Robust Downhill Simplex Method (rDSM)

• rDSM is a robustness-enhanced variant of the Downhill Simplex Method that augments classical geometric moves with explicit degeneracy detection and correction.
• It preserves DSM’s reflection, expansion, contraction, and shrink steps while repairing degenerated simplices to maintain full-dimensional search space.
• The method employs vertex reevaluation to average out noise, ensuring improved convergence reliability in high-dimensional, noisy optimization problems.

Robust Downhill Simplex Method (rDSM) denotes a robustness-oriented variant of the Downhill Simplex Method (DSM), or Nelder–Mead family, for derivative-free minimization, introduced as a software package for optimization problems in high dimensions and in the presence of noise (Wang et al., 7 Sep 2025). It preserves the standard DSM geometric moves—reflection, expansion, contraction, and shrink—but augments them with two explicit mechanisms: simplex degeneracy detection and correction, and reevaluation of long-standing vertices to estimate the real objective value of noisy problems. In the broader simplex literature, closely related robustness ideas appear as random restart when simplex points “get near each other” (Saboori et al., 2012), regular-simplex geometry combined with sufficient decrease and shrinking rules that admit worst-case complexity bounds (Cao et al., 22 Aug 2025), and hybrids that alternate a downhill simplex step with explorative Latin hypercube sampling (Li et al., 2019). Taken together, these works place rDSM within a larger program of making simplex-based search less sensitive to premature convergence, degenerated simplices, and noise-induced spurious minima.

1. Classical descent framework and rDSM problem setting

rDSM solves the unconstrained derivative-free minimization problem

$$\text{Find } x^\star \in \mathbb{R}^n \text{ that minimizes } J(x),$$

where no gradient $\nabla J$ is assumed available and evaluation of $J$ may be noisy (Wang et al., 7 Sep 2025). As in classical DSM, the current state is an $n$-simplex

$S = \{x^{s_1},\dots,x^{s_{n+1}}\},$

with vertices ordered so that

$$J(x^{s_1}) \le \dots \le J(x^{s_n}) \le J(x^{s_{n+1}}),$$

where $x^{s_1}$ is best and $x^{s_{n+1}}$ is worst. The centroid of all but the worst point is

$x^{c} = \frac{1}{n} \sum_{i=1}^{n} x^{s_i}.$

The underlying DSM moves retained by rDSM are the standard Nelder–Mead operations. Reflection uses

$x^{r} = x^{c} + \alpha (x^{c} - x^{s_{n+1}}),$

expansion uses

$\nabla J$0

outside and inside contraction use

$\nabla J$1

and shrink updates

$\nabla J$2

The rDSM software uses the standard Nelder–Mead parameters

$\nabla J$3

The key structural point is that rDSM does not alter the basic DSM decision logic for reflection, expansion, contraction, and shrink. Instead, it wraps that loop with corrective procedures. The main iteration is: run a Nelder–Mead iteration, check degeneracy, call degeneracy correction if required, apply reevaluation logic for long-standing vertices, and proceed to the next iteration (Wang et al., 7 Sep 2025). A common misconception is therefore that rDSM is a wholly different simplex optimizer. In the 2025 formulation, it is more precisely a standard DSM augmented by robustness mechanisms around the basic geometric moves.

2. Degeneracy detection and geometric reinflation

The first robustness mechanism addresses simplex degeneration, described as a major source of loss of effective dimensionality, poor exploration of the search space, and stagnation or premature convergence at non-stationary points (Wang et al., 7 Sep 2025). The method distinguishes two types of degeneracy: an edge-degenerated simplex, in which one edge is much shorter than others, and a volume-degenerated simplex, in which the volume is too small relative to the edge lengths. Because each type may occur without the other, rDSM uses two independent criteria.

The simplex perimeter is defined as

$\nabla J$4

and the $\nabla J$5-dimensional volume is

$\nabla J$6

For edge-based degeneracy, with edge vectors $\nabla J$7 and norms $\nabla J$8,

$\nabla J$9

and the simplex is declared edge-degenerated if

$J$0

For volume-based degeneracy,

$J$1

and the simplex is volume-degenerated if

$J$2

The default thresholds are $J$3 and $J$4.

