Why Calibrate From Cumulative Measurements?
A rate describes how quickly a quantity changes at a particular time. A cumulative measurement records how much has accumulated over an interval. In many applications, totals are easier to observe than instantaneous rates: a meter may report total energy used, a gauge may record collected volume, or an account may show total revenue.
If a model proposes a rate function with unknown parameters, each observed cumulative total gives an equation involving the integral of that rate. Calibrating from cumulative measurements means choosing parameter values so that the model's predicted totals agree with the observations.
The distinction matters. A total over an interval constrains the net accumulation across that interval. It does not, by itself, specify the rate at every instant within it.
From a Rate Model to Cumulative Predictions
Let \(r(t;\boldsymbol{\theta})\) be a rate model, where \(t\) is time and \(\boldsymbol{\theta}\) represents one or more unknown parameters. Suppose the accumulated quantity at the starting time \(t=0\) is \(C_0\). If the rate is continuous on the interval from \(0\) to \(t\), the model predicts the cumulative quantity
The integration variable \(s\) distinguishes the variable being integrated from the endpoint \(t\). If observations give cumulative values \(C_i\) at times \(t_i\), the calibration equations are
These equations may be linear or nonlinear in the unknown parameters. The form depends on the chosen rate model. Before solving them, check that the rate, cumulative quantity, and time units are compatible. For example, a rate in liters per day integrates over days to give liters.
Worked Example: A Rate That Increases Linearly
Worked Example: A Rate That Increases Linearly
A tank initially contains \(10\) liters. Suppose its inflow rate, in liters per day, is modeled by \(r(t)=a+bt\), where \(a\) is measured in liters per day and \(b\) in liters per day squared. Measurements show \(22\) liters at day \(2\) and \(55\) liters at day \(5\). Find \(a\) and \(b\).
The cumulative prediction is
At \(t=2\), the predicted amount must equal \(22\):
Subtract \(10\) and divide by \(2\), giving \(a+b=6\). At \(t=5\), the second observation gives
so \(5a+\frac{25}{2}b=45\). From the first equation, \(a=6-b\). Substitute this into the second:
Then \(a=6-2=4\). Thus the fitted rate is \(r(t)=4+2t\) liters per day. Check both observations using the cumulative formula:
Both totals are reproduced. The parameters also have the expected units: \(a=4\) liters per day and \(b=2\) liters per day squared.
Why Differences Between Totals Are Useful
Subtracting two cumulative measurements removes the accumulation that occurred before the earlier measurement. For any \(t_j>t_i\), the model gives
This follows from additivity of definite integrals: the integral from \(0\) to \(t_j\) is the sum of the integral from \(0\) to \(t_i\) and the integral from \(t_i\) to \(t_j\). Subtracting the first part leaves only accumulation over the later interval. This is useful when the initial cumulative amount is not central to the calibration or when interval totals are easier to interpret.
The first identity explains what cumulative data constrain: the integral of the rate over the observed interval. The derivative identity connects a complete cumulative curve to its instantaneous rate, but a few isolated cumulative observations are not a complete curve.
Worked Example: An Exponentially Decreasing Rate
Worked Example: An Exponentially Decreasing Rate
A process starts with cumulative total \(C_0=0\). Model its rate by \(r(t)=Ae^{-kt}\), where \(A\) is in units per day and \(k\) is in inverse days. The measured totals are \(C(1)=4\) and \(C(3)=7\). Find \(A\) and \(k\), taking time in days.
For \(k\ne 0\), integrating gives
Use the ratio of the two observed totals to eliminate \(A/k\):
Set \(x=e^{-k}\). For a positive decay constant, \(0<x<1\), and \(1-e^{-3k}=1-x^3=(1-x)(1+x+x^2)\). Since \(1-x\ne0\), the ratio equation becomes
Factoring gives \((2x-1)(2x+3)=0\), so the roots are \(x=\frac12\) and \(x=-\frac32\). Only \(x=\frac12\) is possible because \(x=e^{-k}>0\). Therefore \(e^{-k}=\frac12\), so \(k=\ln 2\) per day.
