Mild Threshold Control in Engineering Systems

• Mild threshold control is a concept where smooth, absolutely continuous control policies use an exploding rate at an interior state to render the threshold an inaccessible boundary rather than a reflecting one.
• It is applied in diverse settings—from delay differential equations and apoptosis models to neuromorphic controllers—using polynomial feedback, stochastic tuning, or iterative calibration to achieve near-tight risk regulation.
• Canonical forms range from singular control with diverging rates to graded population responses, highlighting both the method’s versatility and the optimization challenges in dynamically adjusting thresholds.

Searching arXiv for the cited works and closely related threshold-control papers to ground the article. {"^^^^^^^^2query2 dynamics in time-delay systems: polynomial beta-control in a pressing process and connections to blow-up\"","max_results":5,"sort_by":"submittedDate","sort_order":"descending"} {"^^^^^^^^2query2 dynamics in time-delay systems: polynomial beta-control in a pressing process and connections to blow-up","max_results":^^^^2ti:\2query2^^^^} Searching arXiv for papers on threshold control, noisy thresholds, event-triggered control, and related formulations. Mild threshold control is an Editor’s term for a class of threshold-centered control strategies in which the threshold is not implemented solely as a hard switching surface. In the narrowest formal sense, it denotes an absolutely continuous singular-control policy whose rate explodes at an interior state, thereby creating an inaccessible boundary rather than a reflecting one (&&&^^^^2query2^^^^&&&). In a broader synthesis suggested by recent work, it also covers threshold mechanisms that are softened by smooth polynomial feedback, stochastic or heterogeneous thresholds, continuously tuned threshold parameters, or iterative score-threshold calibration that aims for nearly tight risk control rather than conservative over-enforcement (&&&^^^^2ti:\2^^^^&&&, Vilar, 2010, Joshi et al., 31 Dec 2025).

^^^^2ti:\2^^^^. Conceptual scope and terminology

The most explicit formal distinction appears in time-inconsistent singular stochastic control. There, strong threshold control is realized by Skorokhod reflection at a boundary PRESERVED_PLACEHOLDER_^^^^2query2^^^^, with control of the form

PRESERVED_PLACEHOLDER_^^^^2ti:\2^^^^

so the state can hit the boundary and is instantaneously pushed back by local time. By contrast, mild threshold control uses an absolutely continuous control rate $u_\beta(x)$ satisfying

$\lim_{x\uparrow \beta} u_\beta(x)=\infty,$

together with a scale-function condition that makes $\beta$ an entrance-not-exit boundary. The threshold is therefore not reflecting but inaccessible from below (&&&^^^^2query2^^^^&&&).

A broader use of the term is suggested by several papers that do not always use exactly the same phrase. In delayed press control, a threshold in the normalized initial velocity separates safe non-overshoot motion from overshoot, but the implemented policy is smooth and polynomial rather than bang-bang. The same paper explicitly characterizes the strategy as smooth polynomial feedback, soft saturation and switching, operation near but within the threshold, moderate nonlinearity $\beta$, and bounded control effort (&&&^^^^2ti:\2^^^^&&&). In apoptosis, each cell obeys a hard ATP death threshold, yet a population of heterogeneous thresholds produces a graded ensemble response; this yields a soft threshold at the population level although the single-cell rule remains binary (Vilar, 2010). In test-time LLM filtering, multiple risks are regulated by score thresholds chosen to be just strong enough to satisfy prescribed budgets, with finite-sample guarantees and near-tightness rather than large safety margins (Joshi et al., 31 Dec 2025).

Accordingly, mild threshold control is best understood as a family resemblance concept. The common structure is that a threshold still partitions behavior, but the transition around that partition is mediated by smoothing, heterogeneity, asymptotics, calibrated optimization, or dynamic adaptation.

