The Central Idea
Suppose
\[ y=f(x). \]Near \(x=a\), a small input change \(\Delta x\) produces approximately
\[ \Delta y\approx f'(a)\Delta x. \]The derivative
\[ \boxed{f'(a)} \]measures the local sensitivity of the output \(y\) to the input \(x\).
It answers the practical question:
If the input changes a little near \(a\), how strongly should we expect the output to respond?
Sensitivity Is Local
Sensitivity usually depends on where the system is operating.
For
\[ f(x)=x^2, \]the derivative is
\[ f'(x)=2x. \]At \(x=1\),
\[ f'(1)=2, \]while at \(x=10\),
\[ f'(10)=20. \]The same small change in \(x\) has a much stronger effect on \(x^2\) near \(10\) than near \(1\).
The Sign of Sensitivity
The sign of the derivative tells us the direction of response.
If
\[ f'(a)>0, \]then increasing the input slightly tends to increase the output.
If
\[ f'(a)<0, \]then increasing the input slightly tends to decrease the output.
If
\[ f'(a)=0, \]then the first-order response vanishes at that point.
The Magnitude of Sensitivity
The size
\[ |f'(a)| \]measures the strength of the local response without regard to direction.
A large value of \(|f'(a)|\) means the output is locally responsive.
A small value of \(|f'(a)|\) means the output is locally less responsive.
The sign gives direction; the magnitude gives strength.
Absolute Sensitivity
The ordinary derivative
\[ \boxed{ S_{\text{abs}}(x)=f'(x) } \]can be interpreted as absolute sensitivity.
It measures output change per unit input change.
For a small change \(dx\),
\[ dy=f'(x)\,dx. \]This is the differential form of the same sensitivity relationship.
Units of Absolute Sensitivity
Derivative units matter.
If \(x\) is measured in seconds and \(y\) in meters, then
\[ \frac{dy}{dx} \]has units
\[ \frac{\text{meters}}{\text{second}}. \]If \(x\) is dollars of advertising and \(y\) is units sold, then
\[ \frac{dy}{dx} \]has units of units sold per dollar of advertising.
Example: Temperature Sensitivity
Suppose energy consumption is modeled by
\[ E(T)=0.5T^2-8T+120, \]where \(T\) is temperature in degrees and \(E\) is energy use in suitable units.
Differentiate:
\[ E'(T)=T-8. \]At \(T=20\),
\[ E'(20)=12. \]Near \(T=20\), a one-degree increase in temperature changes the modeled energy use by approximately \(12\) output units.
From Sensitivity to Small-Change Prediction
If
\[ f'(a)=12 \]and the input changes by
\[ \Delta x=0.1, \]then
\[ \Delta y \approx 12(0.1) = 1.2. \]Thus the derivative converts local input movement into predicted output movement.
Why Sensitivity and Linearization Are Closely Connected
Linearization gives
\[ f(x) \approx f(a)+f'(a)(x-a). \]The coefficient
\[ f'(a) \]is exactly the local sensitivity.
Linearization is therefore a sensitivity model:
\[ \boxed{ \text{new output} \approx \text{baseline output} + \text{sensitivity}\times\text{input change} }. \]Example: Sensitivity of Circle Area to Radius
A circle has area
\[ A(r)=\pi r^2. \]Differentiate:
\[ \boxed{ \frac{dA}{dr}=2\pi r }. \]This derivative is the sensitivity of area to radius.
At \(r=5\),
\[ \frac{dA}{dr}=10\pi. \]Near radius \(5\), each small unit increase in radius produces about \(10\pi\) square units of area increase per unit of radius.
The Same Formula Can Have Different Sensitivity at Different Scales
For
\[ A(r)=\pi r^2, \]the sensitivity is
\[ A'(r)=2\pi r. \]At \(r=1\),
\[ A'(1)=2\pi. \]At \(r=100\),
\[ A'(100)=200\pi. \]The area becomes far more sensitive in absolute terms as the radius grows.
Absolute Sensitivity Depends on Units
Suppose a length is measured in meters rather than centimeters.
The numerical value of the derivative may change because the input unit has changed.
Therefore, comparing ordinary derivatives across differently scaled variables can be misleading.
This motivates dimensionless or relative sensitivity measures.
Relative Sensitivity
A common dimensionless sensitivity measure compares fractional output change with fractional input change.
For \(x\neq0\) and \(f(x)\neq0\), define
\[ \boxed{ S_{\text{rel}}(x) = \frac{x}{f(x)}f'(x) }. \]This quantity measures approximate percentage output change per percentage input change.
