Coherent Rank-Select in Quantum Algorithms

• The paper introduces a coherent reversible rank-select primitive that maps entangled validity masks to valid bit positions with a sentinel totalization rule to handle out-of-range ranks.
• It employs both a sequential bounded-span scan and a blocked construction to optimize gate count and circuit depth, achieving bounds like O(Nw) and O(N log w).
• The construction integrates into coherent rollout oracles, reversibly mapping basis-state selectors to actions, and plays a role analogous to classical rank-select in succinct indexing.

Searching arXiv for the cited papers to ground the article in current arXiv records. Coherent rank-select is a reversible basis-state selection primitive for quantum algorithms on branch-dependent action spaces. In the formulation studied in "Coherent Rollout Oracles for Finite-Horizon Sequential Decision Problems" (Shukla, 28 Apr 2026), a quantum state register determines an entangled $N$-bit validity mask, a selector register stores a rank $r$, and the primitive returns the position of the $r$-th valid bit or a sentinel when $r$ is out of range. The construction is motivated by coherent rollout for sequential decision problems, where randomness must live in explicit quantum registers and basis-state selectors must be mapped to actions reversibly.

1. Formal definition

Let $\lvert s\rangle$ be a state register whose content defines a validity mask

$M_s=(M_s(0),M_s(1),\dots,M_s(N-1))\in\{0,1\}^N,$

where $M_s(i)=1$ iff action $i$ is valid in branch $s$. Let $\lvert r\rangle$ be a rank register, and let the output register have width

$r$0

so that it can encode positions $r$1, with $r$2 used as a sentinel (Shukla, 28 Apr 2026).

The primitive is defined through the partial function

$r$3

where $r$4, and

$r$5

The coherent rank-select unitary $r$6 acts as

$r$7

A central structural feature is totalization by sentinel: although the underlying notion is the position of the $r$8-th valid bit, the unitary must be total on computational basis states, so the output $r$9 encodes failure. This removes partiality without discarding reversibility.

2. Circuit model and resource measures

The analysis in (Shukla, 28 Apr 2026) distinguishes two reversible-circuit models. In the bounded-fan-in, long-range model, gates act on at most $r$0 qubits, but any subset of qubits may interact. The primary resources are gate count $r$1, ancilla qubits $r$2, and circuit depth $r$3. In the bounded-span scan model, validity bits are laid out on a line at positions $r$4, and each gate of support size at most $r$5 may cross at most $r$6 adjacent prefix cuts. The same measures $r$7, $r$8, and $r$9 are counted, together with total crossing mass of span.

The persistent registers are the state register $r$0, the rank selector $r$1 of width $r$2, the output register $r$3 of width $r$4, and an ancilla pool of size $r$5, reused and cleaned (Shukla, 28 Apr 2026).

This operational setting is stricter than classical selection over a stored mask. The mask may be entangled with the state register, and the implementation must preserve unitarity, return all ancillae to $r$6, and avoid implicit randomness. The abstract of (Shukla, 28 Apr 2026) states this requirement directly: coherent quantum rollout requires a unitary simulator, randomness must live in explicit quantum registers, and basis-state selectors must be mapped to actions reversibly.

3. Sequential bounded-span scan

The first construction in (Shukla, 28 Apr 2026) is a left-to-right reversible scan. It maintains a $r$7-bit running count $r$8 of previously seen $r$9-bits and initializes the output register to the sentinel $\lvert s\rangle$0. For each position $\lvert s\rangle$1, it reads the entangled validity bit $\lvert s\rangle$2, compares $\lvert s\rangle$3 with $\lvert s\rangle$4, forms a match flag $\lvert s\rangle$5, conditionally XORs $\lvert s\rangle$6 into the output, uncomputes the comparison, and conditionally increments $\lvert s\rangle$7. After the forward pass, the increment chain is reversed to uncompute $\lvert s\rangle$8.

Per iteration, the construction uses one $\lvert s\rangle$9-bit reversible equality check with $M_s=(M_s(0),M_s(1),\dots,M_s(N-1))\in\{0,1\}^N,$0 Toffoli/CNOT gates and depth $M_s=(M_s(0),M_s(1),\dots,M_s(N-1))\in\{0,1\}^N,$1, one multi-target XOR of $M_s=(M_s(0),M_s(1),\dots,M_s(N-1))\in\{0,1\}^N,$2 CNOTs and depth $M_s=(M_s(0),M_s(1),\dots,M_s(N-1))\in\{0,1\}^N,$3, and one $M_s=(M_s(0),M_s(1),\dots,M_s(N-1))\in\{0,1\}^N,$4-bit ripple-carry increment with $M_s=(M_s(0),M_s(1),\dots,M_s(N-1))\in\{0,1\}^N,$5 gates and depth $M_s=(M_s(0),M_s(1),\dots,M_s(N-1))\in\{0,1\}^N,$6. The uncompute steps exactly reverse the same cost (Shukla, 28 Apr 2026).

The resulting resource bounds are

$M_s=(M_s(0),M_s(1),\dots,M_s(N-1))\in\{0,1\}^N,$7

The paper further states that this sequential scan is gate-optimal in the bounded-span, bounded-fan-in model. The proof sketch uses a prefix-suffix communication argument: varying the first $M_s=(M_s(0),M_s(1),\dots,M_s(N-1))\in\{0,1\}^N,$8 bits of the mask can force the remainder of the circuit to reveal $M_s=(M_s(0),M_s(1),\dots,M_s(N-1))\in\{0,1\}^N,$9 bits, each cut-crossing gate transmits at most $M_s(i)=1$0 bits, and summing over the $M_s(i)=1$1 cuts yields an $M_s(i)=1$2 lower bound (Shukla, 28 Apr 2026).

