State Transition and Sensitivity Matrices

State perturbations, variational equations, STM propagation, covariance mapping, parameter and control sensitivities, measurement sensitivities, observability, maneuver targeting, numerical differentiation, verification, conditioning, coordinate transformations, and general spacecraft GNC applications.

4.3.8 State Transition and Sensitivity Matrices

The previous section showed that filtering, smoothing, batch estimation, and trajectory reconstruction all depend on relationships between spacecraft states at different times. For example, the RTS smoother uses a transition matrix to transfer future-state information backward, while dynamic estimation requires the sensitivity of later measurements to an earlier reference state.

This leads to one of the most important mathematical tools in spacecraft dynamics and estimation:

\[ \boxed{ \text{State Transition Matrix} \quad\text{and}\quad \text{Sensitivity Matrices} } \]

These matrices answer questions such as:

If I slightly perturb the spacecraft state now, how will that perturbation evolve later?

If I change an uncertain physical parameter, how does the trajectory change?

If a maneuver magnitude changes slightly, how does the final state change?

How sensitive is a future sensor measurement to the initial position, velocity, attitude, bias, or model parameter?

The central idea is

\[ \boxed{ \text{Small Input Perturbation} \rightarrow \text{Dynamic Propagation} \rightarrow \text{Future State Perturbation} } \]

and, more generally,

\[ \boxed{ \text{Sensitivity} = \frac{\partial(\text{output})} {\partial(\text{input})}. } \]

The State Transition Matrix, or STM,

\[ \boxed{ \boldsymbol{\Phi}(t,t_0) = \frac{ \partial\mathbf x(t) }{ \partial\mathbf x(t_0) } } \]

describes the first-order sensitivity of the state at time t to the state at an earlier reference time t_0.

For a small perturbation,

\[ \boxed{ \delta\mathbf x(t) \approx \boldsymbol{\Phi}(t,t_0) \delta\mathbf x(t_0). } \]

That single relationship appears throughout:

4.3.8.1 Why State Transition Matrices Are Needed

Consider a spacecraft state

\[ \mathbf x(t). \]

Suppose the state at reference epoch t_0 is slightly wrong:

\[ \mathbf x(t_0) = \bar{\mathbf x}(t_0) + \delta\mathbf x_0. \]

After propagation, that small initial error produces a later error

\[ \delta\mathbf x(t). \]

The fundamental question is:

\[ \boxed{ \text{How does } \delta\mathbf x_0 \text{ propagate to } \delta\mathbf x(t)? } \]

For sufficiently small perturbations,

\[ \boxed{ \delta\mathbf x(t) = \boldsymbol{\Phi}(t,t_0) \delta\mathbf x_0. } \]

Thus the STM is the linear map between state perturbations at two times.

4.3.8.2 Nonlinear Dynamic System

Consider continuous nonlinear spacecraft dynamics

\[ \boxed{ \dot{\mathbf x} = \mathbf f( \mathbf x, \mathbf u, \boldsymbol{\theta}, t ) } \]

where

The nominal trajectory satisfies

\[ \boxed{ \dot{\bar{\mathbf x}} = \mathbf f( \bar{\mathbf x}, \bar{\mathbf u}, \bar{\boldsymbol{\theta}}, t ). } \]

The purpose of sensitivity analysis is to determine how small changes around this nominal solution evolve.

4.3.8.3 Nominal State and Perturbed State

Write the actual state as

\[ \boxed{ \mathbf x = \bar{\mathbf x} + \delta\mathbf x. } \]

Similarly,

\[ \mathbf u = \bar{\mathbf u} + \delta\mathbf u, \]

and

\[ \boldsymbol{\theta} = \bar{\boldsymbol{\theta}} + \delta\boldsymbol{\theta}. \]

Then

\[ \dot{\mathbf x} = \mathbf f( \bar{\mathbf x} + \delta\mathbf x, \bar{\mathbf u} + \delta\mathbf u, \bar{\boldsymbol{\theta}} + \delta\boldsymbol{\theta}, t ). \]

The goal is to derive an approximate equation for the perturbation

\[ \delta\mathbf x. \]

4.3.8.4 First-Order Linearization of the Dynamics

Expand the nonlinear dynamics around the nominal trajectory.

Using a first-order Taylor expansion,

\[ \begin{aligned} \mathbf f( \bar{\mathbf x}+\delta\mathbf x, \bar{\mathbf u}+\delta\mathbf u, \bar{\boldsymbol{\theta}}+\delta\boldsymbol{\theta} ) \approx{}& \mathbf f( \bar{\mathbf x}, \bar{\mathbf u}, \bar{\boldsymbol{\theta}} ) \\ &+ \frac{\partial\mathbf f}{\partial\mathbf x} \delta\mathbf x + \frac{\partial\mathbf f}{\partial\mathbf u} \delta\mathbf u + \frac{\partial\mathbf f}{\partial\boldsymbol{\theta}} \delta\boldsymbol{\theta}. \end{aligned} \]

Subtracting the nominal dynamics gives

\[ \boxed{ \delta\dot{\mathbf x} = \mathbf A(t)\delta\mathbf x + \mathbf B(t)\delta\mathbf u + \mathbf E(t)\delta\boldsymbol{\theta}. } \]

where

\[ \boxed{ \mathbf A(t) = \left. \frac{\partial\mathbf f} {\partial\mathbf x} \right|_{\bar{\mathbf x}} } \]

\[ \boxed{ \mathbf B(t) = \left. \frac{\partial\mathbf f} {\partial\mathbf u} \right|_{\bar{\mathbf x}} } \]

and

\[ \boxed{ \mathbf E(t) = \left. \frac{\partial\mathbf f} {\partial\boldsymbol{\theta}} \right|_{\bar{\mathbf x}}. } \]

Engineering interpretation: The linearization is local: the matrices \(\mathbf A\), \(\mathbf B\), and \(\mathbf E\) are evaluated about the nominal trajectory and therefore generally vary with time in a nonlinear spacecraft model.

4.3.8.5 The Dynamics Jacobian

The matrix

\[ \boxed{ \mathbf A(t) = \frac{\partial\mathbf f}{\partial\mathbf x} } \]

is the local state Jacobian of the dynamics.

If

\[ \mathbf x\in\mathbb R^n, \]

then

\[ \boxed{ \mathbf A\in\mathbb R^{n\times n}. } \]

Each element

\[ A_{ij} = \frac{\partial f_i}{\partial x_j} \]

describes how the derivative of state component i responds locally to a perturbation in state component j.

Thus,

\[ \boxed{ \mathbf A = \text{Local Dynamic Coupling Matrix}. } \]

4.3.8.6 Variational Equations

Ignoring input and parameter perturbations temporarily,

\[ \delta\mathbf u=\mathbf0, \qquad \delta\boldsymbol{\theta}=\mathbf0, \]

the perturbation dynamics become

\[ \boxed{ \delta\dot{\mathbf x} = \mathbf A(t)\delta\mathbf x. } \]

This system is called the variational equation.

It describes how infinitesimal perturbations evolve along the nominal nonlinear trajectory.

4.3.8.7 Definition of the State Transition Matrix

The STM is formally defined as

\[ \boxed{ \boldsymbol{\Phi}(t,t_0) = \frac{ \partial\mathbf x(t) }{ \partial\mathbf x(t_0) }. } \]

Its element

\[ \Phi_{ij} = \frac{ \partial x_i(t) }{ \partial x_j(t_0) } \]

answers:

How much does state component i at time t change due to a small change in state component j at the reference epoch?

4.3.8.8 Perturbation Propagation

The STM maps an initial perturbation to a later perturbation:

\[ \boxed{ \delta\mathbf x(t) = \boldsymbol{\Phi}(t,t_0) \delta\mathbf x(t_0). } \]

For example,

\[ \delta\mathbf x_0 = \begin{bmatrix} \delta\mathbf r_0\\ \delta\mathbf v_0 \end{bmatrix} \]

may create

\[ \delta\mathbf x(t) = \begin{bmatrix} \delta\mathbf r(t)\\ \delta\mathbf v(t) \end{bmatrix}. \]

The STM tells us exactly how these first-order errors are coupled.

4.3.8.9 Initial Condition of the STM

At the same epoch,

\[ t=t_0. \]

A perturbation maps to itself:

\[ \delta\mathbf x(t_0) = \delta\mathbf x(t_0). \]

Therefore,

\[ \boxed{ \boldsymbol{\Phi}(t_0,t_0) = \mathbf I. } \]

This is the fundamental initial condition for STM integration.

4.3.8.10 Differential Equation for the STM

Starting from

\[ \delta\mathbf x(t) = \boldsymbol{\Phi}(t,t_0) \delta\mathbf x_0, \]

differentiate with respect to time:

\[ \delta\dot{\mathbf x} = \dot{\boldsymbol{\Phi}} \delta\mathbf x_0. \]

But the variational equation gives

\[ \delta\dot{\mathbf x} = \mathbf A(t) \delta\mathbf x = \mathbf A(t) \boldsymbol{\Phi} \delta\mathbf x_0. \]

Therefore,

\[ \boxed{ \dot{\boldsymbol{\Phi}}(t,t_0) = \mathbf A(t) \boldsymbol{\Phi}(t,t_0). } \]

with

\[ \boxed{ \boldsymbol{\Phi}(t_0,t_0) = \mathbf I. } \]

This is the most important differential equation for computing the STM.

Engineering interpretation: This differential equation is what turns the STM from a definition into a computable object. Integrating it alongside the nominal dynamics gives the exact first-order sensitivity associated with the chosen model and reference trajectory.

4.3.8.11 Coupled State and STM Integration

For a nonlinear spacecraft model, the nominal state and STM are usually propagated together.

The augmented differential system is

\[ \boxed{ \begin{aligned} \dot{\mathbf x} &= \mathbf f(\mathbf x,t), \\ \dot{\boldsymbol{\Phi}} &= \mathbf A(\mathbf x,t) \boldsymbol{\Phi}. \end{aligned} } \]

The Jacobian must be evaluated along the evolving nominal trajectory.

Therefore,

\[ \boxed{ \mathbf A(t) = \mathbf A[ \mathbf x(t),t ]. } \]

The STM is trajectory dependent.

4.3.8.12 STM for a Linear Time-Invariant System

If

\[ \dot{\mathbf x} = \mathbf A\mathbf x \]

with constant \mathbf A, then

\[ \boxed{ \boldsymbol{\Phi}(t,t_0) = e^{\mathbf A(t-t_0)}. } \]

The matrix exponential is

\[ e^{\mathbf A\Delta t} = \mathbf I + \mathbf A\Delta t + \frac{\mathbf A^2\Delta t^2}{2!} + \cdots. \]

Thus the familiar linear-system solution is

\[ \boxed{ \mathbf x(t) = \boldsymbol{\Phi}(t,t_0) \mathbf x(t_0). } \]

4.3.8.13 STM for a Time-Varying System

If

\[ \mathbf A=\mathbf A(t), \]

then, in general,

\[ \boldsymbol{\Phi}(t,t_0) \neq e^{\int_{t_0}^{t}\mathbf A(\tau)d\tau} \]

unless the matrices commute appropriately over time.

