Why Normalization Matters
A probability recursion can generate a sequence of nonnegative numbers without making those numbers a probability distribution. For a frequency model, the probabilities of all possible claim counts must add to one. The recursion supplies their relative sizes; normalization determines whether a valid total probability can be assigned.
The key questions are therefore distinct: does the recurrence produce nonnegative terms, and does the sum of those terms converge? Even when the sum is finite, a proposed starting value \(p_0\) gives a valid distribution only if it makes the total exactly one.
From Recursion to a Normalizing Constant
The product expression follows by applying the recurrence repeatedly: \(p_1=p_0(a+b)\), \(p_2=p_0(a+b/2)(a+b)\), and each next probability adds one more factor. Factoring out \(p_0\) leaves weights whose sum controls the total probability:
Thus, if \(W\) is finite and positive, choosing \(p_0=1/W\) makes the total equal to one. If \(W=\infty\), no positive finite choice of \(p_0\) can normalize the sequence: the total is still infinite. If \(p_0=0\), the recursion gives only zero probabilities, not a distribution.
A zero recurrence factor needs careful interpretation. If \(a+b/m=0\) at an integer \(m\), and every factor before it is positive, then \(w_m=0\). All later weights are also zero because the recurrence multiplies the preceding zero. The sequence has finite support and its sum is a finite sum. Factors after termination do not create probabilities beyond the zero term.
Starting from \(w_0=1\), inspect the factors \(a+b/n\). Terms must be nonnegative; a first zero can terminate the sequence if the preceding weights are nonnegative.
For an infinite sequence, determine whether \(\sum w_n\) converges. The limiting ratio \(w_n/w_{n-1}=a+b/n\) is useful except at its boundary value \(a=1\), which needs a separate test.
If the weights are nonnegative and sum to a finite positive \(W\), use \(p_0=1/W\), then verify \(p_0W=1\).
Convergence and Parameter Conditions
For infinite support, the ratio of successive weights is \(w_n/w_{n-1}=a+b/n\). When \(a\) is positive and not equal to one, this ratio tends to \(a\). If \(0<a<1\), it is eventually bounded above by some number strictly less than one, so the tail is bounded by a convergent geometric series. If \(a>1\), the ratios are eventually greater than one, so the weights do not tend to zero and their sum diverges.
The boundary case \(a=1\) cannot be decided from the limiting ratio alone. If the factors stay positive, then $$ w_n=\prod_{k=1}^{n}\left(1+\frac{b}{k}\right). $$ Taking logarithms gives \(\log w_n=\sum_{k=1}^{n}\log(1+b/k)\). For large \(k\), \(\log(1+b/k)=b/k+O(1/k^2)\). Since the sum of \(1/k^2\) converges and \(\sum_{k=1}^{n}1/k\) grows like \(\log n\), it follows that \(w_n\) is proportional in order to \(n^b\). The series of these weights converges only if \(b<-1\). But for \(a=1\), the first factor is \(1+b\); if \(b<-1\), that factor is negative. Thus the admissible, non-terminating \(a=1\) case cannot be normalized. If \(b=-1\), the first factor is zero and the distribution terminates at zero.
For \(a=0\), the weights are \(w_n=b^n/n!\). If \(b\geq 0\), their sum is finite; if \(b<0\), the first weight after \(w_0\) is negative and cannot be a probability weight. If \(a<0\), the factors eventually become negative as \(n\) increases, so a valid sequence with positive weights cannot continue indefinitely. It can, however, terminate at a zero factor. In particular, for \(a<0\) and \(b=-am\), where \(m\) is a positive integer, the factors are positive for \(n<m\) and zero at \(n=m\), producing finite support.
These tests give useful general conclusions, but the actual sequence still needs a nonnegativity check. For example, with \(a>0\), a zero at the first factor (\(b=-a\)) gives the degenerate distribution concentrated at zero. A proposed parameter pair should be assessed through its generated terms, not just by applying a convergence rule to a sequence that already contains a negative probability.
Worked Applications
Worked Example: Normalize a Recursion with \(a=0\)
A hypothetical claim count follows the \((a,b,0)\) recursion with \(a=0\) and \(b=3\). Determine the required \(p_0\), and calculate \(p_1,p_2,p_3\). Give the probabilities to four decimal places.
Method. First generate the weights, then sum them to find the normalizing constant. This avoids treating an arbitrarily chosen starting probability as automatically valid.
Here \(w_0=1\), and the recurrence factor at step \(n\) is \(3/n\). Therefore $$ w_1=3,\qquad w_2=3\left(\frac{3}{2}\right)=\frac{9}{2},\qquad w_3=\frac{9}{2}\left(\frac{3}{3}\right)=\frac{9}{2}. $$ In general, \(w_n=3^n/n!\). The sum of these nonnegative weights is \(e^3\), so $$ p_0=\frac{1}{e^3}=0.0498 $$ to four decimal places.
