Downward Conditional Monotonicity

• Downward conditional monotonicity is a directional property ensuring that adding computation or conditioning information results in non-decreasing outcomes, such as higher prediction quality or maintained stochastic order.
• In anytime neural networks, post-hoc methods like Product Anytime (PA) enforce this property by incrementally improving per-input prediction probabilities across deeper exits.
• In Markov-modulated Poisson processes, the property ensures that the conditional environmental law converges stochastically downward toward an optimal equilibrium measure.

Downward conditional monotonicity is a directional conditional-order property whose precise meaning depends on the modeling framework. In early-exit neural networks for anytime classification, it requires that per-input prediction quality be non-decreasing as exits deepen, typically through the condition $p_{k+1}(y^*\mid x)\ge p_k(y^*\mid x)$ for all inputs and exits (Jazbec et al., 2023). In Markov-modulated Poisson processes, it is a stochastic-order condition for the environmental law given the decreasing event of no arrivals up to time $t$, written $v^*\preceq x_t\preceq y_t$ for all $t>0$ (Stover, 8 Jun 2026). In several adjacent literatures, the exact term does not appear, but closely related notions are developed under names such as lower weak set monotonicity, nonincreasing conditional quantile curves, downward monotone contexts, and non-increasing conditional curves (Che et al., 2019, Birke et al., 2016, Rozanova et al., 2021, Gupta et al., 2019).

1. Principal meanings and terminological scope

In the cited literature, the same phrase names structurally similar but domain-specific properties: a conditional object is required to move only in one direction when either more computation is allocated, more conditioning information is imposed, or a designated argument varies in an ordered way. The common motif is not a single universal formalism, but a one-sided monotonicity requirement under conditioning.

Literature	Conditioned object	Downward meaning
Early-exit neural networks	Per-input prediction quality across exits	Deeper exits should not reduce quality
Markov-modulated Poisson processes	Environmental law given no arrivals	Conditioning on no arrivals drives the law downward toward $v^*$ while preserving stochastic order from below
Adjacent literatures	Choice sets, quantile curves, contexts, or feature-conditional predictions	Closest notions are lower weak set monotonicity, nonincreasing quantile curves, downward monotone contexts, and non-increasing feature effects

The early-exit usage is computational and per-example. The MMPP usage is probabilistic and order-theoretic. In the other cited works, the phrase is either absent or used only as an interpretive mapping. A plausible implication is that “downward conditional monotonicity” is best treated as a family of directional conditional-order constraints rather than a single cross-disciplinary definition.

2. Early-exit architectures and anytime classification

For early-exit neural networks, an EENN defines $M$ exits, each exit $m$ producing a predictive distribution $p_m(y\mid x)$ over $C$ classes from logits $f_m(x)\in\mathbb{R}^C$. The paper defines two per-input quality measures at exit $t$0: the ground-truth probability

$t$1

and the correctness indicator

$t$2

The primary conditional quality measure is $t$3 because it is a continuous signal in $t$4 and is more informative than the binary correctness indicator for per-example monotonicity (Jazbec et al., 2023).

With $t$5 indexing exits and $t$6, downward conditional monotonicity requires

$t$7

For EENNs, the core target is

$t$8

Here “downward” means deeper exits. The paper emphasizes that marginal monotonicity is insufficient for anytime decision-making because averages can improve while some inputs get worse; the stronger conditional, instance-level property is the relevant one for per-example anytime guarantees.

The proposed post-hoc transformation is Product Anytime (PA), a Product-of-Experts modification applied without retraining. The idealized hard PoE ensemble at exit $t$9 is

$v^*\preceq x_t\preceq y_t$0

with threshold $v^*\preceq x_t\preceq y_t$1, positive weights $v^*\preceq x_t\preceq y_t$2, and normalizer

$v^*\preceq x_t\preceq y_t$3

Under the Iverson mapping, if a class $v^*\preceq x_t\preceq y_t$4 is in the final product support, then for every input $v^*\preceq x_t\preceq y_t$5 and every exit $v^*\preceq x_t\preceq y_t$6,

$v^*\preceq x_t\preceq y_t$7

If the true class $v^*\preceq x_t\preceq y_t$8 is in the final support, then $v^*\preceq x_t\preceq y_t$9 is guaranteed to be non-decreasing across exits. The intuition given in the paper is that PoE acts like an “AND” over exits: each exit can only remove classes from support, the support shrinks monotonically with $t>0$0, and the probability of each surviving class increases.

Because the hard PoE is too blunt, the practical PA relaxation replaces the indicator with ReLU:

$t>0$1

with weights fixed as $t>0$2. In log space,

$t>0$3

If supports do not overlap and $t>0$4, PA falls back to the baseline softmax $t>0$5 at the current exit. The paper also introduces a clipped ReLU family,

$t>0$6

to interpolate between the hard Heaviside mapping and ReLU. As $t>0$7, the mapping approaches the hard construction; as $t>0$8, it becomes ReLU.

