Dressed Reparametrizations: Theory & Applications

• Dressed reparametrizations are variable transformations enhanced by additional structures (e.g., soft-photon clouds, equivalent metrics) that ensure physical or statistical invariance.
• They embed constraints such as gauge conditions, metric consistency, and input-output equivalence directly into the reparameterization, improving both theoretical formulation and practical outcomes.
• By coupling the bare change of variables with complementary structures—from Wilson lines in gauge theories to balanced control in qubit optimization—they optimize performance across diverse scientific domains.

“Dressed reparametrizations” (Editor’s term) denotes a family of constructions in which a nominal reparameterization is not a bare change of coordinates or variables, but a re-expression accompanied by the geometric, gauge-theoretic, algorithmic, or dynamical structure needed to preserve the relevant notion of physical or statistical equivalence. In the cited literature, that accompanying structure takes several forms: coherent soft-photon clouds required by Gauss’s law in QED; relational dressings that render local observables gauge invariant in electrodynamics and gravity; fixed algorithmic geometry that turns practical MCMC reparameterizations into metric modifications; input-output preserving transformations that absorb unidentifiable parameters in rational ODE models; diffusion-mediated soft transports for categorical optimization; positive-definite covariance factorizations that induce sparsity on transformed scales; and balanced-control parameterizations that remove the counter-rotating term for dressed qubits (Hirai et al., 2019, Cheung et al., 25 May 2026, Betancourt, 2019, Meshkat et al., 2023, Gourevitch et al., 2 Jan 2026, Rybak et al., 2024, Salhov et al., 27 Jan 2026). This suggests a common motif: the operative object is not the reparameterization alone, but the reparameterization together with the structure that makes it admissible, invariant, or computationally effective.

1. Conceptual schema

A useful way to organize the subject is to distinguish between a complete reparameterization and a reparameterization that leaves some constitutive structure implicit. In the geometric MCMC formulation, a complete reparameterization is a diffeomorphism of the base manifold together with the induced transformations on tangent vectors, covectors, densities, and the metric tensor; under such a transformation, an exact geometric algorithm is invariant (Betancourt, 2019). By contrast, in several of the settings considered here, only part of the system is transformed explicitly, while another part is held fixed, added, or reconstructed.

A comparison across the main domains is summarized below.

Domain	Invariant or preserved structure	Dressing or nontrivial accompaniment
QED asymptotic states	BRST/gauge-invariant physical-state condition	Coherent soft-photon cloud $e^{R}\mathcal H_{free}$
Relational observables in gauge theory/gravity	Gauge invariance under local redundancies	Wilson-line or geodesic dressing
Riemannian MCMC	Full geometric invariance under complete reparameterization	Fixed implementation geometry induces an equivalent metric
Rational ODE models	Same input-output equations	New state and parameter variables built from Lie derivatives
Categorical optimization	Differentiable pathwise transport from noise	Diffusion-based soft reparameterization
Covariance/graphical models	Positive definiteness of $PD(p)$	Lie-theoretic factorization inducing transformed-scale sparsity
Dressed-qubit control	Exact dressed-frame dynamics without CRT	Balanced waveform implementing circular polarization in dressed basis

The comparison should not be read as asserting a single formal theory across all fields. Rather, it identifies a recurrent structural pattern. In each case, the “bare” variable change is insufficient relative to the relevant constraints: gauge constraints in field theory, metric geometry in MCMC, input-output equivalence in identifiability theory, positive-definite geometry in covariance modeling, or dynamical resonance structure in qubit control.

2. Gauge invariance, asymptotic states, and relational observables

In QED, the dressed-state formalism addresses the infrared failure of the ordinary $S$-matrix on bare charged Fock states. The key claim is that asymptotic charged states are not correctly represented by $b^\dagger(\vec p)\lvert 0\rangle$; Gauss’s law requires a charged particle to be accompanied by a long-range electromagnetic field, and the quantum realization of that field is a coherent cloud of soft photons. The central result is that the familiar Faddeev–Kulish dressing need not first be derived by solving an asymptotic Hamiltonian: it follows directly from imposing the correct BRST/gauge-invariant physical-state condition on asymptotic charged states (Hirai et al., 2019).