The requested probabilities are $$ p_1=p_0w_1=\frac{3}{e^3}=0.1494,\qquad p_2=p_0w_2=\frac{9}{2e^3}=0.2240,\qquad p_3=p_0w_3=\frac{9}{2e^3}=0.2240. $$
Check. The complete total is \(p_0\sum_{n=0}^{\infty}w_n=e^{-3}e^3=1\). The rounded displayed probabilities are not intended to sum to exactly one by themselves. The mistake to avoid is adding only \(p_0\) through \(p_3\) and calling that partial sum the normalization.
Worked Example: Normalize a Finite Sequence
A model has \(a=-1/2\) and \(b=2\). Starting from \(w_0=1\), determine whether the recurrence terminates, find the normalizing \(p_0\), and give all positive probabilities.
Method. Generate factors in order. When a factor is zero, the recursion terminates; sum the weights through the last positive term. This prevents extending a finite-support model as if it had an infinite tail.
The factors and weights are $$ \begin{aligned} n=1:&\quad -\frac12+2=\frac32,\quad w_1=\frac32,\\ n=2:&\quad -\frac12+1=\frac12,\quad w_2=\frac34,\\ n=3:&\quad -\frac12+\frac23=\frac16,\quad w_3=\frac18,\\ n=4:&\quad -\frac12+\frac12=0,\quad w_4=0. \end{aligned} $$ Every later weight is zero. The sum is $$ W=1+\frac32+\frac34+\frac18=\frac{27}{8}. $$ Therefore \(p_0=8/27\), and the nonzero probabilities are $$ p_0=\frac{8}{27},\quad p_1=\frac{4}{9},\quad p_2=\frac{2}{9},\quad p_3=\frac{1}{27}. $$
Check. Their sum is \((8+12+6+1)/27=1\), and the next recurrence factor is zero. The common mistake is to stop at \(p_3\) without checking the factor at \(n=4\); that check establishes that there is no further positive mass.
Worked Example: Detect a Proposed Starting Probability That Fails
A hypothetical frequency model uses \(a=0.6\), \(b=0\), and a proposed value \(p_0=0.30\). Determine whether this gives a probability distribution. If not, find the correct \(p_0\), and calculate the correct \(p_3\).
Method. With \(b=0\), each recurrence step multiplies by the same ratio. Sum the resulting geometric series and compare the proposed total with one.
The recurrence gives \(p_n=0.30(0.6)^n\). Since \(0.6<1\), $$ \sum_{n=0}^{\infty}p_n =0.30\sum_{n=0}^{\infty}(0.6)^n =0.30\left(\frac{1}{1-0.6}\right) =0.30(2.5)=0.75. $$ The proposed value is not normalized. The weights sum to \(W=2.5\), so the correct initial probability is \(p_0=1/2.5=0.40\).
With this corrected value, $$ p_3=0.40(0.6)^3=0.40(0.216)=0.0864. $$ Check. The full total is \(0.40(2.5)=1\), and \(p_3\) follows three recurrence steps from \(p_0\). The error to avoid is seeing that the terms decrease and assuming that their total is one; decreasing terms can sum to a value other than one.
Common Traps and Practical Checks
Normalization is a statement about the entire support, not about a few computed probabilities. A finite partial sum below one is expected whenever there is positive probability beyond the last term calculated. Conversely, a partial sum close to one is not proof of exact normalization.
- Confusing weights with probabilities. The recurrence weights start at \(w_0=1\); they become probabilities only after multiplication by the correctly chosen \(p_0\).
- Applying the ratio test at \(a=1\). A limiting ratio of one is inconclusive. The polynomial-order behavior of the weights must be checked, and the admissible non-terminating case does not have a finite total.
- Ignoring negative factors. A convergent formal series is not a probability distribution if any generated term is negative.
- Forgetting termination. A zero factor makes the current and all subsequent weights zero. Check for it before using an infinite-support conclusion.
- Rounding too early. Determine the normalizing constant from the exact sum or a justified infinite-series result, then round the requested probabilities at the end.
Check Your Understanding
For each problem, check both admissibility of the generated terms and normalization of the total.
- For \(a=0\) and \(b=2\), find the value of \(p_0\) that normalizes the recurrence. (Answer: \(p_0=e^{-2}\approx 0.1353\).)
- For \(a=0.4\), \(b=0\), and proposed \(p_0=0.50\), determine the total mass and the correct \(p_0\). (Answer: proposed total \(=0.50/(1-0.4)=0.8333\); correct \(p_0=0.60\).)
- For \(a=-1/3\), \(b=4/3\), find the support and normalizing \(p_0\). (Answer: support \(0,1,2,3\); weights \(1,1,1/3,1/27\); \(W=64/27\), so \(p_0=27/64\).)
- For \(a=1\), \(b=-0.8\), can a positive \(p_0\) normalize the generated sequence? (Answer: No. The terms remain positive and have order \(n^{-0.8}\), whose series diverges.)
- For \(a=0.7\), \(b=0.2\), do the weights have a finite sum, and what condition determines the normalizing \(p_0\)? (Answer: Yes; the ratios tend to \(0.7<1\), so the sum is finite. Set \(p_0=1/\sum_{n=0}^{\infty}w_n\), with \(w_0=1\) and \(w_n=\prod_{k=1}^{n}(0.7+0.2/k)\).)