Continuous-Time Mixed Monotonicity

• Continuous-Time Mixed Monotonicity is defined by a decomposition function that splits a system's dynamics into increasing and decreasing components, with exact recovery on the diagonal.
• The methodology leverages monotone embeddings and sensitivity-based Jacobian bounds to compute tight, rectangular interval enclosures for reachable sets.
• Current research addresses limitations such as conservative over-approximations and high computational costs, while seeking efficient, less conservative tight decompositions.

Searching arXiv for recent and foundational papers on continuous-time mixed monotonicity to ground the article. Continuous-time mixed monotonicity is the property that a continuous-time system can be represented by a decomposition function that is increasing in one argument, decreasing in another, and exact on the diagonal, thereby embedding the original dynamics into a higher-dimensional monotone system. In the literature, the term is used both for vector fields and for flows: one line of work defines a mixed monotone system by requiring the vector field to admit such a decomposition, while another defines it by requiring the flow map $\Phi_t$ itself to be mixed monotone. Under standard regularity, local Lipschitzness, forward completeness, and positive invariance, mixed monotonicity of the vector field implies mixed monotonicity of the flow (Coogan et al., 2015, Yang et al., 2018).

1. Definition through decomposition and embedding

For an autonomous continuous-time system

$\dot x = G(x),\qquad x\in X\subseteq\mathbb{R}^n,$

a standard definition used in continuous-time mixed monotonicity requires a locally Lipschitz decomposition function

$g:X\times X\to\mathbb{R}^n$

such that $g(x,x)=G(x)$, $\frac{\partial g_i}{\partial x_j}(x,y)\ge 0$ for all $i\neq j$, and $\frac{\partial g_i}{\partial y_j}(x,y)\le 0$ for all $i,j$, whenever the derivatives exist (Coogan et al., 2015). In the neural ODE literature the same structure is written as $g(x,\hat x)$, with $g$ increasing in its first argument off-diagonally, decreasing in its second argument, and satisfying $\dot x = G(x),\qquad x\in X\subseteq\mathbb{R}^n,$0 (Sayed et al., 15 Oct 2025).

This decomposition induces the embedding system

$\dot x = G(x),\qquad x\in X\subseteq\mathbb{R}^n,$1

or equivalently

$\dot x = G(x),\qquad x\in X\subseteq\mathbb{R}^n,$2

The diagonal set $\dot x = G(x),\qquad x\in X\subseteq\mathbb{R}^n,$3 is invariant, and the embedding is monotone with respect to the orthant $\dot x = G(x),\qquad x\in X\subseteq\mathbb{R}^n,$4 or, equivalently, the order in which the first copy is ordered increasingly and the second decreasingly (Coogan et al., 2015, Sayed et al., 15 Oct 2025). This is the basic mechanism by which mixed monotonicity transfers monotone-systems tools to non-monotone dynamics.

A central clarification in the continuous-time literature is that mixed monotonicity is not merely monotonicity in another orthant. The traffic-network work explicitly notes that some FIFO traffic networks are not monotone with respect to any orthant, while remaining mixed monotone through the embedding construction (Coogan et al., 2015). This distinction is conceptually important: mixed monotonicity is a property of the decomposition or of the embedding, not of the original system viewed under a coordinate sign flip.

2. Flow maps, sampled-data maps, and the mixed-monotone viewpoint

A second formulation of continuous-time mixed monotonicity proceeds through the flow map. For the uncertain time-varying system

$\dot x = G(x),\qquad x\in X\subseteq\mathbb{R}^n,$5

with flow $\dot x = G(x),\qquad x\in X\subseteq\mathbb{R}^n,$6, reachable set

$\dot x = G(x),\qquad x\in X\subseteq\mathbb{R}^n,$7

and sampled-data map

$\dot x = G(x),\qquad x\in X\subseteq\mathbb{R}^n,$8

the reachability problem at a fixed sampling time $\dot x = G(x),\qquad x\in X\subseteq\mathbb{R}^n,$9 can be reformulated as a mixed-monotonicity question about the discrete-time map $g:X\times X\to\mathbb{R}^n$0 (Meyer et al., 2018).

The key objects are the sensitivities

$g:X\times X\to\mathbb{R}^n$1

which satisfy linear time-varying ODEs along the trajectory. At the sampling time $g:X\times X\to\mathbb{R}^n$2, these sensitivities are exactly the Jacobians of the sampled-data map: $g:X\times X\to\mathbb{R}^n$3 If each sensitivity entry is sign-stable over $g:X\times X\to\mathbb{R}^n$4, then each component of the sampled map is monotone in each state and parameter coordinate, with direction determined by the sign, and the paper proves that the sampled-data map is mixed-monotone in the sense of Coogan & Arcak (2015) (Meyer et al., 2018).

This equivalence gives a precise continuous-time interpretation: sign-stable sensitivities of the continuous-time flow at time $g:X\times X\to\mathbb{R}^n$5 are equivalent to sign-stable Jacobians of the sampled-data map, and therefore to discrete-time mixed monotonicity of that map. The same paper further shows that the sensitivity-based interval over-approximation and the mixed-monotone over-approximation coincide and are both tight under sign-stability (Meyer et al., 2018).

