pith. sign in

arxiv: 2604.18511 · v1 · submitted 2026-04-20 · 🧮 math.OC

An adaptive discretization algorithm for locally optimal experimental design with constraints

Jochen Schmid , Philipp Seufert , Jan Schwientek , Tobias Seidel , Karl-Heinz K\"ufer This is my paper

Pith reviewed 2026-05-10 04:00 UTC · model grok-4.3

classification 🧮 math.OC
keywords optimal experimental designadaptive discretizationconstrained optimizationepsilon-optimalityconvergence analysisiterative algorithmchemical engineeringnonlinear programming
0
0 comments X

The pith

An adaptive discretization algorithm converges to optimal or ε-optimal experimental designs under constraints with reduced computation.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops an iterative algorithm that solves discretized versions of constrained locally optimal experimental design problems and refines the discretization grid by adding approximate violators of a sufficient ε-optimality condition. With ε set to zero the method converges to a true optimum; with positive ε it terminates after finitely many steps at an ε-optimal design. The approach requires solving only smaller-dimensional nonlinear subproblems and only approximately in selected iterations, while supporting a broader range of constraints than earlier algorithms. Numerical tests on chemical engineering problems that include time and yield constraints illustrate the convergence behavior.

Core claim

The algorithm alternates between solving a discretized constrained design problem and adaptively enriching the discretization by inserting points that approximately violate the sufficient ε-optimality condition of the current design; this process is shown to produce convergence to an optimal design when ε equals zero and finite termination at an ε-optimal design when ε is positive.

What carries the argument

The adaptive refinement step that locates and inserts approximate violators of the sufficient ε-optimality condition into the current discretization.

If this is right

  • Nonlinear subproblems remain lower-dimensional and are solved only approximately in the final iterations.
  • The method applies to a larger class of constraints than previous algorithms for constrained experimental design.
  • Finite termination occurs at an ε-optimal design for any positive ε.
  • Convergence to a true optimum is guaranteed when ε is set to zero.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same adaptive-violation refinement idea could be tested on high-dimensional design spaces to check whether the number of added points remains modest.
  • Similar discretization strategies might be examined for related problems such as robust or sequential experimental design.
  • The reduced subproblem size suggests possible use inside real-time or embedded optimization loops where full continuous solves are prohibitive.

Load-bearing premise

An approximate violator of the sufficient ε-optimality condition for the current design can be found or computed in each refinement iteration.

What would settle it

A concrete constrained design problem on which the algorithm either diverges, requires solving the full continuous problem in every iteration, or produces designs whose optimality gap exceeds the prescribed ε after termination.

read the original abstract

We develop a novel iterative algorithm for locally optimal experimental design under constraints, like budget or performance constraints. It is an adaptive discretization algorithm. In every iteration, a discretized version of the constrained-design problem is solved and then the discretization is adaptively refined by adding an approximate violator of a suitable sufficient $\eps$-optimality condition for the current design. We prove that with $\eps = 0$, our algorithm converges to an optimal design and that with $\eps > 0$, our algorithm finitely terminates at an $\eps$-optimal design. Compared to the existing algorithms on constrained experimental design, our algorithm comes with considerably less computational effort because the nonlinear subproblems in our algorithm have a smaller dimension and have to be solved only approximately and only in selected iterations (typically the last few). Additionally, our algorithm covers a considerably larger class of constraints. We demonstrate the good convergence properties of the algorithm on experimental design problems from chemical engineering that feature time and yield constraints.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper develops an adaptive discretization algorithm for locally optimal experimental design problems subject to constraints (e.g., budget, time, or yield). In each iteration a discretized constrained design problem is solved, after which the discretization is refined by adding an approximate violator of a sufficient ε-optimality condition for the current design measure. The authors prove that ε = 0 yields convergence to an optimal design while ε > 0 yields finite termination at an ε-optimal design. They claim the method requires considerably less computational effort than existing constrained-design algorithms because the nonlinear subproblems have smaller dimension and are solved only approximately and only in selected iterations, and that it applies to a broader class of constraints. Numerical results on chemical-engineering design problems with time and yield constraints are presented.