Once degeneracy is detected, rDSM repairs the simplex by moving the worst vertex $J$5 to a new position $J$6 that maximizes simplex volume while preserving simplex perimeter: $J$7 This constrained problem is solved by a Newton–Raphson method. The perimeter constraint keeps the overall scale of the simplex similar, while the volume objective inflates the simplex in collapsed directions. If moving the worst vertex does not remove degeneracy, the algorithm proceeds to move the next worst vertex, and so on, until the simplex is nondegenerate or all points have been corrected (Wang et al., 7 Sep 2025).

In two dimensions, the geometric interpretation is explicit: the locus of points with constant sum of distances to two fixed vertices is an ellipse, so the corrected vertex lies on an ellipse with foci at the fixed vertices, and the algorithm selects the location giving maximal area (Wang et al., 7 Sep 2025). This suggests a useful conceptual reading of rDSM: rather than restarting the search globally, it performs an internal reinitialization that attempts to preserve accumulated information while restoring full-dimensional search geometry.

3. Noise-robust objective estimation by reevaluation

The second robustness mechanism addresses noisy objectives. The paper evaluates rDSM under additive noise,

$J$8

with uniform noise $J$9, where $n$0 or $n$1, and Gaussian noise $n$2, where $n$3 or $n$4 (Wang et al., 7 Sep 2025). The stated concern is that classical DSM reacts to whichever noisy value it sees and can be misled into local, noise-induced minima.

To mitigate this, rDSM assigns to each vertex $n$5 a counter $n$6 that records how many DSM iterations the vertex has remained in the current simplex. The key observation is that most vertices leave the simplex after roughly $n$7 iterations; a vertex that remains significantly longer is likely close to a search center where the algorithm is oscillating. If

$n$8

the algorithm reevaluates $n$9 and replaces the stored objective value by the average of all evaluations at that point: $S = \{x^{s_1},\dots,x^{s_{n+1}}\},$0 where $S = \{x^{s_1},\dots,x^{s_{n+1}}\},$1 are the $S = \{x^{s_1},\dots,x^{s_{n+1}}\},$2 evaluations performed at $S = \{x^{s_1},\dots,x^{s_{n+1}}\},$3 so far. The averaging is intended to reduce the impact of unbiased noise by lowering the variance of the estimate.

This reevaluation step affects the ranking of simplex vertices, and therefore the reflection direction and subsequent move acceptance. In the reported tests, the threshold $S = \{x^{s_1},\dots,x^{s_{n+1}}\},$4 was preferred over a more conservative $S = \{x^{s_1},\dots,x^{s_{n+1}}\},$5: using $S = \{x^{s_1},\dots,x^{s_{n+1}}\},$6 doubled average runtime from $S = \{x^{s_1},\dots,x^{s_{n+1}}\},$7 to $S = \{x^{s_1},\dots,x^{s_{n+1}}\},$8 on a representative 2D noisy test and yielded worse accuracy, converging to a point far from the true optimum (Wang et al., 7 Sep 2025). The mechanism is deliberately selective rather than universal; the method does not reevaluate every point, but concentrates repeated measurement on long-standing vertices that disproportionately influence the local search trajectory.

4. Software implementation and empirical behavior

The rDSM software package is implemented in MATLAB, tested with MATLAB 2021b on Microsoft Windows, and released at https://github.com/tianyubobo/rDSM under a CC-BY-SA license (Wang et al., 7 Sep 2025). Its directory structure separates Main, ObjectiveFunction, Initialization, Optimizer, Visualization, and Output. The optimizer directory contains DSM.m, described as a reimplementation of fminsearch, and rDSM.m, which adds degeneracy correction and reevaluation. The package writes histories such as SimplexHistory.txt, PointsDatabase.txt, and ReevaluationHistory.txt, together with .dat ASCII versions.

