Hidden Quantum Markov Models (HQMM)

• Hidden Quantum Markov Models (HQMMs) are quantum stochastic processes that generalize classical HMMs using quantum channels and density operators to capture complex temporal correlations.
• They offer enhanced memory representation and compression by encoding quantum-coherent effects and non-commuting operator dynamics beyond classical HMM capabilities.
• HQMMs find applications in diverse fields such as neuroscience, quantum transport, and machine learning, leveraging activated measurement protocols and tailored learning algorithms.

Hidden Quantum Markov Models (HQMMs) are a class of formal latent-variable quantum stochastic processes that generalize classical hidden Markov models (HMMs) by replacing finite-state Markov chains with quantum channels on density operators, thereby supporting more general stochastic and quantum-coherent sequence generation. HQMMs are distinct in that their hidden states are quantum density matrices evolving via completely positive trace-preserving (CPTP) maps, with output observed through quantum measurement instruments. This formalism enables a richer representation of temporal correlations and provides enhanced expressive and compressive power for stochastic processes relative to classical HMMs. HQMMs possess well-defined autocovariance and spectral properties and allow seamless integration with physical quantum processes such as open-system evolution and feedback.

1. Quantum Stochastic Process: Formal Structure of HQMMs

An HQMM is specified on a finite-dimensional Hilbert space $\mathcal{H}$ by a triple

• initial state $\rho_0 \in \mathcal{D}(\mathcal{H})$ (density operators),
• a time-homogeneous CPTP map $$\mathcal{E}: \mathcal{D}(\mathcal{H}) \rightarrow \mathcal{D}(\mathcal{H})$$,
• a set of measurement instruments $\{\mathcal{M}_y\}_{y \in \mathcal{Y}}$, one for each output symbol $y$ in a finite alphabet $\mathcal{Y}$.

The dynamics are encoded by Kraus decompositions:

• $$\mathcal{E}(\rho) = \sum_{k=1}^K K_k \rho K_k^\dagger$$, subject to $\sum_{k} K_k^\dagger K_k = I$,
• $$\mathcal{M}_y(\rho) = \sum_{i=1}^{m_y} M_{y,i} \rho M_{y,i}^\dagger$$, with $\sum_{y,i} M_{y,i}^\dagger M_{y,i} = I$.

The sequence of state and output updates at discrete times $\rho_0 \in \mathcal{D}(\mathcal{H})$0 is

$\rho_0 \in \mathcal{D}(\mathcal{H})$1

After observing $\rho_0 \in \mathcal{D}(\mathcal{H})$2, the unnormalized posterior is

$\rho_0 \in \mathcal{D}(\mathcal{H})$3

2. Relationship to Classical HMMs

HQMMs strictly generalize classical HMMs. If all Kraus operators commute and measurements are performed in a fixed basis at each step (the "activated measurement protocol"), the output sequence $\rho_0 \in \mathcal{D}(\mathcal{H})$4 is distributionally equivalent to a classical HMM whose transition and emission statistics are set by the Choi matrices of $\rho_0 \in \mathcal{D}(\mathcal{H})$5 and $\rho_0 \in \mathcal{D}(\mathcal{H})$6. However, in the quantum setting the hidden state is the full density operator $\rho_0 \in \mathcal{D}(\mathcal{H})$7, and the classical notion of a stochastic process trajectory $\rho_0 \in \mathcal{D}(\mathcal{H})$8 is replaced by the evolution of operator-valued memory. The semigroup property $\rho_0 \in \mathcal{D}(\mathcal{H})$9 embodies the quantum Markov property. Not every quantum process admits such a classical reduction; non-commuting Kraus operators and general observables yield genuinely quantum output statistics, inaccessible to classical HMMs. This principal distinction underlies the quantum advantage for memory compression and expressiveness (Paris et al., 2015, Zonnios et al., 2024, O`Neill et al., 2012).

