Local and Global Parameter Effects | Malone Math

The Central Idea

Suppose a model output is written

\[ y=f(x;\theta), \]

where $x$ is an input and $\theta$ is a parameter.

A local parameter effect asks:

What happens to $y$ when $\theta$ changes by a small amount near a selected parameter value?

A global parameter effect asks:

How does the model behave as $\theta$ varies across a broader range?

Why Both Views Matter

A parameter may have a small local effect near one setting but a large global effect over the full range.

Conversely, a parameter may be locally influential but still leave the model in the same qualitative regime globally.

Local and global analysis answer different questions and should not be treated as substitutes.

Parameter Sensitivity

Sensitivity describes how strongly a model output responds to changes in a parameter.

If a small change in $\theta$ produces a large change in $y$, the output is sensitive to $\theta$.

If the output barely changes, the model is relatively insensitive to that parameter near the point being studied.

Local Sensitivity with a Partial Derivative

\[ y=f(x;\theta), \]

then the local effect of $\theta$ can be measured by

\[ \boxed{ \frac{\partial y}{\partial \theta} }. \]

This derivative describes the instantaneous rate at which the output changes as the parameter changes while the other variables are held fixed.

Interpreting the Sign of Sensitivity

\[ \frac{\partial y}{\partial\theta}>0, \]

then increasing $\theta$ locally increases $y$.

\[ \frac{\partial y}{\partial\theta}<0, \]

then increasing $\theta$ locally decreases $y$.

\[ \frac{\partial y}{\partial\theta}=0, \]

then the output is locally flat with respect to that parameter at the chosen point.

Interpreting Sensitivity Magnitude

The magnitude

\[ \left| \frac{\partial y}{\partial\theta} \right| \]

measures the strength of the local response.

A larger magnitude means a given small parameter change produces a larger output change, subject to the units and scale of the quantities involved.

Example: Linear Parameter Effect

Consider

\[ y=\theta x. \]

Then

\[ \frac{\partial y}{\partial\theta}=x. \]

The sensitivity to $\theta$ depends on the input $x$.

At larger values of $x$, the same parameter change produces a larger absolute change in $y$.

Local Change Approximation

For a small parameter change $\Delta\theta$,

\[ \Delta y \approx \frac{\partial y}{\partial\theta} \Delta\theta. \]

This is a first-order approximation.

It converts local sensitivity into an estimated output change.

Example: Local Approximation

Suppose

\[ y=\theta x^2 \]

with

\[ x=4. \]

Then

\[ \frac{\partial y}{\partial\theta} = x^2 = 16. \]

\[ \Delta\theta=0.1, \]

then

\[ \Delta y \approx 16(0.1) = \boxed{1.6}. \]

Finite Parameter Changes

A global analysis often examines finite changes rather than infinitesimal ones.

If the parameter changes from

\[ \theta_1 \]

\[ \theta_2, \]

then the output change is

\[ \Delta y = f(x;\theta_2)-f(x;\theta_1). \]

This difference captures the full model response across that parameter interval.

Finite-Difference Sensitivity

When derivatives are unavailable, local sensitivity can be estimated using a finite difference:

\[ \boxed{ \frac{\Delta y}{\Delta\theta} = \frac{f(x;\theta+\Delta\theta)-f(x;\theta)} {\Delta\theta} }. \]

For sufficiently small $\Delta\theta$, this approximates

\[ \frac{\partial y}{\partial\theta}. \]

Central Difference

A symmetric local estimate is

\[ \boxed{ \frac{\partial y}{\partial\theta} \approx \frac{f(x;\theta+h)-f(x;\theta-h)}{2h} }. \]

This often gives a better local approximation than a one-sided difference for the same small step size.

Absolute Sensitivity Has Units

The derivative

\[ \frac{\partial y}{\partial\theta} \]

has units

\[ \boxed{ \frac{\text{unit of }y} {\text{unit of }\theta} }. \]

This means raw sensitivity values cannot always be compared fairly across parameters with different units or scales.