Now use \(C(1)=4\):
Thus \(A=8\ln 2\) units per day and \(k=\ln 2\) per day. Check the observations by substitution:
The fitted parameters reproduce both totals exactly. The condition \(k\ne0\) was required for the displayed integrated form; the fitted value \(k=\ln 2\) satisfies it.
How Many Measurements Are Enough?
Each cumulative observation supplies an equation, but the number of equations alone does not guarantee that the parameters are uniquely determined. The equations must also distinguish the parameters from one another. If two different parameter choices produce the same observed totals, those measurements cannot decide between them.
For a model with several unknown parameters, calibration therefore involves two questions:
- Do the observations provide enough independent constraints to determine the parameters?
- Does solving the resulting equations produce parameter values that satisfy all the observations and the model's assumptions?
When the equations have no solution, the selected model and exact data cannot all hold simultaneously. When there are multiple solutions, the measurements do not select a unique parameter set. In either case, do not report a single fitted parameter set as if it were uniquely established.
Worked Example: Using Interval Totals to Fit a Quadratic Rate
Worked Example: Using Interval Totals to Fit a Quadratic Rate
A system has initial cumulative amount \(5\) units. Its rate is modeled as \(r(t)=a+bt^2\), with time measured in hours. Measurements give \(C(1)=8\) and \(C(2)=17\). Find \(a\) and \(b\), then interpret the increase from hour \(1\) to hour \(2\).
The cumulative model is
The observation at \(t=1\) gives \(8=5+a+\frac{b}{3}\), or \(a+\frac{b}{3}=3\). At \(t=2\), \(17=5+2a+\frac{8b}{3}\), or \(2a+\frac{8b}{3}=12\). Multiply the first equation by \(3\) and the second by \(3\):
Twice the first equation is \(6a+2b=18\). Subtract it from the second equation to get \(6b=18\), hence \(b=3\). Substitution into \(3a+b=9\) gives \(3a+3=9\), so \(a=2\).
Check the cumulative predictions:
The increase over the interval from hour \(1\) to hour \(2\) is \(17-8=9\) units. The interval-integral identity verifies its meaning:
The interval increase is an integral constraint on the rate over that hour, not a measurement of the rate at one particular time.
What Cumulative Data Can and Cannot Say
An integral combines rate values across time. Different rate functions can therefore have the same integral over an interval, and a finite collection of cumulative totals cannot generally recover every detail of an unrestricted rate function. Calibration is possible because we restrict attention to a specified family of rate functions, such as a linear or exponential family, and then estimate that family's parameters.
There is an important pitfall: treating a cumulative total as if it were an instantaneous rate. For instance, in the first example the amount at day \(2\) was \(22\) liters. It would be incorrect to call \(22\) liters per day the rate at day \(2\), or to set \(r(2)=22\). The units already show the problem: \(22\) liters is a cumulative amount, while \(r(2)\) must have units of liters per day. The correct equation is
which compares like quantities and uses the observed total to constrain the rate over the whole interval.
For a reliable calibration, write the cumulative model first, form one equation for each observed total, solve for the parameters, and substitute the result back into every observation. Keep the initial amount in the equation unless the data are already expressed as changes from that initial amount. Finally, interpret the answer in its units and remember that agreement with the totals establishes agreement at those measurement times—not a direct measurement of the rate at every intermediate time.
Check Your Understanding
Use the integral relationship between rate and cumulative amount to answer these questions.
- If \(r(t)=3+2t\) units per hour and \(C(0)=4\) units, write the model prediction \(C(t)\).
- What equation does an observation \(C(3)=20\) impose on an unknown parameterized rate \(r(t;\boldsymbol{\theta})\), when \(C(0)=2\)?
- Why does the difference \(C(t_2)-C(t_1)\) constrain accumulation over an interval rather than give the rate at one instant?
- A model has three unknown parameters and three cumulative observations. Is that alone enough to guarantee a unique calibration? Explain.
- In what units should \(r(t)\) be reported if \(C(t)\) is measured in kilograms and time in days?
