Overview

A Markov jump linear system (MJLS) is a discrete-time linear system whose dynamics matrices switch according to a finite-state Markov chain — i.e. x_{t+1} = A(θ_t) x_t + B(θ_t) u_t + w_t, with θ_t Markov. The MJLS framework provides the right theoretical home for stability, optimal control, filtering, and H∞ analysis of regime-switching linear-Gaussian models. Key technical objects are coupled Riccati difference / algebraic equations (one Riccati per mode, coupled through the transition probabilities) and the mean-square-stability (MSS) characterization via the spectral radius of an operator on sequences of matrices. The canonical reference is the Costa–Fragoso–Todorov textbook; risk-sensitive variants and impossibility results are due to Moon & Başar.

For the CRE asset-pricing project, MJLS theory is the mathematical backbone for: (a) checking determinacy / stability of the regime-switching macro RE solution, (b) understanding why the asset-pricing Riccati recursions converge (and when they don’t), and (c) providing the theoretical justification for the Rao-Blackwellized particle filter’s per-particle Kalman recursions.

Timeline

  • 1990s — early MJLS results on coupled Riccati equations and MSS (Abou-Kandil et al.; Costa–Fragoso, Bertilsson; Krtolica et al.).
  • 2005 — Costa, Fragoso & Marques: Discrete-Time Markov Jump Linear Systems (Springer textbook, often cited as CFT). The standard reference; consolidates stability, optimal LQR control, filtering, H∞ control, and design techniques.
  • 2016 — Moon & Başar: risk-sensitive MJLS — proves that the clean risk-sensitive / H∞ duality from the jump-free LQG case does not extend to MJLS. No closed-form risk-sensitive optimal controller, no large-deviation or risk-neutral limits.

Seminal works

SOTA tracker

  • MSS characterization: spectral radius of the T operator on H^n sequences; equivalent to existence of a coupled Lyapunov solution. Settled.
  • Optimal LQR control: coupled algebraic Riccati equation (CARE), maximal solution is the stabilizing one. Settled.
  • Filtering with mode known vs mode unknown: mode-known case admits a coupled Kalman recursion (Markov jump filter); mode-unknown case has exponentially growing optimal filter complexity. The latter is the bottleneck that motivates particle filtering.
  • H∞ control: state-feedback case has clean LMI / coupled-Riccati conditions; output-feedback is harder.

Open problems

  • Risk-sensitive MJLS — no closed-form solutions; numerical approximation only (Moon & Başar 2016 negative result).
  • Coupled Riccati existence under degenerate transition matrices (absorbing or near-absorbing regimes) — convergence is fragile, relevant to monetary policy regimes with very persistent states.
  • Identifiability of mode-conditional parameters from observed mixed data is hard; related to the switching-state-estimation literature.

My position

The CRE project does not develop new MJLS theory but uses it as machinery. The most directly relevant chapters of the CFT textbook are:

  • Ch.3 (Stability): justifies the spectral-radius determinacy check.
  • Ch.4 (LQR / coupled Riccati): structural template for the asset-pricing Riccati recursion.
  • Ch.5 (Filtering): the mode-unknown filtering result is exactly why the project needs a particle filter.
  • Appendix A (CARE convergence): provides the convergence-of-Riccati existence theorems the asymptotic CRE pricing behavior depends on.

Research gaps

  • Numerical convergence theory for coupled exponential-quadratic Riccati recursions under near-absorbing regimes is essentially absent; the project encounters this in the form of state-dependent T_bar requirements (Apartment is the slow asset).
  • MJLS-aware particle filter design — the literature treats RBPF as a generic technique; there is little work on MJLS-specific variance-reduction or proposal distributions exploiting the mode-known Kalman structure.

Key people

(populated incrementally during ingest)