Monotone-Separable Loynes' Theorem Extension

• The paper extends Loynes' theorem by introducing a monotone-separable stochastic recurrence equation that rigorously addresses stability in multiserver-job queueing models.
• It employs subadditive ergodic theory to derive explicit stability conditions and convergence properties of the workload process in complex stochastic networks.
• The methodology supports advanced sampling algorithms, enabling efficient performance evaluation and stability assessment in high-dimensional queueing systems.

The monotone-separable extension of Loynes' theorem generalizes classical queue stability results to a broad class of stochastic networks, including multiserver-job queuing models (MJQM) under FCFS policies with stochastic, potentially multi-resource jobs. At its mathematical core is a Stochastic Recurrence Equation (SRE) that encapsulates queue evolution via a monotone-separable mapping, enabling rigorous analysis of workload stationarity, uniqueness, and system stability. The extension admits explicit stability criteria via subadditive ergodic theory and enables advanced sampling algorithms for performance evaluation (Baccelli et al., 28 Jan 2026).

1. Multiserver-Job Queuing Model and Stochastic Recurrence Equation

The FCFS MJQM is defined for $s$ identical servers and infinite queue capacity. Jobs indexed by $n \in \mathbb{Z}$ arrive at epochs $t_n$, with interarrival times $\tau_n = t_{n+1} - t_n$. Job $n$ requests $\alpha_n \in \{1,\ldots,s\}$ servers for service time $\sigma_n > 0$. The system state at arrival is the non-decreasing workload vector $$W_n = (W_n^1 \le \ldots \le W_n^s) \in \mathbb{R}_+^s$$, representing residual work on each server.

The queuing dynamics evolve by the SRE: $$W_{n+1} = \mathcal{R}\big[ S(\alpha_n, W_n) + \sigma_n L(\alpha_n) - \tau_n U \big]^+$$ where:

• $U = (1,\ldots,1) \in \mathbb{R}^s$ (unit decrement per server during $\tau_n$),
• $L(\alpha)$ sets the first $\alpha$ coordinates to 1 (rest zero),
• $S(\alpha, W)^i = \max(W^\alpha, W^i)$ enforces HoL blocking,
• $(\cdot)^+$ applies coordinatewise $\max(\cdot, 0)$,
• $\mathcal{R}(\cdot)$ sorts its argument in non-decreasing order.

Abstractly, $$W_{n+1} = f(W_n; \xi_n), \; \xi_n = (\alpha_n, \sigma_n, \tau_n)$$.

2. Monotone-Separable Framework

A stochastic network is monotone-separable if its SRE mapping $f$ satisfies four structural properties (all comparisons coordinatewise):

• Causality: Activity time $X_{[m,n]} \ge t_n$.
• External Monotonicity: Later arrivals result in later last activity, i.e., replacing $\{t_k\}$ by $\{t'_k\}$ with $t'_k \ge t_k$ yields $X'_{[m,n]} \ge X_{[m,n]}$; equivalently, $W \le W' \implies f(W; \xi) \le f(W'; \xi)$.
• Time-Shift Invariance (Homogeneity): Shifting all arrival times by $c$ shifts $X_{[m,n]}$ by $c$.
• Separability: If $X_{[m,l]} \le t_{l+1}$ (system empties), $X_{[m,n]} = X_{[l+1,n]}$ (future only depends on subsequent arrivals).

Within the MJQM, these properties are realized by the SRE's coordinatewise monotonicity (Lemma 1), and its structure ensures causality, homogeneity, and separability by construction. Thus, MJQM belongs to the monotone-separable class (Theorem 2).

Property	Description	Application in MJQM
Causality	$X_{[m,n]} \ge t_n$	Arrival-indexed activity
External Monotonicity	Later arrivals → later exit	Verified by Lemma 1
Homogeneity	Time-shift invariant	SRE built to preserve
Separability	Empty → reset dependency	Follows from SRE

3. Monotone-Separable Loynes' Theorem

The extension generalizes Loynes' classical result beyond single-server queues. Let $\{\xi_n\}$ be stationary ergodic inputs. The backward Loynes sequence is defined recursively: $$M_0 = 0, \quad M_{n+1} \circ \theta = f(M_n; \xi_0)$$ where $\theta$ is the left-shift operator on the input sequence.

Theorem (Monotone-Separable Loynes):

• The sequence $\{M_n\}$ is non-decreasing and converges a.s. to $$M_\infty(\xi) = \lim_{n \to \infty} M_n(\xi) \in [0, \infty]^s$$ (Coupling-from-the-past).
• $M_\infty$ solves the stationary equation $M = f(M; \xi_0)$ under shift $\theta$.
• Any forward-stated iteration $$W_n = f \circ \cdots \circ f(W_0; \xi_0, \ldots, \xi_{n-1})$$ converges in distribution to $M_\infty$ for any initial $W_0 \in \mathbb{R}_+^s$.
• $M_\infty < \infty$ a.s. iff the system is stable, else $M_\infty = \infty$ a.s.

The theorem requires only stationary, ergodic $\{\xi_n\}$ with finite mean service time $E[\sigma_n] < \infty$ (ensuring classical Loynes principle applies to 1-server marginals).

4. Proof Outline and Mathematical Significance

The proof exploits coordinatewise monotonicity. Initiating the queue empty at time $-n$ and evolving forward yields a workload process $M_n(\xi)$ at time zero that is non-decreasing in $n$. Stationarity and homogeneity of inputs ensure the process law is independent of time origin. By ergodicity, $M_n$ converges a.s. to $M_\infty$. If $M_\infty$ is finite, it satisfies the fixed-point $M_\infty = f(M_\infty; \xi_0)$ and is the unique stationary solution under classic integrability conditions. If infinite in any coordinate, there exists no finite stationary solution.

This extension elevates the applicability of Loynes' theorem to a wide class of queueing networks, particularly those admitting monotone-separable SRE representations such as MJQM.

5. Stability Condition via Saturation Rule

Stability is characterized using the saturation rule, employing a "pile" process (zero interarrival times, $\tau_n \equiv 0$): $$H_0 = 0, \quad H_{n+1} = \mathcal{R}[S(\alpha_n, H_n) + \sigma_n L(\alpha_n)]$$ Define $|H_n|_\infty$ (max coordinate), forming a subadditive process. Kingman's subadditive ergodic theorem asserts: $$\gamma := \lim_{n \to \infty} \frac{|H_n|_\infty}{n}$$ exists a.s. and is constant. For Poisson arrivals of rate $\lambda$, stability ($M_\infty < \infty$ a.s.) holds if and only if: $\lambda < \lambda_c := \frac{1}{\gamma}$ In specific cases (e.g., jobs requiring all $s$ servers with fraction $p_g$), $\gamma$ can be computed via renewal-cycle arguments.

6. Numerical and Algorithmic Implications

The monotone-separable framework provides foundations for sampling algorithms of stationary workloads. Both perfect and sub-perfect samples (SPS) can be drawn, with sub-perfect sampling enabling scalable GPU parallelization for efficient performance metric evaluation. The estimation of $\lambda_c$ uses simulation of the pile process $H_n$, yielding explicit numerical procedures for system stability assessment. These procedures generalize to more complex systems such as MJQM with typed resources (Baccelli et al., 28 Jan 2026).

A plausible implication is that monotone-separable extensions facilitate advanced simulation techniques in high-dimensional queueing systems and support rigorous stability analysis beyond classical models.