3. Temporal Correlations, Spectral Analysis, and Memory Complexity

HQMMs enable computation of output autocovariance and spectral measures. With the output observable

$$\mathcal{E}: \mathcal{D}(\mathcal{H}) \rightarrow \mathcal{D}(\mathcal{H})$$0

the mean output at time $$\mathcal{E}: \mathcal{D}(\mathcal{H}) \rightarrow \mathcal{D}(\mathcal{H})$$1 is $$\mathcal{E}: \mathcal{D}(\mathcal{H}) \rightarrow \mathcal{D}(\mathcal{H})$$2 and the autocovariance at lag $$\mathcal{E}: \mathcal{D}(\mathcal{H}) \rightarrow \mathcal{D}(\mathcal{H})$$3 is

$$\mathcal{E}: \mathcal{D}(\mathcal{H}) \rightarrow \mathcal{D}(\mathcal{H})$$4

For certain energy-driven quantum channels, maximizing entropy under a constraint on activation energy results in noise spectra $$\mathcal{E}: \mathcal{D}(\mathcal{H}) \rightarrow \mathcal{D}(\mathcal{H})$$5, offering a direct quantum stochastic mechanism for observed $$\mathcal{E}: \mathcal{D}(\mathcal{H}) \rightarrow \mathcal{D}(\mathcal{H})$$6-type scaling in biophysical signals (Paris et al., 2015).

Analysis of the transfer superoperator spectrum provides model-invariant lower bounds on memory dimension. If $$\mathcal{E}: \mathcal{D}(\mathcal{H}) \rightarrow \mathcal{D}(\mathcal{H})$$7 is the number of nonzero, distinct eigenvalues of the transfer operator, then the minimal HQMM dimension $$\mathcal{E}: \mathcal{D}(\mathcal{H}) \rightarrow \mathcal{D}(\mathcal{H})$$8 satisfies $$\mathcal{E}: \mathcal{D}(\mathcal{H}) \rightarrow \mathcal{D}(\mathcal{H})$$9, while the minimal classical HMM dimension must satisfy $\{\mathcal{M}_y\}_{y \in \mathcal{Y}}$0, reflecting a quadratic quantum advantage in memory (Zonnios et al., 2024).

Model Class	Min. Memory Bound	Rank Constraint
HQMM (quantum)	$\{\mathcal{M}_y\}_{y \in \mathcal{Y}}$1	Transfer: $\{\mathcal{M}_y\}_{y \in \mathcal{Y}}$2
HMM (classical)	$\{\mathcal{M}_y\}_{y \in \mathcal{Y}}$3	Transfer: $\{\mathcal{M}_y\}_{y \in \mathcal{Y}}$4

This gap is explicitly realized in families where quantum memory encodings allow strict reduction of dimension compared to all classical presentations.

4. Activated Measurement, Causal Structure, and Physical Realizations

The independent activated measurement protocol introduces probabilistic observation at each step: with probability $\{\mathcal{M}_y\}_{y \in \mathcal{Y}}$5 measurement is performed ($\{\mathcal{M}_y\}_{y \in \mathcal{Y}}$6 applied and $\{\mathcal{M}_y\}_{y \in \mathcal{Y}}$7 emitted), otherwise the system evolves via an energy-driven quantum channel $\{\mathcal{M}_y\}_{y \in \mathcal{Y}}$8. This interpolates between fully quantum stochastic evolution ($\{\mathcal{M}_y\}_{y \in \mathcal{Y}}$9) and purely Hamiltonian or Gibbs-preserving evolution ($y$0), enabling models that incorporate both quantum thermodynamics and measurement (Paris et al., 2015).

Recent work formalizes the effect of causal architecture—issuers of the order "emission-then-transition" versus "transition-then-emission"—on the observable process. These orderings coincide for entangled classical HMM lifts but generally yield distinct quantum processes with different temporal correlations and process-tensor entanglement, establishing causal order as an operational quantum resource (Souissi et al., 22 Feb 2026).

Physical realizations include open quantum systems with instantaneous feedback, in which system–environment coupling and measurement-induced stochastic evolution naturally implement the HQMM formalism. This provides a direct connection between HQMMs and quantum feedback control, Markovian master equations, and non-adaptive readout of matrix-product states (Clark et al., 2014, Monras et al., 2010).

