Coherent Reprogramming Methods

• Coherent reprogramming is the controlled reconfiguration of physical or representational substrates while preserving global structural coherence.
• In quantum information, it involves using free operations and majorization protocols to transform coherent states and channels, often quantified by success probabilities and resource metrics.
• The concept extends to hardware control, cellular reprogramming, and model adaptation, enabling structured error mitigation and efficient repurposing across diverse systems.

Coherent reprogramming denotes controlled repurposing or reconfiguration of an existing physical, dynamical, or representational substrate while preserving, exploiting, or certifying coherence. In quantum information, it appears as the reconfiguration of coherent states or channels under constrained free operations, the programming of open-system dynamics, the coherent control of non-unitary operations, and the tailoring of coherent error into effective stochastic noise. In coherent-control hardware, it appears as programmable manipulation of phases, amplitudes, modal correlations, or motional unitaries. In cellular and machine-learning literatures, it denotes globally structured reconfiguration of attractors or representations rather than arbitrary local modification (Liu et al., 2024, Díaz et al., 2018, Lobser et al., 2022).

1. Conceptual range and recurring structure

In the quantum-resource setting, coherent reprogramming is explicitly framed as “the controlled repurposing or reconfiguration of quantum coherence under operational constraints,” with strictly incoherent operations defining the control constraints and majorization defining which coherence patterns can be reconfigured into which others (Liu et al., 2024). In fault-tolerant quantum computation, the same phrase is used more operationally: logical randomized compiling aims to “reprogram how coherent physical/control imperfections appear at the logical level,” so that the effective logical noise is averaged into a stochastic, syndrome-respecting form (Beale et al., 2023). In quantum programming-language semantics, the problem is to define coherent control of arbitrary quantum operations, including non-unitary branches under qcase, in a way that remains meaningful in the presence of iteration (Barsse et al., 14 Jul 2025).

Several biologically oriented papers do not explicitly use the phrase, but they analyze mechanisms that the literature synthesis presents as relevant to it. In one case, reprogramming is described as “collective, attractor-level, phase-structured reprogramming” in a hierarchy of cell cycles (Hannam et al., 2016). In another, pluripotency is approached through “global attraction to the unstable manifold of a saddle,” produced by the interaction of oscillatory gene-expression dynamics and slow epigenetic variables (Matsushita et al., 2021). In spatio-temporal forecasting, a frozen LLM is adapted through “semantic-oriented reprogramming,” in which decomposition and selective vocabulary alignment impose structured compatibility between non-text signals and pretrained textual representations (Wang et al., 2024).

Taken together, these uses suggest a common pattern: coherent reprogramming is less a single formalism than a family of techniques in which global relations among degrees of freedom are preserved or deliberately reshaped. The relevant notion of coherence may be quantum superposition, collective attractor organization, modal correlation, or representation-level consistency, but the recurring theme is structured rather than pointwise reconfiguration.

2. Quantum state and channel transformation

In the resource theory of coherence, the pure-state problem is formulated with respect to a fixed basis $\{|i\rangle\}$. For

$|\psi\rangle=\sum_{i=1}^d \psi_i |i\rangle,$

the coherence vector is the sorted probability vector of squared amplitudes. Under strictly incoherent operations, deterministic pure-state conversion is possible iff

$\psi \prec \phi,$

and, equivalently, iff the cumulative-tail monotones

$C_\ell(\psi)=\sum_{i=\ell}^d \psi_i^\downarrow$

satisfy

$$C_\ell(\psi)\ge C_\ell(\phi),\qquad \forall \ell\in[1,d].$$

When deterministic conversion fails, the maximal success probability is

$$P_{\max}(|\psi\rangle \to |\phi\rangle) = \min_{\ell\in[1,d]}\frac{C_\ell(\psi)}{C_\ell(\phi)}.$$

The same work adapts two optimal majorization-lattice protocols for probabilistic coherence transformation: the greedy protocol, which routes through the optimal common product state $\phi^\vee=\psi\vee\phi$, and the thrifty protocol, which routes through the optimal common resource state $\phi^\wedge=\psi\wedge\phi$. Both attain the same optimal success probability, while the thrifty protocol preserves more coherence on average when the transformation fails. Deterministic and probabilistic coherence transformations between two mixed states are also explored, and probabilistic preprocessing can improve coherence-to-entanglement conversion (Liu et al., 2024).