Normalized Sensitivity

A dimensionless sensitivity can be formed by comparing relative changes:

\[ S_\theta = \frac{\theta}{y} \frac{\partial y}{\partial\theta}. \]

This quantity measures approximately the percent change in $y$ caused by a one-percent change in $\theta$.

Elasticity as Parameter Sensitivity

The normalized sensitivity

\[ S_\theta = \frac{\theta}{y} \frac{\partial y}{\partial\theta} \]

can also be written

\[ \boxed{ S_\theta = \frac{\partial\ln y}{\partial\ln\theta} }. \]

This is an elasticity with respect to the parameter.

Example: Power-Law Sensitivity

Suppose

\[ y=C\theta^p. \]

Then

\[ \frac{\partial y}{\partial\theta} = Cp\theta^{p-1}. \]

The normalized sensitivity is

\[ \frac{\theta}{y} \frac{\partial y}{\partial\theta} = p. \]

Therefore,

\[ \boxed{S_\theta=p}. \]

Why Normalization Helps

Suppose one parameter is measured in millions of dollars and another in percentages.

Their raw derivatives may have very different numerical sizes simply because the units differ.

Dimensionless sensitivity places both effects on a comparable relative scale.

Local Sensitivity Depends on the Operating Point

Consider

\[ y=\theta x^2. \]

Then

\[ \frac{\partial y}{\partial\theta}=x^2. \]

\[ x=1, \]

the sensitivity is $1$.

\[ x=10, \]

the sensitivity is $100$.

A local sensitivity statement should therefore specify where it is being evaluated.

Local Sensitivity Can Change Sign

Consider

\[ y=(\theta-2)^2. \]

Then

\[ \frac{dy}{d\theta} = 2(\theta-2). \]

The derivative is negative when $\theta<2$, zero at $\theta=2$, and positive when $\theta>2$.

The effect of increasing the parameter changes across the parameter range.

Local Flatness Does Not Mean Global Irrelevance

For

\[ y=(\theta-2)^2, \]

the local sensitivity at

\[ \theta=2 \]

is zero.

But moving far from $2$ makes $y$ grow substantially.

Therefore, a zero local derivative does not mean the parameter has no global effect.

What Is Global Sensitivity?

Global sensitivity studies parameter effects across a substantial region of parameter space.

Instead of asking only what happens near one nominal value, we ask how the model behaves over many plausible parameter settings.

Parameter Sweeps

A simple global analysis varies a parameter across a range:

\[ \theta_{\min} \le \theta \le \theta_{\max}. \]

The model is evaluated at many parameter values, producing a parameter sweep.

This reveals trends, thresholds, nonlinear responses, and regime changes.

Example: Parameter Sweep

Suppose

\[ y=\theta^2 \]

and we examine

\[ 0\le\theta\le5. \]

Then

\[ y \]

ranges from

\[ 0 \]

\[ 25. \]

The global response is strongly nonlinear even though a local approximation near one point may look nearly linear.

One-at-a-Time Analysis

In a model with several parameters,

\[ y=f(\theta_1,\theta_2,\ldots,\theta_n), \]

one simple method varies one parameter while holding the others fixed.

This is called one-at-a-time analysis.

It is easy to interpret but may miss interactions among parameters.

Parameter Interactions

Parameters can influence one another's effects.

Consider

\[ y=\alpha\beta x. \]

The sensitivity to $\alpha$ is

\[ \frac{\partial y}{\partial\alpha} = \beta x. \]

Thus the effect of $\alpha$ depends on the value of $\beta$.

Mixed Partial Derivatives

Interaction can be studied locally using a mixed derivative:

\[ \frac{\partial^2 y} {\partial\alpha\,\partial\beta}. \]

If this quantity is nonzero, the local effect of one parameter changes as the other parameter changes.

Global Analysis Reveals Interactions

A global parameter study can vary several parameters together.

This allows the modeler to detect combinations that produce behavior not visible when each parameter is varied separately.