A second post-hoc baseline is CA, which caches the best-so-far prediction using the model confidence

$t>0$9

The cached prediction is overwritten when $v^*$0. CA gives a strong monotonicity bias in the maximum-confidence sense, but it relies on exit calibration and does not enforce product support shrinkage.

The recommended protocol is post-hoc application at test time: compute logits $v^*$1, compute the non-negative factors $v^*$2, accumulate the product incrementally, normalize when the sum is positive, and otherwise fall back to softmax. Optional extensions include adaptive thresholding $v^*$3 for calibration and finetuning with PA’s negative log-likelihood, Softplus activation, and weights $v^*$4 in the last training phase.

3. Markov-modulated Poisson processes and contact processes in random environments

In the MMPP setting, the state space is a totally ordered finite set $v^*$5, the environment is a continuous-time Markov chain $v^*$6 on $v^*$7 with generator $v^*$8, and the arrival process has rate vector $v^*$9. The arrival history up to time $M$0 is denoted $M$1, and the decreasing event of no arrivals on $M$2 is

$M$3

For initial distribution $M$4, the conditional environmental law is

$M$5

where $M$6 (Stover, 8 Jun 2026).

The equilibrium no-arrival distribution is

$M$7

which exists under irreducibility of $M$8 and nonnegativity of $M$9 with at least one positive component. It is the unique strictly positive left eigenvector of $m$0 associated to its minimal-modulus eigenvalue $m$1, normalized by $m$2:

$m$3

Stochastic domination between row probability vectors is written $m$4 when for every threshold $m$5,

$m$6

The MMPP is downward conditionally monotone if, whenever two copies share $m$7 and satisfy $m$8, then

$m$9

for all $p_m(y\mid x)$0, equivalently

$p_m(y\mid x)$1

The paper states that “downward” refers both to the event being decreasing in the space of counting paths and to the fact that $p_m(y\mid x)$2 is always stochastically larger than, but converges down to, $p_m(y\mid x)$3.

For $p_m(y\mid x)$4, with $p_m(y\mid x)$5 irreducible and monotone and $p_m(y\mid x)$6 increasing, DCM holds if and only if for all triples $p_m(y\mid x)$7,

$p_m(y\mid x)$8

All two-state models are DCM, and indeed satisfy the stronger full conditional monotonicity property.

A central consequence is optimal stochastic domination by a homogeneous Poisson process. If $p_m(y\mid x)$9 is a Poisson counting process of rate $C$0, $C$1 is irreducible and monotone, $C$2 is increasing, the MMPP is DCM, and the initial distribution satisfies $C$3, then

$C$4

and

$C$5

The conditional intensity function is

$C$6

and DCM yields the lower bound

$C$7

The same paper applies this machinery to child-CPMRE on $C$8. Each site has an independent environmental chain, infection occurs at rate $C$9 times the number of infected neighbors when the site is healthy and in environment state $f_m(x)\in\mathbb{R}^C$0, recovery occurs at rate $f_m(x)\in\mathbb{R}^C$1 when the site is infected and in state $f_m(x)\in\mathbb{R}^C$2, and environmental transitions follow $f_m(x)\in\mathbb{R}^C$3. Under DCM and suitable initial conditions, the graphical construction gives

$f_m(x)\in\mathbb{R}^C$4

If the infection MMPP is DCM and

$f_m(x)\in\mathbb{R}^C$5

the process survives strongly. If the permutation-reversed recovery MMPP is DCM and

$f_m(x)\in\mathbb{R}^C$6

the process dies out almost surely.

4. Analogues and neighboring formulations

In weak monotone comparative statics, the exact term “Downward Conditional Monotonicity” does not appear. The closest internal notions are lower weak set dominance, written $f_m(x)\in\mathbb{R}^C$7, and lower weak set monotonicity of correspondences (Che et al., 2019). These are the conditions guaranteeing monotonicity of minimal selections when parameters increase. The paper’s lower comparative statics theorem states that if a self-correspondence belongs to the lower monotonicity class $f_m(x)\in\mathbb{R}^C$8 and $f_m(x)\in\mathbb{R}^C$9 for all $t$00, then the fixed-point sets satisfy

$t$01

The mapping supplied in the paper identifies this lower weak set direction as the formal analogue of a downward monotonicity requirement.

In conditional quantile location-scale models, downward monotonicity means that the conditional quantile curve is nonincreasing:

$t$02

The downward-constrained estimator is defined by

$t$03

and the associated independence process is

$t$04

on the trimmed interval for $t$05 (Birke et al., 2016). The asymptotic result for the increasing case is stated under $t$06; for downward monotonicity the data block specifies replacing this by $t$07.

In monotonic deep learning, downward conditional monotonicity with respect to a feature $t$08 means that for all $t$09 and $t$10 that are identical except $t$11, one requires

$t$12

equivalently $t$13 almost everywhere (Gupta et al., 2019). The paper’s point-wise loss for a downward constraint reduces to penalizing positive gradients,

$t$14

inside the combined objective $t$15. The method is explicitly a soft-constraint approach and does not guarantee strict monotonicity at convergence.