The corrected condition differs from the free Gupta–Bleuler constraint because the asymptotic charge density contributes. In the interaction-picture asymptotic limit, the relevant condition is

$$\lim_{t_i\to -\infty}\left[k^\mu a_\mu(k)+e^{i\omega t_i}\tilde{j}^{0I}(t_i,k)\right]\lvert\beta\rangle_0=0.$$

The paper then identifies an operator $\tilde R(t)$ satisfying

$$[k^\mu a_\mu(k),\tilde R(t)] = -e^{i\omega t}\tilde j^{0I}(t,k), \qquad [\tilde j^{0I}(t,k),\tilde R(t)] = 0,$$

so that states in $e^{\tilde R(t)}\mathcal H_{free}$ automatically satisfy the corrected physical-state condition. Using the asymptotic approximation $j^{0I}(t,x)\to j^0_{cl}(t,x)$, the Faddeev–Kulish operator $R(t)$ is shown to satisfy the required commutator relation, and the asymptotic physical Hilbert space becomes $PD(p)$0. The phase $PD(p)$1 is irrelevant for gauge invariance because it commutes with the constraint operator (Hirai et al., 2019).

The same work gives a direct physical interpretation: the dressing operator generates the classical Liénard–Wiechert field of a uniformly moving charge, and the incoming versus outgoing dressings differ by the $PD(p)$2 prescription in a manner that matches retarded versus advanced boundary conditions. It also relates the dressed-state construction to large gauge transformations, asymptotic charges, and the Ward-identity/memory relation $PD(p)$3. Bare Fock states generally lie in different asymptotic-charge sectors, whereas dressed states supply the required hard-plus-soft structure for infrared-finite scattering (Hirai et al., 2019).

A closely related but more general perspective appears in perturbative treatments of electrodynamics and gravity, where the problem is not infrared finiteness of charged asymptotic states but the nonexistence of gauge-invariant local observables. In scalar QED, a charged field is dressed by a Wilson line extending from past infinity,

$PD(p)$4

and in gravity a scalar observable is made relational by evaluating it at a point defined by a geodesic congruence, $PD(p)$5. The resulting dressings are not merely formal completions: they are mathematically equivalent to specific gauge choices, including a $PD(p)$6-gauge in electrodynamics and a dynamical temporal gauge in gravity where the gauge-fixing vector is itself a geodesic fluid (Cheung et al., 25 May 2026).

That equivalence has sharp perturbative consequences. The dressing contributes eikonal denominators $PD(p)$7 in QED and $PD(p)$8 in gravity, incorporates both potential and radiative photons or gravitons, and induces new kinematic singularities in matrix elements and self-energies. The same dressed observable may therefore be computed either by explicit dressing diagrams or by evaluating the undressed correlator in the gauge where the dressing becomes trivial; the two procedures agree exactly (Cheung et al., 25 May 2026). This suggests that, in gauge theory, a dressed reparametrization is best understood as a gauge-invariant re-expression of asymptotic states or operator insertions in variables already adapted to Gauss’s law or relational localization.

3. Incomplete reparameterizations and equivalent metrics in MCMC

In the MCMC setting, the core thesis is that practical advice of the form “reparameterization improves MCMC” does not describe a genuine invariance of the algorithm. What practitioners usually perform is an incomplete reparameterization: they transform the target density while leaving the proposal geometry, fiber measures, and especially the metric fixed or only partially transformed. The result is a different interaction between the target distribution and the algorithm, rather than the same algorithm in different coordinates (Betancourt, 2019).

The complete geometric picture begins with a diffeomorphism

$PD(p)$9

with Jacobian

$S$0

Under complete reparameterization, densities acquire Jacobian factors, tangent vectors push forward, cotangent vectors pull back, and the metric transforms tensorially:

$S$1

The quadratic form is invariant,

$S$2

and exact geometric algorithms are therefore unchanged (Betancourt, 2019).

An incomplete reparameterization transforms only the base coordinates and target density while the implementation continues to use the old metric components as if they lived in the new coordinates. Pulling that fixed implementation geometry back reveals an equivalent metric on the original manifold:

$S$3

This is the formal statement that incompletely reparameterizing the target is equivalent to leaving the target coordinates fixed and modifying the metric geometry directly. The point is especially consequential for Riemannian MCMC because the metric determines geodesic flow on the tangent bundle, Hamiltonian kinetic energy on the cotangent bundle, and the conditional distributions on tangent and cotangent fibers (Betancourt, 2019).

The paper makes the interpretation explicit: transformations such as non-centering are not merely coordinate tricks. They provide a simple way to inject a different effective geometry into a sampler whose software only exposes a fixed metric. This is why reparameterizations can alter proposal behavior, acceptance rates, and exploration efficiency for random walk Metropolis–Hastings, Langevin Monte Carlo or MALA, and Hamiltonian Monte Carlo, despite the formal coordinate invariance of the underlying probability distribution (Betancourt, 2019).