Significance. If the convergence proofs are complete and the violator subproblem is demonstrably cheaper than re-solving the full discretized program, the algorithm would offer a practical advance for constrained experimental design, especially in applications where the design space is continuous and constraints are non-trivial. The adaptive-refinement strategy and the explicit handling of a wider constraint class could reduce the computational barrier that currently limits the use of optimal design in engineering practice.

major comments (3)
  1. [§4 (algorithm and refinement step)] The central efficiency claim (abstract and §4) rests on the assertion that locating an approximate violator of the sufficient ε-optimality condition is cheaper than solving the full discretized nonlinear program. The manuscript must supply either a complexity bound for the violator subproblem or concrete timing data showing that this step does not dominate the overall cost, particularly when the constraint set is defined by time or yield functions.
  2. [§3 (convergence analysis)] The convergence proof for ε = 0 (presumably §3) requires that successive addition of approximate violators produces a sequence whose limit satisfies the first-order optimality condition. The precise statement of the sufficient ε-optimality condition and the argument that an approximate violator guarantees monotonic progress in the discretization measure should be written out explicitly, including the role of the constraint qualification.
  3. [§2 (problem formulation) and §5 (numerical examples)] The claim that the algorithm covers a considerably larger class of constraints than prior methods is load-bearing for the contribution. The exact functional form of the admissible constraints (e.g., the regularity required on the yield or time maps) must be stated, together with a concrete example where an existing method fails but the new algorithm succeeds.
minor comments (2)
  1. [§2] Notation for the design measure and the discretization grid should be introduced once and used consistently; several places switch between μ and ξ without explicit redefinition.
  2. [§5] The numerical section would benefit from a table reporting the number of violator searches performed and the CPU time spent on them versus the time spent on the discretized NLPs, to make the efficiency claim quantitative.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments, which have helped us identify areas for clarification and strengthening of the manuscript. We address each major comment point by point below, indicating the revisions we plan to make.

read point-by-point responses

  1. Referee: [§4 (algorithm and refinement step)] The central efficiency claim (abstract and §4) rests on the assertion that locating an approximate violator of the sufficient ε-optimality condition is cheaper than solving the full discretized nonlinear program. The manuscript must supply either a complexity bound for the violator subproblem or concrete timing data showing that this step does not dominate the overall cost, particularly when the constraint set is defined by time or yield functions.

    Authors: We agree that additional evidence would strengthen the efficiency claim. In the revised manuscript we will add concrete timing data from the numerical experiments in §5, reporting the CPU time spent on violator identification versus the time for solving the discretized nonlinear programs across the chemical-engineering test cases. We will also include a short paragraph in §4 explaining that the violator subproblem is a lower-dimensional, unconstrained (or lightly constrained) optimization over the continuous design space for a fixed measure, which admits cheap approximate solution by gradient ascent and is invoked only in selected iterations; this structure avoids the full constraint Jacobian and barrier terms required by the main subproblem. revision: yes

  2. Referee: [§3 (convergence analysis)] The convergence proof for ε = 0 (presumably §3) requires that successive addition of approximate violators produces a sequence whose limit satisfies the first-order optimality condition. The precise statement of the sufficient ε-optimality condition and the argument that an approximate violator guarantees monotonic progress in the discretization measure should be written out explicitly, including the role of the constraint qualification.

    Authors: We appreciate the request for greater explicitness. In the revision we will restate the sufficient ε-optimality condition at the beginning of §3 as: a measure ξ* is ε-optimal if max_x [φ(x,ξ*) – λ·∇g(ξ*)] ≤ ε, where φ is the sensitivity function and λ are the multipliers for the constraints. We will then expand the convergence argument to show explicitly that an approximate violator (a point where the expression exceeds ε by a positive margin) produces a strict improvement in the discretized objective or a refinement of the support that cannot continue indefinitely; the limit therefore satisfies the first-order stationarity condition. The constraint qualification (Slater’s condition on the admissible constraint set) is used to guarantee existence of the multipliers and to ensure that the monotonic progress in the discretization measure is well-defined; this step will be written out in full. revision: yes

  3. Referee: [§2 (problem formulation) and §5 (numerical examples)] The claim that the algorithm covers a considerably larger class of constraints than prior methods is load-bearing for the contribution. The exact functional form of the admissible constraints (e.g., the regularity required on the yield or time maps) must be stated, together with a concrete example where an existing method fails but the new algorithm succeeds.