Interaction effects are especially important in nonlinear systems.

Threshold Effects

Some parameters matter little until they cross a threshold.

A system may behave one way for

\[ \theta<\theta_c \]

and another way for

\[ \theta>\theta_c. \]

Local analysis far from the threshold may not reveal this global change.

Regime Changes

A regime is a range of conditions under which the model follows a particular qualitative pattern.

Varying parameters globally can reveal transitions between regimes such as:

growth versus decay;
stable versus unstable behavior;
linear versus saturated response;
low-flow versus high-flow behavior;
profitable versus unprofitable operation.

Example: Growth and Decay Parameter

Consider

\[ P(t)=P_0e^{rt}. \]

The parameter $r$ controls a global qualitative change.

\[ r>0, \]

the model grows.

\[ r<0, \]

it decays.

\[ r=0, \]

the model is constant.

Local Sensitivity in Exponential Growth

For

\[ P(t)=P_0e^{rt}, \]

sensitivity to $r$ is

\[ \frac{\partial P}{\partial r} = tP_0e^{rt}. \]

Since

\[ P=P_0e^{rt}, \]

we can write

\[ \boxed{ \frac{\partial P}{\partial r}=tP }. \]

Sensitivity increases with time.

Parameter Effects Can Accumulate Over Time

A rate parameter may have only a modest short-term effect but a large long-term effect.

This is common in exponential, compounding, population, and dynamic models.

Global time horizon is therefore part of parameter-effect analysis.

Local Sensitivity in a Linear Model

Suppose

\[ y=mx+b. \]

Then

\[ \frac{\partial y}{\partial m}=x \]

and

\[ \frac{\partial y}{\partial b}=1. \]

The slope parameter becomes more influential in absolute terms as $|x|$ grows, while the intercept shifts the output by the same absolute amount everywhere.

Global Effects in a Linear Family

Changing $m$ changes the orientation of the entire line.

Changing $b$ translates the entire line vertically.

These are global structural effects that are clearer when the complete family of curves is considered rather than one local output value.

Local Sensitivity in Logistic Growth

Consider

\[ P(t) = \frac{K}{1+Ae^{-rt}}. \]

The parameters $K$, $A$, and $r$ influence different features.

Sensitivity to each parameter changes with time, so no single local derivative tells the whole story.

Global Effects in Logistic Growth

In the logistic model:

$K$ changes the limiting level;
$r$ changes how rapidly the system approaches that level;
$A$ shifts the trajectory relative to its initial condition.

These effects are best understood by examining the complete response curve.

Parameter Robustness

A model output is robust to a parameter when reasonable changes in that parameter do not greatly alter the conclusion.

Robustness is often more important than exact parameter precision.

If a decision remains unchanged across a wide parameter range, the model may support a stronger practical conclusion.

Fragile Conclusions

A conclusion is fragile when a small parameter change reverses or substantially alters it.

For example, if estimated profit changes from positive to negative after a one-percent change in a cost parameter, the decision is highly sensitive to that assumption.

Parameter Effects and Uncertainty

Parameter values are often uncertain.

Sensitivity analysis asks how that uncertainty affects model outputs.

If an uncertain parameter has little effect, greater measurement precision may not be operationally important.

If an uncertain parameter has a large effect, better estimation may be valuable.

Prioritizing Data Collection

Sensitivity analysis can help determine which parameters deserve better data.

A useful priority is often a parameter that is both:

uncertain;
highly influential.

A parameter that is highly uncertain but nearly irrelevant may not justify expensive measurement.

Parameter Effects and Calibration

During calibration, model parameters are adjusted to fit observations.

If several parameter combinations produce nearly identical outputs, the calibration problem may be poorly identified.

Sensitivity analysis can reveal which parameters are actually distinguishable from the available data.

Identifiability and Sensitivity

A parameter that barely affects observable outputs can be difficult to estimate accurately.

Low sensitivity does not always mean the parameter is unimportant scientifically, but it may mean the available experiment cannot identify it reliably.