3. Sufficient conditions and generic existence results

One major line of work studies when a continuous-time system is mixed monotone without constructing the exact flow. For differentiable maps $g:X\times X\to\mathbb{R}^n$6 with partial derivatives bounded in intervals $g:X\times X\to\mathbb{R}^n$7, a Jacobian-bounds construction gives an explicit decomposition function

$g:X\times X\to\mathbb{R}^n$8

where $g:X\times X\to\mathbb{R}^n$9, $g(x,x)=G(x)$0, and $g(x,x)=G(x)$1 are chosen componentwise according to whether each derivative interval is sign-stable positive, sign-unstable “positive,” sign-unstable “negative,” or sign-stable negative (Yang et al., 2018). This strictly generalizes the sign-stable Jacobian condition used in earlier mixed-monotone results.

A particularly strong consequence stated in that work is that any $g(x,x)=G(x)$2 map whose partial derivatives are continuous on a compact domain is mixed monotone on that domain (Yang et al., 2018). The same paper also provides a one-dimensional route via bounded variation: every scalar function of bounded variation on an interval is mixed monotone through a Jordan decomposition into increasing and decreasing parts, and a $g(x,x)=G(x)$3 formula based on the negative part of $g(x,x)=G(x)$4 yields a more explicit decomposition in one dimension (Yang et al., 2018). The paper characterizes these conditions as evidence that mixed monotonicity is a very generic property.

A more aggressive existence result is given for disturbed systems

$g(x,x)=G(x)$5

with locally Lipschitz $g(x,x)=G(x)$6. There, every such system is shown to be mixed-monotone with respect to a decomposition function $g(x,x)=G(x)$7 defined componentwise by a pointwise optimization over the state–disturbance box: $g(x,x)=G(x)$8 in the forward-ordered case, and symmetrically by a maximization in the reverse-ordered case (Abate et al., 2020). This result does not require global sign conditions on the Jacobian; it requires local Lipschitz continuity and obtains a Lipschitz decomposition through optimization.

4. Reachability, tightness, and interval enclosures

The standard reachability consequence of mixed monotonicity is an interval enclosure obtained from the embedding. For a mixed monotone map $g(x,x)=G(x)$9 with decomposition $\frac{\partial g_i}{\partial x_j}(x,y)\ge 0$0, the Coogan–Arcak bound gives

$\frac{\partial g_i}{\partial x_j}(x,y)\ge 0$1

and for a continuous-time system this is applied to the flow map $\frac{\partial g_i}{\partial x_j}(x,y)\ge 0$2 (Yang et al., 2018). In the sampled-data sensitivity framework, sign-stability yields a tight interval enclosure of $\frac{\partial g_i}{\partial x_j}(x,y)\ge 0$3 by evaluating $\frac{\partial g_i}{\partial x_j}(x,y)\ge 0$4 at extremal vertices determined by the signs of the sensitivities. Under that assumption, only up to $\frac{\partial g_i}{\partial x_j}(x,y)\ge 0$5 trajectory simulations are needed to obtain the smallest interval containing the reachable set (Meyer et al., 2018).

The same sampled-data paper weakens sign-stability to bounded sensitivities. If

$\frac{\partial g_i}{\partial x_j}(x,y)\ge 0$6

possibly with intervals containing $\frac{\partial g_i}{\partial x_j}(x,y)\ge 0$7, then a generalized interval over-approximation is obtained by selecting vertices according to the sign of the interval centers and adding compensation terms $\frac{\partial g_i}{\partial x_j}(x,y)\ge 0$8. The construction is derived by defining an auxiliary modified flow

$\frac{\partial g_i}{\partial x_j}(x,y)\ge 0$9

whose sensitivities are sign-stable by design (Meyer et al., 2018). The resulting method still requires at most $i\neq j$0 trajectory simulations once sensitivity bounds are known, so its cost grows linearly in the state dimension, although the enclosure is not tight in general.

The disturbed-system work introduces a different notion of tightness. Among all decomposition functions admissible for the standard embedding construction, the optimization-based $i\neq j$1 is maximal in the forward direction and minimal in the reverse direction, which makes it the tight decomposition function in that framework (Abate et al., 2020). The corresponding embedding system produces the smallest hyperrectangular over-approximation obtainable by the standard mixed-monotonicity tools, and a backward-time construction yields under-approximations of reachable sets when the ordered-pair condition is preserved along the backward embedding (Abate et al., 2020).

A plausible implication is that the literature contains two complementary notions of “tightness.” In the sampled-data sensitivity setting, tightness refers to the smallest interval containing the reachable set under sign-stability (Meyer et al., 2018). In the optimization-based continuous-time setting, tightness refers to minimality within the class of embedding-based hyperrectangular approximations generated by decomposition functions (Abate et al., 2020).