2. Canonical mathematical forms

One canonical form arises in delay differential equations for pressing systems. After nondimensionalization, the generalized $\beta$-control model is

$\frac{du}{dt}(t)=w_0 |1-u(t-1)|^\beta,\qquad t>1,$

with $u(t)=w_0 t$ for $0\le t\le 1$, control parameter PRESERVED_PLACEHOLDER_^^^^2ti:\2query2^^^^, and polynomial degree PRESERVED_PLACEHOLDER_^^^^2ti:\2ti:\2^^^^. Numerical evidence suggests a threshold PRESERVED_PLACEHOLDER_^^^^2ti:\22^^^^ such that PRESERVED_PLACEHOLDER_^^^^2ti:\2^^^^3 yields non-overshoot convergence PRESERVED_PLACEHOLDER_^^^^2ti:\2^^^^4, whereas PRESERVED_PLACEHOLDER_^^^^2ti:\2^^^^5 yields overshoot. The same model admits the transformation PRESERVED_PLACEHOLDER_^^^^2ti:\2^^^^6, under which overshoot corresponds to finite-time blow-up, so the threshold becomes a blow-up threshold as well (&&&^^^^2ti:\2^^^^&&&).

A second form is the noisy-threshold ensemble model for apoptosis. At single-cell level, the rule is crisp: cell PRESERVED_PLACEHOLDER_^^^^2ti:\2^^^^7 dies when the ATP level PRESERVED_PLACEHOLDER_^^^^2ti:\2^^^^8 crosses below its threshold PRESERVED_PLACEHOLDER_^^^^2ti:\2^^^^9. Across the population, $u_\beta(x)$^^^^2query2^^^^ is sampled from a log-normal distribution $u_\beta(x)$^^^^2ti:\2^^^^, so the survival function

$u_\beta(x)$2

is smooth. The induced death rate is

$u_\beta(x)$3

and vanishes when ATP is increasing. The threshold is therefore hard microscopically but mild macroscopically: heterogeneity converts a discrete death event into a graded population response (Vilar, 2010).

A third form is threshold adaptation in spiking control. In the neuromorphic PID controller, Input-Weighted Threshold Adaptation updates an integral neuron’s firing threshold by

$u_\beta(x)$4

with clipping to $u_\beta(x)$5. The threshold becomes a slow internal state that accumulates signed input history. This is mild threshold control in the sense of incremental, bounded, spike-driven threshold modulation rather than abrupt threshold relocation (&&&^^^^2ti:\2query2^^^^&&&).

The narrow singular-control form differs from all three. There, mildness does not mean randomness or tuning; it means replacing singular reflection by an absolutely continuous control whose rate diverges at the boundary, so that the boundary is approached but never hit (&&&^^^^2query2^^^^&&&).

3. Representative control architectures

The literature uses thresholds in markedly different objects: state variables, uncertainty queues, firing thresholds, risk scores, and even control-node counts. The resulting architectures are heterogeneous, but several recurring patterns can be identified.

Domain	Threshold quantity	Control consequence
Delayed pressing (&&&^^^^2ti:\2^^^^&&&)	$u_\beta(x)$6, $u_\beta(x)$7	Safe/non-overshoot polynomial $u_\beta(x)$8-control with saturation
Apoptosis (Vilar, 2010)	Cell-specific ATP threshold $u_\beta(x)$9	Hard single-cell death, graded population survival
Persistent monitoring (&&&^^^^2ti:\2^^^^4&&&)	Matrix $\lim_{x\uparrow \beta} u_\beta(x)=\infty,$^^^^2query2^^^^ on node uncertainties	Stay/leave/routing decisions in a hybrid system
Event-triggered LQ control (&&&^^^^2ti:\2^^^^5&&&)	$\lim_{x\uparrow \beta} u_\beta(x)=\infty,$^^^^2ti:\2^^^^	No transmission and zero input inside a deadzone
Robust data-center control (&&&^^^^2ti:\2^^^^6&&&)	Orthant threshold vectors $\lim_{x\uparrow \beta} u_\beta(x)=\infty,$2	Switch server counts to target configurations
Neuromorphic PID (&&&^^^^2ti:\2query2^^^^&&&)	Adaptive neuronal firing threshold	Integral action via threshold drift
Threshold/XOR Boolean networks (&&&^^^^2ti:\2^^^^8&&&)	Minimal control-node set	Full-state controllability by sparse interventions
Multi-risk LLM filtering (Joshi et al., 31 Dec 2025)	Score thresholds $\lim_{x\uparrow \beta} u_\beta(x)=\infty,$3	Priority-ordered output modification at test time