A closely related process-theoretic perspective studies how a coherent state can be used to implement a quantum channel via maximally incoherent operations. The key quantity is the robustness of coherence of a channel,

$$1+C_R(\mathcal{N}) := \min \left\{\lambda: \ \mathcal{N} \leq \lambda \mathcal{M},\ \mathcal{M}\in\mathrm{MIO}\right\},$$

which reduces to the usual state robustness for constant-output channels. Its logarithm,

$C_{LR}(\mathcal{N})=\log(1+C_R(\mathcal{N})),$

quantifies both the amortized coherence cost of implementation and the zero-error asymptotic cost of implementing many independent copies of a channel. The minimal cosdit size needed for $|\psi\rangle=\sum_{i=1}^d \psi_i |i\rangle,$0-approximate implementation is

$|\psi\rangle=\sum_{i=1}^d \psi_i |i\rangle,$1

while the amortized cost is exactly

$|\psi\rangle=\sum_{i=1}^d \psi_i |i\rangle,$2

The same paper proves that every pure coherent state in dimension larger than $|\psi\rangle=\sum_{i=1}^d \psi_i |i\rangle,$3, however weakly coherent so, is a valuable resource to implement some coherent unitary channel (Díaz et al., 2018).

These two lines of work treat coherent reprogramming at different levels. The first concerns reconfiguration of coherence distributions under a free-operation class; the second concerns using coherence in a program state to realize a target process. A plausible implication is that state-level convertibility and process-level implementability are two complementary forms of the same broader resource-conversion problem.

3. Certification and programmability of quantum dynamics

A distinct problem is whether a programmed family of outputs is genuinely coherent or merely simulates coherence through hidden classical control. For a preparation device $|\psi\rangle=\sum_{i=1}^d \psi_i |i\rangle,$4, the relevant incoherent set $|\psi\rangle=\sum_{i=1}^d \psi_i |i\rangle,$5 consists of families admitting a decomposition

$|\psi\rangle=\sum_{i=1}^d \psi_i |i\rangle,$6

such that, for every $|\psi\rangle=\sum_{i=1}^d \psi_i |i\rangle,$7,

$|\psi\rangle=\sum_{i=1}^d \psi_i |i\rangle,$8

This shows that noncommutativity of the observed $|\psi\rangle=\sum_{i=1}^d \psi_i |i\rangle,$9 is not sufficient for coherence certification under hidden classical control. The paper gives a complete hierarchy of semidefinite programs,

$\psi \prec \phi,$0

a practical SDP relaxation based on block moment matrices, and, for qubits, an exact equivalence between coherence under hidden classical control and joint measurability of the associated unbiased dichotomic POVMs. The qubit method scales to more than one thousand qubits, and the framework extends to coherence-breaking versus coherence-preserving quantum channels (Cobucci et al., 2 Jun 2026).

A further generalization considers programmability of Lindbladian semigroups. Here a fixed retrieval map $\psi \prec \phi,$1 and a time-varying analytic family of program states $\psi \prec \phi,$2 are required to reproduce

$\psi \prec \phi,$3

through

$\psi \prec \phi,$4

For exact physical programmability, the paper derives a necessary condition: $\psi \prec \phi,$5 for some quantum channel $\psi \prec \phi,$6 and $\psi \prec \phi,$7. This rules out coherent generators with non-trivial Hamiltonian part $\psi \prec \phi,$8, and also rules out typical amplitude-damping generators. By contrast, covariant semigroups and fully dissipative Pauli Lindbladians are physically programmable with finite program dimension. For nonphysically programmable cases, explicit finite-resource protocols are constructed in the larger HPTP setting, and an operational programming cost is introduced through the diamond norm of the retrieval map (Jing et al., 9 Dec 2025).

The relation between these two directions is direct. Certification asks whether an apparent programmed family resists all classical hidden-variable explanations; semigroup programmability asks which continuous families of open-system dynamics admit exact programming by a fixed physical processor. In both cases, coherence is treated as a structural property of the entire family, not of isolated outputs alone.