A local optimality criterion is defined by the invariant target geometry

$S$4

together with the covariant Hessian $S$5 and the deviation

$S$6

Under incomplete reparameterization, $S$7 enters the same criterion via $S$8. Reparameterization design thus becomes a geometric matching problem: select a transformation whose induced metric better matches posterior curvature (Betancourt, 2019).

The latent Gaussian model provides the cleanest illustration. In the centered form,

$S$9

weakly informative likelihoods produce a funnel geometry. Under the non-centered transformation

$b^\dagger(\vec p)\lvert 0\rangle$0

the latent prior becomes independent standard normals. For a single latent parameter with $b^\dagger(\vec p)\lvert 0\rangle$1, the Jacobian satisfies $b^\dagger(\vec p)\lvert 0\rangle$2, and starting from the unit metric, the transformed metric is $b^\dagger(\vec p)\lvert 0\rangle$3. For the chosen priors, the paper computes the covariant Hessian and finds exact cancellation,

$b^\dagger(\vec p)\lvert 0\rangle$4

In that precise sense, the non-centered parameterization is geometrically optimal for the model considered (Betancourt, 2019).

4. Reparametrizations that preserve external equivalence in dynamical systems

In rational ODE identifiability theory, reparametrization is organized not by coordinate invariance but by invariance of input-output behavior. The model class is

$b^\dagger(\vec p)\lvert 0\rangle$5

with $b^\dagger(\vec p)\lvert 0\rangle$6 rational over $b^\dagger(\vec p)\lvert 0\rangle$7. The reparametrization problem asks for new variables $b^\dagger(\vec p)\lvert 0\rangle$8 and new parameters $b^\dagger(\vec p)\lvert 0\rangle$9, algebraic over the old parameters, states, and input derivatives, such that there is a new ODE system $$\lim_{t_i\to -\infty}\left[k^\mu a_\mu(k)+e^{i\omega t_i}\tilde{j}^{0I}(t_i,k)\right]\lvert\beta\rangle_0=0.$$0, all parameters $$\lim_{t_i\to -\infty}\left[k^\mu a_\mu(k)+e^{i\omega t_i}\tilde{j}^{0I}(t_i,k)\right]\lvert\beta\rangle_0=0.$$1 are at least locally IO-identifiable, and the original and reparametrized models have the same input-output equations (Meshkat et al., 2023).

The invariant object is the input-output ideal $$\lim_{t_i\to -\infty}\left[k^\mu a_\mu(k)+e^{i\omega t_i}\tilde{j}^{0I}(t_i,k)\right]\lvert\beta\rangle_0=0.$$2, and identifiability is described via the smallest field $$\lim_{t_i\to -\infty}\left[k^\mu a_\mu(k)+e^{i\omega t_i}\tilde{j}^{0I}(t_i,k)\right]\lvert\beta\rangle_0=0.$$3 over which that ideal is generated. A function $$\lim_{t_i\to -\infty}\left[k^\mu a_\mu(k)+e^{i\omega t_i}\tilde{j}^{0I}(t_i,k)\right]\lvert\beta\rangle_0=0.$$4 is globally IO-identifiable if $$\lim_{t_i\to -\infty}\left[k^\mu a_\mu(k)+e^{i\omega t_i}\tilde{j}^{0I}(t_i,k)\right]\lvert\beta\rangle_0=0.$$5, and locally IO-identifiable if $$\lim_{t_i\to -\infty}\left[k^\mu a_\mu(k)+e^{i\omega t_i}\tilde{j}^{0I}(t_i,k)\right]\lvert\beta\rangle_0=0.$$6 is algebraic over $$\lim_{t_i\to -\infty}\left[k^\mu a_\mu(k)+e^{i\omega t_i}\tilde{j}^{0I}(t_i,k)\right]\lvert\beta\rangle_0=0.$$7. Unidentifiability arises when only parameter combinations are determined by the input-output equations, as in

$$\lim_{t_i\to -\infty}\left[k^\mu a_\mu(k)+e^{i\omega t_i}\tilde{j}^{0I}(t_i,k)\right]\lvert\beta\rangle_0=0.$$8

which yields

$$\lim_{t_i\to -\infty}\left[k^\mu a_\mu(k)+e^{i\omega t_i}\tilde{j}^{0I}(t_i,k)\right]\lvert\beta\rangle_0=0.$$9

Only $\tilde R(t)$0 and $\tilde R(t)$1 are then identifiable (Meshkat et al., 2023).