Scenario Analysis

Scenario analysis evaluates a model under selected parameter combinations.

Examples include:

low, baseline, and high demand;
optimistic, expected, and pessimistic growth;
low-cost and high-cost environments;
different policy or operating assumptions.

Scenarios are a practical form of global parameter analysis.

Best-Case and Worst-Case Analysis

If parameters lie within known bounds, one can evaluate combinations that produce favorable or unfavorable outcomes.

This does not assign probabilities to the cases.

It asks what the model permits within the assumed parameter ranges.

Global Parameter Ranges

A parameter study should state its range explicitly:

\[ \theta_{\min} \le \theta \le \theta_{\max}. \]

Conclusions are conditional on this chosen range.

Expanding the range may reveal behavior not seen in a narrower analysis.

Parameter Distributions

When uncertainty is probabilistic, a parameter may be modeled with a distribution rather than a fixed interval.

The model can then be evaluated across many sampled parameter values.

This connects parameter analysis with simulation and uncertainty quantification.

Monte Carlo Parameter Analysis

A Monte Carlo approach repeatedly samples uncertain parameters and computes model outputs.

The resulting output distribution can reveal:

typical outcomes;
extreme outcomes;
parameter-driven variability;
probability of crossing thresholds.

Local Analysis Is Computationally Efficient

Local derivatives can summarize sensitivity with relatively little computation.

They are especially useful when:

changes are expected to remain small;
the model is smooth;
a nominal operating point is meaningful;
many parameters must be screened quickly.

Global Analysis Is More Comprehensive

Global methods require more evaluations but can reveal:

nonlinearity;
thresholds;
multiple regimes;
parameter interactions;
non-monotonic responses;
behavior far from the nominal setting.

Local and Global Analysis Should Work Together

A practical workflow often begins locally and then expands globally.

Local derivatives identify parameters that appear influential near the baseline.

Global analysis then tests whether those conclusions persist across the broader parameter range.

Dimensionless Sensitivity Supports Comparison

Suppose two parameters have different units.

Their normalized sensitivities

\[ S_{\theta_i} = \frac{\theta_i}{y} \frac{\partial y}{\partial\theta_i} \]

provide a dimensionless basis for comparison.

This connects sensitivity analysis with the scaling and nondimensionalization ideas developed earlier in this section.

Sensitivity and Characteristic Scales

If $Y_c$ is a characteristic output scale and $\Theta_c$ a characteristic parameter scale, then

\[ \frac{\Theta_c}{Y_c} \frac{\partial y}{\partial\theta} \]

gives a scale-aware measure of parameter influence.

Scaling prevents large numerical units from being mistaken for large model importance.

Sensitivity and Error Propagation

If $\theta$ has uncertainty $\Delta\theta$, then

\[ \Delta y \approx \frac{\partial y}{\partial\theta} \Delta\theta. \]

Thus the same derivative used for local parameter analysis also describes first-order propagation of parameter uncertainty.

Several Uncertain Parameters

\[ y=f(\theta_1,\theta_2,\ldots,\theta_n), \]

a first-order change is approximately

\[ \Delta y \approx \sum_{i=1}^{n} \frac{\partial y}{\partial\theta_i} \Delta\theta_i. \]

This linearization summarizes how several small parameter changes combine.

Gradient with Respect to Parameters

The local sensitivities can be collected into a parameter gradient:

\[ \nabla_{\theta}y = \left( \frac{\partial y}{\partial\theta_1}, \frac{\partial y}{\partial\theta_2}, \ldots, \frac{\partial y}{\partial\theta_n} \right). \]

This vector describes the local direction in parameter space that changes the output most rapidly.

Sensitivity of a Decision, Not Just an Output

Sometimes the important question is not how much an output changes, but whether a decision changes.

For example, a model may recommend option A whenever

\[ y>0. \]

Parameter analysis should then ask which parameter values cause the model to cross the decision boundary

\[ y=0. \]

Break-Even Parameters

A break-even parameter is a parameter value at which a model output reaches a threshold.