5. Learning Algorithms and Parametric Structure

Learning HQMMs consists in estimating the initial state and Kraus operators subject to CPTP constraints. This is commonly posed as constrained log-likelihood maximization over a Stiefel manifold, where Kraus operator blocks are stacked into a tall matrix $y$1, with $y$2. Techniques include gradient-based optimization on the Stiefel manifold (retraction-based update methods) and blockwise Givens rotations preserving orthonormality (Srinivasan et al., 2017, Adhikary et al., 2019, Adhikary et al., 2019).

The general update rule for observed sequence data $y$3 is

$y$4

This framework supports both model selection (over latent dimension and Kraus multiplicities) and scalable training using mini-batch updates, which are critical for empirical tasks involving long, real-world sequences.

For circular HQMMs (c-HQMMs), a tensor-network learning algorithm based on cyclic locally purified state representations guarantees complete positivity and trace preservation at each update (Javidian et al., 2021).

6. Applications and Impact

HQMMs are deployed in a variety of domains requiring high-fidelity stochastic modeling of sequential data, particularly where classical HMMs are insufficient. For instance:

• Quantum modeling of ion channel kinetics in neuroscience, yielding models for the 1/f$y$5 noise spectrum observed in neural recordings (Paris et al., 2015);
• Quantum simulators for stochastic biophysical processes and quantum transport systems (Li et al., 2023);
• Physical characterization and exact representation of matrix product states and symmetry-protected topological phases, such as the AKLT chain (Souissi et al., 21 Dec 2025).

Empirical studies demonstrate that HQMMs provide improved description accuracy and greater modeling parsimony over classical HMMs for the same memory resources, including in probabilistic safety assessment and machine learning applications (Zaiou et al., 2022, Srinivasan et al., 2017). The quantum structure yields strictly better compression and expressive power, with realized advantages in temporal correlation, diversity of observable processes, and the capability to encode quantum-coherent effects within the latent process.

7. Structural Theorems and Identifiability

Fundamental structural results for HQMMs include:

• Every classical HMM of dimension $y$6 is realizable as an HQMM on a Hilbert space of dimension at most $y$7;
• Any HQMM with commuting Kraus operators is classically reducible, while generic non-commuting HQMMs are strictly quantum;
• Uniqueness up to unitary equivalence holds: two HQMMs generating the same output law are related by a unitary transformation on the hidden space;
• Under linear independence of Kraus operators, HQMM parameters are identifiable (up to the unitary gauge) from observable multi-time statistics (Paris et al., 2015).

These foundational theorems provide justification for the robustness of HQMM identification and the universality of lower bounds based on spectral invariants.

---

References:

• (Paris et al., 2015) Formalized Quantum Stochastic Processes and Hidden Quantum Models with Applications to Neuron Ion Channel Kinetics
• (Zonnios et al., 2024) Memory-minimal quantum generation of stochastic processes: spectral invariants of quantum hidden Markov models
• (Srinivasan et al., 2017) Learning Hidden Quantum Markov Models
• (Souissi et al., 22 Feb 2026) Causal Architecture in Hidden Quantum Markov Models
• (Clark et al., 2014) Hidden Quantum Markov Models and Open Quantum Systems with Instantaneous Feedback
• (Adhikary et al., 2019) Expressiveness and Learning of Hidden Quantum Markov Models
• (Monras et al., 2010) Hidden Quantum Markov Models and non-adaptive read-out of many-body states
• (Li et al., 2023) A new quantum machine learning algorithm: split hidden quantum Markov model inspired by quantum conditional master equation
• (Javidian et al., 2021) Learning Circular Hidden Quantum Markov Models: A Tensor Network Approach
• (Zaiou et al., 2022) A quantum learning approach based on Hidden Markov Models for failure scenarios generation
• (Souissi et al., 21 Dec 2025) A Hidden Quantum Markov model framework for Entanglement and Topological Order in the AKLT Chain