The constructive method proceeds in four steps. First, compute the input-output equations and the rational parametrization from Lie derivatives. Second, replace parameter monomials in that parametrization by new indeterminates $\tilde R(t)$2, and solve the induced polynomial system over the algebraic closure of the identifiable field, selecting a rational solution when possible to obtain global identifiability. Third, reconstruct an ODE realization by defining new state variables and solving

$\tilde R(t)$3

for the new vector field. Fourth, recover the state transformation by solving compatibility equations via Gröbner elimination (Meshkat et al., 2023).

The guarantees are substantial. The procedure solves the local IO-identifiable reparametrization problem for rational ODE models. If the sum of the output orders in the IO-equations equals the model dimension, then the new state variables are algebraic in the old state and parameters. If there exists a globally IO-identifiable reparametrization whose Lie-derivative monomial support is contained in that of the original model, then the algorithm can find it. For linear models, every model has a globally IO-identifiable linear reparametrization obtained by a linear change of variables (Meshkat et al., 2023).

The examples show how the dressing absorbs non-identifiable quantities into new states or parameter combinations. For the two-exponential example, the reparametrized system is

$\tilde R(t)$4

with $\tilde R(t)$5 and $\tilde R(t)$6. For Lotka–Volterra with input,

$\tilde R(t)$7

the identifiable field is $\tilde R(t)$8, and the method constructs

$\tilde R(t)$9

with $$[k^\mu a_\mu(k),\tilde R(t)] = -e^{i\omega t}\tilde j^{0I}(t,k), \qquad [\tilde j^{0I}(t,k),\tilde R(t)] = 0,$$0 and $$[k^\mu a_\mu(k),\tilde R(t)] = -e^{i\omega t}\tilde j^{0I}(t,k), \qquad [\tilde j^{0I}(t,k),\tilde R(t)] = 0,$$1. The non-identifiable parameter $$[k^\mu a_\mu(k),\tilde R(t)] = -e^{i\omega t}\tilde j^{0I}(t,k), \qquad [\tilde j^{0I}(t,k),\tilde R(t)] = 0,$$2 is thus absorbed into the new state variable (Meshkat et al., 2023).

A different but related instance of external-equivalence-preserving reparametrization arises for finite-dimensional Poisson systems

$$[k^\mu a_\mu(k),\tilde R(t)] = -e^{i\omega t}\tilde j^{0I}(t,k), \qquad [\tilde j^{0I}(t,k),\tilde R(t)] = 0,$$3

where a new-time transformation

$$[k^\mu a_\mu(k),\tilde R(t)] = -e^{i\omega t}\tilde j^{0I}(t,k), \qquad [\tilde j^{0I}(t,k),\tilde R(t)] = 0,$$4

produces

$$[k^\mu a_\mu(k),\tilde R(t)] = -e^{i\omega t}\tilde j^{0I}(t,k), \qquad [\tilde j^{0I}(t,k),\tilde R(t)] = 0,$$5

The transformed system remains Poisson if and only if $$[k^\mu a_\mu(k),\tilde R(t)] = -e^{i\omega t}\tilde j^{0I}(t,k), \qquad [\tilde j^{0I}(t,k),\tilde R(t)] = 0,$$6 satisfies the Jacobi identities, which reduces to the linear first-order PDE system

$$[k^\mu a_\mu(k),\tilde R(t)] = -e^{i\omega t}\tilde j^{0I}(t,k), \qquad [\tilde j^{0I}(t,k),\tilde R(t)] = 0,$$7

Two principal admissible families are identified: Casimir-generated factors $$[k^\mu a_\mu(k),\tilde R(t)] = -e^{i\omega t}\tilde j^{0I}(t,k), \qquad [\tilde j^{0I}(t,k),\tilde R(t)] = 0,$$8, and the rank-two case in which every smooth nonvanishing $$[k^\mu a_\mu(k),\tilde R(t)] = -e^{i\omega t}\tilde j^{0I}(t,k), \qquad [\tilde j^{0I}(t,k),\tilde R(t)] = 0,$$9 is admissible. These results lead to a global constructive Darboux reduction for rank-two Poisson systems, obtained from a coordinate transformation built from two arbitrary functions and a complete set of Casimir invariants, followed by the explicit time reparametrization

$e^{\tilde R(t)}\mathcal H_{free}$0

that normalizes the structure matrix to canonical Darboux form (Hernández-Bermejo, 2019).

