Implementation Uncertainty Interval

• Implementation Uncertainty Interval is an interval-valued description capturing the gap between prescribed decisions and their actual realizations under uncertain conditions.
• It applies to robust binary linear programming, predictive control, and interval probability measures by modeling flip changes and bounded parameter deviations.
• Practical insights include efficient offline interval propagation, neural interval learning methods, and tuning through parameters like Γ and δ to balance feasibility and profit.

An implementation uncertainty interval is an interval-valued description of the gap between a prescribed object and its realized counterpart at deployment. In robust binary linear programming, the canonical instance is the realization set for a prescribed binary vector $x\in\{0,1\}^n$, where uncertain components may flip during implementation and are equivalently represented by $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$ with $\Delta_i=1$ on uncertain indices and $\Delta_i=0$ on deterministic indices, subject to $\tilde x\in\{0,1\}^n$ (Ramirez-Calderon et al., 2021). In adjacent interval-based frameworks, the same lower/upper-bound formalism is used for uncertain system matrices, interval probability measures, and interval-valued neural parameters, yielding implementable uncertainty sets that can be propagated, tightened, or calibrated without requiring a single probabilistic parametric model (Quartullo et al., 19 Feb 2026, Basili et al., 2024, Betancourt et al., 2021).

1. Formal meaning of the interval representation

In the decision-variable setting, implementation uncertainty is defined by partitioning the variable indices into $C$, the indices of variables that cannot change at implementation, and $U$, the indices of variables that may flip. The implementation-uncertainty set is

$$U(x)=\{\tilde x\in\{0,1\}^n:\tilde x_i=x_i \text{ for } i\in C;\ \tilde x_i\in\{0,1\} \text{ for } i\in U\}.$$

The equivalent interval form is

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i],$

with $\Delta_i=1$ for $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$0 and $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$1 for $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$2, plus the integrality restriction $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$3. To restrict simultaneous implementation changes, one may impose

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$4

where $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$5 is a budget of implementation uncertainty (Ramirez-Calderon et al., 2021).

This formulation differs explicitly from parameter uncertainty: the uncertainty affects the implemented decision variables rather than model coefficients. The binary nature of the variables is central, because it invalidates the direct use of existing implementation-uncertainty models developed for continuous decision spaces (Ramirez-Calderon et al., 2021).

A more general interval semantics appears in imprecise-probability formalisms. An interval probability measure assigns to each event $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$6 an interval

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$7

such that $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$8 is an ordinary additive probability measure and the interval width is monotone in the sense that if $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$9, then $\Delta_i=1$0 (Basili et al., 2024). This suggests that an implementation uncertainty interval is best understood as an admissible set of realizations, not intrinsically as a confidence interval or posterior credible interval.

2. Robust optimization under implementation uncertainty

For a deterministic BLP

$\Delta_i=1$1

implementation uncertainty yields the robust model

$\Delta_i=1$2

subject to

$\Delta_i=1$3

with optional restriction to $\Delta_i=1$4 satisfying $\Delta_i=1$5. The parameters $\Delta_i=1$6 are right-hand-side relaxations, and $\Delta_i=1$7 controls how many uncertain variables may change at once (Ramirez-Calderon et al., 2021).

Two structural identities linearize the full-worst-case inner maximizations: $\Delta_i=1$8

$\Delta_i=1$9

Hence the uncertain variables $\Delta_i=0$0 for $\Delta_i=0$1 drop out of the master decision, reducing the search from $\Delta_i=0$2 to $\Delta_i=0$3. With a $\Delta_i=0$4-cardinality budget, the reformulation uses a Bertsimas–Sim-type linearization with nonnegative auxiliary variables $\Delta_i=0$5, and the resulting model is a pure MILP in $\Delta_i=0$6 variables $\Delta_i=0$7 plus auxiliaries, polynomial in $\Delta_i=0$8 (Ramirez-Calderon et al., 2021).

The guarantees are correspondingly explicit. With $\Delta_i=0$9, any optimal robust solution satisfies $\tilde x\in\{0,1\}^n$0 for all $\tilde x\in\{0,1\}^n$1. With $\tilde x\in\{0,1\}^n$2, worst-case violation of constraint $\tilde x\in\{0,1\}^n$3 is bounded by $\tilde x\in\{0,1\}^n$4. Under i.i.d. flips with known $\tilde x\in\{0,1\}^n$5, if $\tilde x\in\{0,1\}^n$6 denote the numbers of $\tilde x\in\{0,1\}^n$7 and $\tilde x\in\{0,1\}^n$8 flips, then

$\tilde x\in\{0,1\}^n$9

so the probability that more than $C$0 flips occur is at most

$C$1

In the $C$2 knapsack study, the metrics were average objective-loss $C$3, feasibility rate $C$4, and observed max violation $C$5. The reported findings include feasibility improving from approximately $C$6 to approximately $C$7 for $C$8 under $C$9, at a cost of $U$0–$U$1 profit loss, and the recovery of much of that profit when small $U$2 or smaller $U$3 is allowed (Ramirez-Calderon et al., 2021).