    Authors: We agree that the precise class of constraints should be stated explicitly. In the revised §2 we will specify that admissible constraints are continuous functionals g(ξ) ≤ 0 of the design measure ξ (in the weak* topology), where each g is required to be continuous and the feasible set convex and closed; this includes integral constraints of the form ∫ f_i(x) dξ(x) ≤ c_i with continuous f_i, as well as more general nonlinear maps such as yield or time functions that are continuous in the design variables. As a concrete example we will add in §2 and reference in §5 a nonlinear yield constraint arising in a chemical reactor design problem; existing linear-programming or semidefinite-relaxation methods from the literature cannot accommodate the nonlinearity without conservative outer approximations that become infeasible, whereas our algorithm solves the resulting nonlinear subproblems directly and produces a feasible ε-optimal design. revision: yes

Circularity Check

0 steps flagged

No circularity: convergence proofs and efficiency claims are independent of self-referential definitions or fitted inputs

full rationale

The paper presents an adaptive discretization algorithm for constrained optimal experimental design, with explicit proofs of convergence (ε=0) and finite termination (ε>0) to ε-optimal designs. These results rely on the stated assumption that approximate violators of the ε-optimality condition can be identified in each iteration, but this is an external computational oracle rather than a self-definition or fitted parameter renamed as a prediction. No equations reduce the claimed optimality or efficiency gains to the algorithm's own outputs by construction. The comparison to prior methods (smaller-dimensional subproblems solved approximately in final iterations) is presented as an empirical and structural property of the discretization scheme, not derived from self-citation chains or ansatzes smuggled via prior work. The derivation chain is self-contained against external benchmarks such as standard nonlinear programming solvers and existing constrained-design algorithms.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Based solely on the abstract, the central claim rests on standard assumptions from optimal experimental design theory plus the existence of an approximate violator at each step; no free parameters or new entities are introduced in the provided text.

axioms (1)
  • domain assumption A suitable sufficient ε-optimality condition exists for any feasible design and can be approximately violated by a computable point.
    Invoked to justify the adaptive refinement step that adds new discretization points.

pith-pipeline@v0.9.0 · 5474 in / 1387 out tokens · 34244 ms · 2026-05-10T04:00:15.117292+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

51 extracted references · 51 canonical work pages

  1. [1]

    Amann: Ordinary differential equations

    H. Amann: Ordinary differential equations. An introduction to nonlinear analysis. DeGruyter (1990)

    work page 1990

  2. [2]

    Amann, J

    H. Amann, J. Escher: Analysis I, II, III. Birkhäuser (2005, 2008, 2009)

    work page 2005

  3. [3]

    Atkinson, V.V

    A.C. Atkinson, V.V. Fedorov:The optimum design of experiments in the presence of uncontrolled variability and prior information.In: Y. Dodge, V.V. Fedorov, H. Wynn (eds.): Optimal design and analysis of experiments. North-Holland (1988), 327-344

    work page 1988

  4. [4]

    A.C.Atkinson, A.N.Donev, R.D.Tobias: Optimumexperimentaldesign, withSAS.OxfordUniversity Press (2007)

    work page 2007

  5. [5]

    Atwood:Sequences converging to D-optimal designs of experiments.Ann

    C. Atwood:Sequences converging to D-optimal designs of experiments.Ann. Statist.1(1973), 342-352

    work page 1973

  6. [6]

    Bhatia: Matrix analysis

    R. Bhatia: Matrix analysis. Springer (1997)

    work page 1997

  7. [7]

    Blankenship, J.E

    J.W. Blankenship, J.E. Falk:Infinitely constrained optimization problems.J. Optim. Th. Appl.19 (1976), 261-281

    work page 1976

  8. [8]

    Böhning:A vertex-exchange method in D-optimal design theory.Metrika33(1986), 337-347

    D. Böhning:A vertex-exchange method in D-optimal design theory.Metrika33(1986), 337-347

    work page 1986

  9. [9]

    Bubel, J

    M. Bubel, J. Schmid, V. Kozachynskyi, E. Esche, M. Bortz:Sequential optimal experimental design for vapor-liquid equilibrium modeling.Chem. Eng. Sci. 300 (2024), 120566

    work page 2024

  10. [10]

    Canon, C.D

    M.D. Canon, C.D. Cullum:A tight upper bound on the rate of convergence of the Frank-Wolfe algorithm.SIAM J. Contr.6(1968), 509-516

    work page 1968

  11. [11]

    Cook, L.A

    D. Cook, L.A. Thibodeau:Marginally restricted D-optimal designs.Amer. Statist. Assoc.75(1980), 366-371

    work page 1980

  12. [12]

    Cook, V.V

    D. Cook, V.V. Fedorov:Constrained optimization of experimental design.Statistics26(1995), 129- 178

    work page 1995

  13. [13]

    Daniele, S

    P. Daniele, S. Giuffré, A. Maugeri:Remarks on general infinite dimensional duality with cone and equality constraints.Comm. Appl. Math.13(2009), 567-578

    work page 2009

  14. [14]

    Donato:The infinite dimensional Lagrange multiplier rule for convex optimization problems