4. Blocked construction and lower bounds

When long-range gates are allowed, (Shukla, 28 Apr 2026) gives a blocked construction with asymptotically smaller gate count. Positions are partitioned into $M_s(i)=1$3 blocks of size $M_s(i)=1$4. A global prefix sum $M_s(i)=1$5 and global output register are maintained. For each block $M_s(i)=1$6, the circuit computes the block population

$M_s(i)=1$7

by a balanced reversible adder tree, computes a take flag

$M_s(i)=1$8

and local rank $M_s(i)=1$9, runs a sequential scan within the block to obtain a local index $i$0, conditionally XORs $i$1 into the output, updates $i$2, and then uncomputes the temporary data.

The corresponding bounds are

$i$3

with long-range fan-in used to build the block adder tree in depth $i$4 (Shukla, 28 Apr 2026).

The same work also proves an unconditional lower bound in the broader bounded-fan-in setting: any reversible circuit for $i$5 must have $i$6. The proof sketch considers rank $i$7: the all-zero mask produces output $i$8, while the singleton mask $i$9 produces output $s$0. Hence every input bit $s$1 influences the output, and in a $s$2-fan-in reversible circuit each gate can bring in at most $s$3 new inputs into the backward light cone of an output qubit, forcing $s$4 gates.

Construction or bound	Model	Resource statement
Sequential scan	Bounded-span, bounded-fan-in	$s$5, $s$6, $s$7
Blocked construction	Bounded-fan-in with long-range gates	$s$8, $s$9, $\lvert r\rangle$0
Unconditional lower bound	Any bounded-fan-in model	$\lvert r\rangle$1

These results isolate a basic trade-off. Bounded-span locality enforces an $\lvert r\rangle$2 gate barrier, while long-range interactions permit an $\lvert r\rangle$3 construction but do not remove the $\lvert r\rangle$4 dependence on the mask length.

5. Composition into coherent rollout oracles

In (Shukla, 28 Apr 2026), coherent rank-select is not treated as an isolated primitive; it is the indexing phase of a larger coherent rollout oracle for finite-horizon planning problems. For each round $\lvert r\rangle$5, the oracle composes three phases.

First, $\lvert r\rangle$6 decodes $\lvert r\rangle$7 to a valid action position. Second, a reversible transition circuit $\lvert r\rangle$8 updates the configuration using fresh randomness dice registers that are never discarded, so $\lvert r\rangle$9 remains available; the stated cost is $r$00 gates per round for bounded-neighbors. Third, a reversible predicate $r$01 writes the payoff into one qubit; the stated cost is $r$02 if the task compares two $r$03-bit counts over $r$04 cells (Shukla, 28 Apr 2026).

The total unitary is

$r$05

with total gate and qubit costs

$r$06

The paper states that this oracle satisfies the access model of the best-arm pipeline of Wang et al. Plugging it into amplitude estimation and Dürr–Høyer maximum finding gives an $r$07-correct best-arm algorithm using

$r$08

calls to $r$09 and $r$10, against the classical $r$11 rollout-query lower bound (Shukla, 28 Apr 2026). The same work gives a bounded-influence lifting theorem extending the lower-bound construction from a base configuration to an exponential family of configurations, instantiates the framework on SIR epidemic intervention, includes a stochastic placement-game sanity check, and machine-checks the main results in Lean 4.

6. Relation to classical rank-select structures

Classical rank and select are basic operations on sequences and bitvectors, with applications in compressed text indexes and other space-efficient data structures (Chiu et al., 13 Apr 2026, Chiu et al., 8 Sep 2025). In classical notation for bitvectors, for $r$12,

$r$13

and these operations form the backbone of more complex structures such as wavelet trees (Chiu et al., 8 Sep 2025).

The relationship to coherent rank-select is structural rather than merely terminological. Wavelet forests, as revisited in (Chiu et al., 13 Apr 2026), partition the input sequence into fixed-size blocks, build a separate pointerless canonical Huffman-shaped wavelet tree for each block, and support rank by a three-tier sampled-counts strategy over hyperblocks, superblocks, and blocks. The same paper extends the structure to select queries, reusing the rank-sampling arrays, with select time $r$14 and rank time $r$15, while experiments show that wavelet forests are competitive with, and in most cases outperform, standalone wavelet-tree implementations.

At the bitvector level, "Engineering Select Support for Hybrid Bitvectors" (Chiu et al., 8 Sep 2025) augments a hybrid bitvector that already supports constant-time rank so that it also supports constant-time select while using only $r$16 additional bits beyond the concatenation of the per-block compressed bitstreams. The implementation partitions the bitvector into fixed-size blocks, groups them into superblocks and hyperblocks, and answers $r$17 through a sparse sampling of superblocks plus a block-local select routine specialized to plain, run-length, or minority encoding.

These classical constructions and the coherent primitive solve different problems. Classical designs optimize succinct storage and local decoding on static sequences or bitvectors, whereas coherent rank-select is a reversible unitary on an entangled validity mask with a sentinel totalization rule (Shukla, 28 Apr 2026). A plausible implication is that the quantum primitive occupies, for reversible decision pipelines, a role analogous to the role classical rank/select occupies in succinct indexing: it is a low-level access mechanism on which larger algorithmic frameworks are built.