambda= (1^{n_1}, \ldots, K^{n_K}), \qquad N = \sum_{d=1}^{K} d n_d. λ = ( 1 n 1 , … , K n K ) , N = d = 1 ∑ K d n d .
This means that exactly n d n_d n d distinct output values have d d d preimages. Define the tail counts B j = ∑ d = j K n d B_j = \sum_{d=j}^{K} n_d B j = ∑ d = j K n d and set M = B r ≥ 1 M = B_r \ge1 M = B r ≥ 1 . The alphabet is feasible precisely when q ≥ B 1 q \ge B_1 q ≥ B 1 . An r r r -collision is a list of r r r distinct inputs with the same function value. The algorithm may know the numbers n d n_d n d , but it receives only the ordinary function oracle. It is not given the members of any equality class. Throughout the main result, K K K and r r r are fixed and one collision is requested.
The type affects which partial collisions are useful. A repeated pair is progress toward a triple only if its equality class has another element. With one triple and many two-element classes, most pairs cannot be extended. When eligible classes are abundant, on the other hand, tables of partial collisions can be built and extended by quantum search. The number M M M of target classes alone does not capture both behaviors: the intermediate populations B 2 , … , B r − 1 B_2,\ldots,B_{r-1} B 2 , … , B r − 1 also matter.
This dependence on the type is the subject of an explicit open question of Belovs [8, Section 7]. After proving a tight lower bound for unrestricted r r r -distinctness over sufficiently large alphabets, he asked whether the same bound holds with one r r r -element class, linearly many ( r − 1 ) (r-1) ( r − 1 ) -element classes, and no classes of other sizes. In our notation this is the type ( k n , k + 1 ) (k^n,k+1) ( k n , k + 1 ) with r = k + 1 r=k+1 r = k + 1 . He also asked for a tight characterization for an arbitrary prescribed partition. We resolve both questions for every bounded type, uniformly over all alphabets that realize the type. The resulting formula connects the abundant-collision and sparse-target regimes through the full sequence of intermediate populations.
Our contributions
We determine, up to constant factors, the quantum query complexity of finding an r r r -collision in a function of arbitrary prescribed bounded type. Our characterization gives matching lower and upper bounds throughout the abundant-collision and sparse-target regimes.
For the tail counts B j B_j B j defined above, write
Q r ( λ ) = max 2 ≤ j ≤ r ( N 2 j − 1 ∏ i = 2 j − 1 B i 2 i − 2 M 2 j − 2 ) 1 / ( 2 j − 1 ) , M = B r , (1) Q_r(\lambda)=\max_{2\le j\le r}\left(\frac{N^{2^{j-1}}\prod_{i=2}^{j-1}B_i^{2^{i-2}}}{M^{2^{j-2}}}\right)^{1/(2^j-1)},\qquad M=B_r, \tag*{(1)} Q r ( λ ) = 2 ≤ j ≤ r max ( M 2 j − 2 N 2 j − 1 ∏ i = 2 j − 1 B i 2 i − 2 ) 1/ ( 2 j − 1 ) , M = B r , (1)
where an empty product equals one.
Theorem 1.1 (Quantum search for a prescribed bounded type). Fix K ≥ r ≥ 2 K\ge r\ge2 K ≥ r ≥ 2 . For every sufficiently large N N N , every exact type λ \lambda λ of maximum class size at most K K K with B r ≥ 1 B_r\ge1 B r ≥ 1 , and every alphabet q ≥ B 1 q\ge B_1 q ≥ B 1 , the worst-case quantum query complexity of finding one r r r -collision with probability at least 2 / 3 2/3 2/3 is
Θ K , r ( Q r ( λ ) ) . \Theta_{K,r}(Q_r(\lambda)). Θ K , r ( Q r ( λ )) . The type may be supplied to the algorithm. The lower bound also holds when success is averaged over the uniform distribution of functions of that exact type. The constants are independent of the alphabet and the type.
Although the expression in (1) involves several scales, it is a single characterization valid for every bounded type. Each scale captures a possible obstruction at an intermediate collision order. The following examples illustrate the range of the theorem.
Example 1 (Pure type ( k n ) (k^n) ( k n ) ): Let N = k n N=kn N = k n and fix 2 ≤ r ≤ k 2\le r\le k 2 ≤ r ≤ k . All relevant tail counts equal n = Θ k ( N ) n=\Theta_k(N) n = Θ k ( N ) , so Theorem 1.1 gives
Θ k , r ( N ( 2 r − 1 ) / ( 2 r − 1 ) ) . \Theta_{k,r}\left(N^{(2^r-1)/(2^r-1)}\right). Θ k , r ( N ( 2 r − 1 ) / ( 2 r − 1 ) ) .
This recovers the single-target specialization of our earlier multi-collision bounds [23 ] . When r = k r=k r = k , it also recovers the tight adaptive bound of Peng and Wu [24 ] for promised k k k -to-one finding, with the matching algorithm supplied by Liu and Zhandry [19 ] . Thus both single-target pure-type lower bounds are contained in the characterization for arbitrary bounded types. In particular, the pair-finding scale is Θ ( N 1 / 3 ) \Theta(N^{1/3}) Θ ( N 1/3 ) , and higher collision orders are covered even though the input has no equality classes of intermediate sizes.
Example 2 (Exceptional type ( k n , k + 1 ) (k^n,k+1) ( k n , k + 1 ) ): Here N = k n + k + 1 N=kn+k+1 N = k n + k + 1 , and the task is to find the unique ( k + 1 ) (k+1) ( k + 1 ) -collision. The tail counts are B 2 = ⋯ = B k = n + 1 = Θ k ( N ) B_2=\cdots=B_k=n+1=\Theta_k(N) B 2 = ⋯ = B k = n + 1 = Θ k ( N ) and B k + 1 = 1 B_{k+1}=1 B k + 1 = 1 . Substitution in Theorem 1.1 gives the following consequence.
Corollary 1.2 (Tight search on the exceptional type). For fixed k ≥ 2 k \ge2 k ≥ 2 , the bounded-error function-query complexity of finding the unique ( k + 1 ) (k+1) ( k + 1 ) -collision in type ( k n , k + 1 ) (k^n,k+1) ( k n , k + 1 ) is
Θ k ( N 3 / 4 − 1 / [ 4 ( 2 k + 1 − 1 ) ] ) \Theta_k\left(N^{3/4-1/[4(2^{k+1}-1)]}\right) Θ k ( N 3/4 − 1/ [ 4 ( 2 k + 1 − 1 )] ) for every feasible alphabet q ≥ n + 1 q \ge n+1 q ≥ n + 1 .
Indeed, the j j j th scale in (1) is Θ k ( N 3 / 4 − 1 / [ 4 ( 2 j − 1 ) ] ) \Theta_k\left(N^{3/4-1/[4(2^j-1)]}\right) Θ k ( N 3/4 − 1/ [ 4 ( 2 j − 1 )] ) , which increases with j j j . This resolves the open problem posed by Belovs in [8 Section 7] . His analysis of hard input types uses linearly many classes of every smaller size. Our lower bound attains the same distinctness exponent when all non-target classes have size k k k , with no singletons or other intermediate class sizes. The general theorem likewise imposes no linear-population assumption on any intermediate class size. The bound also holds for the smallest feasible alphabet, q = n + 1 q=n+1 q = n + 1 , so no additional large-alphabet assumption is needed.
Example 3 (Type ( 1 n , 2 b , 3 ) (1^n,2^b,3) ( 1 n , 2 b , 3 ) ): Set N = n + 2 b + 3 N=n+2b+3 N = n + 2 b + 3 . Since M = 1 M=1 M = 1 and B 2 = b + 1 B_2=b+1 B 2 = b + 1 , the tight complexity of finding the unique triple is
Θ ( max { N 2 / 3 , N 4 / 7 ( b + 1 ) 1 / 7 } ) . (2) \Theta\left(\max\left\{N^{2/3},N^{4/7}(b+1)^{1/7}\right\}\right). \tag*{(2)} Θ ( max { N 2/3 , N 4/7 ( b + 1 ) 1/7 } ) . (2) The two terms balance at b + 1 ≍ N 2 / 3 b+1\asymp N^{2/3} b + 1 ≍ N 2/3 . Below this scale, the element-distinctness term dominates; above it, the population of non-extendable pairs determines the complexity. In particular, b = O ( 1 ) b=O(1) b = O ( 1 ) gives Θ ( n 2 / 3 ) \Theta(n^{2/3}) Θ ( n 2/3 ) , the element-distinctness scale, whereas b = Ω ( n ) b=\Omega(n) b = Ω ( n ) gives b = Θ ( N ) b=\Theta(N) b = Θ ( N ) and complexity Θ ( N 5 / 7 ) \Theta(N^{5/7}) Θ ( N 5/7 ) , matching the worst-case three-distinctness exponent. Thus the number of pairs controls a transition between two familiar query bounds.
Beyond this characterization, our contributions include both lower-bound techniques and matching algorithms.
A common framework for collision and distinctness regimes. The complexity formula places the abundant-collision regime associated with collision and set equality and the sparse-target regime associated with k k k -distinctness in one quantitative framework. It identifies how the populations at successive collision orders govern the transition between them. Our lower-bound argument tracks partial records within equality classes, so the intermediate levels remain available even when the input contains no actual classes of those sizes. This permits a uniform analysis of types that previously required different treatments. Our bounds require neither linearly many classes of each intermediate size nor a large-alphabet assumption: they hold uniformly for every q ≥ B 1 q \ge B_1 q ≥ B 1 , including the smallest feasible alphabet q = B 1 q=B_1 q = B 1 , with constants independent of q q q .
Filtered CPO beyond the square-root scale. We introduce a filtered compressed permutation oracle (filtered CPO) process for analyzing complete function queries. After each such query, the process projects the database onto its harmonic subspace. We prove an O ( T / N ) O(T/N) O ( T / N ) comparison with the real function-query evolution, which remains small for T = o ( N ) T=o(N) T = o ( N ) (Corollary 4.19 ). This enables lower bounds beyond the square-root horizon of ordinary CPO soundness [15 ] , including the distinctness scales above. We also prove that the unfiltered process has an intrinsic square-root obstruction even on complete function-query blocks (Proposition 4.16 ). The filtered process, together with hidden capacity extensions and moment estimates in terms of tail counts, supplies the lower bound in Theorem 1.1 .
Matching upper bounds via adaptive learning graphs. We give function-query algorithms attaining O K , r ( Q r ( λ ) ) O_{K,r}(Q_r(\lambda)) O K , r ( Q r ( λ )) on every input of the prescribed type, without logarithmic overhead (Theorem 8.1 ). For scales at most N \sqrt{N} N , sampled collision tables and Grover enumeration suffice. For larger scales, we develop a colored adaptive learning graph (ALG) whose costs depend on the intermediate populations. Ordered restriction of the input roles supplies the required bounds on prefix populations, and a coherent decision-to-search conversion has only constant overhead. These constructions establish the matching upper bound across the full range of bounded types.
An alternative nested quantum-walk algorithm. For three-collision search, we also give a nested quantum random walk (nested QRW) with query complexity
O ~ K ( max { N 2 / 3 M − 1 / 3 , N 4 / 7 B 2 1 / 7 M − 2 / 7 } ) \widetilde{O}_{K}\left(\max\left\{N^{2/3}M^{-1/3},\ N^{4/7}B_{2}^{1/7}M^{-2/7}\right\}\right) O K ( max { N 2/3 M − 1/3 , N 4/7 B 2 1/7 M − 2/7 } )
matching the type-dependent lower bound up to polylogarithmic factors (Theorem A.1). Its main ingredient is a reversible inner-buffer update that retains all eligible pairs, including overlapping pairs. The stationary-distribution, spectral-gap, and initialization analyses are given in Appendix § 12 . This provides a separate quantum-walk construction for the three-collision bound for arbitrary bounded types.
Related work
Collision and set equality. Brassard, Høyer, and Tapp [11 ] gave a quantum algorithm that finds a pair in a k k k -to-one function on N N N inputs using O ( ( N / k ) 1 / 3 ) O((N/k)^{1/3}) O (( N / k ) 1/3 ) expected queries. Aaronson and Shi [1 ] proved matching collision lower bounds by the polynomial method; Ambainis [3 ] subsequently removed the large-range restriction. Belovs and Rosmanis [9 ] gave tight general adversary bounds for collision and set equality under a sufficiently large alphabet assumption. In set equality, two injective functions on n n n -element domains are promised to have equal or disjoint images. In the equal-image case their disjoint union has type ( 2 n ) (2^n) ( 2 n ) , with one preimage in each domain for every value. This explains the connection with the abundant-collision regime; the set-equality promise also specifies how the preimages are distributed between the domains.
Higher-order and multiple collisions. Liu and Zhandry [19 ] obtained the exponent ( 2 r − 1 − 1 ) / ( 2 r − 1 ) (2^{r-1}-1)/(2^r-1) ( 2 r − 1 − 1 ) / ( 2 r − 1 ) for finding one r r r -collision. Their upper-bound construction applies to exact r r r -to-one functions and, more generally, to sufficiently compressing functions. Their matching lower bound uses the compressed oracle for independently sampled random functions. Hamoudi and Magniez [17 ] developed strong success-probability bounds for finding many collision pairs and used them to prove quantum time-space tradeoffs. These random-function lower bounds do not directly transfer to exact types: conditioning on the complete type destroys the independence on which the compression relies. Our lower bound treats the uniform distribution on each prescribed type directly. Our earlier work [23 ] established tight pure-type bounds for collisions in multiple distinct classes. Peng and Wu [24 ] obtained tight adaptive and non-adaptive bounds for promised k k k -to-one finding. The present theorem extends the adaptive single-target characterization from pure types to arbitrary bounded types.
k-distinctness. Ambainis’s quantum walk [4 ] solves element distinctness in O ( N 2 / 3 ) O(N^{2/3}) O ( N 2/3 ) queries, matching the polynomial lower bound, and finds r r r equal elements in O ( N r / ( r + 1 ) ) O(N^{r/(r+1)}) O ( N r / ( r + 1 ) ) queries. Belovs [6 ] improved the latter bound to O ( N 3 / 4 − 1 / [ 4 ( 2 r − 1 ) ] ) O(N^{3/4-1/[4(2^r-1)]}) O ( N 3/4 − 1/ [ 4 ( 2 r − 1 )] ) using adaptive learning graphs. On the lower-bound side, Bun, Kothari, and Thaler [12 ] and Mande, Thaler, and Zhu [21 ] used dual polynomials to obtain exponents approaching 3 / 4 3/4 3/4 as r r r grows. Belovs [8 ] recently proved the matching lower bound for every fixed r r r , assuming an alphabet of size Ω ( N 2 ) \Omega(N^2) Ω ( N 2 ) . His proof uses hard input types with many classes of each smaller size. Our question fixes the exact type: a bound attained on those hard input types need not hold on a specified type with missing class sizes. Theorem 1.1 determines this dependence for bounded types and all feasible alphabets.
Compressed permutation oracles. Rosmanis [25 ] introduced a representation-theoretic decomposition of the permutation-oracle space into knowledge subspaces, recovering the tight permutation-inversion bound under forward queries. Majenz, Malavolta, and Walter [22 ] developed a permutation superposition oracle based on strictly monotone factorizations into transpositions. This gives a nonlocal database representation of a random permutation, supports both forward and inverse queries, and yields search bounds for predicates on a single input-output pair.
Unruh [26 ] proposed a compressed permutation oracle in which inverse queries are implemented by reversing the database; its soundness was left conjectural. Carolan [14 ] modified this construction to preserve injectivity explicitly and proved its soundness. Carolan and Majenz [15 ] improved its soundness horizon from the N 1 / 12 N^{1/12} N 1/12 scale to the tight N 1 / 2 N^{1/2} N 1/2 scale through a representation-theoretic analysis of the compression isometry. We use their completion-frame structure to analyze complete function queries with a harmonic projection after each query. Keeping the function-query interface is essential for types with few target classes, where independent permutation access can reveal additional information. The next subsection explains the filtered comparison and the estimates in terms of tail counts needed for our lower bounds.
Technique overview
Two issues govern the lower bound. First, a useful database representation must remain accurate beyond the square-root query scale. Second, it must distinguish partial collisions that can be extended to a target from those trapped in smaller classes. We address the first issue by filtering complete function queries, and the second by comparing hidden bucket capacities. Moment estimates connect these two parts and identify the hardest intermediate collision order. The upper bound balances the same intermediate populations through sampling and adaptive learning graphs.
Preserving the function-query interface. We represent a function of the prescribed type as f = a ∘ β ∘ π f = a \circ\beta\circ\pi f = a ∘ β ∘ π , where π \pi π is a uniform hidden permutation of slots, β \beta β assigns slots to buckets of the prescribed sizes, and a a a is an injective table of output labels. This samples the exact type uniformly for every feasible alphabet. The algorithm receives only the function oracle; the permutation, bucket assignment, and label table belong to the representation used in the analysis. In the phase-query picture, a complete compressed function query has the form F ′ = C x L C x F' = C_x L C_x F ′ = C x L C x : decompress at the queried input, apply the bucket-label phase, and recompress. Its temporary response register returns to the blank state. Analyzing this complete block preserves the restriction on oracle access that is needed for sparse target classes.
Filtering the database evolution. Completion vectors associate a partial injection with the superposition of permutations extending it. After removing the contributions already represented at smaller degrees, their adjoints identify the harmonic database subspace. Its coefficients have zero sums along each input- and output-deletion fiber. The associated Gram operator gives two useful encodings: V ~ \widetilde{V} V preserves norms, while W W W exactly inverts completion synthesis. These maps differ by O ( T / N ) O(T/N) O ( T / N ) on states of degree at most T T T in the range we use. We follow the query dynamics through W W W and use V ~ \widetilde{V} V at the endpoint to compare probabilities.
An unfiltered complete query can create a component of norm O ( N − 1 / 2 ) O(N^{-1/2}) O ( N − 1/2 ) outside the harmonic subspace. This is a substantive obstruction: Proposition 4.16 gives a distinguishing experiment at T = Θ ( N ) T = \Theta(\sqrt{N}) T = Θ ( N ) using only complete function-query blocks. Our analytical process instead applies the harmonic projector K K K after each complete query, without renormalizing:
ξ 0 = ∣ ∅ ⟩ , ξ t + 1 = K F t U t ξ t , ∥ ξ T − V ~ ψ T ∥ = O ( T / N ) . \xi_0 = |\varnothing\rangle,\qquad\xi_{t+1} = K F_t U_t \xi_t,\qquad\lVert\xi_T - \widetilde{V}\psi_T\rVert= O(T/N). ξ 0 = ∣ ∅ ⟩ , ξ t + 1 = K F t U t ξ t , ∥ ξ T − V ψ T ∥ = O ( T / N ) .
Here U t \mathcal{U}_t U t is the algorithm’s free unitary and ψ T \psi_T ψ T is its real state; the initial vector includes the same algorithm registers. The comparison holds for T ≤ N / 4 T \le N/4 T ≤ N /4 and sufficiently large N N N (Corollary 4.19 ). Its key estimate concerns the return to the harmonic subspace: the upward degree component is exact, while conditioning on fresh and recorded inputs bounds the middle and downward corrections by O ( 1 / ( N − 2 t ) ) O(1/(N-2t)) O ( 1/ ( N − 2 t )) . Summing these corrections gives O ( T / N ) O(T/N) O ( T / N ) , which remains small at all scales needed here.
Transferring a hidden capacity extension. The type ( k n , k + 1 ) (k^n,k+1) ( k n , k + 1 ) illustrates the second part of the proof. Its auxiliary database has n + 1 n+1 n + 1 buckets of capacity k k k ; a uniform hidden mark specifies which receives the extra slot. The source has N N N possible inputs and N − 1 N-1 N − 1 slots and consists of partial injections, rather than a negative distribution of complete functions. Following Belovs’s transfer viewpoint [8 ] , a map J J J carries an observed input-to-bucket assignment from the source to the target. Different capacities give different numbers of physical realizations of that assignment. In normalized bucket-orbit coordinates, this requires the weight
J b ∣ d ˉ ⟩ − = ( k + 1 ) h ( k ) h ∣ d ˉ ⟩ + , b , h = o b ( d ) , J = 1 n + 1 ∑ b = 1 n + 1 ∣ b ⟩ ⊗ J b . J_b\lvert\bar d\rangle_-=\sqrt{\frac{(k+1)_h}{(k)_h}}\lvert\bar d\rangle_{+,b},\qquad h=o_b(d),\qquad J=\frac{1}{\sqrt{n+1}}\sum_{b=1}^{n+1}\lvert b\rangle\otimes J_b. J b ∣ d ˉ ⟩ − = ( k ) h ( k + 1 ) h ∣ d ˉ ⟩ + , b , h = o b ( d ) , J = n + 1 1 b = 1 ∑ n + 1 ∣ b ⟩ ⊗ J b .
The map preserves the label lookup and sends the initial source state to the initial target state. Its image cannot contain all k + 1 k+1 k + 1 records in the marked bucket. Consequently, the failure of J J J to intertwine the query evolutions measures progress toward the exceptional collision.
Let C j C_j C j count buckets with at least j j j records, and let μ j ( t ) \mu_j(t) μ j ( t ) be its unnormalized first moment in the filtered source state. The principal query defect is controlled by N − 1 1 + μ k ( t ) N^{-1}\sqrt{1+\mu_k(t)} N − 1 1 + μ k ( t ) , so full capacity-k k k buckets are the relevant intermediate resource. Filtering requires additional work: J J J need not preserve harmonicity, and harmonic projection need not decrease a collision moment. Deletion identities, estimates for both directions of leakage, and a separate counter-projection bound control these effects. The histories are compared through S = K + J K − S=K^+JK^- S = K + J K − , without assuming that J J J is an isometry. For 0 ≤ t ≤ T ≤ N / 4 0\le t\le T\le N/4 0 ≤ t ≤ T ≤ N /4 , the resulting bounds are
μ j ( t ) = O k ( t 2 1 − j N 1 − 2 1 − j ) , 1 ≤ j ≤ k , \mu_j(t)=O_k\left(\frac{t^{2^{1-j}}}{N^{1-2^{1-j}}}\right),\qquad1\le j\le k, μ j ( t ) = O k ( N 1 − 2 1 − j t 2 1 − j ) , 1 ≤ j ≤ k ,
∥ P ∗ ξ T + ∥ = O k ( T N + T 2 − 2 − k N 3 / 2 − 2 − k ) . \lVert P_*\xi_T^+\rVert=O_k\left(\frac{T}{N}+\frac{T^{2-2^{-k}}}{N^{3/2-2^{-k}}}\right). ∥ P ∗ ξ T + ∥ = O k ( N T + N 3/2 − 2 − k T 2 − 2 − k ) .
where P ∗ P_* P ∗ tests whether the exceptional bucket is fully recorded. After the algorithm fixes its output, a constant number of verification reads converts this amplitude estimate into its success bound. The terminal conversion contributes O k ( T / N + N − 1 / 2 ) O_k(T/N+N^{-1/2}) O k ( T / N + N − 1/2 ) ; its N − 1 / 2 N^{-1/2} N − 1/2 contribution is incurred only a constant number of times.
Why arbitrary types give a maximum. For a general type, the useful intermediate populations are the tail counts B j B_j B j , rather than just M = B n M=B_n M = B n , the number of target classes. We extend the comparison to bounded capacity jumps c → d c\to d c → d . Localizing the defect to progress past occupancy c c c and averaging the hidden mark gives a squared-norm estimate that retains both the source candidate population and the target population.
The moment recursion is proved simultaneously over the target and its auxiliary capacity multisets. Reducing a capacity to j − 1 j-1 j − 1 preserves every lower tail count, allowing the same induction even when some exact class sizes are absent. We propagate the additive projection and transfer errors through every level with their population dependence intact; a coarse bound that discards this information would lose control beyond N \sqrt{N} N .
The proof selects an order maximizing (1) . If target classes are sufficiently abundant, a moment estimate at this order already bounds success. Otherwise, the maximizing inequalities identify an actually present smaller capacity with enough candidate buckets, and a final hidden extension bounds progress into the target classes. This yields the maximum over intermediate orders without requiring linearly many classes of each size. Terminal verification treats the union of target-bucket events directly, avoiding an extra factor M \sqrt{M} M . Every estimate is uniform for each injective label table; averaging the success probabilities therefore requires only q ≥ B 1 q \ge B_1 q ≥ B 1 .
Matching upper bounds. The upper bound chooses raw set sizes so that preparation, successive prefix extensions, and target completion each cost O K , r ( Q r ) O_{K,r}(Q_r) O K , r ( Q r ) . The inequalities defining Q r Q_r Q r ensure that these requirements can be balanced. Realizing the parameters uses different constructions on the two sides of the square-root scale.
When Q r ≤ N Q_r \le\sqrt{N} Q r ≤ N , inputs are assigned to disjoint random roles and sampled at the chosen rates. Querying the first sample gives a table of values; successive Grover enumerations extend equal-value prefixes, and a final search completes an r r r -collision. The parameters give constant probability that a target survives while controlling the intermediate populations. Each enumeration has a fixed total budget, which avoids a logarithmic loss from requiring each individual discovery to succeed with high probability.
When Q r ≥ N Q_r \ge\sqrt{N} Q r ≥ N , a colored adaptive learning graph uses bounds on prefix populations for both positive and negative inputs. We prove feasibility through the dual-adversary constraints, including the fault-correction terms. Ordered restriction of the roles isolates a colorful target with constant probability while preserving enough space for the raw sets and controlling every loading cost. To convert decision into search, parity masks on the first role encode the isolated target input. A coherent compute–phase–uncompute operation followed by a Hadamard transform recovers that input with constant overhead; Grover search finds the remaining preimages in O K , r ( N ) O_{K,r}(\sqrt{N}) O K , r ( N ) additional queries (Lemma 8.9 ). These two constructions establish the matching query bound without logarithmic overhead.
A nested-walk alternative for three-collisions. The appendix gives a separate algorithm matching the three-collision bound up to polylogarithmic factors. Its outer states are lists of pairs from distinct equality classes. An inner walk prepares an admissible queried buffer, from which all eligible pairs are extracted for a reversible batch exchange. The underlying pair pool may contain overlaps, while an admissible buffer contains at most two inputs from each class. Compatible-buffer counts determine a generally nonuniform stationary distribution. We analyze this distribution, the spectral gap, and the cost of preparing the stationary state. Balancing initialization, updates, and third-input checks gives Theorem A.1 .
Organization. Section 2 recalls the CPO, transfer-map, learning-graph, and quantum-walk frameworks. Section 3 defines the oracle representation and the target and auxiliary databases. Section 4 establishes the completion frames, conditioning estimates, and filtered history comparison. Section 5 constructs the capacity-extension map and controls its interaction with harmonic projections and collision counters. Section 6 proves the exceptional-type lower bound, including terminal verification, and Section 7 extends the analysis to arbitrary bounded types. Section 8 gives the matching search algorithms and completes the proof of Theorem 1.1 . Section 9 discusses consequences and further directions. Appendix A contains the alternative nested-walk construction for three-collision search.
Preliminaries
Conventions. We write [ a ] = { 1 , … , a } [a]=\{1,\ldots,a\} [ a ] = { 1 , … , a } , S a S_a S a for its permutation group, and ( u ) j = u ( u − 1 ) ⋯ ( u − j + 1 ) (u)_j=u(u-1)\cdots(u-j+1) ( u ) j = u ( u − 1 ) ⋯ ( u − j + 1 ) with ( u ) 0 = 1 (u)_0=1 ( u ) 0 = 1 . Hilbert spaces are finite-dimensional over C \mathbb{C} C ; the inner product is linear in its second argument. For an operator B B B , B † B^\dagger B † is its adjoint, ker B \ker B ker B its kernel, and ran B \operatorname{ran} B ran B its range. The symbol ∥ ⋅ ∥ \lVert\cdot\rVert ∥ ⋅ ∥ denotes vector norm or operator norm according to its argument. For self-adjoint operators on the same space, B ⪯ C B\preceq C B ⪯ C means that C − B C-B C − B is positive semidefinite. A subspace reduces an operator if both it and its orthogonal complement are invariant. We identify a projection with its inclusion into the ambient space when composing maps. All constants implicit in O k O_k O k may depend on the fixed integer k k k , but not on the number of buckets, the alphabet or the algorithm’s workspace.
Query model. For f : [ N ] → Z q f:[N]\to\mathbb{Z}_q f : [ N ] → Z q , a query applies O f ∣ x , z ⟩ = ∣ x , z + f ( x ) ⟩ O_f|x,z\rangle=|x,z+f(x)\rangle O f ∣ x , z ⟩ = ∣ x , z + f ( x )⟩ . Equivalently, using the Fourier basis in the response register, it applies
P f ∣ x , c ⟩ = ω q c f ( x ) ∣ x , c ⟩ , ω q = e 2 π i / q . P_f|x,c\rangle=\omega_q^{cf(x)}|x,c\rangle,\qquad\omega_q=e^{2\pi i/q}. P f ∣ x , c ⟩ = ω q c f ( x ) ∣ x , c ⟩ , ω q = e 2 π i / q .
The two forms are related by free basis changes. A quantum algorithm alternates queries with input-independent unitaries on its accessible registers; internal randomness and measurements may be purified. We count the worst-case number of queries and require success probability at least 2 / 3 2/3 2/3 . Reference registers used to represent input distributions are inaccessible to the algorithm.
Compressed permutation oracles
We use the compressed permutation oracle of Carolan [14 ] in the formulation of Carolan and Majenz [15 ] . Its database register has orthonormal basis ∣ D ⟩ |D\rangle ∣ D ⟩ indexed by partial injections D : [ N ] ⇀ [ N ] D:[N]\rightharpoonup[N] D : [ N ] ⇀ [ N ] . Write ∣ D ∣ = ∣ dom D ∣ |D|=|\operatorname{dom}D| ∣ D ∣ = ∣ dom D ∣ , and let D [ x ↦ y ] D[x\mapsto y] D [ x ↦ y ] denote an extension at an unused input and output. The database starts in ∣ ∅ ⟩ |\varnothing\rangle ∣ ∅ ⟩ . For a fixed query input x x x , deleting its record partitions this basis into the mutually orthogonal spaces
F x , H = span ( { ∣ H ⟩ } ∪ { ∣ H [ x ↦ y ] ⟩ : y ∉ im H } ) , x ∉ dom H . \mathcal{F}_{x,H}=\operatorname{span}\left(\{|H\rangle\}\cup\{|H[x\mapsto y]\rangle:y\notin\operatorname{im}H\}\right),\qquad x\notin\operatorname{dom}H. F x , H = span ( { ∣ H ⟩} ∪ { ∣ H [ x ↦ y ]⟩ : y ∈ / im H } ) , x ∈ / dom H .
On each such space, the compression unitary C x C_x C x exchanges ∣ H ⟩ |H\rangle ∣ H ⟩ with
∣ u x , H ⟩ = 1 N − ∣ H ∣ ∑ y ∉ im H ∣ H [ x ↦ y ] ⟩ (3) |u_{x,H}\rangle=\frac{1}{\sqrt{N-|H|}}\sum_{y\notin\operatorname{im}H}|H[x\mapsto y]\rangle \tag*{(3)} ∣ u x , H ⟩ = N − ∣ H ∣ 1 y ∈ / im H ∑ ∣ H [ x ↦ y ]⟩ (3)
and fixes the orthogonal complement of their span. Thus C x = C x † C_x=C_x^\dagger C x = C x † and C x 2 = I C_x^2=I C x 2 = I . Introduce a blank response symbol ⊥ \perp ⊥ , and let S y S_y S y exchange ∣ ⊥ ⟩ |\perp\rangle ∣ ⊥ ⟩ and ∣ y ⟩ |y\rangle ∣ y ⟩ while fixing all other response states. Let Q x , y Q_{x,y} Q x , y project onto databases with D ( x ) = y D(x)=y D ( x ) = y , and let Q x , ⊥ Q_{x,\perp} Q x , ⊥ project onto those with x ∉ dom D x\notin\operatorname{dom}D x ∈ / dom D . The lookup and compressed forward-swap query are
U x = I ⊗ Q x , ⊥ + ∑ y ∈ [ N ] S y ⊗ Q x , y , Q x = C x U x C x (4) U_x=I\otimes Q_{x,\perp}+\sum_{y\in[N]}S_y\otimes Q_{x,y},\qquad Q_x=C_xU_xC_x \tag*{(4)} U x = I ⊗ Q x , ⊥ + y ∈ [ N ] ∑ S y ⊗ Q x , y , Q x = C x U x C x (4)
where identities on the response register are implicit in C x C_x C x . Coherent queries are controlled by the input register. An inverse-permutation query is obtained by conjugating the corresponding forward query by the database flip ∣ D ⟩ ↦ ∣ D − 1 ⟩ |D\rangle\mapsto|D^{-1}\rangle ∣ D ⟩ ↦ ∣ D − 1 ⟩ . Each query can increase ∣ D ∣ |D| ∣ D ∣ by at most one, so after T T T queries the database is supported on ∣ D ∣ ≤ T |D|\leq T ∣ D ∣ ≤ T .
For a real permutation π \pi π , the corresponding forward-swap query applies S π ( x ) S_{\pi(x)} S π ( x ) to the response. Initializing a reference register in ( N ! ) − 1 / 2 ∑ π ∈ S N ∣ π ⟩ (N!)^{-1/2}\sum_{\pi\in S_N}|\pi\rangle ( N ! ) − 1/2 ∑ π ∈ S N ∣ π ⟩ purifies a uniform random permutation.
Carolan and Majenz [15 ] show that, for T ≤ N T \le\sqrt{N} T ≤ N , the final accessible states in the real and compressed experiments differ in trace distance by O ( T / N ) O(T/\sqrt{N}) O ( T / N ) , even with forward and inverse queries. In particular, the comparison error vanishes for T = o ( N ) T = o(\sqrt{N}) T = o ( N ) . Our algorithms have only function-query access. § 3 uses the forward-swap construction inside complete function-query blocks and extends its database definition to unequal numbers of inputs and slots. The filtered comparison is established in § 4.4 .
Transfer maps and comparison of query histories
Belovs’s transfer operator represents an algorithm under a chosen input distribution [8 Definition 3.4 and Fact 3.5] . Let F = Z q N \mathcal{F} = \mathbb{Z}_q^N F = Z q N , and let p μ , f p_{\mu,f} p μ , f be a probability distribution on pairs consisting of a hidden label μ \mu μ and a truth table f ∈ F f \in\mathcal{F} f ∈ F . The algorithm queries only f f f . Let X \mathcal{X} X and Y \mathcal{Y} Y be the spaces with orthonormal bases { ∣ f ⟩ } \{\lvert f\rangle\} {∣ f ⟩} and { ∣ μ , f ⟩ : p μ , f > 0 } \{\lvert\mu,f\rangle: p_{\mu,f} > 0\} {∣ μ , f ⟩ : p μ , f > 0 } , respectively. Define Υ : X → Y \Upsilon: \mathcal{X} \to\mathcal{Y} Υ : X → Y by
Υ ∣ f ⟩ = q N / 2 ∑ μ p μ , f ∣ μ , f ⟩ . (5) \Upsilon\lvert f\rangle= q^{N/2}\sum_{\mu}\sqrt{p_{\mu,f}}\lvert\mu,f\rangle. \tag*{(5)} Υ ∣ f ⟩ = q N /2 μ ∑ p μ , f ∣ μ , f ⟩ . (5)
Both input oracles are controlled by the same truth table, so O Y Υ = Υ O X O_{\mathcal{Y}}\Upsilon= \Upsilon O_{\mathcal{X}} O Y Υ = Υ O X , with identities on the algorithm registers understood. The map also commutes with algorithm-only unitaries. It therefore sends the entire history with a uniform truth-table prior to the history with prior p p p , while preserving the function-query interface. The transfer need not be an isometry on arbitrary vectors.
For our database comparisons, the corresponding intertwining identity is approximate. Let A t ± A_t^{\pm} A t ± be contractions on two oracle representations, interleaved with the same algorithm-only unitaries U t U_t U t , and let S S S act only on the oracle registers. Write ξ t + 1 ± = A t ± U t ξ t ± \xi_{t+1}^{\pm} = A_t^{\pm}U_t\xi_t^{\pm} ξ t + 1 ± = A t ± U t ξ t ± , assume S ξ 0 − = ξ 0 + S\xi_0^- = \xi_0^+ S ξ 0 − = ξ 0 + , and set Δ t = A t + S − S A t − \Delta_t = A_t^+S - SA_t^- Δ t = A t + S − S A t − . Subtracting the two updates gives
ξ t + 1 + − S ξ t + 1 − = A t + U t ( ξ t + − S ξ t − ) + Δ t U t ξ t − . \xi_{t+1}^+ - S\xi_{t+1}^- = A_t^+U_t(\xi_t^+ - S\xi_t^-) + \Delta_t U_t\xi_t^-. ξ t + 1 + − S ξ t + 1 − = A t + U t ( ξ t + − S ξ t − ) + Δ t U t ξ t − .
Contractivity and induction imply the elementary hybrid estimate
∥ ξ T + − S ξ T − ∥ ≤ ∑ t = 0 T − 1 ∥ Δ t U t ξ t − ∥ . (6) \left\|\xi_T^+ - S\xi_T^-\right\| \le\sum_{t=0}^{T-1}\left\|\Delta_tU_t\xi_t^-\right\|. \tag*{(6)} ξ T + − S ξ T − ≤ t = 0 ∑ T − 1 Δ t U t ξ t − . (6)
The defect can thus be estimated on the evolving source state, without a factor ∥ S ∥ \lVert S\rVert ∥ S ∥ multiplying the accumulated error.
In our application, S = K + J K − S = K^+JK^- S = K + J K − , where J J J is the weighted capacity-extension map of § 5.1 and K ± K^\pm K ± are the harmonic projections. The source is an auxiliary rectangular space of partial injections, with no associated distribution of total functions. The map S S S is constructed directly on these database spaces; its query defect is bounded in Lemma 5.16.
Learning graphs and the dual adversary bound
Let F : X → { 0 , 1 } F : \mathcal{X} \to\{0,1\} F : X → { 0 , 1 } be nonconstant, where X ⊆ Σ m \mathcal{X} \subseteq\Sigma^m X ⊆ Σ m is a finite promise domain. A 1-certificate for a positive input x x x is a restriction x ∣ S x|_S x ∣ S forcing F = 1 F = 1 F = 1 on all consistent promised inputs. A learning graph is a rooted directed acyclic graph whose vertices carry sets S ⊆ [ m ] S \subseteq[m] S ⊆ [ m ] of loaded inputs; distinct vertices may carry the same set. Its root carries ∅ \varnothing ∅ , and each arc e e e loads one input j ∉ S j \notin S j ∈ / S , passing from S S S to S ∪ { j } S \cup\{j\} S ∪ { j } . In the basic model each arc has a positive, input-independent weight w e w_e w e [7 ] . For every positive x x x , choose a nonnegative unit flow p e ( x ) p_e(x) p e ( x ) from the root to vertices containing a 1-certificate for x x x . The flow is conserved at all other non-root vertices. Define
C 0 = ∑ e w e , C 1 ( x ) = ∑ e p e ( x ) 2 w e , C 1 = max x : F ( x ) = 1 C 1 ( x ) . (7) C_0 = \sum_e w_e,\qquad C_1(x) = \sum_e \frac{p_e(x)^2}{w_e},\qquad C_1 = \max_{x:F(x)=1} C_1(x). \tag*{(7)} C 0 = e ∑ w e , C 1 ( x ) = e ∑ w e p e ( x ) 2 , C 1 = x : F ( x ) = 1 max C 1 ( x ) . (7)
These are the negative and positive complexities. The learning-graph upper bound is Q ( F ) = O ( C 0 C 1 ) Q(F) = O(\sqrt{C_0 C_1}) Q ( F ) = O ( C 0 C 1 ) ; equivalently, rescale all weights to balance C 0 C_0 C 0 and C 1 C_1 C 1 and bound their maximum [6 Section 2.2] .
The underlying dual-adversary formulation also covers the adaptive construction used below. For each input position j j j , let X j X_j X j be a positive semidefinite matrix indexed by X \mathcal{X} X . If
∑ j : x j ≠ y j X j [ x , y ] = 1 whenever F ( x ) = 1 and F ( y ) = 0 , (8) \sum_{j:x_j\ne y_j} X_j[x,y] = 1 \qquad\text{whenever } F(x)=1 \text{ and } F(y)=0, \tag*{(8)} j : x j = y j ∑ X j [ x , y ] = 1 whenever F ( x ) = 1 and F ( y ) = 0 , (8)
then the general adversary characterization gives
Q ( F ) = O ( max z ∈ X ∑ j X j [ z , z ] ) . (9) Q(F) = O\left(\max_{z\in\mathcal{X}} \sum_j X_j[z,z]\right). \tag*{(9)} Q ( F ) = O ( z ∈ X max j ∑ X j [ z , z ] ) . (9)
The constant is independent of the finite input alphabet [18 ] . For an ordinary learning graph, local Gram matrices group inputs by their common restriction to the source set S S S . For a positive x x x and a negative y y y , arcs that first reveal a disagreement form a cut separating the root from every accepting vertex. The unit flow across this cut proves (8) .
The same construction permits a weight depending on the source assignment, provided w e ( x ) = w e ( y ) w_e(x)=w_e(y) w e ( x ) = w e ( y ) whenever x ∣ S = y ∣ S x|_S=y|_S x ∣ S = y ∣ S . The colored construction in § 8.2 uses partially revealed assignments, with ⋆ \star ⋆ marking a loaded input whose value is hidden. Inserting a certificate element can uncover earlier entries, so we use the explicit dual formulation, as in Belovs’s fault-correction construction [6 Sections 3 and 6] . There we write X j = ∑ e l o a d i n g j ∑ α Y e , α X_j=\sum_{e\ \mathrm{loading}\ j}\sum_\alpha Y_{e,\alpha} X j = ∑ e loading j ∑ α Y e , α , construct each Y e , α Y_{e,\alpha} Y e , α as a Gram matrix, and verify (8) directly. Signed off-diagonal entries are compatible with positive semidefiniteness. The diagonal estimates then give (9) .
Quantum walk search
Let P P P be a lazy, irreducible, reversible Markov chain on a finite state space V V V , with stationary distribution π \pi π and spectral gap δ = 1 − λ 2 ( P ) > 0 \delta=1-\lambda_2(P)>0 δ = 1 − λ 2 ( P ) > 0 . Reversibility means π ( v ) P ( v , w ) = π ( w ) P ( w , v ) \pi(v)P(v,w)=\pi(w)P(w,v) π ( v ) P ( v , w ) = π ( w ) P ( w , v ) . Let M ⊆ V \mathcal{M}\subseteq V M ⊆ V be the marked set, and suppose that π ( M ) ≥ ε > 0 \pi(\mathcal{M})\ge\varepsilon>0 π ( M ) ≥ ε > 0 whenever M \mathcal{M} M is nonempty. A quantum walk carries a data structure d ( v ) d(v) d ( v ) containing the queried information associated with its current state. Its stationary state, including an edge register, is
∣ π d ⟩ = ∑ v , w ∈ V π ( v ) P ( v , w ) ∣ v , d ( v ) ⟩ ∣ w , d ( w ) ⟩ . |\pi_d\rangle=\sum_{v,w\in V}\sqrt{\pi(v)P(v,w)}\,|v,d(v)\rangle|w,d(w)\rangle. ∣ π d ⟩ = v , w ∈ V ∑ π ( v ) P ( v , w ) ∣ v , d ( v )⟩ ∣ w , d ( w )⟩ .
Write S S S for its preparation cost, U U U for the cost of implementing the coherent update and its inverse, including the associated data operations, and C C C for a phase flip on marked states that restores its ancillary workspace. The quantum walk search theorem of Magniez, Nayak, Roland, and Santha [20 Theorem 3] gives a bounded-error algorithm that finds a marked state using
O ( S + 1 ε ( U δ + C ) ) (10) O\left(S+\frac{1}{\sqrt{\varepsilon}}\left(\frac{U}{\sqrt{\delta}}+C\right)\right) \tag*{(10)} O ( S + ε 1 ( δ U + C ) ) (10)
queries. This is a finding guarantee. If a marked state represents only a partial certificate, extracting and verifying the requested output must also be included in the cost.
Nested updates. An update may itself require a quantum search. The framework of Childs, Jeffery, Kothari, and Magniez [16 Definition 3.2 and Theorem 3.3] permits quantum data depending on the current vertex and the directed edge. It requires coherent local diffusion and a reversible swap of the edge and all its data, including any garbage produced by the inner search. Suppose the inner walks have uniform marked-mass and gap bounds ε ′ , δ ′ > 0 \varepsilon^\prime,\delta^\prime>0 ε ′ , δ ′ > 0 , update and checking costs U ′ , C ′ U^\prime,C^\prime U ′ , C ′ , and the outer local diffusion with garbage and garbage swap cost at most T l o c T_{\mathrm{loc}} T loc . Under these implementation conditions the effective outer update cost is
U = O ~ ( T l o c + 1 ε ′ ( U ′ δ ′ + C ′ ) ) . U=\widetilde{O}\left(T_{\mathrm{loc}}+\frac{1}{\sqrt{\varepsilon^\prime}}\left(\frac{U^\prime}{\sqrt{\delta^\prime}}+C^\prime\right)\right). U = O ( T loc + ε ′ 1 ( δ ′ U ′ + C ′ ) ) .
The setup cost S S S includes preparation of the inner stationary states and any search needed to initialize the marked-state data. Their preparation can be charged once only when the implementation retains and updates them coherently; any later preparation or uncomputation is charged where it occurs. Approximate reflections are implemented with sufficiently small error for the complete computation, contributing the polylogarithmic factors suppressed by O ~ \widetilde{O} O . All costs here count queries to the original function oracle. Appendix A verifies these conditions for the alternative three-collision construction.
The oracle and the two databases
Sampling the function and specifying query access
Let N − = k m N_{-}=km N − = k m and N = N − + 1 N=N_{-}+1 N = N − + 1 . Partition [ N − ] [N_{-}] [ N − ] into labeled buckets B 1 − , … , B m − B_1^{-},\ldots,B_m^{-} B 1 − , … , B m − of size k k k and let β − : [ N − ] → [ m ] \beta_{-}:[N_{-}]\to[m] β − : [ N − ] → [ m ] be their index map. Independently sample a uniform mark b ∈ [ m ] b\in[m] b ∈ [ m ] , a uniform permutation π ∈ S N \pi\in S_N π ∈ S N and a uniform injection a : [ m ] ↪ Z q a:[m]\hookrightarrow\mathbb{Z}_q a : [ m ] ↪ Z q . Add the slot N N N to bucket b b b . Write B i + , b B_i^{+,b} B i + , b for the resulting buckets and β b : [ N ] → [ m ] \beta_b:[N]\to[m] β b : [ N ] → [ m ] for their index map. Define
f a , b , π ( x ) = a ( β b ( π ( x ) ) ) . (11) f_{a,b,\pi}(x)=a\bigl(\beta_b(\pi(x))\bigr). \tag*{(11)} f a , b , π ( x ) = a ( β b ( π ( x )) ) . (11)
Buckets contain slots ; equality classes contain inputs . The equality class with label a ( i ) a(i) a ( i ) is π − 1 ( B i + , b ) \pi^{-1}(B_i^{+,b}) π − 1 ( B i + , b ) . We may describe the fixed partitions to the analyst; the algorithm receives no oracle for π \pi π , its inverse, the mark or the label table.
Lemma 3.1 (Uniformity). The function in (3.1) is uniform over the stipulated exact type for every q ≥ m q\ge m q ≥ m .
Proof. Fix such an f f f . For a specified mark b b b , the value a ( b ) a(b) a ( b ) must be the unique output of multiplicity k + 1 k+1 k + 1 . The other m − 1 m-1 m − 1 labels may be assigned to the remaining buckets in ( m − 1 ) ! (m-1)! ( m − 1 )! ways. For each assignment, the restriction of π \pi π to each equality class can be any bijection onto the corresponding bucket. Hence m ! ( k + 1 ) ! ( k ! ) m − 1 m!(k+1)!(k!)^{m-1} m ! ( k + 1 )! ( k ! ) m − 1 triples ( a , b , π ) (a,b,\pi) ( a , b , π ) give f f f , independently of f f f . All triples are equiprobable. This counting requires only that an injection into the alphabet exist. □ \square □
We condition on the value of a a a while proving a uniform estimate. In particular a i = a ( i ) a_i=a(i) a i = a ( i ) is fixed in the phase calculations below. We can then average the resulting success probabilities over a a a : if each conditional success amplitude is at most L L L , then the averaged success probability is at most L 2 L^2 L 2 .
Lemma 3.2 (Phase-query form and final readout). It suffices to prove the success bound for complete phase queries, arbitrary free unitaries on finite algorithm registers and a final basis readout. A fixed final measurement can be applied to the real and filtered histories without increasing either the number of queries or their comparison error.
Proof. The function oracle is O f ∣ x , z ⟩ = ∣ x , z + f ( x ) ⟩ O_f|x,z\rangle= |x,z+f(x)\rangle O f ∣ x , z ⟩ = ∣ x , z + f ( x )⟩ on [ N ] × Z q [N]\times\mathbb{Z}_q [ N ] × Z q . For c ∈ Z q c\in\mathbb{Z}_q c ∈ Z q put ∣ c ^ ⟩ = q − 1 / 2 ∑ z exp ( − 2 π i c z / q ) ∣ z ⟩ |\hat{c}\rangle= q^{-1/2}\sum_z\exp(-2\pi icz/q)|z\rangle ∣ c ^ ⟩ = q − 1/2 ∑ z exp ( − 2 π i cz / q ) ∣ z ⟩ . Translation by u u u has eigenvalue ω q c u \omega_q^{cu} ω q c u on this vector, where ω q = exp ( 2 π i / q ) \omega_q=\exp(2\pi i/q) ω q = exp ( 2 π i / q ) . Thus one query in the Fourier basis acts diagonally with phase ω q c a ( β b ( π ( x ) ) ) \omega_q^{c a(\beta_b(\pi(x)))} ω q c a ( β b ( π ( x ))) ; an inverse query changes c c c to − c -c − c . The basis changes are free unitaries. More generally, if u u u indexes a workspace basis, a step may query an input x t ( u ) x_t(u) x t ( u ) and multiply by any unit phase z t , u ( i ) z_{t,u}(i) z t , u ( i ) for bucket label i i i . Neither x t x_t x t nor the table of phases depends on b b b or π \pi π . This is the phase model used below. In particular ordinary function queries, even conditioned on the now fixed a a a , are covered.
Purify internal coins and keep any intermediate measurement result in an auxiliary register; a later choice can be controlled by that register. For a final measurement with effects E j ⪰ 0 E_j\succeq0 E j ⪰ 0 , ∑ j E j = I \sum_j E_j=I ∑ j E j = I , the map v ↦ ∑ j ∣ j ⟩ E j v v\mapsto\sum_j|j\rangle\sqrt{E_j}v v ↦ ∑ j ∣ j ⟩ E j v is an isometry, as is seen by taking its squared norm. Extend it to a unitary on a larger workspace, with its ancillas reserved from the start. The resulting basis outcome determines the reported tuple. A failure symbol can be encoded as a fixed tuple with a repeated input. This final unitary commutes with every reference-register encoding and every database-only outcome projector used in the proof. Acting on both histories therefore preserves their distance and the norm on the target event. Branches that stop early can be padded to the query budget with the character c = 0 c=0 c = 0 ; the bound is for the worst-case number of queries.
Finally, for fixed a a a the real reference state starts in ( m N ! ) − 1 / 2 ∑ b , π ∣ b , π ⟩ (mN!)^{-1/2}\sum_{b,\pi}|b,\pi\rangle ( m N ! ) − 1/2 ∑ b , π ∣ b , π ⟩ . Its component in each such sector evolves by precisely the algorithm for (11) . The weights of successful reports in distinct sectors add, since the sectors are orthogonal. Their sum is the average success probability for fixed a a a . Apply Lemma 3.1 after averaging over a a a to obtain the required distributional interpretation. □ \square □
Counters and the two systems. For a partial injection D D D let o i ( D ) = ∣ im D ∩ B i ∣ o_i(D)=|\operatorname{im}D\cap B_i| o i ( D ) = ∣ im D ∩ B i ∣ be the occupancy of bucket i i i . Put
C j ( D ) = ∣ { i : o i ( D ) ≥ j } ∣ , C ^ j ∣ D ⟩ = C j ( D ) ∣ D ⟩ . C_j(D)=|\{i:o_i(D)\ge j\}|,\qquad\widehat{C}_j|D\rangle=C_j(D)|D\rangle. C j ( D ) = ∣ { i : o i ( D ) ≥ j } ∣ , C j ∣ D ⟩ = C j ( D ) ∣ D ⟩ .
Unless a superscript indicates otherwise, moments of these counters will refer to the auxiliary system with N − N_- N − slots. On target branch b b b let P ∗ b P_*^b P ∗ b project onto o b ( D ) = k + 1 o_b(D)=k+1 o b ( D ) = k + 1 , and put P ∗ = ⨁ b P ∗ b P_*=\bigoplus_bP_*^b P ∗ = ⨁ b P ∗ b . This is the event that the full exceptional bucket is recorded. The auxiliary system has the same N N N input positions and N − = N − 1 N_-=N-1 N − = N − 1 slots; it consists of partial injections only. There is no uniform distribution on total injections [ N ] → [ N − ] [N]\to[N_-] [ N ] → [ N − ] .
Partial injections and complete function-query blocks
Let the query-input set be [ N ] [N] [ N ] and let a range set of size R R R be partitioned into buckets B 1 , … , B m B_1,\ldots,B_m B 1 , … , B m of capacities κ 1 , … , κ m \kappa_1,\ldots,\kappa_m κ 1 , … , κ m . The database Hilbert space has orthonormal basis indexed by partial injections D : [ N ] ⇀ [ R ] D:[N]\rightharpoonup[R] D : [ N ] ⇀ [ R ] . For fixed x x x , deleting the record at x x x decomposes the space into orthogonal subspaces
F x , H = span ( { ∣ H ⟩ } ∪ { ∣ H [ x ↦ y ] ⟩ : y ∉ im H } ) , x ∉ dom H . \mathcal{F}_{x,H}=\operatorname{span}\bigl(\{|H\rangle\}\cup\{|H[x\mapsto y]\rangle:y\notin\operatorname{im}H\}\bigr),\qquad x\notin\operatorname{dom}H. F x , H = span ( { ∣ H ⟩} ∪ { ∣ H [ x ↦ y ]⟩ : y ∈ / im H } ) , x ∈ / dom H .
If s = R − ∣ H ∣ > 0 s=R-|H|>0 s = R − ∣ H ∣ > 0 , let C x C_x C x exchange ∣ H ⟩ |H\rangle ∣ H ⟩ with
∣ u x , H ⟩ = 1 s ∑ y ∉ im H ∣ H [ x ↦ y ] ⟩ |u_{x,H}\rangle=\frac{1}{\sqrt{s}}\sum_{y\notin\operatorname{im}H}|H[x\mapsto y]\rangle ∣ u x , H ⟩ = s 1 y ∈ / im H ∑ ∣ H [ x ↦ y ]⟩
and fix the orthogonal complement of their span. If s = 0 s=0 s = 0 , the local space is one-dimensional and we define C x = I C_x=I C x = I there. This defines a self-adjoint unitary on the entire database space, even when R < N R<N R < N .
For a phase character c ∈ Z q c \in\mathbb{Z}_q c ∈ Z q , define the diagonal unitary
L x , c ∣ D ⟩ = { ∣ D ⟩ , x ∉ dom D , ω q c q i ∣ D ⟩ , D ( x ) ∈ B i , ω q = e 2 π i / q . (12) L_{x,c}|D\rangle=
\begin{cases}
|D\rangle, & x \notin\operatorname{dom} D,\\
\omega_q^{c q_i}|D\rangle, & D(x) \in B_i,
\end{cases}
\qquad\omega_q = e^{2\pi i/q}. \tag*{(12)} L x , c ∣ D ⟩ = { ∣ D ⟩ , ω q c q i ∣ D ⟩ , x ∈ / dom D , D ( x ) ∈ B i , ω q = e 2 π i / q . (12)
and define the complete compressed function-query block by
F x , c = C x L x , c C x , F = ∑ x , c ∣ x , c ⟩ ⟨ x , c ∣ ⊗ F x , c . (13) F_{x,c}=C_xL_{x,c}C_x,\qquad F=\sum_{x,c}|x,c\rangle\langle x,c|\otimes F_{x,c}. \tag*{(13)} F x , c = C x L x , c C x , F = x , c ∑ ∣ x , c ⟩ ⟨ x , c ∣ ⊗ F x , c . (13)
The label table can be retained coherently; all these definitions are then controlled by that table.
The forward-swap construction (4) extends verbatim to this rectangular database, with y ∈ [ R ] y \in[R] y ∈ [ R ] . Let D a D_a D a act on the response register as the identity on ∣ ⊥ ⟩ |\perp\rangle ∣ ⊥ ⟩ and by phase ω q c q i \omega_q^{c q_i} ω q c q i on ∣ y ⟩ |y\rangle ∣ y ⟩ when y ∈ B i y \in B_i y ∈ B i .
Lemma 3.3 (Exact simplification of a complete block). On an initially blank response register,
Q x D a Q x ( ∣ ⊥ ⟩ ⊗ v ) = ∣ ⊥ ⟩ ⊗ F x , c v . Q_xD_aQ_x\left(|\perp\rangle\otimes v\right)=|\perp\rangle\otimes F_{x,c}v. Q x D a Q x ( ∣ ⊥ ⟩ ⊗ v ) = ∣ ⊥ ⟩ ⊗ F x , c v . The response register is exactly blank again at the end of the compressed block. One complete block changes at most the record at x x x ; starting from the empty database, the state after t t t complete blocks is supported on ∣ D ∣ ≤ t |D| \le t ∣ D ∣ ≤ t .
Proof. Since C x C_x C x and D a D_a D a act on different registers, Q x D a Q x = C x U x D a U x C x Q_xD_aQ_x=C_xU_xD_aU_xC_x Q x D a Q x = C x U x D a U x C x . For a fixed database, U x D a U x U_xD_aU_x U x D a U x returns ∣ ⊥ ⟩ |\perp\rangle ∣ ⊥ ⟩ to itself and applies precisely the phase in (12) . The local subspace F x , H F_{x,H} F x , H is preserved, proving the remaining assertions. Controlled superpositions of x , c x,c x , c obey the same identities.
The auxiliary system uses R = N − = k m R=N_-=km R = N − = k m and κ i − = k \kappa_i^-=k κ i − = k for all i i i . It has N = N − + 1 N=N_-+1 N = N − + 1 query-input positions. It is an explicitly defined unitary database evolution, not a distribution of total injections [ N ] → [ N − ] [N]\to[N_-] [ N ] → [ N − ] . The one-dimensional convention above supplies a global definition; no terminal subspace is reached when t ≤ N / 3 t\le N/3 t ≤ N /3 . The target uses R = N R=N R = N and capacities κ + , b \kappa^{+,b} κ + , b , with an inaccessible uniform mark register. Its controlled query is denoted F + = ⨁ b F + , b F^+=\bigoplus_b F^{+,b} F + = ⨁ b F + , b , while the auxiliary query is F − F^- F − .
Normalized bucket-orbit bases
The following quotient representation makes the comparison explicit. For a capacity vector κ \kappa κ , let β \beta β be the map sending each range slot to its bucket index. A partial bucket assignment is a map d : [ N ] ⇀ [ m ] d:[N]\rightharpoonup[m] d : [ N ] ⇀ [ m ] with o i ( d ) = ∣ d − 1 ( i ) ∣ ≤ κ i o_i(d)=|d^{-1}(i)|\le\kappa_i o i ( d ) = ∣ d − 1 ( i ) ∣ ≤ κ i . Define
∣ d ‾ ⟩ κ = 1 ∏ i ( κ i ) o i ( d ) ∑ D : β ∘ D = d ∣ D ⟩ , ( a ) h = a ( a − 1 ) ⋯ ( a − h + 1 ) . (14) |\overline{d}\rangle_\kappa=
\frac{1}{\sqrt{\prod_i(\kappa_i)_{o_i(d)}}}
\sum_{D:\,\beta\circ D=d}|D\rangle,
\qquad
(a)_h=a(a-1)\cdots(a-h+1). \tag*{(14)} ∣ d ⟩ κ = ∏ i ( κ i ) o i ( d ) 1 D : β ∘ D = d ∑ ∣ D ⟩ , ( a ) h = a ( a − 1 ) ⋯ ( a − h + 1 ) . (14)
These vectors are orthonormal and span the subspace invariant under permutations of the range slots inside each bucket. The empty state belongs to this subspace. Complete queries and algorithm-side unitaries preserve it: relabeling slots within a bucket permutes the terms of its uniform extension vector and leaves the lookup phase unchanged. The harmonic projector introduced below also preserves it, since its defining deletion conditions are equivariant under slot relabeling. We use this basis for the capacity comparison; the canonical-frame analysis is on the full physical partial-injection space. We write H H H for a partial bucket assignment as well as for its associated orbit basis vector when no confusion arises.
For a fixed deleted assignment H H H , let a i = κ i − o i ( H ) a_i=\kappa_i-o_i(H) a i = κ i − o i ( H ) and s = ∑ i a i s=\sum_i a_i s = ∑ i a i . The quotient local space has basis ∣ H ⟩ |H\rangle ∣ H ⟩ and ∣ H [ x ↦ i ] ⟩ |H[x\mapsto i]\rangle ∣ H [ x ↦ i ]⟩ for a i > 0 a_i>0 a i > 0 . Its uniform extension vector is
u = ∑ i : a i > 0 a i / s ∣ H [ x ↦ i ] ⟩ . (15) u=\sum_{i:a_i>0}\sqrt{a_i/s}\,|H[x\mapsto i]\rangle. \tag*{(15)} u = i : a i > 0 ∑ a i / s ∣ H [ x ↦ i ]⟩ . (15)
Indeed, each extension in bucket i i i has a i a_i a i available range slots, and each extended injection has a unique deletion at x x x . Applying the physical C x C_x C x to (14) gives the operator exchanging ∣ H ⟩ |H\rangle ∣ H ⟩ and (15) . The lookup phase on the i i i th extension is ω q c a ( i ) \omega_q^{ca(i)} ω q c a ( i ) . In the expression for u u u , a i a_i a i denotes the number of unused slots; a ( i ) a(i) a ( i ) is the function label and has no dependence on that count.
A filtered comparison for function queries
Completion frames and exact query identities
A completion vector is a superposition of all permutations extending a database. Synthesis with these vectors is exact on a complete function query, but its inverse on the relevant subspace and the isometric encoding of that subspace are different maps. We construct both and bound their difference before analyzing query errors. The completion-frame construction is due to Carolan and Majenz [15 ] ; the incidence and conditioning arguments below prove the properties needed here.
Completion frames and harmonic databases
We construct two encodings of a permutation state: one is an isometry; the other exactly inverts completion synthesis on its range. We will need both. We give a counting proof of their comparison, using no classification of representations of the symmetric group.
Let S N S_N S N denote the permutations of [ N ] [N] [ N ] . Put P = C [ S N ] \mathcal{P}=\mathbb{C}[S_N] P = C [ S N ] , with orthonormal basis { ∣ π ⟩ : π ∈ S N } \{|\pi\rangle:\pi\in S_N\} { ∣ π ⟩ : π ∈ S N } . A size-h h h database is a matching of h h h distinct inputs to h h h distinct outputs of [ N ] [N] [ N ] . Let L h \mathcal{L}_h L h be the Hilbert space with these databases as an orthonormal basis, and let L h L_h L h be the projector onto this summand in L = ⨁ h = 0 N L h \mathcal{L}=\bigoplus_{h=0}^{N}\mathcal{L}_h L = ⨁ h = 0 N L h . The normalized completion vector of D D D and the synthesis map E h E_h E h are
∣ T D ⟩ = 1 ( N − h ) ! ∑ π ⊇ D ∣ π ⟩ , E h : L h → P , E h ∣ D ⟩ = ∣ T D ⟩ . (16) |T_D\rangle=\frac{1}{\sqrt{(N-h)!}}\sum_{\pi\supseteq D}|\pi\rangle,\qquad E_h:\mathcal{L}_h\to\mathcal{P},\qquad E_h|D\rangle=|T_D\rangle. \tag*{(16)} ∣ T D ⟩ = ( N − h )! 1 π ⊇ D ∑ ∣ π ⟩ , E h : L h → P , E h ∣ D ⟩ = ∣ T D ⟩ . (16)
Indeed, prescribing D D D leaves a bijection between the two sets of N − h N-h N − h unused coordinates undetermined, so there are ( N − h ) ! (N-h)! ( N − h )! completions. Let H ≤ h \mathcal{H}_{\le h} H ≤ h be the span of completion vectors of sizes at most h h h . For ∣ H ∣ = h − 1 |H|=h-1 ∣ H ∣ = h − 1 and x ∉ dom H x\notin\operatorname{dom}H x ∈ / dom H ,
∑ y ∉ im H ∣ T H [ x ↦ y ] ⟩ = N − h + 1 ∣ T H ⟩ ; \sum_{y\notin\operatorname{im}H}|T_{H[x\mapsto y]}\rangle=\sqrt{N-h+1}\,|T_H\rangle; y ∈ / im H ∑ ∣ T H [ x ↦ y ] ⟩ = N − h + 1 ∣ T H ⟩ ;
the analogous identity holds with a fixed unused output and a sum over inputs. Each identity partitions the completions of H H H according to one coordinate. Iteration shows ran E h = H ≤ h \operatorname{ran}E_h=\mathcal{H}_{\le h} ran E h = H ≤ h . Define H ≤ − 1 = 0 \mathcal{H}_{\le-1}=0 H ≤ − 1 = 0 , let Π h \Pi_h Π h project onto the exact degree H h = H ≤ h ⊖ H ≤ h − 1 \mathcal{H}_h=\mathcal{H}_{\le h}\ominus\mathcal{H}_{\le h-1} H h = H ≤ h ⊖ H ≤ h − 1 , and set
A h = Π h E h : L h → H h , G h = A h A h † : H h → H h . (17) A_h=\Pi_hE_h:\mathcal{L}_h\to\mathcal{H}_h,\qquad G_h=A_hA_h^\dagger:\mathcal{H}_h\to\mathcal{H}_h. \tag*{(17)} A h = Π h E h : L h → H h , G h = A h A h † : H h → H h . (17)
A h A_h A h is onto; hence G h G_h G h is positive definite on H h \mathcal{H}_h H h . The filtration is exhaustive already at h = N − 1 h=N-1 h = N − 1 , since specifying N − 1 N-1 N − 1 values specifies the permutation. Thus H N = 0 \mathcal{H}_N=0 H N = 0 .