    M.B. Donato:The infinite dimensional Lagrange multiplier rule for convex optimization problems. J. Funct. Anal.261(2011), 2083-2093 42

    work page 2011

  15. [15]

    Duarte, A.C

    B.P.M. Duarte, A.C. Atkinson:Algorithms for optimal design of experiments.In: M. Lovric (ed.), International Encyclopedia of Statistical Science. Springer (2025)

    work page 2025

  16. [16]

    Dupuis, R.S

    P. Dupuis, R.S. Ellis: A weak convergence approach to the theory of large deviations. Wiley (1997)

    work page 1997

  17. [17]

    Fedorov: Theory of optimal experiments

    V.V. Fedorov: Theory of optimal experiments. Academic Press (1972)

    work page 1972

  18. [18]

    Fedorov:Optimal design with bounded density: optimization algorithms of the exchange type

    V.V. Fedorov:Optimal design with bounded density: optimization algorithms of the exchange type. J. Statist. Plann. Inference22(1989), 1-13

    work page 1989

  19. [19]

    Fedorov, S.L

    V.V. Fedorov, S.L. Leonov: Optimal design for nonlinear response models. Chapman & Hall (2014)

    work page 2014

  20. [20]

    Gaivoronski:Linearization methods for optimization of functionals which depend on probability measures.In: A

    A. Gaivoronski:Linearization methods for optimization of functionals which depend on probability measures.In: A. Prékopa, R.J.B. Wets (eds.): Stochastic Programming 84 Part II. Mathematical Programming Studies28(1986), 157-181

    work page 1986

  21. [21]

    Harman, L

    R. Harman, L. Filová, P. Richtárik:A ranomized exchange algorithm for computing optimal approx- imate designs of experiments.J. Am. Statist. Assoc.115(2020), 348-361

    work page 2020

  22. [22]

    Herzog, E

    R. Herzog, E. Legler:First-order methods for optimal experimental design problems with bound constraints.arXiv:2004.08084 (2020)

    work page arXiv 2004

  23. [23]

    Huang, M.C

    M.L. Huang, M.C. Hsu:Marginally restricted linear-optimal designs.J. Statist. Plann. Inference35 (1993), 251-266

    work page 1993

  24. [24]

    Jaggi:Revisiting Frank–Wolfe: projection-free sparse convex optimization.Proceedings of the 30th International Conference on Machine Learning (ICML), PMLR (2013), 427-435

    M. Jaggi:Revisiting Frank–Wolfe: projection-free sparse convex optimization.Proceedings of the 30th International Conference on Machine Learning (ICML), PMLR (2013), 427-435

    work page 2013

  25. [25]

    Jahn: Introduction to the theory of nonlinear optimization

    J. Jahn: Introduction to the theory of nonlinear optimization. 3rd edition, Springer (2007)

    work page 2007

  26. [26]

    Kiefer:General equivalence theory for optimum designs (approximate theory).Ann

    J. Kiefer:General equivalence theory for optimum designs (approximate theory).Ann. Stat.2(1974), 849-879

    work page 1974

  27. [27]

    Lacoste-Julien, M

    S. Lacoste-Julien, M. Jaggi:On the global linear convergence of Frank-Wolfe optimization variants. Advances in Neural Information Processing Systems28(2015), 496-504

    work page 2015

  28. [28]

    Mangasarian: Nonlinear programming

    O.L. Mangasarian: Nonlinear programming. McGraw-Hill (1994)

    work page 1994

  29. [29]

    Magnus, H

    J.R. Magnus, H. Neudecker: Matrix differential calculus with applications in statistics and econo- metrics. 3rd edition, Wiley (2007)

    work page 2007

  30. [30]

    Molchanov, S

    I. Molchanov, S. Zuyev:Steepest descent algorithms in a space of measuers.Stat. Comp.12(2002), 115-123

    work page 2002

  31. [31]

    Molchanov, S

    I. Molchanov, S. Zuyev:Optimisation in space of measures and optimal design.ESAIM Prob. Stat.8 (2004), 12-24

    work page 2004

  32. [32]

    Parthasarathy: Probability measures on metric spaces

    K.R. Parthasarathy: Probability measures on metric spaces. Academic Press (1967)

    work page 1967

  33. [33]

    Pronzato:A minimax equivalence theorem for optimum bounded design measures.Stat

    L. Pronzato:A minimax equivalence theorem for optimum bounded design measures.Stat. Prob. Lett.68(2004), 325-331

    work page 2004

  34. [34]

    Pronzato, A

    L. Pronzato, A. Pázman: Design of experiments in nonlinear models. Asymptotic normality, opti- mality criteria and small-sample properties. Springer (2013)

    work page 2013

  35. [35]