Deriving Relative Sensitivity
Start with the differential approximation
\[ dy\approx f'(x)\,dx. \]Divide by \(y=f(x)\):
\[ \frac{dy}{y} \approx \frac{f'(x)}{f(x)}\,dx. \]Multiply and divide by \(x\):
\[ \frac{dy}{y} \approx \frac{x f'(x)}{f(x)} \frac{dx}{x}. \]Therefore,
\[ \boxed{ \frac{dy}{y} \approx S_{\text{rel}}(x)\frac{dx}{x} }. \]Interpreting Relative Sensitivity
If
\[ S_{\text{rel}}(x)=2, \]then a small \(1\%\) increase in the input produces approximately a \(2\%\) increase in the output.
If
\[ S_{\text{rel}}(x)=-0.5, \]then a small \(1\%\) increase in the input produces approximately a \(0.5\%\) decrease in the output.
Elasticity
In economics and other applied fields, the dimensionless quantity
\[ \boxed{ E(x)=\frac{x}{f(x)}f'(x) } \]is often called an elasticity.
It is mathematically the same normalized sensitivity measure.
The terminology depends on context, but the derivative structure is the same.
Example: Relative Sensitivity of a Power Function
Suppose
\[ f(x)=Cx^n, \]where \(C\neq0\).
Then
\[ f'(x)=Cnx^{n-1}. \]Therefore,
\[ \frac{x}{f(x)}f'(x) = \frac{x}{Cx^n} Cnx^{n-1} = n. \]Thus,
\[ \boxed{ S_{\text{rel}}=n }. \]A power law has constant relative sensitivity equal to its exponent.
Example: Area and Volume Scaling
Circle area scales as
\[ A\propto r^2. \]Therefore, its relative sensitivity to radius is
\[ \boxed{2}. \]Sphere volume scales as
\[ V\propto r^3. \]Therefore, its relative sensitivity to radius is
\[ \boxed{3}. \]A small \(1\%\) radius change produces approximately a \(2\%\) area change or \(3\%\) volume change.
Sensitivity Near a Zero Output
Relative sensitivity
\[ \frac{x}{f(x)}f'(x) \]is undefined when
\[ f(x)=0. \]This does not mean the ordinary derivative is meaningless there.
It means percentage change relative to a zero baseline is not defined in the usual way.
Sensitivity Near a Zero Input
Relative sensitivity also requires care when
\[ x=0. \]A fractional input change
\[ \frac{dx}{x} \]is not defined at zero.
Absolute sensitivity may still be perfectly meaningful.
Parameter Sensitivity
Many models depend not only on a variable \(x\), but also on a parameter \(p\):
\[ y=F(x,p). \]We may ask how the output changes when \(p\) changes while \(x\) is held fixed.
The corresponding sensitivity is written
\[ \boxed{ \frac{\partial F}{\partial p} }. \]This is a preview of partial derivatives, which will be developed more fully in multidimensional analysis.
Variable Sensitivity and Parameter Sensitivity
In
\[ y=F(x,p), \]the derivative with respect to \(x\) measures response to the operating variable:
\[ \frac{\partial F}{\partial x}. \]The derivative with respect to \(p\) measures response to the parameter:
\[ \frac{\partial F}{\partial p}. \]The distinction is conceptual: one changes the state or input, while the other changes the model setting.
Example: Exponential Growth Parameter
Suppose
\[ P(t;r)=P_0e^{rt}. \]Here \(t\) is time and \(r\) is a growth-rate parameter.
Sensitivity to time is
\[ \frac{\partial P}{\partial t} = rP_0e^{rt} = rP. \]Sensitivity to the parameter \(r\) is
\[ \boxed{ \frac{\partial P}{\partial r} = tP_0e^{rt} = tP }. \]The effect of uncertainty in \(r\) becomes stronger as time grows.
Local Parameter Effects Can Change with Operating Conditions
In the exponential model
\[ P(t;r)=P_0e^{rt}, \]the sensitivity
\[ \frac{\partial P}{\partial r}=tP \]depends on both time and current population level.
Parameter sensitivity is therefore not necessarily a fixed property of the parameter alone.
Example: A Linear Parameter
Suppose
\[ y=ax+b. \]Sensitivity to \(a\) is
\[ \frac{\partial y}{\partial a}=x. \]Sensitivity to \(b\) is
\[ \frac{\partial y}{\partial b}=1. \]Thus uncertainty in the slope parameter is amplified by the size of \(x\), while uncertainty in the intercept passes directly into the output.