Across both literatures, the reparameterized system is not required to preserve original coordinates or even original state structure. What is preserved is an external invariant: input-output equations in identifiability theory, or Poisson structure in the time-reparameterized Hamiltonian setting.

5. Soft transports and sparsity-inducing transformed scales

For optimization with categorical variables, the obstacle to pathwise differentiation is that an exact map from continuous noise to discrete outcomes is piecewise constant almost everywhere. The diffusion-based construction addresses this by wrapping categorical sampling inside a Gaussian diffusion and deterministic DDIM-style reverse transport. The objective is

$e^{\tilde R(t)}\mathcal H_{free}$1

where the categorical law factorizes over $e^{\tilde R(t)}\mathcal H_{free}$2 variables with $e^{\tilde R(t)}\mathcal H_{free}$3 categories. The forward interpolation is

$e^{\tilde R(t)}\mathcal H_{free}$4

with $e^{\tilde R(t)}\mathcal H_{free}$5 and $e^{\tilde R(t)}\mathcal H_{free}$6 (Gourevitch et al., 2 Jan 2026).

Because $e^{\tilde R(t)}\mathcal H_{free}$7 is one-hot, the posterior denoiser has a closed form. For each row $e^{\tilde R(t)}\mathcal H_{free}$8 and category $e^{\tilde R(t)}\mathcal H_{free}$9,

$j^{0I}(t,x)\to j^0_{cl}(t,x)$0

and the posterior mean is

$j^{0I}(t,x)\to j^0_{cl}(t,x)$1

The deterministic reverse map is

$j^{0I}(t,x)\to j^0_{cl}(t,x)$2

and composing these steps yields a differentiable sampler through which one may backpropagate using a pathwise gradient estimator (Gourevitch et al., 2 Jan 2026).

The construction is explicitly a soft reparameterization rather than an exact transport to the discrete law. With only a single diffusion step, the reparameterized sample reduces to the categorical mean and recovers straight-through-style behavior; the authors also derive a ReinMax-style version so that the one-step case collapses to ReinMax. At the same time, making the relaxation too sharp destroys gradient signal: if the first reverse step $j^{0I}(t,x)\to j^0_{cl}(t,x)$3, then

$j^{0I}(t,x)\to j^0_{cl}(t,x)$4

The method thus exhibits the same bias–gradient-quality trade-off as temperature-based relaxations, but with the trade-off parameterized by the diffusion schedule and number of steps rather than a Gumbel-Softmax temperature (Gourevitch et al., 2 Jan 2026).

A structurally different transformed-scale construction appears in covariance and graphical-model theory. There the problem is that sparsity and positive definiteness are intrinsically linked. The relevant domain is

$j^{0I}(t,x)\to j^0_{cl}(t,x)$5

and the reparametrizations are built so that sparsity on the transformed scale preserves the geometry of the positive-definite cone. Four maps are emphasized:

$j^{0I}(t,x)\to j^0_{cl}(t,x)$6

$j^{0I}(t,x)\to j^0_{cl}(t,x)$7

$j^{0I}(t,x)\to j^0_{cl}(t,x)$8

$j^{0I}(t,x)\to j^0_{cl}(t,x)$9

These induce, respectively, matrix-logarithm, orthogonal-plus-diagonal, lower-triangular, and LDL/regression-graph structures (Rybak et al., 2024).

The richest case is $R(t)$0, with $R(t)$1 and $R(t)$2, which under a causal ordering aligns with chain graphs and specializes to DAGs when chain components are singletons. The paper ties the family of reparametrizations to the Iwasawa decomposition $R(t)$3 of $R(t)$4, with $R(t)$5, $R(t)$6, and $R(t)$7, and interprets transformed-scale sparsity in graphical terms. In the DAG setting, if $R(t)$8 and $R(t)$9, then $PD(p)$00 records direct effects, $PD(p)$01 total effects along all directed paths, and $PD(p)$02 weighted path effects with path length $PD(p)$03 discounted by $PD(p)$04. Hence an approximate zero on the logarithmic scale indicates negligible relatively direct regression effects, even though longer indirect paths may remain (Rybak et al., 2024).

The commonality with diffusion-based categorical reparameterization is not a shared formula but a shared logic. In both cases, the transformed representation is deliberately softer or more structured than the original object: a categorical law is accessed through a differentiable diffusion trajectory, and a covariance matrix is accessed through Lie-theoretic coordinates in which sparsity is meaningful while positive definiteness is automatic.