Call a coefficient vector v ∈ L h v \in\mathcal{L}_h v ∈ L h harmonic if
∑ y ∉ H v H [ x ↦ y ] = 0 , ∑ x ∉ dom H v H [ x ↦ y ] = 0 (18) \sum_{y \notin H} v_{H[x \mapsto y]} = 0,\qquad\sum_{x \notin\operatorname{dom} H} v_{H[x \mapsto y]} = 0 \tag*{(18)} y ∈ / H ∑ v H [ x ↦ y ] = 0 , x ∈ / dom H ∑ v H [ x ↦ y ] = 0 (18)
for every ∣ H ∣ = h − 1 |H| = h-1 ∣ H ∣ = h − 1 and each indicated fresh coordinate. The two conditions are vacuous for h = 0 h = 0 h = 0 . Write K h \mathcal{K}_h K h for the subspace they define. The completion-vector partition above shows immediately that A h † ψ = E h † ψ A_h^{\dagger}\psi= E_h^{\dagger}\psi A h † ψ = E h † ψ is harmonic for every ψ ∈ H h \psi\in\mathcal{H}_h ψ ∈ H h : each row sum is a multiple of the inner product with a lower-degree completion vector. To prove the reverse inclusion and bound G h G_h G h , we now evaluate a small, explicit matrix.
Eliminating coordinates outside a fixed set. For a set of outputs F ⊆ [ N ] F \subseteq[N] F ⊆ [ N ] , suppose v v v is harmonic and is invariant under relabeling the outputs by any permutation fixing F F F pointwise. Fix a database D D D , put j = ∣ im D ∖ F ∣ j = |\operatorname{im} D \setminus F| j = ∣ im D ∖ F ∣ and d F = N − ∣ F ∣ d_F = N - |F| d F = N − ∣ F ∣ . Let E F ( D ) \mathcal{E}_F(D) E F ( D ) be the databases with the same domain as D D D , all outputs in F F F , and the same edges as D D D wherever D D D takes a value in F F F . With ( u ) j = u ( u − 1 ) ⋯ ( u − j + 1 ) (u)_j = u(u-1)\cdots(u-j+1) ( u ) j = u ( u − 1 ) ⋯ ( u − j + 1 ) and ( u ) 0 = 1 (u)_0 = 1 ( u ) 0 = 1 , the first zero-sum identity gives
v D = ( − 1 ) j ( d F ) j ∑ D ′ ∈ E F ( D ) v D ′ . (19) v_D = \frac{(-1)^j}{(d_F)_j}\sum_{D' \in\mathcal{E}_F(D)} v_{D'}. \tag*{(19)} v D = ( d F ) j ( − 1 ) j D ′ ∈ E F ( D ) ∑ v D ′ . (19)
For j = 0 j = 0 j = 0 the sum is v D v_D v D . Otherwise remove any edge whose output lies outside F F F . Of the available replacement outputs, d F − j + 1 d_F-j+1 d F − j + 1 are outside F F F , and every such replacement has coefficient v D v_D v D : a transposition of unused outputs carries it back to the original database. The sum over the available replacements in F F F is therefore − ( d F − j + 1 ) v D -(d_F-j+1)v_D − ( d F − j + 1 ) v D . Apply the induction hypothesis to each of these replacements. The resulting completions partition E F ( D ) \mathcal{E}_F(D) E F ( D ) according to the output assigned to the chosen input, proving (19) . Its denominator is nonzero because j ≤ d F j \le d_F j ≤ d F . Interchanging inputs and outputs gives the corresponding formula for a vector invariant under relabelings fixing a set of inputs.
In particular, if ∣ F ∣ < h |F| < h ∣ F ∣ < h then E F ( D ) \mathcal{E}_F(D) E F ( D ) is always empty, so the only harmonic vector invariant under this stabilizer is zero. For a lower-degree database H H H , average any harmonic v v v over the output permutations fixing F = im H F = \operatorname{im} H F = im H pointwise. Equivariance of completion synthesis means that the averaging does not change ⟨ T H , E h v ⟩ \langle T_H,E_hv\rangle ⟨ T H , E h v ⟩ , since T H T_H T H is fixed. The average remains harmonic and therefore vanishes. We have proved the purity identity
E h v = A h v ∈ H h ( v ∈ K h ) . (20) E_hv = A_hv \in\mathcal{H}_h \qquad(v \in\mathcal{K}_h). \tag*{(20)} E h v = A h v ∈ H h ( v ∈ K h ) . (20)
The Gram matrix on a fixed pair of coordinate sets. Fix S , F ⊆ [ N ] S,F \subseteq[N] S , F ⊆ [ N ] with ∣ S ∣ = ∣ F ∣ = h |S| = |F| = h ∣ S ∣ = ∣ F ∣ = h and put d = N − h d = N-h d = N − h . The base matchings are the bijections G : S → F G:S \to F G : S → F . Suppose v ∈ K h v \in\mathcal{K}_h v ∈ K h is invariant under separate relabelings of inputs fixing S S S pointwise and of outputs fixing F F F pointwise. Write w ( G ) = v G w(G)=v_G w ( G ) = v G for the base coefficients. For any G : S → F G:S \to F G : S → F we claim that
( E h † E h v ) G = ( B w ) ( G ) , (21) \left(E_h^{\dagger}E_hv\right)_G = (Bw)(G), \tag*{(21)} ( E h † E h v ) G = ( B w ) ( G ) , (21)
( B w ) ( G ) = ∑ R ⊆ S ∣ R ∣ ≤ d 1 ( d ) ∣ R ∣ ∑ G ′ : S → F b i j e c t i v e G ′ ∣ S ∖ R = G ∣ S ∖ R w ( G ′ ) . (22) (Bw)(G)=\sum_{\substack{R\subseteq S\\ |R|\le d}}\frac{1}{(d)_{|R|}}\sum_{\substack{G':S\to F\ \mathrm{bijective}\\ G'|_{S\setminus R}=G|_{S\setminus R}}}w(G'). \tag*{(22)} ( B w ) ( G ) = R ⊆ S ∣ R ∣ ≤ d ∑ ( d ) ∣ R ∣ 1 G ′ : S → F bijective G ′ ∣ S ∖ R = G ∣ S ∖ R ∑ w ( G ′ ) . (22)
To see this, fix a permutation π ⊇ G \pi\supseteq G π ⊇ G . Every size-h h h restriction D D D of π \pi π has domain ( S ∖ R ) ∪ U (S\setminus R)\cup U ( S ∖ R ) ∪ U for unique sets R ⊆ S R\subseteq S R ⊆ S and U ⊆ [ N ] ∖ S U\subseteq[N]\setminus S U ⊆ [ N ] ∖ S of equal size j j j . Its edges on S ∖ R S\setminus R S ∖ R agree with G G G , and its values on U U U belong to [ N ] ∖ F [N]\setminus F [ N ] ∖ F . First use (19) to move these j j j values into F F F . The new edges are indexed by the j ! j! j ! bijections U → G ( R ) U \to G(R) U → G ( R ) . For each such choice, apply the input version of (19) to move the same j j j edges onto R R R . The second sum runs exactly once over all G ′ G' G ′ agreeing with G G G on S ∖ R S \setminus R S ∖ R . Thus
v D = j ! ( d ) j 2 ∑ G ′ ∣ S ∖ R = G ∣ S ∖ R w ( G ′ ) . v_D = \frac{j!}{(d)_j^2}\sum_{G'\mid_{S\setminus R}=G\mid_{S\setminus R}} w(G'). v D = ( d ) j 2 j ! G ′ ∣ S ∖ R = G ∣ S ∖ R ∑ w ( G ′ ) .
For fixed R R R , there are ( d j ) \binom{d}{j} ( j d ) choices of U U U , so their total coefficient is 1 / ( d ) j 1/(d)_j 1/ ( d ) j . This proves that ∑ D ⊆ π , ∣ D ∣ = h v D = ( B w ) ( G ) \sum_{D\subseteq\pi,|D|=h}v_D=(Bw)(G) ∑ D ⊆ π , ∣ D ∣ = h v D = ( B w ) ( G ) , independently of the completion π \pi π . Finally, ( E h † E h v ) G (E_h^\dagger E_hv)_G ( E h † E h v ) G is the average of these sums over the d ! d! d ! completions of G G G , by the two factors 1 / d ! 1/\sqrt{d!} 1/ d ! in (4.1). This proves (21) .
For a fixed R R R , the inner sum in (22) acts by an all-ones matrix on each class of bijections with a given restriction to S ∖ R S\setminus R S ∖ R . These blocks are positive semidefinite; R = ∅ R=\varnothing R = ∅ contributes the identity. Hence B ⪰ I B\succeq I B ⪰ I . If h ≤ d h\le d h ≤ d , each row has the same nonnegative sum
∑ j = 0 h ( h j ) j ! ( d ) j = d + 1 d − h + 1 = 1 + h N − 2 h + 1 . (23) \sum_{j=0}^{h}\binom{h}{j}\frac{j!}{(d)_j}
=\frac{d+1}{d-h+1}
=1+\frac{h}{N-2h+1}. \tag*{(23)} j = 0 ∑ h ( j h ) ( d ) j j ! = d − h + 1 d + 1 = 1 + N − 2 h + 1 h . (23)
Indeed, for u j = ( h ) j / ( d ) j u_j=(h)_j/(d)_j u j = ( h ) j / ( d ) j , set v j = ( d − j + 1 ) u j v_j=(d-j+1)u_j v j = ( d − j + 1 ) u j for 0 ≤ j ≤ h 0\le j\le h 0 ≤ j ≤ h and v h + 1 = 0 v_{h+1}=0 v h + 1 = 0 . Then v j − v j + 1 = ( d − h + 1 ) u j v_j-v_{j+1}=(d-h+1)u_j v j − v j + 1 = ( d − h + 1 ) u j , including j = h j=h j = h . Since the matrix is symmetric and nonnegative entrywise, its operator norm is at most this row sum (use a largest coordinate of an eigenvector). The positivity statement itself holds for every h h h .
Lemma 4.1. (Harmonic range and Gram bounds). For every h h h , K h = ran A h † \mathcal{K}_h=\operatorname{ran}A_h^\dagger K h = ran A h † and G h ⪰ I G_h\succeq I G h ⪰ I on H h \mathcal{H}_h H h . For 2 h ≤ N 2h\le N 2 h ≤ N one has
I ⪯ G h ⪯ ( 1 + h N − 2 h + 1 ) I . (24) I\preceq G_h\preceq\left(1+\frac{h}{N-2h+1}\right)I. \tag*{(24)} I ⪯ G h ⪯ ( 1 + N − 2 h + 1 h ) I . (24) In particular G h ⪯ 2 I ⪯ e I G_h\preceq2I\preceq eI G h ⪯ 2 I ⪯ e I if 3 h + 2 ≤ N 3h+2\le N 3 h + 2 ≤ N .
Proof. First suppose that a harmonic vector v v v satisfies E h v = 0 E_hv=0 E h v = 0 . If it is nonzero, choose a matching D 0 D_0 D 0 with v D 0 ≠ 0 v_{D_0}\ne0 v D 0 = 0 . Take S = dom D 0 S=\operatorname{dom}D_0 S = dom D 0 , F = im D 0 F=\operatorname{im}D_0 F = im D 0 and average v v v over the two pointwise stabilizers used above. Their actions preserve harmonicity and commute with synthesis in the sense of relabeling. The resulting vector v ˉ \bar v v ˉ is still annihilated by E h E_h E h and has unchanged coefficient at D 0 D_0 D 0 . Equation (21) gives B w = 0 Bw=0 B w = 0 for its nonzero base vector w w w , contradicting B ⪰ I B\succeq I B ⪰ I . Thus E h E_h E h , hence also A h A_h A h by (20) , is injective on K h \mathcal{K}_h K h . We already proved ran A h † ⊆ K h \operatorname{ran}A_h^\dagger\subseteq\mathcal{K}_h ran A h † ⊆ K h . The orthogonal complement of that range is ker A h \ker A_h ker A h ; no nonzero harmonic vector belongs to this complement. This proves equality.
Now let G h ψ = g ψ G_h\psi=g\psi G h ψ = g ψ with ψ ≠ 0 \psi\ne0 ψ = 0 . Because A h A_h A h is onto, v = A h † ψ ≠ 0 v=A_h^\dagger\psi\ne0 v = A h † ψ = 0 . By purity E h v = A h v = g ψ E_hv=A_hv=g\psi E h v = A h v = g ψ , and E h † ψ = A h † ψ E_h^\dagger\psi=A_h^\dagger\psi E h † ψ = A h † ψ , so E h † E h v = g v E_h^\dagger E_hv=gv E h † E h v = g v . Choose a nonzero coefficient and average as before. The operator E h † E h E_h^\dagger E_h E h † E h commutes with these relabelings, hence B w = g w Bw=gw B w = g w for a nonzero w w w . The lower bound for B B B and, when 2 h ≤ N 2h\le N 2 h ≤ N , its row sum (23) are therefore bounds for every eigenvalue of G h G_h G h . At h = 0 h=0 h = 0 the same argument gives G 0 = I G_0=I G 0 = I . □
The two encodings. Inverses in the following formulas act on H h \mathcal{H}_h H h . Define
W h = A h † G h − 1 , V ~ h = A h † G h − 1 / 2 , K h = W h A h . (25) W_h=A_h^\dagger G_h^{-1},\qquad\widetilde V_h=A_h^\dagger G_h^{-1/2},\qquad K_h=W_hA_h. \tag*{(25)} W h = A h † G h − 1 , V h = A h † G h − 1/2 , K h = W h A h . (25)
K h K_h K h is self-adjoint and idempotent with range ran A h † \operatorname{ran}A_h^\dagger ran A h † ; by Lemma 4.1 it is the orthogonal projector onto the harmonic databases. The polar factor V ~ h \widetilde V_h V h is an isometry, because V ~ h † V ~ h = I \widetilde V_h^\dagger\widetilde V_h=I V h † V h = I , whereas W h = V ~ h G h − 1 / 2 W_h = \widetilde{V}_h G_h^{-1/2} W h = V h G h − 1/2 is a contraction. Define E = ∑ h E h L h E = \sum_h E_hL_h E = ∑ h E h L h , and take the block direct sums A , W , V ~ , K A,W,\widetilde{V},K A , W , V , K of the corresponding degree maps; put Z = I − K Z = I-K Z = I − K . At size N N N take K N = A N = 0 K_N = A_N = 0 K N = A N = 0 . Operators on the database or permutation register act as the identity on the algorithm registers. We have
E W = I , W A = K , A Z = 0 , ( E − A ) K = 0. (26) EW = I,\qquad WA = K,\qquad AZ = 0,\qquad(E-A)K = 0. \tag*{(26)} E W = I , W A = K , A Z = 0 , ( E − A ) K = 0. (26)
Indeed, purity gives E h W h = A h W h = I E_hW_h = A_hW_h = I E h W h = A h W h = I , and the other relations follow from the range of K h K_h K h and purity again. Consequently
T = W E = K + R , R = W ( E − A ) Z , T 2 = T . (27) \mathcal{T} = WE = K + \mathcal{R},\qquad\mathcal{R} = W(E-A)Z,\qquad\mathcal{T}^2 = \mathcal{T}. \tag*{(27)} T = W E = K + R , R = W ( E − A ) Z , T 2 = T . (27)
The identity E W = I EW = I E W = I implies idempotence, but does not imply orthogonality of T \mathcal{T} T .
For Π ≤ s = ∑ h = 0 s Π h \Pi_{\leq s} = \sum_{h=0}^{s}\Pi_h Π ≤ s = ∑ h = 0 s Π h and 2 s ≤ N 2s \leq N 2 s ≤ N the same spectrum gives the endpoint estimate
∥ ( V ~ − W ) Π ≤ s ∥ ≤ s 2 ( N − 2 s + 1 ) . (28) \left\|(\widetilde{V}-W)\Pi_{\leq s}\right\| \leq\frac{s}{2(N-2s+1)}. \tag*{(28)} ( V − W ) Π ≤ s ≤ 2 ( N − 2 s + 1 ) s . (28)
In fact the norm in degree h h h is the maximum of 1 − g − 1 / 2 1-g^{-1/2} 1 − g − 1/2 over its Gram eigenvalues. For g ≥ 1 g \geq1 g ≥ 1 this is at most ( g − 1 ) / 2 (g-1)/2 ( g − 1 ) /2 ; use (24) and maximize over h ≤ s h \leq s h ≤ s . Finally, if a a a is harmonic and supported on sizes h ≤ s h \leq s h ≤ s with 3 s + 2 ≤ N 3s+2 \leq N 3 s + 2 ≤ N , then a = W ψ a = W\psi a = W ψ for ψ = A a \psi= Aa ψ = A a , and ∥ ψ ∥ ≤ 2 ∥ a ∥ ≤ e ∥ a ∥ \|\psi\| \leq\sqrt{2}\|a\| \leq\sqrt{e}\|a\| ∥ ψ ∥ ≤ 2 ∥ a ∥ ≤ e ∥ a ∥ by the same bound.
Table 1. Physical database size and real permutation degree index different spaces.
Exact query identities
Fix an input x x x and phases z y z_y z y indexed by slots y ∈ [ N ] y \in[N] y ∈ [ N ] , with ∣ z y ∣ = 1 |z_y| = 1 ∣ z y ∣ = 1 . On the permutation space let Q y ∣ π ⟩ = 1 { π ( x ) = y } ∣ π ⟩ Q_y|\pi\rangle= \mathbf{1}_{\{\pi(x)=y\}}|\pi\rangle Q y ∣ π ⟩ = 1 { π ( x ) = y } ∣ π ⟩ and U z = ∑ y z y Q y U_z = \sum_y z_yQ_y U z = ∑ y z y Q y . The symbols Q y Q_y Q y here act on permutations, whereas Q x , y d b Q^{\mathrm{db}}_{x,y} Q x , y db below tests a database record. Write M y = C x Q x , y d b C x M_y = C_xQ^{\mathrm{db}}_{x,y}C_x M y = C x Q x , y db C x , and let F z = C x L z C x F_z = C_xL_zC_x F z = C x L z C x be the complete database block in (13) , with slot phases z y z_y z y . All formulas are uniform in these phases.
Lemma 4.2 (Synthesis of one query). On the harmonic subspace,
E F z K = U z E K , W U z ψ = K F z W ψ + R F z W ψ . (29) EF_zK = U_zEK,\qquad WU_z\psi= KF_zW\psi+ RF_zW\psi. \tag*{(29)} E F z K = U z E K , W U z ψ = K F z W ψ + R F z W ψ . (29) The real query and the compressed block have no matrix element between degrees, respectively sizes, differing by more than one.
Proof. In a deleted fiber H H H put s = N − ∣ H ∣ s = N-|H| s = N − ∣ H ∣ , denote its absent coordinate by 0 0 0 , and write a local vector as ( v 0 , ( v y ) y ∉ im H ) (v_0,(v_y)_{y\notin\operatorname{im}H}) ( v 0 , ( v y ) y ∈ / im H ) . Let u = s − 1 / 2 ∑ y ∣ y ⟩ u = s^{-1/2}\sum_{y}|y\rangle u = s − 1/2 ∑ y ∣ y ⟩ be its uniform recorded component. Put w = ∣ 0 ⟩ − u w = |0\rangle-u w = ∣0 ⟩ − u . The compression is the reflection C x = I − w w † C_x=I-ww^\dagger C x = I − w w † . For π ⊇ H \pi\supseteq H π ⊇ H the fiber contributes to the coefficient of E v Ev E v at π \pi π the quantity
( ( N − ∣ H ∣ ) ! ) − 1 / 2 ( v 0 + s v π ( x ) ) . \left((N-\lvert H\rvert)!\right)^{-1/2}\left(v_0+\sqrt{s}\,v_{\pi(x)}\right). ( ( N − ∣ H ∣)! ) − 1/2 ( v 0 + s v π ( x ) ) .
For π ⊉ H \pi\not\supseteq H π ⊇ H the contribution vanishes. The functional in parentheses vanishes on w w w and hence is unchanged by C x C_x C x . If v v v is harmonic, the recorded components in each fiber sum to zero by (18) ; this also holds for a direct sum of homogeneous harmonic vectors. Thus ( C x v ) 0 = 0 (C_xv)_0=0 ( C x v ) 0 = 0 . Applying L z L_z L z after the first C x C_x C x multiplies the functional by exactly z π ( x ) z_{\pi(x)} z π ( x ) . The final C x C_x C x leaves it unchanged. Summing over the fibers proves the first identity. The second follows by applying W W W to the first, since E W = I EW=I E W = I and W E = K + R WE=K+R W E = K + R .
A database fiber consists of just the two sizes ∣ H ∣ , ∣ H ∣ + 1 \lvert H\rvert,\lvert H\rvert+1 ∣ H ∣ , ∣ H ∣ + 1 . On the real side, Q y ∣ T D ⟩ Q_y\lvert T_D\rangle Q y ∣ T D ⟩ is either zero, ∣ T D ⟩ \lvert T_D\rangle ∣ T D ⟩ , or a scalar multiple of ∣ T D [ x ↦ y ] ⟩ \lvert T_D[x\mapsto y]\rangle ∣ T D [ x ↦ y ]⟩ . Thus U z U_z U z and its adjoint both raise the filtration by at most one. Orthogonality to lower degrees then proves the claimed two-sided band bound.
For d = N − h d=N-h d = N − h and fixed x , y x,y x , y , write R x , y ; h 00 R^{00}_{x,y;h} R x , y ; h 00 for the diagonal projector on size-h h h databases missing both x x x and y y y , and R x , y ; h 11 R^{11}_{x,y;h} R x , y ; h 11 for the projector onto those containing the edge ( x , y ) (x,y) ( x , y ) . Define operators on H h \mathcal{H}_h H h by
A y , h = Π h Q y Π h , H y , h 00 = d − 1 A h R x , y ; h 00 A h † , H y , h 11 = A h R x , y ; h 11 A h † . (30) A_{y,h}=\Pi_hQ_y\Pi_h,\qquad H^{00}_{y,h}=d^{-1}A_hR^{00}_{x,y;h}A_h^\dagger,\qquad H^{11}_{y,h}=A_hR^{11}_{x,y;h}A_h^\dagger. \tag*{(30)} A y , h = Π h Q y Π h , H y , h 00 = d − 1 A h R x , y ; h 00 A h † , H y , h 11 = A h R x , y ; h 11 A h † . (30)
A subscript z z z in place of y y y denotes the sum over y y y with coefficients z y z_y z y ; in particular A z , h = Π h U z Π h A_{z,h}=\Pi_hU_z\Pi_h A z , h = Π h U z Π h . Put Q z , h = Π h − 1 U z Π h Q_{z,h}=\Pi_{h-1}U_z\Pi_h Q z , h = Π h − 1 U z Π h and define
D z , h 0 = K h L h F z W h − W h Π h U z Π h , D^0_{z,h}=K_hL_hF_zW_h-W_h\Pi_hU_z\Pi_h, D z , h 0 = K h L h F z W h − W h Π h U z Π h ,
D z , h − = L h − 1 F z W h − W h − 1 Q z , h − . (31) D^-_{z,h}=L_{h-1}F_zW_h-W_{h-1}Q^-_{z,h}. \tag*{(31)} D z , h − = L h − 1 F z W h − W h − 1 Q z , h − . (31)
The second expression is zero at h = 0 h=0 h = 0 .
Lemma 4.3 (The three degree components). For 3 h + 2 ≤ N 3h+2\le N 3 h + 2 ≤ N , the middle and downward corrections are
D z , h 0 = W h ( H z , h 00 G h − 1 + H z , h 11 G h − 1 − A z , h ) , (32) D^0_{z,h}=W_h\left(H^{00}_{z,h}G_h^{-1}+H^{11}_{z,h}G_h^{-1}-A_{z,h}\right), \tag*{(32)} D z , h 0 = W h ( H z , h 00 G h − 1 + H z , h 11 G h − 1 − A z , h ) , (32) D z , h − = W h − 1 ( G h − 1 Q z , h − G h − 1 − Q z , h − ) . (33) D^-_{z,h}=W_{h-1}\left(G_{h-1}Q^-_{z,h}G_h^{-1}-Q^-_{z,h}\right). \tag*{(33)} D z , h − = W h − 1 ( G h − 1 Q z , h − G h − 1 − Q z , h − ) . (33) The upward component has no correction. In particular,
R F z W h = − D z , h 0 − D z , h − . (34) RF_zW_h=-D^0_{z,h}-D^-_{z,h}. \tag*{(34)} R F z W h = − D z , h 0 − D z , h − . (34)
Proof. Take homogeneous harmonic vectors a , c a,c a , c of size h h h . The absent coefficient of C x a C_xa C x a is zero, so on this input F z = ∑ y z y M y F_z=\sum_y z_yM_y F z = ∑ y z y M y . In a fiber with ∣ H ∣ = h \lvert H\rvert=h ∣ H ∣ = h , a a a is a multiple of ∣ H ⟩ \lvert H\rangle ∣ H ⟩ and C x C_x C x sends it to its uniform extensions; measuring slot y y y contributes d − 1 c H ‾ a H d^{-1}\overline{c_H}a_H d − 1 c H a H when that slot is available. In a fiber with ∣ H ∣ = h − 1 \lvert H\rvert=h-1 ∣ H ∣ = h − 1 , both vectors have only recorded components, with sum zero. The compression fixes these vectors, and measuring y y y keeps precisely the term containing ( x , y ) (x,y) ( x , y ) . No other fiber contributes. It follows that
⟨ c , L h M y L h a ⟩ = ⟨ c , ( d − 1 R x , y ; h 00 + R x , y ; h 11 ) a ⟩ . \langle c,L_hM_yL_ha\rangle=\left\langle c,\left(d^{-1}R^{00}_{x,y;h}+R^{11}_{x,y;h}\right)a\right\rangle. ⟨ c , L h M y L h a ⟩ = ⟨ c , ( d − 1 R x , y ; h 00 + R x , y ; h 11 ) a ⟩ .
Test this formula against c = A h † ϕ c=A_h^\dagger\phi c = A h † ϕ , put a = W h ψ = A h † G h − 1 ψ a=W_h\psi=A_h^\dagger G_h^{-1}\psi a = W h ψ = A h † G h − 1 ψ , sum with weights z y z_y z y and use K h = W h A h K_h=W_hA_h K h = W h A h . This gives (32) .
For the downward component let h ≥ 1 h\ge1 h ≥ 1 , ϕ ∈ H h \phi\in\mathcal{H}_h ϕ ∈ H h , and start with a = A h † ϕ a=A_h^\dagger\phi a = A h † ϕ , whose coefficients are ⟨ T D , ϕ ⟩ \langle T_D,\phi\rangle ⟨ T D , ϕ ⟩ . At a size-( h − 1 ) (h-1) ( h − 1 ) matching H H H missing x x x and y y y , the coefficient of M y a M_ya M y a is a H [ x ↦ y ] / N − h + 1 a_{H[x\mapsto y]}/\sqrt{N-h+1} a H [ x ↦ y ] / N − h + 1 . Indeed the first reflection fixes the recorded row of a a a , and the absent coefficient after the last reflection is its uniform average after measuring y y y . This equals ⟨ Q y T H , φ ⟩ \langle Q_yT_H,\varphi\rangle ⟨ Q y T H , φ ⟩ by completion normalization. If y y y is used at another input, both coefficients vanish. If H H H already contains x x x , the physical coefficient is zero because its fiber has no size-h h h member. The real coefficient is zero as well: Q y T H Q_yT_H Q y T H is either zero or a vector of degree at most h − 1 h-1 h − 1 . Thus
L h − 1 M y A h † φ = E h − 1 † Q y φ = A h − 1 † Π h − 1 Q y φ . L_{h-1}M_yA_h^\dagger\varphi=E_{h-1}^\dagger Q_y\varphi=A_{h-1}^\dagger\Pi_{h-1}Q_y\varphi. L h − 1 M y A h † φ = E h − 1 † Q y φ = A h − 1 † Π h − 1 Q y φ .
For the last equality, Q y φ Q_y\varphi Q y φ has no components below degree h − 1 h-1 h − 1 , and a size-( h − 1 ) (h-1) ( h − 1 ) completion column has none above it. Substitute φ = G h − 1 ψ \varphi=G_h^{-1}\psi φ = G h − 1 ψ , sum the phases, and use A h − 1 † = W h − 1 G h − 1 A_{h-1}^\dagger=W_{h-1}G_{h-1} A h − 1 † = W h − 1 G h − 1 . This yields (33) and also proves that the entire physical down block is harmonic.
Finally project the synthesis identity onto degree h + 1 h+1 h + 1 . The sizes below h + 1 h+1 h + 1 have no such component under E E E , and only sizes at most h + 1 h+1 h + 1 occur. Therefore A h + 1 L h + 1 F z W h = Π h + 1 U z Π h A_{h+1}L_{h+1}F_zW_h=\Pi_{h+1}U_z\Pi_h A h + 1 L h + 1 F z W h = Π h + 1 U z Π h . Application of W h + 1 W_{h+1} W h + 1 gives equality of the upward canonical blocks. Subtracting the two remaining blocks from (29) proves (34) .
We will repeatedly use the following elementary inequality: for positive operators B y B_y B y and scalars ∣ z y ∣ ≤ 1 |z_y|\le1 ∣ z y ∣ ≤ 1 ,
∥ ∑ y z y B y ∥ ≤ ∥ ∑ y B y ∥ . (35) \left\|\sum_y z_yB_y\right\|\le\left\|\sum_y B_y\right\|. \tag*{(35)} y ∑ z y B y ≤ y ∑ B y . (35)
Indeed, set S ( ( v y ) y ) = ∑ y B y 1 / 2 v y S((v_y)_y)=\sum_y B_y^{1/2}v_y S (( v y ) y ) = ∑ y B y 1/2 v y . The two sums are respectively S diag ( z y ) S † S\operatorname{diag}(z_y)S^\dagger S diag ( z y ) S † and S S † SS^\dagger S S † .
Conditioning on a permutation entry
We need two estimates for the assertion π ( x ) = y \pi(x)=y π ( x ) = y . On states whose harmonic records know the input x x x , a normalized completion frame is close to the identity. On the complementary states we can delete x x x and y y y and compare the resulting smaller completion frame with the original one. The first estimate uses the energy of a masked vector under deletion; the second uses insertion into a permutation.
A fixed input in a harmonic database
Fix an input x x x . A completion vector can have some database coefficients containing x x x and others missing x x x . Simply keeping the former need not preserve harmonicity. We first quantify this loss in the database basis, and then transfer the estimate to the real permutation space.
The deletion gap. For positive integers a a a , b b b and h ≤ min ( a , b ) h\le\min(a,b) h ≤ min ( a , b ) , let L a , b , h \mathcal{L}_{a,b,h} L a , b , h have as orthonormal basis the size-h h h partial injections from [ a ] [a] [ a ] to [ b ] [b] [ b ] . For h > 0 h>0 h > 0 define the normalized row maps B d B_d B d and B i B_i B i by
( B d v ) H , u = ( b − h + 1 ) − 1 / 2 ∑ y ∉ im H v H [ u ↦ y ] , ( B i v ) H , y = ( a − h + 1 ) − 1 / 2 ∑ u ∉ dom H v H [ u ↦ y ] . (B_dv)_{H,u}=(b-h+1)^{-1/2}\sum_{y\notin\operatorname{im}H}v_{H[u\mapsto y]},\qquad
(B_iv)_{H,y}=(a-h+1)^{-1/2}\sum_{u\notin\operatorname{dom}H}v_{H[u\mapsto y]}. ( B d v ) H , u = ( b − h + 1 ) − 1/2 y ∈ / im H ∑ v H [ u ↦ y ] , ( B i v ) H , y = ( a − h + 1 ) − 1/2 u ∈ / dom H ∑ v H [ u ↦ y ] .
Here ∣ H ∣ = h − 1 |H|=h-1 ∣ H ∣ = h − 1 , and the coordinate named in the row is absent from H H H . Write ∂ = ( B d , B i ) \partial=(B_d,B_i) ∂ = ( B d , B i ) and let K a , b , h K_{a,b,h} K a , b , h project onto its kernel. At h = 0 h=0 h = 0 take ∂ = 0 \partial=0 ∂ = 0 and K a , b , 0 = I K_{a,b,0}=I K a , b , 0 = I . In the square space a = b = N a=b=N a = b = N this harmonic projector is K h K_h K h from the completion construction.
Lemma 4.4 (Deletion gap). Let 1 ≤ h ≤ min ( a , b ) 1\le h\le\min(a,b) 1 ≤ h ≤ min ( a , b ) and 2 h ≤ min ( a , b ) + 1 2h\le\min(a,b)+1 2 h ≤ min ( a , b ) + 1 . Then
∥ ∂ v ∥ 2 ≥ min { a − 2 h + 2 a − h + 1 , b − 2 h + 2 b − h + 1 } ∥ ( I − K a , b , h ) v ∥ 2 . (36) \lVert\partial v\rVert^2\ge\min\left\{\frac{a-2h+2}{a-h+1},\frac{b-2h+2}{b-h+1}\right\}\lVert(I-K_{a,b,h})v\rVert^2. \tag*{(36)} ∥ ∂ v ∥ 2 ≥ min { a − h + 1 a − 2 h + 2 , b − h + 1 b − 2 h + 2 } ∥( I − K a , b , h ) v ∥ 2 . (36)
Proof. Put s = b − h + 1 s=b-h+1 s = b − h + 1 , t = a − h + 1 t=a-h+1 t = a − h + 1 , D = B d † B d D=B_d^\dagger B_d D = B d † B d and F = B i † B i F=B_i^\dagger B_i F = B i † B i . For an input u u u , the domain-deletion rows with missing input u u u have disjoint supports and unit norm: each has exactly s s s extensions, and deleting the record at u u u recovers its unique row. Their Gram on the database space is consequently an orthogonal projector D u D_u D u , and D = ∑ u D u D=\sum_u D_u D = ∑ u D u . Let M u M_u M u be the diagonal mask for u ∈ dom D ′ u\in\operatorname{dom}D' u ∈ dom D ′ on a basis database D ′ D' D ′ . Thus M u D u = D u M_uD_u=D_u M u D u = D u .
To bound D 2 D^2 D 2 , consider two distinct inputs u , v u,v u , v . Both D u w D_uw D u w and D v w D_vw D v w can contribute to their inner product only on matchings containing both inputs. Remove their two records. In each resulting fiber the two remaining outputs range over distinct pairs i , j i,j i , j of an ( s + 1 ) (s+1) ( s + 1 ) -element set. Write w i j w_{ij} w ij for the corresponding coefficients. On the fiber, D u w D_uw D u w has the constant value c j = s − 1 ∑ i ≠ j w i j c_j=s^{-1}\sum_{i\ne j}w_{ij} c j = s − 1 ∑ i = j w ij in column j j j , while D v w D_vw D v w has the constant value r i = s − 1 ∑ j ≠ i w i j r_i=s^{-1}\sum_{j\ne i}w_{ij} r i = s − 1 ∑ j = i w ij in row i i i . Since ∑ i r i = ∑ j c j \sum_i r_i=\sum_j c_j ∑ i r i = ∑ j c j , their inner product on the fiber is
∑ i ≠ j c j ‾ r i = ∣ ∑ i r i ∣ 2 − ∑ i c i ‾ r i . \sum_{i\ne j}\overline{c_j}r_i=\left|\sum_i r_i\right|^2-\sum_i\overline{c_i}r_i. i = j ∑ c j r i = i ∑ r i 2 − i ∑ c i r i .
Its real part is at least − 1 2 ∑ i ( ∣ c i ∣ 2 + ∣ r i ∣ 2 ) -\frac{1}{2}\sum_i(|c_i|^2+|r_i|^2) − 2 1 ∑ i ( ∣ c i ∣ 2 + ∣ r i ∣ 2 ) . The squared norms on the fiber are s ∑ i ∣ c i ∣ 2 s\sum_i|c_i|^2 s ∑ i ∣ c i ∣ 2 and s ∑ i ∣ r i ∣ 2 s\sum_i|r_i|^2 s ∑ i ∣ r i ∣ 2 . Summing the fiber inequalities gives
2 Re ⟨ D u w , D v w ⟩ ≥ − s − 1 ( ∥ M v D u w ∥ 2 + ∥ M u D v w ∥ 2 ) . 2\operatorname{Re}\langle D_uw,D_vw\rangle\ge-s^{-1}\left(\lVert M_vD_uw\rVert^2+\lVert M_uD_vw\rVert^2\right). 2 Re ⟨ D u w , D v w ⟩ ≥ − s − 1 ( ∥ M v D u w ∥ 2 + ∥ M u D v w ∥ 2 ) .
For h = 1 h=1 h = 1 both sides vanish because the common support is empty. Each matching in the support of D u w D_uw D u w has exactly h − 1 h-1 h − 1 other occupied inputs. Expanding ∥ ∑ u D u w ∥ 2 \lVert\sum_uD_uw\rVert^2 ∥ ∑ u D u w ∥ 2 and summing over unordered pairs therefore yields
∥ D w ∥ 2 ≥ ( 1 − h − 1 s ) ∑ u ∥ D u w ∥ 2 = ( 1 − h − 1 s ) ⟨ w , D w ⟩ . \lVert Dw\rVert^2\ge\left(1-\frac{h-1}{s}\right)\sum_u\lVert D_uw\rVert^2=\left(1-\frac{h-1}{s}\right)\langle w,Dw\rangle. ∥ D w ∥ 2 ≥ ( 1 − s h − 1 ) u ∑ ∥ D u w ∥ 2 = ( 1 − s h − 1 ) ⟨ w , D w ⟩ .
Thus every positive eigenvalue of D D D is at least 1 − ( h − 1 ) / s 1-(h-1)/s 1 − ( h − 1 ) / s . Transposing matchings exchanges inputs and outputs and gives the analogous lower bound 1 − ( h − 1 ) / t 1-(h-1)/t 1 − ( h − 1 ) / t for F F F .
It remains to check that these two gaps apply to their joint kernel. Let T o u t T_{\mathrm{out}} T out and T i n T_{\mathrm{in}} T in sum, respectively, the actions on matchings of all unordered output and input transpositions. On a matching, a transposition of two unused coordinates fixes it; one between a used and an unused coordinate replaces one image or one input. Swapping two used outputs gives the same matching as swapping their two preimages. The used–used terms thus cancel in the difference. The operator s D sD sD has diagonal h h h and off-diagonal entry one precisely for one replacement of an output by an unused one; this follows directly by counting shared deletion rows. Likewise t F tF tF describes input replacements. Consequently
T o u t − T i n = s D − t F + [ ( b − h 2 ) − ( a − h 2 ) ] I . T_{\mathrm{out}}-T_{\mathrm{in}}=sD-tF+\left[\binom{b-h}{2}-\binom{a-h}{2}\right]I. T out − T in = sD − tF + [ ( 2 b − h ) − ( 2 a − h ) ] I .
Relabeling either coordinates permutes the rows of B d B_d B d and the columns by the induced relabeling of the database, so D D D commutes with both transposition sums. The identity gives [ D , F ] = 0 [D,F]=0 [ D , F ] = 0 . In a common orthonormal eigenbasis the nonzero eigenvalues of D + F D+F D + F are sums with at least one positive summand. Its kernel is exactly ker δ \ker\delta ker δ , since ⟨ v , ( D + F ) v ⟩ = ∥ δ v ∥ 2 \langle v,(D+F)v\rangle=\lVert\delta v\rVert^2 ⟨ v , ( D + F ) v ⟩ = ∥ δ v ∥ 2 . This proves (36) . □ \square □
Masking one input. From now on a = b = N a=b=N a = b = N . On L h = L N , N , h \mathcal{L}_h=\mathcal{L}_{N,N,h} L h = L N , N , h , use M x M_x M x for the input mask above and M y M^y M y for the mask of y ∈ im D y\in\operatorname{im}D y ∈ im D . Their harmonic compressions are self-adjoint positive contractions. Only the input compression needs the following stronger bound.
Lemma 4.5 (Harmonic occupancy mask). If 2 h ≤ N 2h \le N 2 h ≤ N , the spectrum of M x = K h M x ∣ ran K h M_x = K_hM_x|_{\operatorname{ran} K_h} M x = K h M x ∣ ran K h lies in
{ 0 } ∪ [ 1 − 1 N − 2 h + 2 , 1 ] . \{0\} \cup\left[1-\frac{1}{N-2h+2},1\right]. { 0 } ∪ [ 1 − N − 2 h + 2 1 , 1 ] . Its kernel on ran K h \operatorname{ran} K_h ran K h consists precisely of the harmonic vectors supported on databases missing x x x .
Proof. At h = 0 h=0 h = 0 the mask vanishes. Let h ≥ 1 h \ge1 h ≥ 1 , s = N − h + 1 s=N-h+1 s = N − h + 1 and K h v = v K_hv=v K h v = v . On any domain-deletion row the occupancy of x x x is constant across the row, so B d M x v = 0 B_dM_xv=0 B d M x v = 0 . On an image-deletion row ( H , y ) (H,y) ( H , y ) with x ∈ dom H x \in\operatorname{dom} H x ∈ dom H the row sum likewise vanishes by harmonicity. If x ∉ dom H x \notin\operatorname{dom} H x ∈ / dom H , exactly one term remains, v H [ x ↦ y ] / s v_{H[x\mapsto y]}/\sqrt{s} v H [ x ↦ y ] / s . Deleting the record at x x x counts each matching containing x x x exactly once. Thus
∥ ∂ M x v ∥ 2 = s − 1 ∥ M x v ∥ 2 , ∥ ( I − K h ) M x v ∥ 2 ≤ ∥ M x v ∥ 2 N − 2 h + 2 . (37) \lVert\partial M_xv\rVert^2=s^{-1}\lVert M_xv\rVert^2,\qquad\lVert(I-K_h)M_xv\rVert^2\le\frac{\lVert M_xv\rVert^2}{N-2h+2}. \tag*{(37)} ∥ ∂ M x v ∥ 2 = s − 1 ∥ M x v ∥ 2 , ∥( I − K h ) M x v ∥ 2 ≤ N − 2 h + 2 ∥ M x v ∥ 2 . (37) where the second inequality is Lemma 4.4. The two quadratic forms ∥ M x v ∥ 2 \lVert M_xv\rVert^2 ∥ M x v ∥ 2 and ∥ K h M x v ∥ 2 \lVert K_hM_xv\rVert^2 ∥ K h M x v ∥ 2 equal ⟨ v , M x v ⟩ \langle v,M_xv\rangle ⟨ v , M x v ⟩ and ⟨ v , M x 2 v ⟩ \langle v,M_x^2v\rangle ⟨ v , M x 2 v ⟩ , respectively. Subtracting the loss in (37) gives M x 2 ⪰ ( 1 − 1 / ( N − 2 h + 2 ) ) M x M_x^2 \succeq(1-1/(N-2h+2))M_x M x 2 ⪰ ( 1 − 1/ ( N − 2 h + 2 )) M x . Its eigenvalues are in [ 0 , 1 ] [0,1] [ 0 , 1 ] , proving the assertion. The first quadratic form also proves the kernel statement.
Frames on the permutation space. A masked completion frame has two indices that must be kept separate: the size s s s of the databases used to form it, and the degree j j j of the vectors to which it is applied. Define the following positive operators on P = C [ S N ] \mathcal{P}=\mathbb{C}[S_N] P = C [ S N ] :
J s , x = E s M x E s † , O s , y = E s M y E s † , G s = E s E s † . J_{s,x}=E_sM_xE_s^{\dagger},\qquad O_{s,y}=E_sM^yE_s^{\dagger},\qquad G_s=E_sE_s^{\dagger}. J s , x = E s M x E s † , O s , y = E s M y E s † , G s = E s E s † .
Write Q y ∣ π ⟩ = 1 { π ( x ) = y } ∣ π ⟩ Q_y|\pi\rangle=\mathbf{1}_{\{\pi(x)=y\}}|\pi\rangle Q y ∣ π ⟩ = 1 { π ( x ) = y } ∣ π ⟩ for a fixed x x x .
Lemma 4.6 (Fixed-coordinate frames reduce the degrees). Each of these frames commutes with every degree projector Π j \Pi_j Π j . The full frame G s G_s G s commutes with both masked frames, and J s , x J_{s,x} J s , x commutes with O s ′ , y O_{s',y} O s ′ , y for any sizes s , s ′ s,s' s , s ′ . Moreover [ J s , x , Q y ] = 0 [J_{s,x},Q_y]=0 [ J s , x , Q y ] = 0 for every y y y .
Proof. The ( π , σ ) (\pi,\sigma) ( π , σ ) matrix entry of a frame is ( N − s ) ! − 1 (N-s)!^{-1} ( N − s ) ! − 1 times the number of selected s s s -matchings contained in both permutations. For J s , x J_{s,x} J s , x this count is unchanged if both permutations are multiplied on the left by the same permutation: this relabels outputs and does not change which matchings contain x x x . Its kernel therefore has the form j ( π − 1 σ ) j(\pi^{-1}\sigma) j ( π − 1 σ ) , and the action on a function f f f is ∑ τ j ( τ ) f ( π τ ) \sum_{\tau}j(\tau)f(\pi\tau) ∑ τ j ( τ ) f ( π τ ) . For O s , y O_{s,y} O s , y the kernel is invariant under common right multiplication instead; its action is a linear combination of left translations. For G s G_s G s both invariances hold. Since left and right translations commute, this proves all the asserted frame commutations. Each translation sends a completion column of degree ≤ j \le j ≤ j to a completion column of the same size and is unitary, so it preserves both H ≤ j \mathcal{H}_{\le j} H ≤ j and H ≤ j − 1 ⊥ \mathcal{H}_{\le j-1}^{\perp} H ≤ j − 1 ⊥ , and hence H j \mathcal{H}_j H j . Finally, two permutations sharing a matching containing x x x must agree on x x x . Thus the kernel of J s , x J_{s,x} J s , x has no entries between distinct values of π ( x ) \pi(x) π ( x ) , which proves its commutation with every Q y Q_y Q y .
Restrict the size-h h h frames to H h \mathcal{H}_h H h , writing
J h , x = A h M x A h † , O h , y = A h M y A h † , G h = A h A h † . J_{h,x}=A_hM_xA_h^{\dagger},\qquad O_{h,y}=A_hM^yA_h^{\dagger},\qquad G_h=A_hA_h^{\dagger}. J h , x = A h M x A h † , O h , y = A h M y A h † , G h = A h A h † .
By Lemma 4.6 these are actual restrictions, not just compressions. Set F h , x = ker J h , x \mathcal{F}_{h,x}=\ker J_{h,x} F h , x = ker J h , x (the fresh space) and call F h , x ⊥ \mathcal{F}_{h,x}^{\perp} F h , x ⊥ the known space. Both are subspaces of H h \mathcal{H}_h H h . Indeed, ⟨ ψ , J h , x ψ ⟩ = ∥ M x A h † ψ ∥ 2 \langle\psi,J_{h,x}\psi\rangle=\lVert M_xA_h^{\dagger}\psi\rVert^2 ⟨ ψ , J h , x ψ ⟩ = ∥ M x A h † ψ ∥ 2 : freshness means exactly that no coefficient of the harmonic analysis A h † ψ A_h^{\dagger}\psi A h † ψ records x x x . The word “known” refers to its orthogonal complement, not to the unprojected mask M x M_x M x .
Lemma 4.7 (Normalized frame on the known space). For 2 h ≤ N 2h \le N 2 h ≤ N , put B h , x = J h , x G h − 1 B_{h,x} = J_{h,x}G_h^{-1} B h , x = J h , x G h − 1 . It vanishes on F h , x \mathcal{F}_{h,x} F h , x and is a positive contraction, with
∥ ( I − B h , x ) ∣ F h , x ⊥ ∥ ≤ ( N − 2 h + 2 ) − 1 . (38) \left\|(I-B_{h,x})|_{\mathcal{F}_{h,x}^{\perp}}\right\| \le(N-2h+2)^{-1}. \tag*{(38)} ( I − B h , x ) ∣ F h , x ⊥ ≤ ( N − 2 h + 2 ) − 1 . (38) The degree-space operators G h ± 1 G_h^{\pm1} G h ± 1 , O h , y G h − 1 O_{h,y}G_h^{-1} O h , y G h − 1 and A y , h = Π h Q y ∣ H h A_{y,h} = \Pi_h Q_y|_{\mathcal{H}_h} A y , h = Π h Q y ∣ H h preserve both summands. On the harmonic database space A h † A h A_h^{\dagger}A_h A h † A h commutes with M x M_x M x and with M y = K h M y ∣ ran K h M^y = K_hM^y|_{\operatorname{ran} K_h} M y = K h M y ∣ ran K h ; the latter two operators also commute. For a harmonic v v v , one has
E h K h M x v = E h M x v . (39) E_hK_hM_xv = E_hM_xv. \tag*{(39)} E h K h M x v = E h M x v . (39) Proof. By Lemma 4.6, J h , x J_{h,x} J h , x commutes with G h G_h G h , O h , y O_{h,y} O h , y , and A y , h A_{y,h} A y , h . For the last assertion use the commutation of the full J h , x J_{h,x} J h , x with both Q y Q_y Q y and Π h \Pi_h Π h . Since these frames are self-adjoint, their commuting operators preserve both ker J h , x \ker J_{h,x} ker J h , x and its orthogonal complement; so do G h − 1 G_h^{-1} G h − 1 and O h , y G h − 1 O_{h,y}G_h^{-1} O h , y G h − 1 . In particular the positive operator J h , x G h − 1 J_{h,x}G_h^{-1} J h , x G h − 1 has the same kernel as J h , x J_{h,x} J h , x . Using K h = A h † G h − 1 A h K_h = A_h^{\dagger}G_h^{-1}A_h K h = A h † G h − 1 A h gives the intertwining identity
M x A h † = A h † G h − 1 J h , x = A h † B h , x . (40) M_xA_h^{\dagger} = A_h^{\dagger}G_h^{-1}J_{h,x} = A_h^{\dagger}B_{h,x}. \tag*{(40)} M x A h † = A h † G h − 1 J h , x = A h † B h , x . (40) The map A h † A_h^{\dagger} A h † is injective from H h \mathcal{H}_h H h onto ran K h \operatorname{ran} K_h ran K h . Thus B h , x B_{h,x} B h , x has the spectrum in Lemma 4.5, which gives (38) . With outputs in place of inputs the same calculation intertwines M y M^y M y with O h , y G h − 1 O_{h,y}G_h^{-1} O h , y G h − 1 ; in particular the latter is also a positive contraction. Moreover ( A h † A h ) A h † = A h † G h (A_h^{\dagger}A_h)A_h^{\dagger} = A_h^{\dagger}G_h ( A h † A h ) A h † = A h † G h . Transfer the commutations of G h G_h G h , B h , x B_{h,x} B h , x and O h , y G h − 1 O_{h,y}G_h^{-1} O h , y G h − 1 across the same invertible map to obtain all three database-space commutations. Finally write v = A h † ψ = E h † ψ v = A_h^{\dagger}\psi= E_h^{\dagger}\psi v = A h † ψ = E h † ψ with ψ ∈ H h \psi\in\mathcal{H}_h ψ ∈ H h . Then E h M x v = J h , x ψ ∈ H h E_hM_xv = J_{h,x}\psi\in\mathcal{H}_h E h M x v = J h , x ψ ∈ H h . On the other hand E h K h = A h E_hK_h = A_h E h K h = A h (synthesis on the harmonic space is pure of degree h h h , and A h K h = A h A_hK_h = A_h A h K h = A h ). Thus E h K h M x v = Π h E h M x v = E h M x v E_hK_hM_xv = \Pi_hE_hM_xv = E_hM_xv E h K h M x v = Π h E h M x v = E h M x v . This last identity does not assert that M x v M_xv M x v itself is harmonic.
Conditioning on a fresh input and output
The estimate on the known subspace concerns a recorded input. On its orthogonal complement we need to understand what a literal assertion π ( x ) = y \pi(x)=y π ( x ) = y sees. The complement of this one input and one output is a smaller permutation problem. The completion vectors in the two problems have different normalizations; keeping those normalizations will give both a relative bound and the exact commutation needed for the downward query.
Fix x , y ∈ [ N ] x,y \in[N] x , y ∈ [ N ] , put n = N − 1 n=N-1 n = N − 1 and d = N − h d=N-h d = N − h , and assume 2 h ≤ n 2h \le n 2 h ≤ n . The small input and output sets are X = [ N ] ∖ { x } X=[N]\setminus\{x\} X = [ N ] ∖ { x } and Y = [ N ] ∖ { y } Y=[N]\setminus\{y\} Y = [ N ] ∖ { y } . Identifying them separately with [ n ] [n] [ n ] identifies their bijections with S n S_n S n . Superscripts ( n ) (n) ( n ) and ( N ) (N) ( N ) indicate which universe is meant. Write K h ( S ) = ran K h ( S ) \mathcal{K}_h^{(S)}=\operatorname{ran}K_h^{(S)} K h ( S ) = ran K h ( S ) for the harmonic coefficient space and let ι : L h ( n ) → L h ( N ) \iota:\mathcal{L}_h^{(n)}\to\mathcal{L}_h^{(N)} ι : L h ( n ) → L h ( N ) extend coefficients by zero on matchings that use x x x or y y y . Thus ι † \iota^\dagger ι † simply discards those coordinates. On K h ( n ) \mathcal{K}_h^{(n)} K h ( n ) define
B = ( A h ( n ) † ) A h ( n ) ∣ K h ( n ) , S = E h ( N ) † ι ∣ K h ( n ) , B ^ = S † S . (41) B=(A_h^{(n)\dagger})A_h^{(n)}|_{\mathcal{K}_h^{(n)}},\qquad S=E_h^{(N)\dagger}\iota|_{\mathcal{K}_h^{(n)}},\qquad\widehat{B}=S^\dagger S. \tag*{(41)} B = ( A h ( n ) † ) A h ( n ) ∣ K h ( n ) , S = E h ( N ) † ι ∣ K h ( n ) , B = S † S . (41)
In particular, B B B is a Gram operator on coefficients; G h ( N ) G_h^{(N)} G h ( N ) acts on permutation amplitudes. Put M u = 1 { u ∈ dom D } M_u=\mathbf{1}_{\{u\in\operatorname{dom}D\}} M u = 1 { u ∈ dom D } on the small database, and let M u = K h ( n ) M u ∣ K h ( n ) M_u=K_h^{(n)}M_u|_{\mathcal{K}_h^{(n)}} M u = K h ( n ) M u ∣ K h ( n ) .
Lemma 4.8 (Insertion identity). For every v ∈ L h ( n ) v \in\mathcal{L}^{(n)}_h v ∈ L h ( n ) ,
d ∥ E h ( N ) v ∥ 2 = ∥ E h ( n ) v ∥ 2 + ∑ u ∈ X ∥ E h ( n ) ( I − M u ) v ∥ 2 . (42) d\left\lVert E^{(N)}_h v\right\rVert^2 = \left\lVert E^{(n)}_h v\right\rVert^2 + \sum_{u \in X}\left\lVert E^{(n)}_h(I-M_u)v\right\rVert^2. \tag*{(42)} d E h ( N ) v 2 = E h ( n ) v 2 + u ∈ X ∑ E h ( n ) ( I − M u ) v 2 . (42) On the small harmonic space this gives the operator identity
B ^ = B − 1 d ∑ u ∈ X B ( M u − M u 2 ) . (43) \widehat{B}=B-\frac{1}{d}\sum_{u\in X}B(M_u-M_u^2). \tag*{(43)} B = B − d 1 u ∈ X ∑ B ( M u − M u 2 ) . (43) Consequently B ^ B = B B ^ \widehat{B}B=B\widehat{B} B B = B B and
( 1 − ρ h ) B ⪯ B ^ ⪯ B , ρ h = h d ( n − 2 h + 2 ) . (44) (1-\rho_h)B\preceq\widehat{B}\preceq B,\qquad\rho_h=\frac{h}{d(n-2h+2)}. \tag*{(44)} ( 1 − ρ h ) B ⪯ B ⪯ B , ρ h = d ( n − 2 h + 2 ) h . (44)
Proof. A large permutation is uniquely specified by a pair ( u , σ ) (u,\sigma) ( u , σ ) , where u ∈ X ∪ { x } u\in X\cup\{x\} u ∈ X ∪ { x } and σ : X → Y \sigma:X\to Y σ : X → Y is a bijection. If u = x u=x u = x , extend σ \sigma σ by x ↦ y x\mapsto y x ↦ y . Otherwise replace the edge u ↦ σ ( u ) u\mapsto\sigma(u) u ↦ σ ( u ) by u ↦ y u\mapsto y u ↦ y and x ↦ σ ( u ) x\mapsto\sigma(u) x ↦ σ ( u ) . The inverse reads u = π − 1 ( y ) u=\pi^{-1}(y) u = π − 1 ( y ) , deletes x , y x,y x , y , and restores u ↦ π ( x ) u\mapsto\pi(x) u ↦ π ( x ) if u ≠ x u\ne x u = x . A small matching D D D is completed by this π \pi π precisely when σ ⊇ D \sigma\supseteq D σ ⊇ D and u ∉ dom D u\notin\operatorname{dom}D u ∈ / dom D . Also ( N − h ) ! = d ( n − h ) ! (N-h)!=d(n-h)! ( N − h )! = d ( n − h )! . Thus the coefficient of E h ( N ) v E^{(N)}_h v E h ( N ) v at ( u , σ ) (u,\sigma) ( u , σ ) is d − 1 / 2 d^{-1/2} d − 1/2 times the coefficient of E h ( n ) ( I − M u ) v E^{(n)}_h(I-M_u)v E h ( n ) ( I − M u ) v at σ \sigma σ , with M x = 0 M_x=0 M x = 0 for the omitted input. Summing the squares proves (42) .
For harmonic v v v , Lemma 4.7 gives E h ( n ) M u v = E h ( n ) M u v E^{(n)}_hM_uv=E^{(n)}_hM_uv E h ( n ) M u v = E h ( n ) M u v and B M u = M u B BM_u=M_uB B M u = M u B . This is where purity of the synthesized masked vector matters; M u v M_uv M u v itself need not be harmonic. Every size-h h h matching records exactly h h h inputs, so ∑ u M u = h I \sum_uM_u=hI ∑ u M u = h I . Expanding the right side of (42) on a harmonic vector gives the quadratic form of
( n + 1 ) B − 2 h B + ∑ u B M u 2 = d B − ∑ u B ( M u − M u 2 ) . (n+1)B-2hB+\sum_uBM_u^2=dB-\sum_uB(M_u-M_u^2). ( n + 1 ) B − 2 h B + u ∑ B M u 2 = d B − u ∑ B ( M u − M u 2 ) .
Both sides are self-adjoint; complex polarization yields (43) . Each summand commutes with B B B , proving the commutation with B ^ \widehat{B} B . Finally, Lemma 4.5 gives
0 ⪯ M u − M u 2 ⪯ ( n − 2 h + 2 ) − 1 M u . 0\preceq M_u-M_u^2\preceq(n-2h+2)^{-1}M_u. 0 ⪯ M u − M u 2 ⪯ ( n − 2 h + 2 ) − 1 M u .
Because these operators commute with the positive B B B , the same inequalities hold after multiplication by B B B . Sum over u u u and use (43) to obtain (44) . For h = 0 h=0 h = 0 all masks vanish and B ^ = B \widehat{B}=B B = B . Notice that it is the polarized identity , not the two inequalities in (44) , that supplies commutation.
We next identify the space on which this comparison applies. In the large universe, let M x M_x M x and M y M^y M y be the masks for a recorded input x x x and output y y y . On H h ( N ) \mathcal{H}^{(N)}_h H h ( N ) abbreviate J x = J h , x J_x=J_{h,x} J x = J h , x , J y = O h , y J^y=O_{h,y} J y = O h , y and F x = F h , x \mathcal{F}_x=\mathcal{F}_{h,x} F x = F h , x ; thus put
J x = A h ( N ) M x ( A h ( N ) ) † , J y = A h ( N ) M y ( A h ( N ) ) † , F x = ker J x , F x y = F x ∩ ker J y . J_x=A^{(N)}_hM_x(A^{(N)}_h)^\dagger,\qquad J^y=A^{(N)}_hM^y(A^{(N)}_h)^\dagger,\qquad\mathcal{F}_x=\ker J_x,\quad\mathcal{F}_{xy}=\mathcal{F}_x\cap\ker J^y. J x = A h ( N ) M x ( A h ( N ) ) † , J y = A h ( N ) M y ( A h ( N ) ) † , F x = ker J x , F x y = F x ∩ ker J y .
Let P x y , h P_{xy,h} P x y , h denote the orthogonal projection onto F x y \mathcal{F}_{xy} F x y . The symbol A y , h = Π h Q y Π h A_{y,h}=\Pi_hQ_y\Pi_h A y , h = Π h Q y Π h refers to the degree-h h h assertion for the fixed input x x x ; Q y Q_y Q y is the literal projector 1 { π ( x ) = y } \mathbf{1}_{\{\pi(x)=y\}} 1 { π ( x ) = y } on permutation amplitudes.
Lemma 4.9 (The support obtained by deleting two coordinates). S S S maps K h ( n ) \mathcal{K}^{(n)}_h K h ( n ) bijectively onto F x y \mathcal{F}_{xy} F x y , and
G h ( N ) S = S B ^ . (45) G^{(N)}_hS=S\widehat{B}. \tag*{(45)} G h ( N ) S = S B . (45) On F x \mathcal{F}_x F x , the assertion has support in F x y \mathcal{F}_{xy} F x y : A y , h = P x y , h A y , h P x y , h A_{y,h}=P_{xy,h}A_{y,h}P_{xy,h} A y , h = P x y , h A y , h P x y , h there.
Proof. The image ι K h ( n ) \iota\mathcal{K}_{h}^{(n)} ι K h ( n ) is exactly the large harmonic space supported on matchings avoiding x , y x,y x , y . Indeed, a deletion row with a nonzero summand in the extension by zero must have its retained matching and its fixed coordinate in the smaller rectangle X × Y X \times Y X × Y . For such a row the large sum equals the small sum: all added terms are zero. All other rows are zero throughout. This applies separately to the input and output deletion rows. Conversely, restricting a harmonic vector supported in X × Y X \times Y X × Y therefore satisfies all small deletion rows. This also proves the assertion for h = 0 h=0 h = 0 , when there are no rows. In particular S = A h ( N ) ∣ ι K h ( n ) S=A_{h}^{(N)}|_{\iota\mathcal{K}_{h}^{(n)}} S = A h ( N ) ∣ ι K h ( n ) .
For brevity write A = A h ( N ) A=A_{h}^{(N)} A = A h ( N ) , G = A A † G=AA^{\dagger} G = A A † . For ψ ∈ H ( N ) \psi\in\mathcal{H}^{(N)} ψ ∈ H ( N ) one has ⟨ ψ , J x ψ ⟩ = ∥ M x A † ψ ∥ 2 \langle\psi,J_{x}\psi\rangle=\lVert M_{x}A^{\dagger}\psi\rVert^{2} ⟨ ψ , J x ψ ⟩ = ∥ M x A † ψ ∥ 2 , and likewise for J y J^{y} J y . Thus ψ ∈ F x y \psi\in F_{xy} ψ ∈ F x y precisely when A † ψ ∈ ι K h ( n ) A^{\dagger}\psi\in\iota\mathcal{K}_{h}^{(n)} A † ψ ∈ ι K h ( n ) . On the large harmonic space A † A A^{\dagger}A A † A commutes with both compressed masks by Lemma 4.7 (the output version follows by inversion). The kernel of a compressed mask on that space is the kernel of the original mask there, since ⟨ v , K h ( N ) M x v ⟩ = ∥ M x v ∥ 2 \langle v,K_{h}^{(N)}M_{x}v\rangle=\lVert M_{x}v\rVert^{2} ⟨ v , K h ( N ) M x v ⟩ = ∥ M x v ∥ 2 . Hence A † A A^{\dagger}A A † A preserves ι K h ( n ) \iota\mathcal{K}_{h}^{(n)} ι K h ( n ) , so S v Sv S v belongs to F x y F_{xy} F x y for every small harmonic v v v . Conversely G G G and G − 1 G^{-1} G − 1 preserve F x y F_{xy} F x y by their commutation with J x , J y J_{x},J^{y} J x , J y (Lemma 4.6). For ψ ∈ F x y \psi\in F_{xy} ψ ∈ F x y , we may therefore write A † G − 1 ψ = ι v A^{\dagger}G^{-1}\psi=\iota v A † G − 1 ψ = ι v , whence S v = A A † G − 1 ψ = ψ Sv=AA^{\dagger}G^{-1}\psi=\psi S v = A A † G − 1 ψ = ψ . Injectivity follows from the injectivity of A A A on its harmonic space. On F x y F_{xy} F x y we have S S † = G SS^{\dagger}=G S S † = G : inserting the orthogonal projection onto ι K h ( n ) \iota\mathcal{K}_{h}^{(n)} ι K h ( n ) between A A A and A † A^{\dagger} A † does not change A A † ψ AA^{\dagger}\psi A A † ψ there. Applying this identity to S S S gives (4.30).
It remains to check support of the assertion; this does not follow merely from its energy on F x y F_{xy} F x y . On the full permutation space let the recorded-input frame be E h ( N ) M x ( E h ( N ) ) † E_{h}^{(N)}M_{x}(E_{h}^{(N)})^{\dagger} E h ( N ) M x ( E h ( N ) ) † and define the recorded-output frame analogously. They preserve degree h h h by Lemma 4.6. On a completion column recording x x x , Q y Q_{y} Q y acts as the identity if the column contains x ↦ y x\mapsto y x ↦ y and as zero otherwise. The same holds on a column recording y y y , since a permutation has only one preimage of y y y . Multiplying the rank-one summands of either full frame by Q y Q_{y} Q y , on either side, consequently selects the same fixed-edge frame. Projecting to degree h h h gives
J x A y , h = A y , h J x = J y A y , h = A h ( N ) 1 { D ( x ) = y } ( A h ( N ) ) † . J_{x}A_{y,h}=A_{y,h}J_{x}=J^{y}A_{y,h}=A_{h}^{(N)}1_{\{D(x)=y\}}(A_{h}^{(N)})^{\dagger}. J x A y , h = A y , h J x = J y A y , h = A h ( N ) 1 { D ( x ) = y } ( A h ( N ) ) † .
All three products vanish on F x F_{x} F x . Hence A y , h F x ⊆ F x y A_{y,h}F_{x}\subseteq F_{xy} A y , h F x ⊆ F x y ; self-adjointness on F x F_{x} F x gives the claimed support on both sides.
