Ancillary Representation: A Multidisciplinary View

• Ancillary representation is a strategy that introduces auxiliary objects—ranging from qubits to aggregation trees—to reframe and simplify complex systems.
• It enables non-unitary dynamics in quantum control, supports ancilla-mediated quantum computation, and enhances statistical inference through conditional maps and geometric structures.
• The method leverages auxiliary degrees of freedom to optimize performance metrics such as cooling efficiency, error reduction, and phase sensitivity in diverse applications.

Ancillary representation denotes a family of constructions in which an auxiliary system, auxiliary space, or auxiliary information source is introduced so that the primary object of interest can be manipulated, conditioned, or described through an enlarged formal structure. In the literature surveyed here, this includes shared-ancilla Hilbert-space models for measurement-based cooling, qudit and spin-coherent phase-space buses for quantum computation, collective-spin or double-well ancillas that reshape effective dynamics, quantile-based ancillary contours in likelihood theory, aggregation trees built from side-information distances in multiple testing, auxiliary-task families for representation learning in reinforcement learning, and language agents that act as digital representatives in collective decision-making (Yan et al., 2022, Proctor et al., 2014, Dooley et al., 2014, Testa et al., 2023, Fraser et al., 2010, Li et al., 2024, Farebrother et al., 2023, Jarrett et al., 13 Feb 2025, Fan et al., 2020, Chen et al., 21 Jan 2026).

1. Scope and principal forms

The surveyed literature uses ancillary representation in several non-equivalent but structurally related ways. In each case, an auxiliary object is not the principal target of inference or control, yet it supplies the effective coordinates, conditional maps, or side-information structure through which the primary problem becomes tractable.

Domain	Ancillary object	Operational role
Quantum cooling and control	Qutrit or qubit ancilla	Induces non-unitary maps or modifies open-system dynamics
Ancilla-mediated computation	Qudit or spin-ensemble ancilla	Implements gates through controlled displacements and geometric phases
Quantum many-body engineering	Collective spins or double-well ancilla	Generates effective nonlinearities or conditional Hamiltonians
Likelihood and multiple testing	Ancillary contour or aggregation tree	Organizes conditioning or side-information borrowing
Representation learning and social choice	Auxiliary tasks or language agents	Learns reusable features or proxies individual behavior

In quantum settings, the ancillary object is usually an explicit subsystem with its own Hilbert space, coupled coherently or measured projectively. In statistical settings, the ancillary object is instead a representation of structure in sample space or hypothesis space: a contour, a tree, or a covariate-derived geometry. In machine learning and collective decision-making, ancillary representation appears as a support mechanism for learning or delegation rather than as a physical subsystem.

Taken together, these works suggest that ancillary representation is best understood functionally. The ancillary component enlarges the state space, phase space, or information space; the primary task is then solved through a reduced, conditional, or projected description on the original target system.

2. Measured ancillas and reduced dynamics in quantum control

A particularly explicit formulation appears in simultaneous cooling of two bosonic resonators by repeated measurements on a single ancillary $V$-type qutrit. The total Hilbert space is

$$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$

with ancillary basis $\{|g\rangle,|e\rangle,|f\rangle\}$, where $|g\rangle\leftrightarrow |e\rangle$ couples only to mode $a$ and $|g\rangle\leftrightarrow |f\rangle$ couples only to mode $b$. In the rotating frame, the interaction-picture Hamiltonian is

$$H_I=\delta_e|e\rangle\langle e|+\delta_f|f\rangle\langle f| +g_a(a^\dagger\sigma_{eg}^-+a\sigma_{eg}^+) +g_b(b^\dagger\sigma_{fg}^-+b\sigma_{fg}^+),$$

and repeated projective measurements of the ancilla onto $|g\rangle$ induce the non-unitary resonator map

$V_g(\tau)\equiv \langle g|U(\tau)|g\rangle.$

Because the same ground state $$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$0 is shared by both Jaynes–Cummings channels, a single measurement acts collectively on both resonators. In the two-mode Fock basis, $$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$1 is diagonal with coefficients $$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$2, and the protocol suppresses all excited populations while leaving $$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$3 fixed. The main analytical quantity is the collective thermal Rabi frequency

$$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$4

which yields the optimized interval $$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$5. Under the unequal-time strategy $$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$6, the average population can be reduced by about $$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$7 orders of magnitude in only $$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$8 successful measurements, and the scheme extends to $$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$9 resonators by replacing the qutrit with a multi-level ancilla that has one common ground state and $\{|g\rangle,|e\rangle,|f\rangle\}$0 excited states (Yan et al., 2022).