    Pukelsheim: Optimal design of experiments

    F. Pukelsheim: Optimal design of experiments. SIAM Classics of Applied Mathematics (2006)

    work page 2006

  36. [36]

    Robinson:Stability theory for systems of inequalities in nonlinear programming, part II: dif- ferentiable nonlinear systems.SIAM J

    S.M. Robinson:Stability theory for systems of inequalities in nonlinear programming, part II: dif- ferentiable nonlinear systems.SIAM J. Numer. Anal.13(1976), 497-513

    work page 1976

  37. [37]

    Rockafellar: Convex analysis

    R.T. Rockafellar: Convex analysis. Princeton University Press (1970)

    work page 1970

  38. [38]

    Schittkowski:A robust implementation of a sequential quadratic programming algorithm with successive error restoration.Opt

    K. Schittkowski:A robust implementation of a sequential quadratic programming algorithm with successive error restoration.Opt. Lett.5(2011), 283–296

    work page 2011

  39. [39]

    Schittkowski:NLPQLP: A Fortran implementation of a sequential quadratic programming algo- rithm with distributed non-monotone line search.User’s guide, Version 4.2 (2014)

    K. Schittkowski:NLPQLP: A Fortran implementation of a sequential quadratic programming algo- rithm with distributed non-monotone line search.User’s guide, Version 4.2 (2014)

    work page 2014

  40. [40]

    Schmid, P

    J. Schmid, P. Seufert, M. Bortz:Adaptive discretization algorithms for locally optimal experimental design.arXiv:2406.01541 (2024)

    work page arXiv 2024

  41. [41]

    Silvey, D.H

    S.D. Silvey, D.H. Titterington, B. Torsney:An algorithm for optimal designs on a design space. Comm. Statist. Theory Methods14(1978), 1379-1389 43

    work page 1978

  42. [42]

    Silvey: Optimal design

    S.D. Silvey: Optimal design. An introduction to the theory for parameter estimation. Chapman & Hall (1980)

    work page 1980

  43. [43]

    Uciński, B

    D. Uciński, B. Bogacka:T-optimum designs for multiresponse dynamic heteroscedastic models.In: A. Di Bucchianico, H. Läuter, H.P. Wynn (eds.): Contributions to statistics. Springer (2004), 191-199

    work page 2004

  44. [44]

    Uciński, B

    D. Uciński, B. Bogacka:T-optimum designs for discrimination between two multiresponse dynamic models.J. R. Statist. Soc. B67(2005), 3-18

    work page 2005

  45. [45]

    Vanaret, P

    C. Vanaret, P. Seufert, J. Schwientek, G. Karpov, G. Ryzhakov, I. Oseledets, N. Asprion, M. Bortz: Two-phase approaches to optimal model-based design of experiments: how many experiments and which ones?Comp. Chem. Eng.146(2021), 107218

    work page 2021

  46. [46]

    Virtanen, R

    P. Virtanen, R. Gommers, T.E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser, J. Bright, S.J. van der Walt, M. Brett, J. Wilson, K.J. Millman, N. Mayorov, A.R.J. Nelson, E. Jones, R. Kern, E. Larson, C.J. Carey, İ. Polat, Y. Feng, E.W. Moore, J. VanderPlas, D. Laxalde, J. Perktold, R. Cimrman, I. Henriksen, A.E. Qu...

    work page 2020

  47. [47]

    Wynn:The sequential generation of D-Optimal experimental designs.Ann

    H.P. Wynn:The sequential generation of D-Optimal experimental designs.Ann. Math. Statist.41 (1970), 1655-1664

    work page 1970

  48. [48]

    M. Yang, S. Biedermann, E. Tang:On optimal designs for nonlinear models: a general and efficient algorithm.J. Am. Statist. Assoc.108(2013), 1411-1420

    work page 2013

  49. [49]

    Yu:Monotonic convergence of a general algorithm for computing optimal designs.The Annals of Statistics38(2010), 1593

    Y. Yu:Monotonic convergence of a general algorithm for computing optimal designs.The Annals of Statistics38(2010), 1593

    work page 2010

  50. [50]

    Yu:D-optimal designs via a cocktail algorithm.Stat

    Y. Yu:D-optimal designs via a cocktail algorithm.Stat. Comp.21(2011), 475-481

    work page 2011

  51. [51]

    J. Zowe, S. Kurcyusz:Regularity and stability for the mathematical programming problem in Banach spaces.Appl. Math. Optim.5(1979), 49-62 44

    work page 1979