Lemma 4.10 (Literal conditioning). Let R R R restrict a large permutation amplitude to the fibre π ( x ) = y \pi(x)=y π ( x ) = y , identified with the small bijections X → Y X\to Y X → Y , with counting measure on both sets. Then Q y = R † R Q_{y}=R^{\dagger}R Q y = R † R , R F x ⊆ H h ( n ) RF_{x}\subseteq\mathcal{H}_{h}^{(n)} R F x ⊆ H h ( n ) , and
R S v = d − 1 / 2 E h ( n ) v ( v ∈ K h ( n ) ) . (46) RSv=d^{-1/2}E_{h}^{(n)}v \qquad(v\in\mathcal{K}_{h}^{(n)}). \tag*{(46)} R S v = d − 1/2 E h ( n ) v ( v ∈ K h ( n ) ) . (46)
Proof. The adjoint R † R^{\dagger} R † is extension by zero outside the fibre, by the definition of the counting inner products. Restricting any completion column either kills it or removes the fixed edge x ↦ y x\mapsto y x ↦ y if present; the remaining column has at most the original degree. Hence R ψ ∈ H ≤ h ( n ) R\psi\in\mathcal{H}_{\leq h}^{(n)} R ψ ∈ H ≤ h ( n ) for ψ ∈ F x \psi\in F_{x} ψ ∈ F x . For h ≥ 1 h\geq1 h ≥ 1 , zero-extending a small column of size at most h − 1 h-1 h − 1 gives a scalar multiple of a large column of size at most h h h containing x ↦ y x\mapsto y x ↦ y . If its size is less than h h h , it is orthogonal to v v v by degree; at size h h h , its coefficient against ψ \psi ψ vanishes since M x ( A h ( N ) ) † ψ = 0 M_{x}(A_{h}^{(N)})^{\dagger}\psi=0 M x ( A h ( N ) ) † ψ = 0 . By adjointness R v Rv R v is orthogonal to H ≤ h − 1 ( n ) \mathcal{H}_{\leq h-1}^{(n)} H ≤ h − 1 ( n ) . At h = 0 h=0 h = 0 this space is zero. Finally, on the fibre, every embedded matching has exactly the same satisfaction indicator as its small matching. Its completion coefficient changes by the factorial factor d − 1 / 2 d^{-1/2} d − 1/2 , giving (4.31). More generally, for a large harmonic v ′ v' v ′ supported on matchings missing x x x , columns using y y y cannot contribute on this fibre. Thus R E h ( N ) v ′ = d − 1 / 2 E h ( n ) ι v ′ RE_{h}^{(N)}v'=d^{-1/2}E_{h}^{(n)}\iota v' R E h ( N ) v ′ = d − 1/2 E h ( n ) ι v ′ . The left side has pure degree h h h : E h ( N ) v ′ ∈ F x E_h^{(N)}v' \in\mathcal{F}_x E h ( N ) v ′ ∈ F x by the same mask-kernel argument used in Lemma 4.9 . One can therefore insert K h ( n ) K_h^{(n)} K h ( n ) before synthesis on the right; one cannot simply assume ι † v ′ \iota^\dagger v' ι † v ′ itself is harmonic.
Proposition 4.11 (Fresh assertion and its Gram commutation). On F x \mathcal{F}_x F x , with d = N − h d=N-h d = N − h ,
P x y , h ⪯ d A y , h ⪯ ( 1 + ε h ) P x y , h , [ G h ( N ) , A y , h ] = 0 , ε h = h ( N − 2 h ) ( N − h + 1 ) . (47) P_{xy,h} \preceq dA_{y,h} \preceq(1+\varepsilon_h)P_{xy,h}, \qquad[G_h^{(N)},A_{y,h}]=0, \qquad\varepsilon_h=\frac{h}{(N-2h)(N-h+1)}. \tag*{(47)} P x y , h ⪯ d A y , h ⪯ ( 1 + ε h ) P x y , h , [ G h ( N ) , A y , h ] = 0 , ε h = ( N − 2 h ) ( N − h + 1 ) h . (47) In particular the nonzero eigenvalues of d A y , h dA_{y,h} d A y , h there belong to [ 1 , 1 + ε h ] [1,1+\varepsilon_h] [ 1 , 1 + ε h ] .
Proof. For small harmonic u , v u,v u , v , literal conditioning gives
⟨ S u , A y , h S v ⟩ = ⟨ R S u , R S v ⟩ = d − 1 ⟨ u , B v ⟩ . (48) \langle Su,A_{y,h}Sv\rangle=\langle RSu,RSv\rangle=d^{-1}\langle u,Bv\rangle. \tag*{(48)} ⟨ S u , A y , h S v ⟩ = ⟨ R S u , R S v ⟩ = d − 1 ⟨ u , B v ⟩ . (48) Put ψ = S v \psi=S v ψ = S v . By (44) the quantities ∥ ψ ∥ 2 = ⟨ v , B ^ v ⟩ \|\psi\|^2=\langle v,\widehat{B}v\rangle ∥ ψ ∥ 2 = ⟨ v , B v ⟩ and d ⟨ ψ , A y , h ψ ⟩ = ⟨ v , B v ⟩ d\langle\psi,A_{y,h}\psi\rangle=\langle v,Bv\rangle d ⟨ ψ , A y , h ψ ⟩ = ⟨ v , B v ⟩ satisfy
∥ ψ ∥ 2 ≤ d ⟨ ψ , A y , h ψ ⟩ ≤ ( 1 − ρ h ) − 1 ∥ ψ ∥ 2 . \|\psi\|^2\le d\langle\psi,A_{y,h}\psi\rangle\le(1-\rho_h)^{-1}\|\psi\|^2. ∥ ψ ∥ 2 ≤ d ⟨ ψ , A y , h ψ ⟩ ≤ ( 1 − ρ h ) − 1 ∥ ψ ∥ 2 . Indeed ( 1 − ρ h ) ( 1 + ε h ) = 1 (1-\rho_h)(1+\varepsilon_h)=1 ( 1 − ρ h ) ( 1 + ε h ) = 1 , as one sees from d ( n − 2 h + 2 ) − h = ( N − 2 h ) ( N − h + 1 ) > 0 d(n-2h+2)-h=(N-2h)(N-h+1)>0 d ( n − 2 h + 2 ) − h = ( N − 2 h ) ( N − h + 1 ) > 0 . This proves the sandwich on F x y \mathcal{F}_{xy} F x y , and the support identity of Lemma 4.9 proves it on F x \mathcal{F}_x F x .
For commutation retain the full sesquilinear identity (48) . Since S S S is a bijection onto F x y \mathcal{F}_{xy} F x y and B ^ = S † S \widehat{B}=S^\dagger S B = S † S is invertible,
A y , h S = d − 1 S B ^ − 1 B . A_{y,h}S=d^{-1}S\widehat{B}^{-1}B. A y , h S = d − 1 S B − 1 B . Testing both sides against S u Su S u proves this equality; both ranges lie in F x y \mathcal{F}_{xy} F x y . In the same coordinates G h ( N ) S = S B ^ G_h^{(N)}S=S\widehat{B} G h ( N ) S = S B by (45) . The two operators commute there because B ^ B = B B ^ \widehat{B}B=B\widehat{B} B B = B B . On F x ⊖ F x y \mathcal{F}_x\ominus\mathcal{F}_{xy} F x ⊖ F x y the assertion vanishes, while the Gram preserves this subspace by self-adjointness and invariance of F x , F x y \mathcal{F}_x,\mathcal{F}_{xy} F x , F x y . This completes the commutator proof on all of F x \mathcal{F}_x F x .
The deleted input and output were chosen arbitrarily. More explicitly, independent relabelings a , b a,b a , b act by D ↦ b D a − 1 D\mapsto bDa^{-1} D ↦ b D a − 1 and π ↦ b π a − 1 \pi\mapsto b\pi a^{-1} π ↦ bπ a − 1 . These are coordinate unitaries on both spaces; they preserve D ⊆ π D\subseteq\pi D ⊆ π and the factor ( N − h ) ! (N-h)! ( N − h )! . They therefore intertwine synthesis, its adjoint, both harmonic projections, the masks and the assertions. No comparison between irreducible labels for different outcomes is required.
Comparing a complete function query
The preceding estimates describe a fixed degree. We now put them together across adjacent degrees, first for a whole history of filtered function queries and then for a bounded number of unfiltered terminal reads.
The middle and downward return
We now estimate the two terms in (34) . Fix the query input x x x and write F h \mathcal{F}_h F h for the subspace ker J h , x \ker J_{h,x} ker J h , x of H h \mathcal{H}_h H h , and F h ⊥ \mathcal{F}_h^\perp F h ⊥ for its orthogonal complement in H h \mathcal{H}_h H h . These are the fresh and known subspaces, respectively.
Lemma 4.12 (Middle return). If 3 h + 2 ≤ N 3h+2\le N 3 h + 2 ≤ N , then
∥ D z , h 0 ∥ ≤ e + 1 N − 2 h . \|D_{z,h}^{0}\|\le\frac{e+1}{N-2h}. ∥ D z , h 0 ∥ ≤ N − 2 h e + 1 .
Proof. Set d = N − h d=N-h d = N − h , G = G h G=G_h G = G h , J = J h , x J=J_{h,x} J = J h , x and A y = A y , h A_y=A_{y,h} A y = A y , h as in (30) . Define the output-occupied frame O y = A h 1 { y ∈ im D } A h † O_y=A_h\mathbf{1}_{\{y\in\operatorname{im}D\}}A_h^\dagger O y = A h 1 { y ∈ im D } A h † . Lemma 4.6 gives [ J , A y ] = [ J , O y ] = [ G , J ] = [ G , O y ] = 0 [J,A_y]=[J,O_y]=[G,J]=[G,O_y]=0 [ J , A y ] = [ J , O y ] = [ G , J ] = [ G , O y ] = 0 .
Write A z = ∑ y z y A y A_z=\sum_y z_yA_y A z = ∑ y z y A y . Since it is the compression of U z U_z U z , ∥ A z ∥ ≤ 1 \lVert A_z\rVert\le1 ∥ A z ∥ ≤ 1 . We use the identities
H y , h 11 = A y J , ∑ y H y , h 00 = G − J , ∑ y O y = h G . H^{11}_{y,h}=A_yJ,\qquad\sum_y H^{00}_{y,h}=G-J,\qquad\sum_y O_y=hG. H y , h 11 = A y J , y ∑ H y , h 00 = G − J , y ∑ O y = h G .
For the first, multiply the fixed-input frame by the literal Q y Q_y Q y : among its completion vectors only those recording x ↦ y x\mapsto y x ↦ y survive. Projection onto degree h h h gives the assertion. For the other two, each matching of size h h h missing x x x has d d d unused outputs, and each matching has h h h occupied outputs. All of these operators reduce F h \mathcal{F}_h F h and F h ⊥ \mathcal{F}_h^\perp F h ⊥ . Indeed ψ ∈ F h \psi\in\mathcal{F}_h ψ ∈ F h means A h † ψ A_h^\dagger\psi A h † ψ vanishes on matchings recording x x x . Hence d H y , h 00 ψ = ( G − O y ) ψ dH^{00}_{y,h}\psi=(G-O_y)\psi d H y , h 00 ψ = ( G − O y ) ψ on F h \mathcal{F}_h F h . This expression preserves F h \mathcal{F}_h F h ; self-adjointness gives the invariance of F h ⊥ \mathcal{F}_h^\perp F h ⊥ . The other invariances follow from the commutations just proved.
On F h ⊥ \mathcal{F}_h^\perp F h ⊥ , the fixed-input ratio estimate (38) gives ∥ I − J G − 1 ∥ ≤ η h = ( N − 2 h + 2 ) − 1 \lVert I-JG^{-1}\rVert\le\eta_h=(N-2h+2)^{-1} ∥ I − J G − 1 ∥ ≤ η h = ( N − 2 h + 2 ) − 1 . By the positive-family inequality (35) , ∥ H z , h 00 ∥ ≤ ∥ G − J ∥ ≤ e η h \lVert H^{00}_{z,h}\rVert\le\lVert G-J\rVert\le e\eta_h ∥ H z , h 00 ∥ ≤ ∥ G − J ∥ ≤ e η h on this subspace. Also H z , h 11 G − 1 − A z = A z ( J G − 1 − I ) H^{11}_{z,h}G^{-1}-A_z=A_z(JG^{-1}-I) H z , h 11 G − 1 − A z = A z ( J G − 1 − I ) . As ∥ G − 1 ∥ , ∥ W h ∥ ≤ 1 \lVert G^{-1}\rVert,\lVert W_h\rVert\le1 ∥ G − 1 ∥ , ∥ W h ∥ ≤ 1 , (32) bounds the known part by ( e + 1 ) η h (e+1)\eta_h ( e + 1 ) η h .
On F h \mathcal{F}_h F h , put B y = ( O y G − 1 ) ∣ F h B_y=(O_yG^{-1})|_{\mathcal{F}_h} B y = ( O y G − 1 ) ∣ F h and let P y P_y P y project onto F h ∩ ker O y \mathcal{F}_h\cap\ker O_y F h ∩ ker O y . We have 0 ≤ B y ≤ I 0\le B_y\le I 0 ≤ B y ≤ I , since 0 ≤ O y ≤ G 0\le O_y\le G 0 ≤ O y ≤ G and the two commute. Thus I − B y ≥ P y I-B_y\ge P_y I − B y ≥ P y ; the cross terms vanish since B y P y = 0 B_yP_y=0 B y P y = 0 . The conditioning estimate of Proposition 4.11 gives
P y ⪯ d A y ⪯ ( 1 + ϵ h ) P y , ϵ h = h ( N − 2 h ) ( N − h + 1 ) . P_y\preceq dA_y\preceq(1+\epsilon_h)P_y,\qquad\epsilon_h=\frac{h}{(N-2h)(N-h+1)}. P y ⪯ d A y ⪯ ( 1 + ϵ h ) P y , ϵ h = ( N − 2 h ) ( N − h + 1 ) h .
On this subspace H y , h 11 = 0 H^{11}_{y,h}=0 H y , h 11 = 0 , and H y , h 00 G − 1 = d − 1 ( I − B y ) H^{00}_{y,h}G^{-1}=d^{-1}(I-B_y) H y , h 00 G − 1 = d − 1 ( I − B y ) . Thus the real operator in parentheses in (32) is the signed sum of the differences between two positive families, d − 1 ( I − B y ) d^{-1}(I-B_y) d − 1 ( I − B y ) and A y A_y A y . Both sum to I I I , since ∑ y B y = h I \sum_yB_y=hI ∑ y B y = h I and ∑ y A y = I \sum_yA_y=I ∑ y A y = I . Both dominate P y / d P_y/d P y / d . Subtract this common part before applying the positive-family inequality. The norm is at most
2 ∥ I − 1 d ∑ y P y ∥ ≤ 2 ϵ h 1 + ϵ h ≤ 2 ϵ h . 2\left\|I-\frac{1}{d}\sum_yP_y\right\|\le2\frac{\epsilon_h}{1+\epsilon_h}\le2\epsilon_h. 2 I − d 1 y ∑ P y ≤ 2 1 + ϵ h ϵ h ≤ 2 ϵ h .
This cancellation controls all outcomes at once. Since ϵ h ≤ 1 / ( N − 2 h ) \epsilon_h\le1/(N-2h) ϵ h ≤ 1/ ( N − 2 h ) , the fresh estimate is sufficient. Finally the two restrictions of the real operator have orthogonal ranges; its norm is their maximum. Applying W h W_h W h cannot increase that norm.
To handle neighboring degrees we must use one and the same frame size on both sides of the query. For integers s , j s,j s , j , let J x [ s ; j ] J_x^{[s;j]} J x [ s ; j ] be the restriction to H j \mathcal{H}_j H j of the full-space frame E s 1 { x ∈ dom D } E s † E_s\mathbf{1}_{\{x\in\operatorname{dom}D\}}E_s^\dagger E s 1 { x ∈ dom D } E s † . In particular J x [ s ; s ] = J s , x J_x^{[s;s]}=J_{s,x} J x [ s ; s ] = J s , x . Because the full-space operator preserves all degrees and commutes with every literal at x x x , it satisfies
J x [ s ; i ] Π i U z Π j = Π i U z Π j J x [ s ; j ] . (49) J_x^{[s;i]}\Pi_iU_z\Pi_j=\Pi_iU_z\Pi_jJ_x^{[s;j]}. \tag*{(49)} J x [ s ; i ] Π i U z Π j = Π i U z Π j J x [ s ; j ] . (49)
Lemma 4.13 (Lower frame on the fresh subspace). Let j < N j<N j < N , d = N − j d=N-j d = N − j and A y = A y , j A_y=A_{y,j} A y = A y , j . On F j \mathcal{F}_j F j one has
J x [ j + 1 ; j ] = d ∑ y A y G j A y . J_x^{[j+1;j]}=d\sum_yA_yG_jA_y. J x [ j + 1 ; j ] = d y ∑ A y G j A y . Moreover Q = Π j U z Π j + 1 Q=\Pi_jU_z\Pi_{j+1} Q = Π j U z Π j + 1 has image in F j \mathcal{F}_j F j and vanishes on F j + 1 \mathcal{F}_{j+1} F j + 1 .
Proof. For ϕ ∈ H j + 1 \phi\in\mathcal{H}_{j+1} ϕ ∈ H j + 1 and a matching D D D of size j j j , the coefficient ⟨ T D , Q y ϕ ⟩ \langle T_D,Q_y\phi\rangle ⟨ T D , Q y ϕ ⟩ is zero if D D D records x x x : the literal acts on T D T_D T D as either the identity or zero, and ϕ ⊥ H ≤ j \phi\perp\mathcal{H}_{\le j} ϕ ⊥ H ≤ j . If D D D misses x x x but uses y y y , the coefficient is again zero. In the remaining case it is d − 1 / 2 ⟨ T D ∪ { ( x , y ) } , ϕ ⟩ d^{-1/2}\langle T_{D\cup\{(x,y)\}},\phi\rangle d − 1/2 ⟨ T D ∪ {( x , y )} , ϕ ⟩ , by the normalization of the two completion vectors. Projection onto degree j j j does not change these coefficients, since Q y ϕ Q_y\phi Q y ϕ has no component below j j j . Thus the analysis of Π j Q y ϕ \Pi_jQ_y\phi Π j Q y ϕ vanishes on matchings recording x x x ; it vanishes identically when ϕ ∈ F j + 1 \phi\in\mathcal{F}_{j+1} ϕ ∈ F j + 1 . The size-j j j completion vectors span H ≤ j \mathcal{H}_{\leq j} H ≤ j , so these statements imply the last two claims, after summing over y y y .
For u ∈ F j u\in\mathcal{F}_j u ∈ F j , this same vanishing applied to Π j − 1 Q y u \Pi_{j-1}Q_yu Π j − 1 Q y u shows that Q y u Q_yu Q y u has no component below j j j (at j = 0 j=0 j = 0 there is nothing to show). Its size-j j j analysis therefore equals A j † A y u A_j^\dagger A_yu A j † A y u . This analysis is supported on matchings missing both x x x and y y y : if D D D records x x x , use ⟨ T D , u ⟩ = 0 \langle T_D,u\rangle=0 ⟨ T D , u ⟩ = 0 , and if it misses x x x but uses y y y , Q y T D = 0 Q_yT_D=0 Q y T D = 0 . Now enumerate size-( j + 1 ) (j+1) ( j + 1 ) matchings containing the edge ( x , y ) (x,y) ( x , y ) by deleting that edge. For u , v ∈ F j u,v\in\mathcal{F}_j u , v ∈ F j the resulting frame identity in quadratic-form notation is
⟨ u , J x [ j + 1 ; j ] v ⟩ = d ∑ y ⟨ A j † A y u , A j † A y v ⟩ = d ∑ y ⟨ u , A y G j A y v ⟩ . \langle u,J_x^{[j+1;j]}v\rangle=d\sum_y\langle A_j^\dagger A_yu,A_j^\dagger A_yv\rangle=d\sum_y\langle u,A_yG_jA_yv\rangle. ⟨ u , J x [ j + 1 ; j ] v ⟩ = d y ∑ ⟨ A j † A y u , A j † A y v ⟩ = d y ∑ ⟨ u , A y G j A y v ⟩ .
Here the factor d d d follows from T D ∪ { ( x , y ) } = d Q y T D T_{D\cup\{(x,y)\}}=\sqrt{d}\,Q_yT_D T D ∪ {( x , y )} = d Q y T D . This proves equality of the compressions to F j \mathcal{F}_j F j . They are also restrictions: the right side preserves F j \mathcal{F}_j F j , and the size-( j + 1 ) (j+1) ( j + 1 ) fixed-input frame commutes with the size-j j j fixed-input frame. To see the latter commutation, write their full-space kernels as convolutions supported on the subgroup fixing x x x . The coefficients depend only on the number of other fixed inputs of the relative permutation and are class functions on that subgroup. Convolution by two such functions commutes. Restricting to degree j j j preserves this relation. □ \square □
Lemma 4.14 (Downward return). For h = j + 1 h=j+1 h = j + 1 satisfying 3 h + 2 ≤ N 3h+2\leq N 3 h + 2 ≤ N ,
∥ G j Q G h − 1 − Q ∥ ≤ 4 N − h , ∥ D z , h ∥ ≤ 4 N − h . (50) \lVert G_jQG_h^{-1}-Q\rVert\leq\frac{4}{N-h},\qquad\lVert D_{z,h}\rVert\leq\frac{4}{N-h}. \tag*{(50)} ∥ G j Q G h − 1 − Q ∥ ≤ N − h 4 , ∥ D z , h ∥ ≤ N − h 4 . (50)
Proof. Set R = J h , x G h − 1 R=J_{h,x}G_h^{-1} R = J h , x G h − 1 , S = J x [ h ; j ] G j − 1 S=J_x^{[h;j]}G_j^{-1} S = J x [ h ; j ] G j − 1 , and Δ = G j Q G h − 1 − Q \Delta=G_jQG_h^{-1}-Q Δ = G j Q G h − 1 − Q . Lemma 4.13 and degree-wise invariance under the Gram operators show that Δ \Delta Δ vanishes on F h \mathcal{F}_h F h and takes values in F j \mathcal{F}_j F j . By (49) with s = h s=h s = h ,
S Δ = Q ( R − I ) − ( S − I ) Q . (51) S\Delta=Q(R-I)-(S-I)Q. \tag*{(51)} S Δ = Q ( R − I ) − ( S − I ) Q . (51)
The point is to estimate S S S on the actual range of Δ \Delta Δ , rather than comparing unrelated eigenvalue lists across degrees.
On F j \mathcal{F}_j F j , Proposition 4.11 gives [ A y , G j ] = 0 [A_y,G_j]=0 [ A y , G j ] = 0 . Hence the preceding frame identity reduces to S = d ∑ y A y 2 S=d\sum_yA_y^2 S = d ∑ y A y 2 , d = N − j d=N-j d = N − j . Each A y A_y A y is positive, ∑ y A y = I \sum_yA_y=I ∑ y A y = I , and every nonzero eigenvalue a a a satisfies 1 ≤ d a ≤ 1 + ε j 1\leq da\leq1+\varepsilon_j 1 ≤ d a ≤ 1 + ε j . Functional calculus gives A y ⪯ d A y 2 ⪯ ( 1 + ε j ) A y A_y\preceq dA_y^2\preceq(1+\varepsilon_j)A_y A y ⪯ d A y 2 ⪯ ( 1 + ε j ) A y ; summing and taking norms yields
I ⪯ S ∣ F j ⪯ ( 1 + ε j ) I , ∥ w ∥ ≤ ∥ S w ∥ ( w ∈ F j ) . (52) I\preceq S|_{\mathcal{F}_j}\preceq(1+\varepsilon_j)I,\qquad\lVert w\rVert\leq\lVert Sw\rVert\quad(w\in\mathcal{F}_j). \tag*{(52)} I ⪯ S ∣ F j ⪯ ( 1 + ε j ) I , ∥ w ∥ ≤ ∥ S w ∥ ( w ∈ F j ) . (52)
For v ∈ F h ⊥ v\in\mathcal{F}_h^\perp v ∈ F h ⊥ , use ∥ Q ∥ ≤ 1 \lVert Q\rVert\leq1 ∥ Q ∥ ≤ 1 , ∥ ( R − I ) v ∥ ≤ η h ∥ v ∥ \lVert(R-I)v\rVert\leq\eta_h\lVert v\rVert ∥( R − I ) v ∥ ≤ η h ∥ v ∥ , and Q v , Δ v ∈ F j Qv,\Delta v\in\mathcal{F}_j Q v , Δ v ∈ F j in (51) –(52) to get ∥ Δ v ∥ ≤ ( η h + ε j ) ∥ v ∥ \lVert\Delta v\rVert\leq(\eta_h+\varepsilon_j)\lVert v\rVert ∥ Δ v ∥ ≤ ( η h + ε j ) ∥ v ∥ . Since Δ \Delta Δ annihilates F h \mathcal{F}_h F h , this is its full norm bound. Here η h = 1 / ( N − 2 j ) \eta_h=1/(N-2j) η h = 1/ ( N − 2 j ) , ε j ≤ 1 / ( N − 2 j ) \varepsilon_j\leq1/(N-2j) ε j ≤ 1/ ( N − 2 j ) , and 2 / ( N − 2 j ) ≤ 4 / ( N − j − 1 ) 2/(N-2j)\leq4/(N-j-1) 2/ ( N − 2 j ) ≤ 4/ ( N − j − 1 ) by 3 h + 2 ≤ N 3h+2\leq N 3 h + 2 ≤ N . Finally use (33) and ∥ W j ∥ ≤ 1 \lVert W_j\rVert\leq1 ∥ W j ∥ ≤ 1 . □ \square □
Theorem 4.15 (Complete-phase canonical return). If 3 D + 2 ≤ N 3D+2\leq N 3 D + 2 ≤ N , then
∥ R F z W Π ≤ D ∥ = ∥ ( K F z W − W U z ) Π ≤ D ∥ ≤ e + 5 N − 2 D . (53) \lVert RF_zW\Pi_{\leq D}\rVert=\lVert(KF_zW-WU_z)\Pi_{\leq D}\rVert\leq\frac{e+5}{N-2D}. \tag*{(53)} ∥ R F z W Π ≤ D ∥ = ∥( K F z W − W U z ) Π ≤ D ∥ ≤ N − 2 D e + 5 . (53)
Proof. The canonical upward correction vanishes by the exact query identities. Among the middle blocks, different source degrees have orthogonal target degrees; the same holds among the downward blocks. Thus the norm of either family is its largest block norm. By (34) the two families together cost at most ( e + 1 ) / ( N − 2 D ) + 4 / ( N − 2 D ) (e+1)/(N-2D)+4/(N-2D) ( e + 1 ) / ( N − 2 D ) + 4/ ( N − 2 D ) . Coherent controls by the queried input or by the phase vector give an orthogonal direct sum of such operators. □ \square □
An intrinsic square-root barrier for the unfiltered process
The following obstruction concerns the unfiltered compressed evolution, not merely the accumulation of a particular one-query error estimate. It is witnessed by an algorithm that measures only its own registers.
Proposition 4.16 (Unfiltered CPO has a square-root soundness barrier). Let N N N be even. Fix signs z y ∈ { − 1 , 1 } z_y \in\{-1,1\} z y ∈ { − 1 , 1 } satisfying ∑ y = 1 N z y = 0 \sum_{y=1}^{N} z_y = 0 ∑ y = 1 N z y = 0 , and let F x = C x L z C x F_x = C_xL_zC_x F x = C x L z C x be the corresponding complete compressed phase-query block on square partial-injection databases. For every odd integer s ≥ 1 s \ge1 s ≥ 1 with s 2 ≤ N / 24 s^2 \le N/24 s 2 ≤ N /24 , there is an algorithm using s s s such blocks whose acceptance probability is zero for every real permutation, but is at least
p C P O ≥ s 2 8 N (54) p_{\mathrm{CPO}} \ge\frac{s^2}{8N} \tag*{(54)} p CPO ≥ 8 N s 2 (54) for the unfiltered CPO initialized at the empty database. Each block can be implemented with two forward-swap queries and a known diagonal response operation.
Proof. Let ∣ u ⟩ = N − 1 / 2 ∑ x ∣ x ⟩ |u\rangle= N^{-1/2}\sum_x |x\rangle ∣ u ⟩ = N − 1/2 ∑ x ∣ x ⟩ be the uniform query-input state, and put P = ∣ u ⟩ ⟨ u ∣ ⊗ I P = |u\rangle\langle u| \otimes I P = ∣ u ⟩ ⟨ u ∣ ⊗ I and R = 2 P − I R = 2P-I R = 2 P − I . Both the reflection R R R and the final measurement { P , I − P } \{P,I-P\} { P , I − P } act only on the algorithm’s input register. Write
F = ∑ x ∣ x ⟩ ⟨ x ∣ ⊗ F x , B = ⟨ u ∣ F ∣ u ⟩ = 1 N ∑ x F x . F = \sum_x |x\rangle\langle x| \otimes F_x,\qquad B = \langle u|F|u\rangle= \frac{1}{N}\sum_x F_x. F = x ∑ ∣ x ⟩ ⟨ x ∣ ⊗ F x , B = ⟨ u ∣ F ∣ u ⟩ = N 1 x ∑ F x .
The operator F F F is a self-adjoint unitary, and B B B is a self-adjoint contraction. Starting from ∣ u ⟩ ∣ ∅ ⟩ |u\rangle|\varnothing\rangle ∣ u ⟩ ∣ ∅ ⟩ , the algorithm applies ( R F ) s (RF)^s ( R F ) s and accepts upon outcome P P P .
For a real permutation π \pi π , the phase query is U π ∣ x ⟩ = z π ( x ) ∣ x ⟩ U_\pi|x\rangle= z_{\pi(x)}|x\rangle U π ∣ x ⟩ = z π ( x ) ∣ x ⟩ . Hence ⟨ u ∣ U π ∣ u ⟩ = N − 1 ∑ y z y = 0 \langle u|U_\pi|u\rangle= N^{-1}\sum_y z_y = 0 ⟨ u ∣ U π ∣ u ⟩ = N − 1 ∑ y z y = 0 . The identity below therefore shows that the real acceptance probability is zero for every odd s s s .
To calculate the compressed acceptance probability, let T j T_j T j be the polynomials defined by T 0 ( b ) = 1 T_0(b)=1 T 0 ( b ) = 1 , T 1 ( b ) = b T_1(b)=b T 1 ( b ) = b and T j + 1 ( b ) = 2 b T j ( b ) − T j − 1 ( b ) T_{j+1}(b)=2bT_j(b)-T_{j-1}(b) T j + 1 ( b ) = 2 b T j ( b ) − T j − 1 ( b ) . For V = R F V=RF V = R F we have P ( V + V † ) = 2 P F P = 2 B P P(V+V^\dagger)=2PFP=2BP P ( V + V † ) = 2 P F P = 2 B P , with B B B identified with its action on ran P \operatorname{ran} P ran P . Since V † = V − 1 V^\dagger=V^{-1} V † = V − 1 , multiplication by V j V^j V j gives the recurrence
P V j + 1 P + P V j − 1 P = 2 B P V j P . PV^{j+1}P+PV^{j-1}P=2BPV^jP. P V j + 1 P + P V j − 1 P = 2 B P V j P .
The initial values imply P V j P = T j ( B ) PV^jP=T_j(B) P V j P = T j ( B ) . If ν \nu ν is the spectral measure of B B B in ∣ ∅ ⟩ |\varnothing\rangle ∣ ∅ ⟩ , then, for odd s s s ,
p C P O = ∥ T s ( B ) ∣ ∅ ⟩ ∥ 2 = ∫ − 1 1 sin 2 ( s arcsin b ) d ν ( b ) . (55) p_{\mathrm{CPO}}=\|T_s(B)|\varnothing\rangle\|^2=\int_{-1}^{1}\sin^2(s\arcsin b)\,d\nu(b). \tag*{(55)} p CPO = ∥ T s ( B ) ∣ ∅ ⟩ ∥ 2 = ∫ − 1 1 sin 2 ( s arcsin b ) d ν ( b ) . (55)
Here T j ( b ) = cos ( j arccos b ) T_j(b)=\cos(j\arccos b) T j ( b ) = cos ( j arccos b ) follows from the same recurrence.
We next compute two moments of ν \nu ν . Write ∣ x , y ⟩ |x,y\rangle ∣ x , y ⟩ for the database containing just x ↦ y x\mapsto y x ↦ y . The local reflection and the balanced-sign condition give
B ∣ ∅ ⟩ = N − 3 / 2 ∑ x , y z y ∣ x , y ⟩ , ∫ b 2 d ν ( b ) = 1 N . B|\varnothing\rangle=N^{-3/2}\sum_{x,y}z_y|x,y\rangle,\qquad\int b^2\,d\nu(b)=\frac{1}{N}. B ∣ ∅ ⟩ = N − 3/2 x , y ∑ z y ∣ x , y ⟩ , ∫ b 2 d ν ( b ) = N 1 .
For a size-two matching D = { ( x 1 , y 1 ) , ( x 2 , y 2 ) } D=\{(x_1,y_1),(x_2,y_2)\} D = {( x 1 , y 1 ) , ( x 2 , y 2 )} , put z ( D ) = z y 1 z y 2 z(D)=z_{y_1}z_{y_2} z ( D ) = z y 1 z y 2 . A second application gives
B 2 ∣ ∅ ⟩ = 1 N ∣ ∅ ⟩ − 1 N 5 / 2 ∑ x , y ∣ x , y ⟩ + 2 N 5 / 2 N − 1 ∑ ∣ D ∣ = 2 ( z ( D ) + 1 N − 1 ) ∣ D ⟩ . (56) B^2|\varnothing\rangle=\frac{1}{N}|\varnothing\rangle-\frac{1}{N^{5/2}}\sum_{x,y}|x,y\rangle
+\frac{2}{N^{5/2}\sqrt{N-1}}\sum_{|D|=2}\left(z(D)+\frac{1}{N-1}\right)|D\rangle. \tag*{(56)} B 2 ∣ ∅ ⟩ = N 1 ∣ ∅ ⟩ − N 5/2 1 x , y ∑ ∣ x , y ⟩ + N 5/2 N − 1 2 ∣ D ∣ = 2 ∑ ( z ( D ) + N − 1 1 ) ∣ D ⟩ . (56)
Indeed, querying the input already present in the signed one-record row gives ∑ y z y F x ∣ x , y ⟩ = N ∣ ∅ ⟩ \sum_y z_y F_x|x,y\rangle=\sqrt{N}|\varnothing\rangle ∑ y z y F x ∣ x , y ⟩ = N ∣ ∅ ⟩ . For w ≠ x w\ne x w = x , the mean of the signs available after recording x ↦ y x\mapsto y x ↦ y is − z y / ( N − 1 ) -z_y/(N-1) − z y / ( N − 1 ) , so
F w ∣ x , y ⟩ = − z y N − 1 ∣ x , y ⟩ + 1 N − 1 ∑ v ≠ y ( z v + z y N − 1 ) ∣ ( x , y ) , ( w , v ) ⟩ . F_w|x,y\rangle=-\frac{z_y}{N-1}|x,y\rangle+\frac{1}{\sqrt{N-1}}\sum_{v\ne y}\left(z_v+\frac{z_y}{N-1}\right)|(x,y),(w,v)\rangle. F w ∣ x , y ⟩ = − N − 1 z y ∣ x , y ⟩ + N − 1 1 v = y ∑ ( z v + N − 1 z y ) ∣ ( x , y ) , ( w , v )⟩ .
The two possible insertion orders give the size-two coefficient in (56) . Since ∑ y 1 ≠ y 2 z y 1 z y 2 = − N \sum_{y_1\ne y_2}z_{y_1}z_{y_2}=-N ∑ y 1 = y 2 z y 1 z y 2 = − N , counting the ( N 2 ) \binom{N}{2} ( 2 N ) input pairs yields
∑ ∣ D ∣ = 2 ( z ( D ) + 1 N − 1 ) 2 = N 3 ( N − 2 ) 2 . \sum_{|D|=2}\left(z(D)+\frac{1}{N-1}\right)^2=\frac{N^3(N-2)}{2}. ∣ D ∣ = 2 ∑ ( z ( D ) + N − 1 1 ) 2 = 2 N 3 ( N − 2 ) .
The three sizes in (56) are orthogonal. Consequently
∫ b 4 d ν ( b ) = 1 N 2 + 1 N 3 + 2 ( N − 2 ) N 2 ( N − 1 ) ≤ 3 N 2 . (57) \int b^4\,d\nu(b)=\frac{1}{N^2}+\frac{1}{N^3}+\frac{2(N-2)}{N^2(N-1)}\le\frac{3}{N^2}. \tag*{(57)} ∫ b 4 d ν ( b ) = N 2 1 + N 3 1 + N 2 ( N − 1 ) 2 ( N − 2 ) ≤ N 2 3 . (57)
For ∣ b ∣ ≤ 1 / ( 2 s ) \lvert b\rvert\le1/(2s) ∣ b ∣ ≤ 1/ ( 2 s ) , the inequalities ∣ b ∣ ≤ ∣ arcsin b ∣ ≤ 2 ∣ b ∣ \lvert b\rvert\le\lvert\arcsin b\rvert\le2\lvert b\rvert ∣ b ∣ ≤ ∣ arcsin b ∣ ≤ 2 ∣ b ∣ give ∣ s arcsin b ∣ ≤ 1 \lvert s\arcsin b\rvert\le1 ∣ s arcsin b ∣ ≤ 1 . Using ∣ sin x ∣ ≥ ∣ x ∣ / 2 \lvert\sin x\rvert\ge\lvert x\rvert/2 ∣ sin x ∣ ≥ ∣ x ∣ /2 for ∣ x ∣ ≤ 1 \lvert x\rvert\le1 ∣ x ∣ ≤ 1 , we obtain sin 2 ( s arcsin b ) ≥ s 2 b 2 / 4 \sin^2(s\arcsin b)\ge s^2b^2/4 sin 2 ( s arcsin b ) ≥ s 2 b 2 /4 . The two moment estimates imply
∫ ∣ b ∣ ≤ 1 / ( 2 s ) b 2 d ν ( b ) ≥ 1 N − 4 s 2 ∫ b 4 d ν ( b ) ≥ 1 N ( 1 − 12 s 2 N ) ≥ 1 2 N . \int_{\lvert b\rvert\le1/(2s)} b^2\,d\nu(b)\ge\frac{1}{N}-4s^2\int b^4\,d\nu(b)\ge\frac{1}{N}\left(1-\frac{12s^2}{N}\right)\ge\frac{1}{2N}. ∫ ∣ b ∣ ≤ 1/ ( 2 s ) b 2 d ν ( b ) ≥ N 1 − 4 s 2 ∫ b 4 d ν ( b ) ≥ N 1 ( 1 − N 12 s 2 ) ≥ 2 N 1 .
Substitution into (55) proves (54) .
Corollary 4.17 (Optimal scale of uniform state comparison). For the algorithm above, let ζ s \zeta_s ζ s be its unfiltered CPO state and ψ s \psi_s ψ s its real purified state. For every isometric encoding V d b V_{\mathrm{db}} V db acting only on the real reference register,
∥ ζ s − ( I ⊗ V d b ) ψ s ∥ ≥ s 8 N . \lVert\zeta_s-(I\otimes V_{\mathrm{db}})\psi_s\rVert\ge\frac{s}{\sqrt{8N}}. ∥ ζ s − ( I ⊗ V db ) ψ s ∥ ≥ 8 N s . In particular, a soundness comparison uniform over these algorithms cannot have vanishing error throughout the scale s = Θ ( N ) s=\Theta(\sqrt{N}) s = Θ ( N ) .
Proof. The real final state has P ψ s = 0 P\psi_s=0 P ψ s = 0 , and a reference-only encoding commutes with P P P . Projecting the difference onto P P P and applying Proposition 4.16 gives the norm bound. Choose odd s s s equal to a sufficiently small constant multiple of N \sqrt{N} N to obtain a constant lower bound.
Comparison along a filtered history
Let ψ t \psi_t ψ t denote the real state after t t t queries and let U t U_t U t be the oracle-independent unitary before query t + 1 t+1 t + 1 . With F t F_t F t the complete compressed query at this step, define
ξ 0 = ∣ ∅ ⟩ , ξ t + 1 = K F t U t ξ t . (58) \xi_0 = \lvert\varnothing\rangle,\qquad\xi_{t+1}=K F_t U_t \xi_t. \tag*{(58)} ξ 0 = ∣ ∅ ⟩ , ξ t + 1 = K F t U t ξ t . (58)
The initial vector includes the same algorithm and reference registers as the real initial state. The maps K K K , W W W , V ~ \widetilde{V} V act only on the oracle register; they commute with the intervening free unitaries. Thus ∥ ξ t ∥ ≤ 1 \lVert\xi_t\rVert\le1 ∥ ξ t ∥ ≤ 1 , even though the states are not renormalized.
Corollary 4.19 (Filtered history comparison). For 3 T + 2 ≤ N 3T+2\le N 3 T + 2 ≤ N ,
∥ ξ T − W ψ T ∥ ≤ ( e + 5 ) T N − 2 T , (59) \lVert\xi_T-W\psi_T\rVert\le\frac{(e+5)T}{N-2T}, \tag*{(59)} ∥ ξ T − W ψ T ∥ ≤ N − 2 T ( e + 5 ) T , (59) ∥ ξ T − V ~ ψ T ∥ ≤ ( e + 5 ) T N − 2 T + T 2 ( N − 2 T + 1 ) . (60) \lVert\xi_T-\widetilde{V}\psi_T\rVert\le\frac{(e+5)T}{N-2T}+\frac{T}{2(N-2T+1)}. \tag*{(60)} ∥ ξ T − V ψ T ∥ ≤ N − 2 T ( e + 5 ) T + 2 ( N − 2 T + 1 ) T . (60) In particular the error is O ( T / N ) O(T/N) O ( T / N ) for T ≤ N / 4 T\le N/4 T ≤ N /4 , N ≥ 8 N\ge8 N ≥ 8 .
Proof. The initial real permutation state has degree zero and is sent to the empty database by both encodings. Just before query t + 1 t+1 t + 1 the real state has degree at most t t t ; the free unitary preserves this cutoff. Theorem 4.15 therefore bounds the new discrepancy against W ψ t + 1 W\psi_{t+1} W ψ t + 1 by ( e + 5 ) / ( N − 2 t ) (e+5)/(N-2t) ( e + 5 ) / ( N − 2 t ) , in addition to the preceding discrepancy propagated by the contraction K F t U t K F_t U_t K F t U t . Sum these bounds for 0 ≤ t < T 0\le t<T 0 ≤ t < T and use the endpoint estimate (28) once at the final degree. This proves both inequalities, including T = 0 T=0 T = 0 . □
The unprojected phase and the coherent response
A terminal read is performed without projecting back to the canonical subspace. To handle it we need a one-step estimate for the part discarded by K K K . There is a short way to obtain this estimate from the blocks already analyzed: W † W = G − 1 W^\dagger W=G^{-1} W † W = G − 1 , so a change in the norm of an encoded vector is governed by the change in the inverse Gram operator. Here G − 1 G^{-1} G − 1 denotes ⨁ j G j − 1 \bigoplus_j G_j^{-1} ⨁ j G j − 1 on the real degree decomposition.
Lemma 4.20 (Inverse Gram under one query). For 3 D + 5 ≤ N 3D+5\le N 3 D + 5 ≤ N ,
∥ ( G − 1 U z − U z G − 1 ) Π ≤ D ∥ ≤ 10 N − 2 D − 2 . (61) \left\lVert\left(G^{-1}U_z-U_zG^{-1}\right)\Pi_{\le D}\right\rVert\le\frac{10}{N-2D-2}. \tag*{(61)} ( G − 1 U z − U z G − 1 ) Π ≤ D ≤ N − 2 D − 2 10 . (61)
Proof. Only the middle, downward and upward blocks occur. On degree h h h the middle block is [ G h − 1 , A z ] [G_h^{-1},A_z] [ G h − 1 , A z ] . It vanishes on F h \mathcal{F}_h F h , by the fresh Gram commutation in Proposition 4.11. On F h ⊥ \mathcal{F}_h^\perp F h ⊥ , use an inverse of J = J h , x J=J_{h,x} J = J h , x only on this subspace. Since G h ⪰ I G_h\succeq I G h ⪰ I and J G h − 1 ⪰ ( 1 − η h ) I JG_h^{-1}\succeq(1-\eta_h)I J G h − 1 ⪰ ( 1 − η h ) I there, commutation implies J ⪰ ( 1 − η h ) I J\succeq(1-\eta_h)I J ⪰ ( 1 − η h ) I . Also [ J , A z ] = 0 [J,A_z]=0 [ J , A z ] = 0 . Consequently
∥ G h − 1 − J − 1 ∥ = ∥ J − 1 ( J G h − 1 − I ) ∥ ≤ η h 1 − η h , ∥ [ G h − 1 , A z ] ∥ ≤ 2 N − 2 h + 1 . \left\lVert G_h^{-1}-J^{-1}\right\rVert
=\left\lVert J^{-1}(JG_h^{-1}-I)\right\rVert
\le\frac{\eta_h}{1-\eta_h},
\qquad
\left\lVert[G_h^{-1},A_z]\right\rVert\le\frac{2}{N-2h+1}. G h − 1 − J − 1 = J − 1 ( J G h − 1 − I ) ≤ 1 − η h η h , [ G h − 1 , A z ] ≤ N − 2 h + 1 2 .
For the block Q = Π h − 1 U z Π h Q=\Pi_{h-1}U_z\Pi_h Q = Π h − 1 U z Π h multiply (4.35) by G h − 1 − 1 G_{h-1}^{-1} G h − 1 − 1 to obtain ∥ G h − 1 − 1 Q − Q G h − 1 ∥ ≤ 4 / ( N − h ) \lVert G_{h-1}^{-1}Q-QG_h^{-1}\rVert\le4/(N-h) ∥ G h − 1 − 1 Q − Q G h − 1 ∥ ≤ 4/ ( N − h ) . The upward block from h h h to h + 1 h+1 h + 1 is the adjoint, up to a sign, of this commutator bound for the downward query with conjugated phases, at source degree h + 1 h+1 h + 1 . Its norm is at most 4 / ( N − h − 1 ) 4/(N-h-1) 4/ ( N − h − 1 ) . This is why we require stability also at D + 1 D+1 D + 1 . Within each of the three families different sources give orthogonal targets. Summing their three suprema for h ≤ D h\le D h ≤ D proves (61) . □
Proposition 4.21 (Outgoing square leakage). If 3 D + 5 ≤ N 3D+5\le N 3 D + 5 ≤ N , then
∥ ( I − K ) F z W Π ≤ D ∥ ≤ 8 N − D , ∥ ( I − K ) F z K L ≤ D ∥ ≤ 8 e N − D . \left\|(I-K)F_zW\Pi_{\le D}\right\|\le\frac{8}{\sqrt{N-D}},\qquad\left\|(I-K)F_zKL_{\le D}\right\|\le\frac{8\sqrt{e}}{\sqrt{N-D}}. ∥ ( I − K ) F z W Π ≤ D ∥ ≤ N − D 8 , ∥ ( I − K ) F z K L ≤ D ∥ ≤ N − D 8 e . Also ∥ ( F z W − W U z ) Π ≤ D ∥ ≤ 28 / N − D \left\|(F_zW-WU_z)\Pi_{\le D}\right\|\le28/\sqrt{N-D} ∥ ( F z W − W U z ) Π ≤ D ∥ ≤ 28/ N − D .
Proof. Fix a unit vector ψ ∈ H ≤ D \psi\in\mathcal{H}_{\le D} ψ ∈ H ≤ D , and set r = K F z W ψ − W U z ψ r=KF_zW\psi-WU_z\psi r = K F z W ψ − W U z ψ . By Theorem 4.15 , ∥ r ∥ ≤ ( e + 5 ) / ( N − 2 D ) \|r\|\le(e+5)/(N-2D) ∥ r ∥ ≤ ( e + 5 ) / ( N − 2 D ) . Pythagoras, unitarity of F z F_z F z , and ∥ W U z ψ ∥ ≤ 1 \|WU_z\psi\|\le1 ∥ W U z ψ ∥ ≤ 1 give
∥ ( I − K ) F z W ψ ∥ 2 = ∥ W ψ ∥ 2 − ∥ W U z ψ + r ∥ 2 ≤ ⟨ ψ , ( G − 1 − U z † G − 1 U z ) ψ ⟩ + 2 ∥ r ∥ ≤ 26 N − 2 D − 2 . \begin{aligned}
\left\|(I-K)F_zW\psi\right\|^2=\|W\psi\|^2-\|WU_z\psi+r\|^2 \\
&\le\left\langle\psi,\left(G^{-1}-U_z^\dagger G^{-1}U_z\right)\psi\right\rangle+2\|r\|\le\frac{26}{N-2D-2}.
\end{aligned} ∥ ( I − K ) F z W ψ ∥ 2 = ∥ W ψ ∥ 2 − ∥ W U z ψ + r ∥ 2 ≤ ⟨ ψ , ( G − 1 − U z † G − 1 U z ) ψ ⟩ + 2∥ r ∥ ≤ N − 2 D − 2 26 .
For the final inequality apply Lemma 4.20 and e + 5 < 8 e+5<8 e + 5 < 8 . Since N − D ≤ 2 ( N − 2 D − 2 ) N-D\le2(N-2D-2) N − D ≤ 2 ( N − 2 D − 2 ) , the first claimed bound follows. For the second, write a canonical vector a = W ψ a=W\psi a = W ψ ; ∥ ψ ∥ ≤ e ∥ a ∥ \|\psi\|\le\sqrt{e}\|a\| ∥ ψ ∥ ≤ e ∥ a ∥ on the specified degrees. Finally add ∥ r ∥ \|r\| ∥ r ∥ to the first bound. Here N − D ≤ N − 2 D \sqrt{N-D}\le N-2D N − D ≤ N − 2 D , since N − D ≤ 2 ( N − 2 D − 2 ) N-D\le2(N-2D-2) N − D ≤ 2 ( N − 2 D − 2 ) and N − 2 D ≥ 2 N-2D\ge2 N − 2 D ≥ 2 . This gives even 16 / N − D 16/\sqrt{N-D} 16/ N − D ; the stated constant leaves room in the following estimates.
Introduce an answer register with orthonormal basis indexed by { ⊥ } ∪ [ N ] \{\perp\}\cup[N] { ⊥ } ∪ [ N ] ; S y S_y S y exchanges ⊥ \perp ⊥ and y y y and fixes the other entries. The real forward read is P = ∑ y S y ⊗ Q y P=\sum_y S_y\otimes Q_y P = ∑ y S y ⊗ Q y . The compressed forward read Q Q Q decompresses by C x C_x C x , applies S y S_y S y if the decompressed record at x x x is y y y (and the identity if there is no record), and recompresses. We need a bound for an arbitrary initial response: the verifier can have retained earlier responses and be entangled with them.
Proposition 4.22 (Comparison for a coherent forward read). For 3 D + 5 ≤ N 3D+5\le N 3 D + 5 ≤ N ,
∥ ( Q V ~ − V ~ P ) ( I ⊗ Π ≤ D ) ∥ ≤ 60 N − D + D + 1 N − 2 D − 1 . \left\|(Q\widetilde V-\widetilde V P)(I\otimes\Pi_{\le D})\right\|\le\frac{60}{\sqrt{N-D}}+\frac{D+1}{N-2D-1}. ( Q V − V P ) ( I ⊗ Π ≤ D ) ≤ N − D 60 + N − 2 D − 1 D + 1 . The bound is uniform in the input x x x , including coherent controls.
Proof. First use W W W in both encodings and define outcome defects
B y = ( M y W − W Q y ) ∣ H ≤ D . B_y=(M_yW-WQ_y)|_{\mathcal{H}_{\le D}}. B y = ( M y W − W Q y ) ∣ H ≤ D .
The decompressed unused-answer projector annihilates W W W , by the domain harmonicity relations. Hence ∑ y B y = 0 \sum_y B_y=0 ∑ y B y = 0 and, for signs σ y ∈ { − 1 , 1 } \sigma_y\in\{-1,1\} σ y ∈ { − 1 , 1 } , ∑ y σ y B y = ( F σ W − W U σ ) ∣ H ≤ D \sum_y\sigma_yB_y=(F_\sigma W-WU_\sigma)|_{\mathcal{H}_{\le D}} ∑ y σ y B y = ( F σ W − W U σ ) ∣ H ≤ D . Proposition 4.21 gives this signed sum norm at most δ = 28 / N − D \delta=28/\sqrt{N-D} δ = 28/ N − D , for every sign choice.
The full response defect with W W W is ∑ y S y ⊗ B y \sum_y S_y\otimes B_y ∑ y S y ⊗ B y . On a vector ( v ⊥ , ( v y ) y ) (v_\perp,(v_y)_y) ( v ⊥ , ( v y ) y ) its components are
( ∑ y B y v y , ( B y v ⊥ − B y v y ) y ) , \left(\sum_y B_yv_y,\ (B_yv_\perp-B_yv_y)_y\right), ( y ∑ B y v y , ( B y v ⊥ − B y v y ) y ) ,
where the minus sign uses ∑ y B y = 0 \sum_y B_y=0 ∑ y B y = 0 . For independent uniform signs, expanding the squared norm gives
∑ y ∥ B y v ∥ 2 = E σ ∥ ∑ y σ y B y v ∥ 2 ≤ δ 2 ∥ v ∥ 2 . \sum_y\|B_yv\|^2=\mathbb{E}_{\sigma}\left\|\sum_y\sigma_yB_yv\right\|^2\le\delta^2\|v\|^2. y ∑ ∥ B y v ∥ 2 = E σ y ∑ σ y B y v 2 ≤ δ 2 ∥ v ∥ 2 .
The same holds for B y † B_y^\dagger B y † by taking adjoints of the signed sums. Thus both the column v ↦ ( B y v ) y v\mapsto(B_yv)_y v ↦ ( B y v ) y and the row ( v y ) y ↦ ∑ y B y v y (v_y)_y\mapsto\sum_yB_yv_y ( v y ) y ↦ ∑ y B y v y are contractions up to the factor δ \delta δ . The off-diagonal part of the displayed response defect has norm at most δ \delta δ , since its row and column have disjoint input and output summands. The diagonal part has norm at most max y ∥ B y ∥ ≤ δ \max_y \|B_y\| \le\delta max y ∥ B y ∥ ≤ δ . This proves the comparison with W W W with error 2 δ 2\delta 2 δ , independently of the number of outcomes.
Replace W W W by W ~ \widetilde{W} W once on each side. The starting real degree is at most D D D and the final one at most D + 1 D+1 D + 1 . The two reads are unitary, and the endpoint bound (28) contributes at most
D 2 ( N − 2 D + 1 ) + D + 1 2 ( N − 2 D − 1 ) ≤ D + 1 N − 2 D − 1 . \frac{D}{2(N-2D+1)}+\frac{D+1}{2(N-2D-1)}\leq\frac{D+1}{N-2D-1}. 2 ( N − 2 D + 1 ) D + 2 ( N − 2 D − 1 ) D + 1 ≤ N − 2 D − 1 D + 1 .
Use 2 δ ≤ 60 / N − D 2\delta\leq60/\sqrt{N-D} 2 δ ≤ 60/ N − D . Direct sums over the input and the phase controls preserve each bound. □
The two comparisons play different roles. We sum the canonical return, of order 1 / ( N − 2 D ) 1/(N-2D) 1/ ( N − 2 D ) , throughout the algorithm. The 1 / N − D 1/\sqrt{N-D} 1/ N − D comparison is used only for a bounded number of terminal reads. Summing the latter error throughout the history would lose the range needed for the lower bound.
Capacity extension and harmonic projections
The weighted capacity-extension map
The target has one more slot than the auxiliary database. Carrying an assignment across unchanged preserves all observed function labels, but changes the number of physical databases realizing the assignment. The square root of the ratio of these orbit sizes determines the transfer weight. Throughout this subsection we work on the normalized bucket-orbit spaces of § 3.3 . All identities extend by the identity on the algorithm registers. Let P ∗ P_* P ∗ be the diagonal projector which, in mark branch b b b , tests whether all k + 1 k+1 k + 1 slots of the marked bucket are recorded. Write C j ( d ) = # { i : o i ( d ) ≥ j } C_j(d)=\#\{i:o_i(d)\geq j\} C j ( d ) = # { i : o i ( d ) ≥ j } for a feasible bucket assignment, and use the same notation on physical databases via their induced assignment. Write C ^ j \widehat{C}_j C j for the diagonal operator with these eigenvalues; the underlying space (source, or one marked target branch) will be clear from its argument.
Definition 5.1 (Capacity extension). For a source orbit assignment d d d and a mark b b b , let
J b ∣ d ‾ ⟩ κ − = r o b ( d ) ∣ d ‾ ⟩ κ + , b , r h = ( k + 1 ) h ( k ) h = k + 1 k + 1 − h , 0 ≤ h ≤ k . (62) J_b\lvert\overline{d}\rangle_{\kappa^-}=r_{o_b(d)}\lvert\overline{d}\rangle_{\kappa^+,b},\qquad r_h=\sqrt{\frac{(k+1)_h}{(k)_h}}=\sqrt{\frac{k+1}{k+1-h}},\qquad0\leq h\leq k. \tag*{(62)} J b ∣ d ⟩ κ − = r o b ( d ) ∣ d ⟩ κ + , b , r h = ( k ) h ( k + 1 ) h = k + 1 − h k + 1 , 0 ≤ h ≤ k . (62) Define
J v = 1 m ∑ b = 1 m ∣ b ⟩ ⊗ J b v . (63) Jv=\frac{1}{\sqrt{m}}\sum_{b=1}^{m}\lvert b\rangle\otimes J_bv. \tag*{(63)} J v = m 1 b = 1 ∑ m ∣ b ⟩ ⊗ J b v . (63) On the physical space this means first extracting the source orbit coefficients and then lifting the weighted coefficients to the target orbits. Explicitly, if E − E_- E − and E + , b E_{+,b} E + , b send feasible assignment coordinates to their normalized physical orbit vectors, the physical map in branch b b b is E + , b J b E − † E_{+,b}J_bE_-^\dagger E + , b J b E − † , with J b J_b J b on the right understood in assignment coordinates. It vanishes on the orthogonal complement of the source orbit space. It leaves a retained injective label table unchanged.
Lemma 5.2 (Global properties). The map J J J sends the auxiliary initial state to the target initial state. On the source orbit subspace,
I ≤ J † J ≤ ( k + 1 ) I , P ∗ J = 0 , L + J = J L − . (64) I\leq J^\dagger J\leq(k+1)I,\qquad P_*J=0,\qquad L^+J=JL^-. \tag*{(64)} I ≤ J † J ≤ ( k + 1 ) I , P ∗ J = 0 , L + J = J L − . (64) These statements hold with arbitrary auxiliary registers and with a coherently retained injective function-label table.
Proof. Different source bucket assignments have orthogonal images, and
J † J ∣ d ˉ ⟩ = ( 1 m ∑ b k + 1 k + 1 − o b ( d ) ) ∣ d ˉ ⟩ . J^\dagger J|\bar{d}\rangle= \left(\frac{1}{m}\sum_b \frac{k+1}{k+1-o_b(d)}\right)|\bar{d}\rangle. J † J ∣ d ˉ ⟩ = ( m 1 b ∑ k + 1 − o b ( d ) k + 1 ) ∣ d ˉ ⟩ .
Each summand lies between 1 and k + 1 k+1 k + 1 . For the empty assignment, all weights are 1, which gives the uniform target mark and empty database. A source assignment has o b ( d ) ≤ k o_b(d) \le k o b ( d ) ≤ k , so its image never contains k + 1 k+1 k + 1 records in the marked bucket. Finally, L − L^- L − and L + L^+ L + denote the lookup part of a complete query in the source and the marked target, respectively. Its phase depends only on the algorithm control and, when x x x is recorded, on its bucket label. These are unchanged by every J b J_b J b , proving the intertwining identity.
Relation to the hidden transposition. Fix a mark b b b and relabel its target slots so the unmarked buckets agree with those of the source. Write the marked target bucket as B b ∪ { s 0 } B_b \cup\{s_0\} B b ∪ { s 0 } , with ∣ B b ∣ = k |B_b|=k ∣ B b ∣ = k . Embed a source orbit with h h h marked records into the subspace with no record occupying the added slot. Let Sym b \operatorname{Sym}_b Sym b average permutations of these k + 1 k+1 k + 1 slots. The fraction of the target orbit that avoids s 0 s_0 s 0 is
p h = ( k ) h ( k + 1 ) h = k + 1 − h k + 1 . (65) p_h = \frac{(k)_h}{(k+1)_h} = \frac{k+1-h}{k+1}. \tag*{(65)} p h = ( k + 1 ) h ( k ) h = k + 1 k + 1 − h . (65)
Hence Sym b \operatorname{Sym}_b Sym b sends the embedded source vector to p h \sqrt{p_h} p h times the normalized target orbit, whereas J b J_b J b sends it to p h − 1 / 2 p_h^{-1/2} p h − 1/2 times that orbit. Equivalently, on this source orbit, J b = p h − 1 Sym b J_b=p_h^{-1}\operatorname{Sym}_b J b = p h − 1 Sym b after the natural embedding. This is the normalization supplied by the hidden-transposition argument; it is not an unweighted averaging map.
The local capacity-comparison estimate
We first estimate the failure of the capacity-extension map to intertwine the two local decompression operators. This estimate is stable under tensoring with arbitrary auxiliary registers.
Fix integers k ≥ 2 k \ge2 k ≥ 2 and m ≥ 1 m \ge1 m ≥ 1 , and put N − = k m N_- = km N − = k m and N = N − + 1 N=N_-+1 N = N − + 1 . Let H H H be a source database after deletion of the queried input. Write
h i = o i ( H ) , a i = k − h i ( 1 ≤ i ≤ m ) , s = N − − ∣ H ∣ = ∑ i = 1 m a i , d = s + 1 = N − ∣ H ∣ . h_i=o_i(H), \qquad a_i=k-h_i \quad(1 \le i \le m), \qquad s=N_- - |H|=\sum_{i=1}^{m}a_i, \qquad d=s+1=N-|H|. h i = o i ( H ) , a i = k − h i ( 1 ≤ i ≤ m ) , s = N − − ∣ H ∣ = i = 1 ∑ m a i , d = s + 1 = N − ∣ H ∣.
Here a i a_i a i denotes an available capacity, not a function label. For the local comparison, assume ∣ H ∣ ≤ N / 3 |H| \le N/3 ∣ H ∣ ≤ N /3 . In particular, s ≥ 1 s \ge1 s ≥ 1 , d ≥ 2 d \ge2 d ≥ 2 , and m ≤ d m \le d m ≤ d . Let
I H = { i ∈ [ m ] : a i > 0 } , F H = # { i ∈ [ m ] : a i = 0 } = C k ( H ) . I_H=\{i\in[m]:a_i>0\}, \qquad F_H=\#\{i\in[m]:a_i=0\}=C_k(H). I H = { i ∈ [ m ] : a i > 0 } , F H = # { i ∈ [ m ] : a i = 0 } = C k ( H ) .
The source local space has orthonormal basis ∣ v ⟩ |v\rangle ∣ v ⟩ and ∣ i ⟩ |i\rangle ∣ i ⟩ for i ∈ I H i\in I_H i ∈ I H , where ∣ v ⟩ |v\rangle ∣ v ⟩ denotes the database with the queried input unrecorded. For a target mark b b b , the target local space has orthonormal basis ∣ v ⟩ |v\rangle ∣ v ⟩ and ∣ i ⟩ |i\rangle ∣ i ⟩ for i ∈ I H ∪ { b } i\in I_H\cup\{b\} i ∈ I H ∪ { b } . The corresponding uniform extension vectors in the normalized orbit bases are
u − = ∑ i ∈ I H a i s ∣ i ⟩ , u b + = ∑ i ∈ I H ∪ { b } a i + 1 { i = b } d ∣ i ⟩ . (66) u^-=\sum_{i\in I_H}\sqrt{\frac{a_i}{s}}\,|i\rangle,\qquad
u_b^+=\sum_{i\in I_H\cup\{b\}}\sqrt{\frac{a_i+1_{\{i=b\}}}{d}}\,|i\rangle. \tag*{(66)} u − = i ∈ I H ∑ s a i ∣ i ⟩ , u b + = i ∈ I H ∪ { b } ∑ d a i + 1 { i = b } ∣ i ⟩ . (66)
Both vectors are unit vectors orthogonal to ∣ v ⟩ |v\rangle ∣ v ⟩ . All coordinates excluded by capacity have zero amplitude. A deleted assignment that is itself infeasible in the source likewise has no source component. These facts also hold just after a source query: each local query preserves the span of feasible extensions. For such a vector u u u , define
C ( u ) = I − ∣ v ⟩ ⟨ v ∣ − ∣ u ⟩ ⟨ u ∣ + ∣ v ⟩ ⟨ u ∣ + ∣ u ⟩ ⟨ v ∣ . C(u) = I - |v\rangle\langle v| - |u\rangle\langle u| + |v\rangle\langle u| + |u\rangle\langle v|. C ( u ) = I − ∣ v ⟩ ⟨ v ∣ − ∣ u ⟩ ⟨ u ∣ + ∣ v ⟩ ⟨ u ∣ + ∣ u ⟩ ⟨ v ∣.
Thus C ( u ) C(u) C ( u ) exchanges ∣ v ⟩ |v\rangle ∣ v ⟩ and u u u and fixes their orthogonal complement. We use C − = C ( u − ) C^{-}=C(u^{-}) C − = C ( u − ) and C b + = C ( u b + ) C_b^{+}=C(u_b^{+}) C b + = C ( u b + ) on their respective local spaces. Identical basis labels in different local spaces are understood through their canonical coordinate embeddings.
On this local space the capacity-extension map is
J b ∣ v ⟩ = r b ∣ v ⟩ , r b = k + 1 a b + 1 , J b ∣ i ⟩ = r b ∣ i ⟩ ( i ∈ I H , i ≠ b ) , J b ∣ b ⟩ = r b a b + 1 a b ∣ b ⟩ ( a b > 0 ) . (67) \begin{aligned}
J_b|v\rangle&= r_b|v\rangle,\qquad r_b=\sqrt{\frac{k+1}{a_b+1}},\\
J_b|i\rangle&= r_b|i\rangle\qquad(i\in I_H,\ i\ne b),\\
J_b|b\rangle&= r_b\sqrt{\frac{a_b+1}{a_b}}|b\rangle\qquad(a_b>0). \tag*{(67)}
\end{aligned} J b ∣ v ⟩ J b ∣ i ⟩ J b ∣ b ⟩ = r b ∣ v ⟩ , r b = a b + 1 k + 1 , = r b ∣ i ⟩ ( i ∈ I H , i = b ) , = r b a b a b + 1 ∣ b ⟩ ( a b > 0 ) . (67)
When a b = 0 a_b=0 a b = 0 , the source has no ∣ b ⟩ |b\rangle ∣ b ⟩ coordinate, so the last line is absent. In particular, no division by a b a_b a b is used in that case. Set
D b = C b + J b − J b C − . D_b=C_b^{+}J_b-J_bC^{-}. D b = C b + J b − J b C − .