The STM is therefore normally obtained by directly integrating

\[ \boxed{ \dot{\boldsymbol{\Phi}} = \mathbf A(t)\boldsymbol{\Phi}. } \]

4.3.8.14 STM Composition Property

Suppose

\[ t_0

Then

\[ \delta\mathbf x(t_1) = \boldsymbol{\Phi}(t_1,t_0) \delta\mathbf x(t_0) \]

and

\[ \delta\mathbf x(t_2) = \boldsymbol{\Phi}(t_2,t_1) \delta\mathbf x(t_1). \]

Therefore,

\[ \boxed{ \boldsymbol{\Phi}(t_2,t_0) = \boldsymbol{\Phi}(t_2,t_1) \boldsymbol{\Phi}(t_1,t_0). } \]

This is the semigroup or composition property.

It is essential for multi-stage propagation.

4.3.8.15 STM Inverse Property

Provided the flow is locally invertible,

\[ \boxed{ \boldsymbol{\Phi}(t_0,t) = \boldsymbol{\Phi}(t,t_0)^{-1}. } \]

Therefore,

\[ \boxed{ \boldsymbol{\Phi}(t,t_0) \boldsymbol{\Phi}(t_0,t) = \mathbf I. } \]

This property is useful for verification and backward sensitivity mapping.

4.3.8.16 Identity Property

Another basic property is

\[ \boxed{ \boldsymbol{\Phi}(t,t) = \mathbf I. } \]

Together, the major STM properties are

\[ \boxed{ \begin{aligned} \boldsymbol{\Phi}(t,t)&=\mathbf I, \\ \boldsymbol{\Phi}(t_2,t_0) &= \boldsymbol{\Phi}(t_2,t_1) \boldsymbol{\Phi}(t_1,t_0), \\ \boldsymbol{\Phi}(t_0,t) &= \boldsymbol{\Phi}(t,t_0)^{-1}. \end{aligned} } \]

4.3.8.17 Block Structure for Position and Velocity

For a translational state

\[ \mathbf x = \begin{bmatrix} \mathbf r\\ \mathbf v \end{bmatrix}, \]

the STM can be partitioned as

\[ \boxed{ \boldsymbol{\Phi} = \begin{bmatrix} \boldsymbol{\Phi}_{rr} & \boldsymbol{\Phi}_{rv} \\ \boldsymbol{\Phi}_{vr} & \boldsymbol{\Phi}_{vv} \end{bmatrix}. } \]

These blocks have direct interpretations.

Engineering interpretation: The four translational STM blocks reveal different physical pathways for uncertainty growth. In particular, \(\boldsymbol{\Phi}_{rv}\) explains why even a small velocity error can become a substantial future position error.

4.3.8.18 Position-to-Position Sensitivity

\[ \boxed{ \boldsymbol{\Phi}_{rr} = \frac{ \partial\mathbf r(t) }{ \partial\mathbf r_0 }. } \]

It describes how initial-position perturbations influence later position.

4.3.8.19 Velocity-to-Position Sensitivity

\[ \boxed{ \boldsymbol{\Phi}_{rv} = \frac{ \partial\mathbf r(t) }{ \partial\mathbf v_0 }. } \]

This block is especially important in trajectory targeting because small velocity changes can accumulate into large position changes later.

4.3.8.20 Position-to-Velocity Sensitivity

\[ \boxed{ \boldsymbol{\Phi}_{vr} = \frac{ \partial\mathbf v(t) }{ \partial\mathbf r_0 }. } \]

This describes how initial-position errors change later velocity.

4.3.8.21 Velocity-to-Velocity Sensitivity

\[ \boxed{ \boldsymbol{\Phi}_{vv} = \frac{ \partial\mathbf v(t) }{ \partial\mathbf v_0 }. } \]

It describes how an initial velocity perturbation influences later velocity.

4.3.8.22 Simple Constant-Velocity Example

Consider

\[ \dot{\mathbf r} = \mathbf v, \qquad \dot{\mathbf v} = \mathbf0. \]

Then

\[ \mathbf r(t) = \mathbf r_0 + \mathbf v_0\Delta t, \]

\[ \mathbf v(t) = \mathbf v_0. \]

Therefore,

\[ \boxed{ \boldsymbol{\Phi} = \begin{bmatrix} \mathbf I & \Delta t\,\mathbf I \\ \mathbf0 & \mathbf I \end{bmatrix}. } \]

This clearly shows how velocity uncertainty grows into position uncertainty.

4.3.8.23 STM for Two-Body Orbital Dynamics

For a simplified point-mass gravitational model,

\[ \dot{\mathbf r} = \mathbf v, \]

\[ \dot{\mathbf v} = -\mu \frac{\mathbf r}{r^3}. \]

The state Jacobian is

\[ \boxed{ \mathbf A = \begin{bmatrix} \mathbf0 & \mathbf I\\ \mathbf G & \mathbf0 \end{bmatrix} } \]

where

\[ \mathbf G = \frac{\partial\mathbf a_g}{\partial\mathbf r}. \]

4.3.8.24 Gravity Gradient Matrix

For

\[ \mathbf a_g = -\mu \frac{\mathbf r}{r^3}, \]

the position derivative is

\[ \boxed{ \mathbf G = \mu \left( \frac{ 3\mathbf r\mathbf r^T }{ r^5 } - \frac{\mathbf I}{r^3} \right). } \]

This is the familiar gravity-gradient structure.

Therefore

\[ \boxed{ \mathbf A = \begin{bmatrix} \mathbf0 & \mathbf I\\ \mu \left( \frac{3\mathbf r\mathbf r^T}{r^5} - \frac{\mathbf I}{r^3} \right) & \mathbf0 \end{bmatrix}. } \]

Engineering interpretation: The gravity-gradient matrix is the local curvature of the gravitational acceleration field. It is the term that couples position perturbations into acceleration perturbations in two-body variational dynamics.

4.3.8.25 STM for Relative Motion

For linearized relative orbital motion,

\[ \dot{\mathbf x}_{rel} = \mathbf A_{rel} \mathbf x_{rel}, \]

the STM is

\[ \boxed{ \boldsymbol{\Phi}_{rel} = e^{\mathbf A_{rel}\Delta t}. } \]

For models such as Hill–Clohessy–Wiltshire dynamics, closed-form STMs can be derived.

These matrices are useful for general proximity-motion analysis, targeting, guidance, and uncertainty propagation.

4.3.8.26 General RPO Interpretation

Consider a relative state

\[ \mathbf x_{rel} = \begin{bmatrix} \mathbf r_{rel}\\ \mathbf v_{rel} \end{bmatrix}. \]

A small error at reference epoch,

\[ \delta\mathbf x_{rel,0}, \]

propagates as

\[ \boxed{ \delta\mathbf x_{rel}(t) = \boldsymbol{\Phi}_{rel}(t,t_0) \delta\mathbf x_{rel,0}. } \]

This can answer general questions such as:

This is a general RPO concept and is not tied to any particular mission.

4.3.8.27 Discrete-Time State Transition Matrix

For a discrete system

\[ \boxed{ \mathbf x_{k+1} = \mathbf f_k( \mathbf x_k ) } \]

the local perturbation model is

\[ \boxed{ \delta\mathbf x_{k+1} = \mathbf F_k \delta\mathbf x_k } \]

where

\[ \boxed{ \mathbf F_k = \left. \frac{\partial\mathbf f_k} {\partial\mathbf x_k} \right|_{\bar{\mathbf x}_k}. } \]

Thus \mathbf F_k is the discrete-time STM over one propagation step.

4.3.8.28 Multi-Step Discrete STM

Over several discrete steps,

\[ \delta\mathbf x_N = \mathbf F_{N-1} \cdots \mathbf F_1 \mathbf F_0 \delta\mathbf x_0. \]

Therefore,

\[ \boxed{ \boldsymbol{\Phi}_{N,0} = \mathbf F_{N-1} \mathbf F_{N-2} \cdots \mathbf F_0. } \]

Matrix multiplication order matters.

4.3.8.29 Continuous-Discrete Relationship

For continuous dynamics over a sampling interval,

\[ t_k \rightarrow t_{k+1}, \]

the discrete transition matrix is

\[ \boxed{ \mathbf F_k = \boldsymbol{\Phi}(t_{k+1},t_k). } \]

Thus continuous STM theory and discrete filtering equations are directly connected.

4.3.8.30 Covariance Propagation

If

\[ \delta\mathbf x(t) = \boldsymbol{\Phi}(t,t_0) \delta\mathbf x_0, \]

and

\[ \mathbf P_0 = E[ \delta\mathbf x_0 \delta\mathbf x_0^T ], \]

then

\[ \boxed{ \mathbf P(t) = \boldsymbol{\Phi}(t,t_0) \mathbf P_0 \boldsymbol{\Phi}(t,t_0)^T. } \]

This is deterministic covariance propagation.

4.3.8.31 Why Covariance Uses Two STMs

Covariance is quadratic in the state error.

Since

\[ \delta\mathbf x = \boldsymbol{\Phi}\delta\mathbf x_0, \]

then

\[ \begin{aligned} \mathbf P &= E[ \delta\mathbf x \delta\mathbf x^T ] \\ &= E[ \boldsymbol{\Phi} \delta\mathbf x_0 \delta\mathbf x_0^T \boldsymbol{\Phi}^T ] \\ &= \boldsymbol{\Phi} \mathbf P_0 \boldsymbol{\Phi}^T. \end{aligned} \]

Therefore,

\[ \boxed{ \text{State Perturbation} \rightarrow \boldsymbol{\Phi} } \]

but

\[ \boxed{ \text{Covariance} \rightarrow \boldsymbol{\Phi} \mathbf P \boldsymbol{\Phi}^T. } \]

4.3.8.32 Covariance Propagation with Process Noise

If dynamic uncertainty is present,

\[ \delta\dot{\mathbf x} = \mathbf A\delta\mathbf x + \mathbf G\mathbf w, \]

then covariance propagation must include process noise.

In discrete form,

\[ \boxed{ \mathbf P_{k+1}^{-} = \mathbf F_k \mathbf P_k^{+} \mathbf F_k^T + \mathbf Q_k. } \]

In continuous time,

\[ \boxed{ \dot{\mathbf P} = \mathbf A\mathbf P + \mathbf P\mathbf A^T + \mathbf G \mathbf Q_c \mathbf G^T. } \]

This is the continuous covariance differential equation.

4.3.8.33 STM and Kalman Filtering

The prediction step of a linear or extended Kalman filter uses

\[ \boxed{ \mathbf P_k^{-} = \mathbf F_{k-1} \mathbf P_{k-1}^{+} \mathbf F_{k-1}^T + \mathbf Q_{k-1}. } \]

Here,

\[ \mathbf F_{k-1} \]

is precisely the local state-transition map from one filter epoch to the next.

Thus,

\[ \boxed{ \text{STM} \rightarrow \text{Filter State and Covariance Prediction}. } \]

4.3.8.34 STM and Smoothing

The RTS smoother gain is

\[ \mathbf G_k = \mathbf P_{k|k} \mathbf F_k^T \mathbf P_{k+1|k}^{-1}. \]

Therefore the STM controls how future information is mapped backward.

Conceptually,

\[ \boxed{ \text{Dynamic State Coupling} \rightarrow \mathbf F_k \rightarrow \text{Backward Smoothing Correction}. } \]

This directly connects this section to smoothing and retrodiction.