Several Small Changes at Once
If
\[ y=F(x,p), \]and both \(x\) and \(p\) change slightly, the first-order output change is
\[ \boxed{ dy \approx \frac{\partial F}{\partial x}\,dx + \frac{\partial F}{\partial p}\,dp }. \]Each derivative measures one channel through which the output responds.
This is a multidimensional extension of ordinary differential sensitivity.
Sensitivity and Uncertainty Propagation
Suppose \(x\) has a small uncertainty \(\delta x\).
If
\[ y=f(x), \]then the corresponding first-order output uncertainty is approximately
\[ \boxed{ |\delta y| \approx |f'(x)|\,|\delta x| }. \]A sensitive model amplifies input uncertainty more strongly.
Example: Measurement Sensitivity
Suppose
\[ V=\frac43\pi r^3 \]and the measured radius is
\[ r=10\text{ cm} \]with possible uncertainty
\[ |\delta r|=0.01\text{ cm}. \]Since
\[ \frac{dV}{dr}=4\pi r^2, \]at \(r=10\),
\[ \left|\frac{dV}{dr}\right| = 400\pi. \]Therefore,
\[ |\delta V| \approx 400\pi(0.01) = 4\pi\text{ cm}^3. \]Relative Uncertainty
If \(y=f(x)\), then
\[ \frac{dy}{y} \approx \frac{x f'(x)}{f(x)} \frac{dx}{x}. \]Taking magnitudes gives the approximate relative-uncertainty relationship
\[ \boxed{ \left|\frac{dy}{y}\right| \approx \left| \frac{x f'(x)}{f(x)} \right| \left|\frac{dx}{x}\right| }. \]Relative sensitivity tells us how fractional uncertainty is amplified or reduced.
Conditioning: A Preview
In numerical analysis, sensitivity of an output to small changes in input is closely related to the idea of conditioning.
A problem is locally well-conditioned when small input perturbations produce comparably small output perturbations.
It is poorly conditioned when tiny input changes can produce large output changes.
Relative sensitivity often serves as a local condition measure.
Example: Sensitivity of a Reciprocal
Let
\[ f(x)=\frac1x. \]Then
\[ f'(x)=-\frac1{x^2}. \]Absolute sensitivity has magnitude
\[ \frac1{x^2}. \]As \(x\) approaches zero, the absolute sensitivity becomes very large.
Small absolute input errors near zero can therefore produce large absolute output changes.
Relative Sensitivity of a Reciprocal
For
\[ f(x)=x^{-1}, \]the power-law rule gives
\[ \boxed{ S_{\text{rel}}=-1 }. \]A small \(1\%\) increase in \(x\) produces approximately a \(1\%\) decrease in \(1/x\).
This remains true even though the absolute sensitivity changes strongly with scale.
Absolute and Relative Sensitivity Answer Different Questions
| Measure | Question |
|---|---|
| \(f'(x)\) | How many output units change per input unit? |
| \(\frac{x}{f(x)}f'(x)\) | How many percent of output change per percent of input change? |
Neither measure is universally better. The correct choice depends on the decision or comparison being made.
Sensitivity of an Inverse Relationship
Suppose \(y=f(x)\) is locally invertible and
\[ f'(x)\neq0. \]Then
\[ \frac{dx}{dy} = \frac{1}{dy/dx}. \]Therefore,
\[ \boxed{ \frac{dx}{dy} = \frac1{f'(x)} }. \]Strong forward sensitivity corresponds to weak inverse sensitivity, and vice versa, when the inverse relationship is locally valid.
Example: Inverse Sensitivity
Let
\[ y=x^2 \]on the branch \(x>0\).
Then
\[ \frac{dy}{dx}=2x. \]Hence,
\[ \frac{dx}{dy} = \frac1{2x}. \]At \(x=10\), \(y\) is very sensitive to \(x\), while recovering \(x\) from a small change in \(y\) is comparatively less sensitive.
Zero Derivative and First-Order Insensitivity
If
\[ f'(a)=0, \]then
\[ \Delta y\approx f'(a)\Delta x=0 \]at first order.
This does not mean the function is constant near \(a\).
It means the first-order response vanishes and higher-order terms may dominate.
Example: Sensitivity Near a Minimum
For
\[ f(x)=x^2, \]we have
\[ f'(0)=0. \]A small change from \(x=0\) produces no first-order output change.
But
\[ f(\Delta x)-f(0) = (\Delta x)^2. \]The change is second-order rather than zero.