The reported validation emphasizes three settings. In a 2D linear-gradient problem without an obstacle,

$S = \{x^{s_1},\dots,x^{s_{n+1}}\},$9

DSM starting from $$J(x^{s_1}) \le \dots \le J(x^{s_n}) \le J(x^{s_{n+1}}),$$0 converges rapidly to $$J(x^{s_1}) \le \dots \le J(x^{s_n}) \le J(x^{s_{n+1}}),$$1 with $$J(x^{s_1}) \le \dots \le J(x^{s_n}) \le J(x^{s_{n+1}}),$$2, essentially the global minimum. With an obstacle defined by $$J(x^{s_1}) \le \dots \le J(x^{s_n}) \le J(x^{s_{n+1}}),$$3 in the region $$J(x^{s_1}) \le \dots \le J(x^{s_n}) \le J(x^{s_{n+1}}),$$4, $$J(x^{s_1}) \le \dots \le J(x^{s_n}) \le J(x^{s_{n+1}}),$$5, the simplex becomes flattened along the obstacle boundary and DSM ends at $$J(x^{s_1}) \le \dots \le J(x^{s_n}) \le J(x^{s_{n+1}}),$$6 with $$J(x^{s_1}) \le \dots \le J(x^{s_n}) \le J(x^{s_{n+1}}),$$7. Under the same conditions, rDSM detects degeneracy during iteration $$J(x^{s_1}) \le \dots \le J(x^{s_n}) \le J(x^{s_{n+1}}),$$8, performs degeneracy correction, and converges near $$J(x^{s_1}) \le \dots \le J(x^{s_n}) \le J(x^{s_{n+1}}),$$9 with $x^{s_1}$0 (Wang et al., 7 Sep 2025).

On the noisy version of the same 2D objective with uniform noise $x^{s_1}$1, one run with a maximum of $x^{s_1}$2 iterations ends for DSM at $x^{s_1}$3 with $x^{s_1}$4, while rDSM ends at $x^{s_1}$5 with $x^{s_1}$6. Across $x^{s_1}$7 independent runs, rDSM also shows smaller variance. For uniform $x^{s_1}$8, DSM ends around $x^{s_1}$9 with $x^{s_{n+1}}$0, whereas rDSM ends at $x^{s_{n+1}}$1 with $x^{s_{n+1}}$2. For uniform $x^{s_{n+1}}$3, DSM ends about $x^{s_{n+1}}$4 with $x^{s_{n+1}}$5, while rDSM ends about $x^{s_{n+1}}$6 with $x^{s_{n+1}}$7 (Wang et al., 7 Sep 2025).

A higher-dimensional test is given by the $x^{s_{n+1}}$8-dimensional Rosenbrock function

$x^{s_{n+1}}$9

on $x^{c} = \frac{1}{n} \sum_{i=1}^{n} x^{s_i}.$0, with maximum iterations $x^{c} = \frac{1}{n} \sum_{i=1}^{n} x^{s_i}.$1 and degeneracy thresholds $x^{c} = \frac{1}{n} \sum_{i=1}^{n} x^{s_i}.$2. DSM uses $x^{c} = \frac{1}{n} \sum_{i=1}^{n} x^{s_i}.$3 evaluations and reaches $x^{c} = \frac{1}{n} \sum_{i=1}^{n} x^{s_i}.$4 with cost $x^{c} = \frac{1}{n} \sum_{i=1}^{n} x^{s_i}.$5 in $x^{c} = \frac{1}{n} \sum_{i=1}^{n} x^{s_i}.$6 s. rDSM uses $x^{c} = \frac{1}{n} \sum_{i=1}^{n} x^{s_i}.$7 evaluations, reaches $x^{c} = \frac{1}{n} \sum_{i=1}^{n} x^{s_i}.$8 to numerical precision, attains $x^{c} = \frac{1}{n} \sum_{i=1}^{n} x^{s_i}.$9, and requires $x^{r} = x^{c} + \alpha (x^{c} - x^{s_{n+1}}),$0 s (Wang et al., 7 Sep 2025). The authors explicitly note that rDSM is more robust but can be significantly more expensive than DSM in raw CPU time.

5. Adjacent robust simplex variants and historical lineage

Robustness in downhill simplex methods has been pursued through several distinct mechanisms. The following comparison situates rDSM within that landscape.