Lemma 5.3 (Local decompression defect). For every auxiliary Hilbert space A \mathcal{A} A and every vector φ \varphi φ in the source local space tensored with A \mathcal{A} A ,
1 m ∑ b = 1 m ∥ ( D b ⊗ I A ) φ ∥ 2 ≤ 32 ( k + 1 ) ( 1 + C k ( H ) ) m d ∥ φ ∥ 2 . (68) \frac{1}{m}\sum_{b=1}^{m}\left\|\left(D_b\otimes I_{\mathcal{A}}\right)\varphi\right\|^2\leq\frac{32(k+1)(1+C_k(H))}{md}\|\varphi\|^2. \tag*{(68)} m 1 b = 1 ∑ m ∥ ( D b ⊗ I A ) φ ∥ 2 ≤ m d 32 ( k + 1 ) ( 1 + C k ( H )) ∥ φ ∥ 2 . (68)
Proof. We suppress identity operators on A \mathcal{A} A . Put
ρ = d s , α = 1 − ρ − 1 . \rho=\sqrt{\frac{d}{s}},\qquad\alpha=1-\rho^{-1}. ρ = s d , α = 1 − ρ − 1 .
Since s = d − 1 s=d-1 s = d − 1 and $d\geq 2,
0 ≤ α ≤ 1 d , 0 ≤ ρ − 1 ≤ 2 d , r b 2 ≤ k + 1. (69) 0\leq\alpha\leq\frac{1}{d},\qquad0\leq\rho-1\leq\frac{2}{d},\qquad r_b^2\leq k+1. \tag*{(69)} 0 ≤ α ≤ d 1 , 0 ≤ ρ − 1 ≤ d 2 , r b 2 ≤ k + 1. (69)
Define a vector in the source local space and a vector in the target local space, respectively, by
p b = J b † u b + − r b u − , q b = J b u − − r b u b + . p_b=J_b^{\dagger}u_b^{+}-r_bu^{-},\qquad q_b=J_bu^{-}-r_bu_b^{+}. p b = J b † u b + − r b u − , q b = J b u − − r b u b + .
Expanding the two decompression operators, and using J b ∣ v ⟩ = r b ∣ v ⟩ J_b|v\rangle=r_b|v\rangle J b ∣ v ⟩ = r b ∣ v ⟩ and J b † ∣ v ⟩ = r b ∣ v ⟩ J_b^{\dagger}|v\rangle=r_b|v\rangle J b † ∣ v ⟩ = r b ∣ v ⟩ , gives the exact identity
D b = ( ∣ v ⟩ − ∣ u b + ⟩ ) ⟨ p b ∣ + ∣ q b ⟩ ( ⟨ u − ∣ − ⟨ v ∣ ) . (70) D_b=(|v\rangle-|u_b^{+}\rangle)\langle p_b|+|q_b\rangle(\langle u^{-}|-\langle v|). \tag*{(70)} D b = ( ∣ v ⟩ − ∣ u b + ⟩) ⟨ p b ∣ + ∣ q b ⟩ (⟨ u − ∣ − ⟨ v ∣ ) . (70)
For a b > 0 a_b>0 a b > 0 , direct substitution in (66) –(67) yields
J b u − = r b ρ u b + , p b = − r b α u − + r b a b d ∣ b ⟩ , q b = r b ( ρ − 1 ) u b + . (71) \begin{aligned}
J_bu^{-}&=r_b\rho\,u_b^{+},\\
p_b&=-r_b\alpha\,u^{-}+\frac{r_b}{\sqrt{a_bd}}|b\rangle,\\
q_b&=r_b(\rho-1)u_b^{+}. \tag*{(71)}
\end{aligned} J b u − p b q b = r b ρ u b + , = − r b α u − + a b d r b ∣ b ⟩ , = r b ( ρ − 1 ) u b + . (71)
If a b = 0 a_b=0 a b = 0 , the target has one additional coordinate ∣ b ⟩ |b\rangle ∣ b ⟩ , orthogonal to the embedded source space, and instead
u b + = ρ − 1 u − + d − 1 / 2 ∣ b ⟩ , p b = − r b α u − , q b = r b ( α u − − d − 1 / 2 ∣ b ⟩ ) , r b 2 = k + 1. (72) \begin{aligned}
u_b^+ &= \rho^{-1}u^- + d^{-1/2}|b\rangle,\\
p_b &= -r_b\alpha\,u^-,\\
q_b &= r_b\left(\alpha u^- - d^{-1/2}|b\rangle\right),\qquad r_b^2=k+1. \tag*{(72)}
\end{aligned} u b + p b q b = ρ − 1 u − + d − 1/2 ∣ b ⟩ , = − r b α u − , = r b ( α u − − d − 1/2 ∣ b ⟩ ) , r b 2 = k + 1. (72)
The orthogonality of u − u^- u − and the new coordinate ∣ b ⟩ |b\rangle ∣ b ⟩ in the last case is essential for the estimate below.
Write the auxiliary-register-valued coordinates of the input as
φ = ∣ v ⟩ ⊗ φ v + ∑ i ∈ I H ∣ i ⟩ ⊗ φ i , ξ = ( ⟨ u − ∣ ⊗ I A ) φ . \varphi=|v\rangle\otimes\varphi_v+\sum_{i\in I_H}|i\rangle\otimes\varphi_i,\qquad\xi=(\langle u^-|\otimes I_A)\varphi. φ = ∣ v ⟩ ⊗ φ v + i ∈ I H ∑ ∣ i ⟩ ⊗ φ i , ξ = (⟨ u − ∣ ⊗ I A ) φ .
Then ∥ ξ ∥ ≤ ∥ φ ∥ \|\xi\|\le\|\varphi\| ∥ ξ ∥ ≤ ∥ φ ∥ and ∑ i ∈ I H ∥ φ i ∥ 2 ≤ ∥ φ ∥ 2 \sum_{i\in I_H}\|\varphi_i\|^2\le\|\varphi\|^2 ∑ i ∈ I H ∥ φ i ∥ 2 ≤ ∥ φ ∥ 2 . Equations (71) and (72) imply
( ⟨ p b ∣ ⊗ I A ) φ = − r b α ξ + { r b a b d φ b , a b > 0 , 0 , a b = 0. (\langle p_b|\otimes I_A)\varphi=-r_b\alpha\,\xi+
\begin{cases}
\dfrac{r_b}{\sqrt{a_bd}}\varphi_b, & a_b>0,\\
0, & a_b=0.
\end{cases} (⟨ p b ∣ ⊗ I A ) φ = − r b α ξ + ⎩ ⎨ ⎧ a b d r b φ b , 0 , a b > 0 , a b = 0.
Using ∥ x + y ∥ 2 ≤ 2 ∥ x ∥ 2 + 2 ∥ y ∥ 2 \|x+y\|^2\le2\|x\|^2+2\|y\|^2 ∥ x + y ∥ 2 ≤ 2∥ x ∥ 2 + 2∥ y ∥ 2 , we obtain
1 m ∑ b ∥ ( ⟨ p b ∣ ⊗ I A ) φ ∥ 2 ≤ 2 ( k + 1 ) α 2 ∥ ξ ∥ 2 + 2 ( k + 1 ) m d ∑ b ∈ I H ∥ φ b ∥ 2 a b ≤ 2 ( k + 1 ) ( 1 d 2 + 1 m d ) ∥ φ ∥ 2 . (73) \begin{aligned}
\frac{1}{m}\sum_b\|(\langle p_b|\otimes I_A)\varphi\|^2
&\le2(k+1)\alpha^2\|\xi\|^2+\frac{2(k+1)}{md}\sum_{b\in I_H}\frac{\|\varphi_b\|^2}{a_b}\\
&\le2(k+1)\left(\frac{1}{d^2}+\frac{1}{md}\right)\|\varphi\|^2. \tag*{(73)}
\end{aligned} m 1 b ∑ ∥ (⟨ p b ∣ ⊗ I A ) φ ∥ 2 ≤ 2 ( k + 1 ) α 2 ∥ ξ ∥ 2 + m d 2 ( k + 1 ) b ∈ I H ∑ a b ∥ φ b ∥ 2 ≤ 2 ( k + 1 ) ( d 2 1 + m d 1 ) ∥ φ ∥ 2 . (73)
In particular, the second term is summed over orthogonal source coordinates. It is not estimated by m m m copies of the full input norm.
For a b > 0 a_b>0 a b > 0 , (71) gives ∥ q b ∥ 2 ≤ 4 ( k + 1 ) / d 2 \|q_b\|^2\le4(k+1)/d^2 ∥ q b ∥ 2 ≤ 4 ( k + 1 ) / d 2 . For a b = 0 a_b=0 a b = 0 , (72) gives
∥ q b ∥ 2 = ( k + 1 ) ( α 2 + 1 d ) ≤ ( k + 1 ) ( 1 d 2 + 1 d ) . \|q_b\|^2=(k+1)\left(\alpha^2+\frac{1}{d}\right)\le(k+1)\left(\frac{1}{d^2}+\frac{1}{d}\right). ∥ q b ∥ 2 = ( k + 1 ) ( α 2 + d 1 ) ≤ ( k + 1 ) ( d 2 1 + d 1 ) .
Consequently,
1 m ∑ b ∥ q b ∥ 2 ≤ 4 ( k + 1 ) d 2 + ( k + 1 ) C k ( H ) m d . (74) \frac{1}{m}\sum_b\|q_b\|^2\le\frac{4(k+1)}{d^2}+\frac{(k+1)C_k(H)}{md}. \tag*{(74)} m 1 b ∑ ∥ q b ∥ 2 ≤ d 2 4 ( k + 1 ) + m d ( k + 1 ) C k ( H ) . (74)
Finally, ∥ ∣ v ⟩ − ∣ u b + ⟩ ∥ = 2 \||v\rangle-|u_b^+\rangle\|=\sqrt{2} ∥∣ v ⟩ − ∣ u b + ⟩ ∥ = 2 and
∥ ( ( ⟨ u − ∣ − ⟨ v ∣ ) ⊗ I A ) φ ∥ ≤ 2 ∥ φ ∥ . \|((\langle u^-|-\langle v|)\otimes I_A)\varphi\|\le\sqrt{2}\,\|\varphi\|. ∥ ((⟨ u − ∣ − ⟨ v ∣ ) ⊗ I A ) φ ∥ ≤ 2 ∥ φ ∥.
Applying (70) , followed by (73) and (74) , yields
1 m ∑ b ∥ D b φ ∥ 2 ≤ 4 m ∑ b ∥ ( ⟨ p b ∣ ⊗ I A ) φ ∥ 2 + 4 m ∑ b ∥ q b ∥ 2 ∥ φ ∥ 2 ≤ ( k + 1 ) ( 24 d 2 + 8 + 4 C k ( H ) m d ) ∥ φ ∥ 2 ≤ 32 ( k + 1 ) ( 1 + C k ( H ) ) m d ∥ φ ∥ 2 . (75) \begin{aligned}
\frac{1}{m}\sum_b\|D_b\varphi\|^2
&\le\frac{4}{m}\sum_b\|(\langle p_b|\otimes I_A)\varphi\|^2+\frac{4}{m}\sum_b\|q_b\|^2\|\varphi\|^2\\
&\le(k+1)\left(\frac{24}{d^2}+\frac{8+4C_k(H)}{md}\right)\|\varphi\|^2\\
&\le\frac{32(k+1)(1+C_k(H))}{md}\|\varphi\|^2. \tag*{(75)}
\end{aligned} m 1 b ∑ ∥ D b φ ∥ 2 ≤ m 4 b ∑ ∥ (⟨ p b ∣ ⊗ I A ) φ ∥ 2 + m 4 b ∑ ∥ q b ∥ 2 ∥ φ ∥ 2 ≤ ( k + 1 ) ( d 2 24 + m d 8 + 4 C k ( H ) ) ∥ φ ∥ 2 ≤ m d 32 ( k + 1 ) ( 1 + C k ( H )) ∥ φ ∥ 2 . (75)
The last step uses m ≤ d m\le d m ≤ d and 32 + 4 C k ( H ) ≤ 32 ( 1 + C k ( H ) ) 32+4C_k(H)\le32(1+C_k(H)) 32 + 4 C k ( H ) ≤ 32 ( 1 + C k ( H )) .
We next pass from decompression to a complete function-query block. The phase operation causes no additional comparison error: a source record assigned to bucket i i i and its image under J b J_b J b have the same function label, including when i = b i=b i = b .
Corollary 5.4 (Local complete-query defect). Let L − L^- L − and L b + L_b^+ L b + be unitary lookup operators on the source and target local spaces, with common auxiliary registers, such that
L b + J b = J b L − . L_b^+J_b=J_bL^-. L b + J b = J b L − . Suppose that L − L^- L − is the same source operator for every b b b . Define
F − = C − L − C − , F b + = C b + L b + C b + . F^-=C^-L^-C^- ,\qquad F_b^+=C_b^+L_b^+C_b^+. F − = C − L − C − , F b + = C b + L b + C b + . Then, for every source vector φ \varphi φ , including arbitrary auxiliary registers,
1 m ∑ b ∥ ( F b + J b − J b F − ) φ ∥ 2 ≤ 128 ( k + 1 ) ( 1 + C k ( H ) ) m d ∥ φ ∥ 2 . (76) \frac{1}{m}\sum_b\left\|\left(F_b^+J_b-J_bF^-\right)\varphi\right\|^2\leq\frac{128(k+1)(1+C_k(H))}{md}\|\varphi\|^2. \tag*{(76)} m 1 b ∑ ( F b + J b − J b F − ) φ 2 ≤ m d 128 ( k + 1 ) ( 1 + C k ( H )) ∥ φ ∥ 2 . (76)
Proof. The intertwining assumption and the definition of D b D_b D b give
F b + J b − J b F − = C b + L b + D b + D b L − C − . (77) F_b^+J_b-J_bF^-=C_b^+L_b^+D_b+D_bL^-C^-. \tag*{(77)} F b + J b − J b F − = C b + L b + D b + D b L − C − . (77)
Take the direct sum over b b b , normalized by m − 1 / 2 m^{-1/2} m − 1/2 . The first term has the same direct-sum norm as m − 1 / 2 ⨁ b D b φ m^{-1/2}\bigoplus_bD_b\varphi m − 1/2 ⨁ b D b φ , because each C b + L b + C_b^+L_b^+ C b + L b + is unitary. Apply Lemma 5.3 to this term. Apply the same lemma to the second term with the common source vector L − C − φ L^-C^-\varphi L − C − φ , whose norm equals ∥ φ ∥ \|\varphi\| ∥ φ ∥ . The triangle inequality gives twice the square root of the bound in (68) . Squaring proves (76) . □
Remark 5.5. For a function phase query, the lookup operators act by the phase associated with bucket i i i on coordinate ∣ i ⟩ |i\rangle ∣ i ⟩ , and by the identity on ∣ v ⟩ |v\rangle ∣ v ⟩ . Thus the intertwining hypothesis holds even when the function-label table and the phase character are retained as quantum registers. All preceding estimates are uniform over these registers. They also apply to superpositions of queried inputs: different queried inputs and deletion databases H H H define orthogonal local components, and the mark register makes the different target branches b b b orthogonal.
A state-dependent complete-query defect
The local estimate yields a state-dependent query comparison in terms of the auxiliary collision counter.
Lemma 5.6 (Global one-query comparison). For any bucket-invariant source vector v v v supported on ∣ D ∣ ≤ t ≤ N / 3 |D|\leq t\leq N/3 ∣ D ∣ ≤ t ≤ N /3 ,
∥ ( F + J − J F − ) v ∥ ≤ K k ( 0 ) N ∥ v ∥ 2 + ⟨ v , C ^ k v ⟩ , (78) \left\|\left(F^+J-JF^-\right)v\right\|\leq\frac{K_k^{(0)}}{N}\sqrt{\|v\|^2+\langle v,\widehat C_kv\rangle}, \tag*{(78)} ( F + J − J F − ) v ≤ N K k ( 0 ) ∥ v ∥ 2 + ⟨ v , C k v ⟩ , (78) where K k ( 0 ) K_k^{(0)} K k ( 0 ) depends only on k k k .
Proof. Decompose v v v into orthogonal queried-input, phase-character, and deleted-assignment components v x , c , H v_{x,c,H} v x , c , H . Their images under the query defect remain orthogonal, including the target mark register. Corollary 5.4 therefore gives
∥ ( F + J − J F − ) v ∥ 2 ≤ 128 ( k + 1 ) m ( N − t ) ∑ x , c , H ( 1 + C k ( H ) ) ∥ v x , c , H ∥ 2 . (79) \left\|\left(F^+J-JF^-\right)v\right\|^2\leq\frac{128(k+1)}{m(N-t)}\sum_{x,c,H}(1+C_k(H))\|v_{x,c,H}\|^2. \tag*{(79)} ( F + J − J F − ) v 2 ≤ m ( N − t ) 128 ( k + 1 ) x , c , H ∑ ( 1 + C k ( H )) ∥ v x , c , H ∥ 2 . (79)
Within each component, C k ( H ) ≤ C k ( D ) C_k(H) \le C_k(D) C k ( H ) ≤ C k ( D ) . Since m = ( N − 1 ) / k ≥ N / ( 2 k ) m = (N - 1)/k \ge N/(2k) m = ( N − 1 ) / k ≥ N / ( 2 k ) and N − t ≥ 2 N / 3 N - t \ge2N/3 N − t ≥ 2 N /3 , the right-hand side is at most
384 k ( k + 1 ) N 2 ( ∥ v ∥ 2 + ⟨ v , C ~ k v ⟩ ) . \frac{384k(k+1)}{N^2}\left(\lVert v\rVert^2 + \langle v,\widetilde{C}_k v\rangle\right). N 2 384 k ( k + 1 ) ( ∥ v ∥ 2 + ⟨ v , C k v ⟩ ) .
Thus, for example, K k ( 0 ) = 384 k ( k + 1 ) K_k^{(0)} = \sqrt{384k(k+1)} K k ( 0 ) = 384 k ( k + 1 ) is sufficient.
Harmonic projections and the capacity-extension map
We next show that inserting canonical projections preserves the estimates needed for the capacity comparison. The source projection will be defined directly on rectangular partial matchings. In particular, this construction does not posit a total injection from N N N inputs to N − = N − 1 N_- = N - 1 N − = N − 1 outputs.
Deletion maps and square harmonicity
Use the normalized deletion maps of § 4.3.1 , now with their dimensions displayed: B A , B , h d o m = B d d B^{\mathrm{dom}}_{A,B,h} = B^{\mathrm{d}}_d B A , B , h dom = B d d and B A , B , h i m = B i i B^{\mathrm{im}}_{A,B,h} = B^{\mathrm{i}}_i B A , B , h im = B i i . As before, ∂ A , B , h \partial_{A,B,h} ∂ A , B , h is their direct sum, K A , B , h K_{A,B,h} K A , B , h projects onto the joint kernel, and Z A , B , h = I − K A , B , h Z_{A,B,h} = I - K_{A,B,h} Z A , B , h = I − K A , B , h . In degree zero ∂ = 0 \partial= 0 ∂ = 0 and K = I K = I K = I . We use superscripts + + + and − - − for ( A , B ) = ( N , N ) (A,B) = (N,N) ( A , B ) = ( N , N ) and ( N , N − ) (N,N_-) ( N , N − ) , respectively, and write K ± = ⨁ h K h ± K^\pm= \bigoplus_h K^\pm_h K ± = ⨁ h K h ± and Z ± = I − K ± Z^\pm= I - K^\pm Z ± = I − K ± . A subscript ≤ D \le D ≤ D denotes the sum of the size blocks up to D D D ; L h ± L_h^\pm L h ± and L ≤ D ± L_{\le D}^\pm L ≤ D ± denote the projectors onto the entire size-h h h block and the sum of these physical size blocks, respectively.
For square spaces the harmonic subspace is the canonical image in each size block, as shown in the completion-frame construction. Lemma 4.4, applied with A = B = N A = B = N A = B = N , yields the following slightly weakened version of its gap:
( ∂ h + ) † ∂ h + ⪰ N − 2 h + 1 N − h + 1 Z h + ( 1 ≤ h ≤ N / 2 ) . (80) (\partial_h^+)^\dagger\partial_h^+ \succeq\frac{N - 2h + 1}{N - h + 1}Z_h^+
\qquad(1 \le h \le N/2). \tag*{(80)} ( ∂ h + ) † ∂ h + ⪰ N − h + 1 N − 2 h + 1 Z h + ( 1 ≤ h ≤ N /2 ) . (80)
We also need the following consequence: testing whether the output contains a given set almost preserves harmonicity. This is specific to a fixed set of slots; below we prove a separate estimate for a collision counter.
Lemma 5.7 (A fixed target and the canonical projection). Let R ⊆ [ N ] R \subseteq[N] R ⊆ [ N ] have size r ≥ 1 r \ge1 r ≥ 1 , and let P R = 1 { R ⊆ im D } P_R = 1_{\{R \subseteq\operatorname{im} D\}} P R = 1 { R ⊆ im D } on the square database space. For a ∈ ran K h + a \in\operatorname{ran} K_h^+ a ∈ ran K h + and h ≤ N / 2 h \le N/2 h ≤ N /2 ,
∥ Z h + P R a ∥ 2 ≤ r N − 2 h + 1 ∥ P R a ∥ 2 . (81) \lVert Z_h^+ P_R a\rVert^2 \le\frac{r}{N - 2h + 1}\lVert P_R a\rVert^2. \tag*{(81)} ∥ Z h + P R a ∥ 2 ≤ N − 2 h + 1 r ∥ P R a ∥ 2 . (81) Consequently, if v v v is supported on sizes at most D ≤ N / 2 D \le N/2 D ≤ N /2 , then
∥ P R K + v ∥ ≤ ∥ P R v ∥ + r N − 2 D + 1 ∥ Z + v ∥ . (82) \lVert P_R K^+v\rVert\le\lVert P_Rv\rVert+ \sqrt{\frac{r}{N - 2D + 1}}\lVert Z^+v\rVert. \tag*{(82)} ∥ P R K + v ∥ ≤ ∥ P R v ∥ + N − 2 D + 1 r ∥ Z + v ∥ . (82)
Proof. The assertion is immediate when h < r h < r h < r . For an image-deletion row ( H , y ) (H,y) ( H , y ) , whether H [ x ↦ y ] H[x \mapsto y] H [ x ↦ y ] covers R R R is independent of x x x . The image-harmonic zero sum therefore gives B h + , i m P R a = 0 B_h^{+,{\rm im}}P_Ra = 0 B h + , im P R a = 0 . For a domain-deletion row ( H , x ) (H,x) ( H , x ) , the row is zero if H H H already covers R R R , by harmonicity, and also if H H H misses at least two elements of R R R . If it misses exactly one, say y ∗ y_* y ∗ , its value is a H [ x ↦ y ∗ ] / N − h + 1 a_{H[x\mapsto y_*]}/\sqrt{N-h+1} a H [ x ↦ y ∗ ] / N − h + 1 . Each degree-h h h database covering R R R occurs in exactly r r r such rows, obtained by deleting one of its records whose image lies in R R R . Hence
∥ ∂ h + P R a ∥ 2 = r N − h + 1 ∥ P R a ∥ 2 . (83) \lVert\partial_h^+P_Ra\rVert^2 = \frac{r}{N-h+1}\lVert P_Ra\rVert^2. \tag*{(83)} ∥ ∂ h + P R a ∥ 2 = N − h + 1 r ∥ P R a ∥ 2 . (83)
Equation (5.18) proves (81) .
Orthogonal summation over the size blocks gives ∥ Z + P R a ∥ ≤ ε ∥ P R a ∥ \lVert Z^{+}P_{R}a\rVert\le\varepsilon\lVert P_{R}a\rVert ∥ Z + P R a ∥ ≤ ε ∥ P R a ∥ for canonical a a a of degree at most D D D , where ε = r / ( N − 2 D + 1 ) \varepsilon= \sqrt{r/(N-2D+1)} ε = r / ( N − 2 D + 1 ) . Write a = K + v a=K^{+}v a = K + v and b = Z + v b=Z^{+}v b = Z + v . Since b b b is orthogonal to the canonical space,
∥ P R a ∥ 2 = Re ⟨ P R a , P R v ⟩ − Re ⟨ Z + P R a , b ⟩ ≤ ∥ P R a ∥ ( ∥ P R v ∥ + ε ∥ b ∥ ) . \lVert P_{R}a\rVert^{2}=\operatorname{Re}\langle P_{R}a,P_{R}v\rangle-\operatorname{Re}\langle Z^{+}P_{R}a,b\rangle\le\lVert P_{R}a\rVert(\lVert P_{R}v\rVert+\varepsilon\lVert b\rVert). ∥ P R a ∥ 2 = Re ⟨ P R a , P R v ⟩ − Re ⟨ Z + P R a , b ⟩ ≤ ∥ P R a ∥ (∥ P R v ∥ + ε ∥ b ∥) .
Division by ∥ P R a ∥ \lVert P_{R}a\rVert ∥ P R a ∥ proves the result when this norm is nonzero; the zero case is immediate.
The rectangular projection from a square embedding
Add a distinguished slot s ∗ s_{*} s ∗ to the source range, and let i h : L N , N − , h → L N , N , h i_h:\mathcal{L}_{N,N-,h}\to\mathcal{L}_{N,N,h} i h : L N , N − , h → L N , N , h be the basis embedding whose images never use s ∗ s_{*} s ∗ . In this construction only, write P d u m = 1 { s ∗ ∈ im D } P_{\mathrm{dum}}=\mathbf{1}_{\{s_{*}\in\operatorname{im}D\}} P dum = 1 { s ∗ ∈ im D } and P 0 = I − P d u m P_{0}=I-P_{\mathrm{dum}} P 0 = I − P dum . This one-slot event is different from the full marked-bucket event P ∗ P_{*} P ∗ used in the capacity comparison.
Lemma 5.8 (Intersection and angle). For 0 ≤ h < N / 2 0\le h<N/2 0 ≤ h < N /2 ,
i h ran K h − = ran K h + ∩ ran P 0 . (84) i_h\operatorname{ran}K_h^{-}=\operatorname{ran}K_h^{+}\cap\operatorname{ran}P_{0}. \tag*{(84)} i h ran K h − = ran K h + ∩ ran P 0 . (84) If 1 ≤ h < N / 2 1\le h<N/2 1 ≤ h < N /2 and v ∈ L N , N − , h v\in\mathcal{L}_{N,N-,h} v ∈ L N , N − , h , then
∥ Z h − v ∥ ≤ ( 1 − 1 N − 2 h + 1 ) − 1 / 2 ∥ Z h + i h v ∥ . (85) \lVert Z_h^{-}v\rVert\le\left(1-\frac{1}{N-2h+1}\right)^{-1/2}\lVert Z_h^{+}i_hv\rVert. \tag*{(85)} ∥ Z h − v ∥ ≤ ( 1 − N − 2 h + 1 1 ) − 1/2 ∥ Z h + i h v ∥ . (85) At h = 0 h=0 h = 0 both invalid subspaces are zero.
Proof. Embedding a rectangular vector adds zero coefficients for all databases using s ∗ s_{*} s ∗ . Its square zero-sum equations are exactly the rectangular equations, together with equations whose coefficients are all zero. The different positive normalizing factors in the domain equations do not change their kernels. This proves (5.21) in both directions.
On ran K h + \operatorname{ran}K_h^{+} ran K h + , let B = K h + P d u m K h + B=K_h^{+}P_{\mathrm{dum}}K_h^{+} B = K h + P dum K h + and ε h 2 = 1 / ( N − 2 h + 1 ) \varepsilon_h^{2}=1/(N-2h+1) ε h 2 = 1/ ( N − 2 h + 1 ) . Lemma 5.7 gives
B − B 2 ⪯ ε h 2 B . B-B^{2}\preceq\varepsilon_h^{2}B. B − B 2 ⪯ ε h 2 B .
Every nonzero eigenvalue of B B B therefore lies in [ 1 − ε h 2 , 1 ] [1-\varepsilon_h^{2},1] [ 1 − ε h 2 , 1 ] . It follows that K h + P 0 K h + K_h^{+}P_{0}K_h^{+} K h + P 0 K h + has eigenvalue 1 1 1 precisely on the intersection in (5.21), and its remaining eigenvalues are at most ε h 2 \varepsilon_h^{2} ε h 2 . The nonzero singular-value squares of K h + P 0 K_h^{+}P_{0} K h + P 0 and its adjoint coincide. Consequently, for a vector w w w in ran P 0 \operatorname{ran}P_{0} ran P 0 orthogonal to the intersection,
∥ K h + w ∥ 2 ≤ ε h 2 ∥ w ∥ 2 , ∥ Z h + w ∥ 2 ≥ ( 1 − ε h 2 ) ∥ w ∥ 2 . \lVert K_h^{+}w\rVert^{2}\le\varepsilon_h^{2}\lVert w\rVert^{2},\qquad\lVert Z_h^{+}w\rVert^{2}\ge(1-\varepsilon_h^{2})\lVert w\rVert^{2}. ∥ K h + w ∥ 2 ≤ ε h 2 ∥ w ∥ 2 , ∥ Z h + w ∥ 2 ≥ ( 1 − ε h 2 ) ∥ w ∥ 2 .
Apply this to w = i h Z h − v w=i_hZ_h^{-}v w = i h Z h − v . The vector i h K h − v i_hK_h^{-}v i h K h − v belongs to the square canonical space, so Z h + w = Z h + i h v Z_h^{+}w=Z_h^{+}i_hv Z h + w = Z h + i h v . This proves (85) .
Corollary 5.9 (A sufficient rectangular incidence gap). For 1 ≤ h < N / 2 1\le h<N/2 1 ≤ h < N /2 ,
( ∂ h ) † ∂ h ⪰ N − 2 h N − h + 1 Z h − . (86) (\partial_h)^{\dagger}\partial_h\succeq\frac{N-2h}{N-h+1}Z_h^{-}. \tag*{(86)} ( ∂ h ) † ∂ h ⪰ N − h + 1 N − 2 h Z h − . (86)
Proof. Apply Lemma 4.4 with A = N A=N A = N and B = N − 1 B=N-1 B = N − 1 . Its coefficient is ( N − 2 h + 1 ) / ( N − h ) (N-2h+1)/(N-h) ( N − 2 h + 1 ) / ( N − h ) , which is at least the coefficient in (86) .
Outgoing and incoming leakage
The square leakage estimate was proved for a complete phase query on a permutation. Here we transfer it to the rectangular auxiliary space and record its adjoint form. The estimates hold for the inverse query, coherent control of the queried input and phase, and retained label tables.
Lemma 5.10 (Leakage of a complete query). Suppose 3 D + 5 ≤ N 3D+5\le N 3 D + 5 ≤ N and N − 2 D − 1 > 1 N-2D-1>1 N − 2 D − 1 > 1 . Define
b D + = 8 e N − D , (87) b_D^+ = \frac{8\sqrt{e}}{\sqrt{N-D}}, \tag*{(87)} b D + = N − D 8 e , (87) b D − = ( 1 − 1 N − 2 D − 1 ) − 1 / 2 ( b D + + 2 2 N − D ) . (88) b_D^- = \left(1-\frac{1}{N-2D-1}\right)^{-1/2}\left(b_D^+ + 2\sqrt{\frac{2}{N-D}}\right). \tag*{(88)} b D − = ( 1 − N − 2 D − 1 1 ) − 1/2 ( b D + + 2 N − D 2 ) . (88) Then
∥ Z ± F ± K ≤ D ± ∥ ≤ b D ± . (89) \lVert Z^{\pm}F^{\pm}K_{\leq D}^{\pm}\rVert\leq b_D^{\pm}. \tag*{(89)} ∥ Z ± F ± K ≤ D ± ∥ ≤ b D ± . (89) Taking adjoints and using one-step size bandedness also gives
∥ K ≤ D ± F ± Z ± L ≤ D ± ∥ ≤ b D + 1 ± (90) \lVert K_{\leq D}^{\pm}F^{\pm}Z^{\pm}L_{\leq D}^{\pm}\rVert\leq b_{D+1}^{\pm} \tag*{(90)} ∥ K ≤ D ± F ± Z ± L ≤ D ± ∥ ≤ b D + 1 ± (90) provided the hypotheses for the outgoing estimate also hold at D + 1 D+1 D + 1 . In particular, all these bounds are O ( N − 1 / 2 ) O(N^{-1/2}) O ( N − 1/2 ) for D ≤ N / 4 D\leq N/4 D ≤ N /4 and sufficiently large N N N .
Proof. In the square case this is Proposition 4.21 , applied to a unit-modulus phase, with the coherent controls retained. Its proof uses the canonical return estimate before any capacity transfer.
For the rectangular bound, fix a query address and a deleted database H H H , and put s = N − − ∣ H ∣ s=N_- - |H| s = N − − ∣ H ∣ . Give the added square slot any phase of modulus one, for example 1. The square and rectangular lookup phases then intertwine under the embedding. Their local normalized extension vectors satisfy
u + = a i u − + b ∣ s ∗ ⟩ , a = s s + 1 , b = 1 s + 1 . u^+ = a\,i u^- + b\lvert s_*\rangle,\qquad a=\sqrt{\frac{s}{s+1}},\qquad b=\frac{1}{\sqrt{s+1}}. u + = a i u − + b ∣ s ∗ ⟩ , a = s + 1 s , b = s + 1 1 . Let i = ⨁ h i h i=\bigoplus_h i_h i = ⨁ h i h and work in one deleted-database component. The compression operators agree under i i i on the source subspace orthogonal to the vacuum ∣ v ⟩ \lvert v\rangle ∣ v ⟩ and u − u^- u − . On their span, ( C + i − i C − ) ∣ v ⟩ = u + − i u − (C^+i-iC^-)\lvert v\rangle=u^+-iu^- ( C + i − i C − ) ∣ v ⟩ = u + − i u − and ( C + i − i C − ) u − = i u − − a u + − ( 1 − a ) ∣ v ⟩ (C^+i-iC^-)u^-=iu^- - au^+-(1-a)\lvert v\rangle ( C + i − i C − ) u − = i u − − a u + − ( 1 − a ) ∣ v ⟩ . Thus the two columns have Gram matrix
( 2 ( 1 − a ) − b 2 − b 2 2 ( 1 − a ) ) . \begin{pmatrix}
2(1-a) & -b^2 \\
-b^2 & 2(1-a)
\end{pmatrix}. ( 2 ( 1 − a ) − b 2 − b 2 2 ( 1 − a ) ) . Its largest eigenvalue is 2 ( 1 − a ) + b 2 = b 2 ( 1 + 2 / ( 1 + a ) ) 2(1-a)+b^2=b^2(1+2/(1+a)) 2 ( 1 − a ) + b 2 = b 2 ( 1 + 2/ ( 1 + a )) . Since 2 / ( 1 + a ) = 1 + ( 1 − a ) / ( 1 + a ) ≤ 1 + 1 / ( s + 1 ) 2/(1+a)=1+(1-a)/(1+a)\leq1+1/(s+1) 2/ ( 1 + a ) = 1 + ( 1 − a ) / ( 1 + a ) ≤ 1 + 1/ ( s + 1 ) , this eigenvalue is at most 2 / ( s + 1 ) + 1 / ( s + 1 ) 2 ≤ 2 / s 2/(s+1)+1/(s+1)^2\leq2/s 2/ ( s + 1 ) + 1/ ( s + 1 ) 2 ≤ 2/ s . Expanding the two compression factors, using exact intertwining of the lookup phases, therefore gives
∥ ( F + i − i F − ) L ≤ D ∥ ≤ 2 2 N − D . (91) \lVert(F^+i-iF^-)L_{\leq D}\rVert\leq2\sqrt{\frac{2}{N-D}}. \tag*{(91)} ∥( F + i − i F − ) L ≤ D ∥ ≤ 2 N − D 2 . (91) The second compression factor stays in the same deleted-database component, so the same lower bound on s s s applies to both terms.
For a ∈ ran K ≤ D − a\in\operatorname{ran}K_{\leq D}^- a ∈ ran K ≤ D − , equation (5.21) makes i a ia ia square canonical. The square bound and (91) imply
∥ Z + i F − a ∥ ≤ ( b D + + 2 2 N − D ) ∥ a ∥ . \lVert Z^+iF^-a\rVert\leq\left(b_D^+ + 2\sqrt{\frac{2}{N-D}}\right)\lVert a\rVert. ∥ Z + i F − a ∥ ≤ ( b D + + 2 N − D 2 ) ∥ a ∥ .
The rectangular output has size at most D + 1 D+1 D + 1 . Apply Lemma 5.8 at that cutoff to obtain (88) . Finally, the adjoint of an incoming block is an outgoing block for F † F^\dagger F † . A query changes size by at most one, so its canonical source can be restricted to size at most D + 1 D+1 D + 1 . This proves (90) . Coherent controls and auxiliary registers are handled by orthogonal direct sums throughout.
Collision counters under projection
Let κ \kappa κ bound the bucket capacities. We use C j C_j C j and C ^ j \widehat{C}_j C j as in Section 5.1, for 1 ≤ j ≤ κ 1 \le j \le\kappa 1 ≤ j ≤ κ . The common harmonic kernel is invariant under every permutation of input positions and every permutation of range slots. Its orthogonal projector therefore commutes with permutations within each bucket. In particular, applying K ± K^\pm K ± preserves the normalized bucket-orbit subspace of Section 3.3.
Lemma 5.11 (Counter boundary estimate). Let a a a be harmonic at size h h h , with range size B B B , and let 2 ≤ j ≤ κ 2 \le j \le\kappa 2 ≤ j ≤ κ . Then
∥ ∂ h C ^ j a ∥ 2 ≤ j ( κ − j + 1 ) B − h + 1 ⟨ a , C ^ j C ^ j − 1 a ⟩ . (92) \lVert\partial_h \widehat{C}_j a\rVert^2 \le\frac{j(\kappa-j+1)}{B-h+1}\langle a,\widehat{C}_j\widehat{C}_{j-1}a\rangle. \tag*{(92)} ∥ ∂ h C j a ∥ 2 ≤ B − h + 1 j ( κ − j + 1 ) ⟨ a , C j C j − 1 a ⟩ . (92) In the square case this implies
∥ Z h + C ^ j a ∥ 2 ≤ j ( κ − j + 1 ) N − 2 h + 1 ⟨ a , C ^ j C ^ j − 1 a ⟩ . (93) \lVert Z_h^+\widehat{C}_j a\rVert^2 \le\frac{j(\kappa-j+1)}{N-2h+1}\langle a,\widehat{C}_j\widehat{C}_{j-1}a\rangle. \tag*{(93)} ∥ Z h + C j a ∥ 2 ≤ N − 2 h + 1 j ( κ − j + 1 ) ⟨ a , C j C j − 1 a ⟩ . (93) In the rectangular source case it implies
∥ Z h − C ^ j a ∥ 2 ≤ j ( κ − j + 1 ) ( N − h + 1 ) ( N − h ) ( N − 2 h ) ⟨ a , C ^ j C ^ j − 1 a ⟩ . (94) \lVert Z_h^-\widehat{C}_j a\rVert^2 \le\frac{j(\kappa-j+1)(N-h+1)}{(N-h)(N-2h)}\langle a,\widehat{C}_j\widehat{C}_{j-1}a\rangle. \tag*{(94)} ∥ Z h − C j a ∥ 2 ≤ ( N − h ) ( N − 2 h ) j ( κ − j + 1 ) ( N − h + 1 ) ⟨ a , C j C j − 1 a ⟩ . (94) Proof. The image-deletion component vanishes because C j ( H [ x ↦ y ] ) C_j(H[x\mapsto y]) C j ( H [ x ↦ y ]) is independent of x x x . In a domain-deletion row, subtract C j ( H ) C_j(H) C j ( H ) times the harmonic row sum. Only outputs in buckets having occupancy exactly j − 1 j-1 j − 1 in H H H remain. The number of such available slots is at most ( κ − j + 1 ) C j − 1 ( H ) (\kappa-j+1)C_{j-1}(H) ( κ − j + 1 ) C j − 1 ( H ) , so Cauchy–Schwarz bounds this row’s squared norm by that number times the sum of the corresponding squared coefficients, divided by B − h + 1 B-h+1 B − h + 1 .
For a final database D D D , these deletions remove one of the j j j records in a bucket of occupancy exactly j j j . Thus D D D is counted j B = j ( D ) jB_{=j}(D) j B = j ( D ) times, where B = j ( D ) ≤ C j ( D ) B_{=j}(D)\le C_j(D) B = j ( D ) ≤ C j ( D ) ; also C j − 1 ( H ) ≤ C j − 1 ( D ) C_{j-1}(H)\le C_{j-1}(D) C j − 1 ( H ) ≤ C j − 1 ( D ) . Summing the rows proves (92) . Apply the square gap (5.18) or the rectangular gap (86) to obtain the remaining assertions.
Corollary 5.12 (One-sided counter comparison). In the rectangular source, let v v v be supported on sizes at most D < N / 2 D<N/2 D < N /2 , and put
a = K − v , b = Z − v , q j = ⟨ a , C ^ j a ⟩ , p j = ⟨ v , C ^ j v ⟩ , δ = ∥ b ∥ . a=K^-v,\qquad b=Z^-v,\qquad q_j=\langle a,\widehat{C}_j a\rangle,\qquad p_j=\langle v,\widehat{C}_jv\rangle,\qquad\delta=\lVert b\rVert. a = K − v , b = Z − v , q j = ⟨ a , C j a ⟩ , p j = ⟨ v , C j v ⟩ , δ = ∥ b ∥ . For 2 ≤ j ≤ κ 2\le j\le\kappa 2 ≤ j ≤ κ ,
q j ≤ p j + 2 δ 2 ( κ − j + 1 ) D N − 2 D q j − 1 . (95) q_j\le p_j+2\delta\sqrt{\frac{2(\kappa-j+1)D}{N-2D}}\,q_{j-1}. \tag*{(95)} q j ≤ p j + 2 δ N − 2 D 2 ( κ − j + 1 ) D q j − 1 . (95) For the square projection the same statement holds with the coefficient inside the square root replaced by ( κ − j + 1 ) D / ( N − 2 D + 1 ) (\kappa-j+1)D/(N-2D+1) ( κ − j + 1 ) D / ( N − 2 D + 1 ) .
Proof. Orthogonal summation of (94) over sizes up to D D D , and ( N − h + 1 ) / ( N − h ) ≤ 2 (N-h+1)/(N-h)\le2 ( N − h + 1 ) / ( N − h ) ≤ 2 , give
∥ Z − C ^ j a ∥ 2 ≤ 2 j ( κ − j + 1 ) N − 2 D ⟨ a , C ^ j C ^ j − 1 a ⟩ ≤ 2 ( κ − j + 1 ) D N − 2 D q j − 1 . (96) \lVert Z^-\widehat{C}_j a\rVert^2\le\frac{2j(\kappa-j+1)}{N-2D}\langle a,\widehat{C}_j\widehat{C}_{j-1}a\rangle\le\frac{2(\kappa-j+1)D}{N-2D}q_{j-1}. \tag*{(96)} ∥ Z − C j a ∥ 2 ≤ N − 2 D 2 j ( κ − j + 1 ) ⟨ a , C j C j − 1 a ⟩ ≤ N − 2 D 2 ( κ − j + 1 ) D q j − 1 . (96)
since C j ≤ D / j C_j \le D/j C j ≤ D / j . The exact expansion
q j − p j = − 2 Re ⟨ Z − C ^ j a , b ⟩ − ⟨ b , C ^ j b ⟩ q_j-p_j=-2\operatorname{Re}\langle Z-\widehat{C}_j a,b\rangle-\langle b,\widehat{C}_j b\rangle q j − p j = − 2 Re ⟨ Z − C j a , b ⟩ − ⟨ b , C j b ⟩
and positivity of C j C_j C j prove (95) . Use (93) for the square version.
Compatibility of the weighted extension
We return to the capacity-extension map J J J from (62) –(63) . It preserves database size and satisfies P ∗ J = 0 P_*J=0 P ∗ J = 0 , where P ∗ P_* P ∗ now denotes the event that the marked bucket contains all k + 1 k+1 k + 1 records. We will estimate both off-diagonal blocks of J J J relative to the two harmonic projections. The second estimate is necessary because J J J is not an isometry; the first estimate alone does not imply it.
Here are the incidence equations in the normalized bucket-orbit basis. If H H H has size h − 1 h-1 h − 1 and available bucket capacities a i a_i a i , the unnormalized domain row is
∑ i : a i > 0 a i u H [ x ↦ i ] , (97) \sum_{i:a_i>0}\sqrt{a_i}\,u_{H[x\mapsto i]}, \tag*{(97)} i : a i > 0 ∑ a i u H [ x ↦ i ] , (97)
and its contribution to the squared incidence norm is divided by the range-side normalization B − h + 1 B-h+1 B − h + 1 . The image zero sums are
∑ x ∉ dom H u H [ x ↦ i ] = 0. (98) \sum_{x\notin\operatorname{dom} H}u_{H[x\mapsto i]}=0. \tag*{(98)} x ∈ / dom H ∑ u H [ x ↦ i ] = 0. (98)
To verify the normalization, let w ( H ) = ∏ i ( κ i ) o i ( H ) w(H)=\prod_i(\kappa_i)_{o_i(H)} w ( H ) = ∏ i ( κ i ) o i ( H ) . A physical extension in bucket i i i has coefficient u H i / w ( H ) a i u_{H i}/\sqrt{w(H)a_i} u H i / w ( H ) a i . Summing its a i a_i a i available slots gives a i u H i / w ( H ) \sqrt{a_i}u_{H i}/\sqrt{w(H)} a i u H i / w ( H ) . Summing squared physical rows over the w ( H ) w(H) w ( H ) realizations of H H H cancels this last denominator. For image rows, summing also over the a i a_i a i choices of the fixed output cancels the factor a i a_i a i in the squared coefficient. This proves (5.33) and (98) , including their norm interpretation.
Lemma 5.13 (Forward compatibility). For u ∈ ran K h − u\in\operatorname{ran}K_h^- u ∈ ran K h − in the bucket-orbit subspace and h ≤ N / 2 h\le N/2 h ≤ N /2 ,
∥ Z h + J u ∥ 2 ≤ f h 2 ∥ u ∥ 2 , f h 2 : = ( k + 1 ) h 2 m ( N − 2 h + 1 ) . (99) \lVert Z_h^+Ju\rVert^2\le f_h^2\lVert u\rVert^2,\qquad f_h^2:=\frac{(k+1)h}{2m(N-2h+1)}. \tag*{(99)} ∥ Z h + J u ∥ 2 ≤ f h 2 ∥ u ∥ 2 , f h 2 := 2 m ( N − 2 h + 1 ) ( k + 1 ) h . (99) Set f 0 = 0 f_0=0 f 0 = 0 .
Proof. Fix a source deleted assignment H H H , a query input x x x , and a mark b b b . Put a i = k − o i ( H ) a_i=k-o_i(H) a i = k − o i ( H ) and r = r o b ( H ) r=r_{o_b}(H) r = r o b ( H ) . If a b > 0 a_b>0 a b > 0 , the target domain row of J b u J_bu J b u is
r ( ∑ i ≠ b a i u H i + a b + 1 a b u H b ) = r a b u H b , r\left(\sum_{i\ne b}\sqrt{a_i}\,u_{H i}+\frac{a_b+1}{\sqrt{a_b}}u_{H b}\right)=\frac{r}{\sqrt{a_b}}u_{H b}, r i = b ∑ a i u H i + a b a b + 1 u H b = a b r u H b ,
where source harmonicity was used in the equality. If a b = 0 a_b=0 a b = 0 , the row is zero: the source has no b b b -extension, and J b J_b J b has zero coefficient on the target extension that would complete the marked bucket. Target deleted assignments that are not source-feasible likewise give zero rows. All image-harmonic rows vanish because the extension weight in J b J_b J b is independent of the newly added input.
For a source database d d d with marked occupancy s = o b ( d ) ≥ 1 s=o_b(d)\ge1 s = o b ( d ) ≥ 1 , each deletion of one of its s s s marked records contributes coefficient squared
r s − 1 2 k + 1 − s = k + 1 ( k + 2 − s ) ( k + 1 − s ) ≤ k + 1 2 . \frac{r_{s-1}^2}{k+1-s}=\frac{k+1}{(k+2-s)(k+1-s)}\le\frac{k+1}{2}. k + 1 − s r s − 1 2 = ( k + 2 − s ) ( k + 1 − s ) k + 1 ≤ 2 k + 1 .
There are h h h records in all. Summing the domain-row norms over the orthogonal mark branches, with the factor 1 / m 1/m 1/ m from J J J , gives
∥ ∂ h + J u ∥ 2 ≤ ( k + 1 ) h 2 m ( N − h + 1 ) ∥ u ∥ 2 . \left\|\partial_h^+ J u\right\|^2 \le\frac{(k+1)h}{2m(N-h+1)}\|u\|^2. ∂ h + J u 2 ≤ 2 m ( N − h + 1 ) ( k + 1 ) h ∥ u ∥ 2 .
The square incidence gap proves (99) .
Lemma 5.14 (Reverse compatibility). Let v = ⨁ b v b ∈ ran K h + v=\bigoplus_b v^b\in\operatorname{ran}K_h^+ v = ⨁ b v b ∈ ran K h + belong to the controlled target bucket-orbit subspace, and suppose h < N / 2 h<N/2 h < N /2 . Let C k C_k C k count buckets with occupancy at least k k k in each branch. Then
∥ Z h − J † v ∥ 2 ≤ ( k + 1 ) 2 ( N − h + 1 ) m ( N − h ) ( N − 2 h ) ⟨ P ∗ v , C ^ k P ∗ v ⟩ . (100) \left\|Z_h^-J^\dagger v\right\|^2\le\frac{(k+1)^2(N-h+1)}{m(N-h)(N-2h)}\langle P_*v,\widehat C_kP_*v\rangle. \tag*{(100)} Z h − J † v 2 ≤ m ( N − h ) ( N − 2 h ) ( k + 1 ) 2 ( N − h + 1 ) ⟨ P ∗ v , C k P ∗ v ⟩ . (100) In particular,
∥ K h + J Z h − ∥ ≤ g h , g h 2 : = ( k + 1 ) 2 h ( N − h + 1 ) k m ( N − h ) ( N − 2 h ) . (101) \left\|K_h^+JZ_h^-\right\|\le g_h,\qquad
g_h^2:=\frac{(k+1)^2h(N-h+1)}{km(N-h)(N-2h)}. \tag*{(101)} K h + J Z h − ≤ g h , g h 2 := k m ( N − h ) ( N − 2 h ) ( k + 1 ) 2 h ( N − h + 1 ) . (101) Here the operator is between the corresponding bucket-orbit subspaces, and g 0 = 0 g_0=0 g 0 = 0 . If P ∗ v = 0 P_*v=0 P ∗ v = 0 , then J † v J^\dagger v J † v is exactly source harmonic.
Proof. Fix a source deleted assignment H H H and input x x x , and again put a i = k − o i ( H ) a_i=k-o_i(H) a i = k − o i ( H ) and r = r o b ( H ) r=r_{\mathrm{ob}}(H) r = r ob ( H ) . If a b > 0 a_b>0 a b > 0 , the source domain row of J b † v b J_b^\dagger v^b J b † v b is
r ( ∑ i ≠ b a i v H i b + a b + 1 v H b b ) = 0 , r\left(\sum_{i\ne b}\sqrt{a_i}\,v^b_{H_i}+\sqrt{a_b+1}\,v^b_{H_b}\right)=0, r i = b ∑ a i v H i b + a b + 1 v H b b = 0 ,
by target harmonicity. If a b = 0 a_b=0 a b = 0 , the source omits the last target extension. The row is consequently − k + 1 v H b b -\sqrt{k+1}\,v^b_{H_b} − k + 1 v H b b . Image harmonicity is preserved in both cases, since the weights are independent of x x x .
Unlike the forward direction, different marks can interfere after J † J^\dagger J † . The normalized source domain row is exactly
− k + 1 m ( N − h ) ∑ b : o b ( H ) = k v H [ x ↦ b ] b . -\sqrt{\frac{k+1}{m(N-h)}}\sum_{b:o_b(H)=k}v^b_{H[x\mapsto b]}. − m ( N − h ) k + 1 b : o b ( H ) = k ∑ v H [ x ↦ b ] b .
Cauchy–Schwarz bounds the squared sum by C k ( H ) ∑ b : o b ( H ) = k ∥ v H [ x ↦ b ] b ∥ 2 C_k(H)\sum_{b:o_b(H)=k}\|v^b_{H[x\mapsto b]}\|^2 C k ( H ) ∑ b : o b ( H ) = k ∥ v H [ x ↦ b ] b ∥ 2 . Each target database with a full marked bucket is counted exactly k + 1 k+1 k + 1 times, once for each deletion of a marked record. Deleting such a record leaves its number of buckets of occupancy at least k k k unchanged. Therefore
∥ ∂ h − J † v ∥ 2 ≤ ( k + 1 ) 2 m ( N − h ) ⟨ P ∗ v , C ^ k P ∗ v ⟩ . \left\|\partial_h^-J^\dagger v\right\|^2\le\frac{(k+1)^2}{m(N-h)}\langle P_*v,\widehat C_kP_*v\rangle. ∂ h − J † v 2 ≤ m ( N − h ) ( k + 1 ) 2 ⟨ P ∗ v , C k P ∗ v ⟩ .
The rectangular gap (86) proves (100) . Since C k ≤ h / k C_k\le h/k C k ≤ h / k , taking adjoints gives (101) . When P ∗ v = 0 P_*v=0 P ∗ v = 0 , both deletion components of J † v J^\dagger v J † v vanish, proving the last assertion. □ \square □
Corollary 5.15 (Endpoint leakage of the projected extension). Let S = K + J K − S=K^+JK^- S = K + J K − . On source size h < N / 2 h<N/2 h < N /2 ,
∥ P ∗ S L h − ∥ ≤ e h , e h 2 : = ( k + 1 ) 2 h 2 m ( N − 2 h + 1 ) 2 . (102) \left\|P_*SL_h^-\right\|\le e_h,\qquad
e_h^2:=\frac{(k+1)^2h}{2m(N-2h+1)^2}. \tag*{(102)} P ∗ S L h − ≤ e h , e h 2 := 2 m ( N − 2 h + 1 ) 2 ( k + 1 ) 2 h . (102) In particular, e h = O k ( h / N 3 / 2 ) e_h=O_k(\sqrt{h}/N^{3/2}) e h = O k ( h / N 3/2 ) for h ≤ N / 4 h\le N/4 h ≤ N /4 .
Proof. For u ∈ ran K h − u \in\operatorname{ran} K_h^{-} u ∈ ran K h − , apply (82) to J u Ju J u , separately in each mark branch, with R R R its k + 1 k+1 k + 1 marked slots. The term P ∗ J u P_*Ju P ∗ J u vanishes, and Lemma 5.13 bounds the remaining term. Thus
∥ P ∗ K h + J u ∥ ≤ k + 1 N − 2 h + 1 f h ∥ u ∥ = e h ∥ u ∥ . \lVert P_*K_h^{+}Ju\rVert\le\sqrt{\frac{k+1}{N-2h+1}}f_h\lVert u\rVert= e_h\lVert u\rVert. ∥ P ∗ K h + J u ∥ ≤ N − 2 h + 1 k + 1 f h ∥ u ∥ = e h ∥ u ∥ .
□
Lemma 5.16 (Defect between the filtered queries). Let A ± = K ± F ± K ± A^{\pm}=K^{\pm}F^{\pm}K^{\pm} A ± = K ± F ± K ± and S = K + J K − S=K^{+}JK^{-} S = K + J K − . Then
A + S − S A − = K + ( F + J − J F − ) K − − K + F + Z + J K − + K + J Z − F − K − . (103) A^{+}S-SA^{-}=K^{+}(F^{+}J-JF^{-})K^{-}-K^{+}F^{+}Z^{+}JK^{-}+K^{+}JZ^{-}F^{-}K^{-}. \tag*{(103)} A + S − S A − = K + ( F + J − J F − ) K − − K + F + Z + J K − + K + J Z − F − K − . (103)
For u ∈ ran K ≤ D − u\in\operatorname{ran}K_{\leq D}^{-} u ∈ ran K ≤ D − in the bucket-orbit subspace , the sum of the norms of the last two terms applied to u u u is at most
( b D + 1 + f D + g D + 1 b D − ) ∥ u ∥ . (104) \left(b_{D+1}^{+}f_D+g_{D+1}b_D^{-}\right)\lVert u\rVert. \tag*{(104)} ( b D + 1 + f D + g D + 1 b D − ) ∥ u ∥ . (104)
Consequently, for D ≤ N / 4 D\leq N/4 D ≤ N /4 and sufficiently large N N N ,
∥ ( A + S − S A − ) u ∥ ≤ ∥ ( F + J − J F − ) u ∥ + C k D + 1 N 3 / 2 ∥ u ∥ , (105) \lVert(A^{+}S-SA^{-})u\rVert\leq\lVert(F^{+}J-JF^{-})u\rVert+\frac{C_k\sqrt{D+1}}{N^{3/2}}\lVert u\rVert, \tag*{(105)} ∥( A + S − S A − ) u ∥ ≤ ∥( F + J − J F − ) u ∥ + N 3/2 C k D + 1 ∥ u ∥ , (105)
with C k C_k C k depending only on k k k .
Proof. Insert K + = I − Z + K^{+}=I-Z^{+} K + = I − Z + immediately before J J J in the first product, and K − = I − Z − K^{-}=I-Z^{-} K − = I − Z − immediately after J J J in the second. Adding and subtracting K + J F − K − K^{+}JF^{-}K^{-} K + J F − K − gives (103) exactly. The vector Z + J u Z^{+}Ju Z + J u has size at most D D D . The incoming bound (90) and Lemma 5.13 bound the second term by b D + 1 + f D ∥ u ∥ b_{D+1}^{+}f_D\lVert u\rVert b D + 1 + f D ∥ u ∥ . The vector Z − F − u Z^{-}F^{-}u Z − F − u has size at most D + 1 D+1 D + 1 . Lemma 5.14 and the rectangular outgoing bound give g D + 1 b D − ∥ u ∥ g_{D+1}b_D^{-}\lVert u\rVert g D + 1 b D − ∥ u ∥ for the third term. Size blocks are orthogonal, and J , K ± J,K^{\pm} J , K ± preserve size, so the uniform bounds up to the indicated cutoffs are the maxima of the displayed constants. These constants are increasing in their cutoffs in the range used here.
For D ≤ N / 4 D\leq N/4 D ≤ N /4 , f D , g D + 1 = O k ( D + 1 / N ) f_D,g_{D+1}=O_k(\sqrt{D+1}/N) f D , g D + 1 = O k ( D + 1 / N ) and b D ± , b D + 1 ± = O ( N − 1 / 2 ) b_D^{\pm},b_{D+1}^{\pm}=O(N^{-1/2}) b D ± , b D + 1 ± = O ( N − 1/2 ) for sufficiently large N N N . The first term in (103) is bounded by the unprojected capacity defect because K + K^{+} K + is a contraction. This proves (105) . All steps allow arbitrary auxiliary registers and controlled queries. The reverse estimate already accounts for coherent summation over the hidden mark. □
Lower bound for the exceptional type
Filtered moments and the exceptional-bucket amplitude
We now combine the local capacity estimate with the projection estimates. This requires checking the moment recurrence for the filtered source itself; a bound for an unfiltered process would not suffice.
A one-query moment estimate before projection
We use C ^ j \widehat{C}_j C j for the diagonal operator with eigenvalues C j ( d ) C_j(d) C j ( d ) defined above. The value on a deleted assignment H H H does not include a record at the queried input.
Lemma 6.1 (Local counter gain). Let v v v be a source vector of norm at most one, supported on database sizes at most D < N − D<N_{-} D < N − , in the bucket-invariant subspace, with arbitrary auxiliary registers. For a complete source phase query F − F^{-} F − and 2 ≤ j ≤ k 2\leq j\leq k 2 ≤ j ≤ k ,
∣ ⟨ F − v , C ^ j F − v ⟩ − ⟨ v , C ^ j v ⟩ ∣ ≤ 2 2 ( k − j + 1 ) ⟨ v , C ^ j − 1 v ⟩ N − − D . (106) \left|\langle F^{-}v,\widehat{C}_jF^{-}v\rangle-\langle v,\widehat{C}_jv\rangle\right|
\leq2\sqrt{2(k-j+1)}\sqrt{\frac{\langle v,\widehat{C}_{j-1}v\rangle}{N_{-}-D}}. \tag*{(106)} ⟨ F − v , C j F − v ⟩ − ⟨ v , C j v ⟩ ≤ 2 2 ( k − j + 1 ) N − − D ⟨ v , C j − 1 v ⟩ . (106)
Proof. Fix a queried input, phase character, label table, and deleted assignment H H H . Let b j ( H ) b_j(H) b j ( H ) count buckets with occupancy exactly j − 1 j-1 j − 1 . On this local space, C ^ j = C j ( H ) I + B \widehat{C}_j=C_j(H)I+B C j = C j ( H ) I + B , where B B B projects onto extensions in those buckets. For the normalized extension vector u u u ,
p H : = ∥ B u ∥ 2 = ( k − j + 1 ) b j ( H ) N − − ∣ H ∣ . p_H:=\lVert Bu\rVert^2=\frac{(k-j+1)b_j(H)}{N_- - |H|}. p H := ∥ B u ∥ 2 = N − − ∣ H ∣ ( k − j + 1 ) b j ( H ) .
Writing v 0 = ∣ H ⟩ v_0=|H\rangle v 0 = ∣ H ⟩ and w = v 0 − u w=v_0-u w = v 0 − u , the compression reflection is C x = I − ∣ w ⟩ ⟨ w ∣ C_x=I-|w\rangle\langle w| C x = I − ∣ w ⟩ ⟨ w ∣ . Direct expansion gives
[ B , C x ] = ∣ B u ⟩ ⟨ w ∣ − ∣ w ⟩ ⟨ B u ∣ , ∥ [ B , C x ] ∥ 2 = 2 p H − p H 2 ≤ 2 p H . [B,C_x]=|Bu\rangle\langle w|-|w\rangle\langle Bu|,\qquad\lVert[B,C_x]\rVert^2=2p_H-p_H^2\le2p_H. [ B , C x ] = ∣ B u ⟩ ⟨ w ∣ − ∣ w ⟩ ⟨ B u ∣ , ∥[ B , C x ] ∥ 2 = 2 p H − p H 2 ≤ 2 p H .
Indeed, the rank-two skew-adjoint operator has squared singular value ∥ w ∥ 2 ∥ B u ∥ 2 − ∣ ⟨ w , B u ⟩ ∣ 2 = 2 p H − p H 2 \lVert w\rVert^2\lVert Bu\rVert^2-|\langle w,Bu\rangle|^2=2p_H-p_H^2 ∥ w ∥ 2 ∥ B u ∥ 2 − ∣ ⟨ w , B u ⟩ ∣ 2 = 2 p H − p H 2 . The lookup commutes with B B B , so ∥ [ B , F − ] ∥ ≤ 2 2 p H \lVert[B,F^-]\rVert\le2\sqrt{2p_H} ∥[ B , F − ]∥ ≤ 2 2 p H .
Decompose v v v into the orthogonal controlled-input, character, table, and deletion components v H v_H v H , suppressing the other indices. The identity ( F − ) † C ^ j F − − C ^ j = ( F − ) † [ C ^ j , F − ] (F^-)^\dagger\widehat{C}_jF^--\widehat{C}_j=(F^-)^\dagger[\widehat{C}_j,F^-] ( F − ) † C j F − − C j = ( F − ) † [ C j , F − ] and the local norm bound show that the absolute gain is at most
2 2 ( k − j + 1 ) N − − D ‾ ∑ H b j ( H ) ∥ v H ∥ 2 . \frac{2\sqrt{2(k-j+1)}}{\sqrt{N_- - \overline{D}}}\sum_H\sqrt{b_j(H)}\lVert v_H\rVert^2. N − − D 2 2 ( k − j + 1 ) H ∑ b j ( H ) ∥ v H ∥ 2 .
Every database in the H H H -component satisfies b j ( H ) ≤ C j − 1 ( d ) b_j(H)\le C_{j-1}(d) b j ( H ) ≤ C j − 1 ( d ) . Cauchy–Schwarz and ∑ H ∥ v H ∥ 2 ≤ 1 \sum_H\lVert v_H\rVert^2\le1 ∑ H ∥ v H ∥ 2 ≤ 1 give (106) .
The filtered recurrence
Let the source start in the empty database and evolve as
ξ t + 1 − = K − F t − U t ξ t − , ξ 0 − = ∣ ∅ ⟩ , \xi^-_{t+1}=K^-F^-_t\mathcal{U}_t\xi^-_t,\qquad\xi^-_0=|\varnothing\rangle, ξ t + 1 − = K − F t − U t ξ t − , ξ 0 − = ∣ ∅ ⟩ ,
where U t \mathcal{U}_t U t is an arbitrary algorithm-only unitary; a tensor factor equal to the algorithm’s initial vector (of norm at most one) is implicit in ξ 0 − \xi^-_0 ξ 0 − . These states have norm at most one and database size at most t t t . We do not renormalize after projecting. The compression commutes with every permutation within a bucket, the lookup is constant on each bucket, and K − K^- K − preserves invariance. Starting from the empty database, the whole trajectory thus stays in the orbit space on which the transfer estimates hold. Define the unnormalized first moment
μ j ( t ) = ⟨ ξ t − , C ^ j ξ t − ⟩ . \mu_j(t)=\langle\xi^-_t,\widehat{C}_j\xi^-_t\rangle. μ j ( t ) = ⟨ ξ t − , C j ξ t − ⟩ .