4. Coherent control of errors, subroutines, and proofs

In fault-tolerant quantum computation, coherent reprogramming often refers to rewriting the effective error model rather than the target computation. Logical Randomized Compiling inserts random stabilizers and logical twirling operations into encoded gadgets so that coherent errors do not remain as off-diagonal amplitudes between syndrome sectors. The central equalities are

$\psi \prec \phi,$9

for random stabilizer averaging, and

$C_\ell(\psi)=\sum_{i=\ell}^d \psi_i^\downarrow$0

for corrected twirling of a noisy encoded gate. The result is logical/subspace dephasing between cospaces, and, for suitable twirling groups, stochastic logical Weyl noise within cospaces. The method is applied to logical gates, state preparation, measurement, syndrome extraction, and idling, with the claim that it does not significantly increase logical-circuit depth and usually does not lead to any increase in depth (Beale et al., 2023).

A complementary diagnostic perspective appears in cycle-level error reconstruction. There, coherent and decoherent contributions are separated by folding a hard cycle $C_\ell(\psi)=\sum_{i=\ell}^d \psi_i^\downarrow$1 times before Pauli twirling. For small noise, repeated-channel Pauli fidelity obeys

$C_\ell(\psi)=\sum_{i=\ell}^d \psi_i^\downarrow$2

so coherent contributions appear in the quadratic term and decoherent contributions in the linear term. Proof-of-concept experiments on ibmq_guadalupe, ibmq_manila, and ibmq_montreal reported substantial coherent errors biased in $C_\ell(\psi)=\sum_{i=\ell}^d \psi_i^\downarrow$3 (Carignan-Dugas et al., 2023).

At the language-theoretic level, coherent control is formalized through a quantum conditional

$C_\ell(\psi)=\sum_{i=\ell}^d \psi_i^\downarrow$4

an operational semantics based on appropriate Kraus decompositions, and a denotational semantics based on vacuum-extensions. The denotational qcase acts on pairs $C_\ell(\psi)=\sum_{i=\ell}^d \psi_i^\downarrow$5, where $C_\ell(\psi)=\sum_{i=\ell}^d \psi_i^\downarrow$6 is the ordinary quantum operation and $C_\ell(\psi)=\sum_{i=\ell}^d \psi_i^\downarrow$7 is the transformation matrix retained by the vacuum-extension formalism: $C_\ell(\psi)=\sum_{i=\ell}^d \psi_i^\downarrow$8 The language is universal for vacuum-extensions, the operational and denotational semantics are adequate, and the denotational semantics is fully abstract for observational equivalence (Barsse et al., 14 Jul 2025).

A cryptographic use of the phrase is even more literal. In the quantum random oracle model, “coherent reprogramming” denotes a measure-and-reprogram framework in which the simulator stores queried points coherently rather than measuring query registers immediately. The resulting lifting theorem improves the loss to

$C_\ell(\psi)=\sum_{i=\ell}^d \psi_i^\downarrow$9

which the paper contrasts with an earlier $$C_\ell(\psi)\ge C_\ell(\phi),\qquad \forall \ell\in[1,d].$$0-type bound. This supports direct product theorems, salted-game hardness, and multi-instance search results by reducing the final analysis to classical reasoning about uniformly random image tuples (Cojocaru et al., 11 Sep 2025).

5. Hardware embodiments

In trapped-ion control, coherent reprogramming is implemented as a systems capability rather than as an abstract resource theory. The Octet architecture uses hardware-native phase bookkeeping, pulse-level modularization, multi-stage compressed representations, an embedded pulse compiler, and in-situ mutation of compressed gate data. The waveform model

$$C_\ell(\psi)\ge C_\ell(\phi),\qquad \forall \ell\in[1,d].$$1

makes virtual $$C_\ell(\psi)\ge C_\ell(\phi),\qquad \forall \ell\in[1,d].$$2 rotations and frame updates hardware primitives. Reported figures include branching latency of about $$C_\ell(\psi)\ge C_\ell(\phi),\qquad \forall \ell\in[1,d].$$3 ns, full mutation time of about $$C_\ell(\psi)\ge C_\ell(\phi),\qquad \forall \ell\in[1,d].$$4 for gates with $$C_\ell(\psi)\ge C_\ell(\phi),\qquad \forall \ell\in[1,d].$$5 spline knots, and storage of over $$C_\ell(\psi)\ge C_\ell(\phi),\qquad \forall \ell\in[1,d].$$6 gates in 128 MB (Lobser et al., 2022).