3. Interval uncertainty propagation in predictive control

In robust MPC for linear discrete-time systems, the uncertain model is written as

$U$4

with

$U$5

or equivalently

$U$6

The propagation of this interval matrix uncertainty is bounded by over-approximating each power $U$7, or $U$8 under stabilizing feedback, by a matrix zonotope

$U$9

If $$U(x)=\{\tilde x\in\{0,1\}^n:\tilde x_i=x_i \text{ for } i\in C;\ \tilde x_i\in\{0,1\} \text{ for } i\in U\}.$$0 bounds $$U(x)=\{\tilde x\in\{0,1\}^n:\tilde x_i=x_i \text{ for } i\in C;\ \tilde x_i\in\{0,1\} \text{ for } i\in U\}.$$1, then $$U(x)=\{\tilde x\in\{0,1\}^n:\tilde x_i=x_i \text{ for } i\in C;\ \tilde x_i\in\{0,1\} \text{ for } i\in U\}.$$2 is over-approximated by $$U(x)=\{\tilde x\in\{0,1\}^n:\tilde x_i=x_i \text{ for } i\in C;\ \tilde x_i\in\{0,1\} \text{ for } i\in U\}.$$3, whose center is $$U(x)=\{\tilde x\in\{0,1\}^n:\tilde x_i=x_i \text{ for } i\in C;\ \tilde x_i\in\{0,1\} \text{ for } i\in U\}.$$4, whose old generators are $$U(x)=\{\tilde x\in\{0,1\}^n:\tilde x_i=x_i \text{ for } i\in C;\ \tilde x_i\in\{0,1\} \text{ for } i\in U\}.$$5, and whose new generators are obtained from the entry-wise decomposition of

$$U(x)=\{\tilde x\in\{0,1\}^n:\tilde x_i=x_i \text{ for } i\in C;\ \tilde x_i\in\{0,1\} \text{ for } i\in U\}.$$6

The core enclosure is

$$U(x)=\{\tilde x\in\{0,1\}^n:\tilde x_i=x_i \text{ for } i\in C;\ \tilde x_i\in\{0,1\} \text{ for } i\in U\}.$$7

and the precomputed interval bound is

$$U(x)=\{\tilde x\in\{0,1\}^n:\tilde x_i=x_i \text{ for } i\in C;\ \tilde x_i\in\{0,1\} \text{ for } i\in U\}.$$8

These bounds enter a variable-horizon MPC with nominal model

$$U(x)=\{\tilde x\in\{0,1\}^n:\tilde x_i=x_i \text{ for } i\in C;\ \tilde x_i\in\{0,1\} \text{ for } i\in U\}.$$9

and interval tube

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i],$0

The optimization minimizes

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i],$1

subject to

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i],$2

Theorems 1–2 provide recursive feasibility and robust asymptotic stability, and the stability certificate includes the shrinking-horizon decrease

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i],$3

A central implementation point is that the interval bounds are computed offline, making the online computational load independent of the number of uncertain parameters (Quartullo et al., 19 Feb 2026).

A data-driven variant constructs the interval matrix directly from input-state data. With

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i],$4

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i],$5

and bounded disturbance $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i],$6, the consistent system set satisfies

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i],$7

After partitioning $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i],$8 and $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i],$9, the method precomputes

$\Delta_i=1$0

forms

$\Delta_i=1$1

and uses online tubes

$\Delta_i=1$2

The resulting scheme is recursively feasible and practically stable; after entering the terminal regime, the state converges to

$\Delta_i=1$3

a bounded invariant limit set (Quartullo et al., 16 Mar 2026). This suggests that implementation uncertainty intervals in control are not restricted to a priori modeling; they can also be extracted from finite data and then propagated by offline-computable interval operators.

4. Probabilistic and inferential constructions of uncertainty intervals

In imprecise-probability conditioning, one starts from a convex set $\Delta_i=1$4 of probabilities on a finite universe $\Delta_i=1$5, conditions on an event $\Delta_i=1$6, renormalizes the surviving distributions, and interprets the normalizing factor $\Delta_i=1$7 as a possibility weight on each conditional. For a query event $\Delta_i=1$8, the resulting collection $\Delta_i=1$9 is converted into an uncertainty interval $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$00 either by a Sugeno integral or by a Choquet integral. With possibility measure $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$01 and dual necessity measure $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$02,

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$03

and

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$04

with $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$05 and $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$06. Sorting the normalized points costs $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$07; after sorting, Sugeno’s max–min scan and Choquet’s piecewise trapezoid sum take $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$08. The paper emphasizes that using only the extreme points can make the intervals artificially narrow and discontinuous under tiny perturbations of $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$09, whereas using the full set $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$10 restores continuity (Moral, 2013).