4.3.8.35 Measurement Sensitivity to Current State

For

\[ \mathbf z_k = h_k( \mathbf x_k ) + \mathbf v_k, \]

the measurement Jacobian is

\[ \boxed{ \mathbf H_k = \left. \frac{\partial h_k} {\partial\mathbf x_k} \right|. } \]

It gives measurement sensitivity to the state at the measurement epoch.

4.3.8.36 Measurement Sensitivity to Reference State

In batch estimation, the unknown may instead be a reference state

\[ \mathbf x_0. \]

Using the chain rule,

\[ \boxed{ \frac{\partial h_k} {\partial\mathbf x_0} = \frac{\partial h_k} {\partial\mathbf x_k} \frac{\partial\mathbf x_k} {\partial\mathbf x_0}. } \]

Therefore,

\[ \boxed{ \mathbf G_k = \mathbf H_k \boldsymbol{\Phi}_{k,0}. } \]

This is the dynamic measurement sensitivity matrix.

Engineering interpretation: This chain-rule expression is fundamental in orbit determination: the sensor is sensitive to the state at measurement time, while the STM maps that local sensitivity back to the solve-for state at the reference epoch.

4.3.8.37 Importance in Batch Estimation

Each measurement at time t_k contributes sensitivity

\[ \boxed{ \mathbf G_k = \mathbf H_k \boldsymbol{\Phi}_{k,0}. } \]

Stacking them gives

\[ \boxed{ \mathbf G = \begin{bmatrix} \mathbf H_1\boldsymbol{\Phi}_{1,0}\\ \mathbf H_2\boldsymbol{\Phi}_{2,0}\\ \vdots\\ \mathbf H_N\boldsymbol{\Phi}_{N,0} \end{bmatrix}. } \]

The normal matrix then becomes

\[ \boxed{ \mathbf N = \mathbf G^T \mathbf R^{-1} \mathbf G. } \]

Thus,

\[ \boxed{ \text{STM} \rightarrow \text{Dynamic Design Matrix} \rightarrow \text{Batch State Correction}. } \]

4.3.8.38 Why the STM Matters in Orbit Determination

Suppose an observation is taken long after the reference epoch.

The measurement is sensitive directly to

\[ \mathbf x(t_k). \]

But the estimator may solve for

\[ \mathbf x(t_0). \]

The STM supplies the connection:

\[ \boxed{ \mathbf x(t_0) \rightarrow \mathbf x(t_k) \rightarrow \mathbf z_k. } \]

Without this sensitivity mapping, the estimator would not know how to correct the reference state from later observations.

4.3.8.39 Parameter Sensitivity Matrix

Suppose the dynamics depend on parameter vector

\[ \boldsymbol{\theta}. \]

Define the parameter sensitivity matrix

\[ \boxed{ \mathbf S_\theta(t) = \frac{ \partial\mathbf x(t) }{ \partial\boldsymbol{\theta} }. } \]

If

\[ \mathbf x\in\mathbb R^n \]

and

\[ \boldsymbol{\theta}\in\mathbb R^p, \]

then

\[ \boxed{ \mathbf S_\theta \in \mathbb R^{n\times p}. } \]

4.3.8.40 Meaning of Parameter Sensitivity

An element

\[ \left[ \mathbf S_\theta \right]_{ij} = \frac{ \partial x_i(t) }{ \partial\theta_j } \]

answers:

If parameter \theta_j changes slightly, how does state component x_i at time t change?

This is fundamental for estimating parameters such as:

4.3.8.41 Differential Equation for Parameter Sensitivity

Starting from

\[ \dot{\mathbf x} = \mathbf f( \mathbf x, \boldsymbol{\theta}, t ), \]

differentiate with respect to \boldsymbol{\theta}:

\[ \frac{d}{dt} \left( \frac{\partial\mathbf x} {\partial\boldsymbol{\theta}} \right) = \frac{\partial\mathbf f}{\partial\mathbf x} \frac{\partial\mathbf x}{\partial\boldsymbol{\theta}} + \frac{\partial\mathbf f}{\partial\boldsymbol{\theta}}. \]

Therefore,

\[ \boxed{ \dot{\mathbf S}_\theta = \mathbf A \mathbf S_\theta + \mathbf E. } \]

where

\[ \boxed{ \mathbf E = \frac{\partial\mathbf f} {\partial\boldsymbol{\theta}}. } \]

Engineering interpretation: Parameter sensitivities are forced variational equations. Unlike the STM, which evolves from an identity initial condition, parameter sensitivity also contains the direct forcing term \(\mathbf E=\partial\mathbf f/\partial\boldsymbol{\theta}\).

4.3.8.42 Initial Condition for Parameter Sensitivity

If the initial state does not depend on the parameter,

\[ \boxed{ \mathbf S_\theta(t_0) = \mathbf0. } \]

But if

\[ \mathbf x_0 = \mathbf x_0( \boldsymbol{\theta} ), \]

then

\[ \boxed{ \mathbf S_\theta(t_0) = \frac{\partial\mathbf x_0} {\partial\boldsymbol{\theta}}. } \]

This distinction matters in augmented estimation problems.

4.3.8.43 Integral Form of Parameter Sensitivity

The solution of

\[ \dot{\mathbf S}_\theta = \mathbf A \mathbf S_\theta + \mathbf E \]

can be written as

\[ \boxed{ \mathbf S_\theta(t) = \boldsymbol{\Phi}(t,t_0) \mathbf S_\theta(t_0) + \int_{t_0}^{t} \boldsymbol{\Phi}(t,\tau) \mathbf E(\tau) \,d\tau. } \]

This equation shows that parameter sensitivity is accumulated continuously along the trajectory.

4.3.8.44 Measurement Sensitivity to Parameters

A measurement can depend on a parameter directly or indirectly through the trajectory.

Using the chain rule,

\[ \boxed{ \frac{\partial h} {\partial\boldsymbol{\theta}} = \frac{\partial h} {\partial\mathbf x} \frac{\partial\mathbf x} {\partial\boldsymbol{\theta}} + \left. \frac{\partial h} {\partial\boldsymbol{\theta}} \right|_{\text{direct}}. } \]

Thus,

\[ \boxed{ \mathbf H_{\theta,k} = \mathbf H_{x,k} \mathbf S_{\theta,k} + \mathbf H_{\theta,k}^{direct}. } \]

4.3.8.45 Direct versus Indirect Parameter Sensitivity

A parameter may affect a measurement in two ways.

Indirect effect

\[ \theta \rightarrow \text{trajectory} \rightarrow \text{measurement}. \]

Direct effect

\[ \theta \rightarrow \text{measurement model directly}. \]

Both terms must be included.

Ignoring one can lead to incorrect parameter estimation.

4.3.8.46 Augmented State Approach

Instead of maintaining a separate parameter sensitivity matrix, parameters can be included in an augmented state:

\[ \boxed{ \mathbf x_a = \begin{bmatrix} \mathbf x\\ \boldsymbol{\theta} \end{bmatrix}. } \]

For constant parameters,

\[ \boxed{ \dot{\boldsymbol{\theta}} = \mathbf0. } \]

Then an augmented STM can propagate state and parameter sensitivity together.

4.3.8.47 Augmented Dynamics Jacobian

For

\[ \dot{\mathbf x} = \mathbf f( \mathbf x,\boldsymbol{\theta} ), \]

and

\[ \dot{\boldsymbol{\theta}} = \mathbf0, \]

the augmented Jacobian is

\[ \boxed{ \mathbf A_a = \begin{bmatrix} \mathbf A & \mathbf E \\ \mathbf0 & \mathbf0 \end{bmatrix}. } \]

The resulting augmented STM contains both ordinary state-transition and parameter-sensitivity blocks.

4.3.8.48 Augmented STM Structure

The augmented STM may be written as

\[ \boxed{ \boldsymbol{\Phi}_a = \begin{bmatrix} \boldsymbol{\Phi}_{xx} & \boldsymbol{\Phi}_{x\theta} \\ \boldsymbol{\Phi}_{\theta x} & \boldsymbol{\Phi}_{\theta\theta} \end{bmatrix}. } \]

For constant independent parameters,

\[ \boldsymbol{\Phi}_{\theta x} = \mathbf0, \]

and

\[ \boldsymbol{\Phi}_{\theta\theta} = \mathbf I. \]

The block

\[ \boxed{ \boldsymbol{\Phi}_{x\theta} } \]

contains trajectory sensitivity to the parameters.

4.3.8.49 Control-Input Sensitivity

Now consider a control input

\[ \mathbf u. \]

Define

\[ \boxed{ \mathbf S_u(t) = \frac{ \partial\mathbf x(t) }{ \partial\mathbf u }. } \]

For continuous distributed input perturbations,

\[ \delta\dot{\mathbf x} = \mathbf A\delta\mathbf x + \mathbf B\delta\mathbf u. \]

The corresponding response is

\[ \boxed{ \delta\mathbf x(t) = \boldsymbol{\Phi}(t,t_0) \delta\mathbf x_0 + \int_{t_0}^{t} \boldsymbol{\Phi}(t,\tau) \mathbf B(\tau) \delta\mathbf u(\tau) \,d\tau. } \]

Engineering interpretation: Control sensitivity and parameter sensitivity have the same mathematical structure: both describe how an external quantity changes the propagated trajectory. The difference is in physical interpretation and how that quantity enters the dynamics.

4.3.8.50 Physical Meaning of Control Sensitivity

The control-sensitivity relationship answers:

If the commanded acceleration, thrust, torque, burn duration, or other control changes slightly, how will the later state change?

This is important in

4.3.8.51 Maneuver Sensitivity

Suppose an impulsive velocity maneuver is applied at t_m:

\[ \mathbf v(t_m^+) = \mathbf v(t_m^-) + \Delta\mathbf v. \]

Then, for t>t_m,

\[ \boxed{ \frac{ \partial\mathbf x(t) }{ \partial\Delta\mathbf v } = \boldsymbol{\Phi}(t,t_m) \mathbf B_{\Delta v}. } \]

For a translational state,

\[ \mathbf B_{\Delta v} = \begin{bmatrix} \mathbf0\\ \mathbf I \end{bmatrix}. \]

4.3.8.52 Maneuver Targeting

Suppose a desired terminal position is

\[ \mathbf r_f^\star. \]

Current prediction gives

\[ \mathbf r_f. \]

The terminal error is

\[ \Delta\mathbf r_f = \mathbf r_f^\star-\mathbf r_f. \]

If

\[ \mathbf S_{\Delta v} = \frac{ \partial\mathbf r_f }{ \partial\Delta\mathbf v }, \]

then a first-order maneuver correction may satisfy

\[ \boxed{ \mathbf S_{\Delta v} \delta(\Delta\mathbf v) \approx \Delta\mathbf r_f. } \]

This is a basic targeting principle.

4.3.8.53 Targeting Matrix

More generally, define a terminal constraint

\[ \mathbf c( \mathbf x_f )=\mathbf0. \]

Its sensitivity to maneuver variables \mathbf u is

\[ \boxed{ \mathbf T = \frac{ \partial\mathbf c }{ \partial\mathbf u } = \frac{ \partial\mathbf c }{ \partial\mathbf x_f } \frac{ \partial\mathbf x_f }{ \partial\mathbf u }. } \]

This matrix may be called a targeting or sensitivity matrix.