A closely related but dynamically distinct use of an ancilla appears in open-system quantum speed limits. There the ancillary system is a second qubit $\{|g\rangle,|e\rangle,|f\rangle\}$1, coherently coupled to the system qubit $\{|g\rangle,|e\rangle,|f\rangle\}$2 by

$\{|g\rangle,|e\rangle,|f\rangle\}$3

The ancilla is not traced out as part of the environment, but its hopping interaction changes the reduced evolution of qubit $\{|g\rangle,|e\rangle,|f\rangle\}$4 in Lorentzian non-Markovian reservoirs. For independent environments, a common environment, and the case where only the system couples to the bath, the reduced state of qubit $\{|g\rangle,|e\rangle,|f\rangle\}$5 remains diagonal with excited-state amplitude $\{|g\rangle,|e\rangle,|f\rangle\}$6, and the quantum speed limit time is written in terms of $\{|g\rangle,|e\rangle,|f\rangle\}$7 and $\{|g\rangle,|e\rangle,|f\rangle\}$8. Across all three configurations, increasing the system–environment coupling shortens the speed limit time, while increasing $\{|g\rangle,|e\rangle,|f\rangle\}$9 produces periodic oscillations. The common-environment configuration yields the shortest quantum speed limit time, showing that ancillary representation here means embedding the system in a larger coherent structure whose internal exchange and shared environmental pathways accelerate the reduced dynamics (Fan et al., 2020).

3. Phase-space ancillas in ancilla-mediated quantum computation

In ancilla-mediated quantum computation, ancillary representation is realized as a phase-space description of register-register interactions. A $|g\rangle\leftrightarrow |e\rangle$0-dimensional ancillary qudit has discrete phase space $|g\rangle\leftrightarrow |e\rangle$1, generalized Pauli operators $|g\rangle\leftrightarrow |e\rangle$2 and $|g\rangle\leftrightarrow |e\rangle$3, and displacement operator

$|g\rangle\leftrightarrow |e\rangle$4

Controlled displacements between a register qubit and the ancilla,

$|g\rangle\leftrightarrow |e\rangle$5

move the ancilla around closed loops in discrete phase space. The resulting geometric phase implements controlled-phase gates while returning the ancilla to its initial state. The basic loop identity,

$|g\rangle\leftrightarrow |e\rangle$6

is the discrete analogue of qubus geometry. Because multiple controlled displacements can be concatenated before the ancilla is disentangled, the extra degrees of freedom of the ancilla reduce ancilla-register interaction counts for structured gate sequences (Proctor et al., 2014).

The same paper develops a spin-ensemble ancillary model based on spin coherent states. Here the ancilla is an ensemble of $|g\rangle\leftrightarrow |e\rangle$7 spins restricted to the symmetric subspace, with displacement operator on the Bloch sphere

$|g\rangle\leftrightarrow |e\rangle$8

Closed loops on the Bloch sphere accumulate spherical geometric phases, and the controlled version

$|g\rangle\leftrightarrow |e\rangle$9

again mediates entangling gates while returning the ancilla to $a$0. In the large-$a$1 limit,

$a$2

so the spin-ensemble ancilla converges to the continuous-variable qubus architecture. Ancillary representation is therefore literal in this setting: multi-qubit gates are represented as paths in the ancilla’s phase space, and the effective gate is determined by loop geometry rather than by direct inter-register coupling (Proctor et al., 2014).