Higher Derivatives Refine Sensitivity Information
The first derivative gives first-order sensitivity.
The second derivative tells us how sensitivity itself changes:
\[ f''(x) = \frac{d}{dx}f'(x). \]Thus:
- \(f''>0\): sensitivity is increasing;
- \(f''<0\): sensitivity is decreasing.
Higher derivatives provide still finer information about changing response.
Sensitivity and Concavity
When
\[ f''(x)>0, \]the first derivative rises as \(x\) increases.
In sensitivity language, the output becomes progressively more responsive in the positive direction.
When
\[ f''(x)<0, \]the first derivative falls as \(x\) increases.
Sensitivity and Critical Points
Critical points often occur where the first-order sensitivity vanishes:
\[ f'(c)=0. \]Optimization frequently searches for such locations because a small first-order perturbation in the decision variable produces no first-order change in the objective.
The second derivative then helps determine whether the point behaves like a local maximum, local minimum, or neither.
Sensitivity Can Be Directional
For a one-variable differentiable function, local sensitivity is captured by one derivative.
For a model with several inputs, sensitivity can differ by direction.
A model may be highly sensitive to one parameter and weakly sensitive to another.
This is one reason multidimensional models require partial derivatives and gradients.
Example: Two Parameter Effects
Suppose
\[ Q(a,b)=a^2b. \]Sensitivity to \(a\) is
\[ \frac{\partial Q}{\partial a}=2ab. \]Sensitivity to \(b\) is
\[ \frac{\partial Q}{\partial b}=a^2. \]At different operating values of \(a\) and \(b\), the dominant sensitivity can change.
Comparing Parameter Importance
Raw partial derivatives can be difficult to compare when parameters use different units.
A normalized parameter sensitivity can be written
\[ \boxed{ S_p = \frac{p}{F}\frac{\partial F}{\partial p} } \]when \(p\neq0\) and \(F\neq0\).
This expresses approximate percentage output change per percentage parameter change.
Sensitivity Is Not Causality
A derivative describes how a mathematical output changes with respect to an input inside a specified model.
It does not by itself prove that changing the input causes the observed real-world change.
Causal interpretation requires assumptions about the model, measurement, confounding, experimental design, or mechanism.
Sensitivity Depends on the Model
Suppose two models fit the same observations:
\[ y=f(x) \]and
\[ y=g(x). \]They may have different derivatives at the same input.
Therefore, sensitivity conclusions inherit the assumptions and limitations of the model used to compute them.
Local Sensitivity Is Not Global Behavior
A large derivative at one point does not imply large sensitivity everywhere.
Likewise, a small derivative at one point does not imply that the function is globally insensitive.
Sensitivity should be evaluated over the region relevant to the decision or application.
Sensitivity Profiles
Instead of evaluating the derivative at one point, we can study
\[ f'(x) \]over an interval.
This produces a sensitivity profile showing where the model is:
- most responsive;
- least responsive;
- positively responsive;
- negatively responsive;
- first-order insensitive.
Example: Logistic Sensitivity
Consider the logistic form
\[ P(t)=\frac{K}{1+Ae^{-rt}}. \]Its derivative can be written
\[ \boxed{ P'(t) = rP\left(1-\frac{P}{K}\right) }. \]The growth sensitivity to time is small when \(P\) is near \(0\) or near the carrying capacity \(K\), and larger in the middle region.
Sensitivity can therefore vary systematically through the evolution of a model.
Sensitivity and Thresholds
In decision systems, a quantity may be especially important when the model is near a threshold.
Even a modest derivative can matter if a small output change crosses a critical boundary.
Sensitivity analysis should therefore consider both:
- the numerical size of the derivative;
- the consequences of the resulting output movement.
Sensitivity and Robustness
A model or decision is often called robust when reasonable input or parameter variations do not change the conclusion very much.
Small derivative magnitudes can support local robustness.
Large derivative magnitudes can signal fragility and motivate tighter measurement, better parameter estimation, or additional scenario analysis.
Example: Revenue Sensitivity
Suppose revenue depends on price through
\[ R(p)=p(1000-20p). \]Expand:
\[ R(p)=1000p-20p^2. \]Differentiate:
\[ R'(p)=1000-40p. \]At \(p=20\),
\[ R'(20)=200. \]Near \(p=20\), a small one-unit increase in price changes modeled revenue by approximately \(200\) revenue units per price unit.