Suppose profit is

\[ \Pi(p)=pq-C. \]

Break-even occurs when

\[ \Pi(p)=0. \]

Solving gives

\[ p=\frac{C}{q}. \]

This parameter threshold can be more decision-relevant than a local derivative.

Example: Cost Sensitivity

Suppose profit is

\[ \Pi = pq-cq-F, \]

where $c$ is variable cost per unit.

Then

\[ \frac{\partial\Pi}{\partial c} = -q. \]

If $q=10{,}000$, then a \$1 increase in unit cost lowers modeled profit by

\[ \boxed{\$10{,}000}. \]

Example: Price Sensitivity

Using

\[ \Pi = pq-cq-F, \]

the local sensitivity to price is

\[ \frac{\partial\Pi}{\partial p} = q. \]

At fixed quantity, a \$1 increase in price raises modeled profit by $q$ dollars.

In a more realistic model where demand depends on price, this effect can become nonlinear and globally more complicated.

Common Error: Treating Local Sensitivity as Global Truth

A derivative describes behavior near one point.

It should not automatically be extrapolated across a wide parameter range.

Nonlinear models can bend, flatten, reverse direction, or cross thresholds.

Common Error: Comparing Raw Sensitivities with Different Units

A derivative measured in dollars per percentage point cannot be compared directly with one measured in dollars per kilogram without considering parameter scales.

Normalize when cross-parameter comparison is the goal.

Common Error: Varying One Parameter and Ignoring Interactions

One-at-a-time analysis may miss situations in which parameters reinforce or offset one another.

Global multiparameter analysis is needed when interactions are important.

Common Error: Choosing Unrealistic Parameter Ranges

A global sensitivity result depends on the parameter ranges being explored.

Extremely wide or implausible ranges may overstate possible behavior.

Ranges should be justified by data, constraints, design limits, or credible scenarios.

Common Error: Ranking Sensitivity Without Considering Uncertainty

A parameter can be highly sensitive but known almost exactly.

Another can be moderately sensitive but highly uncertain.

Practical risk depends on both influence and uncertainty.

Common Error: Ignoring the Output of Interest

A parameter may strongly affect one model output and barely affect another.

Sensitivity rankings should therefore specify the output, time horizon, and operating point being analyzed.

A Reliable Local Sensitivity Procedure

Identify the model output of interest.
Choose the nominal parameter values and operating point.
Compute or estimate $\partial y/\partial\theta$.
Interpret the sign of the derivative.
Interpret its magnitude together with units.
Normalize the sensitivity if comparing parameters with different scales.
Use $\Delta y\approx(\partial y/\partial\theta)\Delta\theta$ for small changes.
Check whether the approximation remains valid for the intended parameter change.

A Reliable Global Parameter Procedure

Identify the parameters to vary.
Choose credible parameter ranges or distributions.
Identify the outputs and decisions to monitor.
Evaluate the model across the parameter space.
Look for monotonic trends, curvature, thresholds, and regime changes.
Examine interactions among parameters.
Compare output variation with operational tolerances or decision thresholds.
Identify robust and fragile conclusions.
Use the results to prioritize calibration, data collection, or model refinement.

Example: Full Local Workflow

Suppose

\[ y=\theta^2x \]

with

\[ x=5 \]

and

\[ \theta=3. \]

The local sensitivity is

\[ \frac{\partial y}{\partial\theta} = 2\theta x. \]

At the selected point,

\[ \frac{\partial y}{\partial\theta} = 2(3)(5) = 30. \]

For a small change

\[ \Delta\theta=0.02, \]

the predicted output change is

\[ \Delta y \approx 30(0.02) = \boxed{0.6}. \]

Example: Full Global Workflow

Consider

\[ y=(\theta-2)^2 \]

over

\[ 0\le\theta\le5. \]

The local derivative is

\[ \frac{dy}{d\theta} = 2(\theta-2). \]

The derivative changes sign at

\[ \theta=2. \]

Globally, the output decreases as $\theta$ approaches $2$, reaches its minimum there, and then increases.