4. Ancillas as effective interaction landscapes in state engineering, tunneling, and metrology

A different use of ancillary representation appears when the ancilla is itself a many-body resource that converts bilinear couplings into effective nonlinear dynamics. In a model of two collective spin ensembles $a$3 and $a$4, the ancillary system $a$5 is a symmetric spin of size $a$6 coupled to the primary ensemble $a$7 through

$a$8

Depending on the initial spin coherent state and ancilla preparation, the evolution produces spin cat states, spin squeezed states, or multiple cat states. In the $a$9 regime, the approximate Hamiltonian

$|g\rangle\leftrightarrow |f\rangle$0

yields one-axis twisting on $|g\rangle\leftrightarrow |f\rangle$1 when the ancilla is initialized in a $|g\rangle\leftrightarrow |f\rangle$2 eigenstate. For maximal $|g\rangle\leftrightarrow |f\rangle$3, the effective strength scales as $|g\rangle\leftrightarrow |f\rangle$4, so cat-generation and squeezing times shrink by a factor $|g\rangle\leftrightarrow |f\rangle$5. The ancillary spin ensemble is therefore represented explicitly through collective SU(2) operators and Dicke states, but operationally it acts as an entangling bus and a tunable source of nonlinear dynamics (Dooley et al., 2014).

In catalyzed tunneling, the ancillary system is another double-well Bose–Hubbard system of the same physical type as the primary tunneling system. The joint Hilbert space

$|g\rangle\leftrightarrow |f\rangle$6

has dimension $|g\rangle\leftrightarrow |f\rangle$7, and the interaction

$|g\rangle\leftrightarrow |f\rangle$8

makes each ancilla Fock state $|g\rangle\leftrightarrow |f\rangle$9 define an effective system Hamiltonian

$b$0

in the single-particle limit. The ancilla basis therefore represents a family $b$1 of conditional detunings. By optimizing ancilla parameters $b$2, coupling $b$3, initial ancilla density matrix, and the evolution time with automatic differentiation and ADAM, the protocol can raise the maximum tunneling probability from $b$4 to $b$5 in the one-boson benchmark, and to $b$6 for all studied interacting pairs $b$7 in the noiseless many-body setting. Here ancillary representation is an enlarged configuration space and a learned representation of effective barriers and resonant channels (Testa et al., 2023).

A measurement-based metrology protocol with a spin ensemble and one ancillary qubit makes the same structural idea explicit in a weak-coupling XXZ model,

$b$8

After joint evolution and projective measurement of the ancilla in the $b$9 basis, the probe evolves along two parallel paths. Under the optimized weak-coupling condition

$$H_I=\delta_e|e\rangle\langle e|+\delta_f|f\rangle\langle f| +g_a(a^\dagger\sigma_{eg}^-+a\sigma_{eg}^+) +g_b(b^\dagger\sigma_{fg}^-+b\sigma_{fg}^+),$$0

and suitable $$H_I=\delta_e|e\rangle\langle e|+\delta_f|f\rangle\langle f| +g_a(a^\dagger\sigma_{eg}^-+a\sigma_{eg}^+) +g_b(b^\dagger\sigma_{fg}^-+b\sigma_{fg}^+),$$1, those two paths implement a superposition of $$H_I=\delta_e|e\rangle\langle e|+\delta_f|f\rangle\langle f| +g_a(a^\dagger\sigma_{eg}^-+a\sigma_{eg}^+) +g_b(b^\dagger\sigma_{fg}^-+b\sigma_{fg}^+),$$2 and $$H_I=\delta_e|e\rangle\langle e|+\delta_f|f\rangle\langle f| +g_a(a^\dagger\sigma_{eg}^-+a\sigma_{eg}^+) +g_b(b^\dagger\sigma_{fg}^-+b\sigma_{fg}^+),$$3, transforming polarized or thermal probe states into two-component states with large separation in the $$H_I=\delta_e|e\rangle\langle e|+\delta_f|f\rangle\langle f| +g_a(a^\dagger\sigma_{eg}^-+a\sigma_{eg}^+) +g_b(b^\dagger\sigma_{fg}^-+b\sigma_{fg}^+),$$4 eigenspace. The resulting QFI for optimized polarized probes is

$$H_I=\delta_e|e\rangle\langle e|+\delta_f|f\rangle\langle f| +g_a(a^\dagger\sigma_{eg}^-+a\sigma_{eg}^+) +g_b(b^\dagger\sigma_{fg}^-+b\sigma_{fg}^+),$$5

and asymptotic quadratic scaling persists for thermal probes. With parity detection on the ancilla or the probe, the phase sensitivity approaches the Heisenberg limit. In this setting the ancilla is a compact controller of evolution branches, and its unconditional measurement replaces direct generation of GHZ-like states or squeezing Hamiltonians (Chen et al., 21 Jan 2026).

