D Optimality

Definition

D-optimality (D-opt) is a scalar summary of an estimator’s error-covariance matrix drawn from
optimal-experiment theory: it quantifies the volume of the uncertainty ellipsoid, and minimizing it
selects the trajectory / experiment expected to yield the most confident belief. Papachristos et al.
(2019) use it as the BeliefGain metric in the inner (uncertainty-optimization) layer of a
receding-horizon planner: each candidate path is belief-propagated through the EKF, and the path with
the smallest D-opt value of the propagated pose-and-feature covariance is executed. The cited source
operates on a micro aerial vehicle (a hexarotor; aerial regime, planned on the reduced state
$[x,y,z,\psi ]$ with roll/pitch held near zero), not a space robot — the criterion itself is
platform-agnostic, but the belief-propagation model and the free-flying base dynamics differ from our case.

Key Equations

Symbols per notation.md.

Let $\mathbf{\Sigma }$ denote the estimator error-covariance over the $d=l_p+l_f$ pose-and-feature
states ($\mathbf{\Sigma }$ for covariance follows the convention on
fisher_information_matrix; not yet in notation.md). The D-optimality
metric used by the source (its Eq. 6, the Kiefer (1974) / Carrillo et al. (2012) unifying form) is the
geometric mean of the eigenvalues of $\mathbf{\Sigma }$:

$D_{\mathrm{opt}}(\mathbf{\Sigma })=\exp \left(\log \right[\det (\mathbf{\Sigma }){]}^{1/d})=[\det (\mathbf{\Sigma }){]}^{1/d}.$

The planner then assigns this as the belief gain of each admissible branch and minimizes over paths (the
source’s Eq. 7; here $\alpha$ is the branch/path index as written in the source, not the CVaR
confidence level $\alpha$ reserved in notation.md — disambiguate locally):

$g_{\alpha }^{\mathcal{B}}\equiv \mathbf{BeliefGain}(p_{\alpha }^{\mathcal{B}})=D_{\mathrm{opt}}({\mathbf{\Sigma }}_{p,f}(p_{\alpha }^{\mathcal{B}})),p^{\star }=\arg {\min }_{\alpha }{g}_{\alpha }^{\mathcal{B}}.$

The $1/d$ normalization is the dimensionless variant adopted to address the bare $\det (\mathbf{\Sigma })$
collapsing toward zero and to support task-completion checking; it equals the geometric mean of the
ellipsoid semi-axes squared.

Source Support

• papachristos2019localization — selects D-opt (after
  Carrillo et al. 2012, Kiefer 1974) as the BeliefGain scalar that ranks belief-propagated paths in
  the inner layer of a receding-horizon exploration planner; gives the normalized form (Eq. 6) and its
  $\mathcal{O}(l_p+l_f)$ diagonal-only cost. Aerial-robot (MAV) regime.

Related Topics

• fisher_information_matrix — the dual object: D-opt on the covariance
  $\mathbf{\Sigma }$ corresponds (via the Cramér–Rao bound $\mathbf{\Sigma }⪰{\mathbf{F}}^{-1}$)
  to a $\det$-based criterion on the FIM $\mathbf{F}$; D-opt is the determinant member of the
  A/D/E-optimality family.
• measurement_uncertainty — D-opt is the scalar readout of exactly this
  uncertainty, condensing the covariance ellipsoid into one number for path ranking.
• next_best_view — D-opt is the objective that turns view/path selection into an
  uncertainty-minimizing choice; here it re-ranks paths reaching a viewpoint already chosen for
  exploration gain (see also inspection_nbv).
• active_parameter_learning — same optimal-design lineage: D-opt is one
  acquisition criterion for choosing motions that shrink estimate uncertainty.
• inertial_parameter_identification — a D-optimal trajectory
  design directly targets the parameter covariance, the relevant link for identifying a target
  satellite’s inertial parameters in our free-flying inspection task.

Open Questions

• The source applies D-opt to a localization-and-mapping belief (robot pose + environment
  landmarks) on an MAV. Does the same criterion remain the right scalar when the estimated quantity is
  instead a target’s inertial parameters observed by a free-flying manipulator?
• Eq. 6 evaluates D-opt on diagonal terms only for $\mathcal{O}(d)$ cost; how much is lost by
  ignoring cross-covariances when the FFSM’s base–arm coupling makes pose and parameter errors
  strongly correlated?
• D-opt ranks the volume of the ellipsoid; for an inspection task with a dominant worst-axis error,
  is an E-optimal (worst-eigenvalue) or risk-aware (CVaR) criterion better aligned with our risk layer?