Integrated photonics provides a second hardware realization. In a hexagonal mesh of $$C_\ell(\psi)\ge C_\ell(\phi),\qquad \forall \ell\in[1,d].$$7 Mach-Zehnder interferometers with $$C_\ell(\psi)\ge C_\ell(\phi),\qquad \forall \ell\in[1,d].$$8 accessible perimeter optical channels, a two-mode coherence matrix

$$C_\ell(\psi)\ge C_\ell(\phi),\qquad \forall \ell\in[1,d].$$9

is synthesized from incoherent input by first using a non-unitary attenuation stage

$$P_{\max}(|\psi\rangle \to |\phi\rangle) = \min_{\ell\in[1,d]}\frac{C_\ell(\psi)}{C_\ell(\phi)}.$$0

to set the degree of spatial coherence

$$P_{\max}(|\psi\rangle \to |\phi\rangle) = \min_{\ell\in[1,d]}\frac{C_\ell(\psi)}{C_\ell(\phi)}.$$1

and then applying a programmable $$P_{\max}(|\psi\rangle \to |\phi\rangle) = \min_{\ell\in[1,d]}\frac{C_\ell(\psi)}{C_\ell(\phi)}.$$2 unitary

$$P_{\max}(|\psi\rangle \to |\phi\rangle) = \min_{\ell\in[1,d]}\frac{C_\ell(\psi)}{C_\ell(\phi)}.$$3

Reconstruction is performed via spatial Stokes parameters,

$$P_{\max}(|\psi\rangle \to |\phi\rangle) = \min_{\ell\in[1,d]}\frac{C_\ell(\psi)}{C_\ell(\phi)}.$$4

Across the reported coherent and partially coherent cases, the mixed-state fidelity satisfied $$P_{\max}(|\psi\rangle \to |\phi\rangle) = \min_{\ell\in[1,d]}\frac{C_\ell(\psi)}{C_\ell(\phi)}.$$5 (Hashemi et al., 14 Jan 2026).

For ultracold atoms in optical lattices, coherent rearrangement is implemented as arbitrary unitary control of single-particle motional states. Using an analogy with discrete linear optics and the Clements scheme, any $$P_{\max}(|\psi\rangle \to |\phi\rangle) = \min_{\ell\in[1,d]}\frac{C_\ell(\psi)}{C_\ell(\phi)}.$$6-mode unitary is decomposed into nearest-neighbor two-site blocks

$$P_{\max}(|\psi\rangle \to |\phi\rangle) = \min_{\ell\in[1,d]}\frac{C_\ell(\psi)}{C_\ell(\phi)}.$$7

and each block is compiled into native double-well gates through

$$P_{\max}(|\psi\rangle \to |\phi\rangle) = \min_{\ell\in[1,d]}\frac{C_\ell(\psi)}{C_\ell(\phi)}.$$8

The scheme enables subroutines including the Discrete Fourier Transform and finite-time simulation of non-native Hamiltonians. In two dimensions, all-to-all atomic rearrangement is achieved with circuit depth

$$P_{\max}(|\psi\rangle \to |\phi\rangle) = \min_{\ell\in[1,d]}\frac{C_\ell(\psi)}{C_\ell(\phi)}.$$9

and the error analysis identifies addressing crosstalk at the level $\phi^\vee=\psi\vee\phi$0 as a key requirement for high-fidelity larger circuits (Roth et al., 4 Mar 2026).

These platform-specific realizations share a common architectural trait: the programmable layer is narrow and structured. Global tunnel pulses plus local phases, selected vocabulary embeddings plus cross-attention, or fixed hardware with mutable compressed pulselets are sufficient when the control model is chosen to preserve coherent relations.

6. Cellular and dynamical-systems meanings

In cellular reprogramming, coherence is not quantum coherence but coordinated motion through a structured state space. An early formulation is the epigenetic state network, in which stable cell phenotypes are fixed-point attractors, edges are saddle-mediated transition routes with Wentzell-Freidlin rates, and dominant pathways are extracted by a minimum spanning tree. Applied to fibroblast-to-cardiomyocyte conversion, the framework predicts three major routes: one via iPSC and subsequent differentiation, one indirect transdifferentiation route via cardiac progenitor, and one direct route through a fibroblast-cardiomyocyte double-positive intermediate FC (Wang et al., 2012).