5. Ancillary contours and ancillary trees in statistical inference

In classical likelihood theory, ancillary representation is formulated geometrically rather than physically. The central object is a sample-space contour whose conditional distribution is free of the parameter to second order. Starting from a vector quantile function

$$H_I=\delta_e|e\rangle\langle e|+\delta_f|f\rangle\langle f| +g_a(a^\dagger\sigma_{eg}^-+a\sigma_{eg}^+) +g_b(b^\dagger\sigma_{fg}^-+b\sigma_{fg}^+),$$6

one fixes the fitted reference value $$H_I=\delta_e|e\rangle\langle e|+\delta_f|f\rangle\langle f| +g_a(a^\dagger\sigma_{eg}^-+a\sigma_{eg}^+) +g_b(b^\dagger\sigma_{fg}^-+b\sigma_{fg}^+),$$7 determined by the observed data $$H_I=\delta_e|e\rangle\langle e|+\delta_f|f\rangle\langle f| +g_a(a^\dagger\sigma_{eg}^-+a\sigma_{eg}^+) +g_b(b^\dagger\sigma_{fg}^-+b\sigma_{fg}^+),$$8 and the MLE $$H_I=\delta_e|e\rangle\langle e|+\delta_f|f\rangle\langle f| +g_a(a^\dagger\sigma_{eg}^-+a\sigma_{eg}^+) +g_b(b^\dagger\sigma_{fg}^-+b\sigma_{fg}^+),$$9, and defines the intrinsic ancillary contour

$|g\rangle$0

A Taylor expansion of the quantile function yields tangent array $|g\rangle$1 and curvature array $|g\rangle$2, so locally

$|g\rangle$3

For scalar parameters the vector field integrates directly to ancillary contours; for vector parameters the Frobenius conditions need not hold globally, but the paper shows that they hold in a restricted way sufficient to verify second-order ancillary contours in moderate deviations. Exact ancillaries in location–scale and transformation models are recovered as special cases, while the general construction supports numerical or Monte Carlo calculation of $|g\rangle$4-values and confidence quantiles (Fraser et al., 2010).

In multiple testing, ancillary information is represented not by a statistic with parameter-free distribution but by an aggregation tree built from a distance matrix over hypotheses. DART2 assumes $|g\rangle$5 hypotheses and constructs

$|g\rangle$6

where $|g\rangle$7 and higher layers merge nearby hypotheses according to ancillary distances. For each node $|g\rangle$8, DART2 tests

$|g\rangle$9

using the Stouffer aggregate

$V_g(\tau)\equiv \langle g|U(\tau)|g\rangle.$0

The procedure combines a screening stage over the tree with a refining stage based on robust node-specific thresholds

$V_g(\tau)\equiv \langle g|U(\tau)|g\rangle.$1

where

$V_g(\tau)\equiv \langle g|U(\tau)|g\rangle.$2

The key design goal is robustness when ancillary information is helpful, non-informative, or misleading. Under the sparsity condition $V_g(\tau)\equiv \langle g|U(\tau)|g\rangle.$3 with $V_g(\tau)\equiv \langle g|U(\tau)|g\rangle.$4, the method asymptotically controls FDR, and when the tree is informative it improves power relative to BH. Here ancillary representation is explicitly geometric: side information becomes a tree, and the tree determines which groups are aggregated, screened, and refined (Li et al., 2024).

6. Auxiliary-task representations and digital representatives

In deep reinforcement learning, ancillary representation appears as representation learning by auxiliary tasks. Proto-Value Networks define a shared feature map $V_g(\tau)\equiv \langle g|U(\tau)|g\rangle.$5 and study the multi-task loss

$V_g(\tau)\equiv \langle g|U(\tau)|g\rangle.$6

where $V_g(\tau)\equiv \langle g|U(\tau)|g\rangle.$7 is an auxiliary-task matrix. The paper constructs a family of auxiliary tasks from the successor measure

$V_g(\tau)\equiv \langle g|U(\tau)|g\rangle.$8

which is the value function for binary reward $V_g(\tau)\equiv \langle g|U(\tau)|g\rangle.$9. With subsets $$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$00 drawn procedurally, auxiliary tasks form an essentially infinite source of information about the environment. The learned representation can therefore be understood as extending proto-value functions to deep reinforcement learning. In experiments on the Arcade Learning Environment, the resulting features support performance comparable to established algorithms using only linear approximation and a small number of interactions with the environment’s reward function. The ancillary component is not a separate physical system but a procedurally generated family of side tasks that determines the geometry of the representation space (Farebrother et al., 2023).