4.3.8.54 Sensitivity to Burn Time

A maneuver outcome may depend not only on impulse magnitude but also on burn timing.

Define

\[ \boxed{ \mathbf S_t = \frac{ \partial\mathbf x_f }{ \partial t_m }. } \]

A small timing error produces

\[ \boxed{ \delta\mathbf x_f \approx \mathbf S_t \delta t_m. } \]

This is important when maneuver timing uncertainty affects terminal conditions.

4.3.8.55 Sensitivity to Burn Duration

For finite burns, the final state may depend on burn duration

\[ T_b. \]

Define

\[ \boxed{ \mathbf S_T = \frac{ \partial\mathbf x_f }{ \partial T_b }. } \]

This can be used to assess timing-error effects or refine finite-burn targeting.

4.3.8.56 Sensitivity to Thrust Magnitude

Let

\[ T \]

be thrust magnitude.

Then

\[ \boxed{ \mathbf S_T = \frac{ \partial\mathbf x_f }{ \partial T }. } \]

This is useful for:

4.3.8.57 Sensitivity to Thrust Direction

If the thrust direction is parameterized by angle vector

\[ \boldsymbol{\alpha}, \]

define

\[ \boxed{ \mathbf S_\alpha = \frac{ \partial\mathbf x_f }{ \partial\boldsymbol{\alpha} }. } \]

Misalignment uncertainty can then be approximated by

\[ \boxed{ \delta\mathbf x_f \approx \mathbf S_\alpha \delta\boldsymbol{\alpha}. } \]

4.3.8.58 Sensitivity to Sensor Bias

Suppose measurement bias is

\[ \mathbf b. \]

For

\[ \mathbf z = h(\mathbf x) + \mathbf b + \mathbf v, \]

the direct bias sensitivity is

\[ \boxed{ \frac{\partial\mathbf z}{\partial\mathbf b} = \mathbf I. } \]

When bias is estimated jointly with the state, its indirect dynamic correlations may also affect the complete design matrix.

4.3.8.59 Trajectory Sensitivity

A general trajectory sensitivity is

\[ \boxed{ \frac{ \partial\mathbf x(t) }{ \partial\boldsymbol{\eta} } } \]

where \boldsymbol{\eta} may represent:

Thus the STM is one member of a broader family of trajectory sensitivity matrices.

4.3.8.60 Sensitivity Matrix as a General Jacobian

Conceptually,

\[ \boxed{ \mathbf S = \frac{ \partial(\text{trajectory quantity}) }{ \partial(\text{uncertain or controllable quantity}) }. } \]

The STM is specifically

\[ \boxed{ \boldsymbol{\Phi} = \frac{ \partial\mathbf x(t) }{ \partial\mathbf x_0 }. } \]

Parameter sensitivity is

\[ \boxed{ \mathbf S_\theta = \frac{ \partial\mathbf x(t) }{ \partial\boldsymbol{\theta} }. } \]

Control sensitivity is

\[ \boxed{ \mathbf S_u = \frac{ \partial\mathbf x(t) }{ \partial\mathbf u }. } \]

4.3.8.61 Local Nature of Sensitivity

Sensitivity matrices are generally first-order quantities.

For

\[ \mathbf x(\boldsymbol{\eta}+\delta\boldsymbol{\eta}), \]

the approximation is

\[ \boxed{ \mathbf x( \boldsymbol{\eta}+\delta\boldsymbol{\eta} ) \approx \mathbf x(\boldsymbol{\eta}) + \mathbf S \delta\boldsymbol{\eta}. } \]

This is accurate only when the perturbation is sufficiently small and the system is locally smooth.

Engineering interpretation: A sensitivity derivative is trustworthy only within the neighbourhood in which the first-order approximation is valid. Validation over the actual uncertainty or control range is therefore essential before using it for requirements or targeting.

4.3.8.62 Higher-Order Effects

If perturbations are large,

\[ \mathbf S \delta\boldsymbol{\eta} \]

may not accurately predict trajectory changes.

Higher-order terms become important:

\[ \boxed{ \delta\mathbf x \approx \mathbf S \delta\boldsymbol{\eta} + \frac12 \mathcal H[ \delta\boldsymbol{\eta}, \delta\boldsymbol{\eta} ] + \cdots. } \]

Here \mathcal H represents second-order sensitivity information.

4.3.8.63 When First-Order STM Propagation Can Fail

First-order propagation may become inaccurate when:

Thus,

\[ \boxed{ \text{STM Accuracy} \text{ must be checked over the actual uncertainty range}. } \]

4.3.8.64 Numerical Computation of the STM

The STM can be computed in several ways:

Each approach has advantages and limitations.

4.3.8.65 Variational-Equation Method

The standard high-quality method is to integrate

\[ \boxed{ \dot{\boldsymbol{\Phi}} = \mathbf A(t) \boldsymbol{\Phi} } \]

alongside

\[ \dot{\mathbf x} = \mathbf f(\mathbf x,t). \]

Advantages include:

The main requirement is an accurate dynamic Jacobian.

4.3.8.66 Finite-Difference STM

The STM may also be approximated column by column.

For state component j,

\[ \boxed{ \boldsymbol{\Phi}_{:,j} \approx \frac{ \mathbf x( t; \mathbf x_0+\Delta x_j\mathbf e_j ) - \mathbf x( t; \mathbf x_0-\Delta x_j\mathbf e_j ) }{ 2\Delta x_j }. } \]

This is a central-difference approximation.

4.3.8.67 Interpretation of Finite-Difference STM

Each column is obtained by:

  1. perturbing one initial-state component,
  2. propagating forward,
  3. comparing the final state.

Thus,

\[ \boxed{ \text{One Initial-State Perturbation} \rightarrow \text{One STM Column}. } \]

For an n-state system, central differences require approximately 2n additional trajectory propagations.

4.3.8.68 Finite-Difference Step Size

The perturbation

\[ \Delta x_j \]

must be chosen carefully.

Too large:

\[ \boxed{ \text{Nonlinear truncation error dominates}. } \]

Too small:

\[ \boxed{ \text{Floating-point and integration noise dominate}. } \]

Thus,

\[ \boxed{ \text{Good FD Step} = \text{Balance Between Linearization and Numerical Precision}. } \]

Engineering interpretation: Finite-difference step-size studies are an important verification tool: a useful derivative should remain approximately stable over a reasonable range of perturbation sizes before roundoff or nonlinearity begins to dominate.

4.3.8.69 Forward Difference

A simpler approximation is

\[ \boxed{ \frac{ \partial\mathbf x(t) }{ \partial x_{0,j} } \approx \frac{ \mathbf x(t;\mathbf x_0+\Delta x_j\mathbf e_j) - \mathbf x(t;\mathbf x_0) }{ \Delta x_j }. } \]

It requires fewer propagations but is generally less accurate than central differences.

4.3.8.70 Complex-Step Differentiation

For compatible smooth numerical code, derivatives may be approximated using complex-step differentiation.

For scalar dependence,

\[ \boxed{ \frac{\partial f}{\partial x} \approx \frac{ \operatorname{Im} f(x+i\epsilon) }{ \epsilon }. } \]

This can provide very high derivative accuracy without subtractive cancellation.

However, all model functions must properly support complex arithmetic.

4.3.8.71 Automatic Differentiation

Automatic differentiation can generate derivatives of

\[ \mathbf f \]

and sometimes the complete numerical propagation.

Advantages include:

It can be especially useful for large nonlinear dynamics.

4.3.8.72 Analytical Jacobians

An analytical dynamics Jacobian provides explicit expressions for

\[ \mathbf A = \frac{\partial\mathbf f}{\partial\mathbf x}. \]

Advantages include:

But analytical derivatives must be carefully verified.

4.3.8.73 Jacobian Verification

A recommended procedure is:

\[ \boxed{ \mathbf A_{\text{analytic}} \quad\text{versus}\quad \mathbf A_{\text{finite difference}}. } \]

Similarly compare

\[ \boxed{ \boldsymbol{\Phi}_{\text{variational}} \quad\text{versus}\quad \boldsymbol{\Phi}_{\text{finite difference}}. } \]

Agreement should be checked over representative trajectories.

4.3.8.74 STM Verification with Perturbation Propagation

Choose a small test perturbation

\[ \delta\mathbf x_0. \]

The linear prediction is

\[ \boxed{ \delta\mathbf x_{\Phi}(t) = \boldsymbol{\Phi}(t,t_0) \delta\mathbf x_0. } \]

The nonlinear truth perturbation is

\[ \boxed{ \delta\mathbf x_{NL}(t) = \mathbf x( t;\mathbf x_0+\delta\mathbf x_0 ) - \mathbf x( t;\mathbf x_0 ). } \]

Compare them.

4.3.8.75 STM Verification Error

Define

\[ \boxed{ \mathbf e_\Phi = \delta\mathbf x_{NL} - \delta\mathbf x_\Phi. } \]

A normalized metric might be

\[ \boxed{ \epsilon_\Phi = \frac{ \| \mathbf e_\Phi \| }{ \| \delta\mathbf x_{NL} \| }. } \]

For sufficiently small perturbations, the error should decrease approximately with perturbation magnitude according to the neglected higher-order terms.

4.3.8.76 Composition Verification

Choose three times

\[ t_0

Verify

\[ \boxed{ \boldsymbol{\Phi}(t_2,t_0) \approx \boldsymbol{\Phi}(t_2,t_1) \boldsymbol{\Phi}(t_1,t_0). } \]

This is a useful implementation check.

4.3.8.77 Inverse Verification

Check

\[ \boxed{ \boldsymbol{\Phi}(t,t_0) \boldsymbol{\Phi}(t_0,t) \approx \mathbf I. } \]

Large deviations may indicate:

4.3.8.78 Initial-Condition Verification

At

\[ t=t_0, \]

verify

\[ \boxed{ \boldsymbol{\Phi}(t_0,t_0) = \mathbf I. } \]

This seems simple, but incorrect initialization is a common implementation mistake.

4.3.8.79 Numerical Integration Tolerance

Because the STM can contain rapidly growing or shrinking elements, its integration may be more sensitive than state integration.

A practical test is to tighten the integrator tolerance and confirm that:

change negligibly.

Thus,

\[ \boxed{ \text{Converged State Integration} \not\Rightarrow \text{Automatically Converged Sensitivity Integration}. } \]

4.3.8.80 Conditioning of the STM

The STM itself may become poorly conditioned.

Define

\[ \boxed{ \kappa(\boldsymbol{\Phi}) = \| \boldsymbol{\Phi} \| \| \boldsymbol{\Phi}^{-1} \|. } \]

A large condition number means some perturbation directions grow or shrink much more strongly than others.

This can affect:

Engineering interpretation: Poor conditioning does not necessarily mean the dynamics are numerically wrong. It can indicate that the chosen coordinates contain directions with very different amplification, which is why scaling and singular-value analysis are important.

4.3.8.81 Singular Values of the STM

Use SVD:

\[ \boxed{ \boldsymbol{\Phi} = \mathbf U \boldsymbol{\Sigma} \mathbf V^T. } \]

The largest singular value identifies the perturbation direction with greatest first-order amplification.

The smallest identifies a strongly contracted direction.

Thus singular values reveal dynamic sensitivity directions.

4.3.8.82 Perturbation Amplification

If

\[ \sigma_{\max}( \boldsymbol{\Phi} ) \gg1, \]

some initial-state errors can grow strongly.