In persistent monitoring on graphs, each agent carries an $\lim_{x\uparrow \beta} u_\beta(x)=\infty,$4 threshold matrix $\lim_{x\uparrow \beta} u_\beta(x)=\infty,$5. The diagonal entries $\lim_{x\uparrow \beta} u_\beta(x)=\infty,$6 determine when a node’s uncertainty has been reduced enough to permit departure, while off-diagonal entries $\lim_{x\uparrow \beta} u_\beta(x)=\infty,$7 determine when neighboring nodes become urgent enough to attract the agent. The policy is hybrid and event-driven, and the threshold parameters are optimized online by Infinitesimal Perturbation Analysis (&&&^^^^2ti:\2^^^^4&&&).

In event-triggered linear-quadratic control with limited control actions, the no-transmission set

$\lim_{x\uparrow \beta} u_\beta(x)=\infty,$8

forces $\lim_{x\uparrow \beta} u_\beta(x)=\infty,$9 whenever the state lies inside the deadzone. Here mildness is not smoothing of the control law but selective inaction: the system is allowed to drift open loop when the state is small enough (&&&^^^^2ti:\2^^^^5&&&).

In robust data-center control, threshold policies are defined on server-capacity states. In one dimension, lower and upper thresholds determine whether the controller turns servers on, turns them off, or holds the current capacity; in multiple dimensions, threshold vectors indexed by orthants play the same role. Under convexity assumptions on the QoS cost, the robust MDP admits optimal threshold policies (&&&^^^^2ti:\2^^^^6&&&).

Threshold/XOR Boolean networks introduce a different notion of mildness: not smoothness in state feedback, but small intervention sets. For XOR-BNs, controllability is equivalent to a linear span condition over $\beta$^^^^2query2^^^^; for threshold and majority-type BNs, the paper derives lower bounds and explicit constructions showing when the number of control nodes can be kept small (&&&^^^^2ti:\2^^^^8&&&).

4. Optimization, controllability, and computational structure

A central theme in recent work is that threshold values are not merely hand-designed; they are optimization variables.

In persistent monitoring, the threshold matrices $\beta$^^^^2ti:\2^^^^ induce a hybrid system whose event times depend on threshold crossings. IPA yields piecewise-constant state sensitivities and an event-driven gradient estimator for the average uncertainty cost. Thresholds are then updated by projected gradient descent. In the single-agent case, the IPA gradient is monotone, implying the optimal diagonal thresholds are zero: the agent should reduce node uncertainty to zero before departing. In multi-agent cases, optimal thresholds can remain strictly positive, giving genuinely milder policies in which agents leave before full depletion because other agents can complete service (&&&^^^^2ti:\2^^^^4&&&).

Event-triggered linear systems present a non-convex threshold-constrained optimization problem. Exact synthesis requires solving exponentially many quadratic programs indexed by threshold-region sequences, while proposed greedy and ADMM heuristics provide tractable approximations. In receding-horizon form, the resulting MPC law is uniformly practically asymptotically stable, with an ultimate bound that scales with the threshold parameter $\beta$2 (&&&^^^^2ti:\2^^^^5&&&).

For XOR-BNs, the optimization problem becomes algebraic. A general XOR-BN with system matrix $\beta$3 and control-node set $\beta$4 is controllable if and only if

$\beta$5

This characterization supports polynomial-time algorithms for computing control-node sets and explicit control signals. The same paper proves a best-case construction: for any $\beta$6 and any odd $\beta$7, there exists a controllable $\beta$8-node $\beta$9-$\beta$^^^^2query2^^^^-XOR-BN with one control node (&&&^^^^2ti:\2^^^^8&&&).

In MultiRisk, thresholds are optimized sequentially under multiple simultaneous constraints. MULTIRISK-BASE performs empirical generalized inversions of risk functions, while MULTIRISK modifies the empirical risks with finite-sample corrections under exchangeability. The guarantee is simultaneous control

$\beta$^^^^2ti:\2^^^^

and, under additional conditions, nearly tight lower bounds of order $\beta$2. Mildness here means that thresholds use almost all of each risk budget rather than enforcing a much stricter regime than requested (Joshi et al., 31 Dec 2025).

5. Threshold dynamics, stochasticity, and criticality

Several papers complicate the common intuition that threshold control is inherently deterministic and piecewise constant.