In collective decision-making, representation is formalized directly as proxy participation by a digital agent. The base model is an episodic process with participants $$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$01, outcome space $$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$02, policy profile $$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$03, mechanism $$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$04, and outcome map

$$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$05

Digital representation is defined by replacing a human policy $$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$06 with a model policy $$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$07 such that outcome-level behavior is preserved. The paper distinguishes conditional equivalence, transition-based equivalence, and trajectory or payoff-based equivalence, with the latter defined through repeated Bellman operators: $$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$08 In the empirical case study, language agents are fine-tuned to represent participants during the critique step of a consensus-finding mechanism, using base information plus optional past opinions, past critiques, demographic profile, and position scores. The best-performing variant uses Base+O+C, and fine-tuned models yield little or no degradation in payoff when substituting critiques. Here ancillary representation becomes digital representation: a learned proxy policy makes an absent individual’s preferences present in the process (Jarrett et al., 13 Feb 2025).

7. Structural motifs, divergences, and limitations

Taken together, these works suggest a recurring architecture. There is a primary system, dataset, or decision process; there is an ancillary object external to that primary target; there is a joint map on the enlarged structure; and there is a reduced, conditional, or projected effect on the primary object obtained by measurement, partial trace, conditioning, aggregation, or linear readout. The ancillary object may be a qutrit with a shared ground state, a qudit phase space, a collective spin, a second Bose–Hubbard system, a quantile contour, a hypothesis tree, an infinite auxiliary-task family, or a language-agent proxy (Yan et al., 2022, Proctor et al., 2014, Dooley et al., 2014, Testa et al., 2023, Fraser et al., 2010, Li et al., 2024, Farebrother et al., 2023, Jarrett et al., 13 Feb 2025).

The differences are equally important. Quantum-control and metrology papers treat ancillary representation as a concrete extension of Hilbert space and Hamiltonian dynamics; likelihood theory treats it as a sample-space partition with second-order invariance; multiple testing treats it as a geometry on hypotheses induced by distances; reinforcement learning treats it as a support mechanism for feature learning; and digital representation treats it as outcome-equivalent proxy behavior. A plausible implication is that ancillary representation is less a single technical object than a general strategy for relocating difficult structure into an auxiliary degree of freedom, where it can be measured, optimized, conditioned upon, or otherwise exploited.

The limitations are domain-specific. Measurement-based cooling assumes the rotating-wave approximation, ideal projective measurements, and neglect of decoherence during each interval (Yan et al., 2022). Quantum-speed-limit analysis is restricted to Lorentzian environments and the single-excitation sector (Fan et al., 2020). Ancilla-mediated quantum computation with finite-dimensional phase spaces inherits discrete-angle or finite-$$\mathcal{H}=\mathcal{H}_q\otimes\mathcal{H}_a\otimes\mathcal{H}_b,$$09 constraints (Proctor et al., 2014). Collective-spin state engineering relies on symmetric couplings and short-time approximations (Dooley et al., 2014). Tunneling catalysis remains sensitive to the learned ancillary Hamiltonian, even though the enhancement is reported to be minimally affected by noise and decoherence (Testa et al., 2023). Quantile-based ancillary contours require continuity and smoothness, and the vector-parameter case only satisfies a restricted form of Frobenius consistency (Fraser et al., 2010). DART2 is asymptotic and depends on the constructed tree, although it is designed to remain robust when the side information is misleading (Li et al., 2024). Proto-Value Networks exhibit a network-capacity versus task-count trade-off (Farebrother et al., 2023). Digital representatives are studied primarily for simulation and mechanism design rather than as direct replacements for human accountability, and the paper explicitly identifies risks of misrepresentation, bias amplification, and unequal representation quality (Jarrett et al., 13 Feb 2025).

Across these literatures, ancillary representation remains a unifying but context-dependent idea: an auxiliary entity is introduced not as the final object of interest, but as the representation through which the effective dynamics, inference problem, or collective process becomes manipulable.