Proposition 6.2 (Filtered first moments). *For sufficiently large N N N and 0 ≤ t ≤ T ≤ N / 4 0\le t\le T\le N/4 0 ≤ t ≤ T ≤ N /4 ,
μ j ( t ) ≤ C k , j t α j N β j , α j = 2 − 2 1 − j , β j = 1 − 2 1 − j , 1 ≤ j ≤ k , \mu_j(t)\le C_{k,j}\frac{t^{\alpha_j}}{N^{\beta_j}},\qquad\alpha_j=2-2^{1-j},\qquad\beta_j=1-2^{1-j},\qquad1\le j\le k, μ j ( t ) ≤ C k , j N β j t α j , α j = 2 − 2 1 − j , β j = 1 − 2 1 − j , 1 ≤ j ≤ k , where C k , j C_{k,j} C k , j depends only on k , j k,j k , j .*
Proof. The first layer satisfies μ 1 ( t ) ≤ t \mu_1(t)\le t μ 1 ( t ) ≤ t , since the process is contractive and has degree at most t t t . For j ≥ 2 j\ge2 j ≥ 2 , put v = F t − U t ξ t − v=F^-_t\mathcal{U}_t\xi^-_t v = F t − U t ξ t − and D = t + 1 D=t+1 D = t + 1 . The outgoing estimate of Lemma 5.10 gives δ : = ∥ Z − v ∥ ≤ C / N \delta:=\lVert Z^-v\rVert\le C/\sqrt{N} δ := ∥ Z − v ∥ ≤ C / N . The counter projection estimate (95) and Lemma 6.1 imply
μ j ( t + 1 ) ≤ μ j ( t ) + C k N μ j − 1 ( t ) \mu_j(t+1)\le\mu_j(t)+\frac{C_k}{\sqrt{N}}\sqrt{\mu_{j-1}(t)} μ j ( t + 1 ) ≤ μ j ( t ) + N C k μ j − 1 ( t )
+ C k t + 1 N μ j − 1 ( t + 1 ) . (107) \qquad+\frac{C_k\sqrt{t+1}}{N}\sqrt{\mu_{j-1}(t+1)}. \tag*{(107)} + N C k t + 1 μ j − 1 ( t + 1 ) . (107)
Here the projection estimate uses the already projected vector K − v = ξ t + 1 − K^{-}v=\xi_{t+1}^{-} K − v = ξ t + 1 − . Its same-time lower-layer term is handled by induction on the layer, not by an assumption that this term is bounded by the preceding-time moment.
Because t + 1 ≤ N / 4 t+1\le N/4 t + 1 ≤ N /4 and N − = N − 1 N_{-}=N-1 N − = N − 1 , increasing C k C_k C k gives
μ j ( t + 1 ) ≤ μ j ( t ) + C k N − ( μ j − 1 ( t ) + μ j − 1 ( t + 1 ) ) . \mu_j(t+1)\le\mu_j(t)+\frac{C_k}{\sqrt{N_{-}}}\left(\sqrt{\mu_{j-1}(t)}+\sqrt{\mu_{j-1}(t+1)}\right). μ j ( t + 1 ) ≤ μ j ( t ) + N − C k ( μ j − 1 ( t ) + μ j − 1 ( t + 1 ) ) .
Assume (6.2) for layer j − 1 j-1 j − 1 throughout 0 ≤ t ≤ T 0\le t\le T 0 ≤ t ≤ T . Since μ j ( 0 ) = 0 \mu_j(0)=0 μ j ( 0 ) = 0 , summation and ∑ s = 0 t − 1 ( s + 1 ) α j − 1 / 2 ≤ t 1 + α j − 1 / 2 \sum_{s=0}^{t-1}(s+1)^{\alpha_{j-1}/2}\le t^{1+\alpha_{j-1}/2} ∑ s = 0 t − 1 ( s + 1 ) α j − 1 /2 ≤ t 1 + α j − 1 /2 give the claim with C k , j = 2 C k C k , j − 1 C_{k,j}=2C_k\sqrt{C_{k,j-1}} C k , j = 2 C k C k , j − 1 . The identities α j = 1 + α j − 1 / 2 \alpha_j=1+\alpha_{j-1}/2 α j = 1 + α j − 1 /2 and β j = ( 1 + β j − 1 ) / 2 \beta_j=(1+\beta_{j-1})/2 β j = ( 1 + β j − 1 ) /2 complete the induction.
Comparison through the projected capacity map
Run the same algorithm against the filtered target process, ξ t + 1 + = K + F t + U t ξ t + \xi_{t+1}^{+}=K^{+}F_t^{+}U_t\xi_t^{+} ξ t + 1 + = K + F t + U t ξ t + , starting with the uniform hidden mark and empty database. Define
S = K + J K − , A t ± = K ± F t ± K ± . S=K^{+}JK^{-},\qquad A_t^{\pm}=K^{\pm}F_t^{\pm}K^{\pm}. S = K + J K − , A t ± = K ± F t ± K ± .
Both A t ± A_t^{\pm} A t ± are contractions, and S ξ 0 − = ξ 0 + S\xi_0^{-}=\xi_0^{+} S ξ 0 − = ξ 0 + .
Theorem 6.3 (Filtered exceptional-bucket amplitude). For sufficiently large N N N and 0 ≤ T ≤ N / 4 0\le T\le N/4 0 ≤ T ≤ N /4 .
∥ P ∗ ξ T + ∥ ≤ C k ( T N + T 2 − 2 − k N 3 / 2 − 2 − k ) . (108) \left\lVert P_{*}\xi_T^{+}\right\rVert\le C_k\left(\frac{T}{N}+\frac{T^{2-2^{-k}}}{N^{3/2-2^{-k}}}\right). \tag*{(108)} P ∗ ξ T + ≤ C k ( N T + N 3/2 − 2 − k T 2 − 2 − k ) . (108) The estimate is uniform in the alphabet and the injective label table.
Proof. Apply Lemma 5.16 and the raw comparison (78) to the canonical prequery vector U t ξ t − U_t\xi_t^{-} U t ξ t − . The unitary acts on algorithm registers alone and therefore commutes with S S S , K ± K^{\pm} K ± and the counters. Consequently,
∥ ( A t + S − S A t − ) U t ξ t − ∥ ≤ C k N 1 + μ k ( t ) + C k t + 1 N 3 / 2 . \left\lVert(A_t^{+}S-SA_t^{-})U_t\xi_t^{-}\right\rVert\le\frac{C_k}{N}\sqrt{1+\mu_k(t)}+\frac{C_k\sqrt{t+1}}{N^{3/2}}. ( A t + S − S A t − ) U t ξ t − ≤ N C k 1 + μ k ( t ) + N 3/2 C k t + 1 . The equality
ξ t + 1 + − S ξ t + 1 − = A t + U t ( ξ t + − S ξ t − ) + ( A t + S − S A t − ) U t ξ t − \xi_{t+1}^{+}-S\xi_{t+1}^{-}=A_t^{+}U_t(\xi_t^{+}-S\xi_t^{-})+(A_t^{+}S-SA_t^{-})U_t\xi_t^{-} ξ t + 1 + − S ξ t + 1 − = A t + U t ( ξ t + − S ξ t − ) + ( A t + S − S A t − ) U t ξ t − now telescopes without a multiplicative factor ∥ J ∥ \lVert J\rVert ∥ J ∥ . Using Proposition 6.2 yields
∥ ξ T + − S ξ T − ∥ ≤ C k N ∑ t < T 1 + μ k ( t ) + C k N 3 / 2 ∑ t < T t + 1 \left\lVert\xi_T^{+}-S\xi_T^{-}\right\rVert\le\frac{C_k}{N}\sum_{t<T}\sqrt{1+\mu_k(t)}+\frac{C_k}{N^{3/2}}\sum_{t<T}\sqrt{t+1} ξ T + − S ξ T − ≤ N C k t < T ∑ 1 + μ k ( t ) + N 3/2 C k t < T ∑ t + 1 ≤ C k ( T N + T 2 − 2 − k N 3 / 2 − 2 − k + ( T N ) 3 / 2 ) . \le C_k\left(\frac{T}{N}+\frac{T^{2-2^{-k}}}{N^{3/2-2^{-k}}}+\left(\frac{T}{N}\right)^{3/2}\right). ≤ C k ( N T + N 3/2 − 2 − k T 2 − 2 − k + ( N T ) 3/2 ) . Here 1 + α k / 2 = 2 − 2 − k 1+\alpha_k/2=2-2^{-k} 1 + α k /2 = 2 − 2 − k and 1 + β k / 2 = 3 / 2 − 2 − k 1+\beta_k/2=3/2-2^{-k} 1 + β k /2 = 3/2 − 2 − k . The endpoint estimate of Corollary 5.15 gives ∥ P ∗ S ξ T − ∥ = O k ( T / N 3 / 2 ) \left\lVert P_{*}S\xi_T^{-}\right\rVert=O_k(\sqrt{T}/N^{3/2}) P ∗ S ξ T − = O k ( T / N 3/2 ) . For 1 ≤ T ≤ N / 4 1\le T\le N/4 1 ≤ T ≤ N /4 this term and ( T / N ) 3 / 2 (T/N)^{3/2} ( T / N ) 3/2 are bounded by O k ( T / N ) O_k(T/N) O k ( T / N ) . At T = 0 T=0 T = 0 the target amplitude is zero. The triangle inequality proves (108) .
From database amplitude to collision probability
Write r = k + 1 r=k+1 r = k + 1 . The database projector P ∗ P_{*} P ∗ tests whether all r r r slots of the exceptional bucket have been recorded. A successful algorithm, however, reports inputs rather than entries of the database. We connect the two events by reading the reported inputs twice. The first pass stores their permutation images; the second tests whether those stored pairs are present. Both passes are part of the proof, and are appended only after the algorithm has chosen its output.
Recording and checking a list
Let L ≤ d L_{\le d} L ≤ d project onto the span of physical target databases with at most d d d entries. The mark, the input and any additional registers may be in superposition. We work on the full physical database: a read of an individual permutation image need not preserve the bucket-orbit subspace.
Lemma 6.4 (Growth in one physical read). For a controlled forward-swap compressed query Q Q Q and d < N d < N d < N ,
∥ P ∗ Q ( I − P ∗ ) L ≤ d ∥ ≤ 2 2 N − d . (109) \left\|P_*Q(I-P_*)L_{\le d}\right\| \le2\sqrt{\frac{2}{N-d}}. \tag*{(109)} ∥ P ∗ Q ( I − P ∗ ) L ≤ d ∥ ≤ 2 N − d 2 . (109) In particular, if v = L ≤ T v v=L_{\le T}v v = L ≤ T v and v ( s ) v^{(s)} v ( s ) is obtained by appending s s s such queries and register permutations used to store their responses, then, for T + s < N T+s<N T + s < N ,
∥ P ∗ v ( s ) ∥ ≤ ∥ P ∗ v ∥ + 2 s 2 N − T − s ∥ v ∥ . (110) \left\|P_*v^{(s)}\right\| \le\left\|P_*v\right\|+2s\sqrt{\frac{2}{N-T-s}}\left\|v\right\|. \tag*{(110)} P ∗ v ( s ) ≤ ∥ P ∗ v ∥ + 2 s N − T − s 2 ∥ v ∥ . (110)
Proof. Fix the mark b b b , the input x x x , and a deletion database H H H with x ∉ dom H x\notin\operatorname{dom}H x ∈ / dom H . On F x , H \mathcal{F}_{x,H} F x , H , write s H = N − ∣ H ∣ s_H=N-|H| s H = N − ∣ H ∣ . The target projector is the identity if H H H already covers B b + , b B_b^{+,b} B b + , b ; it is zero if H H H misses at least two of its slots. If H H H misses exactly one slot y y y , it is the rank-one projector Q e Q_e Q e onto e = ∣ H [ x ↦ y ] ⟩ e=|H[x\mapsto y]\rangle e = ∣ H [ x ↦ y ]⟩ . In this last case ⟨ e , C x e ⟩ = 1 − 1 / s H \langle e,C_xe\rangle=1-1/s_H ⟨ e , C x e ⟩ = 1 − 1/ s H . Since C x C_x C x is a self-adjoint unitary, writing C x C_x C x in the splitting Q e ⊕ ( I − Q e ) Q_e\oplus(I-Q_e) Q e ⊕ ( I − Q e ) gives
∥ [ Q e , C x ] ∥ = ∥ ( I − Q e ) C x e ∥ = 1 − ( 1 − 1 s H ) 2 ≤ 2 s H . \left\|[Q_e,C_x]\right\|=\left\|(I-Q_e)C_xe\right\|=\sqrt{1-\left(1-\frac{1}{s_H}\right)^2}\le\sqrt{\frac{2}{s_H}}. ∥ [ Q e , C x ] ∥ = ∥ ( I − Q e ) C x e ∥ = 1 − ( 1 − s H 1 ) 2 ≤ s H 2 .
In the other cases the commutator is zero. The slot-controlled response swap U x U_x U x commutes with P ∗ P_* P ∗ , so
[ P ∗ , C x U x C x ] = [ P ∗ , C x ] U x C x + C x U x [ P ∗ , C x ] . [P_*,C_xU_xC_x]=[P_*,C_x]U_xC_x+C_xU_x[P_*,C_x]. [ P ∗ , C x U x C x ] = [ P ∗ , C x ] U x C x + C x U x [ P ∗ , C x ] .
Every local space reached from L ≤ d L_{\le d} L ≤ d has ∣ H ∣ ≤ d |H|\le d ∣ H ∣ ≤ d . Taking the orthogonal sum over ( b , x , H ) (b,x,H) ( b , x , H ) proves (6.5); there is no factor for the number of controls. Each read increases the database size by at most one. Before read i i i it is at most T + i − 1 T+i-1 T + i − 1 . Split its input at P ∗ P_* P ∗ , use the contraction of the already accepted component and apply (6.5) to the other. The register permutations do not change the target or the norm. Summing the s s s errors proves (6.6). □
For x , y ∈ [ N ] x,y\in[N] x , y ∈ [ N ] , let Q x , y Q_{x,y} Q x , y be the diagonal projector testing D ( x ) = y D(x)=y D ( x ) = y and let M x , y = C x Q x , y C x M_{x,y}=C_xQ_{x,y}C_x M x , y = C x Q x , y C x . A physical swap with a blank response followed by selecting response y y y applies exactly M x , y M_{x,y} M x , y to the database: between the two occurrences of C x C_x C x , the swap writes y y y precisely on the range of Q x , y Q_{x,y} Q x , y .
Lemma 6.5 (Checking retained images). Suppose a state w = L ≤ d w w=L_{\le d}w w = L ≤ d w , d < N d<N d < N , retains a reported list of inputs x 1 , … , x r x_1,\ldots,x_r x 1 , … , x r and first-pass responses y 1 , … , y r y_1,\ldots,y_r y 1 , … , y r . Reject incomplete lists (any y i = ⊥ y_i=\perp y i =⊥ ), repeated inputs, and lists whose decoded labels are not all equal. Read the same inputs again into fresh response registers and accept only if the second list equals the first. The norm of the accepted component is at most
∥ P ∗ w ∥ + r 2 N − d ∥ w ∥ . (111) \left\|P_*w\right\|+r\sqrt{\frac{2}{N-d}}\left\|w\right\|. \tag*{(111)} ∥ P ∗ w ∥ + r N − d 2 ∥ w ∥ . (111)
Proof. On a deletion fiber with y ∉ H y\notin H y ∈ / H , Q x , y Q_{x,y} Q x , y and M x , y M_{x,y} M x , y are the rank-one projections onto e = ∣ H [ x ↦ y ] ⟩ e=|H[x\mapsto y]\rangle e = ∣ H [ x ↦ y ]⟩ and C x e C_xe C x e , respectively. Their overlap is 1 − 1 / s H 1-1/s_H 1 − 1/ s H . For unit vectors with overlap of modulus a a a , the difference of the two rank-one projectors has norm 1 − a 2 \sqrt{1-a^2} 1 − a 2 : on their span its trace is zero and the trace of its square is 2 ( 1 − a 2 ) 2(1-a^2) 2 ( 1 − a 2 ) . If y ∈ im H y\in\operatorname{im}H y ∈ im H , both projectors vanish. Taking the sum over the fibers gives
∥ ( M x , y − Q x , y ) L ≤ d ∥ ≤ 2 N − d . (112) \left\|(M_{x,y}-Q_{x,y})L_{\le d}\right\|\le\sqrt{\frac{2}{N-d}}. \tag*{(112)} ∥ ( M x , y − Q x , y ) L ≤ d ∥ ≤ N − d 2 . (112)
For a fixed complete first-pass list set M i = M x i , y i M_i=M_{x_i,y_i} M i = M x i , y i and Q i = Q x i , y i Q_i=Q_{x_i,y_i} Q i = Q x i , y i . The selected second-pass component is M r ⋯ M 1 w M_r\cdots M_1w M r ⋯ M 1 w on that register branch. Compare it to the product of the literal record tests:
M r ⋯ M 1 − Q r ⋯ Q 1 = ∑ i = 1 r M r ⋯ M i + 1 ( M i − Q i ) Q i − 1 ⋯ Q 1 . M_r\cdots M_1-Q_r\cdots Q_1=\sum_{i=1}^{r}M_r\cdots M_{i+1}(M_i-Q_i)Q_{i-1}\cdots Q_1. M r ⋯ M 1 − Q r ⋯ Q 1 = i = 1 ∑ r M r ⋯ M i + 1 ( M i − Q i ) Q i − 1 ⋯ Q 1 .
The right factors preserve database size and the left factors are contractions. Thus
∥ ( M r ⋯ M 1 − Q r ⋯ Q 1 ) L ≤ d ∥ ≤ r 2 N − d . (113) \left\|(M_r\cdots M_1-Q_r\cdots Q_1)_{L\leq d}\right\|\leq r\sqrt{\frac{2}{N-d}}. \tag*{(113)} ∥ ( M r ⋯ M 1 − Q r ⋯ Q 1 ) L ≤ d ∥ ≤ r N − d 2 . (113)
No commutativity of the M i M_i M i is needed. It remains to bound the literal term on each nonrejected branch. If two y i y_i y i coincide, the literal product is zero, because a partial injection cannot assign the same image to two distinct inputs. Otherwise label injectivity puts the r r r distinct slots in one bucket. All ordinary buckets have only r − 1 r-1 r − 1 slots, so these slots exhaust the exceptional bucket and the literal projector is dominated by P ∗ P_* P ∗ . The stored register branches are orthogonal; the uniform operator bound (113) therefore applies unchanged to their direct sum. This proves (111) without conditioning or renormalizing the state.
Comparison through the reads
Let ψ \psi ψ be a real state whose permutation slices have completion degree at most T T T and ∥ ψ ∥ ≤ 1 \|\psi\|\leq1 ∥ ψ ∥ ≤ 1 . Let v v v be any physical state supported on L ≤ T L_{\leq T} L ≤ T with ∥ v ∥ ≤ 1 \|v\|\leq1 ∥ v ∥ ≤ 1 . Write ε p h = ∥ v − V ~ ψ ∥ \varepsilon_{\mathrm{ph}}=\|v-\widetilde{V}\psi\| ε ph = ∥ v − V ψ ∥ . In particular the completed phase-query history satisfies
∥ ξ T + − V ~ ψ T ∥ ≤ C T / N ( 4 T ≤ N , N ≥ 8 ) . (114) \left\|\xi_T^+-\widetilde{V}\psi_T\right\|\leq CT/N\qquad(4T\leq N,\ N\geq8). \tag*{(114)} ξ T + − V ψ T ≤ C T / N ( 4 T ≤ N , N ≥ 8 ) . (114)
by (60) . A final workspace-only operation, as in Lemma 3.2, leaves this distance and the P ∗ P_* P ∗ amplitude unchanged.
For an integer d d d with 3 d + 5 ≤ N 3d+5\leq N 3 d + 5 ≤ N , put
e N ( d ) = 60 N − d + d + 1 N − 2 d − 1 . (115) e_N(d)=\frac{60}{\sqrt{N-d}}+\frac{d+1}{N-2d-1}. \tag*{(115)} e N ( d ) = N − d 60 + N − 2 d − 1 d + 1 . (115)
Proposition 4.22 gives a coherent forward-swap comparison with error at most e N ( d ) ∥ ϕ ∥ e_N(d)\|\phi\| e N ( d ) ∥ ϕ ∥ on any real state ϕ \phi ϕ of degree at most d d d , including a query controlled by previously stored outcomes. Consequently, if s s s further swaps are performed on both sides, D = T + s D=T+s D = T + s , and 3 D + 5 ≤ N 3D+5\leq N 3 D + 5 ≤ N , the resulting states obey
∥ v ( s ) − V ~ ψ ( s ) ∥ ≤ ε p h + s e N ( D ) . (116) \left\|v^{(s)}-\widetilde{V}\psi^{(s)}\right\|\leq\varepsilon_{\mathrm{ph}}+se_N(D). \tag*{(116)} v ( s ) − V ψ ( s ) ≤ ε ph + s e N ( D ) . (116)
Indeed, before the i i i th read the real degree is at most T + i − 1 ≤ D T+i-1\leq D T + i − 1 ≤ D . At that step insert the physical query applied to the encoded real state. Unitarity bounds the propagated discrepancy by its previous value; the new local error is at most e N ( D ) e_N(D) e N ( D ) since the real norm is at most one. Swapping a response into its unused storage register commutes with V ~ \widetilde{V} V . Iterating proves (116) . No full-history swap estimate has been used for the first T T T phase queries.
Proposition 6.6 (Terminal success conversion). Let p p p be the Born weight for reporting r r r distinct inputs with equal labels in ψ \psi ψ . For D = T + 2 r D=T+2r D = T + 2 r and 3 D + 5 ≤ N 3D+5\leq N 3 D + 5 ≤ N ,
p ≤ ∥ P ∗ ξ ∥ + 6 r 2 N − T − r + ∥ v − V ~ ψ ∥ + 2 r e N ( D ) . (117) \sqrt{p}\leq\|P_*\xi\|+6r\sqrt{\frac{2}{N-T-r}}+\|v-\widetilde{V}\psi\|+2re_N(D). \tag*{(117)} p ≤ ∥ P ∗ ξ ∥ + 6 r N − T − r 2 + ∥ v − V ψ ∥ + 2 r e N ( D ) . (117) In particular, for N ≥ 32 ( k + 2 ) N\geq32(k+2) N ≥ 32 ( k + 2 ) and 4 T ≤ N 4T\leq N 4 T ≤ N , the corresponding real and filtered histories satisfy
p ≤ ∥ P ∗ ξ T + ∥ + O k ( T / N + N − 1 / 2 ) . (118) \sqrt{p}\leq\|P_*\xi_T^+\|+O_k\left(T/N+N^{-1/2}\right). \tag*{(118)} p ≤ ∥ P ∗ ξ T + ∥ + O k ( T / N + N − 1/2 ) . (118)
Proof. Store both lists of r r r swap responses, keeping every slot including a possible blank. On a fixed real branch with permutation π \pi π , both passes return the same list ( π ( x i ) ) i = 1 r (\pi(x_i))_{i=1}^r ( π ( x i ) ) i = 1 r . Thus the acceptance test of Lemma 6.5 selects exactly the original collision event on that branch. More explicitly, if c c c denotes the terminal register configuration (including the mark) and u ( c , π ) u(c,\pi) u ( c , π ) its amplitude, the squared norm accepted in the real experiment is
∑ c , π ∣ u ( c , π ) ∣ 2 1 { the reported inputs are distinct and have equal labels on ( c , π ) } = p . \sum_{c,\pi} \lvert u(c,\pi)\rvert^2 \mathbf{1}_{\{\text{the reported inputs are distinct and have equal labels on }(c,\pi)\}}=p. c , π ∑ ∣ u ( c , π ) ∣ 2 1 { the reported inputs are distinct and have equal labels on ( c , π )} = p .
The final test acts on the stored registers, not on the permutation or on the physical database. It therefore commutes with the permutation-register isometry V ~ \widetilde{V} V . Applying the test to both sides of (116) shows that p \sqrt{p} p is at most the physical acceptance norm plus ∥ v − V ~ ψ ∥ + 2 r e N ( D ) \lVert v-\widetilde{V}\psi\rVert+2r e_N(D) ∥ v − V ψ ∥ + 2 r e N ( D ) .
On the physical side let w w w denote the state after the first r r r reads. It has size support at most T + r T+r T + r , and
∥ P ∗ w ∥ ≤ ∥ P ∗ v ∥ + 2 r 2 N − T − r (119) \lVert P_*w\rVert\le\lVert P_*v\rVert+2r\sqrt{\frac{2}{N-T-r}} \tag*{(119)} ∥ P ∗ w ∥ ≤ ∥ P ∗ v ∥ + 2 r N − T − r 2 (119)
by Lemma 6.4 . Applying Lemma 6.5 with d = T + r d=T+r d = T + r adds at most another r 2 / ( N − T − r ) r\sqrt{2/(N-T-r)} r 2/ ( N − T − r ) . This even gives coefficient 3 r 3r 3 r in place of 6 r 6r 6 r ; we retain the looser bound (117) to match the formal estimate. It is the verifier, not an assumed commutation with P ∗ P_* P ∗ , that permits the use of the database event.
Finally N ≥ 32 ( r + 1 ) N\ge32(r+1) N ≥ 32 ( r + 1 ) and T ≤ N / 4 T\le N/4 T ≤ N /4 imply 3 D + 5 ≤ N 3D+5\le N 3 D + 5 ≤ N , N − D ≥ N / 2 N-D\ge N/2 N − D ≥ N /2 , and N − 2 D − 1 ≥ N / 4 N-2D-1\ge N/4 N − 2 D − 1 ≥ N /4 . Hence e N ( D ) ≤ 60 2 / N + 4 ( T + 2 r + 1 ) / N e_N(D)\le60\sqrt{2/N}+4(T+2r+1)/N e N ( D ) ≤ 60 2/ N + 4 ( T + 2 r + 1 ) / N . Substitute (114) into (117) to get (118) .
The query lower bound
We retain an explicit query–success tradeoff for the exceptional type.
Theorem 6.7 (Exceptional collision search). Fix k ≥ 2 k\ge2 k ≥ 2 , let n ≥ 0 n\ge0 n ≥ 0 , and put N = k n + k + 1 N=kn+k+1 N = k n + k + 1 . For every q ≥ n + 1 q\ge n+1 q ≥ n + 1 , sample f : [ N ] → Z q f:[N]\to\mathbb{Z}_q f : [ N ] → Z q uniformly from exact type ( k n , k + 1 ) (k^n,k+1) ( k n , k + 1 ) . If a quantum algorithm uses at most T T T function queries and outputs the unique ( k + 1 ) (k+1) ( k + 1 ) -collision with probability p p p , then, for sufficiently large N N N and 0 ≤ T ≤ N / 4 0\le T\le N/4 0 ≤ T ≤ N /4 ,
p ≤ C k ( N − 1 / 2 + T N + T 2 − 2 − k N 3 / 2 − 2 − k ) (120) \sqrt{p}\le C_k\left(N^{-1/2}+\frac{T}{N}+\frac{T^{2-2^{-k}}}{N^{3/2-2^{-k}}}\right) \tag*{(120)} p ≤ C k ( N − 1/2 + N T + N 3/2 − 2 − k T 2 − 2 − k ) (120) In particular, for every fixed p 0 ∈ ( 0 , 1 ) p_0\in(0,1) p 0 ∈ ( 0 , 1 ) , success probability at least p 0 p_0 p 0 requires
T = Ω k , p 0 ( N γ k ) , γ k = 3 / 2 − 2 − k 2 − 2 − k = 3 4 − 1 4 ( 2 k + 1 − 1 ) . (121) T=\Omega_{k,p_0}(N^{\gamma_k}),\qquad\gamma_k=\frac{3/2-2^{-k}}{2-2^{-k}}=\frac{3}{4}-\frac{1}{4(2^{k+1}-1)}. \tag*{(121)} T = Ω k , p 0 ( N γ k ) , γ k = 2 − 2 − k 3/2 − 2 − k = 4 3 − 4 ( 2 k + 1 − 1 ) 1 . (121) The constants are independent of q q q .
Proof of Theorem 6.7. The exceptional-bucket amplitude bound (108) and the terminal conversion (118) give, for each fixed injective label table,
p ≤ C k ( N − 1 / 2 + T N + T 2 − 2 − k N 3 / 2 − 2 − k ) . \sqrt{p}\le C_k\left(N^{-1/2}+\frac{T}{N}+\frac{T^{2-2^{-k}}}{N^{3/2-2^{-k}}}\right). p ≤ C k ( N − 1/2 + N T + N 3/2 − 2 − k T 2 − 2 − k ) .
The bound is uniform in the table. Squaring first and averaging the conditional probabilities gives the same amplitude bound for the overall experiment. The exact-type sampling and the query-model identification in Lemma 3.2 give (120) for the stated function problem.
Write a = 2 − 2 − k a = 2 - 2^{-k} a = 2 − 2 − k and b = 3 / 2 − 2 − k b = 3/2 - 2^{-k} b = 3/2 − 2 − k , so that γ k = b / a < 1 \gamma_k = b/a < 1 γ k = b / a < 1 . At T ≤ c N γ k T \le cN^{\gamma_k} T ≤ c N γ k the bound reads
p ≤ C k ( N − 1 / 2 + c N γ k − 1 + c a ) . \sqrt{p} \le C_k(N^{-1/2} + cN^{\gamma_k-1} + c^a). p ≤ C k ( N − 1/2 + c N γ k − 1 + c a ) .
Choose c = c ( k , p 0 ) > 0 c = c(k,p_0) > 0 c = c ( k , p 0 ) > 0 so that C k c a < p 0 / 2 C_kc^a < \sqrt{p_0}/2 C k c a < p 0 /2 . The first two terms tend to zero uniformly over the indicated query counts, and c N γ k ≤ N / 4 cN^{\gamma_k} \le N/4 c N γ k ≤ N /4 for large N N N . Thus success probability p ≥ p 0 p \ge p_0 p ≥ p 0 requires (121) . Any algorithm succeeding with probability at least p 0 p_0 p 0 on every input of the type also succeeds with that probability on average under its uniform distribution. This proves the worst-case lower bound as well.
Lower bounds for arbitrary bounded types
We now allow arbitrary bounded class sizes and any positive number of eligible classes. In this section K K K is the fixed maximum class size, r r r is the requested collision order, and M M M denotes the number of eligible classes. In particular, M M M is not a database range size. The lower bound depends on the intermediate tail counts B 2 , … , B r B_2,\ldots,B_r B 2 , … , B r . Fix integers K ≥ r ≥ 2 K \ge r \ge2 K ≥ r ≥ 2 . For an exact type
λ = ( 1 n 1 , … , K n K ) , N = ∑ d = 1 K d n d , \lambda= (1^{n_1},\ldots,K^{n_K}), \qquad N = \sum_{d=1}^{K} d n_d, λ = ( 1 n 1 , … , K n K ) , N = d = 1 ∑ K d n d ,
define its tail counts and number of eligible classes by
B j = ∑ d = j K n d ( 1 ≤ j ≤ K ) , M = B r ≥ 1. (122) B_j = \sum_{d=j}^{K} n_d \quad(1 \le j \le K), \qquad M = B_r \ge1. \tag*{(122)} B j = d = j ∑ K n d ( 1 ≤ j ≤ K ) , M = B r ≥ 1. (122)
For 2 ≤ j ≤ r 2 \le j \le r 2 ≤ j ≤ r put
F j , M = ( N 2 j − 1 ∏ i = 2 j − 1 B i 2 i − 2 M 2 j − 2 ) 1 / ( 2 j − 1 ) , Q = max 2 ≤ j ≤ r F j , M . (123) F_{j,M} = \left(\frac{N^{2j-1}\prod_{i=2}^{j-1}B_i^{2i-2}}{M^{2j-2}}\right)^{1/(2j-1)}, \qquad Q = \max_{2\le j\le r} F_{j,M}. \tag*{(123)} F j , M = ( M 2 j − 2 N 2 j − 1 ∏ i = 2 j − 1 B i 2 i − 2 ) 1/ ( 2 j − 1 ) , Q = 2 ≤ j ≤ r max F j , M . (123)
Empty products are one. A class is an equality class of inputs; a bucket is a set of hidden permutation slots representing that class. The algorithm receives only the function oracle O f ∣ x , z ⟩ = ∣ x , z + f ( x ) ⟩ O_f\lvert x,z\rangle= \lvert x,z+f(x)\rangle O f ∣ x , z ⟩ = ∣ x , z + f ( x )⟩ , with values in Z q \mathbb{Z}_q Z q .
Theorem 7.1 (A lower bound for every bounded exact type). Fix p 0 ∈ ( 0 , 1 ) p_0 \in(0,1) p 0 ∈ ( 0 , 1 ) . Uniformly over the types in (122) , all q ≥ B 1 q \ge B_1 q ≥ B 1 , and all sufficiently large N N N , finding one r r r -collision with success probability at least p 0 p_0 p 0 on a uniformly random function of type λ \lambda λ requires
Ω K , r , p 0 ( Q ) \Omega_{K,r,p_0}(Q) Ω K , r , p 0 ( Q ) function queries. The type may be given to the algorithm. Consequently the same lower bound holds for worst-case bounded-error search over that exact type.
The theorem permits any number M M M of eligible classes while requesting one collision. The matching search construction is given in Theorem 8.1 .
Corollary 7.2 (Tight complexity with a unique largest class). For fixed r r r , suppose λ = ( 1 n 1 , … , ( r − 1 ) n r − 1 , r ) \lambda= (1^{n_1},\ldots,(r-1)^{n_{r-1}},r) λ = ( 1 n 1 , … , ( r − 1 ) n r − 1 , r ) , so that B j = 1 + ∑ d = j r − 1 n d B_j = 1 + \sum_{d=j}^{r-1} n_d B j = 1 + ∑ d = j r − 1 n d and M = 1 M = 1 M = 1 . For every q ≥ B 1 q \ge B_1 q ≥ B 1 , the worst-case quantum query complexity of finding the unique r r r -collision is
Θ r ( max 2 ≤ j ≤ r [ N 2 j − 1 ∏ i = 2 j − 1 B i 2 i − 2 ] 1 / ( 2 j − 1 ) ) . (124) \Theta_r\left(\max_{2\le j\le r}\left[N^{2j-1}\prod_{i=2}^{j-1}B_i^{2i-2}\right]^{1/(2j-1)}\right). \tag*{(124)} Θ r 2 ≤ j ≤ r max [ N 2 j − 1 i = 2 ∏ j − 1 B i 2 i − 2 ] 1/ ( 2 j − 1 ) . (124)
The lower bound holds on the uniform distribution of the type. No condition requires an intermediate size to occur even once.
Proof. Substitute M = 1 M=1 M = 1 into Theorem 7.1 and combine it with the upper bound of Theorem 8.1 . □
For example, for λ = ( 1 n 1 , 2 n 2 , 3 n 3 , 4 ) \lambda=(1^{n_1},2^{n_2},3^{n_3},4) λ = ( 1 n 1 , 2 n 2 , 3 n 3 , 4 ) ,
Q = max { N 2 / 3 , N 4 / 7 B 2 1 / 7 , N 8 / 15 B 2 1 / 15 B 3 2 / 15 } . (125) Q=\max\{N^{2/3},\ N^{4/7}B_2^{1/7},\ N^{8/15}B_2^{1/15}B_3^{2/15}\}. \tag*{(125)} Q = max { N 2/3 , N 4/7 B 2 1/7 , N 8/15 B 2 1/15 B 3 2/15 } . (125)
Here B 2 = n 2 + n 3 + 1 B_2=n_2+n_3+1 B 2 = n 2 + n 3 + 1 and B 3 = n 3 + 1 B_3=n_3+1 B 3 = n 3 + 1 . If n 2 = 0 n_2=0 n 2 = 0 , the second and third scales still exist: partial records of three-element classes supply the intermediate progress events.
The proof uses the complete-query return and terminal read estimates proved above. We first extend the auxiliary database estimates to a bounded deficit in the number of slots, then compare hidden assignments of capacity multisets and propagate all errors through the intermediate levels.
Database conventions and bounded-deficit estimates
The database basis consists of partial injections D : [ N ] ⇀ [ R ] D:[N]\rightharpoonup[R] D : [ N ] ⇀ [ R ] . A range of size R = N − e R=N-e R = N − e , where 0 ≤ e = O K ( 1 ) 0\le e=O_K(1) 0 ≤ e = O K ( 1 ) , is partitioned into buckets of capacities at most K K K . Auxiliary rectangular databases do not represent a distribution of total injections into fewer than N N N slots. A bucket assignment d d d records the bucket, rather than the precise slot, assigned to each recorded input. The normalized orbit vector ∣ d ˉ ⟩ |\bar d\rangle ∣ d ˉ ⟩ is uniform over the physical injections realizing it. If the capacities are κ i \kappa_i κ i , its orbit size is ∏ i ( κ i ) o i ( d ) \prod_i(\kappa_i)_{o_i(d)} ∏ i ( κ i ) o i ( d ) , where o i ( d ) = ∣ d − 1 ( i ) ∣ o_i(d)=|d^{-1}(i)| o i ( d ) = ∣ d − 1 ( i ) ∣ and ( a ) u (a)_u ( a ) u is the falling factorial.
Let K R K_R K R project onto the simultaneous kernel of domain and image deletion, degree by degree, and let Z R = I − K R Z_R=I-K_R Z R = I − K R . Explicitly, in physical degree h h h , their normalized deletion rows are
( ∂ d v ) H , x = ∑ y ∉ im H v H [ x ↦ y ] R − h + 1 , ( ∂ i v ) H , y = ∑ x ∉ dom H v H [ x ↦ y ] N − h + 1 . (\partial_d v)_{H,x}=\frac{\sum_{y\notin\operatorname{im}H}v_{H[x\mapsto y]}}{\sqrt{R-h+1}},\qquad
(\partial_i v)_{H,y}=\frac{\sum_{x\notin\operatorname{dom}H}v_{H[x\mapsto y]}}{\sqrt{N-h+1}}. ( ∂ d v ) H , x = R − h + 1 ∑ y ∈ / im H v H [ x ↦ y ] , ( ∂ i v ) H , y = N − h + 1 ∑ x ∈ / dom H v H [ x ↦ y ] .
Here ∣ H ∣ = h − 1 |H|=h-1 ∣ H ∣ = h − 1 . In degree zero K R = I K_R=I K R = I .
For a fixed query input x x x , delete its record to obtain a database H H H . Write v H = ∣ H ⟩ v_H=|H\rangle v H = ∣ H ⟩ and let u H u_H u H be the uniform normalized superposition of its extensions at x x x . The decompression operator is
C x = I − ∣ v H − u H ⟩ ⟨ v H − u H ∣ C_x=I-|v_H-u_H\rangle\langle v_H-u_H| C x = I − ∣ v H − u H ⟩ ⟨ v H − u H ∣
on this deletion component. It exchanges v H v_H v H and u H u_H u H . A complete phase-query block is F = C x L C x F=C_xLC_x F = C x L C x , where lookup L L L applies the phase ω q z a ( i ) \omega_q^{za(i)} ω q z a ( i ) to a record in bucket i i i and the identity to an unrecorded query input, where ω q = e 2 π i / q \omega_q=e^{2\pi i/q} ω q = e 2 π i / q . The injection a : [ B 1 ] ↪ Z q a:[B_1]\hookrightarrow\mathbb{Z}_q a : [ B 1 ] ↪ Z q gives distinct bucket labels. The input and phase character may be controlled coherently. A filtered update is K R F K R K_RFK_R K R F K R , interleaved with arbitrary algorithm-only unitaries. These updates are contractions, and their state at time t t t has degree at most t t t . No projected state is renormalized.
All capacity assignments to the label indices are hidden and uniformly random among assignments of the specified capacity multiset. They are retained in reference registers. For a fixed label table, even allowing algorithm-only unitaries to depend on that table and the capacity multiset leaves the following estimates valid. They cannot depend on the hidden capacity assignment or a newly introduced mark. For fixed a a a , a uniform capacity assignment followed by a uniform slot permutation yields a uniform exact-type function with image a ( [ B 1 ] ) a([B_1]) a ([ B 1 ]) : every function has exactly ∏ d ( d ! ) n d \prod_d(d!)^{n_d} ∏ d ( d ! ) n d slot realizations for its uniquely determined assignment. Averaging over a uniform injection a a a gives the uniform type on Z q \mathbb{Z}_q Z q .
For a filtered state ξ t = ∑ d ∣ ξ t , d ⟩ ∣ d ˉ ⟩ \xi_t=\sum_d|\xi_{t,d}\rangle|\bar d\rangle ξ t = ∑ d ∣ ξ t , d ⟩ ∣ d ˉ ⟩ , including its hidden capacity branches, define
C j ( d ) = ∣ { i : o i ( d ) ≥ j } ∣ , μ j ( t ) = ∑ d C j ( d ) ∥ ξ t , d ∥ 2 . C_j(d)=|\{i:o_i(d)\ge j\}|,\qquad\mu_j(t)=\sum_d C_j(d)\|\xi_{t,d}\|^2. C j ( d ) = ∣ { i : o i ( d ) ≥ j } ∣ , μ j ( t ) = d ∑ C j ( d ) ∥ ξ t , d ∥ 2 .
The coefficients are unnormalized algorithm-register vectors. A superscript will restrict the counter to a specified capacity class. The basic degree bound gives μ 1 ( t ) ≤ t \mu_1(t)\le t μ 1 ( t ) ≤ t .
We need the incoming and outgoing estimates for every bounded range deficit, rather than only the one-slot deficit used for the exceptional family. The following argument makes this extension explicit.
Lemma 7.3 (Bounded-deficit leakage). Fix a constant e 0 e_0 e 0 . For N − e 0 ≤ R ≤ N N-e_0\le R\le N N − e 0 ≤ R ≤ N , sufficiently large N N N , and t ≤ N / 4 t\le N/4 t ≤ N /4 , complete phase queries satisfy
∥ Z R F K R ∥ ≤ t + ∥ K R F Z R ∥ ≤ t ≤ C e 0 N − 1 / 2 . (126) \|Z_R F K_R\|_{\le t}+\|K_R F Z_R\|_{\le t}\le C_{e_0}N^{-1/2}. \tag*{(126)} ∥ Z R F K R ∥ ≤ t + ∥ K R F Z R ∥ ≤ t ≤ C e 0 N − 1/2 . (126) The norm subscript restricts the physical input degree. The estimate is uniform in the phases, coherent query controls, and auxiliary registers.
Proof. Lemma 4.4 applies directly to ([N]\mathbin{\rightharpoonup}[R]$. Its coefficient is ( R − 2 h + 2 ) / ( R − h + 1 ) (R-2h+2)/(R-h+1) ( R − 2 h + 2 ) / ( R − h + 1 ) in degree h h h , hence is bounded below by a positive constant in the stated range.
Let i R i_R i R be the basis embedding from range R R R into range R + 1 R+1 R + 1 , avoiding the added slot. The zero-sum equations give
i R ran K R , h = ran K R + 1 , h ∩ ran P 0 . i_R\operatorname{ran}K_{R,h}=\operatorname{ran}K_{R+1,h}\cap\operatorname{ran}P_0. i R ran K R , h = ran K R + 1 , h ∩ ran P 0 .
where P 0 P_0 P 0 tests avoidance of that slot. To quantify its angle, let P = I − P 0 P=I-P_0 P = I − P 0 and let v v v be harmonic in the larger space. The image deletion rows of P v Pv P v vanish. A domain row is nonzero only if its extension uses the added slot, so each contributing database occurs once. Therefore
∥ ∂ P v ∥ 2 = ∥ P v ∥ 2 R − h + 2 , ∥ Z R + 1 , h P v ∥ 2 ≤ ∥ P v ∥ 2 R − 2 h + 3 . \|\partial Pv\|^2=\frac{\|Pv\|^2}{R-h+2},\qquad\|Z_{R+1,h}Pv\|^2\le\frac{\|Pv\|^2}{R-2h+3}. ∥ ∂ P v ∥ 2 = R − h + 2 ∥ P v ∥ 2 , ∥ Z R + 1 , h P v ∥ 2 ≤ R − 2 h + 3 ∥ P v ∥ 2 .
Set B = K R + 1 , h P K R + 1 , h B=K_{R+1,h}PK_{R+1,h} B = K R + 1 , h P K R + 1 , h on the harmonic space and ϵ 2 = ( R − 2 h + 3 ) − 1 \epsilon^2=(R-2h+3)^{-1} ϵ 2 = ( R − 2 h + 3 ) − 1 . The preceding inequality is B − B 2 ⪯ ϵ 2 B B-B^2\preceq\epsilon^2B B − B 2 ⪯ ϵ 2 B . Thus the nonzero eigenvalues of B B B are at least 1 − ϵ 2 1-\epsilon^2 1 − ϵ 2 , and the eigenvalues of K R + 1 , h P 0 K R + 1 , h K_{R+1,h}P_0K_{R+1,h} K R + 1 , h P 0 K R + 1 , h off the displayed intersection are at most ϵ 2 \epsilon^2 ϵ 2 . Comparing the singular values of K R + 1 , h P 0 K_{R+1,h}P_0 K R + 1 , h P 0 and its adjoint gives
∥ Z R , h w ∥ ≤ ( 1 − ϵ 2 ) − 1 / 2 ∥ Z R + 1 , h i R w ∥ . \|Z_{R,h}w\|\le(1-\epsilon^2)^{-1/2}\|Z_{R+1,h}i_Rw\|. ∥ Z R , h w ∥ ≤ ( 1 − ϵ 2 ) − 1/2 ∥ Z R + 1 , h i R w ∥.
At degree zero both invalid components vanish. Iterating over the at most e 0 e_0 e 0 omitted slots gives a uniformly bounded angle factor.
Embed the range-R R R database into the square space by the resulting map i i i , and assign arbitrary unit phases to the added slots. Lookup intertwines exactly. In a deletion component, put s = R − ∣ H ∣ s=R-|H| s = R − ∣ H ∣ and e = N − R e=N-R e = N − R . Its uniform vectors satisfy
u N = a i u R + b u n e w , a = s s + e , b = e s + e . u_N=aiu_R+bu_{\mathrm{new}},\qquad a=\sqrt{\frac{s}{s+e}},\quad b=\sqrt{\frac{e}{s+e}}. u N = ai u R + b u new , a = s + e s , b = s + e e .
The compression difference vanishes off the source span of the blank coordinate and u R u_R u R . On those two coordinates its Gram matrix is
( 2 ( 1 − a ) − b 2 − b 2 2 ( 1 − a ) ) . \begin{pmatrix}
2(1-a) & -b^2\\
-b^2 & 2(1-a)
\end{pmatrix}. ( 2 ( 1 − a ) − b 2 − b 2 2 ( 1 − a ) ) .
whose largest eigenvalue is at most 3 b 2 3b^2 3 b 2 . Expanding the two compression factors of a query gives
∥ ( F N i − i F R ) L ≤ t ∥ ≤ 2 3 e R − t . \|(F_N\mathbf{i}-\mathbf{i}F_R)_{L\le t}\| \le2\sqrt{\frac{3e}{R-t}}. ∥ ( F N i − i F R ) L ≤ t ∥ ≤ 2 R − t 3 e .
The embedded rectangular harmonic vector is square harmonic. Apply Proposition 4.21 , the last display, and the bounded angle at output degree t + 1 t+1 t + 1 to obtain the outgoing estimate. Apply the same estimate to the inverse phase at cutoff t + 1 t+1 t + 1 and take adjoints to obtain the incoming estimate; a complete query changes degree by at most one. For sufficiently large N N N the square cutoff conditions hold throughout. Orthogonal sums handle all query controls.
We next record the moment inequality used both in dense layers and for auxiliary capacity multisets. It also holds when the counters are restricted to any specified set of buckets.
Lemma 7.4 (Ordinary filtered moment gain). For every 2 ≤ j ≤ K 2 \le j \le K 2 ≤ j ≤ K , bounded deficit e 0 e_0 e 0 , and t ≤ N / 8 t \le N/8 t ≤ N /8 ,
μ j ( t + 1 ) − μ j ( t ) ≤ C K , e 0 N ( μ j − 1 ( t ) + μ j − 1 ( t + 1 ) ) . (127) \mu_j(t+1)-\mu_j(t)\le\frac{C_{K,e_0}}{\sqrt{N}}\left(\sqrt{\mu_{j-1}(t)}+\sqrt{\mu_{j-1}(t+1)}\right). \tag*{(127)} μ j ( t + 1 ) − μ j ( t ) ≤ N C K , e 0 ( μ j − 1 ( t ) + μ j − 1 ( t + 1 ) ) . (127)
Proof. On a deletion component, write C ^ j = C j ( H ) I + B \widehat{C}_j=C_j(H)I+B C j = C j ( H ) I + B , where B B B tests an extension into a counted bucket of occupancy j − 1 j-1 j − 1 . There are at most ( K − j + 1 ) C j − 1 ( H ) (K-j+1)C_{j-1}(H) ( K − j + 1 ) C j − 1 ( H ) available slots of this kind. For the uniform extension vector u u u , put a H = ∥ B u ∥ 2 a_H=\|Bu\|^2 a H = ∥ B u ∥ 2 . The rank-two calculation of Lemma 6.1 gives
∥ [ B , C x ] ∥ 2 = 2 a H − a H 2 , a H ≤ ( K − j + 1 ) C j − 1 ( H ) R − ∣ H ∣ . \|[B,C_x]\|^2=2a_H-a_H^2,\qquad a_H\le\frac{(K-j+1)C_{j-1}(H)}{R-|H|}. ∥ [ B , C x ] ∥ 2 = 2 a H − a H 2 , a H ≤ R − ∣ H ∣ ( K − j + 1 ) C j − 1 ( H ) .
Lookup commutes with B B B . Cauchy–Schwarz over the orthogonal deletion components therefore bounds the unprojected moment gain by C K μ j − 1 ( t ) / ( R − t ) C_K\sqrt{\mu_{j-1}(t)/(R-t)} C K μ j − 1 ( t ) / ( R − t ) .
For completeness, the projection correction has the same form for every range R R R considered here. For a harmonic degree-h h h vector a a a , the domain-row calculation in Lemma 5.11 and the rectangular deletion gap give
∥ Z R C ^ j a ∥ 2 ≤ j ( K − j + 1 ) R − 2 h + 2 ⟨ a , C ^ j C ^ j − 1 a ⟩ . \|Z_R\widehat{C}_j a\|^2\le\frac{j(K-j+1)}{R-2h+2}\langle a,\widehat{C}_j\widehat{C}_{j-1}a\rangle. ∥ Z R C j a ∥ 2 ≤ R − 2 h + 2 j ( K − j + 1 ) ⟨ a , C j C j − 1 a ⟩ .
This calculation is unchanged for a restricted bucket set: only its boundary slots enter the row sum. If v = F ξ t v=F\xi_t v = F ξ t , a = K R v a=K_Rv a = K R v , z = Z R v z=Z_Rv z = Z R v , and D = t + 1 D=t+1 D = t + 1 , positivity and the identity
⟨ a , C ^ j a ⟩ − ⟨ v , C ^ j v ⟩ = − 2 Re ⟨ Z R C ^ j a , z ⟩ − ⟨ z , C ^ j z ⟩ \langle a,\widehat{C}_j a\rangle-\langle v,\widehat{C}_jv\rangle=-2\operatorname{Re}\langle Z_R\widehat{C}_ja,z\rangle-\langle z,\widehat{C}_jz\rangle ⟨ a , C j a ⟩ − ⟨ v , C j v ⟩ = − 2 Re ⟨ Z R C j a , z ⟩ − ⟨ z , C j z ⟩
bound the correction by
2 ∥ z ∥ ( K − j + 1 ) D R − 2 D + 2 μ j − 1 ( t + 1 ) . 2\|z\|\sqrt{\frac{(K-j+1)D}{R-2D+2}\mu_{j-1}(t+1)}. 2∥ z ∥ R − 2 D + 2 ( K − j + 1 ) D μ j − 1 ( t + 1 ) .
Here C j ≤ D / j C_j\le D/j C j ≤ D / j was used. Lemma 7.3 gives ∥ z ∥ ≤ C e 0 / N \|z\|\le C_{e_0}/\sqrt{N} ∥ z ∥ ≤ C e 0 / N , so this term is at most C K , e 0 t + 1 μ j − 1 ( t + 1 ) / N C_{K,e_0}\sqrt{t+1}\sqrt{\mu_{j-1}(t+1)}/N C K , e 0 t + 1 μ j − 1 ( t + 1 ) / N . Since t + 1 = O ( N ) t+1=O(N) t + 1 = O ( N ) and R − t = Θ ( N ) R-t=\Theta(N) R − t = Θ ( N ) , it is absorbed into (7.6). ∎
A hidden bounded-capacity jump
Consider two capacity multisets related by increasing one capacity-c c c bucket to d d d , with $1 \le c < d \le K$. Write m m m for the number of capacity-c c c buckets in the source and p p p for the number of capacity-d d d buckets in the target. A source capacity assignment and one of its m m m candidates correspond bijectively to a target assignment and one of its p p p marked d d d -buckets, so
∣ M − ∣ m = ∣ M + ∣ p . (128) |\mathcal{M}_{-}|m = |\mathcal{M}_{+}|p. \tag*{(128)} ∣ M − ∣ m = ∣ M + ∣ p . (128)
In particular, after forgetting the mark the target has the uniform distribution over assignments of its capacity multiset, and the mark is independent of its execution conditional on that capacity assignment. More explicitly, identify each physical branch with its normalized bucket-orbit coordinates. Slot relabelings intertwine compression, lookup and the harmonic projection, so the target query matrices depend only on the target capacities and label table, not on which eligible bucket was designated as the mark.
Write K − K_{-} K − and K + K_{+} K + for the harmonic projections of the source and target ranges. Their filtered query blocks are A ± = K ± F ± K ± A_{\pm}=K_{\pm}F_{\pm}K_{\pm} A ± = K ± F ± K ± . For each mark b b b , define the capacity map and its normalized direct sum by
J b ∣ d 0 ⟩ − = ( d ) o b ( d 0 ) ( c ) o b ( d 0 ) ∣ d 0 ⟩ + , b , J = 1 m ⨁ b J b , S = K + J K − . J_b|d_0\rangle_{-}=\sqrt{\frac{(d)_{o_b(d_0)}}{(c)_{o_b(d_0)}}}|d_0\rangle_{+,b},\qquad J=\frac{1}{\sqrt{m}}\bigoplus_b J_b,\qquad S=K_{+}JK_{-}. J b ∣ d 0 ⟩ − = ( c ) o b ( d 0 ) ( d ) o b ( d 0 ) ∣ d 0 ⟩ + , b , J = m 1 b ⨁ J b , S = K + J K − .
The assignment symbol d 0 d_0 d 0 here is unrelated to the capacity d d d . All weights are bounded by a constant depending only on K K K . On the complement of the source orbit space J b J_b J b is zero. The harmonic projections preserve the orbit spaces. The map S S S sends the initial uniform source empty state to the initial uniformly marked target empty state, by (128) .
Lemma 7.5 (Localized bounded jump). Suppose 0 ≤ t ≤ N / 8 0 \le t \le N/8 0 ≤ t ≤ N /8 . Let μ c ( t ) \mu_c^{(t)} μ c ( t ) be the source moment of full capacity-c c c candidate buckets. On the source prequery state, the one-query filtered comparison error has the form K + v t + r t K_{+}v_t+r_t K + v t + r t with
Ξ t v t = v t , ∥ v t ∥ ≤ C K μ c ( t ) m N , ∥ r t ∥ ≤ C K m N . \Xi_t v_t=v_t,\qquad\lVert v_t\rVert\le C_K\sqrt{\frac{\mu_c^{(t)}}{mN}},\qquad\lVert r_t\rVert\le\frac{C_K}{\sqrt{mN}}. Ξ t v t = v t , ∥ v t ∥ ≤ C K m N μ c ( t ) , ∥ r t ∥ ≤ m N C K . Here Ξ t \Xi_t Ξ t requires the queried input to be recorded in the marked target bucket at occupancy c + 1 c+1 c + 1 . For the broader event P P P of marked occupancy at least c + 1 c+1 c + 1 , the coarse endpoint satisfies
∥ P S ξ − t ∥ ≤ C K t m N . \lVert P S \xi_{-}^{t}\rVert\le C_K\sqrt{\frac{t}{mN}}. ∥ P S ξ − t ∥ ≤ C K m N t . Proof. We calculate the defect on one deletion component, then sum over its independent query and hidden-parameter controls. In a deletion component let a i a_i a i be the remaining source capacities, s = ∑ i a i s=\sum_i a_i s = ∑ i a i , and δ = d − c \delta=d-c δ = d − c . Set ρ = ( s + δ ) / s \rho=\sqrt{(s+\delta)/s} ρ = ( s + δ ) / s and α = 1 − ρ − 1 \alpha=1-\rho^{-1} α = 1 − ρ − 1 . Write v v v for the blank coordinate, ∣ i ⟩ |i\rangle ∣ i ⟩ for a normalized bucket extension, and define r u = ( d ) u / ( c ) u r_u=\sqrt{(d)_u/(c)_u} r u = ( d ) u / ( c ) u for 0 ≤ u ≤ c 0\le u\le c 0 ≤ u ≤ c , with r ~ b = r o b ( H ) \widetilde r_b=r_{o_b(H)} r b = r o b ( H ) . Let u − u_{-} u − and u + u_{+} u + be the normalized uniform extension vectors of the source and target in this component. The exact rank-two identity is
C + J b − J b C − = ( ∣ v ⟩ − ∣ u + ⟩ ) ⟨ p b ∣ + ∣ q b ⟩ ( ⟨ u − ∣ − ⟨ v ∣ ) . C_{+}J_b-J_bC_{-}=(|v\rangle-|u_{+}\rangle)\langle p_b|+|q_b\rangle(\langle u_{-}|-\langle v|). C + J b − J b C − = ( ∣ v ⟩ − ∣ u + ⟩) ⟨ p b ∣ + ∣ q b ⟩ (⟨ u − ∣ − ⟨ v ∣ ) . For a b > 0 a_b>0 a b > 0 its vectors are
p b = − r ~ b α u − + r ~ b δ a b ( s + δ ) ∣ b ⟩ , q b = r ~ b ( ρ − 1 ) u + . p_b=-\widetilde r_b\alpha u_{-}+\frac{\widetilde r_b\delta}{\sqrt{a_b(s+\delta)}}|b\rangle,\qquad q_b=\widetilde r_b(\rho-1)u_{+}. p b = − r b α u − + a b ( s + δ ) r b δ ∣ b ⟩ , q b = r b ( ρ − 1 ) u + .
For a b = 0 a_b=0 a b = 0 they are
p b = − r ~ b α u − , q b = r ~ b ( α u − − δ s + δ ∣ b ⟩ ) . p_b=-\widetilde r_b\alpha u_-,\qquad q_b=\widetilde r_b\left(\alpha u_--\sqrt{\frac{\delta}{s+\delta}}|b\rangle\right). p b = − r b α u − , q b = r b ( α u − − s + δ δ ∣ b ⟩ ) .
Here α , ρ − 1 = O K ( N − 1 ) \alpha,\rho-1=O_K(N^{-1}) α , ρ − 1 = O K ( N − 1 ) . Separate the full-candidate term
G b C = − 1 { a b = 0 } r c δ s + δ ∣ b ⟩ ( ⟨ u − ∣ − ⟨ v ∣ ) . G_b^C=-\mathbf{1}_{\{a_b=0\}}r_c\sqrt{\frac{\delta}{s+\delta}}|b\rangle\left(\langle u_-|-\langle v|\right). G b C = − 1 { a b = 0 } r c s + δ δ ∣ b ⟩ ( ⟨ u − ∣ − ⟨ v ∣ ) .
Its range satisfies Ξ t \Xi_t Ξ t . Its squared direct-sum norm is bounded by the source full-candidate count divided by m N mN m N . After removing this term, the remaining averaged squared defect is O K ( N − 2 + ( m N ) − 1 ) O_K(N^{-2}+(mN)^{-1}) O K ( N − 2 + ( m N ) − 1 ) times the local source squared norm: the potentially larger nonfull term has bra ⟨ b ∣ \langle b| ⟨ b ∣ , whose squared source coordinates sum to at most the source squared norm.
Since L + J b = J b L − L_+J_b=J_bL_- L + J b = J b L − , the complete-query leading term can be chosen as
G b F = L + G b C + G b C L − C − . G_b^F=L_+G_b^C+G_b^C L_-C_-. G b F = L + G b C + G b C L − C − .
It still satisfies Ξ t G b F = G b F \Xi_tG_b^F=G_b^F Ξ t G b F = G b F . The extra term ( C + − I ) L + G b C (C_+-I)L_+G_b^C ( C + − I ) L + G b C has norm O K ( N − 1 / 2 ) ∥ G b C ∥ O_K(N^{-1/2})\lVert G_b^C\rVert O K ( N − 1/2 ) ∥ G b C ∥ . Orthogonal summation over the query controls, deletion components and marks gives the asserted raw bounds. We also require both harmonic-compatibility estimates, including the coherent reverse sum over marks:
∥ Z + J K − ∥ ≤ t + ∥ K + J Z − ∥ ≤ t ≤ C K t m N . (129) \lVert Z_+JK_-\rVert_{\leq t}+\lVert K_+JZ_-\rVert_{\leq t}\leq C_K\sqrt{\frac{t}{mN}}. \tag*{(129)} ∥ Z + J K − ∥ ≤ t + ∥ K + J Z − ∥ ≤ t ≤ C K m N t . (129)
Let R − R_- R − and R + = R − + δ R_+=R_-+\delta R + = R − + δ be the two range sizes. On an orbit assignment H H H of degree h − 1 h-1 h − 1 , the unnormalized domain row is ∑ i : a i > 0 a i u H i \sum_{i:a_i>0}\sqrt{a_i}\,u_{H_i} ∑ i : a i > 0 a i u H i ; its squared norm is divided by R ± − h + 1 R_\pm-h+1 R ± − h + 1 in the corresponding normalized row map. For source-harmonic u u u , the target row of J b u J_bu J b u is, when a b > 0 a_b>0 a b > 0 ,
r ~ b ( ∑ i ≠ b a i u H i + a b + δ a b u H b ) = δ r ~ b a b u H b . \widetilde r_b\left(\sum_{i\neq b}\sqrt{a_i}\,u_{H_i}+\frac{a_b+\delta}{\sqrt{a_b}}u_{H_b}\right)=\frac{\delta\widetilde r_b}{\sqrt{a_b}}u_{H_b}. r b i = b ∑ a i u H i + a b a b + δ u H b = a b δ r b u H b .
When a b = 0 a_b=0 a b = 0 it is zero, since the new target extension has zero source coefficient. All image rows vanish: the capacity weight is constant over the inputs summed by each such row. The displayed coefficient is bounded in terms of K K K ; summing over marks and deletions counts each source database at most h h h times. Hence
∥ ∂ + J u ∥ 2 ≤ C K h m ( R + − h + 1 ) ∥ u ∥ 2 . \lVert\partial_+Ju\rVert^2\leq\frac{C_Kh}{m(R_+-h+1)}\lVert u\rVert^2. ∥ ∂ + J u ∥ 2 ≤ m ( R + − h + 1 ) C K h ∥ u ∥ 2 .
The deletion gap proves the forward bound.
For the reverse bound, let w = ⨁ b w b w=\bigoplus_b w^b w = ⨁ b w b be target harmonic. When a b > 0 a_b>0 a b > 0 , the source domain row of J b † w b J_b^\dagger w^b J b † w b is
r ~ b ( ∑ i ≠ b a i w H i b + a b + δ w H b b ) = 0. \widetilde r_b\left(\sum_{i\neq b}\sqrt{a_i}\,w^b_{H_i}+\sqrt{a_b+\delta}\,w^b_{H_b}\right)=0. r b i = b ∑ a i w H i b + a b + δ w H b b = 0.
When a b = 0 a_b=0 a b = 0 , it equals − r c δ w H b b -r_c\sqrt{\delta}\,w^b_{H_b} − r c δ w H b b . After the direct-sum adjoint, the normalized row is therefore
− r c δ m ( R − − h + 1 ) ∑ b : ob ( H ) = c w H b b . -r_c\sqrt{\frac{\delta}{m(R_--h+1)}}\sum_{\substack{b:\operatorname{ob}(H)=c}}w^b_{H_b}. − r c m ( R − − h + 1 ) δ b : ob ( H ) = c ∑ w H b b .
There are at most h / c h/c h / c summands. Cauchy–Schwarz, followed by summing over rows, counts each contributing target database at most c + 1 c+1 c + 1 times, once for each deletion from marked occupancy c + 1 c+1 c + 1 . Rows with larger marked occupancy are not source-feasible, and image rows again vanish. Thus
∥ ∂ − J † w ∥ 2 ≤ C K h m ( R − − h + 1 ) ∥ w ∥ 2 . \lVert\partial_- J^\dagger w\rVert^2 \le\frac{C_K h}{m(R_- - h + 1)}\lVert w\rVert^2. ∥ ∂ − J † w ∥ 2 ≤ m ( R − − h + 1 ) C K h ∥ w ∥ 2 .
The deletion gap and adjoints prove the second bound in (7.8). With hidden capacity assignments, each marked target branch has one parent; the reverse summation is made separately in each parent block. Distinct parents and degrees are orthogonal, which proves the same bounds with all reference registers. Insert them and (7.5) into the exact identity
A + S − S A − = K + ( F + J − J F − ) K − − K + F + Z + J K − + K + J Z − F − K − . A_+S-SA_- = K_+(F_+J-JF_-)K_- - K_+F_+Z_+JK_- + K_+JZ_-F_-K_-. A + S − S A − = K + ( F + J − J F − ) K − − K + F + Z + J K − + K + J Z − F − K − .
The filtering remainder is O K ( t + 1 / ( N m ) ) O_K(\sqrt{t+1}/(N\sqrt{m})) O K ( t + 1 / ( N m )) . Together with the raw bounds, μ c − ( t ) ≤ m \mu_c^-(t) \le m μ c − ( t ) ≤ m , and t ≤ N / 8 t \le N/8 t ≤ N /8 , this is at most C K / m N C_K/\sqrt{mN} C K / m N . Finally P J = 0 PJ=0 P J = 0 , so the forward compatibility bound implies ∥ P S ξ − ∥ ≤ ∥ Z + J K − ξ − ∥ \lVert PS\xi_-\rVert\le\lVert Z_+JK_-\xi_-\rVert ∥ P S ξ − ∥ ≤ ∥ Z + J K − ξ − ∥ as claimed. □ \square □
Lemma 7.6 (Squared transfer estimate). Suppose p ≤ m p \le m p ≤ m and 0 ≤ T ≤ N / 8 0 \le T \le N/8 0 ≤ T ≤ N /8 . The target moment counting capacity-d d d buckets with occupancy at least c + 1 c+1 c + 1 satisfies
μ c + 1 ( d ) ( T ) ≤ C K p m N ∑ t < T μ c − ( t ) + C K p T 2 m N . (130) \mu_{c+1}^{(d)}(T) \le C_K\sqrt{\frac{p}{mN}}\sum_{t<T}\sqrt{\mu_c^-(t)}+C_K\frac{pT^2}{mN}. \tag*{(130)} μ c + 1 ( d ) ( T ) ≤ C K m N p t < T ∑ μ c − ( t ) + C K m N p T 2 . (130) The states may have arbitrary algorithm registers and hidden source capacity assignments.
