Home / Curated / Mathematical Methods / Active CMA covariance update

Active CMA covariance update

Mathematical Methods AIMech:ActiveCMACovarianceUpdate DRAFT

A matrix-valued covariance-update component from the original 2006 active CMA construction that combines retained search covariance and an evolution-path outer product with an equal-weight best-minus-worst mutation-step second moment.

Identity Notes & Concept Scope This record isolates the original 2006 equal-weight signed covariance component, with best-minus-worst old-mean moments and its combined path update. It is distinct from the maintained 2001 CMA base, 2003 nonnegative rank-mu blend, whole Active-CMA-ES, restart controllers and later negative-weight safeguards. Fresh postmerge collection at 1031c6b1f1f4492fbebec6eee1db5a6c68672d9e checked all 737 native and 442 pinned AIO record identities and bytes. Four searches included ignored and hidden files, yielding 1,463 occurrences grouped into 26 exact contents; root read all 26 complete contents. No normalized ID, label, alias or scoped-synonym collision was found for the explicit probes. The matches are carried active-covariance leads, unrelated Gaussian-process cross-covariance language, and existing IPOP boundary context. This finding is bounded to the pinned inventory and inspected semantic neighborhoods, not all AI literature. Existing owner 201 is resolved in place with later variants retained.

Claims & Grounded Evidence

6 Claims

Mechanism #signed-old-mean-mutation-moments

For integer dimension n>=1, positive scale sigma and current positive-definite C=B*D^2*B^T, with orthogonal B and positive diagonal D, write each mutation step as u_i=(offspring_i-old_mean)/sigma=B*D*z_i, where the z_i are independent standard-normal vectors. The source ranks lambda=4*mu offspring, with integer mu>=1, and forms Z=(1/mu)*sum_best_mu(u_i*u_i^T)-(1/mu)*sum_worst_mu(u_i*u_i^T). The two groups have equal weight magnitudes and opposite signs; intermediate ranks do not enter Z. Both moments use the old sampling mean and scale. They are not covariances centered on their selected-group means, and the negative covariance weights do not replace the source mean update with negative recombination weights.

supports Source: Improving Evolution Strategies through Active Covariance Matrix Adaptation · Locator: Section II, PDF p. 2, sampling and ranking steps 1–4; Section III, PDF p. 3, Equation 5 and the distinction from weighted mean recombination.

Mechanism #active-path-and-population-update

The source accumulates the selected mean step into p_new=(1-c_C)*p_old+sqrt(mu*c_C*(2-c_C))*mean_best_mu(u_i), with c_C=4/(n+4), and uses C_new=(1-c_cov)*C+c_cov*p_new*p_new^T+beta*Z. The updated path represents accumulated selected mean-step information, whereas Z uses the current generation’s individual best and worst steps. Equation 6 replaces the baseline hybrid coefficient arrangement; it is not obtained by retaining the old alpha mixture unchanged while merely assigning a negative sign to an extra step.

supports Source: Improving Evolution Strategies through Active Covariance Matrix Adaptation · Locator: Section II, PDF p. 2, step 5 and Equation 2; Section III, PDF p. 3, Equations 5–6.

Scope #source-rates-and-empirical-beta

For this active rule, the paper sets c_cov=2/(n+sqrt(2))^2 and beta=(4*mu-2)/((n+12)^2+4*mu). The beta expression was fitted after multiplying experimentally selected values by 0.6, using ellipsoid tests at n in {2,3,5,10,20,40,80} and mu in {2,n}, then rounding the fitted coefficients. This is an empirical parameter prescription, not a universally optimal rate or a guarantee that every updated matrix remains positive definite.

supports Source: Improving Evolution Strategies through Active Covariance Matrix Adaptation · Locator: Section III, PDF p. 3, c_cov expression following Equation 6; PDF p. 4, beta fitting discussion and Equation 7.

Scope #signed-moment-and-definiteness-boundary

Z is symmetric but can be indefinite, with rank at most min(n,2*mu). The negative summand subtracts the worst-step second moment relative to the same update without that summand; the full update need not decrease every direction. A bounded curator hand example uses n=mu=2, lambda=8, C=I, old_mean=0, sigma=1 and p_old=0, ranking by squared Euclidean norm. Let the best steps be ±(1/2)e2, the middle steps ±e1 and ±2e1, and the worst steps ±6e1. Then p_new=0 and Z=diag(-36,1/4). With the source rates, the first diagonal of C_new is 2*sqrt(2)-52/17<0. Thus even these parameter settings do not preserve positive definiteness for every supplied generation.

supports Source: Improving Evolution Strategies through Active Covariance Matrix Adaptation · Locator: Section III, PDF p. 3, Equations 5–6; PDF p. 4, Equation 7 and explicit positive-definiteness warning. The rank bound, directional qualification and two-dimensional example are curator linear algebra from these formulas, not reported author experiments.