A related interval-probability formalism defines

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$11

where $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$12 is a degree of indecision and $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$13 is the set of indecisive eventualities associated with $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$14. In the special case $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$15,

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$16

so the interval width is exactly $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$17. Updating on an event $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$18 is given by

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$19

with

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$20

which yields a non-Bayesian interval update rule (Basili et al., 2024).

Maximum-uncertainty reconstruction provides a third route. Given marginal interval distributions, one infers the least informative consistent joint interval distribution by maximizing total width

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$21

subject to normalization and marginal-consistency constraints. The resulting optimum is the unique maximum-$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$22 element of the extension, and can also be recovered by solving two LPs per atomic event to obtain $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$23 and $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$24 (Pittarelli, 2013). These constructions show that uncertainty intervals may arise from conditioning, updating, or reconstruction, not only from perturbing an implemented decision vector.

5. Neural-network implementations of interval uncertainty

Deep interval learning implements interval uncertainty by replacing every real-valued object with an interval-valued analogue. In a DINN, an input sample $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$25 is embedded as

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$26

weights and biases become $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$27 and $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$28, and each layer executes

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$29

For $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$30 and $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$31,

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$32

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$33

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$34

The paper uses an interval-MSE loss and an interval Adam procedure, I-Adam. Practical notes include outward rounding, smaller initial radii than midpoints, gradient clipping on interval endpoints, and the statement that computational cost is roughly $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$35 a real-valued net; monotonicity of $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$36 is described as crucial because non-monotonic activations produce badly over-estimated intervals (Betancourt et al., 2021).

For SysID, INNs convert pretrained parameters $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$37 into intervals

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$38

with $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$39 enforced by reparameterization. The framework extends LSTM and Neural ODE architectures into ILSTM and INODE, and trains them with the RQR-W loss

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$40

where

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$41

A later formulation separates two strategies: Cascade INN (C-INN), which first trains a crisp NN and then optimizes interval margins only, and Joint INN (J-INN), which jointly optimizes point accuracy and interval quality using

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$42

with GradNorm balancing. Calibration is assessed by PICP, PINAW, and CWC. Across the reported four SysID datasets, C-INN slightly outperforms J-INN in RMSE, while J-INN consistently yields higher PICP and lower CWC. The same study introduces channel-wise elasticity,

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$43

to analyze how uncertainty is distributed over channels (Ferah et al., 26 Apr 2025, Ferah et al., 12 May 2026).

In classification, CreINNs output classwise probability intervals $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$44 and define the credal set

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$45

Total, aleatoric, and epistemic uncertainty are then computed as entropy extrema over $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$46, with epistemic uncertainty given by $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$47 (Wang et al., 2024). For surrogate-based interval propagation, three direct approaches are reported: naive interval DIP, bound-propagation methods such as IBP and CROWN, and INNs with interval weights. The reported conclusion is that all direct methods are approximately $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$48–$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$49 faster than optimization-based propagation, while CROWN often yields the lowest $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$50 under ideal interval training and INNs may over-estimate due to dependency (Faza et al., 22 Mar 2026).

6. Guarantees, calibration, and recurrent misunderstandings

A recurrent issue is whether an interval is a robustness set, a prediction interval, or an imprecise-probability envelope. In LLM-as-a-judge, conformal prediction constructs a continuous prediction interval from a single evaluation run. With calibration nonconformity scores

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$51

the split-conformal quantile $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$52 yields

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$53

and therefore the interval

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$54

For discrete $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$55-point ratings, the ordinal boundary adjustment

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$56

maps the continuous band to valid labels, and the midpoint

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$57

is proposed as a low-bias alternative to the raw score and weighted average (Sheng et al., 23 Sep 2025). This is a calibrated prediction interval, not a worst-case implementation set.

In regression, DPIN separates mean estimation and PI estimation into two stages. The PI head is trained with

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$58

where $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$59 enforces

$\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$60

to meet the target coverage. The reported results state $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$61 to $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$62 error reduction while maintaining $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$63 PICP for $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$64 out of $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$65 UCI benchmark datasets, together with $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$66 to $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$67 error reduction in active learning (Patel et al., 2022). Here, interval width is explicitly optimized against coverage.

Several pitfalls recur across the literature. Artificially narrow intervals can be produced if one conditions a convex set of probabilities and keeps only extreme normalized conditionals rather than the full $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$68 (Moral, 2013). Over-estimation can arise in INNs because of dependency and wrapping effects in interval arithmetic (Faza et al., 22 Mar 2026). In robust optimization, $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$69 and $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$70 are not error bars but tuning parameters that dial down conservatism relative to the full worst-case model (Ramirez-Calderon et al., 2021). This suggests that interval interpretation is domain-specific: identical syntax $\tilde x_i\in[x_i-\Delta_i,\;x_i+\Delta_i]$71 may encode feasibility guarantees, calibration guarantees, or admissible imprecision, but not the same semantics.