6. Balanced control and dressed qubits

In dressed-qubit control, the dressing is not a state-space completion as in QED but a control-parameterization strategy. Standard dressed-qubit protocols define a dressed qubit by resonantly driving a bare two-level system,

$PD(p)$05

which under the bare rotating-wave approximation yields

$PD(p)$06

Further control of this dressed qubit is usually implemented by linearly polarized modulation in the dressed basis, which decomposes into co-rotating and counter-rotating components. The counter-rotating term forces operation in the regime $PD(p)$07, where a dressed rotating-wave approximation is valid (Salhov et al., 27 Jan 2026).

The alternative protocol keeps only a single physical coupling axis in the laboratory frame but writes the waveform as

$PD(p)$08

After transforming to the interaction picture with respect to $PD(p)$09 and assuming the bare RWA, this yields the balanced-control Hamiltonian

$PD(p)$10

No extra $PD(p)$11- or $PD(p)$12-axis hardware is required (Salhov et al., 27 Jan 2026).

To define the dressed qubit, one sets $PD(p)$13. The key reparameterized control choice is then

$PD(p)$14

which is circularly polarized in the dressed frame. If

$PD(p)$15

the counter-rotating term cancels exactly in the second interaction picture, and the resulting Hamiltonian is

$PD(p)$16

No dressed rotating-wave approximation is required (Salhov et al., 27 Jan 2026).

The practical consequences are extensive. Because the CRT is removed rather than perturbatively neglected, the protocol can operate at $PD(p)$17, allowing dressed gates as fast as bare-qubit gates while reducing CRT-induced control errors. In a trapped-ion two-qubit benchmark, rewriting an earlier phase-modulated drive in the new circularly polarized dressed form yields a $PD(p)$18 reduction in infidelity, to about $PD(p)$19. For sensing, the protocol achieves more than an order-of-magnitude sensitivity improvement and a wider operating range in $PD(p)$20. For atomic-clock stabilization, the ability to use $PD(p)$21 without CRT-induced degradation yields about a 5-fold improvement in coherence time over the previously used magic-angle detuned double-drive (Salhov et al., 27 Jan 2026).

The same work supplies a Floquet-based coherence analysis and an approximate coherence-time expression, together with the scaling $PD(p)$22. Here the “dressed reparametrization” is explicitly a redesign of the control variables $PD(p)$23 and their phases so that the dressed Hamiltonian is implemented as an exact rotating field rather than a linear one. The general implication is that, even in control theory, the operative transformation is not merely a re-labeling of drive parameters but a structural recoding of the waveform that changes the effective dynamics.

7. Interpretation and scope

Across these literatures, dressed reparametrizations share three recurring features. First, they preserve an invariant object that is more fundamental than the original parametrization: BRST/Gauss constraints, asymptotic charge sectors, input-output equations, Poisson structure, the positive-definite cone, or exact dressed-frame dynamics. Second, they supplement or reinterpret the variable change with additional structure: soft clouds, Wilson lines, geodesic congruences, equivalent metrics, new state variables, diffusion trajectories, or Lie-theoretic factorizations. Third, they change the operational problem. A bare charged particle becomes an IR-inadmissible asymptotic state, a target-only MCMC reparameterization becomes a new effective algorithm, an unidentifiable ODE model becomes an identifiable realization, and a linearly driven dressed qubit becomes a CRT-free programmable qubit (Hirai et al., 2019, Betancourt, 2019, Meshkat et al., 2023, Gourevitch et al., 2 Jan 2026, Rybak et al., 2024, Salhov et al., 27 Jan 2026).

A common misconception is that reparameterization is always innocuous because it “should leave the system invariant.” The cited work on MCMC explicitly rejects that conclusion for practical implementations: incomplete reparameterizations alter the metric geometry seen by the sampler (Betancourt, 2019). The gauge-theory literature similarly rejects the idea that dressing is optional ornamentation: in QED it is required by the correct asymptotic gauge condition, and in gravity it is required because bare local operators are not observables (Hirai et al., 2019, Cheung et al., 25 May 2026). In identifiability theory, the relevant equivalence is not sameness of coordinates but indistinguishability at the input-output level (Meshkat et al., 2023).

The term therefore names a family resemblance rather than a single doctrine. In each setting, the decisive move is to replace a naked parametrization by one “dressed” with the structures that the original formulation had treated as external, fixed, or implicit. This suggests a general methodological lesson: where a problem is constrained by gauge redundancy, geometry, observability, positivity, or resonance structure, the most effective reparameterization is often the one that internalizes those constraints into the parametrization itself.