If

\[ \sigma_{\min} \ll1, \]

some perturbations contract strongly.

Therefore,

\[ \boxed{ \text{STM Singular Values} \rightarrow \text{Local Error Growth and Contraction}. } \]

4.3.8.83 Local Stability Interpretation

The STM contains local information about trajectory stability.

If small perturbations grow rapidly, the nominal trajectory is locally sensitive.

However, finite-time STM behaviour should not automatically be interpreted as asymptotic stability.

That requires additional dynamical-systems analysis.

4.3.8.84 STM Determinant

The determinant

\[ \det\boldsymbol{\Phi} \]

measures local volume expansion or contraction of perturbation space.

For certain Hamiltonian orbital dynamics, special structure may exist.

But interpretation depends strongly on the underlying coordinate system and model.

4.3.8.85 State Scaling and STM Scaling

Suppose position is measured in kilometres while velocity is measured in kilometres per second, and other states have very different magnitudes.

Then raw STM elements may have very different numerical scales.

Introduce

\[ \mathbf x = \mathbf S_x \tilde{\mathbf x}. \]

The scaled STM becomes

\[ \boxed{ \tilde{\boldsymbol{\Phi}} = \mathbf S_x^{-1} \boldsymbol{\Phi} \mathbf S_x. } \]

Scaling can improve numerical conditioning and interpretation.

4.3.8.86 Units of STM Elements

STM elements are not always dimensionless.

For example,

\[ \boldsymbol{\Phi}_{rv} = \frac{ \partial\mathbf r }{ \partial\mathbf v_0 } \]

has units of time.

Therefore comparing raw STM elements without considering units can be misleading.

Scaled or nondimensional sensitivity measures may be more meaningful.

4.3.8.87 Dimensionless Sensitivities

A normalized sensitivity can be defined as

\[ \boxed{ S_{ij}^{rel} = \frac{x_j} {y_i} \frac{ \partial y_i }{ \partial x_j }. } \]

This measures fractional output change per fractional input change.

It can help compare sensitivity between variables with different physical units.

4.3.8.88 Sensitivity and Observability

Measurement observability depends on both:

For measurement k,

\[ \boxed{ \mathbf G_k = \mathbf H_k \boldsymbol{\Phi}_{k,0}. } \]

Thus a state that is not directly measured may become observable because dynamics transfer its influence into measured quantities.

4.3.8.89 Dynamic Observability Matrix

For a discrete linear system,

\[ \mathbf x_{k+1} = \mathbf F\mathbf x_k, \]

\[ \mathbf z_k = \mathbf H\mathbf x_k, \]

the observability matrix is

\[ \boxed{ \mathcal O = \begin{bmatrix} \mathbf H\\ \mathbf H\mathbf F\\ \mathbf H\mathbf F^2\\ \vdots\\ \mathbf H\mathbf F^{n-1} \end{bmatrix}. } \]

Each row block is essentially a measurement sensitivity propagated to the reference state.

4.3.8.90 Connection to the Information Matrix

The information matrix in batch estimation is

\[ \boxed{ \mathbf N = \sum_k \mathbf G_k^T \mathbf R_k^{-1} \mathbf G_k. } \]

Since

\[ \mathbf G_k = \mathbf H_k \boldsymbol{\Phi}_{k,0}, \]

we obtain

\[ \boxed{ \mathbf N = \sum_k \boldsymbol{\Phi}_{k,0}^T \mathbf H_k^T \mathbf R_k^{-1} \mathbf H_k \boldsymbol{\Phi}_{k,0}. } \]

This equation explicitly shows how dynamics, measurement geometry, and sensor uncertainty combine to determine estimation information.

Engineering interpretation: This equation unifies dynamics and sensing. A measurement can be accurate yet provide little information about a state direction if \(\mathbf H_k\boldsymbol{\Phi}_{k,0}\) is weak in that direction.

4.3.8.91 Sensitivity and Identifiability

For parameter \theta, if

\[ \frac{\partial h} {\partial\theta} \approx0, \]

the measurements contain little information about it.

If two parameter-sensitivity columns are nearly dependent,

\[ \mathbf g_{\theta_1} \approx c \mathbf g_{\theta_2}, \]

then the parameters may be difficult to distinguish.

Thus sensitivity matrices provide direct insight into identifiability.

4.3.8.92 Parameter Correlation

Suppose two parameters create similar trajectory perturbations:

\[ \mathbf S_{\theta_1} \approx c \mathbf S_{\theta_2}. \]

Then their measurement signatures will also be similar.

This can produce:

Therefore,

\[ \boxed{ \text{Similar Sensitivity Signatures} \rightarrow \text{Weak Parameter Separation}. } \]

4.3.8.93 Sensitivity and Experiment Design

Sensitivity analysis can help decide:

Thus,

\[ \boxed{ \text{Sensitivity Analysis} \rightarrow \text{Measurement and Maneuver Design}. } \]

This turns estimation theory into mission-design information.

4.3.8.94 Sensitivity and Sensor Selection

Suppose several sensors measure different functions.

Their Jacobians are

\[ \mathbf H_1, \mathbf H_2, \ldots \]

and dynamic sensitivities are

\[ \mathbf H_i\boldsymbol{\Phi}. \]

The sensor providing the smallest raw noise is not necessarily the most informative for every state direction.

Information depends on both sensitivity and covariance.

4.3.8.95 Sensitivity and Guidance

Guidance requires understanding how control changes future conditions.

For terminal quantity

\[ \mathbf y_f = g( \mathbf x_f ), \]

control sensitivity is

\[ \boxed{ \frac{ \partial\mathbf y_f }{ \partial\mathbf u } = \frac{ \partial g }{ \partial\mathbf x_f } \frac{ \partial\mathbf x_f }{ \partial\mathbf u }. } \]

This matrix can be used in iterative guidance or targeting.

4.3.8.96 Sensitivity and Trajectory Optimization

Trajectory optimization frequently requires gradients such as

\[ \boxed{ \frac{\partial J}{\partial\mathbf u}, \qquad \frac{\partial\mathbf c}{\partial\mathbf u}. } \]

These can be built from state and control sensitivities.

Thus sensitivity propagation is fundamental not only to estimation but also to optimization.

4.3.8.97 Forward Sensitivity Method

The direct variational approach propagates one sensitivity column for every parameter.

For p parameters,

\[ \mathbf S_\theta \in \mathbb R^{n\times p}. \]

This is efficient when p is relatively small.

It is known as forward sensitivity propagation.

4.3.8.98 Adjoint Sensitivity Concept

If there are many parameters but only a few scalar outputs or cost functions, an adjoint method can be more efficient.

Instead of propagating sensitivity with respect to every parameter individually, an adjoint variable is integrated and used to construct gradients.

Conceptually,

\[ \boxed{ \text{Forward Sensitivity} \rightarrow \text{Cost Scales with Number of Parameters} } \]

whereas

\[ \boxed{ \text{Adjoint Sensitivity} \rightarrow \text{Often Efficient for Few Outputs and Many Parameters}. } \]

4.3.8.99 Why Adjoint Methods Matter

Suppose an optimizer has hundreds of control parameters but only one cost

\[ J. \]

Forward sensitivity may require hundreds of derivative propagations.

Adjoint methods can often obtain

\[ \frac{\partial J}{\partial\boldsymbol{\theta}} \]

with much less dependence on parameter dimension.

This becomes important in large trajectory-optimization problems.

4.3.8.100 Discontinuities and the STM

If the trajectory contains a discontinuity, the STM cannot simply be propagated through it as though dynamics were smooth.

Examples include:

A jump relation must be applied.

Engineering interpretation: For hybrid systems, sensitivity propagation must respect event boundaries. Treating a discontinuity as a smooth segment can corrupt both the STM and any covariance or targeting result derived from it.

4.3.8.101 Jump Matrix

Suppose

\[ \mathbf x^{+} = \mathbf g( \mathbf x^{-} ). \]

Then the perturbation satisfies

\[ \boxed{ \delta\mathbf x^{+} = \mathbf J_g \delta\mathbf x^{-} } \]

where

\[ \boxed{ \mathbf J_g = \frac{\partial\mathbf g} {\partial\mathbf x^{-}}. } \]

Thus the STM across the event becomes

\[ \boxed{ \boldsymbol{\Phi}^{+} = \mathbf J_g \boldsymbol{\Phi}^{-}. } \]

4.3.8.102 Impulsive Maneuver Jump Matrix

For a simple known velocity impulse,

\[ \mathbf r^{+} = \mathbf r^{-}, \]

\[ \mathbf v^{+} = \mathbf v^{-} + \Delta\mathbf v. \]

If \Delta\mathbf v is known and independent of state, the state jump Jacobian is simply

\[ \boxed{ \mathbf J_g = \mathbf I. } \]

But if the maneuver depends on the state or attitude, additional derivatives appear.

4.3.8.103 Sensitivity to Event Time

When an event time changes, the system follows one dynamics model for a slightly different duration.

Event-time sensitivity therefore requires special treatment.

A small timing perturbation can affect the state both through:

This is especially important in burn timing and mode-switch problems.

4.3.8.104 Hybrid Systems

A spacecraft may alternate among dynamic modes:

\[ f^{(1)}, \quad f^{(2)}, \quad \ldots \]

The complete sensitivity propagation becomes

\[ \boxed{ \boldsymbol{\Phi} = \boldsymbol{\Phi}_m \mathbf J_{m-1} \boldsymbol{\Phi}_{m-1} \cdots \mathbf J_1 \boldsymbol{\Phi}_1. } \]

This combines continuous evolution with discrete jump mappings.

4.3.8.105 Attitude-State Sensitivity

For attitude dynamics, the state may contain

\[ \mathbf q, \qquad \boldsymbol{\omega}. \]

Because quaternion attitude lies on a constrained manifold,

\[ \|\mathbf q\|=1, \]

a naive Euclidean STM in four quaternion components may be undesirable.

A local attitude-error representation is often preferable.

4.3.8.106 Multiplicative Attitude Error

Represent attitude error through a small rotation vector

\[ \delta\boldsymbol{\theta}. \]

Then the error state may be

\[ \boxed{ \delta\mathbf x = \begin{bmatrix} \delta\boldsymbol{\theta}\\ \delta\boldsymbol{\omega}\\ \cdots \end{bmatrix}. } \]

The STM then maps physically meaningful local attitude errors.

This is consistent with error-state and multiplicative filtering approaches.

4.3.8.107 Coupled Translational and Attitude Sensitivity

For a fully coupled spacecraft,

\[ \mathbf x = \begin{bmatrix} \mathbf r\\ \mathbf v\\ \mathbf q\\ \boldsymbol{\omega}\\ \cdots \end{bmatrix}. \]

Then

\[ \mathbf A \]

may contain off-diagonal terms coupling translation and attitude.

Examples include:

Thus full 6-DOF sensitivity may require a coupled STM.

4.3.8.108 Sensor-to-State Sensitivity

A sensor may produce

\[ \mathbf z = h( \mathbf x ). \]

The measurement Jacobian

\[ \mathbf H = \frac{\partial h}{\partial\mathbf x} \]

converts state perturbation to measurement perturbation:

\[ \boxed{ \delta\mathbf z = \mathbf H \delta\mathbf x. } \]

With an STM,

\[ \boxed{ \delta\mathbf z_k = \mathbf H_k \boldsymbol{\Phi}_{k,0} \delta\mathbf x_0. } \]

This is the complete state-to-measurement sensitivity chain.