5. Structured constructions and application domains

Continuous-time mixed monotonicity has been developed in several structured settings. In partial first-in-first-out traffic models, the continuous-time network dynamics

$i\neq j$2

are not generally monotone because congestion on one outgoing link can decrease FIFO flow to another while increasing non-FIFO flow. The decomposition is built by evaluating FIFO terms on hybrid states $i\neq j$3 that read adjacent-link densities from the second argument $i\neq j$4, while retaining cooperative terms in the first argument $i\neq j$5. Under the paper’s locality and sign assumptions, this yields a mixed-monotone decomposition for the full traffic network (Coogan et al., 2015). The same embedding is then used to establish convergence to an equilibrium in a diverging-junction example.

For polynomial functions, the theory has been pushed in a different direction. Every univariate polynomial is shown to be mixed monotone globally with a polynomial decomposition function constructed from a Gram matrix representation of the derivative. If

$i\neq j$6

and $i\neq j$7 with $i\neq j$8, then

$i\neq j$9

where $\frac{\partial g_i}{\partial y_j}(x,y)\le 0$0 and $\frac{\partial g_i}{\partial y_j}(x,y)\le 0$1, is a global polynomial decomposition (Tahir, 2023). The paper further states that such polynomial decompositions can be much tighter than local decomposition functions constructed using local Jacobian bounds, and it explicitly connects this result to continuous-time mixed-monotone embeddings for scalar polynomial ODEs and decoupled polynomial vector fields (Tahir, 2023).

Neural ODE verification has recently adopted a direct continuous-time mixed-monotone embedding. For $\frac{\partial g_i}{\partial y_j}(x,y)\le 0$2, Jacobian interval bounds $\frac{\partial g_i}{\partial y_j}(x,y)\le 0$3 on a reachable tube are used to define an off-diagonal shifting matrix $\frac{\partial g_i}{\partial y_j}(x,y)\le 0$4 so that $\frac{\partial g_i}{\partial y_j}(x,y)\le 0$5 is sign-stable. The decomposition is then written componentwise as

$\frac{\partial g_i}{\partial y_j}(x,y)\le 0$6

with $\frac{\partial g_i}{\partial y_j}(x,y)\le 0$7 if $\frac{\partial g_i}{\partial y_j}(x,y)\le 0$8 and $\frac{\partial g_i}{\partial y_j}(x,y)\le 0$9 otherwise (Sayed et al., 15 Oct 2025). Because the paper assumes smooth neural ODE vector fields, it states that any such neural ODE is mixed monotone on any compact subset, with a decomposition $i,j$0 constructed in this way. The resulting interval reachability methods, implemented in TIRA with single-step, incremental, and boundary-based variants, provide sound over-approximations while explicitly trading tightness for efficiency (Sayed et al., 15 Oct 2025).

6. Limitations, misconceptions, and open problems

The literature is consistent in emphasizing that mixed monotonicity is broad but not automatically computationally sharp. The Jacobian-bounds and bounded-variation conditions show that mixed monotonicity is generic on compact domains or intervals, but the corresponding decomposition functions may be too conservative for quantitative reachability unless derivative bounds are informative (Yang et al., 2018). The optimization-based tight decomposition $i,j$1 removes this conservatism relative to other embedding-based decompositions, but evaluating $i,j$2 requires solving optimization problems pointwise in the extended state–disturbance space, and the paper notes that the number of pieces in a closed-form expression can grow exponentially with state and disturbance dimension (Abate et al., 2020).

Another recurring limitation is geometric. The sampled-data sensitivity method and the neural ODE methods both return rectangular over-approximations. One paper explicitly notes that the current approach yields only rectangular over-approximations and that more refined shapes such as zonotopes or polytopes could be more accurate but require other methods (Meyer et al., 2018). The neural ODE paper makes the same point in different language: interval boxes are inherently conservative because they cannot align with the geometry of spirals or rotated ellipsoids, even though they are extremely simple geometrically and computationally (Sayed et al., 15 Oct 2025).

There is also a misconception that incrementalization or boundary treatment necessarily improves results. In the neural ODE experiments, incremental and single-step continuous-time mixed monotonicity produced identical interval over-approximations, but incremental was much slower; boundary-based reachability likewise gave similar or identical boxes while incurring higher runtimes on the reported examples (Sayed et al., 15 Oct 2025). This suggests that the principal benefit of continuous-time mixed monotonicity in those settings is computational simplicity rather than systematic geometric refinement.

Open problems are stated with similar precision across the papers. For sampled-data reachability, efficient, guaranteed tight bounds on sensitivities for large systems and long times remain a central problem (Meyer et al., 2018). For tight decomposition theory, efficient computation or approximation of the optimization-based $i,j$3 is an obvious unresolved issue (Abate et al., 2020). For neural ODEs, the role of richer partitioning strategies, adaptive tube refinement, and extension to input-varying neural ODEs is left open (Sayed et al., 15 Oct 2025). Taken together, these results suggest that continuous-time mixed monotonicity is best understood as a unifying structural principle—one that connects vector-field decompositions, flow-map sensitivities, monotone embeddings, and interval reachability—while leaving substantial room for sharper geometry and more scalable computation.