Revenue Sensitivity Changes Sign
For
\[ R'(p)=1000-40p, \]the derivative is positive when
\[ p<25, \]zero when
\[ p=25, \]and negative when
\[ p>25. \]The sign change shows that increasing price helps revenue below the critical price but hurts revenue above it within this model.
Sensitivity and Optimization Are Connected
Optimization asks where an objective stops improving in the first-order sense.
At an interior differentiable optimum,
\[ f'(x)=0 \]is a common necessary condition.
In sensitivity language, the objective is first-order insensitive to small decision changes at that point.
A Reliable Procedure for Absolute Sensitivity
- Identify the output \(y=f(x)\).
- Differentiate to find \(f'(x)\).
- Evaluate \(f'(x)\) at the operating point.
- Interpret the sign as response direction.
- Interpret the magnitude as response strength.
- State derivative units.
- Use \(dy\approx f'(x)\,dx\) for small-change prediction when needed.
A Reliable Procedure for Relative Sensitivity
- Compute \(f(x)\) and \(f'(x)\).
- Verify \(x\neq0\) and \(f(x)\neq0\).
- Form \(S_{\text{rel}}=\frac{x}{f(x)}f'(x)\).
- Interpret the sign as the direction of percentage response.
- Interpret the magnitude as percentage amplification or reduction.
- Use \(\frac{dy}{y}\approx S_{\text{rel}}\frac{dx}{x}\) for small relative changes.
A Reliable Procedure for Parameter Sensitivity
- Write the model with its parameter visible, such as \(y=F(x,p)\).
- Choose the parameter whose effect is being studied.
- Differentiate with respect to that parameter while holding other independent quantities fixed.
- Evaluate the derivative at the operating point.
- Interpret the sign, magnitude, and units.
- Normalize the sensitivity if percentage comparisons are more useful.
Common Error: Comparing Raw Derivatives with Different Units
Suppose one parameter is measured in dollars and another in seconds.
Their raw derivatives have different units and may not be directly comparable.
A normalized sensitivity can provide a more meaningful comparison.
Common Error: Ignoring the Sign
Sensitivity magnitude alone does not tell us whether the output increases or decreases.
A derivative of
\[ -8 \]and a derivative of
\[ 8 \]have the same absolute strength but opposite response directions.
Common Error: Treating Sensitivity as Constant
Unless the derivative itself is constant, sensitivity varies with the operating point.
For
\[ f(x)=x^2, \]sensitivity is
\[ 2x, \]so it changes continuously with \(x\).
Common Error: Using Relative Sensitivity at a Zero Baseline
The formula
\[ \frac{x}{f(x)}f'(x) \]requires
\[ x\neq0 \]and
\[ f(x)\neq0. \]If either baseline is zero, percentage interpretation requires a different approach.
Common Error: Treating a Derivative as a Causal Effect
A model derivative measures change within the mathematical relationship.
It does not automatically establish that manipulating the input will cause the predicted output change in the real world.
Causal claims require additional evidence and assumptions.
Common Error: Applying a Local Sensitivity Too Far Away
The approximation
\[ \Delta y\approx f'(a)\Delta x \]is local.
If \(\Delta x\) is large, the derivative may change substantially over the interval and the first-order estimate may become poor.
Practice 1: Absolute Sensitivity
Let
\[ f(x)=x^3. \]Find the absolute sensitivity at \(x=2\).
Practice 2: Small-Change Estimate
For
\[ f(x)=x^3, \]estimate the change in \(f\) when \(x\) increases from \(2\) to \(2.01\).
Practice 3: Sign Interpretation
A model satisfies
\[ f'(a)=-6. \]Interpret the sign and magnitude of the local sensitivity.
Practice 4: Relative Sensitivity
Find the relative sensitivity of
\[ f(x)=5x^4. \]Practice 5: Percentage Interpretation
A model has relative sensitivity
\[ S_{\text{rel}}=1.5. \]Estimate the percentage output change produced by a small \(2\%\) increase in the input.
Practice 6: Circle Sensitivity
For
\[ A=\pi r^2, \]find the absolute sensitivity of area to radius at \(r=6\), and state the relative sensitivity.
Practice 7: Parameter Sensitivity
Suppose
\[ y=ax^2+b. \]Find the sensitivity of \(y\) to \(a\) and to \(b\).
Practice 8: Inverse Sensitivity
If
\[ y=x^3 \]and \(x>0\), find
\[ \frac{dx}{dy} \]at \(x=2\).