4.3.8.109 General Range Sensitivity

For range

\[ \rho = \|\mathbf r\|, \]

the measurement Jacobian is

\[ \boxed{ \frac{\partial\rho}{\partial\mathbf r} = \frac{\mathbf r^T}{\|\mathbf r\|}. } \]

Sensitivity to the reference state becomes

\[ \boxed{ \frac{\partial\rho_k} {\partial\mathbf x_0} = \mathbf H_{\rho,k} \boldsymbol{\Phi}_{k,0}. } \]

4.3.8.110 General Range-Rate Sensitivity

Range rate is

\[ \dot\rho = \frac{ \mathbf r^T\mathbf v }{ \rho }. \]

Its Jacobian contains both position and velocity terms.

Therefore range-rate measurements often provide complementary information to range measurements.

When propagated through the STM, this can improve reference-state observability.

4.3.8.111 Line-of-Sight Sensitivity

For

\[ \mathbf u = \frac{\mathbf r}{r}, \]

the derivative with respect to \mathbf r is

\[ \boxed{ \frac{\partial\mathbf u} {\partial\mathbf r} = \frac{1}{r} \left( \mathbf I - \mathbf u\mathbf u^T \right). } \]

This matrix projects perturbations into directions perpendicular to the line of sight.

Thus an angular measurement is locally insensitive to motion directly along the line of sight.

4.3.8.112 Why Changing Geometry Helps

At one instant, LOS measurements may poorly constrain range.

But as geometry evolves,

\[ \mathbf H_k \]

changes.

Combined with

\[ \boldsymbol{\Phi}_{k,0}, \]

the stacked design matrix can become full rank.

Therefore,

\[ \boxed{ \text{Changing Geometry} + \text{Dynamics} \rightarrow \text{Improved Observability}. } \]

4.3.8.113 Sensitivity to Bias and Calibration States

If a measurement contains calibration parameters,

\[ \mathbf z = h( \mathbf x, \mathbf b, \mathbf s, \boldsymbol{\alpha} ), \]

the complete Jacobian may include

\[ \boxed{ \mathbf H = \begin{bmatrix} \frac{\partial h}{\partial\mathbf x} & \frac{\partial h}{\partial\mathbf b} & \frac{\partial h}{\partial\mathbf s} & \frac{\partial h}{\partial\boldsymbol{\alpha}} \end{bmatrix}. } \]

This is essential for joint state and calibration estimation.

4.3.8.114 Sensitivity and Monte Carlo Analysis

Sensitivity analysis provides a first-order alternative to brute-force Monte Carlo.

If uncertain parameters have covariance

\[ \mathbf P_\eta, \]

and

\[ \delta\mathbf y = \mathbf S_\eta \delta\boldsymbol{\eta}, \]

then approximately

\[ \boxed{ \mathbf P_y = \mathbf S_\eta \mathbf P_\eta \mathbf S_\eta^T. } \]

This predicts uncertainty propagation efficiently.

Engineering interpretation: This first-order covariance mapping is one of the fastest ways to estimate uncertainty contributions, but Monte Carlo remains important for checking whether the local linear approximation remains valid over the actual uncertainty distribution.

4.3.8.115 First-Order Covariance versus Monte Carlo

First-order covariance propagation is fast but local.

Monte Carlo captures more nonlinear effects.

Therefore,

\[ \boxed{ \text{Sensitivity/Covariance} \rightarrow \text{Fast Local Approximation} } \]

while

\[ \boxed{ \text{Monte Carlo} \rightarrow \text{More General Nonlinear Statistical Validation}. } \]

They complement rather than replace one another.

4.3.8.116 Sensitivity Ranking

One may rank parameters by how strongly they affect a chosen output.

For scalar output y,

\[ \boxed{ S_j = \left| \frac{\partial y} {\partial\theta_j} \right|. } \]

However, raw derivatives may be misleading when parameters have different units or uncertainty ranges.

Normalized or uncertainty-weighted sensitivities are often preferable.

4.3.8.117 Uncertainty-Weighted Sensitivity

If parameter \theta_j has standard deviation \sigma_{\theta_j}, define

\[ \boxed{ S_j^{unc} = \left| \frac{\partial y} {\partial\theta_j} \right| \sigma_{\theta_j}. } \]

This estimates the typical output uncertainty contributed by that parameter.

It is often more useful than the derivative alone.

4.3.8.118 Sensitivity Does Not Equal Importance Automatically

A parameter can have:

Therefore physical importance depends on both sensitivity and expected parameter variation.

Thus,

\[ \boxed{ \text{Impact} \approx \text{Sensitivity} \times \text{Uncertainty Scale}. } \]

4.3.8.119 Time History of Sensitivity

Sensitivity should often be examined as a function of time:

\[ \boxed{ \mathbf S(t). } \]

A parameter may have little influence early in the trajectory but strong influence later.

This is useful for:

4.3.8.120 Sensitivity Peaks

If

\[ \left\| \frac{\partial\mathbf y(t)} {\partial\theta} \right\| \]

has a strong peak at certain times, measurements near those epochs may be especially informative about \theta.

Thus,

\[ \boxed{ \text{Sensitivity-Time History} \rightarrow \text{Potential Measurement Scheduling Information}. } \]

4.3.8.121 Sensitivity Heat Maps

For many outputs and parameters, construct

\[ \boxed{ \mathbf S_{ij} = \frac{ \partial y_i }{ \partial\theta_j }. } \]

A heat map can reveal:

This is especially useful before a large estimation study.

4.3.8.122 Sensitivity and Requirement Allocation

Suppose final-position uncertainty must satisfy

\[ \sigma_{r_f} < \sigma_{req}. \]

Sensitivity can identify how much each uncertainty source contributes.

For parameter covariance \mathbf P_\theta,

\[ \boxed{ \mathbf P_{r_f,\theta} = \mathbf S_{\theta} \mathbf P_\theta \mathbf S_{\theta}^T. } \]

This can support allocation of sensor, actuator, or model requirements.

4.3.8.123 Sensitivity and Navigation Requirement Design

Suppose initial navigation error is

\[ \mathbf P_0. \]

Terminal covariance is approximately

\[ \boxed{ \mathbf P_f = \boldsymbol{\Phi}_{f,0} \mathbf P_0 \boldsymbol{\Phi}_{f,0}^T. } \]

If a terminal accuracy requirement is imposed, one can work backward conceptually to determine acceptable initial-state uncertainty.

This connects dynamics directly to navigation requirements.

Engineering interpretation: Requirement design can therefore be performed in both directions: propagate an assumed navigation covariance forward to check a terminal requirement, or use terminal sensitivity to identify which initial uncertainty components require tighter allocation.

4.3.8.124 Backward Requirement Mapping

Given a terminal sensitivity direction \mathbf c, terminal variance is

\[ \boxed{ \sigma_c^2 = \mathbf c^T \mathbf P_f \mathbf c. } \]

Using

\[ \mathbf P_f = \boldsymbol{\Phi} \mathbf P_0 \boldsymbol{\Phi}^T, \]

we obtain

\[ \boxed{ \sigma_c^2 = ( \boldsymbol{\Phi}^T\mathbf c )^T \mathbf P_0 ( \boldsymbol{\Phi}^T\mathbf c ). } \]

This shows which initial-state directions drive the terminal requirement.

4.3.8.125 Sensitivity in General RPO Navigation Requirements

In a general RPO context, suppose final relative position accuracy must satisfy a tolerance.

The relative-motion STM can be used to determine how uncertainty in

\[ \delta\mathbf r_0, \qquad \delta\mathbf v_0 \]

maps into final relative-position uncertainty.

This helps identify whether the navigation requirement should focus more strongly on:

Again, this is a general method rather than a mission-specific result.

4.3.8.126 Sensitivity to Process Parameters

Suppose trajectory acceleration is

\[ \mathbf a = \mathbf a( \mathbf x, \theta ). \]

Then

\[ \boxed{ \frac{\partial\mathbf a} {\partial\theta} } \]

drives the parameter-sensitivity equation.

If that derivative remains almost zero throughout the trajectory, the parameter is unlikely to be observable from the trajectory.

4.3.8.127 Example: Drag-Coefficient Sensitivity

For a drag model

\[ \mathbf a_D = -\frac12 \rho C_D \frac{A}{m} v_{rel}^2 \hat{\mathbf v}_{rel}, \]

the derivative with respect to C_D is approximately

\[ \boxed{ \frac{\partial\mathbf a_D} {\partial C_D} = \frac{\mathbf a_D}{C_D} } \]

when other quantities are held fixed.

Thus atmospheric density and velocity determine when C_D becomes dynamically observable.

4.3.8.128 Example: Ballistic-Coefficient Sensitivity

If drag acceleration is parameterized through ballistic coefficient

\[ B = \frac{m}{C_DA}, \]

then

\[ \mathbf a_D \propto -\frac{1}{B}. \]

Therefore

\[ \boxed{ \frac{\partial\mathbf a_D} {\partial B} \propto \frac{1}{B^2}. } \]

The trajectory sensitivity to B is obtained by integrating the associated parameter variational equations.

4.3.8.129 Example: Gravity-Parameter Sensitivity

For

\[ \mathbf a = -\mu \frac{\mathbf r}{r^3}, \]

the acceleration sensitivity to \mu is

\[ \boxed{ \frac{\partial\mathbf a} {\partial\mu} = - \frac{\mathbf r}{r^3}. } \]

This drives

\[ \frac{\partial\mathbf x}{\partial\mu}. \]

Thus long orbital arcs can accumulate measurable sensitivity to gravitational parameters.

4.3.8.130 STM and Multiple Shooting

In trajectory optimization, a trajectory may be divided into segments.

Each segment has its own state and STM.

Continuity constraints are imposed:

\[ \boxed{ \mathbf x_{i+1}^{start} - \phi_i( \mathbf x_i^{start} ) = \mathbf0. } \]

The Jacobian of this continuity condition contains the segment STM.

Therefore STM propagation is central to multiple-shooting methods.

4.3.8.131 STM and Differential Correction

In differential correction, the residual correction equation contains the dynamic Jacobian

\[ \mathbf H_k \boldsymbol{\Phi}_{k,0}. \]

Thus,

\[ \boxed{ \text{Differential Correction} = \text{Measurement Jacobian} \times \text{Dynamic Sensitivity}. } \]

This connects Sections 4.3.3, 4.3.5, and 4.3.8 directly.

4.3.8.132 STM and Smoother Cross-Covariance

The smoother uses

\[ \boxed{ \mathbf P_{k,k+1|k} = \mathbf P_{k|k} \mathbf F_k^T. } \]

Since

\[ \mathbf F_k = \boldsymbol{\Phi}(t_{k+1},t_k), \]

the same dynamic sensitivity underlying batch estimation also determines the cross-covariance needed for backward smoothing.

4.3.8.133 STM and Reachability

Control sensitivity is closely related to controllability or reachability.

For

\[ \delta\dot{\mathbf x} = \mathbf A\delta\mathbf x + \mathbf B\delta\mathbf u, \]

the control effect is

\[ \boxed{ \delta\mathbf x(t) = \int \boldsymbol{\Phi}(t,\tau) \mathbf B(\tau) \delta\mathbf u(\tau) d\tau. } \]

This shows how control authority is transported through system dynamics.