Proof. Put e t = ξ t + − S ξ t − e_t=\xi_t^+-S\xi_{t}^- e t = ξ t + − S ξ t − and E t = ∥ e t ∥ E_t=\lVert e_t\rVert E t = ∥ e t ∥ . Then e 0 = 0 e_0=0 e 0 = 0 and, including the prequery algorithm unitary U t U_t U t ,
e t + 1 = a t + K + v t + r t , a t = A + , t U t e t , ∥ a t ∥ ≤ E t . e_{t+1}=a_t+K_+v_t+r_t,\qquad a_t=A_{+,t}U_te_t,\qquad\lVert a_t\rVert\le E_t. e t + 1 = a t + K + v t + r t , a t = A + , t U t e t , ∥ a t ∥ ≤ E t .
The mark is uniform among p p p target d d d -buckets and has no effect on their unmarked execution. For a fixed queried input and database, at most one such mark satisfies Ξ t \Xi_t Ξ t . Therefore ∥ Ξ t ξ t + 1 + ∥ ≤ p − 1 / 2 \lVert\Xi_t\xi_{t+1}^+\rVert\le p^{-1/2} ∥ Ξ t ξ t + 1 + ∥ ≤ p − 1/2 . Since a t a_t a t is harmonic and v t v_t v t is localized,
⟨ a t , K + v t ⟩ = ⟨ Ξ t a t , v t ⟩ , ∥ Ξ t a t ∥ ≤ p − 1 / 2 + η t + 1 + g t + ρ t , \langle a_t,K_+v_t\rangle=\langle\Xi_ta_t,v_t\rangle,\qquad\lVert\Xi_ta_t\rVert\le p^{-1/2}+\eta_{t+1}+g_t+\rho_t, ⟨ a t , K + v t ⟩ = ⟨ Ξ t a t , v t ⟩ , ∥ Ξ t a t ∥ ≤ p − 1/2 + η t + 1 + g t + ρ t ,
where g t = ∥ v t ∥ g_t=\lVert v_t\rVert g t = ∥ v t ∥ , ρ t = ∥ r t ∥ \rho_t=\lVert r_t\rVert ρ t = ∥ r t ∥ and η u = C K u / ( m N ) \eta_u=C_K\sqrt{u/(mN)} η u = C K u / ( m N ) . We used a t = ξ t + 1 + − S ξ t + 1 − − K + v t − r t a_t=\xi_{t+1}^+-S\xi_{t+1}^--K_+v_t-r_t a t = ξ t + 1 + − S ξ t + 1 − − K + v t − r t and Ξ t ⪯ P \Xi_t\preceq P Ξ t ⪯ P . Consequently
E t + 1 2 ≤ E t 2 + 2 g t p + 2 g t η t + 1 + 3 g t 2 + 4 g t ρ t + 2 E t ρ t + ρ t 2 . E_{t+1}^2\le E_t^2+\frac{2g_t}{\sqrt{p}}+2g_t\eta_{t+1}+3g_t^2+4g_t\rho_t+2E_t\rho_t+\rho_t^2. E t + 1 2 ≤ E t 2 + p 2 g t + 2 g t η t + 1 + 3 g t 2 + 4 g t ρ t + 2 E t ρ t + ρ t 2 .
Write A c − ( t ) = μ c − ( t ) / m ≤ 1 A_c^-(t)=\sqrt{\mu_c^-(t)/m}\le1 A c − ( t ) = μ c − ( t ) / m ≤ 1 , so g t ≤ C K A c − ( t ) / N g_t\le C_KA_c^-(t)/\sqrt{N} g t ≤ C K A c − ( t ) / N . The terms g t 2 , g t η t + 1 , g t ρ t g_t^2,g_t\eta_{t+1},g_t\rho_t g t 2 , g t η t + 1 , g t ρ t are each at most C K g t / p C_Kg_t/\sqrt{p} C K g t / p , using p ≤ m p\le m p ≤ m , p ≤ N p\le N p ≤ N and t ≤ N / 8 t\le N/8 t ≤ N /8 . For a nonnegative recurrence E t + 1 2 ≤ E t 2 + f t + 2 E t ρ t + ρ t 2 E_{t+1}^2\le E_t^2+f_t+2E_t\rho_t+\rho_t^2 E t + 1 2 ≤ E t 2 + f t + 2 E t ρ t + ρ t 2 , put E ∗ = max t ≤ T E t E_*=\max_{t\le T}E_t E ∗ = max t ≤ T E t , F T = ∑ t < T f t F_T=\sum_{t<T}f_t F T = ∑ t < T f t , and R T = ∑ t < T ρ t R_T=\sum_{t<T}\rho_t R T = ∑ t < T ρ t . Summing up to an index attaining E ∗ E_* E ∗ gives E ∗ 2 ≤ F T + 2 E ∗ R T + R T 2 E_*^2\le F_T+2E_*R_T+R_T^2 E ∗ 2 ≤ F T + 2 E ∗ R T + R T 2 . Hence
E ∗ ≤ R T + F T + 2 R T 2 ≤ F T + ( 1 + 2 ) R T , E ∗ 2 ≤ 2 F T + 12 R T 2 . E_*\le R_T+\sqrt{F_T+2R_T^2}\le\sqrt{F_T}+(1+\sqrt{2})R_T,\qquad E_*^2\le2F_T+12R_T^2. E ∗ ≤ R T + F T + 2 R T 2 ≤ F T + ( 1 + 2 ) R T , E ∗ 2 ≤ 2 F T + 12 R T 2 .
Thus
E T 2 ≤ C K [ 1 N p ∑ t < T A c − ( t ) + T 2 m N ] . E_T^2\le C_K\left[\frac{1}{\sqrt{Np}}\sum_{t<T}A_c^-(t)+\frac{T^2}{mN}\right]. E T 2 ≤ C K [ N p 1 t < T ∑ A c − ( t ) + m N T 2 ] .
Uniform mark averaging gives μ c + 1 ( d ) ( T ) / p = ∥ P ξ T + ∥ 2 \mu_{c+1}^{(d)}(T)/p=\lVert P\xi_T^+\rVert^2 μ c + 1 ( d ) ( T ) / p = ∥ P ξ T + ∥ 2 . Use ∥ P ξ T + ∥ ≤ E T + η T \lVert P\xi_T^+\rVert\le E_T+\eta_T ∥ P ξ T + ∥ ≤ E T + η T , square, and multiply by p p p . For integer T ≥ 1 T\ge1 T ≥ 1 the endpoint term is absorbed in p T 2 / ( m N ) pT^2/(mN) p T 2 / ( m N ) ; the case T = 0 T=0 T = 0 is immediate. □ \square □
A uniform recursion over capacity multisets
We now derive one recursion that handles both abundant and missing intermediate sizes. Its source and target capacity multisets can differ, so the induction is carried out simultaneously over the required capacity multisets.
Proposition 7.7 (Uniform moment recursion over capacity multisets). For each j ≥ 2 j \ge2 j ≥ 2 with B j > 0 B_j > 0 B j > 0 and 0 ≤ T ≤ N / 8 0 \le T \le N/8 0 ≤ T ≤ N /8 , an upper bound for the target moment is
μ j ( T ) ≤ C K B j N B j − 1 ∑ t < T ( V j − 1 ( t ) + V j − 1 ( t + 1 ) ) + C K B j T 2 N B j − 1 . (131) \mu_j(T) \le C_K \sqrt{\frac{B_j}{N B_{j-1}}} \sum_{t<T} \left(\sqrt{V_{j-1}(t)}+\sqrt{V_{j-1}(t+1)}\right) + C_K \frac{B_jT^2}{N B_{j-1}}. \tag*{(131)} μ j ( T ) ≤ C K N B j − 1 B j t < T ∑ ( V j − 1 ( t ) + V j − 1 ( t + 1 ) ) + C K N B j − 1 B j T 2 . (131) Here V j − 1 V_{j-1} V j − 1 is any uniform upper bound on the ( j − 1 ) (j-1) ( j − 1 ) st moment for the target capacity multiset and the auxiliary capacity multisets obtained by reducing one capacity d ≥ j d \ge j d ≥ j to j − 1 j-1 j − 1 . Every such auxiliary capacity multiset has exactly the same B i B_i B i for i ≤ j − 1 i \le j-1 i ≤ j − 1 .
Proof. If B j ≥ B j − 1 / 2 B_j \ge B_{j-1}/2 B j ≥ B j − 1 /2 , the ordinary filtered recurrence (127) implies the result, with no need for the additive term. Otherwise n j − 1 = B j − 1 − B j > B j − 1 / 2 n_{j-1}=B_{j-1}-B_j>B_{j-1}/2 n j − 1 = B j − 1 − B j > B j − 1 /2 . For each d ≥ j d \ge j d ≥ j with n d > 0 n_d>0 n d > 0 , apply Lemma 7.6 to the source obtained by reducing one d d d -bucket to j − 1 j-1 j − 1 . Its candidate pool is m = n j − 1 + 1 ≍ B j − 1 m=n_{j-1}+1\asymp B_{j-1} m = n j − 1 + 1 ≍ B j − 1 and the target marking pool is p = n d ≤ B j < m p=n_d\le B_j<m p = n d ≤ B j < m . The source full-candidate count is at most its total level-( j − 1 ) (j-1) ( j − 1 ) count. Summing (7.9) over d ≥ j d \ge j d ≥ j , and using ∑ d ≥ j n d ≤ K B j \sum_{d\ge j}\sqrt{n_d}\le\sqrt{K B_j} ∑ d ≥ j n d ≤ K B j , proves the bound. Reducing d d d to j − 1 j-1 j − 1 leaves every smaller capacity-tail count unchanged, proving the last assertion. □
Recursive descent has at most K K K levels and removes at most K K K slots at a step. All auxiliary ranges therefore have deficit at most K 2 K^2 K 2 from the starting square range. This is why the bounded-deficit estimates apply throughout. One may formulate the induction over all ancestors reachable by these descents; it does not assert that a fixed range-deficit class is closed under arbitrarily many reductions.
The additive errors can be propagated explicitly without first restricting T T T to a proposed lower-bound scale. Put α j = 2 − 2 1 − j \alpha_j=2-2^{1-j} α j = 2 − 2 1 − j .
Lemma 7.8 (Moments with all propagated errors). For j ≥ 2 j\ge2 j ≥ 2 with B j > 0 B_j>0 B j > 0 , define
L j ( T ) = T α j N ∏ i = 2 j B i 2 i − j − 1 , (132) L_j(T)=\frac{T^{\alpha_j}}{N\prod_{i=2}^{j}B_i^{2^{i-j-1}}}, \tag*{(132)} L j ( T ) = N ∏ i = 2 j B i 2 i − j − 1 T α j , (132) E l , j ( T ) = T 2 N B l − 1 2 l − j ∏ i = l + 1 j B i 2 i − j − 1 ( 2 ≤ l ≤ j ) . (133) E_{l,j}(T)=\frac{T^2}{N B_{l-1}^{2^{l-j}}\prod_{i=l+1}^{j}B_i^{2^{i-j-1}}}
\qquad(2\le l\le j). \tag*{(133)} E l , j ( T ) = N B l − 1 2 l − j ∏ i = l + 1 j B i 2 i − j − 1 T 2 ( 2 ≤ l ≤ j ) . (133) Then, uniformly over the required capacity multisets and T ≤ N / 8 T\le N/8 T ≤ N /8 ,
μ j ( T ) B j ≤ C K ( L j ( T ) + ∑ l = 2 j E l , j ( T ) ) . (134) \frac{\mu_j(T)}{B_j}\le C_K\left(L_j(T)+\sum_{l=2}^{j}E_{l,j}(T)\right). \tag*{(134)} B j μ j ( T ) ≤ C K ( L j ( T ) + l = 2 ∑ j E l , j ( T ) ) . (134) For level one use μ 1 ( T ) / B 1 ≤ C K T / N \mu_1(T)/B_1\le C_KT/N μ 1 ( T ) / B 1 ≤ C K T / N .
Proof. The range is N − O K ( 1 ) N-O_K(1) N − O K ( 1 ) and capacities are bounded, so B 1 = Θ K ( N ) B_1=\Theta_K(N) B 1 = Θ K ( N ) . Set A j ( T ) = μ j ( T ) / B j A_j(T)=\sqrt{\mu_j(T)/B_j} A j ( T ) = μ j ( T ) / B j . Dividing (131) by B j B_j B j gives
A j ( T ) 2 ≤ C K [ 1 N B j ∑ t < T ( A j − 1 s r c ( t ) + A j − 1 s r c ( t + 1 ) ) + T 2 N B j − 1 ] . A_j(T)^2\le C_K\left[\frac{1}{\sqrt{N B_j}}\sum_{t<T}\left(A^{\mathrm{src}}_{j-1}(t)+A^{\mathrm{src}}_{j-1}(t+1)\right)+\frac{T^2}{N B_{j-1}}\right]. A j ( T ) 2 ≤ C K [ N B j 1 t < T ∑ ( A j − 1 src ( t ) + A j − 1 src ( t + 1 ) ) + N B j − 1 T 2 ] .
where any source is covered by the uniform induction hypothesis and has unchanged lower tails. Starting with C K T / N C_KT/N C K T / N , its leading term is L j L_j L j , because α j = 1 + α j − 1 / 2 \alpha_j = 1 + \alpha_{j-1}/2 α j = 1 + α j − 1 /2 . An error born at level l l l is initially T 2 / ( N B l − 1 ) T^2/(NB_{l-1}) T 2 / ( N B l − 1 ) . Taking its square root and summing over time at the next level preserves the power T 2 T^2 T 2 , halves all existing bucket exponents, and adds exponent 1 / 2 1/2 1/2 on the new B j B_j B j . This gives exactly (133) . The inequality u 1 + ⋯ + u a ≤ ∑ i u i \sqrt{u_1}+\cdots+\sqrt{u_a}\le\sum_i\sqrt{u_i} u 1 + ⋯ + u a ≤ ∑ i u i and ∑ t < T ( t + 1 ) u ≤ T u + 1 \sum_{t<T}(t+1)^u\le T^{u+1} ∑ t < T ( t + 1 ) u ≤ T u + 1 for u ≥ 0 u\ge0 u ≥ 0 justify each induction step, including its same-time term. There are at most K K K terms and levels, so the constant depends only on K K K .
The denominator exponents in each E l , j E_{l,j} E l , j sum to one. This fact will control their contribution after any remaining capacity transfer. It also gives the coarse uniform bound μ j ( T ) ≤ C K ( B j L j ( T ) + T 2 / N ) \mu_j(T)\le C_K(B_jL_j(T)+T^2/N) μ j ( T ) ≤ C K ( B j L j ( T ) + T 2 / N ) , but keeping the individual errors is essential beyond the square-root query scale.
Converting a union of database events to success
Let P r P_r P r project onto databases having at least one bucket with r r r recorded slots. This event is the union over all eligible buckets; it has no distinguished mark.
Lemma 7.9 (Union-event success conversion). Let p s u c c p_{\mathrm{succ}} p succ be the real success probability of a T T T -query function algorithm and ξ T \xi_T ξ T its square filtered history. For fixed K , r K,r K , r , sufficiently large N N N , and T ≤ N / 8 T\le N/8 T ≤ N /8 ,
p s u c c ≤ ∥ P r ξ T ∥ + C K , r ( T N + T + 1 N ) . (135) \sqrt{p_{\mathrm{succ}}}\le\lVert P_r\xi_T\rVert+C_{K,r}\left(\frac{T}{N}+\sqrt{\frac{T+1}{N}}\right). \tag*{(135)} p succ ≤ ∥ P r ξ T ∥ + C K , r ( N T + N T + 1 ) . (135) The comparison is uniform over the injective label table and the hidden capacity assignments. There is no factor M \sqrt{M} M .
Proof. Corollary 4.19 bounds the distance between the filtered history and the isometrically encoded real history by C T / N CT/N C T / N . Its complete-phase estimate is uniform over all slot phases, so it applies to every capacity multiset and injective label table here. Keep the reported r r r inputs and use two passes of r r r forward-swap reads as in Lemma 6.5 . We give the boundary estimate for the larger event P r P_r P r explicitly.
On a physical deletion component with residual database H H H , if H H H already contains an r r r -collision then P r P_r P r is the identity and its commutator with decompression is zero. Otherwise the event tests extensions into buckets of occupancy r − 1 r-1 r − 1 . There are at most ∣ H ∣ / ( r − 1 ) |H|/(r-1) ∣ H ∣/ ( r − 1 ) such buckets, each with at most K − r + 1 K-r+1 K − r + 1 available slots. Writing u u u for the uniform extension vector and a = ∥ P r u ∥ 2 a=\lVert P_ru\rVert^2 a = ∥ P r u ∥ 2 ,
a ≤ ( K − r + 1 ) ∣ H ∣ ( r − 1 ) ( N − ∣ H ∣ ) , ∥ [ P r , C x ] ∥ = a ( 2 − a ) . a\le\frac{(K-r+1)|H|}{(r-1)(N-|H|)},\qquad\lVert[P_r,C_x]\rVert=\sqrt{a(2-a)}. a ≤ ( r − 1 ) ( N − ∣ H ∣ ) ( K − r + 1 ) ∣ H ∣ , ∥[ P r , C x ]∥ = a ( 2 − a ) .
The compressed swap is decompression, a lookup diagonal in the physical database, and recompression. Its boundary norm is therefore at most 2 ∥ [ P r , C x ] ∥ = O K ( ( T + 1 ) / N ) 2\lVert[P_r,C_x]\rVert=O_K(\sqrt{(T+1)/N}) 2 ∥[ P r , C x ]∥ = O K ( ( T + 1 ) / N ) on the relevant degree cutoff. Orthogonal summation over query controls and deletion components preserves this bound.
After the constant number of reads, the literal verifier accepts only if the reported distinct inputs are recorded at distinct slots in one bucket. Injective bucket labels make its event a subevent of P r P_r P r . The product comparison (113) , with P r P_r P r dominating its literal term, contributes O r ( N − 1 / 2 ) O_r(N^{-1/2}) O r ( N − 1/2 ) . Proposition 4.22 and the hybrid sum (116) through the constant number of reads contribute O r ( N − 1 / 2 + ( T + 1 ) / N ) O_r(N^{-1/2}+(T+1)/N) O r ( N − 1/2 + ( T + 1 ) / N ) . Add the boundary changes and the initial O ( T / N ) O(T/N) O ( T / N ) history error to obtain (135) . Every estimate holds in direct sums over the hidden parameters and stored output registers; none conditions on or renormalizes a mark. □
Selecting the relevant intermediate level
The maximum in (123) specifies the intermediate order that must remain hard. We now show that this order can always be analyzed using either the full moment or a capacity layer that is actually present in the input.
Proof of Theorem 7.1 . Choose a maximizing index p p p , so F p , M = Q F_{p,M}=Q F p , M = Q , and put D = M Q 4 / N 2 D=MQ^4/N^2 D = M Q 4 / N 2 . If p ≥ 3 p\ge3 p ≥ 3 , comparison with F p − 1 , M F_{p-1,M} F p − 1 , M gives B p − 1 ≥ D B_{p-1}\ge D B p − 1 ≥ D . For p = 2 p=2 p = 2 , the same statement holds up to a K K K -dependent constant because B 1 = Θ K ( N ) B_1=\Theta_K(N) B 1 = Θ K ( N ) and D = Q = N 2 / 3 M − 1 / 3 D=Q=N^{2/3}M^{-1/3} D = Q = N 2/3 M − 1/3 . If p < r p<r p < r , comparison with F p + 1 , M F_{p+1,M} F p + 1 , M gives B p ≤ D B_p\le D B p ≤ D .
Suppose first that M ≥ B p − 1 / ( 4 K ) M\ge B_{p-1}/(4K) M ≥ B p − 1 / ( 4 K ) . Then M ≤ B p ≤ B p − 1 ≤ 4 K M M\le B_p\le B_{p-1}\le4KM M ≤ B p ≤ B p − 1 ≤ 4 K M , and the peak inequality implies Q = O K ( N ) Q=O_K(\sqrt{N}) Q = O K ( N ) . Since all earlier tails are at least B p B_p B p ,
B p E ℓ , p ( T ) ≤ T 2 N . B_pE_{\ell,p}(T)\le\frac{T^2}{N}. B p E ℓ , p ( T ) ≤ N T 2 .
Also
B p L p ( T ) = B p M ( T F p , M ) α p . B_pL_p(T)=\sqrt{\frac{B_p}{M}}\left(\frac{T}{F_{p,M}}\right)^{\alpha_p}. B p L p ( T ) = M B p ( F p , M T ) α p .
It follows that μ p ( ε Q ) ≤ C K ( ε α p + ε 2 ) \mu_p(\varepsilon Q)\le C_K(\varepsilon^{\alpha_p}+\varepsilon^2) μ p ( εQ ) ≤ C K ( ε α p + ε 2 ) . Any recorded r r r -collision entails at least one occupied-p p p bucket, so Markov’s inequality bounds the target database probability by this quantity.
Otherwise M < B p − 1 / ( 4 K ) M<B_{p-1}/(4K) M < B p − 1 / ( 4 K ) . Among the capacities c ∈ { p − 1 , … , r − 1 } c\in\{p-1,\ldots,r-1\} c ∈ { p − 1 , … , r − 1 } choose one with
n c ≥ B p − 1 − M K ≥ B p − 1 2 K , m = n c + 1. n_c\ge\frac{B_{p-1}-M}{K}\ge\frac{B_{p-1}}{2K},\qquad m=n_c+1. n c ≥ K B p − 1 − M ≥ 2 K B p − 1 , m = n c + 1.
Thus m ≍ K B c ≍ K B p − 1 m\asymp_K B_c\asymp_K B_{p-1} m ≍ K B c ≍ K B p − 1 , m > M m>M m > M , and every tail through c c c is at least a constant multiple of max { M , D } \max\{M,D\} max { M , D } . Moreover F c + 1 , M ≍ K Q F_{c+1,M}\asymp_K Q F c + 1 , M ≍ K Q . If c = p − 1 c=p-1 c = p − 1 this is immediate. Otherwise p < r p<r p < r , and B p ≤ D B_p\le D B p ≤ D while B i ≥ m B_i\ge m B i ≥ m for p ≤ i ≤ c p\le i\le c p ≤ i ≤ c . Hence those B i B_i B i are all within K K K -dependent constant factors of D D D . Starting with F p , M = Q F_{p,M}=Q F p , M = Q , iterate
F j + 1 , M α j + 1 = N 1 / 2 M − 1 / 4 B j 1 / 4 F j , M α j / 2 . F_{j+1,M}^{\alpha_{j+1}}=N^{1/2}M^{-1/4}B_j^{1/4}F_{j,M}^{\alpha_j/2}. F j + 1 , M α j + 1 = N 1/2 M − 1/4 B j 1/4 F j , M α j /2 .
The definition of D D D proves the claimed comparison at each step.
For each target capacity d ≥ r d\ge r d ≥ r with n d > 0 n_d>0 n d > 0 , lower one such bucket to capacity c c c . Its lower tails B i B_i B i , i ≤ c i\le c i ≤ c , are exactly unchanged. The parent has m = n c + 1 m=n_c+1 m = n c + 1 candidate buckets, and the final mark is uniform among the n d n_d n d target buckets of capacity d d d . Apply the localized energy estimate for c → d c\to d c → d , and sum over d d d . Writing W c + 1 ( T ) W_{c+1}(T) W c + 1 ( T ) for the expected number of capacity-at-least-r r r buckets having at least c + 1 c+1 c + 1 records, this gives
W c + 1 ( T ) ≤ C K M N ∑ t < T A c s r c ( t ) + C K M T 2 m N . W_{c+1}(T)\le C_K\sqrt{\frac{M}{N}\sum_{t<T}A_c^{\mathrm{src}}(t)}+C_K\frac{MT^2}{mN}. W c + 1 ( T ) ≤ C K N M t < T ∑ A c src ( t ) + C K m N M T 2 .
Lemma 7.8 gives a source moment bound uniform over the different parents and uses the unchanged tails through c c c . Its leading contribution is C K ( T / F c + 1 , M ) α c + 1 C_K(T/F_{c+1,M})^{\alpha_{c+1}} C K ( T / F c + 1 , M ) α c + 1 . Each propagated error has denominator exponents summing to one; after the last square root their exponents sum to one half. Since all involved tails are at least a constant multiple of m m m , their combined contribution is at most
C K T 2 N M m . C_K\frac{T^2}{N}\sqrt{\frac{M}{m}}. C K N T 2 m M .
At T = ε Q T=\varepsilon Q T = εQ , this and the final local remainder are O K ( ε 2 ) O_K(\varepsilon^2) O K ( ε 2 ) . Indeed, use m ≥ a K M m \geq a_KM m ≥ a K M if Q ≤ N Q \leq\sqrt{N} Q ≤ N , and m ≥ a K M Q 4 / N 2 m \geq a_KMQ^4/N^2 m ≥ a K M Q 4 / N 2 if Q ≥ N Q \geq\sqrt{N} Q ≥ N . Consequently
W c + 1 ( ε Q ) ≤ C K ( ε α c + 1 + ε 2 ) . W_{c+1}(\varepsilon Q) \leq C_K(\varepsilon^{\alpha c+1}+\varepsilon^2). W c + 1 ( εQ ) ≤ C K ( ε α c + 1 + ε 2 ) .
Every recorded target r r r -collision entails one of these events because c + 1 ≤ r c+1 \leq r c + 1 ≤ r . Markov again bounds its database probability.
For sufficiently small ε = ε ( K , r , p 0 ) > 0 \varepsilon=\varepsilon(K,r,p_0)>0 ε = ε ( K , r , p 0 ) > 0 , the preceding database probability bounds give ∥ P r ξ T ∥ < p 0 / 2 \lVert P_r\xi_T\rVert< \sqrt{p_0}/2 ∥ P r ξ T ∥ < p 0 /2 at T ≤ ε Q T \leq\varepsilon Q T ≤ εQ . Since B i ≤ N B_i \leq N B i ≤ N and M ≥ 1 M \geq1 M ≥ 1 , we have
Q ≤ N γ K , γ K = 1 − 2 K − 2 2 K − 1 < 1. Q \leq N^{\gamma_K}, \qquad\gamma_K=1-\frac{2^{K-2}}{2^K-1}<1. Q ≤ N γ K , γ K = 1 − 2 K − 1 2 K − 2 < 1.
Hence the query budget is at most N / 8 N/8 N /8 for large N N N , and the error in Lemma 7.9 is uniformly o ( 1 ) o(1) o ( 1 ) . This rules out real success probability at least p 0 p_0 p 0 . All estimates are uniform for each injective label table. Squaring before averaging over that table yields the uniform exact-type lower bound for every q ≥ B 1 q \geq B_1 q ≥ B 1 . An algorithm with the stated success guarantee on every input has that guarantee on average, which also proves the worst-case lower bound.
Matching search upper bounds
We now match the type-dependent lower bound by algorithms that use only the function oracle. Throughout this section, use B j B_j B j , M = B r M=B_r M = B r , and Q Q Q from (122) –(123) . The type may be supplied to the algorithm, but the equality classes and their input positions are hidden. The inverse of the addition oracle is obtained by conjugating its response register with negation, so using an inverse call does not require an additional kind of access.
Theorem 8.1 (Type-dependent search upper bound). For every fixed K ≥ r ≥ 2 K \geq r \geq2 K ≥ r ≥ 2 , every sufficiently large N N N , and every bounded exact type with B r ≥ 1 B_r \geq1 B r ≥ 1 , a quantum algorithm finds one r r r -collision with probability at least 2 / 3 2/3 2/3 on every function of that type using O K , r ( Q ) O_{K,r}(Q) O K , r ( Q ) function queries. The implicit constant is uniform over all feasible alphabets q ≥ B 1 q \geq B_1 q ≥ B 1 .
The constructions on the two sides of Q = N Q=\sqrt{N} Q = N use the same preparation-cost expression. For raw set sizes a 1 , … , a r − 1 a_1,\ldots,a_{r-1} a 1 , … , a r − 1 , it is
a 1 + ∑ j = 2 r − 1 a j B j ∏ i < j a i N j / 2 , (136) a_1+\sum_{j=2}^{r-1}\frac{a_j\sqrt{B_j\prod_{i<j}a_i}}{N^{j/2}}, \tag*{(136)} a 1 + j = 2 ∑ r − 1 N j /2 a j B j ∏ i < j a i , (136)
and the target-completion scale is N r / 2 / M ∏ i < r a i N^{r/2}/\sqrt{M\prod_{i<r}a_i} N r /2 / M ∏ i < r a i . We first choose parameters making these quantities O K , r ( Q ) O_{K,r}(Q) O K , r ( Q ) . For Q ≤ N Q \leq\sqrt{N} Q ≤ N , sampled tables and successive Grover enumeration realize the parameters directly. For Q ≥ N Q \geq\sqrt{N} Q ≥ N , ordered rolewise restriction and a colored adaptive learning graph suffice. A coherent decision-to-search conversion avoids logarithmic overhead.
Direct search below the square-root scale
The regime below the square-root scale admits a direct search algorithm. It uses classical tables and Grover enumeration of matching values in independently sampled, disjoint input roles. In particular, it requires neither a quantum walk on tuple lists nor a coherent update of those lists.
We first choose the sampling parameters so that every stage has the desired query cost and a constant expected number of target classes survives all stages.
Lemma 8.2 (Sampling parameters). Suppose Q ≤ N Q \le\sqrt{N} Q ≤ N . There are real parameters Q ≤ a i ≤ N Q \le a_i \le N Q ≤ a i ≤ N for 1 ≤ i < r 1 \le i < r 1 ≤ i < r , with a 1 = Q a_1 = Q a 1 = Q , such that, on writing
z j = ∏ i = 1 j a i N j , z_j = \frac{\prod_{i=1}^{j} a_i}{N^j}, z j = N j ∏ i = 1 j a i , one has
a j B j z j ≤ Q ( 2 ≤ j < r ) , N M z r − 1 ≤ Q . (137) \sqrt{a_j B_j z_j} \le Q \qquad(2 \le j < r), \qquad\sqrt{\frac{N}{M z_{r-1}}} \le Q. \tag*{(137)} a j B j z j ≤ Q ( 2 ≤ j < r ) , M z r − 1 N ≤ Q . (137) Moreover z 1 ≥ ⋯ ≥ z r − 1 z_1 \ge\cdots\ge z_{r-1} z 1 ≥ ⋯ ≥ z r − 1 and B j z j ≥ 1 B_j z_j \ge1 B j z j ≥ 1 for 2 ≤ j < r 2 \le j < r 2 ≤ j < r , while M z r − 1 ≥ 1 M z_{r-1} \ge1 M z r − 1 ≥ 1 .
Proof. Put z 1 = Q / N z_1 = Q/N z 1 = Q / N , and for 2 ≤ j < r 2 \le j < r 2 ≤ j < r define
z j = min { z j − 1 , Q z j − 1 N B j } , a j = N z j / z j − 1 . z_j = \min\left\{z_{j-1}, \frac{Q\sqrt{z_{j-1}}}{\sqrt{N B_j}}\right\}, \qquad a_j = Nz_j/z_{j-1}. z j = min { z j − 1 , N B j Q z j − 1 } , a j = N z j / z j − 1 .
These definitions imply a j ≤ N a_j \le N a j ≤ N and
a j B j z j = z j N B j z j − 1 ≤ Q . \sqrt{a_j B_j z_j} = z_j \sqrt{\frac{N B_j}{z_{j-1}}} \le Q. a j B j z j = z j z j − 1 N B j ≤ Q .
They also imply a j ≥ Q a_j \ge Q a j ≥ Q : the uncapped expression for a j a_j a j is Q N / ( B j z j − 1 ) ≥ Q Q\sqrt{N/(B_j z_{j-1})} \ge Q Q N / ( B j z j − 1 ) ≥ Q , and N ≥ Q N \ge Q N ≥ Q . Expanding the minima expresses z r − 1 z_{r-1} z r − 1 as the minimum, over increasing subsequences 2 ≤ i 1 < ⋯ < i m ≤ r − 1 2 \le i_1 < \cdots< i_m \le r - 1 2 ≤ i 1 < ⋯ < i m ≤ r − 1 , including the empty subsequence, of
Q 2 − 2 − m N ∏ t = 1 m B i t 2 − ( m − t + 1 ) . \frac{Q^{2-2^{-m}}}{N\prod_{t=1}^{m} B_{i_t}^{2^{-(m-t+1)}}}. N ∏ t = 1 m B i t 2 − ( m − t + 1 ) Q 2 − 2 − m .
For a fixed m m m , monotonicity of the B i B_i B i makes this expression smallest at i t = t + 1 i_t = t+1 i t = t + 1 . The j = m + 2 j=m+2 j = m + 2 term in (123) gives
Q 4 − 2 − m ≥ N 2 M ∏ t = 1 m B t + 1 2 − ( m − t + 1 ) . Q^{4-2^{-m}} \ge\frac{N^2}{M}\prod_{t=1}^{m} B_{t+1}^{2^{-(m-t+1)}}. Q 4 − 2 − m ≥ M N 2 t = 1 ∏ m B t + 1 2 − ( m − t + 1 ) .
Consequently z r − 1 ≥ N / ( M Q 2 ) z_{r-1} \ge N/(M Q^2) z r − 1 ≥ N / ( M Q 2 ) , which proves the final cost bound. Since the z j z_j z j are nonincreasing, B j ≥ M B_j \ge M B j ≥ M , and Q 2 ≤ N Q^2 \le N Q 2 ≤ N , all the stated lower bounds on the expected counts follow. This argument includes r = 2 r=2 r = 2 , when the subsequence is empty.
The enumeration subroutine needs a fixed total budget, rather than a separate high-confidence guarantee for every discovered input.
Lemma 8.3 (Fixed-budget enumeration). Let a known set have size a a a , and suppose a query-computable predicate marks h ≤ h max h \le h_{\max} h ≤ h m a x of its positions, where h max ≥ 1 h_{\max} \ge1 h m a x ≥ 1 is known. For any fixed ε > 0 \varepsilon> 0 ε > 0 , all marked positions can be found with probability at least 1 − ε 1-\varepsilon 1 − ε using a worst-case budget O ε ( a h max ) O_{\varepsilon}(\sqrt{a h_{\max}}) O ε ( a h m a x ) predicate queries .
Proof. If a = 0 a=0 a = 0 , return the empty list. Otherwise repeatedly run verified randomized Grover search [10 ] , excluding already found positions, until a single predetermined total budget is exhausted. Only verified positions are added to the list. When u > 0 u>0 u > 0 marked positions remain, the expected time to discover the next one is O ( a / u ) O(\sqrt{a/u}) O ( a / u ) . Thus the expected time T a l l T_{\mathrm{all}} T all at which all h h h positions have been found satisfies
E T a l l ≤ C ∑ u = 1 h a / u ≤ 2 C a h ≤ 2 C a h max . \mathbb{E}T_{\mathrm{all}} \le C\sum_{u=1}^{h}\sqrt{a/u} \le2C\sqrt{ah} \le2C\sqrt{ah_{\max}}. E T all ≤ C u = 1 ∑ h a / u ≤ 2 C ah ≤ 2 C a h m a x .
For h = 0 h=0 h = 0 the list is already complete. No test certifying that the remaining set is empty is needed. Markov’s inequality gives Pr [ T a l l > D a h max ] ≤ 2 C / D \Pr[T_{\mathrm{all}}>D\sqrt{ah_{\max}}]\le2C/D Pr [ T all > D a h m a x ] ≤ 2 C / D . Choose D ≥ 2 C / ε D\ge2C/\varepsilon D ≥ 2 C / ε and stop at that budget, even when the promise on h h h fails. This proves the stated guarantee and the worst-case query bound.
Theorem 8.4 (Search below the square-root scale). Fix K ≥ r ≥ 2 K\ge r\ge2 K ≥ r ≥ 2 . For every sufficiently large N N N and every input of maximum class size at most K K K satisfying Q ≤ N Q\le\sqrt{N} Q ≤ N , an algorithm finds an r r r -collision with probability at least 2 / 3 2/3 2/3 using O K , r ( Q ) O_{K,r}(Q) O K , r ( Q ) function queries. Its guarantee holds for every such function and is uniform over all feasible alphabets.
Proof. Use the parameters from Lemma 8.2. Independently assign each input a uniform role in { 0 , 1 , … , r } \{0,1,\ldots,r\} { 0 , 1 , … , r } . For 1 ≤ i < r 1\le i<r 1 ≤ i < r , retain each input assigned role i i i , independently, with probability a i / N a_i/N a i / N ; call the resulting public input set A i A_i A i . Let A r A_r A r contain all inputs assigned role r r r . The sets A i A_i A i are pairwise disjoint. Their indices are known without querying f f f .
Query all of A 1 A_1 A 1 . At stage j j j , for 2 ≤ j < r 2\le j<r 2 ≤ j < r , enumerate the inputs of A j A_j A j whose values occur in the preceding table, storing their values and one prefix tuple for each value found. A prefix tuple has one input from each of A 1 , … , A j A_1,\ldots,A_j A 1 , … , A j . If a value occurs several times in a role, retain any one representative for constructing its prefix tuple. Finally, use Grover search on A r A_r A r for an input whose value occurs in the final prefix table. The stored prefix and this input give r r r distinct inputs, because their roles are distinct.
We analyze the ideal tables obtained when each enumeration is complete. Let H j H_j H j be the number of marked inputs in A j A_j A j at stage j j j . For a class to contribute to H j H_j H j , it must contain distinct selected inputs in all of the first j j j roles. A union bound over ordered j j j -tuples within each class gives
E H j ≤ ( K ) j ( r + 1 ) j B j z j ≤ C K , r B j z j ( 2 ≤ j < r ) . (138) \mathbb{E}H_j \le\frac{(K)_j}{(r+1)^j}B_jz_j \le C_{K,r}B_jz_j \qquad(2\le j<r). \tag*{(138)} E H j ≤ ( r + 1 ) j ( K ) j B j z j ≤ C K , r B j z j ( 2 ≤ j < r ) . (138) An unsuccessful earlier enumeration can only reduce the later marked counts, so this ideal count also bounds the count seen by the actual algorithm.
For each of the M M M target classes, fix r r r distinct inputs for the analysis. Let its favorable event mean that these inputs are assigned, in order, roles 1 , … , r 1,\ldots,r 1 , … , r , and that the first r − 1 r-1 r − 1 are retained in their corresponding sets. These events are independent across classes, each with probability c r z r − 1 c_rz_{r-1} c r z r − 1 , where c r = ( r + 1 ) − r c_r=(r+1)^{-r} c r = ( r + 1 ) − r . If Y Y Y counts favorable classes, then
λ : = E Y = c r M z r − 1 ≥ c r , E Y 2 ≤ λ + λ 2 . \lambda:=\mathbb{E}Y=c_rMz_{r-1}\ge c_r,\qquad\mathbb{E}Y^2\le\lambda+\lambda^2. λ := E Y = c r M z r − 1 ≥ c r , E Y 2 ≤ λ + λ 2 . The second-moment inequality yields
Pr [ Y ≥ λ / 2 ] ≥ λ 4 ( 1 + λ ) ≥ p ∗ : = c r 4 ( 1 + c r ) > 0. (139) \Pr[Y\ge\lambda/2]\ge\frac{\lambda}{4(1+\lambda)}\ge p_*:=\frac{c_r}{4(1+c_r)}>0. \tag*{(139)} Pr [ Y ≥ λ /2 ] ≥ 4 ( 1 + λ ) λ ≥ p ∗ := 4 ( 1 + c r ) c r > 0. (139) Every favorable class supplies a marked input to the final search whenever all prefix enumerations are complete.
Choose a constant C H = C H ( K , r ) C_H=C_H(K,r) C H = C H ( K , r ) sufficiently large that Markov’s inequality and (8.3) give
Pr [ H j > C H B j z j for some 2 ≤ j < r ] ≤ p ∗ / 4. \Pr[H_j>C_HB_jz_j\text{ for some }2\le j<r]\le p_*/4. Pr [ H j > C H B j z j for some 2 ≤ j < r ] ≤ p ∗ /4.
Also, ∣ A i ∣ |A_i| ∣ A i ∣ is binomial with mean a i / ( r + 1 ) a_i/(r+1) a i / ( r + 1 ) . Since a i ≥ Q ≥ N 1 / 3 a_i \ge Q \ge N^{1/3} a i ≥ Q ≥ N 1/3 , Chernoff bounds imply, for all sufficiently large N N N ,
Pr [ ∣ A i ∣ > 2 a i for some 1 ≤ i < r ] ≤ p ∗ / 4. \Pr[|A_i| > 2a_i\text{ for some }1 \le i < r] \le p_*/4. Pr [ ∣ A i ∣ > 2 a i for some 1 ≤ i < r ] ≤ p ∗ /4.
Here Q ≥ N 1 / 3 Q \ge N^{1/3} Q ≥ N 1/3 follows already from the j = 2 j=2 j = 2 term and M ≤ N M \le N M ≤ N . No independence between these upper-count events and the target event is required. Subtracting their failure probabilities from (139) shows that
Y ≥ λ / 2 , ∣ A i ∣ ≤ 2 a i , H j ≤ C H B j z j (140) Y \ge\lambda/2,\quad|A_i| \le2a_i,\quad H_j \le C_HB_jz_j \tag*{(140)} Y ≥ λ /2 , ∣ A i ∣ ≤ 2 a i , H j ≤ C H B j z j (140)
hold jointly with probability at least p ∗ / 2 p_*/2 p ∗ /2 .
Abort an attempt if any public size bound ∣ A i ∣ ≤ 2 a i |A_i| \le2a_i ∣ A i ∣ ≤ 2 a i fails. For each enumeration, apply Lemma 8.3 with h max = ⌈ C H B j z j ⌉ h_{\max} = \lceil C_HB_jz_j\rceil h m a x = ⌈ C H B j z j ⌉ and failure probability at most p ∗ / ( 8 r ) p_*/(8r) p ∗ / ( 8 r ) . Testing whether a value occurs in the preceding table uses O ( 1 ) O(1) O ( 1 ) function queries; table operations do not add queries. The resulting query cost is
O K , r ( ∣ A j ∣ ( C H B j z j + 1 ) ) = O K , r ( a j B j z j ) = O K , r ( Q ) . O_{K,r}\left(\sqrt{|A_j|(C_HB_jz_j+1)}\right)
= O_{K,r}\left(\sqrt{a_jB_jz_j}\right)
= O_{K,r}(Q). O K , r ( ∣ A j ∣ ( C H B j z j + 1 ) ) = O K , r ( a j B j z j ) = O K , r ( Q ) .
The first equality uses B j z j ≥ 1 B_jz_j \ge1 B j z j ≥ 1 . This budget is fixed even on inputs to the subroutine with more marked positions; only its success guarantee is conditional on the count bound. Conditional on (140) , a union bound over the fixed number of stages leaves enumeration failure probability at most p ∗ / 8 p_*/8 p ∗ /8 .
The final randomized Grover search is given a fixed budget O r ( N / λ ) O_r(\sqrt{N/\lambda}) O r ( N / λ ) , which is O r ( Q ) O_r(Q) O r ( Q ) by Lemma 8.2 . When there are at least λ / 2 \lambda/2 λ /2 marked inputs, it succeeds with probability at least 2 / 3 2/3 2/3 after adjusting the constant. This statement also covers λ / 2 < 1 \lambda/2 < 1 λ /2 < 1 : the number of marked inputs is integral, so the favorable event then guarantees at least one. Every reported solution is verified.
Combining (140) , the enumeration guarantee, and the final search gives a positive success probability depending only on K , r K,r K , r . A constant number of independent attempts raises it to 2 / 3 2/3 2/3 . Querying A 1 A_1 A 1 costs at most 2 Q 2Q 2 Q , there are only r − 2 r-2 r − 2 enumeration stages, and all budgets are fixed. The resulting worst-case query bound is therefore O K , r ( Q ) O_{K,r}(Q) O K , r ( Q ) .
A colored learning-graph bound with prescribed prefix populations
We give an explicit dual-adversary construction for the colored problem. It adapts the fault correction in Belovs [6, Section 6], with separate input domains for the certificate positions and a fault budget determined by the maximum multiplicity. In particular, the argument below verifies the off-diagonal constraints; the upper bound does not follow merely by substituting the domain sizes in an existing complexity formula.
Lemma 8.5 (Colored search bound with prescribed prefix populations). Let U 1 , … , U r U_1,\ldots,U_r U 1 , … , U r be disjoint sets, with ∣ U i ∣ = v i |U_i|=v_i ∣ U i ∣ = v i , and let X \mathcal{X} X be a finite promise domain of inputs on their union, with values in a finite alphabet E \mathcal{E} E . Every equivalence class of every input has size at most a fixed K ≥ r K \ge r K ≥ r . Let F ( x ) = 1 F(x)=1 F ( x ) = 1 if and only if there is an equal-value tuple in U 1 × ⋯ × U r U_1 \times\cdots\times U_r U 1 × ⋯ × U r . For 2 ≤ j < r 2 \le j < r 2 ≤ j < r , suppose every x ∈ X x \in\mathcal{X} x ∈ X has at most C j C_j C j equal-value tuples in U 1 × ⋯ × U j U_1 \times\cdots\times U_j U 1 × ⋯ × U j . Put b = K + 1 b=K+1 b = K + 1 and L i = b i − 1 ( 2 b − 1 ) L_i=b^{i-1}(2^b-1) L i = b i − 1 ( 2 b − 1 ) . For positive integers a i a_i a i satisfying L i a i ≤ v i / 4 L_i a_i \le v_i/4 L i a i ≤ v i /4 for 1 ≤ i < r 1 \le i < r 1 ≤ i < r , the bounded-error quantum query complexity of F F F is
O K , r ( a 1 + ∑ j = 2 r − 1 a j C j ∏ i < j a i ∏ i ≤ j v i + ∑ s = 1 r ∏ i ≤ s v i ∏ i < s a i ) . (141) O_{K,r}\left(a_1+\sum_{j=2}^{r-1}a_j\sqrt{\frac{C_j\prod_{i<j}a_i}{\prod_{i\le j}v_i}}+\sum_{s=1}^{r}\sqrt{\frac{\prod_{i\le s}v_i}{\prod_{i<s}a_i}}\right). \tag*{(141)} O K , r ( a 1 + j = 2 ∑ r − 1 a j ∏ i ≤ j v i C j ∏ i < j a i + s = 1 ∑ r ∏ i < s a i ∏ i ≤ s v i ) . (141) The prefix bounds must hold on both positive and negative inputs. Uniqueness of the colored r r r -tuple is not required.
If the positive promise domain is empty, the conclusion is immediate. Otherwise, choose a single colored certificate ( c 1 ( x ) , … , c r ( x ) ) (c_1(x),\ldots,c_r(x)) ( c 1 ( x ) , … , c r ( x )) for every positive input x x x , by an arbitrary fixed deterministic rule. Write ν x = x c 1 ( x ) \nu_x=x_{c_1(x)} ν x = x c 1 ( x ) and Z x = { t : x t = ν x } Z_x=\{t:x_t=\nu_x\} Z x = { t : x t = ν x } for its entire equality class. This choice is used only to define the feasible matrices; no certificate or equality-class information is supplied to the algorithm.
Raw sets and certificate stages. Let R \mathcal{R} R consist of families of pairwise disjoint sets
R i ( d 1 , … , d i − 1 , J ) ⊆ U i , i < r , d h ∈ [ b ] , ∅ ≠ J ⊆ [ b ] , R_i(d_1,\ldots,d_{i-1},J)\subseteq U_i,\qquad i<r,\quad d_h\in[b],\quad\varnothing\ne J\subseteq[b], R i ( d 1 , … , d i − 1 , J ) ⊆ U i , i < r , d h ∈ [ b ] , ∅ = J ⊆ [ b ] ,
with every set at level i i i having size a i a_i a i . There are L i L_i L i sets at that level. For each family R R R , fix an order of its elements in which all level-i i i elements precede all level-( i + 1 ) (i+1) ( i + 1 ) elements. The setup indices are the successive queries in this order; an index retains the full tag R R R , even when only part of the family is loaded.
Call R R R consistent with a positive input x x x if ⋃ R \bigcup R ⋃ R is disjoint from Z x Z_x Z x . Thus consistency excludes the whole selected equality class, not just the chosen certificate. Let R x \mathcal{R}_x R x be the consistent families and set
N R = ∣ R ∣ , N R ( x ) = ∣ R x ∣ , q 0 = N R − 1 , γ x = N R N R ( x ) , p x = N R ( x ) − 1 . (142) N_{\mathcal{R}}=|\mathcal{R}|,\qquad N_{\mathcal{R}}(x)=|\mathcal{R}_x|,\qquad q_0=N_{\mathcal{R}}^{-1},\qquad\gamma_x=\frac{N_{\mathcal{R}}}{N_{\mathcal{R}}(x)},\qquad p_x=N_{\mathcal{R}}(x)^{-1}. \tag*{(142)} N R = ∣ R ∣ , N R ( x ) = ∣ R x ∣ , q 0 = N R − 1 , γ x = N R ( x ) N R , p x = N R ( x ) − 1 . (142)
In particular, q 0 γ x = p x q_0\gamma_x=p_x q 0 γ x = p x . The volume condition on the raw sets implies 1 ≤ γ x = O K , r ( 1 ) 1\le\gamma_x=O_{K,r}(1) 1 ≤ γ x = O K , r ( 1 ) ; this is also proved in the diagonal calculation below.
For a nonempty J ⊆ [ b ] J\subseteq[b] J ⊆ [ b ] , put μ ( J ) = min J \mu(J)=\min J μ ( J ) = min J . Given R ∈ R x R\in\mathcal{R}_x R ∈ R x and nonempty subsets J 1 , … , J s − 1 ⊆ [ b ] J_1,\ldots,J_{s-1}\subseteq[b] J 1 , … , J s − 1 ⊆ [ b ] , form a family S S S by adjoining c i ( x ) c_i(x) c i ( x ) to the cell
R i ( μ ( J 1 ) , … , μ ( J i − 1 ) , J i ) ( 1 ≤ i < s ) . R_i\bigl(\mu(J_1),\ldots,\mu(J_{i-1}),J_i\bigr)\qquad(1\le i<s). R i ( μ ( J 1 ) , … , μ ( J i − 1 ) , J i ) ( 1 ≤ i < s ) .
Denote this family by R [ J 1 ← c 1 ( x ) , … , J s − 1 ← c s − 1 ( x ) ] R[J_1\leftarrow c_1(x),\ldots,J_{s-1}\leftarrow c_{s-1}(x)] R [ J 1 ← c 1 ( x ) , … , J s − 1 ← c s − 1 ( x )] . The certificate-stage indices at stage s s s are pairs ( S , j ) (S,j) ( S , j ) with j ∈ U s ∖ ⋃ S j\in U_s\setminus\bigcup S j ∈ U s ∖ ⋃ S . The enlarged cells in S S S determine J 1 , … , J s − 1 J_1,\ldots,J_{s-1} J 1 , … , J s − 1 . The index ( S , c s ( x ) ) (S,c_s(x)) ( S , c s ( x )) is taken by x x x and R R R for every such branch sequence. For fixed x x x , each taken index determines R R R uniquely: remove c i ( x ) c_i(x) c i ( x ) from its enlarged cell for every i < s i<s i < s . All indices may be included, even if no positive input takes them.
These are indices for the matrices constructed below. Their use does not require interpreting the signed correction as an ordinary nonnegative flow.
Partially revealed assignments. For any loaded family S S S and input z z z , define an assignment α z S : ⋃ S → Σ ∪ { ⋆ } \alpha_z^S:\bigcup S\to\Sigma\cup\{\star\} α z S : ⋃ S → Σ ∪ { ⋆ } recursively by level. Every element of a level-1 cell is revealed, so its assignment is its input value. For i > 1 i>1 i > 1 and t ∈ S i ( d 1 , … , d i − 1 , J ) t\in S_i(d_1,\ldots,d_{i-1},J) t ∈ S i ( d 1 , … , d i − 1 , J ) , reveal z t z_t z t precisely when it occurs among the revealed values in
⋃ J ′ ∋ d i − 1 S i − 1 ( d 1 , … , d i − 2 , J ′ ) . (143) \bigcup_{J'\ni d_{i-1}}S_{i-1}(d_1,\ldots,d_{i-2},J'). \tag*{(143)} J ′ ∋ d i − 1 ⋃ S i − 1 ( d 1 , … , d i − 2 , J ′ ) . (143)
Otherwise put α z S ( t ) = ⋆ \alpha_z^S(t)=\star α z S ( t ) = ⋆ . The symbol ⋆ \star ⋆ is not an input value. Write x ∼ S y x\sim_S y x ∼ S y when α x S = α y S \alpha_x^S=\alpha_y^S α x S = α y S .
During setup, a newly loaded element can change only its own assignment: all lower levels are already loaded, and there are no loaded higher levels. At certificate stages assignments are recomputed, so the insertion of a certificate element can uncover previously hidden raw elements. The signed construction below accounts for exactly this effect.
Positive semidefinite matrices. We construct real positive semidefinite matrices X j X_j X j , indexed by X \mathcal{X} X , such that
∑ j : x j ≠ y j X j [ x , y ] = 1 ( F ( x ) = 1 , F ( y ) = 0 ) . (144) \sum_{j:x_j\ne y_j}X_j[x,y]=1\qquad\bigl(F(x)=1,\ F(y)=0\bigr). \tag*{(144)} j : x j = y j ∑ X j [ x , y ] = 1 ( F ( x ) = 1 , F ( y ) = 0 ) . (144)
By (9) , these matrices give a query upper bound within a universal constant of max z ∑ j X j [ z , z ] \max_z \sum_j X_j[z,z] max z ∑ j X j [ z , z ] .
Fix an index e e e loading input j j j , let S S S be its source family, and fix an assignment α \alpha α on S S S . Let δ e ( z ) \delta_e(z) δ e ( z ) mean that z z z is positive and takes index e e e for some consistent raw family. Every vector defined below is zero unless α z S = α \alpha^S_z=\alpha α z S = α ; its positive coordinates are also zero unless δ e ( z ) \delta_e(z) δ e ( z ) holds. These support conditions are understood in all displayed formulas. All weights are positive and constant within the stage and case under consideration.
For a level-1 setup index use the vector
u e , α ( z ) = { γ z / w , F ( z ) = 1 , δ e ( z ) , w , F ( z ) = 0 , Y e , α = q 0 u e , α u e , α ∗ . (145) u_{e,\alpha}(z)=
\begin{cases}
\gamma_z/\sqrt{w}, & F(z)=1,\ \delta_e(z),\\
\sqrt{w}, & F(z)=0,
\end{cases}
\qquad
Y_{e,\alpha}=q_0u_{e,\alpha}u_{e,\alpha}^{*}. \tag*{(145)} u e , α ( z ) = { γ z / w , w , F ( z ) = 1 , δ e ( z ) , F ( z ) = 0 , Y e , α = q 0 u e , α u e , α ∗ . (145)
For a level-i i i setup index with i > 1 i>1 i > 1 , let χ e ( z ) \chi_e(z) χ e ( z ) indicate that z j z_j z j lies in the revealed parent-value set specified by (143) and α \alpha α . Define two vectors by
u e , α ( z ) = { γ z / w 1 , F ( z ) = 1 , δ e ( z ) , χ e ( z ) = 1 , w 1 , F ( z ) = 0 , 0 , otherwise , u_{e,\alpha}(z)=
\begin{cases}
\gamma_z/\sqrt{w_1}, & F(z)=1,\ \delta_e(z),\ \chi_e(z)=1,\\
\sqrt{w_1}, & F(z)=0,\\
0, & \text{otherwise},
\end{cases} u e , α ( z ) = ⎩ ⎨ ⎧ γ z / w 1 , w 1 , 0 , F ( z ) = 1 , δ e ( z ) , χ e ( z ) = 1 , F ( z ) = 0 , otherwise ,
v e , α ( z ) = { γ z / w 0 , F ( z ) = 1 , δ e ( z ) , χ e ( z ) = 0 , w 0 , F ( z ) = 0 , χ e ( z ) = 1 , 0 , otherwise , Y e , α = q 0 ( u u ∗ + v v ∗ ) . (146) v_{e,\alpha}(z)=
\begin{cases}
\gamma_z/\sqrt{w_0}, & F(z)=1,\ \delta_e(z),\ \chi_e(z)=0,\\
\sqrt{w_0}, & F(z)=0,\ \chi_e(z)=1,\\
0, & \text{otherwise},
\end{cases}
\qquad
Y_{e,\alpha}=q_0(uu^{*}+vv^{*}). \tag*{(146)} v e , α ( z ) = ⎩ ⎨ ⎧ γ z / w 0 , w 0 , 0 , F ( z ) = 1 , δ e ( z ) , χ e ( z ) = 0 , F ( z ) = 0 , χ e ( z ) = 1 , otherwise , Y e , α = q 0 ( u u ∗ + v v ∗ ) . (146)
Consequently, for a positive x x x taking e e e and a negative y y y with x ∼ S y x\sim_S y x ∼ S y , the off-diagonal contribution is p x p_x p x if at least one of x j , y j x_j,y_j x j , y j is revealed after the setup query, and is zero if both are hidden.
For a stage-s s s certificate index with branch prefix J 1 , … , J s − 1 J_1,\ldots,J_{s-1} J 1 , … , J s − 1 , put
σ s = ( − 1 ) s − 1 + ∑ i < s ∣ J i ∣ , u e , α ( z ) = { w s , F ( z ) = 1 , δ e ( z ) , σ s w s , F ( z ) = 0 , Y e , α = q 0 u u ∗ . (147) \sigma_s=(-1)^{s-1+\sum_{i<s}|J_i|},
\qquad
u_{e,\alpha}(z)=
\begin{cases}
\sqrt{w_s}, & F(z)=1,\ \delta_e(z),\\
\sigma_s\sqrt{w_s}, & F(z)=0,
\end{cases}
\qquad
Y_{e,\alpha}=q_0uu^{*}. \tag*{(147)} σ s = ( − 1 ) s − 1 + ∑ i < s ∣ J i ∣ , u e , α ( z ) = { w s , σ s w s , F ( z ) = 1 , δ e ( z ) , F ( z ) = 0 , Y e , α = q 0 u u ∗ . (147)
Its positive-negative entry is σ s p x \sigma_s p_x σ s p x when the source assignments agree, and zero otherwise. Finally define
X j = ∑ e l o a d i n g j ∑ α Y e , α . (148) X_j=\sum_{e\ \mathrm{loading}\ j}\sum_{\alpha}Y_{e,\alpha}. \tag*{(148)} X j = e loading j ∑ α ∑ Y e , α . (148)
Every summand is a Gram matrix of one or two real vectors. In particular, negative off-diagonal signs in (147) do not compromise positive semidefiniteness.
The off-diagonal identity. Fix a positive x x x , a negative y y y , and R ∈ R x R\in R_x R ∈ R x . We prove that the indices taken for this choice of x , R x,R x , R contribute exactly p x p_x p x to the left side of (144) . Let s s s be the first certificate position with x c s ( x ) ≠ y c s ( x ) x_{c_s(x)}\ne y_{c_s(x)} x c s ( x ) = y c s ( x ) . Such a position exists because otherwise y y y has the same colored certificate as x x x .
First, observe two persistence properties. Before position s s s , all inserted certificate values equal ν x \nu_x ν x on both inputs. Because no raw element of x x x has value ν x \nu_x ν x , these insertions do not change any raw assignment on x x x . On y y y they can only uncover entries with value ν x \nu_x ν x . Hence they cannot repair an existing disagreement on a raw element. At position s s s , the inserted c s ( x ) c_s(x) c s ( x ) is revealed on x x x , along its chosen branch, with value ν x \nu_x ν x . Its assignment on y y y is either hidden or a different value. This disagreement persists at every later certificate stage. The same argument applies if a disagreement is first created by uncovering a raw entry with value ν x \nu_x ν x .
If x ̸ ∼ R y x\not\sim_R y x ∼ R y , there is a first setup query after which the assignments disagree. The queried true values differ, and at least one is revealed. Thus this query contributes p x p_x p x by (145) or (146) . Every earlier query either has equal true values or hides both values, and contributes zero. Every later setup source disagrees; the persistence properties show that every certificate source also disagrees. This proves the required contribution in this case.
Now suppose x ∼ R y x \sim_R y x ∼ R y . All setup contributions vanish. Certificate stages earlier than s s s query equal true values, and stages later than s s s have disagreeing source assignments. It remains to sum the stage-s s s contributions over its branch prefixes.
For 0 ≤ h < s 0 \le h < s 0 ≤ h < s and fixed nonempty J 1 , … , J h J_1,\ldots,J_h J 1 , … , J h , let
S h = R [ J 1 ← c 1 ( x ) , … , J h ← c h ( x ) ] , S_h = R[J_1 \leftarrow c_1(x),\ldots,J_h \leftarrow c_h(x)], S h = R [ J 1 ← c 1 ( x ) , … , J h ← c h ( x )] ,
and let C ( S h ) \mathcal{C}(S_h) C ( S h ) be the sum of stage-s s s entries obtained by all continuations J h + 1 , … , J s − 1 J_{h+1},\ldots,J_{s-1} J h + 1 , … , J s − 1 . We claim
C ( S h ) = { ( − 1 ) h + ∑ i ≤ h ∣ J i ∣ p x , x ∼ S h y , 0 , x ̸ ∼ S h y . (149) \mathcal{C}(S_h) =
\begin{cases}
(-1)^{h+\sum_{i\le h}|J_i|}p_x, & x \sim_{S_h} y,\\
0, & x \not\sim_{S_h} y.
\end{cases} \tag*{(149)} C ( S h ) = { ( − 1 ) h + ∑ i ≤ h ∣ J i ∣ p x , 0 , x ∼ S h y , x ∼ S h y . (149)
For h = s − 1 h=s-1 h = s − 1 , this is exactly (147) . We prove the remaining cases by downward induction on h h h . If the current source disagrees, persistence gives zero for every continuation.
Suppose instead that the current source agrees. A raw input t t t with y t = ν x y_t=\nu_x y t = ν x will be called a fault. There are at most K K K such inputs, because the multiplicity bound holds for y y y . For each d ∈ [ b ] d\in[b] d ∈ [ b ] , consider the row group at the next raw level
⋃ ∅ ≠ J ⊆ [ b ] R h + 2 ( μ ( J 1 ) , … , μ ( J h ) , d , J ) . (150) \bigcup_{\varnothing\ne J\subseteq[b]} R_{h+2}\bigl(\mu(J_1),\ldots,\mu(J_h),d,J\bigr). \tag*{(150)} ∅ = J ⊆ [ b ] ⋃ R h + 2 ( μ ( J 1 ) , … , μ ( J h ) , d , J ) . (150)
Let I ⊆ [ b ] I\subseteq[b] I ⊆ [ b ] be the indices of groups containing no fault. If h + 2 = r h+2=r h + 2 = r , there is no such raw level, and we set I = [ b ] I=[b] I = [ b ] . Otherwise the groups are disjoint, so ∣ I ∣ ≥ b − K ≥ 1 |I|\ge b-K\ge1 ∣ I ∣ ≥ b − K ≥ 1 .
Adjoining c h + 1 ( x ) c_{h+1}(x) c h + 1 ( x ) in the cell indexed by J h + 1 J_{h+1} J h + 1 can newly reveal ν x \nu_x ν x only in groups with indices in J h + 1 J_{h+1} J h + 1 . Every fault there was hidden before the insertion, since the source assignments agree and the corresponding x x x value is not ν x \nu_x ν x . Thus the updated assignments agree exactly when ∅ ≠ J h + 1 ⊆ I \varnothing\ne J_{h+1}\subseteq I ∅ = J h + 1 ⊆ I . If there are no faults in these immediate child groups, no raw value changes its visibility, so there is no subsequent cascade to higher raw levels. This also explains the stated boundary convention when no further raw level exists.
Applying the induction hypothesis to each child gives
C ( S h ) = ∑ ∅ ≠ J ⊆ I ( − 1 ) h + 1 + ∑ i ≤ h ∣ J i ∣ + ∣ J ∣ p x = ( − 1 ) h + ∑ i ≤ h ∣ J i ∣ p x , \begin{aligned}
\mathcal{C}(S_h)=\sum_{\varnothing\ne J\subseteq I}(-1)^{h+1+\sum_{i\le h}|J_i|+|J|}p_x \\
&=(-1)^{h+\sum_{i\le h}|J_i|}p_x,
\end{aligned} C ( S h ) = ∅ = J ⊆ I ∑ ( − 1 ) h + 1 + ∑ i ≤ h ∣ J i ∣ + ∣ J ∣ p x = ( − 1 ) h + ∑ i ≤ h ∣ J i ∣ p x ,
because I I I is nonempty and ∑ ∅ ≠ J ⊆ I ( − 1 ) ∣ J ∣ = − 1 \sum_{\varnothing\ne J\subseteq I}(-1)^{|J|}=-1 ∑ ∅ = J ⊆ I ( − 1 ) ∣ J ∣ = − 1 . This proves (149) . Taking h = 0 h=0 h = 0 yields the desired contribution p x p_x p x .
Finally, for fixed x x x , no taken index is associated with two different consistent families R R R . Summing the per-family identity over R x \mathcal{R}_x R x therefore yields N R ( x ) p x = 1 N_{\mathcal{R}}(x)p_x=1 N R ( x ) p x = 1 , proving (144) exactly.
Cost. The next proposition bounds the diagonal sum of these explicit matrices. Combined with (144) and positive semidefiniteness, it proves Lemma 8.5 by the dual adversary characterization. The positive-dependent factor γ x \gamma_x γ x is essential to exact normalization; its boundedness ensures that it changes the diagonal estimate by only a constant depending on K , r K,r K , r .
Diagonal costs of the colored construction
We keep the domain sizes of the roles separate when estimating the positive-semidefinite matrices of the preceding construction. In particular, the counting argument below uses complete colored prefixes, rather than the number of uncolored pairs or of shorter prefixes.
Fix constants K ≥ r ≥ 2 K \ge r \ge2 K ≥ r ≥ 2 . Let U 1 , … , U r U_1,\ldots,U_r U 1 , … , U r be disjoint public sets, with ∣ U i ∣ = v i |U_i| = v_i ∣ U i ∣ = v i , and assume that every equality class has size at most K K K . For an input z z z , write
C j ( z ) = ∣ { ( u 1 , … , u j ) ∈ U 1 × ⋯ × U j : z u 1 = ⋯ = z u j } ∣ , 2 ≤ j < r . C_j(z) = \left|\left\{(u_1,\ldots,u_j) \in U_1 \times\cdots\times U_j : z_{u_1} = \cdots= z_{u_j}\right\}\right|,\qquad2 \le j < r. C j ( z ) = { ( u 1 , … , u j ) ∈ U 1 × ⋯ × U j : z u 1 = ⋯ = z u j } , 2 ≤ j < r .