In neural natural language inference, downward monotonicity is formulated as antitonicity of a context. If $t$16 and $t$17 is downward monotone, then

$t$18

is licensed (Rozanova et al., 2021). The same directional reversal is described for the material conditional: the antecedent of $t$19 is antitone and the consequent is monotone, so if $t$20 and $t$21, then

$t$22

(Chen, 2021). Here the “downward” direction concerns entailment reversal under substitution in designated positions.

In comparison-based preference learning, the exact term again does not appear, but the paper maps it to “downward conditional monotonicity in pairwise odds/logits” and to the stronger notion of “downward conditional monotonicity in individual probability” (Bareilles et al., 10 Jun 2025). Under mild assumptions, the paper proves local pairwise monotonicity. The stronger individual-probability monotonicity can fail unless fully pairwise monotonicity or demanding structural conditions hold.

5. Guarantees, diagnostics, and failure modes

The early-exit literature distinguishes formal guarantees from empirical enforcement. For PA with ReLU, the paper measures monotonicity violations through the maximum probability decrease

$t$23

and for thresholds $t$24 counts

$t$25

On CIFAR-100 with baseline MSDNet, about $t$26 of test points have a drop $t$27 somewhere; with PA, the violation curve collapses drastically. On ImageNet, PA exhibits effectively no drop beyond $t$28 for any test example across MSDNet, IMTA, and DViT backbones. For correctness monotonicity on CIFAR-100 with MSDNet, the reported values are: baseline $t$29, $t$30; PA $t$31, $t$32; CA $t$33, $t$34 (Jazbec et al., 2023).

The same paper reports that PA preserves average test accuracy per exit across all datasets and models, and on CIFAR-100 often improves MSDNet’s accuracy. It also reports decreasing entropy per exit, decreasing conformal predictive set sizes under RAPS, and substantial reductions in overthinking measures. At the same time, PA can raise early-exit ECE because of underconfidence bias; the suggested mitigations are adaptive thresholds, finetuning with the PA objective and Softplus activation, and the fallback to softmax when conflicting experts produce zero products.

For MMPPs, the central guarantee is exact: the DCM inequality is necessary and sufficient for $t$35 under the stated monotonicity assumptions, and all two-state models are DCM (Stover, 8 Jun 2026). The paper also stresses several limitations. DCM requires a finite-state environment, a monotone generator $t$36, and monotone rate vectors; the environment processes at distinct sites are independent; and DCM is sufficient but not necessary for the optimal Poisson domination at $t$37. For $t$38, DCM does not imply full conditional monotonicity in general.

In comparison-based preference learning, the failure mode is different. The paper reports that after accounting for a preference for response $t$39 over $t$40, the model may actually decrease the probability and reward of generating $t$41 (Bareilles et al., 10 Jun 2025). The local guarantee is on pairwise score differences or log-odds, not on individual probabilities. Downward versions are obtained by flipping the roles of $t$42 and $t$43 or by conditioning on evidence against $t$44. This makes explicit the distinction between local pairwise monotonicity and stronger individual-probability monotonicity.

In gradient-penalized deep networks, residual violations are expected because the method is a soft penalty rather than an architectural guarantee (Gupta et al., 2019). The paper contrasts this with deep lattice networks, which guarantee monotonicity but tend to produce step-wise conditional curves. The trade-off is stated directly: as $t$45 increases, monotonicity improves but AUC may decrease if the data violates the assumed monotone trend.

6. Significance and conceptual boundaries

In anytime computation, downward conditional monotonicity is described as essential because the environment can interrupt at any time and deeper exits should never harm per-example quality (Jazbec et al., 2023). The property directly ties additional computation to non-decreasing per-input prediction quality and to progressively more confident uncertainty behavior.

In random-environment point processes, DCM is significant because it yields a uniform lower bound on the conditional intensity function, identifies the optimal Poisson rate $t$46 through the eigenpair of $t$47, and enables direct comparison arguments for survival and extinction of contact processes in random environments (Stover, 8 Jun 2026). The property therefore functions as a bridge from a time-varying modulated process to a homogeneous Poisson proxy.

In the adjacent literatures, the same directional intuition reappears under different formalisms. Lower weak set monotonicity tracks minimal selections in comparative statics; constrained rearrangement methods test whether a conditional quantile curve is globally nonincreasing; monotonicity-aware NLI models encode entailment reversal in downward contexts; and gradient penalties or pairwise-logit conditions attempt to control how predictions move when inputs or preference evidence change (Che et al., 2019, Birke et al., 2016, Rozanova et al., 2021, Chen, 2021, Gupta et al., 2019, Bareilles et al., 10 Jun 2025).

A plausible implication is that the unifying content of downward conditional monotonicity is directional robustness under conditioning: more depth, stronger conditioning, or movement along an ordered argument should not reverse the designated order relation. What differs across the literature is the object being ordered—probability mass on the true class, stochastic order on environmental states, minimal fixed points, quantile curves, entailment relations, or pairwise log-odds—and the strength of the guarantee, which ranges from exact theorems to empirical regularization.