4.3.8.134 Controllability Gramian

For a linear system, a controllability Gramian may be defined as

\[ \boxed{ \mathbf W_c = \int_{t_0}^{t_f} \boldsymbol{\Phi}(t_f,\tau) \mathbf B(\tau) \mathbf B(\tau)^T \boldsymbol{\Phi}(t_f,\tau)^T d\tau. } \]

It describes how strongly the control can influence different state directions.

Thus STM theory connects directly to control-system properties.

4.3.8.135 Observability Gramian

Similarly, an observability Gramian can be written as

\[ \boxed{ \mathbf W_o = \int_{t_0}^{t_f} \boldsymbol{\Phi}(t,t_0)^T \mathbf H(t)^T \mathbf H(t) \boldsymbol{\Phi}(t,t_0) dt. } \]

With measurement weighting,

\[ \mathbf H^T\mathbf R^{-1}\mathbf H \]

appears.

This closely resembles the estimation information matrix.

4.3.8.136 Information, Observability, and Sensitivity

The following concepts are therefore deeply connected:

\[ \boxed{ \text{STM} \rightarrow \text{Dynamic Sensitivity} \rightarrow \text{Measurement Sensitivity} \rightarrow \text{Information} \rightarrow \text{Observability}. } \]

They are not separate topics.

They are different views of how perturbations propagate and become measurable.

4.3.8.137 Time-Varying Sensitivity and Mission Phases

A spacecraft trajectory may contain:

Sensitivity may change dramatically between these phases.

Therefore it can be useful to analyse

\[ \boldsymbol{\Phi}(t,t_0) \]

and

\[ \mathbf S_\theta(t) \]

phase by phase rather than only at final time.

4.3.8.138 Phase-to-Phase STM

If a trajectory has several phases,

\[ t_0 \rightarrow t_1 \rightarrow t_2 \rightarrow t_3, \]

then

\[ \boxed{ \boldsymbol{\Phi}_{3,0} = \boldsymbol{\Phi}_{3,2} \boldsymbol{\Phi}_{2,1} \boldsymbol{\Phi}_{1,0}. } \]

This is valuable for tracing where sensitivity grows.

4.3.8.139 Sensitivity Contribution by Phase

Suppose total terminal sensitivity is large.

By examining each phase STM, one may identify whether the amplification originates during:

This supports design improvement.

4.3.8.140 STM Reset at Reference Epoch

When changing the estimation reference epoch to t_r,

\[ \boxed{ \boldsymbol{\Phi}(t_r,t_r)=\mathbf I. } \]

Sensitivities must then be redefined relative to the new reference state.

Care is required when comparing covariances or normal matrices across reference epochs.

4.3.8.141 Reference Epoch and Conditioning

Choosing a reference epoch near the middle of an observation arc can sometimes improve numerical conditioning because sensitivities need not grow over the entire interval in only one time direction.

However, the optimal choice depends on the estimation formulation.

Reference epoch is therefore partly a numerical-design decision.

4.3.8.142 Forward versus Backward STM Propagation

A forward STM maps

\[ t_0\rightarrow t. \]

A backward STM maps

\[ t\rightarrow t_0. \]

Since

\[ \boldsymbol{\Phi}(t_0,t) = \boldsymbol{\Phi}(t,t_0)^{-1}, \]

one may obtain the backward map from the inverse, but direct backward integration may sometimes be preferable numerically.

4.3.8.143 Numerical Danger of Explicit Inversion

If the STM is ill conditioned,

\[ \boldsymbol{\Phi}^{-1} \]

may be numerically unstable.

Therefore one should often solve linear systems instead of explicitly forming the inverse.

This is the same numerical principle encountered in least squares and covariance propagation.

4.3.8.144 Sparsity of Large Sensitivity Systems

For large augmented systems, many derivatives may be zero.

Examples include parameters affecting only particular forces or sensors.

The matrices

\[ \mathbf A, \quad \mathbf E, \quad \mathbf H \]

may therefore be sparse.

Exploiting sparsity reduces computational cost.

4.3.8.145 Memory Requirements

An n\times n STM contains

\[ n^2 \]

elements.

For a 6-state orbit,

\[ 36 \]

elements are modest.

But a large augmented navigation state can produce a much larger STM.

Therefore one may choose between:

4.3.8.146 Selected Sensitivity Propagation

If only sensitivity to a few parameters is needed, there is no requirement to propagate a full augmented STM.

Instead propagate

\[ \boxed{ \mathbf S_\theta } \]

only for selected parameters.

This can reduce computational load significantly.

4.3.8.147 Coordinate-Frame Dependence

STM values depend on the coordinates used to represent the state.

For example, an STM expressed in:

will look different.

Therefore,

\[ \boxed{ \text{STM Is Coordinate-Representation Dependent}. } \]

The physical perturbation mapping is equivalent only after consistent coordinate transformations.

Engineering interpretation: Because an STM is a derivative with respect to coordinates, its numerical values change when the coordinates change. Physical conclusions should therefore be based on consistently transformed perturbations, covariances, and sensitivities.

4.3.8.148 STM Transformation Between Coordinate Systems

Suppose

\[ \mathbf y = g(\mathbf x). \]

Define coordinate Jacobian

\[ \mathbf T = \frac{\partial\mathbf y}{\partial\mathbf x}. \]

Then approximately,

\[ \boxed{ \boldsymbol{\Phi}_y(t,t_0) = \mathbf T(t) \boldsymbol{\Phi}_x(t,t_0) \mathbf T(t_0)^{-1}. } \]

This allows STM transformation between coordinate representations.

4.3.8.149 Covariance Transformation and STM

Similarly, coordinate covariance transforms as

\[ \boxed{ \mathbf P_y = \mathbf T \mathbf P_x \mathbf T^T. } \]

Thus coordinate transformation Jacobians and state transition matrices follow closely related rules.

4.3.8.150 STM in Orbital Elements

If orbital elements are used,

\[ \mathbf e = [ a,e,i,\Omega,\omega,\nu ]^T, \]

one may define

\[ \boxed{ \boldsymbol{\Phi}_e = \frac{ \partial\mathbf e(t) }{ \partial\mathbf e(t_0) }. } \]

However, singularities in classical orbital elements can make sensitivity analysis difficult near special orbital configurations.

Nonsingular element sets may be preferable.

4.3.8.151 Coordinate Singularities

Sensitivity matrices can become numerically misleading near coordinate singularities even when physical dynamics remain regular.

Examples include classical orbital-element singularities.

Therefore,

\[ \boxed{ \text{Large Sensitivity} \text{ may reflect} \text{coordinate singularity rather than physical instability}. } \]

Interpretation should always consider the state representation.

4.3.8.152 Nonlinear Error Geometry

Attitude and some orbital representations live on manifolds rather than unconstrained Euclidean space.

In such cases, sensitivity should preferably be defined in local tangent-space coordinates.

This avoids interpreting invalid Euclidean perturbations as physical states.

4.3.8.153 Sensitivity to Time

Time itself may be treated as an uncertain parameter.

Define

\[ \boxed{ \mathbf S_t = \frac{ \partial\mathbf x }{ \partial t }. } \]

Since

\[ \frac{d\mathbf x}{dt} = \mathbf f, \]

local time sensitivity is closely related to the state derivative.

This becomes important in timestamp-error estimation.

4.3.8.154 Time-Bias Measurement Sensitivity

If a measurement has timing bias

\[ \Delta t, \]

then approximately

\[ h[ \mathbf x(t+\Delta t) ] \approx h[ \mathbf x(t) ] + \frac{\partial h}{\partial\mathbf x} \dot{\mathbf x} \Delta t. \]

Therefore,

\[ \boxed{ \frac{\partial h}{\partial\Delta t} \approx \mathbf H \dot{\mathbf x}. } \]

This gives a simple timing-sensitivity relation.

4.3.8.155 Sensitivity to Frame Misalignment

Suppose a sensor frame is misaligned by small angle

\[ \delta\boldsymbol{\alpha}. \]

A measurement vector transformed by a rotation matrix will have sensitivity

\[ \boxed{ \frac{\partial\mathbf z} {\partial\boldsymbol{\alpha}}. } \]

Such derivatives can be included in the estimation design matrix for alignment calibration.

4.3.8.156 Sensitivity to Scale Factors

Suppose a sensor reports

\[ \mathbf z = (\mathbf I+\mathbf S) \mathbf z_{true} + \mathbf b. \]

Sensitivity to scale-factor parameter s_i can be obtained analytically.

Thus scale-factor calibration can be combined with trajectory sensitivity if sufficient measurement excitation exists.

4.3.8.157 Local Sensitivity versus Global Sensitivity

STM-based analysis is a local sensitivity method.

It asks:

\[ \boxed{ \text{What happens for infinitesimal or small perturbations near this nominal trajectory?} } \]

Global sensitivity methods instead explore large parameter ranges and nonlinear interactions.

Examples include:

Both serve different purposes.

4.3.8.158 Why Local Sensitivity Is Still Extremely Valuable

Even though local, STM methods are:

Therefore they remain foundational in spacecraft dynamics.

4.3.8.159 Practical STM Development Workflow

A disciplined STM implementation workflow is:

  1. Define the state vector.
  2. Define the nonlinear dynamics.
  3. Define coordinate frames and units.
  4. Derive

\[ \mathbf A = \frac{\partial\mathbf f}{\partial\mathbf x}. \]

  1. Check dimensions.
  2. Initialize

\[ \boldsymbol{\Phi}_0=\mathbf I. \]

  1. Integrate state and STM together.
  2. Check

\[ \boldsymbol{\Phi}(t_0,t_0)=\mathbf I. \]

  1. Compare selected STM columns with finite differences.
  2. Test several perturbation magnitudes.
  3. Check composition property.
  4. Check inverse property when appropriate.
  5. Tighten integrator tolerance.
  6. Check sensitivity convergence.
  7. Examine STM singular values.
  8. Examine condition number.
  9. Apply state scaling if required.
  10. Propagate covariance.
  11. Compare covariance with Monte Carlo.
  12. Validate over representative trajectory regimes.

Engineering interpretation: The verification steps are not optional bookkeeping. An STM can look smooth and numerically well behaved while still containing a sign, ordering, unit, or frame error that later corrupts estimation and covariance results.

4.3.8.160 Practical Parameter-Sensitivity Workflow

For parameter vector

\[ \boldsymbol{\theta}, \]

use:

  1. Define the parameters.
  2. Derive

\[ \mathbf E = \frac{\partial\mathbf f} {\partial\boldsymbol{\theta}}. \]

  1. Initialize

\[ \mathbf S_\theta(t_0). \]

  1. Integrate

\[ \dot{\mathbf S}_\theta = \mathbf A \mathbf S_\theta + \mathbf E. \]

  1. Compare with finite-difference parameter perturbations.
  2. Compute measurement sensitivity.
  3. Examine rank.
  4. Examine parameter correlations.
  5. Examine sensitivity versus time.
  6. Determine whether the measurement arc can identify the parameter.
  7. Compare predicted uncertainty with Monte Carlo.
  8. Check physical plausibility.