Let C j ≥ 1 C_j \ge1 C j ≥ 1 be specified upper bounds on C j ( z ) C_j(z) C j ( z ) throughout the promise, for both positive and negative inputs. Positive inputs have at least one colorful target tuple, and negative inputs have none. For each positive input, fix one such tuple deterministically and use its equality class as the chosen target class in the construction.
Put b = K + 1 b = K + 1 b = K + 1 . At level i < r i < r i < r , the fault-tolerant construction has L i = b i − 1 ( 2 b − 1 ) = O K , r ( 1 ) L_i = b^{i-1}(2^b - 1) = O_{K,r}(1) L i = b i − 1 ( 2 b − 1 ) = O K , r ( 1 ) labeled raw-set copies, each of size a i a_i a i . All copies at one level are disjoint subsets of U i U_i U i . In the following diagonal estimates, the a i a_i a i are positive integers, and
v i ≥ 4 K , L i a i ≤ v i / 4 ( 1 ≤ i < r ) . (151) v_i \ge4K,\qquad L_i a_i \le v_i/4\qquad(1 \le i < r). \tag*{(151)} v i ≥ 4 K , L i a i ≤ v i /4 ( 1 ≤ i < r ) . (151)
Copies at different levels are automatically disjoint because their roles are disjoint. The constants below may depend on K , r K,r K , r and hence on the numbers of fault copies, but not on v i , a i , C j v_i,a_i,C_j v i , a i , C j .
Proposition 8.6 (Diagonal costs of the colored construction). Under the preceding assumptions, the feasible matrices of the colored construction have maximum diagonal objective at most
O K , r ( a 1 + ∑ j = 2 r − 1 a j C j ∏ i < j a i ∏ i ≤ j v i + ∑ j = 1 r ∏ i ≤ j v i ∏ i < j a i ) . (152) O_{K,r}\left(a_1+\sum_{j=2}^{r-1}a_j\sqrt{\frac{C_j\prod_{i<j}a_i}{\prod_{i\le j}v_i}}+\sum_{j=1}^{r}\sqrt{\frac{\prod_{i\le j}v_i}{\prod_{i<j}a_i}}\right). \tag*{(152)} O K , r ( a 1 + j = 2 ∑ r − 1 a j ∏ i ≤ j v i C j ∏ i < j a i + j = 1 ∑ r ∏ i < j a i ∏ i ≤ j v i ) . (152)
Proof. A common normalization preserves positivity. Let R \mathcal{R} R be the set of all raw configurations. Its cardinality and the corresponding reference weight are
N R = ∏ i < r ( v i ) L i a i ( a i ! ) L i , q 0 = N R − 1 . N_{\mathcal{R}}=\prod_{i<r}\frac{(v_i)_{L_i a_i}}{(a_i!)^{L_i}},\qquad q_0=N_{\mathcal{R}}^{-1}. N R = i < r ∏ ( a i ! ) L i ( v i ) L i a i , q 0 = N R − 1 .
For a positive input x x x , let Z x Z_x Z x be its chosen target equality class and put d i ( x ) = ∣ Z x ∩ U i ∣ d_i(x)=|Z_x\cap U_i| d i ( x ) = ∣ Z x ∩ U i ∣ . We use only configurations avoiding the whole class Z x Z_x Z x , not merely the selected target tuple. Writing R x \mathcal{R}_x R x for these configurations, we have
N R ( x ) : = ∣ R x ∣ = ∏ i < r ( v i − d i ( x ) ) L i a i ( a i ! ) L i , γ x = N R N R ( x ) . N_{\mathcal{R}}(x):=|\mathcal{R}_x|=\prod_{i<r}\frac{(v_i-d_i(x))_{L_i a_i}}{(a_i!)^{L_i}},\qquad\gamma_x=\frac{N_{\mathcal{R}}}{N_{\mathcal{R}}(x)}. N R ( x ) := ∣ R x ∣ = i < r ∏ ( a i ! ) L i ( v i − d i ( x ) ) L i a i , γ x = N R ( x ) N R .
The normalization is uniformly bounded. Indeed,
1 ≤ γ x = ∏ i < r ( v i ) d i ( x ) ( v i − L i a i ) d i ( x ) ≤ 2 ∑ i < r d i ( x ) ≤ 2 K . (153) 1\le\gamma_x=\prod_{i<r}\frac{(v_i)_{d_i(x)}}{(v_i-L_i a_i)_{d_i(x)}}\le2^{\sum_{i<r}d_i(x)}\le2^K. \tag*{(153)} 1 ≤ γ x = i < r ∏ ( v i − L i a i ) d i ( x ) ( v i ) d i ( x ) ≤ 2 ∑ i < r d i ( x ) ≤ 2 K . (153)
Here each denominator factor is at least v i / 2 v_i/2 v i /2 , by (151) , and the target class has at most K K K elements. Use the local Gram factors in (145) , (146) , and (147) , with common prefactor q 0 q_0 q 0 . Multiply every positive-input coordinate of every Gram vector by γ x \gamma_x γ x ; leave negative-input coordinates unchanged. Equivalently, replace each local matrix X X X by D X D DXD D X D , where D D D is diagonal with D x , x = γ x D_{x,x}=\gamma_x D x , x = γ x on positive inputs and D y , y = 1 D_{y,y}=1 D y , y = 1 on negative inputs. The modified matrices remain positive semidefinite. A positive–negative entry associated with one admissible configuration is now weighted by
q 0 γ x = N R ( x ) − 1 . q_0\gamma_x=N_{\mathcal{R}}(x)^{-1}. q 0 γ x = N R ( x ) − 1 .
Thus the feasibility calculation, performed separately for each R ∈ R x R\in\mathcal{R}_x R ∈ R x , has exactly the required normalization. On the diagonal, the positive prefactor is q 0 γ x 2 q_0\gamma_x^2 q 0 γ x 2 and the negative prefactor is q 0 q_0 q 0 .
Counting uncovered raw elements. Let V i ( R ) ⊆ U i V_i(R)\subseteq U_i V i ( R ) ⊆ U i be the union of the L i L_i L i raw copies at level i i i , so that ∣ V i ( R ) ∣ = L i a i |V_i(R)|=L_i a_i ∣ V i ( R ) ∣ = L i a i . Let H j ( z , R ) H_j(z,R) H j ( z , R ) denote the number of elements uncovered during setup stage I.j j j , summed over all copies at that level. Every such element is the final element of a chain of equal values containing one raw element from each of the roles 1 , … , j 1,\ldots,j 1 , … , j . This remains true with the fault-copy dependencies: those dependencies restrict which previous copies can supply a match, and cannot create an uncovered element without a complete colored prefix. Consequently,
H j ( z , R ) ≤ ∑ ( u 1 , … , u j ) ∈ U 1 × ⋯ × U j z u 1 = ⋯ = z u j ∏ i ≤ j 1 { u i ∈ V i ( R ) } . H_j(z,R)\leq\sum_{\substack{(u_1,\ldots,u_j)\in U_1\times\cdots\times U_j\\ z_{u_1}=\cdots=z_{u_j}}}\prod_{i\leq j}\mathbf{1}_{\{u_i\in V_i(R)\}}. H j ( z , R ) ≤ ( u 1 , … , u j ) ∈ U 1 × ⋯ × U j z u 1 = ⋯ = z u j ∑ i ≤ j ∏ 1 { u i ∈ V i ( R )} .
For uniform R ∈ R R\in\mathcal{R} R ∈ R , the unions V i ( R ) V_i(R) V i ( R ) are independent across roles and are uniform subsets of the indicated sizes. Define
A j = C j ∏ i ≤ j a i v i . (154) A_j=C_j\prod_{i\leq j}\frac{a_i}{v_i}. \tag*{(154)} A j = C j i ≤ j ∏ v i a i . (154)
It follows that, uniformly over the promised inputs,
E R ∈ R H j ( z , R ) ≤ C j ∏ i ≤ j L i a i v i = O K , r ( A j ) . (155) \mathbb{E}_{R\in\mathcal{R}}H_j(z,R)\leq C_j\prod_{i\leq j}\frac{L_i a_i}{v_i}=O_{K,r}(A_j). \tag*{(155)} E R ∈ R H j ( z , R ) ≤ C j i ≤ j ∏ v i L i a i = O K , r ( A j ) . (155)
Conditioning on whole-class avoidance changes this upper bound by at most γ x ≤ 2 K \gamma_x\leq2^K γ x ≤ 2 K . Hence E R ∈ R x H j ( x , R ) = O K , r ( A j ) \mathbb{E}_{R\in\mathcal{R}_x}H_j(x,R)=O_{K,r}(A_j) E R ∈ R x H j ( x , R ) = O K , r ( A j ) as well. All raw setup stages precede certificate insertion, so no inserted target element contributes an additional shorter-prefix term to (155) .
Setup-stage diagonal bounds. For stage I.j j j , with 2 ≤ j < r 2\leq j<r 2 ≤ j < r , let w 0 , j , w 1 , j > 0 w_{0,j},w_{1,j}>0 w 0 , j , w 1 , j > 0 be its two weights. Equation (146) gives the following diagonal contributions, before normalization: on a negative input the contribution is w 1 , j w_{1,j} w 1 , j for every loaded element, plus w 0 , j w_{0,j} w 0 , j when that element is uncovered; on a positive input it is 1 / w 1 , j 1/w_{1,j} 1/ w 1 , j for an uncovered element and 1 / w 0 , j 1/w_{0,j} 1/ w 0 , j for a covered element. Each input satisfies exactly one assignment at a fixed arc, so summing over assignment labels introduces no factor depending on the output alphabet. Summing over configurations and using (153) –(155) therefore gives
D 0 , j ≤ O K , r ( A j w 0 , j + a j w 1 , j ) , D_{0,j}\leq O_{K,r}(A_jw_{0,j}+a_jw_{1,j}), D 0 , j ≤ O K , r ( A j w 0 , j + a j w 1 , j ) ,
D 1 , j ≤ O K , r ( A j / w 1 , j + a j / w 0 , j ) . D_{1,j}\leq O_{K,r}(A_j/w_{1,j}+a_j/w_{0,j}). D 1 , j ≤ O K , r ( A j / w 1 , j + a j / w 0 , j ) .
For example, the positive sum is γ x \gamma_x γ x times the conditional expectation over R x \mathcal{R}_x R x of the indicated reciprocal-weight expression; the bounded factor γ x \gamma_x γ x has already been included above. Taking
w 0 , j = a j / A j , w 1 , j = A j / a j w_{0,j}=\sqrt{a_j/A_j},\qquad w_{1,j}=\sqrt{A_j/a_j} w 0 , j = a j / A j , w 1 , j = A j / a j
bounds both diagonal costs by
O K , r ( a j A j ) = O K , r ( a j C j ∏ i < j a i ∏ i ≤ j v i ) . (156) O_{K,r}\left(\sqrt{a_jA_j}\right)=O_{K,r}\left(a_j\sqrt{\frac{C_j\prod_{i<j}a_i}{\prod_{i\leq j}v_i}}\right). \tag*{(156)} O K , r ( a j A j ) = O K , r ( a j ∏ i ≤ j v i C j ∏ i < j a i ) . (156)
Stage I.1 has L 1 a 1 = O K , r ( a 1 ) L_1a_1 = O_{K,r}(a_1) L 1 a 1 = O K , r ( a 1 ) loads per configuration; setting its single weight to 1 gives cost O K , r ( a 1 ) O_{K,r}(a_1) O K , r ( a 1 ) .
Certificate-stage diagonal bounds. At stage II.j j j , fix one of the O K , r ( 1 ) O_{K,r}(1) O K , r ( 1 ) fault-branch patterns. For each i < j i<j i < j , one specified raw copy at level i i i has been enlarged from a i a_i a i to a i + 1 a_i+1 a i + 1 . The ratio between the number of enlarged level-i i i configurations and the number of raw level-i i i configurations is exactly
( v i ) L i a i + 1 / [ ( a i + 1 ) ! ( a i ! ) L i − 1 ] ( v i ) L i a i / ( a i ! ) L i = v i − L i a i a i + 1 ≤ v i a i . \frac{(v_i)_{L_i}a_{i+1}/[(a_i+1)!(a_i!)^{L_i-1}]}{(v_i)_{L_i}a_i/(a_i!)^{L_i}}=\frac{v_i-L_i a_i}{a_i+1}\leq\frac{v_i}{a_i}. ( v i ) L i a i / ( a i ! ) L i ( v i ) L i a i + 1 / [( a i + 1 )! ( a i ! ) L i − 1 ] = a i + 1 v i − L i a i ≤ a i v i .
There are at most v j v_j v j choices for the element loaded at stage II.j j j . Summing over the constant number of branch patterns, the total number of its arcs is therefore at most
O K , r ( N R v j ∏ i < j v i a i ) . O_{K,r}\left(N_R v_j\prod_{i<j}\frac{v_i}{a_i}\right). O K , r ( N R v j i < j ∏ a i v i ) .
For a fixed positive input, each R ∈ R x R\in\mathcal R_x R ∈ R x contributes only O K , r ( 1 ) O_{K,r}(1) O K , r ( 1 ) taken arcs at this stage. The signs in (147) do not affect their diagonals. With stage weight w j w_j w j , the two diagonal costs consequently satisfy
D 0 , j ≤ O K , r ( w j 2 v j ∏ i < j v i a i ) , D 1 , j ≤ O K , r ( w j − 1 ) , D_{0,j}\leq O_{K,r}\left(w_j^2v_j\prod_{i<j}\frac{v_i}{a_i}\right),\qquad D_{1,j}\leq O_{K,r}(w_j^{-1}), D 0 , j ≤ O K , r ( w j 2 v j i < j ∏ a i v i ) , D 1 , j ≤ O K , r ( w j − 1 ) ,
where the positive estimate uses q 0 2 γ x 2 N R ( x ) = γ x = O K ( 1 ) q_0^2\gamma_x^2N_R(x)=\gamma_x=O_K(1) q 0 2 γ x 2 N R ( x ) = γ x = O K ( 1 ) . Set
S j = ∏ i ≤ j v i ∏ i ≤ j a i , w j = S j − 1 / 2 . S_j=\frac{\prod_{i\leq j}v_i}{\prod_{i\leq j}a_i},\qquad w_j=S_j^{-1/2}. S j = ∏ i ≤ j a i ∏ i ≤ j v i , w j = S j − 1/2 .
The cost of certificate stage II.j j j is then O K , r ( S j ) O_{K,r}(\sqrt{S_j}) O K , r ( S j ) . This calculation applies also to j = r j=r j = r , where there is no raw level-r r r set.
Summing the maximum diagonal costs of all stages proves (152) . □
Thus the full colored-prefix bounds C j C_j C j control every setup stage, including all its fault copies. No additional count of shorter prefixes or of uncolored collisions is needed in the diagonal estimate. The fault-copy constants must depend on the global multiplicity bound K K K : a negative colorful instance can have more than r − 1 r-1 r − 1 occurrences of one value even though it lacks a colorful r r r -tuple.
The statement was made for integer parameters to expose the exact configuration counts. Real parameters at least 1 can first be scaled by a sufficiently small constant depending on K , r K,r K , r and then rounded up. Provided the role sizes tend to infinity and the unrounded parameters are at most the corresponding role sizes, this ensures (151) and changes every term in (152) by only a constant factor.
Ordered rolewise restriction above the square-root scale
We now assume Q ≥ N Q\geq\sqrt{N} Q ≥ N . The colored decision bound allows us to retain different fractions of the input in the different certificate roles. Shrinking these roles in order preserves the preparation costs and controls every certificate-loading stage.
Lemma 8.7 (Raw parameters in the larger-query regime). For a sufficiently large constant C = C ( K , r ) C=C(K,r) C = C ( K , r ) , there are positive integer parameters a i a_i a i and R = C Q R=CQ R = C Q such that
a i ≤ c K , r ′ N , ∏ i < r a i ≍ K , r N r M R 2 , (157) a_i\le c'_{K,r}N,\qquad\prod_{i<r}a_i\asymp_{K,r}\frac{N^r}{MR^2}, \tag*{(157)} a i ≤ c K , r ′ N , i < r ∏ a i ≍ K , r M R 2 N r , (157) a 1 + ∑ j = 2 r − 1 a j B j ∏ i < j a i N j / 2 ≤ O K , r ( Q ) , N j M ∏ i < j a i ≤ O K , r ( Q ) ( 1 ≤ j ≤ r ) . (158) a_1+\sum_{j=2}^{r-1}\frac{a_j\sqrt{B_j\prod_{i<j}a_i}}{N^{j/2}}\le O_{K,r}(Q),\qquad\sqrt{\frac{N^j}{M\prod_{i<j}a_i}}\le O_{K,r}(Q)\quad(1\le j\le r). \tag*{(158)} a 1 + j = 2 ∑ r − 1 N j /2 a j B j ∏ i < j a i ≤ O K , r ( Q ) , M ∏ i < j a i N j ≤ O K , r ( Q ) ( 1 ≤ j ≤ r ) . (158) The constant c K , r ′ c'_{K,r} c K , r ′ can be made arbitrarily small by increasing C C C . Every a i a_i a i tends to infinity as a fixed positive power of N N N .
Proof. The recurrence in the proof of Lemma 8.2 does not use Q ≤ N Q\le\sqrt{N} Q ≤ N until its final assertion about expected counts. It therefore gives real parameters Q ≤ a ˉ i ≤ N Q\le\bar a_i\le N Q ≤ a ˉ i ≤ N with a ˉ 1 = Q \bar a_1=Q a ˉ 1 = Q , preparation cost O K , r ( Q ) O_{K,r}(Q) O K , r ( Q ) , and
A : = ∏ i < r a ˉ i ≥ N r M Q 2 . A:=\prod_{i<r}\bar a_i\ge\frac{N^r}{MQ^2}. A := i < r ∏ a ˉ i ≥ M Q 2 N r .
Put
θ = ( N r M R 2 A ) 1 / ( r − 1 ) , a i ∘ = θ a ˉ i . \theta=\left(\frac{N^r}{MR^2A}\right)^{1/(r-1)},\qquad a_i^\circ=\theta\bar a_i. θ = ( M R 2 A N r ) 1/ ( r − 1 ) , a i ∘ = θ a ˉ i .
Then θ ≤ C − 2 / ( r − 1 ) \theta\le C^{-2/(r-1)} θ ≤ C − 2/ ( r − 1 ) , the product of the a i ∘ a_i^\circ a i ∘ is exactly N r / ( M R 2 ) N^r/(MR^2) N r / ( M R 2 ) , and every preparation term decreases. The final certificate term equals R R R . All preceding certificate terms are no larger, because a i ∘ ≤ N a_i^\circ\le N a i ∘ ≤ N .
For completeness, these real parameters are large enough to round. The inequalities B i ≤ N B_i\le N B i ≤ N and Q ≥ N Q\ge\sqrt{N} Q ≥ N imply
M ≤ N 1 − 2 1 − r . (159) M\le N^{1-2^{1-r}}. \tag*{(159)} M ≤ N 1 − 2 1 − r . (159)
Indeed, take an index j j j attaining the maximum in (123) , bound its numerator by N 3 ⋅ 2 j − 2 − 1 N^{3\cdot2^{j-2}-1} N 3 ⋅ 2 j − 2 − 1 , and compare with N ( 2 j − 1 ) / 2 N^{(2^j-1)/2} N ( 2 j − 1 ) /2 . For r ≥ 3 r\ge3 r ≥ 3 , using A ≤ N r − 1 A\le N^{r-1} A ≤ N r − 1 gives
a i ∘ ≥ C − 2 / ( r − 1 ) ( N / M ) 1 / ( r − 1 ) Q ( r − 3 ) / ( r − 1 ) . a_i^\circ\ge C^{-2/(r-1)}(N/M)^{1/(r-1)}Q^{(r-3)/(r-1)}. a i ∘ ≥ C − 2/ ( r − 1 ) ( N / M ) 1/ ( r − 1 ) Q ( r − 3 ) / ( r − 1 ) .
This grows polynomially by (159) . For r = 2 r=2 r = 2 , directly a 1 ∘ = Q / C 2 a_1^\circ=Q/C^2 a 1 ∘ = Q / C 2 . Take a i = ⌊ a i ∘ ⌋ a_i=\lfloor a_i^\circ\rfloor a i = ⌊ a i ∘ ⌋ . Their product changes by 1 + o ( 1 ) 1+o(1) 1 + o ( 1 ) , setup costs decrease, and all certificate costs change by 1 + o ( 1 ) 1+o(1) 1 + o ( 1 ) . This proves the assertions.
Fix ρ 0 = 1 / ( 2 r ) \rho_0=1/(2r) ρ 0 = 1/ ( 2 r ) and then choose a sufficiently small constant 0 < κ < ρ 0 r 0<\kappa<\rho_0^r 0 < κ < ρ 0 r depending only on K , r K,r K , r . Choose a large constant D = D ( K , r ) D=D(K,r) D = D ( K , r ) , large enough to accommodate all raw-set copies in the colored construction, and choose the constant C C C in Lemma 8.7 still larger if necessary. Starting with every role probability equal to ρ 0 \rho_0 ρ 0 , decrease the first role probability to D a 1 / N Da_1/N D a 1 / N , then the second to D a 2 / N Da_2/N D a 2 / N , and so on, stopping as soon as
∏ i = 1 r ρ i = κ / M . (160) \prod_{i=1}^{r}\rho_i=\kappa/M. \tag*{(160)} i = 1 ∏ r ρ i = κ / M . (160)
The last role need not be decreased. In fact, the product with the first r − 1 r-1 r − 1 roles at their caps is
ρ 0 D r − 1 ∏ i < r a i N r − 1 ≤ ρ 0 D r − 1 N M C 2 Q 2 ≤ κ M . \rho_0D^{r-1}\frac{\prod_{i<r}a_i}{N^{r-1}}\le\frac{\rho_0D^{r-1}N}{MC^2Q^2}\le\frac{\kappa}{M}. ρ 0 D r − 1 N r − 1 ∏ i < r a i ≤ M C 2 Q 2 ρ 0 D r − 1 N ≤ M κ .
for sufficiently large C C C . All caps are below ρ 0 \rho_0 ρ 0 by the same choice. Thus the stopping point exists.
Assign each original input independently to role i i i with probability ρ i \rho_i ρ i , and discard it otherwise. The total retained probability is at most 1 / 2 1/2 1/2 . Write U i U_i U i for the resulting public role and v i = ∣ U i ∣ v_i = |U_i| v i = ∣ U i ∣ . The role assignment is internal randomization of the algorithm; it reveals no equality class.
The next statement supplies both a target and the prefix bounds used by the colored decision algorithm. The stronger unique-target event is convenient for the reduction, although the colored algorithm itself accepts any positive number of targets.
Lemma 8.8 (Isolation and prefix populations). With probability Ω K , r ( 1 ) \Omega_{K,r}(1) Ω K , r ( 1 ) , the roles contain exactly one colored r r r -tuple, their sizes satisfy N ρ i / 2 ≤ v i ≤ 2 N ρ i N\rho_i/2 \le v_i \le2N\rho_i N ρ i /2 ≤ v i ≤ 2 N ρ i , and simultaneously
C j ( f ) ≤ A K , r B j ∏ i ≤ j ρ i ( 2 ≤ j < r ) . (161) C_j(f) \le A_{K,r}B_j \prod_{i\le j}\rho_i \qquad(2 \le j < r). \tag*{(161)} C j ( f ) ≤ A K , r B j i ≤ j ∏ ρ i ( 2 ≤ j < r ) . (161) Here C j ( f ) C_j(f) C j ( f ) counts equal-value tuples in U 1 × ⋯ × U j U_1 \times\cdots\times U_j U 1 × ⋯ × U j .
Proof. For an original class of size d ≥ r d \ge r d ≥ r , the probability that it meets every role is at most ( d r ) ∏ i ρ i \binom{d}{r}\prod_i \rho_i ( r d ) ∏ i ρ i . Its probability of having exactly one element in every role, with all others discarded, is
( d r ) ∏ i ρ i ( 1 − ∑ i ρ i ) d − r = Θ K , r ( ∏ i ρ i ) . \binom{d}{r}\prod_i \rho_i\left(1-\sum_i\rho_i\right)^{d-r}
= \Theta_{K,r}\left(\prod_i\rho_i\right). ( r d ) i ∏ ρ i ( 1 − i ∑ ρ i ) d − r = Θ K , r ( i ∏ ρ i ) . The role assignments in different classes are independent. By taking κ \kappa κ sufficiently small, the expected number of classes meeting every role is at most 1 / 2 1/2 1/2 . For each choice of a unique surviving class, multiply its exact-one-per-role probability by the probability that no other class meets every role. The latter is at least 1 / 2 1/2 1/2 by a union bound and independence. These events are disjoint for different choices of the surviving class, so their sum has probability Ω K , r ( M ∏ i ρ i ) = Ω K , r ( 1 ) \Omega_{K,r}(M\prod_i\rho_i) = \Omega_{K,r}(1) Ω K , r ( M ∏ i ρ i ) = Ω K , r ( 1 ) .
Linearity of expectation gives
E C j ( f ) = ( ∑ d ≥ j ( d j ) n d ) ∏ i ≤ j ρ i ≤ ( K ) j B j ∏ i ≤ j ρ i . \mathbb{E}C_j(f) = \left(\sum_{d\ge j}\binom{d}{j}n_d\right)\prod_{i\le j}\rho_i \le(K)_jB_j\prod_{i\le j}\rho_i. E C j ( f ) = d ≥ j ∑ ( j d ) n d i ≤ j ∏ ρ i ≤ ( K ) j B j i ≤ j ∏ ρ i . Markov’s inequality and a union bound, with a sufficiently large constant A K , r A_{K,r} A K , r , make the probability of violating any prefix bound smaller than half the isolation probability. Finally ρ i ≥ ∏ u ρ u = κ / M \rho_i \ge\prod_u\rho_u = \kappa/M ρ i ≥ ∏ u ρ u = κ / M , so (8.24) implies N ρ i ≥ κ N 1 − r N\rho_i \ge\kappa N^{1-r} N ρ i ≥ κ N 1 − r . Chernoff bounds give all the stated role-size estimates with probability 1 − o ( 1 ) 1-o(1) 1 − o ( 1 ) . Intersecting the events proves the lemma.
Before applying the decision algorithm, we establish a conversion that exploits the isolated colorful tuple. The role domains remain fixed, so every virtual input is covered by the same decision circuit.
Lemma 8.9 (Coherent search for an isolated colorful target). Fix disjoint public roles U 1 , … , U r ⊆ [ N ] U_1,\ldots,U_r \subseteq[N] U 1 , … , U r ⊆ [ N ] and public bounds C j C_j C j on their colored prefix counts. Suppose a bounded-error decision algorithm uses at most D d e c D_{\mathrm{dec}} D dec queries on every input in the promise of Lemma 8.5, over any prescribed finite alphabet. On any such input with exactly one colorful r r r -tuple, its first-role input can be found with constant success probability using O ( D d e c ) O(D_{\mathrm{dec}}) O ( D dec ) original function queries. The algorithm has a fixed query budget on all inputs, including those outside the promise.
Proof. Let x ∗ x_\ast x ∗ be the first-role input of the unique tuple. Put m = ⌈ log 2 N ⌉ m=\lceil\log_2 N\rceil m = ⌈ log 2 N ⌉ and encode x ∈ [ N ] x\in[N] x ∈ [ N ] by the m m m -bit representation enc ( x ) \operatorname{enc}(x) enc ( x ) of x − 1 x-1 x − 1 . For v ∈ { 0 , 1 } m v\in\{0,1\}^{m} v ∈ { 0 , 1 } m , define a virtual input on the unchanged role domains by
f v ( x ) = { ( 1 , x ) , x ∈ U 1 and v ⋅ enc ( x ) = 0 ( m o d 2 ) , ( 0 , f ( x ) ) , otherwise . (162) f_v(x)=
\begin{cases}
(1,x), & x\in U_1\text{ and }v\cdot\operatorname{enc}(x)=0\pmod{2},\\
(0,f(x)), & \text{otherwise}.
\end{cases} \tag*{(162)} f v ( x ) = { ( 1 , x ) , ( 0 , f ( x )) , x ∈ U 1 and v ⋅ enc ( x ) = 0 ( mod 2 ) , otherwise . (162)
The tagged alphabet is the disjoint union ( { 0 } × Z q ) ⊔ ( { 1 } × [ N ] ) \left(\{0\}\times\mathbb{Z}_q\right)\sqcup\left(\{1\}\times[N]\right) ( { 0 } × Z q ) ⊔ ( { 1 } × [ N ] ) . Every dummy value occurs once. Masking therefore creates no new colored collision, decreases every colored prefix count, and preserves the maximum class-size bound K K K . Since the original colorful tuple is unique, the decision answer on f v f_v f v is
b ( v ) = v ⋅ s ( m o d 2 ) , s = enc ( x ∗ ) . (163) b(v)=v\cdot s\pmod{2},\qquad s=\operatorname{enc}(x_\ast). \tag*{(163)} b ( v ) = v ⋅ s ( mod 2 ) , s = enc ( x ∗ ) . (163)
Choose one decision circuit A A A for the full fixed promise consisting of all inputs on these roles with class sizes at most K K K and prefix counts at most the supplied C j C_j C j . This circuit works for every mask v v v ; its non-query operations do not depend on v v v . Amplify its error to a fixed constant δ < 1 / 100 \delta<1/100 δ < 1/100 , using only a constant number of repetitions, and purify its workspace and measurements. Denote the resulting unitary circuit with virtual oracle f v f_v f v by A v A_v A v .
A virtual query can be implemented coherently with a constant number of queries to f f f : compute f ( x ) f(x) f ( x ) into an ancillary register, compute the public mask and the tagged value in (8.27), add that value to the virtual response register, and uncompute the ancilla. The register v v v is preserved. The inverse circuit has the same constant-overhead simulation. Thus the controlled family A v A_v A v is available without assuming controlled access to an unknown black-box unitary beyond the stated function-query model.
Let Z a n s Z_{\mathrm{ans}} Z ans multiply answer 1 1 1 by − 1 -1 − 1 and leave answer 0 0 0 unchanged, and set U v = A v † Z a n s A v U_v=A_v^\dagger Z_{\mathrm{ans}}A_v U v = A v † Z ans A v . The component of A v ∣ 0 ⟩ A_v|0\rangle A v ∣0 ⟩ with an incorrect answer has norm at most δ \sqrt{\delta} δ . Consequently
∥ U v ∣ 0 ⟩ − ( − 1 ) b ( v ) ∣ 0 ⟩ ∥ ≤ 2 δ . (164) \left\|U_v|0\rangle-(-1)^{b(v)}|0\rangle\right\|\le2\sqrt{\delta}. \tag*{(164)} U v ∣0 ⟩ − ( − 1 ) b ( v ) ∣0 ⟩ ≤ 2 δ . (164)
Apply this compute–phase–uncompute circuit to a uniform superposition of v v v . Orthogonality of the preserved control-register sectors gives
∥ 2 − m / 2 ∑ v ∣ v ⟩ U v ∣ 0 ⟩ − 2 − m / 2 ∑ v ( − 1 ) v ⋅ s ∣ v ⟩ ∣ 0 ⟩ ∥ 2 ≤ 4 δ . (165) \left\|2^{-m/2}\sum_v|v\rangle U_v|0\rangle-2^{-m/2}\sum_v(-1)^{v\cdot s}|v\rangle|0\rangle\right\|^2\le4\delta. \tag*{(165)} 2 − m /2 v ∑ ∣ v ⟩ U v ∣0 ⟩ − 2 − m /2 v ∑ ( − 1 ) v ⋅ s ∣ v ⟩ ∣0 ⟩ 2 ≤ 4 δ . (165)
In particular, the number 2 m 2^m 2 m of masks introduces no error factor. The Bernstein–Vazirani Hadamard transform [13 ] on the control register maps the second state exactly to ∣ s ⟩ ∣ 0 ⟩ |s\rangle|0\rangle ∣ s ⟩ ∣0 ⟩ . The measured control therefore returns s s s with probability at least 1 − 4 δ 1-4\delta 1 − 4 δ . Decode it as x ∗ x_\ast x ∗ , rejecting an out-of-range index or an index outside U 1 U_1 U 1 . The circuit makes O ( D d e c ) O(D_{\mathrm{dec}}) O ( D dec ) function queries. Its budget is fixed even on an input for which the promise or uniqueness fails. □
We next check the costs of every stage, including the stages before the last restricted role.
Proposition 8.10 (Search above the square-root scale). For Q ≥ N Q\ge\sqrt{N} Q ≥ N , every function of the prescribed bounded tuple admits an r r r -collision search algorithm using O K , r ( Q ) O_{K,r}(Q) O K , r ( Q ) function queries and succeeding with probability at least 2 / 3 2/3 2/3 .
Proof. Use Lemma 8.7 and the random role assignment above. Abort any trial whose public role sizes violate N ρ i / 2 ≤ v i ≤ 2 N ρ i N\rho_i/2\le v_i\le2N\rho_i N ρ i /2 ≤ v i ≤ 2 N ρ i . The constant D D D ensures that every remaining trial satisfies the raw-set capacity conditions of the colored construction. Supply the bounds ∣ A K , r B j ∏ i ≤ j ρ i ∣ |A_{K,r}B_j\prod_{i\le j}\rho_i| ∣ A K , r B j ∏ i ≤ j ρ i ∣ to
that algorithm. The ceiling changes them by only a constant factor: B j ≥ M B_j \ge M B j ≥ M , and ∏ i ≤ j ρ i ≥ ∏ i ≤ r ρ i = κ / M \prod_{i \le j} \rho_i \ge\prod_{i \le r} \rho_i = \kappa/M ∏ i ≤ j ρ i ≥ ∏ i ≤ r ρ i = κ / M .
On the event of Lemma 8.8 , these are valid prefix bounds. In each setup term of (152) , the role probabilities cancel between the prefix population and the role sizes. Thus that term is
O K , r ( a j B j ∏ i < j a i N j / 2 ) = O K , r ( Q ) . O_{K,r}\left(\frac{a_j\sqrt{B_j\prod_{i<j}a_i}}{N^{j/2}}\right)=O_{K,r}(Q). O K , r N j /2 a j B j ∏ i < j a i = O K , r ( Q ) .
Let t t t be the last role whose probability was decreased. If j ≤ t j \le t j ≤ t , all earlier probabilities are ρ i = D a i / N \rho_i = Da_i/N ρ i = D a i / N and hence
∏ i ≤ j v i ∏ i < j a i = O K , r ( N ) . \frac{\prod_{i\le j}v_i}{\prod_{i<j}a_i}=O_{K,r}(N). ∏ i < j a i ∏ i ≤ j v i = O K , r ( N ) .
Its certificate-stage cost is O K , r ( N ) = O K , r ( Q ) O_{K,r}(\sqrt{N})=O_{K,r}(Q) O K , r ( N ) = O K , r ( Q ) . If j > t j>t j > t , every probability after j j j is ρ 0 \rho_0 ρ 0 , so (160) gives
∏ i ≤ j v i ∏ i < j a i = O K , r ( N j M ∏ i < j a i ) ≤ O K , r ( Q 2 ) \frac{\prod_{i\le j}v_i}{\prod_{i<j}a_i}=O_{K,r}\left(\frac{N^j}{M\prod_{i<j}a_i}\right)\le O_{K,r}(Q^2) ∏ i < j a i ∏ i ≤ j v i = O K , r ( M ∏ i < j a i N j ) ≤ O K , r ( Q 2 )
by (158) . Proposition 8.6 therefore gives a bounded-error colored decision algorithm with O K , r ( Q ) O_{K,r}(Q) O K , r ( Q ) queries.
On the event of Lemma 8.8 , the colorful target is unique. Apply Lemma 8.9 to the fixed colored promise with the supplied prefix bounds. It finds its first-role input x x x with constant success probability in O K , r ( Q ) O_{K,r}(Q) O K , r ( Q ) function queries. Query f ( x ) f(x) f ( x ) and use verified randomized Grover search, excluding previously found inputs, to find r − 1 r-1 r − 1 further preimages in the original domain. There are only a fixed number of these searches, and their total fixed budget can be chosen O K , r ( N ) ≤ O K , r ( Q ) O_{K,r}(\sqrt{N}) \le O_{K,r}(Q) O K , r ( N ) ≤ O K , r ( Q ) with constant conditional success probability. Verify all r r r distinct reported inputs.
All circuits are run with fixed query budgets, even on trials outside the prefix-count promise, and every final collision is verified. The unknown prefix-count event need not be checked by the algorithm. On the event in Lemma 8.8 , the search succeeds with a fixed positive probability. A constant number of trials raises total success to 2 / 3 2/3 2/3 . The worst-case query cost is O K , r ( Q ) O_{K,r}(Q) O K , r ( Q ) , with no logarithmic search-conversion factor.
Combining the two regimes
Proof of Theorem 8.1 . The type determines Q Q Q . If Q ≤ N Q \le\sqrt{N} Q ≤ N , apply Theorem 8.4 . Otherwise apply Proposition 8.10 . Both are bounded-query search algorithms that succeed on every input of the stated type, use only function queries, and have constants independent of the output alphabet. This proves the claimed bound.
The colored learning-graph construction adapts the corrected construction in [6 ] ; the role-dependent tuple counts and the ordered restriction are the additional ingredients used here. The proof does not require a multidimensional quantum walk that outputs a marked vertex. It also does not establish a time bound for implementing the adversary-based decision algorithm. For fixed K , r K,r K , r , the large number of correction branches is a constant in the query model.
Only one collision is requested. An upper bound for outputting many distinct target classes would require a separate enumeration or reuse analysis and is not asserted by Theorem 8.1 .
Proof of Theorem 1.1. The quantity Q Q Q in (123) is exactly Q r ( λ ) Q_r(\lambda) Q r ( λ ) in (1) . Theorem 7.1, with p 0 = 2 / 3 p_0 = 2/3 p 0 = 2/3 , gives the lower bound under the uniform distribution of the type, and hence for worst-case bounded-error search. Theorem 8.1 gives an algorithm with the matching query count on every input of that type. Both estimates are uniform over all feasible alphabets.
Consequences and further directions
The complexity formula separates several sources of difficulty that a single eligible-bucket count cannot distinguish. An input can contain many partial collisions while containing few classes in which those collisions can be extended. The lower bound selects the most restrictive intermediate order. The upper bound allocates raw sets to balance the same sequence of intermediate populations.
The exceptional type illustrates why oracle access must remain explicit. In a public slot representation of ( k n , k + 1 ) (k^n,k+1) ( k n , k + 1 ) , independent forward permutation queries allow the exceptional slots to be searched for in O k ( N ) O_k(\sqrt{N}) O k ( N ) queries. The larger function-query lower bound uses the restriction to complete function queries. Filtering controls the canonical return of such a query; it does not establish improved soundness for arbitrary unfiltered permutation-query algorithms. Indeed, Proposition 4.16 shows that the unfiltered process has a square-root soundness obstruction even for complete function queries. Its harmonic projection is thus a substantive change to the compared evolution, not merely an improvement in the bound for the same simulation.
Further extensions require additional work. First, outputting collisions in several distinct classes requires tracking depletion and the reuse of intermediate data. The single-target matching theorem does not provide that analysis. Second, the learning-graph result is a query bound. An equally efficient implementation in time for every bounded type would require suitable data structures and coherent updates. The three-collision construction in the appendix gives a concrete alternative update mechanism, but is not an all-order implementation theorem. Allowing the maximum class size or collision order to grow with N N N would also require explicit control of the constants suppressed throughout this paper.
AI methodology
GPT-6 Astra was used in developing the proofs in this work. The authors proposed the main ideas: combining compressed permutation techniques with Belovs’s transfer map; identifying the square-root obstruction in the unfiltered CPO simulation, which motivated the filtered CPO approach; obtaining sharper upper bounds through more careful control of intermediate parameters in adaptive learning graphs and nested quantum walks; and characterizing the lower bound for arbitrary types as the maximum of several type-dependent bounds. Through iterative discussions with the authors, GPT-6 Astra developed the detailed proofs for all of these components.
GPT-6 Astra also assisted with English exposition and the preparation of references. The introduction and technical overview combine text written by the authors with model-generated drafts that the authors edited and revised. The authors take full responsibility for the final manuscript, including its mathematical arguments and references.
References
An alternative nested-walk construction for three-collision search
The general upper bound in Theorem 8.1 uses a learning graph above the square-root scale. This appendix records a different construction for three-collisions. Its distinct ingredient is an all-pairs inner-buffer update that remains reversible when eligible pairs overlap. We give the stationary-distribution, spectral-gap, and initialization arguments explicitly. This alternative incurs polylogarithmic overhead from coherent walk operations; it is not needed for the logarithm-free bound in the main text.
Problem, conclusion, and scope
Let f : [ N ] → Z q f : [N] \to\mathbb{Z}_q f : [ N ] → Z q be any function with exact type λ = ( 1 n 1 , … , K n K ) \lambda= (1^{n_1}, \ldots, K^{n_K}) λ = ( 1 n 1 , … , K n K ) , where K ≥ 3 K \ge3 K ≥ 3 is fixed. Its equality classes are the nonempty sets f − 1 ( y ) f^{-1}(y) f − 1 ( y ) . Set
B = ∑ d = 2 K ( d 2 ) n d , M = ∑ d = 3 K n d ≥ 1. B = \sum_{d=2}^{K} \binom{d}{2} n_d,\qquad M = \sum_{d=3}^{K} n_d \ge1. B = d = 2 ∑ K ( 2 d ) n d , M = d = 3 ∑ K n d ≥ 1.
Thus B B B is the number of unordered equal pairs, and M M M is the number of classes that contain a three-collision. In particular, 3 M ≤ B ≤ ( K − 1 ) N / 2 3M \le B \le(K-1)N/2 3 M ≤ B ≤ ( K − 1 ) N /2 and B = Θ K ( B 2 ) B = \Theta_K(B_2) B = Θ K ( B 2 ) for B 2 = ∑ d ≥ 2 n d B_2 = \sum_{d\ge2} n_d B 2 = ∑ d ≥ 2 n d . The algorithm knows the type, and has only ordinary function-query access. It is not given the members of any equality class or access to a hidden permutation.
Theorem A.1. For fixed K K K , a three-collision can be found with probability at least 2 / 3 2/3 2/3 , for every such function, using
O ~ K ( Q ) function queries , Q = max { N 2 / 3 M 1 / 3 , N 4 / 7 B 1 / 7 M 2 / 7 } . \widetilde{O}_K(Q)\ \text{function queries},\qquad Q = \max\left\{\frac{N^{2/3}}{M^{1/3}}, \frac{N^{4/7}B^{1/7}}{M^{2/7}}\right\}. O K ( Q ) function queries , Q = max { M 1/3 N 2/3 , M 2/7 N 4/7 B 1/7 } . The bound is uniform over every feasible alphabet. The tilde suppresses polylogarithmic factors in N N N arising from coherent search and approximate reflections.
The theorem concerns search, not only a decision promise. Its quantum-walk compilation uses the finding theorem for nested updates of Childs, Jeffery, Kothari, and Magniez [16 Definition 3.2 and Theorem 3.3] . The modifications below are additional arguments; the type-dependent theorem is not a theorem quoted from that paper. Theorem 7.1 gives the matching type-dependent lower bound. The construction in this appendix is specific to three-collision search; it does not assert a higher-order nested-walk extension.
Parameters and the two search regimes
Write Q 2 = N 2 / 3 M − 1 / 3 Q_2 = N^{2/3}M^{-1/3} Q 2 = N 2/3 M − 1/3 and Q 3 = N 4 / 7 B 1 / 7 M − 2 / 7 Q_3 = N^{4/7}B^{1/7}M^{-2/7} Q 3 = N 4/7 B 1/7 M − 2/7 . Temporarily ignore constant factors and integer rounding, and put
s 0 = Q , L 0 = max { 1 , min { Q 3 / 2 B N , B Q N } } . (166) s_0 = Q,\qquad L_0 = \max\left\{1,\min\left\{\frac{Q^{3/2}\sqrt{B}}{N},\frac{BQ}{N}\right\}\right\}. \tag*{(166)} s 0 = Q , L 0 = max { 1 , min { N Q 3/2 B , N B Q } } . (166)
The following elementary identities explain both the desired costs and the normalization bounds used later.
Lemma A.2. Up to constants depending only on K K K , these parameters satisfy
s 0 + N L 0 B s 0 + N M s 0 + N B M L 0 = O K ( Q ) , s_0+\frac{NL_0}{\sqrt{Bs_0}}+\frac{N}{\sqrt{Ms_0}}+\sqrt{\frac{NB}{ML_0}}=O_K(Q), s 0 + B s 0 N L 0 + M s 0 N + M L 0 N B = O K ( Q ) ,
min ( M , L 0 ) s 0 N = O K ( 1 ) , \min(M,L_0)\frac{s_0}{N}=O_K(1), min ( M , L 0 ) N s 0 = O K ( 1 ) ,
μ 0 L 0 B = O K ( 1 ) , M s 0 3 N 3 = o ( min ( 1 , μ 0 ) ) , μ 0 = B s 0 2 N 2 . \frac{\mu_0L_0}{B}=O_K(1),\qquad\frac{Ms_0^3}{N^3}=o(\min(1,\mu_0)),\qquad\mu_0=\frac{Bs_0^2}{N^2}. B μ 0 L 0 = O K ( 1 ) , N 3 M s 0 3 = o ( min ( 1 , μ 0 )) , μ 0 = N 2 B s 0 2 .
Also Q = O K ( N 5 / 7 ) Q=O_K(N^{5/7}) Q = O K ( N 5/7 ) , and μ 0 / L 0 = o ( 1 ) \mu_0/L_0=o(1) μ 0 / L 0 = o ( 1 ) whenever L 0 → ∞ L_0\to\infty L 0 → ∞ .
Proof. The condition Q 3 ≥ Q 2 Q_3\ge Q_2 Q 3 ≥ Q 2 is equivalent to B 3 M ≥ N 2 B^3M\ge N^2 B 3 M ≥ N 2 and to Q 3 ≤ B Q_3\le B Q 3 ≤ B . In this case the first expression in the inner minimum in (A.1) is smaller. Its square is Q 3 B / N 2 ≥ B / M Q^3B/N^2\ge B/M Q 3 B / N 2 ≥ B / M , so it is at least one. Balancing this choice of L 0 L_0 L 0 gives the first inequality. Moreover,
( M Q N ) ( L 0 Q N ) = B N , μ 0 L 0 B = B M N . \left(\frac{MQ}{N}\right)\left(\frac{L_0Q}{N}\right)=\frac{B}{N},\qquad\frac{\mu_0L_0}{B}=\frac{B}{MN}. ( N M Q ) ( N L 0 Q ) = N B , B μ 0 L 0 = M N B .
If Q 2 ≥ Q 3 Q_2\ge Q_3 Q 2 ≥ Q 3 , then B ≤ Q B\le Q B ≤ Q and the second expression in the minimum applies. When it is at least one the same two identities hold, using Q 3 = N 2 / M Q^3=N^2/M Q 3 = N 2 / M ; the setup term is B Q ≤ Q \sqrt{BQ}\le Q B Q ≤ Q . When it is below one, set L 0 = 1 L_0=1 L 0 = 1 . Then N / B Q ≤ Q N/\sqrt{BQ}\le Q N / B Q ≤ Q since B ≥ M B\ge M B ≥ M , and N B / M ≤ N / M Q ≤ Q \sqrt{NB/M}\le N/\sqrt{MQ}\le Q N B / M ≤ N / M Q ≤ Q since B Q < N BQ<N B Q < N . The remaining normalization assertions follow from the displayed identities and B = O K ( N ) B=O_K(N) B = O K ( N ) . In the Q 2 Q_2 Q 2 branch, M s 0 3 / N 3 = 1 / N Ms_0^3/N^3=1/N M s 0 3 / N 3 = 1/ N ; in the Q 3 Q_3 Q 3 branch it is N − 9 / 7 B 3 / 7 M 1 / 7 = O K ( N − 5 / 7 ) N^{-9/7}B^{3/7}M^{1/7}=O_K(N^{-5/7}) N − 9/7 B 3/7 M 1/7 = O K ( N − 5/7 ) . When μ 0 < 1 \mu_0<1 μ 0 < 1 , its ratio to μ 0 \mu_0 μ 0 is M s 0 / ( B N ) = O K ( s 0 / N ) = o ( 1 ) Ms_0/(BN)=O_K(s_0/N)=o(1) M s 0 / ( B N ) = O K ( s 0 / N ) = o ( 1 ) . Finally, μ 0 / L 0 \mu_0/L_0 μ 0 / L 0 is Q / N Q/N Q / N or B Q / N \sqrt{BQ}/N B Q / N , respectively, unless L 0 = 1 L_0=1 L 0 = 1 . □
In the walk construction take s = ⌊ a K s 0 ⌋ s=\lfloor a_Ks_0\rfloor s = ⌊ a K s 0 ⌋ and L = max { 1 , ⌊ b K L 0 ⌋ } L=\max\{1,\lfloor b_KL_0\rfloor\} L = max { 1 , ⌊ b K L 0 ⌋} for sufficiently small positive constants. They can be chosen in order so that the setup has many more available partner inputs than L L L , and m max L / B m_{\max}L/B m m a x L / B is below the constant required by the spectral-gap lemma. Every cost in Lemma A.2 changes only by a K K K -dependent factor. Cases where L 0 L_0 L 0 is bounded by a sufficiently large K K K -dependent constant use L = 1 L=1 L = 1 and the separate initialization below; this also changes the costs only by a constant.
For Q ≤ N Q\le\sqrt{N} Q ≤ N , Theorem 8.4 with r = 3 r=3 r = 3 already provides a direct search algorithm: its tail-count scale is no larger than Q Q Q , because B 2 ≤ B B_2\le B B 2 ≤ B . Since also B = O K ( B 2 ) B=O_K(B_2) B = O K ( B 2 ) , both scales agree up to a K K K -dependent constant. The nested walk addresses Q > N Q>\sqrt{N} Q > N .
A symmetric update for lists from distinct equality classes
This subsection gives an exact implementation of the outer update. The stationary distribution is allowed to depend on the number of admissible inner buffers. This removes the need to assume that all collision pairs in the input are disjoint.
Fix a function f : [ N ] → Z q f:[N]\to\mathbb{Z}_q f : [ N ] → Z q whose equality classes have sizes at most a fixed constant K K K . Let M M M be the number of classes of size at least three. Fix disjoint sets A 1 , A 2 ⊆ [ N ] A_1,A_2\subseteq[N] A 1 , A 2 ⊆ [ N ] and a fixed set Z ⊆ Z q \mathcal{Z}\subseteq\mathbb{Z}_q Z ⊆ Z q of masked values. Define the eligible pair pool
P = { { x , y } : x ∈ A 1 , y ∈ A 2 , f ( x ) = f ( y ) ∉ Z } , B = ∣ P ∣ . \mathcal{P}=\{\{x,y\}:x\in A_1,\ y\in A_2,\ f(x)=f(y)\notin\mathcal{Z}\},\qquad B=|\mathcal{P}|. P = {{ x , y } : x ∈ A 1 , y ∈ A 2 , f ( x ) = f ( y ) ∈ / Z } , B = ∣ P ∣.
For the one-pair initialization branch, P \mathcal{P} P may instead be the full pool of equal pairs. The proofs below apply verbatim: the cross-pair restriction is then simply omitted. The mask changes only which pairs are eligible. In particular, inputs whose values belong to Z \mathcal{Z} Z are not excluded from an inner buffer merely because of the mask. All queried values are stored as their actual values under f f f .
Write w b w_b w b for the number of eligible pairs in equality class b b b . Then 0 ≤ w b ≤ W K : = ( K 2 ) 0 \le w_b \le W_K := \binom{K}{2} 0 ≤ w b ≤ W K := ( 2 K ) . An eligible class of size two has exactly one eligible pair. Let Ω L \Omega_L Ω L be the family of sets S S S of L L L eligible pairs, with at most one pair from each equality class. For any such pair set W W W , let end ( W ) \operatorname{end}(W) end ( W ) be the union of its endpoints and let F ( W ) F(W) F ( W ) be the set of its function values. Thus ∣ end ( S ) ∣ = 2 L |\operatorname{end}(S)| = 2L ∣ end ( S ) ∣ = 2 L for S ∈ Ω L S \in\Omega_L S ∈ Ω L .
Fix an integer buffer size s s s and a nonempty set of permitted batch sizes
J ⊆ { 1 , … , min { L , ⌊ s / 2 ⌋ } } . \mathcal{J} \subseteq\{1,\ldots,\min\{L,\lfloor s/2\rfloor\}\}. J ⊆ { 1 , … , min { L , ⌊ s /2 ⌋}} .
For an input set R R R , put
J ( R ) = { e ∈ P : e ⊆ R } , X ( R ) = ∣ J ( R ) ∣ . \mathcal{J}(R)=\{e\in\mathcal{P}:e\subseteq R\},\qquad X(R)=|\mathcal{J}(R)|. J ( R ) = { e ∈ P : e ⊆ R } , X ( R ) = ∣ J ( R ) ∣.
For S ∈ Ω L S\in\Omega_L S ∈ Ω L , a buffer is good if
Good S = { R ⊆ [ N ] ∖ end ( S ) : ∣ R ∣ = s , f ( R ) ∩ F ( S ) = ∅ , \operatorname{Good}_S=\left\{R\subseteq[N]\setminus\operatorname{end}(S): |R|=s,\ f(R)\cap F(S)=\varnothing,
\right. Good S = { R ⊆ [ N ] ∖ end ( S ) : ∣ R ∣ = s , f ( R ) ∩ F ( S ) = ∅ ,
each equality class meets R in at most two inputs , X ( R ) ∈ J } . \left.\text{each equality class meets }R\text{ in at most two inputs},\quad X(R)\in\mathcal{J}\right\}. each equality class meets R in at most two inputs , X ( R ) ∈ J } .
Let Z S = ∣ Good S ∣ Z_S=|\operatorname{Good}_S| Z S = ∣ Good S ∣ . The occupancy condition implies that J ( R ) \mathcal{J}(R) J ( R ) consists of disjoint pairs from distinct equality classes. Their classes are also disjoint from those represented by S S S .
For a pair set W W W with distinct values and an integer u ≥ 0 u\ge0 u ≥ 0 , define the residual-buffer family
F W , u = { C ⊆ [ N ] ∖ end ( W ) : ∣ C ∣ = u , f ( C ) ∩ F ( W ) = ∅ , \mathcal{F}_{W,u}=\left\{C\subseteq[N]\setminus\operatorname{end}(W): |C|=u,\ f(C)\cap F(W)=\varnothing,
\right. F W , u = { C ⊆ [ N ] ∖ end ( W ) : ∣ C ∣ = u , f ( C ) ∩ F ( W ) = ∅ ,
each equality class meets C in at most two inputs , J ( C ) = ∅ } . \left.\text{each equality class meets }C\text{ in at most two inputs},\quad\mathcal{J}(C)=\varnothing\right\}. each equality class meets C in at most two inputs , J ( C ) = ∅ } .
Write G W ( u ) = ∣ F W , u ∣ G_W(u)=|\mathcal{F}_{W,u}| G W ( u ) = ∣ F W , u ∣ . Here J ( C ) = ∅ \mathcal{J}(C)=\varnothing J ( C ) = ∅ forbids eligible cross-pairs only. It does not forbid, for example, a pair whose value is masked.
Lemma A.3 (Exact local diffusion and symmetric garbage). Suppose Z S > 0 Z_S>0 Z S > 0 for every S ∈ Ω L S\in\Omega_L S ∈ Ω L . There is a Markov chain P P P on Ω L \Omega_L Ω L with the following transition probabilities. If S ∪ S ′ S\cup S' S ∪ S ′ represents distinct equality classes and m = ∣ S ∖ S ′ ∣ = ∣ S ′ ∖ S ∣ ∈ J m=|S\setminus S'|=|S'\setminus S|\in\mathcal{J} m = ∣ S ∖ S ′ ∣ = ∣ S ′ ∖ S ∣ ∈ J , then
P ( S , S ′ ) = G S ∪ S ′ ( s − 2 m ) Z S ( L m ) ; (167) P(S,S')=\frac{G_{S\cup S'}(s-2m)}{Z_S\binom{L}{m}}; \tag*{(167)} P ( S , S ′ ) = Z S ( m L ) G S ∪ S ′ ( s − 2 m ) ; (167) otherwise P ( S , S ′ ) = 0 P(S,S')=0 P ( S , S ′ ) = 0 . Its stationary distribution is
π ( S ) = Z S ∑ U ∈ Ω L Z U . (168) \pi(S)=\frac{Z_S}{\sum_{U\in\Omega_L}Z_U}. \tag*{(168)} π ( S ) = ∑ U ∈ Ω L Z U Z S . (168) Given the queried values in S S S and a coherent uniform good buffer, the local diffusion associated with P P P uses no further function queries. Its edge garbage is invariant under exchanging S S S and S ′ S' S ′ , so the corresponding garbage swap also uses no further queries.
Proof. For R ∈ Good S R\in\operatorname{Good}_S R ∈ Good S , let m = X ( R ) m=X(R) m = X ( R ) and extract all the pairs J = J ( R ) J=\mathcal{J}(R) J = J ( R ) . Set C = R ∖ end ( J ) C=R\setminus\operatorname{end}(J) C = R ∖ end ( J ) . Every class represented by J J J has exactly two inputs in R R R , so C C C contains no further input from any of these classes. It also avoids the classes of S S S . Consequently,
C ∈ F S ∪ J , s − 2 m . C\in\mathcal{F}_{S\cup J,s-2m}. C ∈ F S ∪ J , s − 2 m .
Choose I ⊆ S I \subseteq S I ⊆ S , ∣ I ∣ = m |I|=m ∣ I ∣ = m , uniformly, and put S ′ = ( S ∖ I ) ∪ J S'=(S\setminus I)\cup J S ′ = ( S ∖ I ) ∪ J . Then S ∪ J = S ∪ S ′ S\cup J=S\cup S' S ∪ J = S ∪ S ′ .
Conversely, fix S ′ S' S ′ occurring in (A.2) and C ∈ F S ∪ S ′ , s − 2 m C\in\mathcal{F}_{S\cup S',s-2m} C ∈ F S ∪ S ′ , s − 2 m . The sets
I = S ∖ S ′ , J = S ′ ∖ S , R = C ∪ end ( J ) I=S\setminus S',\qquad J=S'\setminus S,\qquad R=C\cup\operatorname{end}(J) I = S ∖ S ′ , J = S ′ ∖ S , R = C ∪ end ( J )
recover the original choices uniquely. The buffer C C C avoids every class represented by S ∪ S ′ S\cup S' S ∪ S ′ . Adding end ( J ) \operatorname{end}(J) end ( J ) therefore creates exactly the m m m eligible pairs J J J , with occupancy exactly two in their classes, and creates no other eligible pair. Thus R ∈ Good S R\in\operatorname{Good}_S R ∈ Good S . This proves a bijection between the choices ( R , I ) (R,I) ( R , I ) and the resulting pairs ( S ′ , C ) (S',C) ( S ′ , C ) .
Include the queried function values in the notation for each basis state. Starting with the normalized uniform good-buffer state, the query-free extraction and uniform choice of I I I give
1 Z S ∑ R ∈ Good S ∣ S , R ⟩ ⟼ ∑ S ′ G S ∪ S ′ ( s − 2 m ) Z S ( L m ) ∣ S , S ′ ⟩ ∣ ψ S ∪ S ′ , m ⟩ . \frac{1}{\sqrt{Z_S}}\sum_{R\in\operatorname{Good}_S}|S,R\rangle
\longmapsto
\sum_{S'}\sqrt{\frac{G_{S\cup S'}(s-2m)}{Z_S\binom{L}{m}}}|S,S'\rangle|\psi_{S\cup S',m}\rangle. Z S 1 R ∈ Good S ∑ ∣ S , R ⟩ ⟼ S ′ ∑ Z S ( m L ) G S ∪ S ′ ( s − 2 m ) ∣ S , S ′ ⟩ ∣ ψ S ∪ S ′ , m ⟩ .
where terms with zero residual count are omitted and
∣ ψ W , m ⟩ = 1 G W ( s − 2 m ) ∑ C ∈ F W , s − 2 m ∣ C , f ∣ C ⟩ . |\psi_{W,m}\rangle=\frac{1}{\sqrt{G_W(s-2m)}}\sum_{C\in\mathcal{F}_{W,s-2m}}|C,f|C\rangle. ∣ ψ W , m ⟩ = G W ( s − 2 m ) 1 C ∈ F W , s − 2 m ∑ ∣ C , f ∣ C ⟩ .
The bijection proves that the coefficients in (A.2) sum to one. It also makes the extraction reversible: from S , S ′ , C S,S',C S , S ′ , C one recovers both R R R and the chosen I I I . All values needed for these operations were already stored.
Detailed balance follows from the exact identity
π ( S ) P ( S , S ′ ) = G S ∪ S ′ ( s − 2 m ) ( L m ) ∑ U ∈ Ω L Z U = π ( S ′ ) P ( S ′ , S ) . \pi(S)P(S,S')=\frac{G_{S\cup S'}(s-2m)}{\binom{L}{m}\sum_{U\in\Omega_L}Z_U}=\pi(S')P(S',S). π ( S ) P ( S , S ′ ) = ( m L ) ∑ U ∈ Ω L Z U G S ∪ S ′ ( s − 2 m ) = π ( S ′ ) P ( S ′ , S ) .
Finally, the normalized garbage state depends only on S ∪ S ′ S\cup S' S ∪ S ′ and m m m . Swapping the two outer registers leaves this state unchanged. This proves the asserted zero-query garbage swap. □ \square □
One may replace P P P by ( I + P ) / 2 (I+P)/2 ( I + P ) /2 to make the chain lazy. This does not change its stationary distribution and changes its gap and implementation costs by at most constant factors.
Uniform comparison of residual-buffer counts. The next elementary switching bound controls the dependence of G W ( u ) G_W(u) G W ( u ) on the selected equality classes.
Lemma A.4 (Deleting a few positions from a residual buffer space). Let a domain of n n n inputs be partitioned into classes of size at most $K . Consider its size-u u u subsets satisfying an occupancy cap of two and containing no pair from a fixed eligible cross-pair pool . If n > K u n>Ku n > K u , the probability that a uniform such subset contains a specified input is at most
u n − K u . \frac{u}{n-Ku}. n − K u u . In particular, suppose u = o ( N ) u=o(N) u = o ( N ) , n = Θ ( N ) n=\Theta(N) n = Θ ( N ) , and D = o ( N ) D=o(N) D = o ( N ) . Deleting any D D D positions changes the number of admissible subsets by a multiplicative factor between exp ( − O K ( D u / N ) ) \exp(-O_K(Du/N)) exp ( − O K ( D u / N )) and 1 1 1 , provided the remaining domain still has size Θ ( N ) \Theta(N) Θ ( N ) .
Proof. Fix an input x x x . For every admissible set C C C containing x x x , replace x x x by an input y y y from a class not represented in C C C . At least n − K u n-Ku n − K u choices of y y y are available. The replacement is admissible: removal preserves both constraints, and the new input lies in an otherwise empty class. For any resulting set not containing x x x , at most u u u choices of its element y y y can invert this operation. If a a a and b b b count admissible sets containing and not containing x x x , respectively, then
( n − K u ) a ≤ u b . (n-Ku)a \le ub. ( n − K u ) a ≤ u b .
This implies the displayed marginal bound. Apply that bound successively while deleting the prescribed positions. The ratio of the final count to the original count is at least
∏ j = 0 D − 1 ( 1 − u n − j − K u ) . \prod_{j=0}^{D-1}\left(1-\frac{u}{n-j-Ku}\right). j = 0 ∏ D − 1 ( 1 − n − j − K u u ) .
For sufficiently large N N N , the denominators have order N N N and the subtracted fractions are o ( 1 ) o(1) o ( 1 ) . Taking logarithms gives the claimed lower bound. Deletion cannot increase the count. □ \square □
Lemma A.5 (Residual counts are comparable). Assume s = o ( N ) s=o(N) s = o ( N ) , L = o ( N ) L=o(N) L = o ( N ) , and
min { M , L } s N ≤ A (169) \frac{\min\{M,L\}s}{N}\le A \tag*{(169)} N min { M , L } s ≤ A (169) for a fixed constant A A A . Fix m ∈ J m\in\mathcal{J} m ∈ J and put u = s − 2 m u=s-2m u = s − 2 m . For any two pair sets W , W ′ W,W' W , W ′ of size L + m L+m L + m with distinct equality classes,
G W ( u ) = Θ K , A ( G W ′ ( u ) ) . G_W(u)=\Theta_{K,A}(G_{W'}(u)). G W ( u ) = Θ K , A ( G W ′ ( u )) . The constants are uniform over the fixed cross-pair pool and its fixed value mask.
Proof. All eligible classes of size two have the same residual-buffer constraint: both of their inputs cannot be selected together. In particular, their counting polynomial is 1 + 2 z 1+2z 1 + 2 z , independently of their actual input positions and output labels. Thus, after relabeling whole classes of this kind, the residual domains for W W W and W ′ W' W ′ can be aligned on all but
D = O K ( min { M , L + m } ) D=O_K(\min\{M,L+m\}) D = O K ( min { M , L + m })
positions. To see this explicitly, keep the exceptional classes of size at least three that remain in both domains. Delete the exceptional classes that remain in only one domain, and match the numbers of remaining eligible size-two classes by deleting the excess from the other domain. The number of exceptional selected classes is at most min { M , L + m } \min\{M,L+m\} min { M , L + m } , and each has at most K K K inputs. All other classes, including masked classes, are unchanged. This yields isomorphic common residual domains after deleting at most O K ( min { M , L + m } ) O_K(\min\{M,L+m\}) O K ( min { M , L + m }) positions from each.
The full residual domains have size N − O K ( L + m ) = N − o ( N ) N-O_K(L+m)=N-o(N) N − O K ( L + m ) = N − o ( N ) . Lemma A.4 compares each count to that on the common domain. Since m ≤ L m\le L m ≤ L and u ≤ s u\le s u ≤ s ,
D u N = O K ( min { M , L } s N ) = O K ( A ) . \frac{Du}{N}=O_K\left(\frac{\min\{M,L\}s}{N}\right)=O_K(A). N D u = O K ( N min { M , L } s ) = O K ( A ) .