A single local slope cannot summarize this entire behavior.

Practice 1: Local Sensitivity

\[ y=\theta x, \]

find

\[ \frac{\partial y}{\partial\theta}. \]

Practice 2: Evaluate Sensitivity

Using

\[ y=\theta x, \]

evaluate the sensitivity at

\[ x=12. \]

Practice 3: Local Change

If the sensitivity is $12$ and the parameter changes by $0.05$, estimate the output change.

Practice 4: Power-Law Normalized Sensitivity

\[ y=C\theta^4, \]

find the normalized sensitivity

\[ S_\theta. \]

Practice 5: Sign Change

For

\[ y=(\theta-3)^2, \]

find where the local sensitivity changes sign.

Practice 6: Linear Model Parameters

For

\[ y=mx+b, \]

find the sensitivities to $m$ and $b$.

Practice 7: Exponential Sensitivity

For

\[ P=P_0e^{rt}, \]

find

\[ \frac{\partial P}{\partial r}. \]

Practice 8: One-at-a-Time Limitation

Why can varying one parameter at a time fail to reveal the full behavior of a multiparameter model?

Practice 9: Robustness

A decision remains unchanged while a parameter varies by $\pm20\%$. What does this suggest about the decision with respect to that parameter?

Practice 10: Threshold Analysis

Suppose

\[ y=5\theta-20. \]

Find the parameter value at which the output crosses the threshold

\[ y=0. \]

Practice Answers

Practice 1

\[ \boxed{ \frac{\partial y}{\partial\theta}=x }. \]

Practice 2

\[ \frac{\partial y}{\partial\theta} = x = \boxed{12}. \]

Practice 3

\[ \Delta y \approx 12(0.05) = \boxed{0.6}. \]

Practice 4

Since

\[ y=C\theta^4, \]

the normalized sensitivity equals the exponent:

\[ \boxed{S_\theta=4}. \]

Practice 5

\[ \frac{dy}{d\theta} = 2(\theta-3). \]

The sensitivity changes sign at

\[ \boxed{\theta=3}. \]

Practice 6

\[ \boxed{ \frac{\partial y}{\partial m}=x } \]

and

\[ \boxed{ \frac{\partial y}{\partial b}=1 }. \]

Practice 7

\[ \boxed{ \frac{\partial P}{\partial r} = tP_0e^{rt} = tP }. \]

Practice 8

One-at-a-time analysis holds all other parameters fixed, so it can miss interactions in which the effect of one parameter depends on the values of others.

Practice 9

The decision appears relatively robust to that parameter over the tested range.

Practice 10

Set

\[ 5\theta-20=0. \]

Then

\[ 5\theta=20, \]

\[ \boxed{\theta=4}. \]

What to Remember

Local and global parameter analysis reveal different dimensions of model behavior.

local sensitivity measures response near a selected parameter value;
partial derivatives provide a natural local sensitivity measure;
finite differences can estimate sensitivity numerically;
normalized sensitivities allow relative comparison across parameters;
global analysis explores broad parameter ranges and multiple regimes;
parameter sweeps reveal curvature, thresholds, and non-monotonic behavior;
multiparameter analysis can reveal interactions;
sensitivity and uncertainty together determine practical parameter importance;
robust conclusions remain stable across plausible parameter changes;
fragile conclusions should be treated cautiously and may require better data or stronger modeling.

The key modeling habit is: do not ask only whether a parameter matters—ask where, how much, and across what range it matters.

Where We Go Next

Local and global parameter effects complete the core toolkit for Quantities, Parameters, Scale, and Dimension. Together, measurable quantities, dimensional consistency, derived units, parameters, characteristic scales, nondimensionalization, scaling laws, orders of magnitude, proportional reasoning, precision, and sensitivity provide the language needed to build and evaluate mathematical models before more advanced analysis begins.