4.3.8.161 Practical Measurement-Sensitivity Workflow

For each observation epoch k:

  1. Propagate the state.
  2. Propagate the STM.
  3. Compute predicted measurement.
  4. Compute

\[ \mathbf H_k = \frac{\partial h_k}{\partial\mathbf x_k}. \]

  1. Form

\[ \boxed{ \mathbf G_k = \mathbf H_k \boldsymbol{\Phi}_{k,0}. } \]

  1. Add direct parameter terms.
  2. Stack the design matrix.
  3. Weight using

\[ \mathbf R_k^{-1}. \]

  1. Examine rank.
  2. Examine singular values.
  3. Examine information contributions.
  4. Use the result in estimation or observability analysis.

4.3.8.162 Complete Sensitivity Architecture

\[ \boxed{ \begin{array}{c} \text{Nominal State and Parameters} \\ \downarrow\\ \dot{\mathbf x} = \mathbf f( \mathbf x,\boldsymbol{\theta},\mathbf u,t ) \\ \downarrow\\ \text{Dynamics Jacobians} \\ \mathbf A = \frac{\partial\mathbf f}{\partial\mathbf x}, \qquad \mathbf B = \frac{\partial\mathbf f}{\partial\mathbf u}, \qquad \mathbf E = \frac{\partial\mathbf f}{\partial\boldsymbol{\theta}} \\ \downarrow\\ \text{Variational Equations} \\ \dot{\boldsymbol{\Phi}} = \mathbf A\boldsymbol{\Phi} \\ \dot{\mathbf S}_\theta = \mathbf A\mathbf S_\theta+\mathbf E \\ \downarrow\\ \text{Trajectory Sensitivity} \\ \downarrow\\ \text{Measurement Jacobian} \\ \mathbf H = \frac{\partial h}{\partial\mathbf x} \\ \downarrow\\ \text{Dynamic Measurement Sensitivity} \\ \mathbf G = \mathbf H\boldsymbol{\Phi} \\ \downarrow\\ \text{Estimation / Covariance / Guidance / Targeting / Observability} \end{array} } \]

4.3.8.163 Common Implementation Errors

Common problems include:

These errors can produce apparently reasonable trajectories while corrupting covariance and estimation results.

4.3.8.164 Debugging Checklist

If the STM appears wrong, check:

  1. Is

\[ \Phi(t_0,t_0)=I? \]

  1. Is the state ordering identical in \mathbf x, \mathbf A, and \Phi?
  2. Are units consistent?
  3. Are frames consistent?
  4. Does one STM column match a finite-difference perturbation?
  5. Does central differencing agree with forward differencing?
  6. Does changing perturbation size produce convergence?
  7. Does tightening integrator tolerance change the result?
  8. Does composition hold?
  9. Does the inverse property hold?
  10. Are maneuver jumps handled?
  11. Are parameter sensitivities initialized correctly?
  12. Is the dynamics Jacobian evaluated along the correct nominal trajectory?

4.3.8.165 Key Equations

Nonlinear dynamics

\[ \boxed{ \dot{\mathbf x} = \mathbf f( \mathbf x,\mathbf u,\boldsymbol{\theta},t ) } \]

Perturbation dynamics

\[ \boxed{ \delta\dot{\mathbf x} = \mathbf A\delta\mathbf x + \mathbf B\delta\mathbf u + \mathbf E\delta\boldsymbol{\theta} } \]

Dynamics Jacobian

\[ \boxed{ \mathbf A = \frac{\partial\mathbf f}{\partial\mathbf x} } \]

STM definition

\[ \boxed{ \boldsymbol{\Phi}(t,t_0) = \frac{ \partial\mathbf x(t) }{ \partial\mathbf x(t_0) } } \]

Perturbation propagation

\[ \boxed{ \delta\mathbf x(t) = \boldsymbol{\Phi}(t,t_0) \delta\mathbf x_0 } \]

STM differential equation

\[ \boxed{ \dot{\boldsymbol{\Phi}} = \mathbf A\boldsymbol{\Phi} } \]

STM initial condition

\[ \boxed{ \boldsymbol{\Phi}(t_0,t_0) = \mathbf I } \]

Composition property

\[ \boxed{ \boldsymbol{\Phi}(t_2,t_0) = \boldsymbol{\Phi}(t_2,t_1) \boldsymbol{\Phi}(t_1,t_0) } \]

Inverse property

\[ \boxed{ \boldsymbol{\Phi}(t_0,t) = \boldsymbol{\Phi}(t,t_0)^{-1} } \]

Deterministic covariance propagation

\[ \boxed{ \mathbf P(t) = \boldsymbol{\Phi} \mathbf P_0 \boldsymbol{\Phi}^T } \]

Discrete covariance prediction

\[ \boxed{ \mathbf P_{k+1}^{-} = \mathbf F_k \mathbf P_k^{+} \mathbf F_k^T + \mathbf Q_k } \]

Parameter sensitivity

\[ \boxed{ \mathbf S_\theta = \frac{ \partial\mathbf x }{ \partial\boldsymbol{\theta} } } \]

Parameter-sensitivity differential equation

\[ \boxed{ \dot{\mathbf S}_\theta = \mathbf A \mathbf S_\theta + \mathbf E } \]

Measurement Jacobian

\[ \boxed{ \mathbf H_k = \frac{ \partial h_k }{ \partial\mathbf x_k } } \]

Dynamic measurement sensitivity

\[ \boxed{ \mathbf G_k = \mathbf H_k \boldsymbol{\Phi}_{k,0} } \]

Parameter measurement sensitivity

\[ \boxed{ \frac{\partial h_k} {\partial\boldsymbol{\theta}} = \mathbf H_k \mathbf S_{\theta,k} + \left. \frac{\partial h_k} {\partial\boldsymbol{\theta}} \right|_{direct} } \]

Control perturbation response

\[ \boxed{ \delta\mathbf x(t) = \boldsymbol{\Phi}(t,t_0)\delta\mathbf x_0 + \int_{t_0}^{t} \boldsymbol{\Phi}(t,\tau) \mathbf B(\tau) \delta\mathbf u(\tau) \,d\tau } \]

Finite-difference STM column

\[ \boxed{ \boldsymbol{\Phi}_{:,j} \approx \frac{ \mathbf x( t;\mathbf x_0+\Delta x_j\mathbf e_j ) - \mathbf x( t;\mathbf x_0-\Delta x_j\mathbf e_j ) }{ 2\Delta x_j } } \]

4.3.8.166 Key Takeaways

The State Transition Matrix is

\[ \boxed{ \boldsymbol{\Phi}(t,t_0) = \frac{ \partial\mathbf x(t) }{ \partial\mathbf x(t_0) }. } \]

It describes how small perturbations in the reference state propagate through spacecraft dynamics:

\[ \boxed{ \delta\mathbf x(t) = \boldsymbol{\Phi}(t,t_0) \delta\mathbf x_0. } \]

For nonlinear dynamics,

\[ \boxed{ \dot{\mathbf x} = \mathbf f(\mathbf x,t), } \]

the local perturbation system is

\[ \boxed{ \delta\dot{\mathbf x} = \mathbf A(t)\delta\mathbf x, } \]

with

\[ \boxed{ \mathbf A = \frac{\partial\mathbf f}{\partial\mathbf x}. } \]

The STM therefore satisfies

\[ \boxed{ \dot{\boldsymbol{\Phi}} = \mathbf A \boldsymbol{\Phi}, \qquad \boldsymbol{\Phi}(t_0,t_0)=\mathbf I. } \]

Its most important structural properties are

\[ \boxed{ \boldsymbol{\Phi}(t,t)=\mathbf I, } \]

\[ \boxed{ \boldsymbol{\Phi}(t_2,t_0) = \boldsymbol{\Phi}(t_2,t_1) \boldsymbol{\Phi}(t_1,t_0), } \]

and

\[ \boxed{ \boldsymbol{\Phi}(t_0,t) = \boldsymbol{\Phi}(t,t_0)^{-1}. } \]

The STM is the foundation of covariance propagation:

\[ \boxed{ \mathbf P(t) = \boldsymbol{\Phi} \mathbf P_0 \boldsymbol{\Phi}^T. } \]

It also connects later measurements to a reference state:

\[ \boxed{ \mathbf G_k = \mathbf H_k \boldsymbol{\Phi}_{k,0}. } \]

Thus,

\[ \boxed{ \text{Measurement Sensitivity} = \text{Sensor Sensitivity} \times \text{Dynamic Sensitivity}. } \]

Parameter sensitivity is

\[ \boxed{ \mathbf S_\theta = \frac{\partial\mathbf x}{\partial\boldsymbol{\theta}}, } \]

and satisfies

\[ \boxed{ \dot{\mathbf S}_\theta = \mathbf A \mathbf S_\theta + \mathbf E. } \]

Therefore the same sensitivity framework can describe:

Sensitivity matrices provide the mathematical link among

\[ \boxed{ \text{Dynamics} \leftrightarrow \text{Estimation} \leftrightarrow \text{Covariance} \leftrightarrow \text{Observability} \leftrightarrow \text{Guidance} \leftrightarrow \text{Trajectory Optimization}. } \]

A crucial engineering principle is:

\[ \boxed{ \text{The STM Is a Local Linear Description of a Nonlinear Trajectory}. } \]

It should therefore be verified against nonlinear propagation and finite differences.

Another key point is:

\[ \boxed{ \text{Large Sensitivity Does Not Automatically Mean Large Uncertainty}. } \]

Actual uncertainty impact depends on both

\[ \boxed{ \text{Sensitivity} \times \text{Magnitude of Input Uncertainty}. } \]

Finally,

\[ \boxed{ \text{Sensitivity Analysis Does Not Create Information}. } \]

It tells us how existing uncertainty, measurement information, control authority, and physical parameters propagate through the spacecraft system.

4.3.8.167 What Comes Next

The STM tells us how perturbations propagate:

\[ \delta\mathbf x(t) = \boldsymbol{\Phi} \delta\mathbf x_0. \]

Therefore if the reference state has covariance

\[ \mathbf P_0, \]

the later covariance becomes

\[ \mathbf P(t) = \boldsymbol{\Phi} \mathbf P_0 \boldsymbol{\Phi}^T. \]

But spacecraft estimation requires much more than simply propagating one covariance matrix.

We must understand:

This leads directly to

\[ \boxed{ \textbf{4.3.9 Covariance Analysis and Error Propagation} } \]

where the next section should develop:

Next: 4.3.9 Covariance Analysis and Error Propagation

State-transition and sensitivity matrices show how small perturbations propagate through the dynamics. The next step is to understand how entire uncertainty distributions and covariance structures evolve, transform, accumulate process noise, and reveal coupling between state components.

The next section develops variance and covariance, cross-covariance, covariance-matrix properties, STM-based covariance propagation, continuous and discrete process-noise accumulation, nonlinear Jacobian error propagation, uncertainty ellipses and ellipsoids, principal axes, confidence regions, correlation coefficients, covariance conditioning and consistency, Monte Carlo verification, and general spacecraft navigation applications.

4.3.9 Covariance Analysis and Error Propagation

Covariance definition, error propagation, process noise, coordinate transformations, uncertainty ellipsoids, correlation, confidence regions, cross-covariance, numerical conditioning, consistency, Monte Carlo validation, and navigation use.

Continue to Covariance Analysis and Error Propagation →