A more detailed statistical-physics model adds transcription-factor binding, protein synthesis and degradation, and slow histone-state changes. The resulting epigenetic landscape contains basin minima generated in fixed histone states, and slow histone dynamics creates narrow valleys connecting differentiated and pluripotent states. These valleys explain both the existence of an intermediary state and the heterogeneity of latencies in reprogramming. Changing the Yamanaka-factor mechanism alters the pathway to bypass the barrier, thereby accelerating reprogramming and reducing latency heterogeneity (Ashwin et al., 2014).

A distinct attractor-based model represents cell types as cyclic attractors corresponding to stages of the cell cycle. Reprogramming then becomes a transition between hierarchically related limit cycles. The two mechanisms identified are cycle-specific directed perturbations and noise-induced switching. The threshold perturbation fraction $\phi^\vee=\psi\vee\phi$1 depends strongly on which cell-cycle phase is targeted, and the most effective intervention is applied just before the phase where stem and daughter cycles are most similar (Hannam et al., 2016).

The slow-fast theory of pluripotency makes the geometric picture sharper. Gene-expression variables $\phi^\vee=\psi\vee\phi$2 evolve quickly according to

$\phi^\vee=\psi\vee\phi$3

while epigenetic variables evolve slowly via

$\phi^\vee=\psi\vee\phi$4

In the reduced dynamics, pluripotency is a saddle near $\phi^\vee=\psi\vee\phi$5, and successful reprogramming is interpreted as global attraction toward the unstable manifold of that saddle. The paper reports the same qualitative mechanism in the repressilator, in random gene regulatory networks, and in a five-gene embryonic stem-cell network (Matsushita et al., 2021).

Taken together, these works describe coherent reprogramming in biology as collective, low-dimensional, and pathway-constrained. The coherence lies in shared attractor geometry, synchronized oscillatory modes, bottleneck states, and phase-specific control rather than in microscopic quantum superposition.

7. Representational and cross-domain machine learning

In machine learning, model reprogramming is defined as repurposing and reusing a well-developed pre-trained model from a source domain to solve tasks in a target domain without model finetuning, by introducing an input transformation layer and an output mapping layer (Chen, 2022). This usage is not identical to quantum or cellular coherence, but it preserves the idea of structured reuse of an existing substrate rather than rebuilding the model internally.

A concrete example is RePST, which adapts a frozen GPT-2 backbone to spatio-temporal forecasting. The input

$\phi^\vee=\psi\vee\phi$6

is first normalized and decomposed into dynamic components

$\phi^\vee=\psi\vee\phi$7

and a reconstructed dominant signal $\phi^\vee=\psi\vee\phi$8. These are patched and embedded, then aligned to the language-model space through a selective discrete reprogramming scheme. A learnable mask over the pretrained embedding matrix

$\phi^\vee=\psi\vee\phi$9

defines an expanded spatio-temporal vocabulary $\phi^\wedge=\psi\wedge\phi$0, and cross-attention produces reprogrammed embeddings

$\phi^\wedge=\psi\wedge\phi$1

The model is trained only with forecasting MAE,

$\phi^\wedge=\psi\wedge\phi$2

and was reported to outperform twelve state-of-the-art baseline methods, particularly in data-scarce scenarios (Wang et al., 2024).

This suggests a representational notion of coherent reprogramming: the source model is not altered internally, but the interface presented to it is made semantically and structurally consistent with the model’s pretrained geometry. The resulting coherence is architectural and representational rather than explicitly resource-theoretic.

Across these literatures, coherent reprogramming consistently denotes more than mere retargeting. It denotes reconfiguration that preserves or exploits structure: majorization order in coherence theory, syndrome-sector structure in fault tolerance, analytic program families in open-system dynamics, phase bookkeeping in hardware control, attractor geometry in cellular reprogramming, or semantic decomposition in model adaptation. The specific mathematics varies sharply, but the shared objective is the same: to repurpose an existing coherent substrate without reducing it to incoherent local replacement.