Variant	Robustness mechanism	Stated emphasis
Improved Downhill Simplex Search	Random restart when points “get near each other”	K-means initialization
RSSM	Regular simplex, sufficient decrease, reflection and shrink	Worst-case complexity
EGM	Alternating DSM and Latin hypercube sampling, plus degeneration repair	Avoiding suboptimal local minima
rDSM	Degeneracy correction and reevaluation of long-standing points	Noisy and high-dimensional problems

The 2012 “Improved Downhill Simplex Search” uses the classical Nelder–Mead structure but adds a single explicit modification: when simplex points “get near each other,” generate $x^{r} = x^{c} + \alpha (x^{c} - x^{s_{n+1}}),$1 random numbers and continue with a new simplex of $x^{r} = x^{c} + \alpha (x^{c} - x^{s_{n+1}}),$2 points (Saboori et al., 2012). Its role is to reduce deterministic trapping and improve the selection of the first $x^{r} = x^{c} + \alpha (x^{c} - x^{s_{n+1}}),$3 means in k-means clustering. The paper reports better cluster-quality indicators than plain k-means, including higher Dunn index and lower Jagota index, but it does not specify a formal criterion for “near,” does not define the restart distribution precisely, and does not provide convergence or stability proofs.

The 2025 Regular Simplicial Search Method (RSSM) pursues a different notion of robustness by maintaining a regular simplex at all times and restricting the algorithm to reflection and shrink (Cao et al., 22 Aug 2025). Reflection is accepted only when a sufficient decrease condition holds,

$x^{r} = x^{c} + \alpha (x^{c} - x^{s_{n+1}}),$4

or, in the theoretical form,

$x^{r} = x^{c} + \alpha (x^{c} - x^{s_{n+1}}),$5

Under smoothness assumptions, RSSM establishes worst-case complexity bounds in nonconvex, convex, and strongly convex cases. In this setting, robustness refers primarily to geometry control, exclusion of degeneracy, and provable convergence rates rather than noise handling.

The 2019 Explorative Gradient Method (EGM) offers a third model of robustness. It alternates one exploitive downhill simplex step with one explorative Latin hypercube sampling iteration and also includes simplex degeneration detection and repair (Li et al., 2019). On a multimodal analytical function, EGM is reported to have failure rate $x^{r} = x^{c} + \alpha (x^{c} - x^{s_{n+1}}),$6, compared with $x^{r} = x^{c} + \alpha (x^{c} - x^{s_{n+1}}),$7 for plain DSM and $x^{r} = x^{c} + \alpha (x^{c} - x^{s_{n+1}}),$8 for a random-restart simplex scheme, under $x^{r} = x^{c} + \alpha (x^{c} - x^{s_{n+1}}),$9 independent runs and a $\nabla J$00-evaluation budget. This suggests a robustness strategy based on explicitly budgeting exploration and exploitation, rather than correcting noisy objective values or reinflating degenerate simplices at fixed perimeter.

6. Theoretical status, limitations, and scope

The rDSM paper is primarily software- and methodology-oriented and does not provide rigorous convergence proofs (Wang et al., 7 Sep 2025). The argument for robustness is empirical: on smooth deterministic and noisy test problems, degeneracy correction prevents collapse into lower dimension, and reevaluation averages out random noise at vertices that significantly influence the search. The underlying DSM structure is preserved, so the method inherits the practical strengths and weaknesses of Nelder–Mead while attempting to mitigate two particular failure modes.

Several limitations are stated directly. rDSM does not solve the fundamental scaling problem in high-dimensional space. It improves robustness and may increase the applicability of DSM to higher dimensions, but CPU time may grow prohibitive, and future work is suggested in dimensionality reduction and hybrid solvers (Wang et al., 7 Sep 2025). It also handles only simple bounds through large penalties outside the feasible region, rather than general nonlinear constraints. In extremely noisy cases, even averaging may not suffice, and in multimodal landscapes rDSM can still get stuck in local minima.

A broader reading of the simplex literature suggests that “robustness” is not a single property but a family of design goals. In one line of work, robustness means escape from local minima by restarting or injecting randomness when simplex collapse is detected (Saboori et al., 2012). In another, it means regular geometry, sufficient decrease, and worst-case complexity guarantees (Cao et al., 22 Aug 2025). In rDSM, it means geometric correction of degenerated simplices and statistical correction of noisy function values (Wang et al., 7 Sep 2025). EGM adds yet another meaning: scheduled global exploration to avoid suboptimal local minima while preserving fast local simplex descent (Li et al., 2019). A plausible implication is that future “robust downhill simplex” designs will combine these axes—geometry control, noise handling, and explicit exploration—rather than treat them as mutually exclusive alternatives.