In virtual stick balancing, the classical event-driven hypothesis would activate control once the deviation exceeds a threshold. The reported data instead show broad, approximately exponential action-point distributions, and the proposed model treats activation as intrinsically stochastic. A smooth gating function,

$\beta$3

suppresses deterministic control near $\beta$4, while noise drives escape from the passive state. This does not furnish a fixed threshold in the usual sense; a plausible implication is a probabilistic soft-threshold interpretation, where activation likelihood increases with error but no single trigger value exists (Zgonnikov et al., 2014).

In excitable neural networks, control is most effective near criticality even though criticality is the noisiest regime. With additive global feedback of the form

$\beta$5

errors are minimal for the widest range of target firing rates near $\beta$6, while optimal control is slightly away from criticality for heterogeneous degree distributions. Threshold-like regulation is therefore easiest precisely where intrinsic fluctuations are largest (Finlinson et al., 2019).

The R-tipping literature contributes a different threshold object: a safe forcing speed. For the asymptotically autonomous scalar ODE $\beta$7, the paper proves that for each forcing arclength

$\beta$8

there exists a critical speed $\beta$9 such that any tipping forcing must satisfy $\beta$^^^^2query2^^^^. The bound is tight, and $\beta$^^^^2ti:\2^^^^ is strictly decreasing in $\beta$2. The threshold is therefore a minimal rate barrier separating safe and tipping trajectories (&&&3^^^^2query2^^^^&&&).

Supra-threshold LOD control in graphics inverts another common assumption. Classical perceptual LOD schemes rely on threshold visibility and typically remove more detail as eccentricity increases or contrast drops. The reported experiments instead find that supra-threshold LOD must support a task-dependent level of reliable perceptibility; above that level, perceptibility of manipulations should be minimized, while below it perceptibility must be maximized and LOD should be improved as eccentricity rises or contrast drops. The paper explicitly states that this directly contradicts prevailing threshold-based LOD control schemes (&&&3^^^^2ti:\2^^^^&&&).

6. Limits, controversies, and open directions

The concept is broad, but its mathematical status is uneven across subfields. In delayed pressing, the threshold $\beta$3 separating overshoot and non-overshoot is formulated as Conjecture 3.^^^^2ti:\2^^^^, supported by numerical monotonicity and bisection estimates; even the empirical fit

$\beta$4

is explicitly numerical rather than proved (&&&^^^^2ti:\2^^^^&&&). This makes the control law practically useful but leaves the threshold theory incomplete.

In human intermittent control, the authors argue against noise-affected fixed-threshold models, yet also note that more complex deterministic threshold models are not exhaustively ruled out. The controversy is therefore not whether thresholds matter at all, but whether control activation is better modeled as deterministic crossing or intrinsic stochasticity (Zgonnikov et al., 2014).

The narrow singular-control notion of mild threshold control remains highly specialized. The 2^^^^2query22^^^^5 stochastic-control paper presents the first example of a singular control problem whose equilibrium uses an exploding absolutely continuous rate to create an inaccessible boundary, and does so precisely because no strong threshold equilibrium exists for certain parameter values (&&&^^^^2query2^^^^&&&). A plausible implication is that other time-inconsistent singular problems may also require enlarging the control class beyond reflection and jumps.

For multi-risk thresholding in generative AI, guarantees depend on exchangeability, bounded costs, and the validity of the chosen scores as proxies for the actual risks of interest. The algorithm can make threshold selection nearly tight relative to those scores, but it does not solve score misspecification or distribution shift by itself (Joshi et al., 31 Dec 2025).

Finally, minimal intervention results in Boolean networks show that threshold control cannot be arbitrarily mild in general. Majority-type threshold BNs satisfy lower bounds on the number of control nodes, whereas best-case XOR constructions can be controlled by a single node. This establishes a structural divide: some architectures admit vanishing intervention fractions, while others impose irreducible control budgets (&&&^^^^2ti:\2^^^^8&&&).

Taken together, these results suggest that mild threshold control is not a single method but a research program. Its unifying problem is how to preserve the decisiveness of thresholding while avoiding the brittleness of hard switching, excessive conservatism, or unattainable control requirements.