Scope #neutral-selection-and-zero-path-scope

With a fixed pre-update C, independent zero-mean Gaussian mutation steps with covariance C, and ranking randomized independently of those steps, the best and worst groups have equal expected second moments, so E[Z]=0. This need not hold under fitness-based selection. With p_old=0 under the same neutral law, the first path has E[p_new*p_new^T]=c_C*(2-c_C)*C, giving E[C_new]=[1-c_cov*(1-c_C)^2]*C. This curator first-step calculation qualifies the source’s stationarity motivation: zero initialization is different from a stationary-path idealization. It is not a claim about the full adaptive run.

supports Source: Improving Evolution Strategies through Active Covariance Matrix Adaptation · Locator: Section II, PDF p. 2, zero-path initialization and step 5; Section III, PDF p. 3, random-selection discussion between Equations 5–6. The zero-path expectation is curator algebra under the stated frozen-law assumptions.

Scope #original-component-and-host-boundary

This identity is the original 2006 equal-weight signed covariance-update component, not the whole Active-CMA-ES optimizer. The source host uses unweighted best-offspring mean recombination and a separate cumulative step-size rule with a scalar exponential. That host operation is outside the defining Equations 5–6 component. The new subtractive moment differs from the reviewed positive rank-one or nonnegative hybrid covariance rules. The paper leaves weighted recombination for future work; arbitrary weighted-negative updates and later safeguards are not identified with this original construction.

supports Source: Improving Evolution Strategies through Active Covariance Matrix Adaptation · Locator: Section II, PDF pp. 2–3, Equations 1–4 and step 7; Section III, PDF p. 3, Equations 5–6 and exclusion of weighted mean recombination; Section V, PDF pp. 7–8, conclusions and future weighted-recombination work.

⚠️ Curated Limitations & Negative Results

The scope is the original equal-weight best/worst moment and combined covariance rule. The whole Active-CMA-ES optimizer, later unequal-weight negative updates, the distinct (1+1) active method, restart policies and modern implementations are not exact aliases of this component.
A positive-definite input matrix does not ensure a positive-definite output. The source says a detected loss can be fixed but does not specify an executable repair. Eigenvalue clipping, step normalization, negative-weight rescaling or exponential covariance updates are not silently added. Good conditioning and stable finite-precision eigendecomposition are separate concerns.
A rank or tie policy is required to define the selected groups in general. The source does not fully specify a tie convention. In the hand example, objective ties occur only within pairs whose order does not alter the top or bottom group. This example is not a sampled failure rate or a benchmark replication.
The neutral-selection expectation and zero-path calculation use explicitly fixed pre-update covariance and independent random ranking. They do not certify martingale behavior or stationarity throughout an adaptive run. The calculation is curator hand algebra, not an author erratum or an audit of the separate 2005 thesis argument cited by the paper.
The beta fit and reported comparisons concern the source’s specified tests. Neither the empirical absence of covariance failure in those runs nor the selected rates establish universal stability, speedup, objective improvement or global convergence. No optimizer, author code or benchmark was run, and numerical plot results are not curated here.
The primary bytes are an eight-page archived CiteSeer copy, not certified as identical to an author-hosted or publisher copy. Locators use PDF pages; the file’s build date is not the publication date, and differing external proceedings pagination is not resolved by the PDF labels. All eight native-text pages and the defining pages 1–4 visually were inspected in supporting research. The cited thesis and later active-method papers were not fully inspected.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:MathematicalFunction` Mathematical Function	related_match	Equations 5–6 define an explicit matrix-valued rule from the current covariance, supplied updated path, ranked mutation steps and scalar rates to a new symmetric matrix. This specific mathematical construction supports the relation, rather than generic computability of a whole optimizer. It asserts neither equivalence nor an upstream subclass edge, scalar activation semantics or universal positive-definite output. The host’s separate exponential step-size operation is not imported.

Evidence Sources & Primary Literature

1 Source

Source ID	Publication / Document	Type / Access	Reference & Checksum
`jastrebski2006`	Improving Evolution Strategies through Active Covariance Matrix Adaptation Grahame A. Jastrebski, Dirk V. Arnold	primary paper full_text	ID: https://doi.org/10.1109/CEC.2006.1688662 sha256:12328a58db10…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-15T06:52:10Z

Updated: 2026-10-15T06:52:10Z

Source File: data/curated/ActiveCMACovarianceUpdate.yaml