Practice 9: Uncertainty Propagation
Let
\[ y=x^2. \]At \(x=10\), the input uncertainty is
\[ |\delta x|=0.02. \]Estimate the corresponding output uncertainty.
Practice 10: Identify the Error
A student says, “The derivative of this model is \(20\), so the model has sensitivity \(20\) everywhere.”
Explain when that statement is justified and when it is not.
Practice Answers
Practice 1
For
\[ f(x)=x^3, \]the derivative is
\[ f'(x)=3x^2. \]At \(x=2\),
\[ \boxed{ f'(2)=12 }. \]The absolute sensitivity is \(12\) output units per input unit.
Practice 2
At \(x=2\),
\[ f'(2)=12. \]The input change is
\[ \Delta x=0.01. \]Therefore,
\[ \Delta f \approx 12(0.01) = 0.12. \]Thus,
\[ \boxed{ \Delta f\approx0.12 }. \]Practice 3
The negative sign means increasing the input slightly tends to decrease the output.
The magnitude \(6\) means a small one-unit input increase produces approximately \(6\) output units of decrease near the operating point.
Practice 4
For
\[ f(x)=5x^4, \]we have
\[ f'(x)=20x^3. \]Therefore,
\[ S_{\text{rel}} = \frac{x}{5x^4}(20x^3) = 4. \]Thus,
\[ \boxed{ S_{\text{rel}}=4 }. \]Practice 5
Use
\[ \frac{\Delta y}{y} \approx S_{\text{rel}} \frac{\Delta x}{x}. \]Therefore,
\[ 1.5(2\%)=3\%. \]The output increases by approximately
\[ \boxed{3\%}. \]Practice 6
For
\[ A=\pi r^2, \]the absolute sensitivity is
\[ \frac{dA}{dr}=2\pi r. \]At \(r=6\),
\[ \boxed{ \frac{dA}{dr}=12\pi }. \]Since area is proportional to \(r^2\), the relative sensitivity is
\[ \boxed{2}. \]Practice 7
For
\[ y=ax^2+b, \]sensitivity to \(a\) is
\[ \boxed{ \frac{\partial y}{\partial a}=x^2 } \]and sensitivity to \(b\) is
\[ \boxed{ \frac{\partial y}{\partial b}=1 }. \]Practice 8
Since
\[ y=x^3, \]we have
\[ \frac{dy}{dx}=3x^2. \]Therefore,
\[ \frac{dx}{dy} = \frac1{3x^2}. \]At \(x=2\),
\[ \boxed{ \frac{dx}{dy}=\frac1{12} }. \]Practice 9
For
\[ y=x^2, \]the sensitivity is
\[ \frac{dy}{dx}=2x. \]At \(x=10\),
\[ \left|\frac{dy}{dx}\right|=20. \]Therefore,
\[ |\delta y| \approx 20(0.02) = 0.4. \]Thus,
\[ \boxed{ |\delta y|\approx0.4 }. \]Practice 10
The statement is justified only if the derivative is actually the constant function
\[ f'(x)=20 \]throughout the region being discussed.
If the derivative equals \(20\) only at one operating point, then the sensitivity is \(20\) only locally at that point. Sensitivity may differ elsewhere.
What to Remember
The derivative is a local sensitivity measure:
\[ \boxed{ \Delta y\approx f'(x)\Delta x }. \]The sign of \(f'(x)\) gives response direction, while \(|f'(x)|\) gives response strength.
Ordinary derivative sensitivity has units:
\[ \boxed{ S_{\text{abs}}(x)=f'(x) }. \]Relative sensitivity is dimensionless:
\[ \boxed{ S_{\text{rel}}(x) = \frac{x}{f(x)}f'(x) }. \]For small relative changes,
\[ \boxed{ \frac{dy}{y} \approx S_{\text{rel}}(x)\frac{dx}{x} }. \]Sensitivity analysis can be applied to:
- operating variables;
- model parameters;
- measurement uncertainty;
- inverse relationships;
- optimization;
- robustness and conditioning;
- small-change prediction.
The derivative answers a practical question:
\[ \boxed{ \text{How much local output response should we expect from a small input change?} }. \]Where We Go Next
Sensitivity completes the differential-analysis viewpoint developed in this section. We began with difference quotients and limits, built the derivative and its rules, extended differentiation to implicit and parametric relationships, interpreted higher derivatives, tangent and normal geometry, related rates, linearization, critical behavior, and curve structure, and finally interpreted derivatives as local response information.
These tools now provide a general language for analyzing continuous change throughout the remaining Modeling & Analysis curriculum.