Both comparisons therefore lose at most a constant factor. □ \square □
Preparing and reflecting about good inner buffers. The inner state is obtained from an ordinary Johnson walk; none of the validity constraints has to be built into its transition rule. This distinction is useful because the ordinary walk has a known uniform stationary state and gap of order 1 / s 1/s 1/ s .
Lemma A.6 (Density of good buffers). Assume K K K and A A A are fixed, s ≥ 2 s \ge2 s ≥ 2 , s = o ( N ) s = o(N) s = o ( N ) , L = o ( N ) L = o(N) L = o ( N ) , (A.4), and B ≥ 2 W K L B \ge2W_KL B ≥ 2 W K L . Put
μ = B s 2 N 2 , ν = M s 3 N 3 , \mu= \frac{Bs^2}{N^2}, \qquad\nu= \frac{Ms^3}{N^3}, μ = N 2 B s 2 , ν = N 3 M s 3 , and suppose ν = o ( min { 1 , μ } ) \nu= o(\min\{1,\mu\}) ν = o ( min { 1 , μ }) . There are constants μ 0 , c − , c + > 0 \mu_0,c_-,c_+ > 0 μ 0 , c − , c + > 0 , depending only on K K K , with the following property. Choose the batch window to be { 1 } \{1\} { 1 } when μ < μ 0 \mu< \mu_0 μ < μ 0 , and otherwise to be
J = { m ∈ Z : max { 1 , ⌈ c − μ ⌉ } ≤ m ≤ ⌊ c + μ ⌋ } . \mathcal{J} = \{m \in\mathbb{Z} : \max\{1,\lceil c_-\mu\rceil\} \le m \le\lfloor c_+\mu\rfloor\}. J = { m ∈ Z : max { 1 , ⌈ c − μ ⌉} ≤ m ≤ ⌊ c + μ ⌋} . Assume this window is contained in { 1 , … , min { L , ⌊ s / 2 ⌋ } } \{1,\ldots,\min\{L,\lfloor s/2\rfloor\}\} { 1 , … , min { L , ⌊ s /2 ⌋}} . Then, uniformly over S ∈ Ω L S \in\Omega_L S ∈ Ω L ,
Z S ( N − 2 L s ) = Θ K , A ( min { 1 , μ } ) . (170) \frac{Z_S}{\binom{N-2L}{s}} = \Theta_{K,A}(\min\{1,\mu\}). \tag*{(170)} ( s N − 2 L ) Z S = Θ K , A ( min { 1 , μ }) . (170) In particular, the stationary distribution in (A.3) is pointwise comparable to the uniform distribution on Ω L \Omega_L Ω L .
Proof. A raw buffer is a uniform s s s -subset of [ N ] ∖ end ( S ) [N] \setminus\operatorname{end}(S) [ N ] ∖ end ( S ) . Beyond the 2 L 2L 2 L known endpoints, the classes represented by S S S contain at most
D S ≤ ( K − 2 ) min { M , L } D_S \le(K-2)\min\{M,L\} D S ≤ ( K − 2 ) min { M , L } further inputs. The probability of avoiding these inputs is
( N − 2 L − D S s ) ( N − 2 L s ) = Ω K , A ( 1 ) , \frac{\binom{N-2L-D_S}{s}}{\binom{N-2L}{s}} = \Omega_{K,A}(1), ( s N − 2 L ) ( s N − 2 L − D S ) = Ω K , A ( 1 ) , by the product formula for the ratio, s , L = o ( N ) s,L = o(N) s , L = o ( N ) , and (A.4). Condition on this avoidance. The buffer is now uniform on a domain U S U_S U S of size n S = N − O K ( L ) ∼ N n_S = N - O_K(L) \sim N n S = N − O K ( L ) ∼ N obtained by deleting the complete classes represented by S S S .
Let B S B_S B S be the number of eligible pairs remaining on U S U_S U S . The assumption B ≥ 2 W K L B \ge2W_KL B ≥ 2 W K L gives B / 2 ≤ B S ≤ B B/2 \le B_S \le B B /2 ≤ B S ≤ B . For the eligible-pair count X = X ( R ) X = X(R) X = X ( R ) ,
E X = B S ( s 2 ) ( n S 2 ) = Θ K ( μ ) . \mathbb{E}X = B_S\frac{\binom{s}{2}}{\binom{n_S}{2}} = \Theta_K(\mu). E X = B S ( 2 n S ) ( 2 s ) = Θ K ( μ ) . Pairs on disjoint input sets contribute O K ( μ 2 ) O_K(\mu^2) O K ( μ 2 ) to E [ X ( X − 1 ) ] \mathbb{E}[X(X-1)] E [ X ( X − 1 )] . Each eligible pair overlaps only O K ( 1 ) O_K(1) O K ( 1 ) other eligible pairs; an overlapping pair of pairs requires three sampled inputs. Consequently,
E [ X ( X − 1 ) ] = O K ( μ 2 + μ s N ) , E X 2 = O K ( μ + μ 2 ) . (171) \mathbb{E}[X(X-1)] = O_K\left(\mu^2+\mu\frac{s}{N}\right), \qquad\mathbb{E}X^2 = O_K(\mu+\mu^2). \tag*{(171)} E [ X ( X − 1 )] = O K ( μ 2 + μ N s ) , E X 2 = O K ( μ + μ 2 ) . (171) All these estimates are uniform in S S S .
For a sufficiently small fixed μ 0 \mu_0 μ 0 , if μ < μ 0 \mu< \mu_0 μ < μ 0 then
Pr [ X = 1 ] ≥ E X − E [ X ( X − 1 ) ] = Ω K ( μ ) . \Pr[X=1] \ge\mathbb{E}X-\mathbb{E}[X(X-1)] = \Omega_K(\mu). Pr [ X = 1 ] ≥ E X − E [ X ( X − 1 )] = Ω K ( μ ) . Here the first inequality follows pointwise from 1 { x = 1 } ≥ x − x ( x − 1 ) \mathbf{1}_{\{x=1\}} \ge x-x(x-1) 1 { x = 1 } ≥ x − x ( x − 1 ) for nonnegative integers x x x . If μ ≥ μ 0 \mu\ge\mu_0 μ ≥ μ 0 , the second-moment inequality gives a constant lower bound on Pr [ X ≥ c − μ ] \Pr[X \ge c_-\mu] Pr [ X ≥ c − μ ] for a suitable c − > 0 c_- > 0 c − > 0 . Markov’s inequality makes Pr [ X > c + μ ] \Pr[X > c_+\mu] Pr [ X > c + μ ] smaller than half that lower bound by choosing
c + c_{+} c + sufficiently large. The constant can also be increased so that the integer window is nonempty throughout μ ≥ μ 0 \mu\ge\mu_0 μ ≥ μ 0 . Thus
Pr [ X ∈ J ] = Ω K ( min { 1 , μ } ) . \Pr[X \in\mathcal{J}] = \Omega_K(\min\{1,\mu\}). Pr [ X ∈ J ] = Ω K ( min { 1 , μ }) .
There are at most ( K 3 ) M \binom{K}{3}M ( 3 K ) M triples within equality classes. A union bound therefore gives
Pr [ some class has occupancy at least three in R ] = O K ( ν ) = o ( min { 1 , μ } ) . \Pr[\text{some class has occupancy at least three in }R] = O_K(\nu) = o(\min\{1,\mu\}). Pr [ some class has occupancy at least three in R ] = O K ( ν ) = o ( min { 1 , μ }) .
Removing these buffers preserves the preceding lower bound. Multiplication by the probability of avoiding the extra inputs proves the lower bound in (A.5). The upper bound follows from Pr [ X ≥ 1 ] ≤ min { 1 , E X } \Pr[X \ge1] \le\min\{1,\mathbb{E}X\} Pr [ X ≥ 1 ] ≤ min { 1 , E X } , also without conditioning on the avoidance event. Finally, all Z S Z_S Z S are comparable to the same quantity in (A.5), which proves the claim about π \pi π .
On the raw Johnson graph J ( N − 2 L , s ) J(N-2L,s) J ( N − 2 L , s ) , setup costs s s s function queries, an update costs O ( 1 ) O(1) O ( 1 ) queries, and the good-buffer predicate costs no further queries once the buffer values are stored. Its gap is Θ ( 1 / s ) \Theta(1/s) Θ ( 1/ s ) for s = o ( N ) s=o(N) s = o ( N ) . Thus standard coherent quantum-walk search and its inverse prepare, and reflect about, the normalized good-buffer state with the inner search cost
O ~ K , A ( s min { 1 , μ } ) = O ~ K , A ( max { s , N B s } ) . (172) \widetilde{O}_{K,A}\left(\sqrt{\frac{s}{\min\{1,\mu\}}}\right)
=
\widetilde{O}_{K,A}\left(\max\left\{\sqrt{s},\frac{N}{\sqrt{Bs}}\right\}\right). \tag*{(172)} O K , A ( min { 1 , μ } s ) = O K , A ( max { s , B s N } ) . (172)
The one-time initial preparation also includes the s s s setup queries. The logarithmic factors allow the coherent operations to be implemented to the inverse-polynomial accuracy required by the enclosing walk. Lemma A.3 then supplies the local diffusion and garbage swap exactly, without additional function queries.
The outer walk on pairs from distinct classes
We first analyze the outer chain independently of its quantum implementation. Let E = ⨆ b = 1 H E b E=\bigsqcup_{b=1}^{H}E_b E = ⨆ b = 1 H E b , where E b E_b E b is the set of pair choices in class b b b , and write
1 ≤ w b : = ∣ E b ∣ ≤ W , B = ∑ b = 1 H w b . 1\le w_b:=|E_b|\le W,\qquad B=\sum_{b=1}^{H}w_b. 1 ≤ w b := ∣ E b ∣ ≤ W , B = b = 1 ∑ H w b .
For bounded multiplicities one may take W = ( K 2 ) W=\binom{K}{2} W = ( 2 K ) . For 1 ≤ L ≤ H 1\le L\le H 1 ≤ L ≤ H , put
Ω L = { S ⊆ E : ∣ S ∣ = L , ∣ S ∩ E b ∣ ≤ 1 for every b } . \Omega_L=\{S\subseteq E:|S|=L,\ |S\cap E_b|\le1\text{ for every }b\}. Ω L = { S ⊆ E : ∣ S ∣ = L , ∣ S ∩ E b ∣ ≤ 1 for every b } .
Let cl ( S ) \operatorname{cl}(S) cl ( S ) denote the classes represented by S S S , and let π 0 \pi_0 π 0 be uniform on Ω L \Omega_L Ω L . Under π 0 \pi_0 π 0 , a class set A A A has probability proportional to w ( A ) : = ∏ b ∈ A w b w(A):=\prod_{b\in A}w_b w ( A ) := ∏ b ∈ A w b ; conditional on A A A , the item in each selected class is uniform. We use e j e_j e j for an elementary symmetric polynomial, with e 0 = 1 e_0=1 e 0 = 1 and e j = 0 e_j=0 e j = 0 outside its possible range, and set
Z j = e j ( w 1 , … , w H ) , Z j ( R ) = e j ( ( w b ) b ∉ cl ( R ) ) , z j ( S ) = Z j ( S ) . Z_j=e_j(w_1,\ldots,w_H),\qquad Z_j(R)=e_j((w_b)_{b\notin\operatorname{cl}(R)}),\qquad z_j(S)=Z_j(S). Z j = e j ( w 1 , … , w H ) , Z j ( R ) = e j (( w b ) b ∈ / cl ( R ) ) , z j ( S ) = Z j ( S ) .
For a reversible Markov kernel P P P with stationary measure π \pi π , write
E π , P ( f ) = 1 2 ∑ S , T π ( S ) P ( S , T ) ( f ( S ) − f ( T ) ) 2 . \mathcal{E}_{\pi,P}(f)=\frac{1}{2}\sum_{S,T}\pi(S)P(S,T)(f(S)-f(T))^2. E π , P ( f ) = 2 1 S , T ∑ π ( S ) P ( S , T ) ( f ( S ) − f ( T ) ) 2 .
The Poincaré spectral gap is the largest γ \gamma γ for which E π , P ( f ) ≥ γ Var π f \mathcal{E}_{\pi,P}(f)\ge\gamma\operatorname{Var}_{\pi}f E π , P ( f ) ≥ γ Var π f for all real f f f . When the kernel is not already lazy, we explicitly pass to ( I + P ) / 2 (I+P)/2 ( I + P ) /2 before using an absolute spectral gap for quantum search.
Weighted subsets and the marked fraction
The following coupling will also be used to compare two batch updates.
Lemma A.7 (Nested weighted subsets). For positive weights on a finite set, let ν j ( A ) = w ( A ) / e j ( w ) \nu_j(A) = w(A)/e_j(w) ν j ( A ) = w ( A ) / e j ( w ) on its j j j -subsets. For each feasible j ≥ 1 j \ge1 j ≥ 1 , there is a coupling of A j − 1 ∼ ν j − 1 A_{j-1} \sim\nu_{j-1} A j − 1 ∼ ν j − 1 and A j ∼ ν j A_j \sim\nu_j A j ∼ ν j with A j − 1 ⊆ A j A_{j-1} \subseteq A_j A j − 1 ⊆ A j .
Proof. The elementary symmetric polynomials satisfy e i 2 ≥ e i − 1 e i + 1 e_i^2 \ge e_{i-1}e_{i+1} e i 2 ≥ e i − 1 e i + 1 . For completeness, this follows by adding the weights one at a time. If the old coefficient sequence is ( a i ) (a_i) ( a i ) , adjoining a weight w w w gives b i = a i + w a i − 1 b_i = a_i + wa_{i-1} b i = a i + w a i − 1 , and
b i 2 − b i − 1 b i + 1 = ( a i 2 − a i − 1 a i + 1 ) + w ( a i a i − 1 − a i − 2 a i + 1 ) + w 2 ( a i − 1 2 − a i − 2 a i ) ≥ 0. b_i^2 - b_{i-1}b_{i+1} = (a_i^2 - a_{i-1}a_{i+1}) + w(a_ia_{i-1} - a_{i-2}a_{i+1}) + w^2(a_{i-1}^2 - a_{i-2}a_i) \ge0. b i 2 − b i − 1 b i + 1 = ( a i 2 − a i − 1 a i + 1 ) + w ( a i a i − 1 − a i − 2 a i + 1 ) + w 2 ( a i − 1 2 − a i − 2 a i ) ≥ 0.
The middle term is nonnegative because the ratios of consecutive positive terms of a log-concave sequence are nonincreasing; the boundary cases use the zero coefficients outside its support.
We now induct on the number of elements. If the last element has weight w w w , its inclusion probability under ν j \nu_j ν j is
p j = w e j − 1 e j + w e j − 1 . p_j = \frac{we_{j-1}}{e_j + we_{j-1}}. p j = e j + w e j − 1 w e j − 1 .
where these polynomials concern the remaining elements. Log-concavity implies p j − 1 ≤ p j p_{j-1} \le p_j p j − 1 ≤ p j . Couple the two last-element indicators monotonically. If both are absent, apply induction to sizes j − 1 , j j-1,j j − 1 , j on the remaining set. If both are present, apply it to sizes j − 2 , j − 1 j-2,j-1 j − 2 , j − 1 . If the indicator changes from absent to present, choose the same ( j − 1 ) (j-1) ( j − 1 ) -subset of the remaining elements. Each case has the required conditional marginals and preserves containment; zero-probability boundary cases can be omitted.
Suppose that M M M of the H H H classes are eligible, and mark an outer state if it represents at least one eligible class. This event has the required probability even though different classes have different numbers of pair choices.
Lemma A.8 (Marked fraction). For every 1 ≤ L ≤ H 1 \le L \le H 1 ≤ L ≤ H , the marked fraction under π 0 \pi_0 π 0 satisfies
1 W + 1 min { 1 , M L H } ≤ ε 0 ≤ W min { 1 , M L H } . \frac{1}{W+1}\min\left\{1,\frac{ML}{H}\right\} \le\varepsilon_0 \le W\min\left\{1,\frac{ML}{H}\right\}. W + 1 1 min { 1 , H M L } ≤ ε 0 ≤ W min { 1 , H M L } . In particular, it is Θ W ( min { 1 , M L / B } ) \Theta_W(\min\{1,ML/B\}) Θ W ( min { 1 , M L / B }) .
Proof. Let I b I_b I b indicate inclusion of class b b b . Its probability is
p b = w b e L − 1 ( w − b ) e L ( w − b ) + w b e L − 1 ( w − b ) . p_b = \frac{w_b e_{L-1}(w_{-b})}{e_L(w_{-b}) + w_b e_{L-1}(w_{-b})}. p b = e L ( w − b ) + w b e L − 1 ( w − b ) w b e L − 1 ( w − b ) .
The identity
L e L ( w − b ) = ∑ A ⊆ [ H ] ∖ { b } ∣ A ∣ = L − 1 w ( A ) ∑ c ∉ A ∪ { b } w c L e_L(w_{-b}) = \sum_{\substack{A \subseteq[H]\setminus\{b\}\\ |A|=L-1}} w(A) \sum_{c \notin A\cup\{b\}} w_c L e L ( w − b ) = A ⊆ [ H ] ∖ { b } ∣ A ∣ = L − 1 ∑ w ( A ) c ∈ / A ∪ { b } ∑ w c
shows that
H − L L ≤ e L ( w − b ) e L − 1 ( w − b ) ≤ W ( H − L ) L . \frac{H-L}{L} \le\frac{e_L(w_{-b})}{e_{L-1}(w_{-b})} \le\frac{W(H-L)}{L}. L H − L ≤ e L − 1 ( w − b ) e L ( w − b ) ≤ L W ( H − L ) .
Consequently L / ( W H ) ≤ p b ≤ W L / H L/(WH) \le p_b \le WL/H L / ( W H ) ≤ p b ≤ W L / H . The indicators are pairwise negatively correlated. Indeed, for distinct b , c b,c b , c put a i = e i ( w − { b , c } ) a_i=e_i(w_{-\{b,c\}}) a i = e i ( w − { b , c } ) . Clearing denominators reduces E ( I b I c ) ≤ p b p c \mathbb{E}(I_bI_c)\le p_bp_c E ( I b I c ) ≤ p b p c to a L − 2 a L ≤ a L − 1 2 a_{L-2}a_L\le a_{L-1}^2 a L − 2 a L ≤ a L − 1 2 , which was established in the preceding proof. Thus, if X X X counts eligible classes represented in the state,
M L W H ≤ E X ≤ W M L H , E X 2 ≤ E X + ( E X ) 2 . \frac{ML}{WH}\le\mathbb{E}X\le\frac{WML}{H},\qquad\mathbb{E}X^2\le\mathbb{E}X+(\mathbb{E}X)^2. W H M L ≤ E X ≤ H W M L , E X 2 ≤ E X + ( E X ) 2 .
Cauchy–Schwarz and the union bound give
E X 1 + E X ≤ Pr ( X > 0 ) ≤ min { 1 , E X } . \frac{\mathbb{E}X}{1+\mathbb{E}X}\le\Pr(X>0)\le\min\{1,\mathbb{E}X\}. 1 + E X E X ≤ Pr ( X > 0 ) ≤ min { 1 , E X } .
The claimed constants follow. Finally, H ≤ B ≤ W H H\le B\le WH H ≤ B ≤ W H .
Batch heat-bath updates
We next obtain a gap bound for the chain that will serve as a comparison chain. For 1 ≤ m ≤ L 1\le m\le L 1 ≤ m ≤ L , the batch heat-bath kernel P m h b P_m^{\mathrm{hb}} P m hb removes a uniform m m m -subset of the current state and then chooses a uniform completion in Ω L \Omega_L Ω L of the retained set. It is reversible for π 0 \pi_0 π 0 : its transition probability is a sum, over retained sets R ⊆ S ∩ T R\subseteq S\cap T R ⊆ S ∩ T of size L − m L-m L − m , of 1 / ( ( L m ) Z m ( R ) ) 1/\left(\binom{L}{m}Z_m(R)\right) 1/ ( ( m L ) Z m ( R ) ) , which is symmetric in S , T S,T S , T .
Lemma A.9 (Batch heat-bath gap). If L ≤ B / ( 3 W ) L\le B/(3W) L ≤ B / ( 3 W ) , then every nonconstant eigenvalue of P m h b P_m^{\mathrm{hb}} P m hb has absolute value at most 1 − m / ( 2 L ) 1-m/(2L) 1 − m / ( 2 L ) . For L = 1 L=1 L = 1 , the chain is independent uniform resampling and has gap one, without the displayed restriction.
Proof. Use the exchange metric d ( S , T ) = L − ∣ S ∩ T ∣ d(S,T)=L-|S\cap T| d ( S , T ) = L − ∣ S ∩ T ∣ . It is the graph distance for valid single-item exchanges: first correct the chosen item in classes common to the two states, and then replace classes present only on one side. We give a coupling for adjacent states S = R 0 ∪ { u } S=R_0\cup\{u\} S = R 0 ∪ { u } and T = R 0 ∪ { v } T=R_0\cup\{v\} T = R 0 ∪ { v } . Couple the removal sets by identifying u u u with v v v and leaving all common items fixed. With probability m / L m/L m / L , the differing items are removed; identical completions then coalesce. Otherwise, the retained states still differ by u , v u,v u , v . If these items have the same class, their completion spaces coincide, and identical completions preserve distance one.
Suppose that their classes are different. Let C C C be the classes available to both completions. The first completion can additionally use the class of v v v , and the second the class of u u u . Their class weights are denoted by w v , w u w_v,w_u w v , w u in this paragraph. Their probabilities of using these extra classes are
p v = w v e m − 1 ( C ) e m ( C ) + w v e m − 1 ( C ) , p u = w u e m − 1 ( C ) e m ( C ) + w u e m − 1 ( C ) . p_v=\frac{w_v e_{m-1}(C)}{e_m(C)+w_v e_{m-1}(C)},\qquad
p_u=\frac{w_u e_{m-1}(C)}{e_m(C)+w_u e_{m-1}(C)}. p v = e m ( C ) + w v e m − 1 ( C ) w v e m − 1 ( C ) , p u = e m ( C ) + w u e m − 1 ( C ) w u e m − 1 ( C ) .
The elementary-symmetric-polynomial identity used above gives
e m ( C ) e m − 1 ( C ) ≥ ∑ b ∈ C w b − ( m − 1 ) W m . \frac{e_m(C)}{e_{m-1}(C)}\ge\frac{\sum_{b\in C}w_b-(m-1)W}{m}. e m − 1 ( C ) e m ( C ) ≥ m ∑ b ∈ C w b − ( m − 1 ) W .
At most L − m + 1 L-m+1 L − m + 1 classes are excluded in defining C C C , so
max { p u , p v } ≤ m W B − W L . \max\{p_u,p_v\}\le\frac{mW}{B-WL}. max { p u , p v } ≤ B − W L mW .
Couple the two extra-class indicators maximally. If neither extra class is used, choose the same weighted m m m -subset of C C C . The final distance is one. If both are used, choose the same weighted ( m − 1 ) (m-1) ( m − 1 ) -subset of C C C ; the final distance is at most two. If exactly one is used, apply Lemma A.7 to couple the common-class subsets of sizes m − 1 m-1 m − 1 and m m m by containment. Couple pair choices identically in shared classes. The final distance is again at most two. Therefore
E d ( S ′ , T ′ ) ≤ ( 1 − m L ) ( 1 + m W B − W L ) \mathbb{E}d(S',T') \leq\left(1-\frac{m}{L}\right)\left(1+\frac{mW}{B-WL}\right) E d ( S ′ , T ′ ) ≤ ( 1 − L m ) ( 1 + B − W L mW )
≤ 1 − m L + m W B − W L ≤ 1 − m 2 L . \leq1-\frac{m}{L}+\frac{mW}{B-WL}\leq1-\frac{m}{2L}. ≤ 1 − L m + B − W L mW ≤ 1 − 2 L m .
The last inequality uses 3 W L ≤ B 3WL\leq B 3 W L ≤ B . Composing the adjacent-state couplings along shortest paths gives the same contraction for arbitrary states. Equivalently, the Lipschitz seminorm of P m h b f P_m^{\mathrm{hb}}f P m hb f contracts by this factor. Applying that assertion to any nonconstant eigenfunction proves the eigenvalue bound. If L = 1 L=1 L = 1 , removing the sole item leaves the empty state and the subsequent completion is an independent uniform sample, proving the final assertion.
Strict exchanges
The actual update inserts pairs from classes absent from the entire old list. We therefore remove the heat-bath transitions that reuse a deleted class. Write S ∼ m T S\sim_m T S ∼ m T if ∣ S ∩ T ∣ = L − m |S\cap T|=L-m ∣ S ∩ T ∣ = L − m and every class represented by T ∖ S T\setminus S T ∖ S is absent from S S S . This condition is symmetric. The removed and inserted sets are uniquely determined on such an edge.
Lemma A.10 (Strict-exchange comparison). Suppose 1 ≤ m ≤ L 1\leq m\leq L 1 ≤ m ≤ L and
m L B ≤ 1 100 W . (173) \frac{mL}{B}\leq\frac{1}{100W}. \tag*{(173)} B m L ≤ 100 W 1 . (173) The strict-exchange kernel
P m s t r ( S , T ) = 1 { S ∼ m T } ( L m ) z m ( S ) P_m^{\mathrm{str}}(S,T)=\frac{\mathbf{1}\{S\sim_m T\}}{\binom{L}{m}z_m(S)} P m str ( S , T ) = ( m L ) z m ( S ) 1 { S ∼ m T } is reversible for π m s t r ( S ) ∝ z m ( S ) \pi_m^{\mathrm{str}}(S)\propto z_m(S) π m str ( S ) ∝ z m ( S ) , and
49 50 ≤ z m ( S ) Z m ≤ 1. (174) \frac{49}{50}\leq\frac{z_m(S)}{Z_m}\leq1. \tag*{(174)} 50 49 ≤ Z m z m ( S ) ≤ 1. (174) Its Poincaré gap is at least m / ( 4 L ) m/(4L) m / ( 4 L ) , and the absolute gap of its lazy version is at least m / ( 8 L ) m/(8L) m / ( 8 L ) . More precisely, the symmetric substochastic kernel
K m ( S , T ) = 1 { S ∼ m T } ( L m ) Z m K_m(S,T)=\frac{\mathbf{1}\{S\sim_m T\}}{\binom{L}{m}Z_m} K m ( S , T ) = ( m L ) Z m 1 { S ∼ m T } satisfies
1 2 ∑ S , T π 0 ( S ) K m ( S , T ) ( f ( S ) − f ( T ) ) 2 ≥ m 4 L Var π 0 f . (175) \frac{1}{2}\sum_{S,T}\pi_0(S)K_m(S,T)(f(S)-f(T))^2\geq\frac{m}{4L}\operatorname{Var}_{\pi_0}f. \tag*{(175)} 2 1 S , T ∑ π 0 ( S ) K m ( S , T ) ( f ( S ) − f ( T ) ) 2 ≥ 4 L m Var π 0 f . (175)
Proof. Under the weighted m m m -subset distribution on all classes, the inclusion probability of a fixed class is at most
m W B − W ( m − 1 ) ≤ 2 m W B . \frac{mW}{B-W(m-1)}\leq\frac{2mW}{B}. B − W ( m − 1 ) mW ≤ B 2 mW .
This follows directly by bounding the average remaining weight in the elementary-symmetric-polynomial identity. A union bound over the L L L classes of S S S proves (A.9). Condition (A.8) implies L ≤ B / ( 3 W ) L \le B/(3W) L ≤ B / ( 3 W ) . In a heat-bath step, call reuse of any removed class bad. Conditional on the retained set, the inclusion probability of one removed class is at most
m W B − W ( L − 1 ) ≤ 2 m W B . \frac{mW}{B-W(L-1)} \le\frac{2mW}{B}. B − W ( L − 1 ) mW ≤ B 2 mW .
The total bad-step probability is consequently at most α = 2 W m 2 / B \alpha= 2Wm^2/B α = 2 W m 2 / B . The bad-edge kernel is symmetric. Centering f f f at its π 0 \pi_0 π 0 -mean and using ( f ( S ) − f ( T ) ) 2 ≤ 2 f ( S ) 2 + 2 f ( T ) 2 (f(S)-f(T))^2 \le2f(S)^2+2f(T)^2 ( f ( S ) − f ( T ) ) 2 ≤ 2 f ( S ) 2 + 2 f ( T ) 2 bounds its Dirichlet form by 2 α Var π 0 f 2\alpha\operatorname{Var}_{\pi_0}f 2 α Var π 0 f . Lemma A.9 therefore gives
E g o o d ( f ) ≥ ( m 2 L − 4 W m 2 B ) Var π 0 f . \mathcal{E}_{\mathrm{good}}(f) \ge\left(\frac{m}{2L}-\frac{4Wm^2}{B}\right)\operatorname{Var}_{\pi_0}f. E good ( f ) ≥ ( 2 L m − B 4 W m 2 ) Var π 0 f .
where the good kernel retains only strict exchanges.
On a strict edge the retained set is necessarily S ∩ T S\cap T S ∩ T , so the good heat-bath transition is 1 / ( L m ) Z m ( S ∩ T ) 1/\binom{L}{m}Z_m(S\cap T) 1/ ( m L ) Z m ( S ∩ T ) . The preceding union bound, applied to its L − m L-m L − m classes, gives Z m ( S ∩ T ) / Z m ≥ 49 / 50 Z_m(S\cap T)/Z_m \ge49/50 Z m ( S ∩ T ) / Z m ≥ 49/50 . Consequently
E π 0 , K m ( f ) ≥ 49 50 E g o o d ( f ) ≥ 49 50 ( 1 2 − 1 25 ) m L Var π 0 f ≥ m 4 L Var π 0 f . \begin{aligned}
\mathcal{E}_{\pi_0,K_m}(f) \ge\frac{49}{50}\mathcal{E}_{\mathrm{good}}(f) \\
&\ge\frac{49}{50}\left(\frac{1}{2}-\frac{1}{25}\right)\frac{m}{L}\operatorname{Var}_{\pi_0}f \ge\frac{m}{4L}\operatorname{Var}_{\pi_0}f.
\end{aligned} E π 0 , K m ( f ) ≥ 50 49 E good ( f ) ≥ 50 49 ( 2 1 − 25 1 ) L m Var π 0 f ≥ 4 L m Var π 0 f .
Here the notation for the Dirichlet form also applies to the substochastic kernel; adding its missing mass on the diagonal does not change that form.
The row sum of K m K_m K m is d m ( S ) = z m ( S ) / Z m d_m(S)=z_m(S)/Z_m d m ( S ) = z m ( S ) / Z m . Put d ˉ m = E π 0 d m \bar d_m=\mathbb{E}_{\pi_0}d_m d ˉ m = E π 0 d m . Conditioning on an accepted step gives P m s t r P_m^{\mathrm{str}} P m str , with stationary density d m / d ˉ m d_m/\bar d_m d m / d ˉ m relative to π 0 \pi_0 π 0 . Thus
E π m s t r , P m s t r ( f ) = E π 0 , K m ( f ) d ˉ m , Var π m s t r f ≤ Var π 0 f d ˉ m . \mathcal{E}_{\pi_m^{\mathrm{str}},P_m^{\mathrm{str}}}(f)=\frac{\mathcal{E}_{\pi_0,K_m}(f)}{\bar d_m}, \qquad\operatorname{Var}_{\pi_m^{\mathrm{str}}}f\le\frac{\operatorname{Var}_{\pi_0}f}{\bar d_m}. E π m str , P m str ( f ) = d ˉ m E π 0 , K m ( f ) , Var π m str f ≤ d ˉ m Var π 0 f .
The second inequality follows by choosing the π 0 \pi_0 π 0 -mean as the centering constant and using d m ≤ 1 d_m\le1 d m ≤ 1 . These inequalities prove the Poincaré gap. Lazification makes all eigenvalues nonnegative and divides the gap by two. □
The exact buffer-weighted chain
We finally allow the batch size and the residual-buffer count to depend on the transition. This comparison does not assume that buffer counts agree across outer states.
Theorem A.11 (Buffer-weighted variable-batch gap). Let I ⊆ { 1 , … , L } \mathcal{I}\subseteq\{1,\ldots,L\} I ⊆ { 1 , … , L } be nonempty, with endpoints m min m_{\min} m m i n and m max m_{\max} m m a x , and assume
m max L B ≤ 1 100 W . \frac{m_{\max}L}{B}\le\frac{1}{100W}. B m m a x L ≤ 100 W 1 . For every m ∈ I m\in\mathcal{I} m ∈ I , suppose that a positive symmetric buffer count on strict exchange edges satisfies
g m ≤ G S ∪ T ( s − 2 m ) ≤ C 0 g m ( S ∼ m T ) , (176) g_m\le G_{S\cup T}(s-2m)\le C_0g_m \qquad(S\sim_m T), \tag*{(176)} g m ≤ G S ∪ T ( s − 2 m ) ≤ C 0 g m ( S ∼ m T ) , (176) where g m > 0 g_m>0 g m > 0 and C 0 ≥ 1 C_0\ge1 C 0 ≥ 1 is independent of m , S , T m,S,T m , S , T . Define conductances and degrees by
c ( S , T ) = ∑ m ∈ I 1 { S ∼ m T } ( L m ) G S ∪ T ( s − 2 m ) , c(S,T)=\sum_{m\in\mathcal{I}}\frac{\mathbf{1}_{\{S\sim_m T\}}}{\binom{L}{m}}G_{S\cup T}(s-2m), c ( S , T ) = m ∈ I ∑ ( m L ) 1 { S ∼ m T } G S ∪ T ( s − 2 m ) , A ( S ) = ∑ T c ( S , T ) , P ( S , T ) = c ( S , T ) A ( S ) . A(S)=\sum_T c(S,T), \qquad P(S,T)=\frac{c(S,T)}{A(S)}. A ( S ) = T ∑ c ( S , T ) , P ( S , T ) = A ( S ) c ( S , T ) .
Then P P P is reversible for π ( S ) ∝ A ( S ) \pi(S) \propto A(S) π ( S ) ∝ A ( S ) , and
49 50 C 0 ≤ π ( S ) π 0 ( S ) ≤ 50 C 0 49 . (177) \frac{49}{50C_0} \le\frac{\pi(S)}{\pi_0(S)} \le\frac{50C_0}{49}. \tag*{(177)} 50 C 0 49 ≤ π 0 ( S ) π ( S ) ≤ 49 50 C 0 . (177)
The absolute spectral gap of the lazy chain ( I + P ) / 2 (I+P)/2 ( I + P ) /2 is at least
49 400 C 0 m min L . (178) \frac{49}{400C_0}\frac{m_{\min}}{L}. \tag*{(178)} 400 C 0 49 L m m i n . (178)
Its marked fraction is Θ W , C 0 ( min { 1 , M L / B } ) \Theta_{W,C_0}(\min\{1,ML/B\}) Θ W , C 0 ( min { 1 , M L / B }) .
Proof. Symmetry of the conductances proves reversibility. For fixed m m m , each inserted set J J J can be paired with exactly ( L m ) \binom{L}{m} ( m L ) removed sets, and S ∪ T = S ∪ J S \cup T = S \cup J S ∪ T = S ∪ J does not depend on the removed set. In particular, the degree contribution is exactly
A m ( S ) = ∑ J ∈ Ω m cl ( J ) ∩ cl ( S ) = ∅ G S ∪ J ( s − 2 m ) . (179) A_m(S)=\sum_{\substack{J\in\Omega_m\\ \operatorname{cl}(J)\cap\operatorname{cl}(S)=\varnothing}}G_{S\cup J}(s-2m). \tag*{(179)} A m ( S ) = J ∈ Ω m cl ( J ) ∩ cl ( S ) = ∅ ∑ G S ∪ J ( s − 2 m ) . (179)
Compare first with the baseline conductances
c 0 ( S , T ) = ∑ m ∈ I g m ( L m ) 1 { S ∼ m T } . c^0(S,T)=\sum_{m\in\mathcal{I}}\frac{g_m}{\binom{L}{m}}\mathbf{1}\{S\sim_m T\}. c 0 ( S , T ) = m ∈ I ∑ ( m L ) g m 1 { S ∼ m T } .
Their degrees are A 0 ( S ) = ∑ m g m z m ( S ) A^0(S)=\sum_m g_m z_m(S) A 0 ( S ) = ∑ m g m z m ( S ) . With Z ∗ = ∑ m g m Z m Z_*=\sum_m g_m Z_m Z ∗ = ∑ m g m Z m , Lemma A.10 gives
49 50 Z ∗ ≤ A 0 ( S ) ≤ Z ∗ . (180) \frac{49}{50}Z_* \le A^0(S)\le Z_*. \tag*{(180)} 50 49 Z ∗ ≤ A 0 ( S ) ≤ Z ∗ . (180)
Let P 0 P^0 P 0 and π 0 \pi^0 π 0 be the associated normalized kernel and stationary measure, and put D 0 = ∑ S A 0 ( S ) D^0=\sum_S A^0(S) D 0 = ∑ S A 0 ( S ) . Equation (A.10) implies
E π 0 , P 0 ( f ) = ∣ Ω L ∣ D 0 ∑ m g m Z m E π 0 , K m ( f ) ≥ m min 4 L Var π 0 f . \mathcal{E}_{\pi^0,P^0}(f)=\frac{|\Omega_L|}{D^0}\sum_m g_m Z_m\mathcal{E}_{\pi^0,K_m}(f)
\ge\frac{m_{\min}}{4L}\operatorname{Var}_{\pi^0}f. E π 0 , P 0 ( f ) = D 0 ∣ Ω L ∣ m ∑ g m Z m E π 0 , K m ( f ) ≥ 4 L m m i n Var π 0 f .
because D 0 ≤ ∣ Ω L ∣ Z ∗ D^0\le|\Omega_L|Z_* D 0 ≤ ∣ Ω L ∣ Z ∗ . Also π 0 / π 0 ≤ 50 / 49 \pi^0/\pi_0\le50/49 π 0 / π 0 ≤ 50/49 by (A.15), so
gap ( P 0 ) ≥ 49 200 m min L . \operatorname{gap}(P^0)\ge\frac{49}{200}\frac{m_{\min}}{L}. gap ( P 0 ) ≥ 200 49 L m m i n .
Assumption (A.11) gives c 0 ≤ c ≤ C 0 c 0 c^0\le c\le C_0c^0 c 0 ≤ c ≤ C 0 c 0 . Writing D = ∑ S A ( S ) D=\sum_S A(S) D = ∑ S A ( S ) , direct comparison of Dirichlet forms and stationary variances yields
E π , P ( f ) ≥ D 0 D E π 0 , P 0 ( f ) , Var π f ≤ C 0 D 0 D Var π 0 f . \mathcal{E}_{\pi,P}(f)\ge\frac{D^0}{D}\mathcal{E}_{\pi^0,P^0}(f),\qquad
\operatorname{Var}_{\pi}f\le C_0\frac{D^0}{D}\operatorname{Var}_{\pi^0}f. E π , P ( f ) ≥ D D 0 E π 0 , P 0 ( f ) , Var π f ≤ C 0 D D 0 Var π 0 f .
For the variance inequality, center at the π 0 \pi^0 π 0 -mean and use A ( S ) ≤ C 0 A 0 ( S ) A(S)\le C_0A^0(S) A ( S ) ≤ C 0 A 0 ( S ) . The Poincaré gap of P P P is therefore at least 49 m min / ( 200 C 0 L ) 49m_{\min}/(200C_0L) 49 m m i n / ( 200 C 0 L ) . Lazification proves (A.13).
Finally, 49 Z ∗ / 50 ≤ A ( S ) ≤ C 0 Z ∗ 49Z_*/50\le A(S)\le C_0Z_* 49 Z ∗ /50 ≤ A ( S ) ≤ C 0 Z ∗ proves the density bounds (A.12). Applying those bounds to the marked event and using Lemma A.8 proves its stated probability. Lazification changes neither the stationary measure nor the marked event. □
The theorem supplies the required outer parameters whenever the exact quantum update realizes these conductances and the buffer comparison (A.11) holds. It does not by itself construct the conditional completion states or their coherent update. Parameter regimes with only boundedly many pair classes can be handled separately instead of imposing a strict exchange when no new class is available.
Preparing the outer state without a uniqueness promise
The update construction uses a stationary distribution proportional to the number of admissible inner buffers. We do not assume that the initial list produced by a sequence of searches has this distribution. Instead, we first obtain a state that is uniform on the unmarked lists, and then project onto the stationary state of the implemented walk. The one preliminary test of whether a list is marked costs O ~ ( N ) \widetilde{O}(\sqrt{N}) O ( N ) ; this is affordable in the present regime Q > N Q > \sqrt{N} Q > N .
Write B = ∑ d ( d 2 ) n d B = \sum_d \binom{d}{2}n_d B = ∑ d ( 2 d ) n d and M = ∑ d ≥ 3 n d M = \sum_{d\ge3} n_d M = ∑ d ≥ 3 n d . Throughout this subsection the maximum class size is the fixed constant K K K , and
Q = max { N 2 / 3 M − 1 / 3 , N 4 / 7 B 1 / 7 M − 2 / 7 } . Q = \max\{N^{2/3}M^{-1/3},\ N^{4/7}B^{1/7}M^{-2/7}\}. Q = max { N 2/3 M − 1/3 , N 4/7 B 1/7 M − 2/7 } .
The choices of s = Θ K ( Q ) s = \Theta_K(Q) s = Θ K ( Q ) and L L L in the parameter calculation satisfy L ≤ c K B s / N L \le c_K Bs/N L ≤ c K B s / N outside the one-pair initialization regime. We first treat list sizes exceeding a sufficiently large fixed K K K -dependent cutoff; the estimates below approach probability one when L → ∞ L \to\infty L → ∞ , and exceed a fixed constant close to one uniformly above that cutoff. Smaller list sizes are handled separately below.
Independently place each input into one of two parts A ~ \widetilde{A} A and C C C , with equal probabilities. Their sizes are Θ ( N ) \Theta(N) Θ ( N ) except with exponentially small probability. Prepare a uniform queried s s s -subset I I I of A ~ \widetilde{A} A . A candidate partner is an input y ∈ C y \in C y ∈ C whose value occurs exactly once in I I I . If a value has already been selected, exclude that value from later searches. Thus repeated values in the seed are ignored; the seed is not required to be collision-free.
For a seed I I I , let
H ( I ) = ∣ { y ∈ C : ∣ { x ∈ I : f ( x ) = f ( y ) } ∣ = 1 } ∣ . H(I) = \left|\left\{y \in C : \left|\{x \in I : f(x) = f(y)\}\right| = 1\right\}\right|. H ( I ) = ∣ { y ∈ C : ∣ { x ∈ I : f ( x ) = f ( y )} ∣ = 1 } ∣ .
Use coherent quantum searches to select L L L partners of distinct values, keeping each selected input in superposition. Each partner has a uniquely determined anchor in I I I . The selection can be implemented with error N − c N^{-c} N − c per stage, for an arbitrary fixed c > 0 c > 0 c > 0 , whenever H ( I ) ≥ h 0 + K L H(I) \ge h_0 + KL H ( I ) ≥ h 0 + K L , where h 0 = c K ′ B s / N h_0 = c'_K Bs/N h 0 = c K ′ B s / N . The cost per stage is O ~ K ( N / B s ) \widetilde{O}_K(N/\sqrt{Bs}) O K ( N / B s ) . Validity flags may be measured, but the partner registers are not measured. The algorithm declares failure on an invalid stage. All search operations are symmetric among inputs satisfying their current search predicate. Consequently, their successful amplitudes are equal on those inputs, including any phase depending on the number of candidates. This observation will also account for seeds on which the prescribed high-success guarantee does not apply.
Let S S S be the resulting list of L L L pairs and let F = I ∖ anchors ( S ) F = I \setminus\operatorname{anchors}(S) F = I ∖ anchors ( S ) . Sort the list reversibly, retaining its ordering register, and measure F F F together with its queried values. On an unmarked outcome the ordering register is a known uniform state and can then be erased without queries. Define
A = A ~ ∖ F . A = \widetilde{A} \setminus F. A = A ∖ F .
The pair universe for the ensuing walk consists of equal-value pairs across A × C A \times C A × C whose values do not occur in F F F . A fixed mask implements this restriction: on the remaining inputs use the virtual function
g F ( x ) = { ( 0 , f ( x ) ) , f ( x ) ∉ f ( F ) , ( 1 , x ) , f ( x ) ∈ f ( F ) . (181) g_F(x) =
\begin{cases}
(0,f(x)), & f(x) \notin f(F),\\
(1,x), & f(x) \in f(F).
\end{cases} \tag*{(181)} g F ( x ) = { ( 0 , f ( x )) , ( 1 , x ) , f ( x ) ∈ / f ( F ) , f ( x ) ∈ f ( F ) . (181)
The two tags denote disjoint alphabets. A query to g F g_F g F uses a constant number of queries to f f f . In particular, inputs whose values are masked remain available to the inner buffer as singletons. We do not require the buffer to avoid every class touched by F F F . The pair universe is restricted to A × C A \times C A × C , but the inner buffer is sampled from the full original domain, including the masked positions, apart from the known endpoints of its current outer list. Its ambient size is therefore always N − 2 L N - 2L N − 2 L . The final third-input search also uses the original domain of f f f .
Lemma A.12 (Uniformity after an unmarked outcome ). Fix the measured set F F F and the classical transcript of the partner-search procedure. On the component in which every selected pair belongs to a class of size exactly two in the original input, the amplitudes of all valid outer lists are equal. Thus an ideal unmarked projection produces the uniform superposition of unmarked lists in the modified pair universe.
Proof. For an unmarked list S S S , every selected anchor belongs to a true two-element class, whose other element is its selected partner in C C C . This class has no member in F F F . Conversely, every unmarked list in the modified pair universe determines the unique seed I = F ∪ anchors ( S ) I=F\cup\operatorname{anchors}(S) I = F ∪ anchors ( S ) .
Adding these L L L anchors to F F F contributes exactly one candidate partner per anchor and does not change the candidate status of any other value. Hence
H ( I ) = H ( F ) + L . H(I)=H(F)+L. H ( I ) = H ( F ) + L .
After j j j selected partners, the candidate count is exactly H ( F ) + L − j H(F)+L-j H ( F ) + L − j . In an ideal uniform search, the amplitude of any ordered list is therefore proportional to
( ( H ( F ) + L ) L ) − 1 / 2 . \bigl((H(F)+L)L\bigr)^{-1/2}. ( ( H ( F ) + L ) L ) − 1/2 .
For an implemented symmetric search, an additional successful amplitude, phase, or transcript factor can depend on these candidate counts, but not on the particular unmarked list. The initial seed amplitudes are equal as well. All orders of a fixed unmarked list have the same amplitude, so their order register can be erased by a query-free change of basis. Measuring F F F leaves equal amplitudes on all unmarked lists, as claimed. No assertion about the amplitudes of marked lists is needed.
The test used to realize the unmarked projection is a verified Grover search for an input outside the stored endpoints whose value belongs to the list. It costs O ~ ( N ) \widetilde O(\sqrt{N}) O ( N ) , never returns an incorrect witness, and can have an arbitrarily small fixed failure probability on every marked list. On an unmarked list its failure workspace is independent of that list. If the marked component has constant weight, this preliminary test already succeeds with constant probability. Otherwise its failure branch has constant weight and, by Lemma A.12 , is arbitrarily close to the uniform unmarked state. This argument does not condition a small absolute error on an event of vanishing probability.
:::
Lemma A.13 (Retention of the input populations ). With probability bounded below by a positive constant depending only on K K K , the modified pair universe contains Θ K ( B ) \Theta_K(B) Θ K ( B ) pairs and has Θ K ( M ) \Theta_K(M) Θ K ( M ) classes that extend to a three-collision in the original input. For L → ∞ L\to\infty L → ∞ , the partner-selection procedure succeeds with probability 1 − o ( 1 ) 1-o(1) 1 − o ( 1 ) on the seeds needed for this conclusion.
Proof. The random two-coloring is independent across original equality classes. A class of size at least two crosses the two parts with probability at least 1 / 2 1/2 1/2 , and its number of cross pairs is at most the constant ( K 2 ) \binom{K}{2} ( 2 K ) . The mean number of cross pairs is B / 2 B/2 B /2 , and its variance is O K ( B ) O_K(B) O K ( B ) . Thus the pair count is Θ K ( B ) \Theta_K(B) Θ K ( B ) with probability 1 − o ( 1 ) 1-o(1) 1 − o ( 1 ) when B → ∞ B\to\infty B → ∞ . The number of eligible classes crossing the parts has mean at least M / 2 M/2 M /2 and second moment O K ( M + M 2 ) O_K(M+M^2) O K ( M + M 2 ) . It is therefore at least a fixed positive multiple of M M M with constant probability, including when M M M is bounded.
For a uniform seed of size s s s in a part of size Θ ( N ) \Theta(N) Θ ( N ) , the probability of touching any fixed class is O K ( s / N ) O_K(s/N) O K ( s / N ) . Therefore the expected number of target classes touched by I I I is O K ( M s / N ) O_K(Ms/N) O K ( M s / N ) , and the expected number of candidate pairs belonging to a class touched by I I I is O K ( B s / N ) O_K(Bs/N) O K ( B s / N ) . Since s = o ( N ) s=o(N) s = o ( N ) and F ⊆ I F\subseteq I F ⊆ I , Markov’s inequality shows that masking f ( F ) f(F) f ( F ) loses o ( M ) o(M) o ( M ) target classes and
::: o ( B ) o(B) o ( B ) pairs in probability. For bounded M M M this says, more strongly, that no target class is lost with probability 1 − o ( 1 ) 1-o(1) 1 − o ( 1 ) .
For completeness, put Y = ∑ x ∈ I deg C ( x ) Y=\sum_{x\in I}\deg_C(x) Y = ∑ x ∈ I deg C ( x ) , where deg C ( x ) \deg_C(x) deg C ( x ) is the number of equal-value inputs in C C C . On partitions with Θ K ( B ) \Theta_K(B) Θ K ( B ) cross pairs,
E Y = Θ K ( B s / N ) , Var Y ≤ ( K − 1 ) E Y . \mathbb{E}Y=\Theta_K(Bs/N),\qquad\operatorname{Var}Y\le(K-1)\mathbb{E}Y. E Y = Θ K ( B s / N ) , Var Y ≤ ( K − 1 ) E Y .
The variance bound follows from the negative correlation of fixed-size sampling indicators and 0 ≤ deg C ( x ) ≤ K − 1 0\le\deg_C(x)\le K-1 0 ≤ deg C ( x ) ≤ K − 1 . If a class has a a a inputs in A ˉ \bar{A} A ˉ , b b b in C C C , and seed occupancy X X X , its contribution to Y − H ( I ) Y-H(I) Y − H ( I ) is b X 1 { X ≥ 2 } ≤ b X ( X − 1 ) bX\mathbf{1}_{\{X\ge2\}}\le bX(X-1) b X 1 { X ≥ 2 } ≤ b X ( X − 1 ) . Summing the hypergeometric second factorial moments therefore gives
E ( Y − H ( I ) ) = O K ( B s 2 / N 2 ) = O K ( s / N ) E Y . \mathbb{E}(Y-H(I))=O_K(Bs^2/N^2)=O_K(s/N)\mathbb{E}Y. E ( Y − H ( I )) = O K ( B s 2 / N 2 ) = O K ( s / N ) E Y .
Chebyshev’s inequality for Y Y Y and Markov’s inequality for Y − H ( I ) Y-H(I) Y − H ( I ) imply H ( I ) ≥ E Y / 2 H(I)\ge\mathbb{E}Y/2 H ( I ) ≥ E Y /2 except with probability O K ( 1 / E Y + s / N ) O_K(1/\mathbb{E}Y+s/N) O K ( 1/ E Y + s / N ) . Since B s / N ≥ L / c K → ∞ Bs/N\ge L/c_K\to\infty B s / N ≥ L / c K → ∞ , the partner count is at least c K B s / N + K L c_KBs/N+K L c K B s / N + K L with probability 1 − o ( 1 ) 1-o(1) 1 − o ( 1 ) after choosing the constant in the constraint on L L L sufficiently small. Removing one selected value removes at most K K K candidates, so the same lower bound needed by each subsequent search holds.
The partner searches are controlled by the seed. Averaging over their outcomes leaves the diagonal seed distribution unchanged. Moreover, the later map is injective in the relevant information: I = F ∪ anchors ( S ) I=F\cup\operatorname{anchors}(S) I = F ∪ anchors ( S ) . Thus a bad-retention event for F F F cannot acquire more unconditional probability than the corresponding bad event for I I I . This justifies combining the population estimates with the measurements and validity flags in the actual preparation procedure. □
Lemma A.14 (Projection onto the implemented stationary state). Suppose the implemented outer walk has stationary distribution π \pi π pointwise comparable, within K K K -dependent constants, to the uniform distribution on its outer lists, and its stationary marked mass is at most a fixed constant smaller than one. From the uniform unmarked state one can prepare the implemented stationary walk state with constant success probability. In the parameter range used here the additional query cost is O ~ K ( Q ) \widetilde{O}_K(Q) O K ( Q ) .
Proof. Let ρ S \rho_S ρ S be the good-buffer probability for an outer list S S S . The inner analysis gives, uniformly in S S S ,
ρ S = Θ K ( min { 1 , μ } ) , μ = B s 2 / N 2 . \rho_S=\Theta_K(\min\{1,\mu\}),\qquad\mu=Bs^2/N^2. ρ S = Θ K ( min { 1 , μ }) , μ = B s 2 / N 2 .
First prepare the uniform queried inner buffer, at cost O ( s ) O(s) O ( s ) . If μ \mu μ is small, use an ordinary Grover amplification with one common safe iteration count and then measure the good-buffer flag. The iteration count is chosen using uniform upper and lower bounds on ρ S \rho_S ρ S , so it does not overshoot: successful amplitudes are positive and bounded below by a fixed constant on every S S S . If ρ S \rho_S ρ S is already bounded below, merely measure this flag. The reflection about the uniform inner state is implemented by the ordinary Johnson walk. The cost is
O ~ K ( s + s min { 1 , μ } ) . \widetilde{O}_K\left(s+\frac{\sqrt{s}}{\sqrt{\min\{1,\mu\}}}\right). O K ( s + min { 1 , μ } s ) .
This use of a positive successful amplitude is deliberate; arbitrary uncontrolled S S S -dependent phases would not justify a stationary overlap bound.
Consequently, after a successful flag the outer coefficients on the unmarked lists are positive and comparable to uniform coefficients, and each list carries its normalized good-buffer state. Apply the query-free local diffusion to put this state in the row-star space of the implemented walk. Pointwise comparison of π \pi π with uniform measure, together with the constant unmarked mass, gives constant overlap with the stationary walk state.
For the lazy reversible walk, phase estimation projects onto this state at cost O ~ ( U / δ ) \widetilde{O}(U/\sqrt{\delta}) O ( U / δ ) . Within the row-star space, irreducibility makes the stationary state the only zero-phase component; zero-phase states orthogonal to both star spaces have zero initial overlap. The inner reflection, as in the nested-update construction of [16 ] , and the outer phase estimation are implemented to inverse-polynomial precision, which adds only logarithmic factors.
The update and gap bounds established for the buffer construction are
U = O ~ K ( s min { 1 , μ } ) , δ = Ω K ( max { 1 , μ } L ) . U = \widetilde{O}_{K}\left(\frac{\sqrt{s}}{\sqrt{\min\{1,\mu\}}}\right), \qquad\delta= \Omega_{K}\left(\frac{\max\{1,\mu\}}{L}\right). U = O K ( min { 1 , μ } s ) , δ = Ω K ( L max { 1 , μ } ) .
They give
U δ = O ~ K ( N L B s ) ≤ O ~ K ( Q ) . \frac{U}{\sqrt{\delta}} = \widetilde{O}_{K}\left(\frac{N\sqrt{L}}{\sqrt{Bs}}\right) \leq\widetilde{O}_{K}(Q). δ U = O K ( B s N L ) ≤ O K ( Q ) .
The last inequality follows from the parameter bounds, or from M L / B = O K ( 1 ) ML/B = O_{K}(1) M L / B = O K ( 1 ) and N / M s = O K ( Q ) N/\sqrt{M}s = O_{K}(Q) N / M s = O K ( Q ) . The other displayed initialization costs are also O ~ K ( Q ) \widetilde{O}_{K}(Q) O K ( Q ) . Thus a constant number of attempts suffices in expectation; a constant-factor cutoff gives a bounded-query procedure with constant success probability. □
Bounded list sizes. If the unrounded optimal list size remains bounded, use L = 1 L=1 L = 1 and the full, uncolored pool of equal pairs. At these parameters μ = B s 2 / N 2 = o ( 1 ) \mu= Bs^{2}/N^{2} = o(1) μ = B s 2 / N 2 = o ( 1 ) for s = Θ K ( Q ) s = \Theta_{K}(Q) s = Θ K ( Q ) . Prepare a uniform queried s s s -subset and search for a subset containing exactly one equal pair. After coherently extracting that pair, the joint state differs by
O K ( B s 2 / N 2 + s / N ) = o ( 1 ) O_{K}\left(\sqrt{Bs^{2}/N^{2}+s/N}\right)=o(1) O K ( B s 2 / N 2 + s / N ) = o ( 1 )
from a uniform pair together with an unrestricted uniform residual buffer avoiding its two endpoints. Indeed, under the latter ideal state another disjoint pair occurs with probability O K ( B s 2 / N 2 ) O_{K}(Bs^{2}/N^{2}) O K ( B s 2 / N 2 ) , and another pair overlapping the selected pair occurs with probability O K ( s / N ) O_{K}(s/N) O K ( s / N ) . Projection onto the complement of these events gives exactly the extracted state, by the bijection between an exactly-one-pair subset and its pair and residual subset. Unquery and erase the residual buffer, and subsequently initialize the inner data as in Lemma A.14 . The query cost is
O ~ K ( s + N B s ) = O ~ K ( Q ) , \widetilde{O}_{K}\left(s+\frac{N}{\sqrt{Bs}}\right)=\widetilde{O}_{K}(Q), O K ( s + B s N ) = O K ( Q ) ,
because B ≥ M B \geq M B ≥ M and Q 3 ≥ N 2 / M Q^{3} \geq N^{2}/M Q 3 ≥ N 2 / M .
If M / B M/B M / B is at least a sufficiently small fixed constant, measuring this almost uniform pair and running the third-input search already solves the problem in O ~ K ( Q ) \widetilde{O}_{K}(Q) O K ( Q ) queries. Otherwise the constant can be chosen so that B B B is large enough for the strict one-pair exchange estimate, and the stationary marked mass is bounded away from one. The preceding stationary projection and search apply. This also treats the finite-candidate cases without presuming that a nontrivial strict-exchange graph exists.
Combining the partner-selection cost, the preliminary third-input test, and Lemma A.14 , the outer initialization in the regime Q > N Q > \sqrt{N} Q > N has total cost
O ~ K ( s + N L B s + N + Q ) = O ~ K ( Q ) . \widetilde{O}_{K}\left(s+\frac{NL}{\sqrt{Bs}}+\sqrt{N}+Q\right)=\widetilde{O}_{K}(Q). O K ( s + B s N L + N + Q ) = O K ( Q ) .
All constructions use only function queries and the values already stored by the algorithm.
Completion of the three-collision search bound
We now combine the exact update with its implementation costs. In this subsection B B B again denotes the original total pair count. A cross-pair pool constructed by initialization has size B ′ = Θ K ( B ) B' = \Theta_K(B) B ′ = Θ K ( B ) and contains pairs from M ′ = Θ K ( M ) M' = \Theta_K(M) M ′ = Θ K ( M ) target classes on the successful constant-probability preparation event. The estimates in the preceding sections are uniform over every such pool and its fixed mask.
Proof of Theorem A.1. Theorem 8.4 covers Q ≤ N Q \le\sqrt{N} Q ≤ N . Suppose Q > N Q > \sqrt{N} Q > N . Take the constant-scaled parameters from Subsection A.2 and put μ = B s 2 / N 2 \mu= Bs^2/N^2 μ = B s 2 / N 2 . The constants defining the batch window can be chosen uniformly for B ′ = Θ K ( B ) B' = \Theta_K(B) B ′ = Θ K ( B ) .
First dispose of a harmless bounded-ratio case. If M / B M/B M / B is bounded below by a sufficiently small positive constant depending only on K K K , then B = Θ K ( M ) B = \Theta_K(M) B = Θ K ( M ) and Q > N Q > \sqrt{N} Q > N implies B = O K ( N ) B = O_K(\sqrt{N}) B = O K ( N ) . Prepare a nearly uniform equal pair by the sparse-buffer construction of Subsection A.5, and measure it. The probability that its class has size at least three is Ω K ( M / B ) = Ω K ( 1 ) \Omega_K(M/B) = \Omega_K(1) Ω K ( M / B ) = Ω K ( 1 ) . A Grover search for a third input costs O ( N ) = O ( Q ) O(\sqrt{N}) = O(Q) O ( N ) = O ( Q ) , while pair preparation costs at most O ~ K ( Q ) \widetilde{O}_K(Q) O K ( Q ) . Constant repetition proves the claim. This case includes bounded B B B .
In the remaining case M / B M/B M / B is sufficiently small. For unbounded L 0 L_0 L 0 , choose s = a K Q s = a_KQ s = a K Q and L = b K L 0 L = b_KL_0 L = b K L 0 with suitable fixed constants. For bounded L 0 L_0 L 0 , use L = 1 L = 1 L = 1 instead. Lemma A.2 implies, uniformly in these regimes,
s , L = o ( N ) , min ( M , L ) s N = O K ( 1 ) , M s 3 N 3 = o ( min ( 1 , μ ) ) . s,L = o(N), \qquad\frac{\min(M,L)s}{N} = O_K(1), \qquad\frac{Ms^3}{N^3} = o(\min(1,\mu)). s , L = o ( N ) , N min ( M , L ) s = O K ( 1 ) , N 3 M s 3 = o ( min ( 1 , μ )) .
The scaled constants ensure B ′ ≥ 2 W K L B' \ge2W_KL B ′ ≥ 2 W K L and m max L / B ′ ≤ 1 / ( 100 W K ) m_{\max}L/B' \le1/(100W_K) m m a x L / B ′ ≤ 1/ ( 100 W K ) . If the unscaled L 0 L_0 L 0 is not bounded, μ / L = o ( 1 ) \mu/L = o(1) μ / L = o ( 1 ) , so the window in Lemma A.6 lies within { 1 , … , min ( L , ⌊ s / 2 ⌋ ) } \{1,\ldots,\min(L,\lfloor s/2\rfloor)\} { 1 , … , min ( L , ⌊ s /2 ⌋)} for large N N N . If L 0 L_0 L 0 is bounded, then μ = o ( 1 ) \mu= o(1) μ = o ( 1 ) and the window is simply { 1 } \{1\} { 1 } . Its smallest and largest batch sizes have order max ( 1 , μ ) \max(1,\mu) max ( 1 , μ ) , with constants depending only on K K K .
Lemma A.3 realizes the exact local diffusion and garbage swap, using no new queries. Lemma A.5 verifies the normalizer hypothesis of Theorem A.11 . The lazy outer walk therefore satisfies
δ = Ω K ( max ( 1 , μ ) L ) , ϵ = Θ K ( M L B ) , π ≲ K π 0 . \delta= \Omega_K\left(\frac{\max(1,\mu)}{L}\right), \qquad\epsilon= \Theta_K\left(\frac{ML}{B}\right), \qquad\pi\lesssim_K \pi_0. δ = Ω K ( L max ( 1 , μ ) ) , ϵ = Θ K ( B M L ) , π ≲ K π 0 .
Here M L / B ML/B M L / B is at most a sufficiently small fixed constant by the choice of b K b_K b K (and by the bounded-ratio case just removed when L = 1 L = 1 L = 1 ). In particular, the unmarked stationary mass is bounded below. All inner marked-state reflections have cost
U = O ~ K ( s min ( 1 , μ ) ) . U = \widetilde{O}_K\left(\sqrt{\frac{s}{\min(1,\mu)}}\right). U = O K ( min ( 1 , μ ) s ) .
The checking procedure searches the original domain for an input whose function value is represented by the pair list and whose index is outside its stored endpoints. Because the list uses distinct classes, this is exactly the condition that some stored pair belongs to a target class. It costs C = O ~ ( N ) C = \widetilde{O}(\sqrt{N}) C = O ( N ) ; a successful check also supplies a verifiable third input.
Subsection A.5 prepares the actual stationary walk state, with constant success probability, within
O ~ K ( s + N L B s + N + U δ ) . \widetilde{O}_K\left(s + \frac{NL}{\sqrt{B}s} + \sqrt{N} + \frac{U}{\sqrt{\delta}}\right). O K ( s + B s N L + N + δ U ) .
The first three terms are O ~ K ( Q ) \widetilde{O}_K(Q) O K ( Q ) , and the fourth is dominated by the eventual search cost. Apply the finding theorem for nested updates [16 Theorem 3.3] . Since
min ( 1 , μ ) max ( 1 , μ ) = μ , (182) \min(1,\mu)\max(1,\mu) = \mu, \tag*{(182)} min ( 1 , μ ) max ( 1 , μ ) = μ , (182)
the update contribution becomes
U ε δ = O ~ K ( B s M μ ) = O ~ K ( N M s ) \frac{U}{\sqrt{\varepsilon\delta}}=\widetilde{O}_K\left(\sqrt{\frac{Bs}{M\mu}}\right)=\widetilde{O}_K\left(\sqrt{\frac{N}{Ms}}\right) ε δ U = O K ( M μ B s ) = O K ( M s N )
and the checking contribution is
C ε = O ~ K ( N B ‾ M L ) . \frac{C}{\sqrt{\varepsilon}}=\widetilde{O}_K\left(\sqrt{\frac{N\overline{B}}{ML}}\right). ε C = O K M L N B .
Lemma A.2 bounds both by O ~ K ( Q ) \widetilde{O}_K(Q) O K ( Q ) . The initialization, search, and reported third-input verification have a fixed positive total success probability. Independent repetition raises it to 2/3. All subroutines have fixed query budgets. Any expected-cost preparation is cut off at a sufficiently large constant multiple of its expectation; Markov’s inequality preserves a fixed positive success probability. They use only queries to f f f and classical information obtained from those queries. None of the bounds depends on q q q beyond realizability.
What the argument establishes. The new step is the choice to extract every eligible pair in a good inner buffer. The resulting symmetric residual state determines a reversible outer chain with a generally nonuniform stationary distribution. Counting and initialization are adapted to that distribution rather than identifying it with a Johnson stationary state. This proves a three-collision search upper bound with the two type-dependent scales in Theorem A.1 . Extending the construction to higher collision orders requires additional levels of coherent data and their own update analysis.