The inverse normal distribution calculator above runs the normal distribution backwards. Instead of taking a value and returning the probability below it, it takes a probability and returns the value that sits at it. This is the percent point function, also called the quantile function or the inverse cumulative distribution function, and it is what you need whenever the question starts with a proportion — the top 10%, the middle 95%, the cut-off that leaves 1% in the tail — rather than with a measurement.
Arb Digital publishes this alongside the forward tool because the two questions arrive in different work. Forward: "how unusual is this reading?" Backward: "what reading would be unusual?" The second is the one that sets thresholds, specification limits, service-level targets and stock reorder points, and it is the harder of the two to do by hand because the normal CDF has no closed-form inverse. It has to be computed numerically, which is precisely why a tool is worth having.
What This Inverse Normal Distribution Calculator Does
You supply a probability, a mean and a standard deviation, and it returns the x value at that point on the distribution, together with the underlying z-score. Three interpretations of p are available: area to the left, area to the right, and a central interval that splits the remaining area evenly between the two tails and gives you both endpoints. It also draws the tail split as bars so it is obvious which region you actually asked for.
Two neighbouring tools on the site are worth separating from this one. Our normal distribution calculator goes the other way — value in, probability out — and if you already have an x and want to know the area below it, that is the page you want. Our critical value calculator looks superficially similar but does a narrower and different job: it returns hypothesis-testing cut-offs on standard reference distributions, covering z, t, chi-square and F from a significance level and degrees of freedom. This page only handles the normal, but it handles it on any mean and standard deviation, which is what you need for a real measured quantity rather than a test statistic. If your question is "what t value do I compare against at alpha 0.05 with 12 degrees of freedom", use the critical value tool. If it is "what score does the top 5% of this population start at", use this one.
How to Use It
- Enter p as a decimal. 0.9 rather than 90. The tool works strictly inside 0 and 1, since the normal distribution is unbounded and the endpoints correspond to infinity.
- Say which area p describes. Getting left and right the wrong way round is the single most common error here and it reflects the answer across the mean.
- Set the mean and standard deviation of your actual quantity. Leave them at 0 and 1 only if you genuinely want a z value.
- Read the z-score as well as the x. The z is portable: it applies to any normal distribution with the same tail probability, so it is the number worth remembering.
- Sanity-check against the empirical rule. A central 95% interval should come out close to two standard deviations either side of the mean; if it does not, an input is wrong.
The Formula and How It's Calculated
The calculation happens in two steps. First the standard normal quantile z is found such that Φ(z) = p, where Φ is the standard normal cumulative distribution function. Then it is rescaled to your distribution with x = μ + z·σ. That second step is trivial; the first is not. As the NIST/SEMATECH e-Handbook page on the normal distribution states plainly, the percent point function of the normal distribution has no simple closed form and must be computed numerically.
This tool uses a rational approximation for the quantile followed by Newton refinement steps against a high-accuracy complementary error function, which brings the result to the limit of double-precision display well beyond the decimals shown. That is why the familiar textbook values come out exactly as expected: p = 0.975 gives z = 1.959964, p = 0.95 gives 1.644854, and p = 0.99 gives 2.326348.
Run the default. With p = 0.9 on the left tail, the standard normal quantile is z = 1.281552. Multiply by σ = 15 to get 19.223, then add μ = 100, and the answer is x = 119.22. In words: on a normal distribution centred at 100 with a spread of 15, ninety percent of the population falls below about 119.2 and ten percent above it. Switch to a central interval at p = 0.95 and the z becomes ±1.959964, giving endpoints of 70.60 and 129.40 — a width of 58.8, or very nearly four standard deviations.
Left Tail, Right Tail and Central: Three Different Questions
The same p produces three different answers depending on which region it names, and mixing them up is the fastest route to a wrong threshold. With p = 0.95 on the defaults: left tail gives x = 124.67, meaning 95% of values fall below it. Right tail gives x = 75.33, meaning 95% of values fall above it. Central gives the pair 70.60 to 129.40, containing the middle 95% with 2.5% left in each tail.
The relationship between the first two is a reflection: the right-tail answer for p is the left-tail answer for 1 − p, mirrored about the mean. The central case is not a reflection of either — its endpoints come from the quantiles at (1 − p)/2 and (1 + p)/2, so the 95% central interval uses the 97.5th percentile rather than the 95th. That factor-of-two-in-the-tail is exactly the same distinction as one-tailed versus two-tailed testing, and it is why a 95% confidence interval uses 1.96 while a one-sided 95% bound uses 1.645.
Where the Backwards Question Actually Comes Up
Setting a threshold is the classic use. If delivery times are roughly normal with a mean of 4.2 days and a standard deviation of 0.8, and you want to publish a promise you will meet 95% of the time, the answer is the 95th percentile: 4.2 + 1.645 × 0.8 = 5.5 days. Promising the mean would leave you failing almost half of all orders, which is why "average delivery time" makes such a poor service commitment.
Inventory works the same way. Safety stock is set by choosing a service level, converting it to a z with this calculation, and multiplying by the standard deviation of demand over the lead time. Moving from a 95% service level to 99% raises z from 1.645 to 2.326 — a 41% increase in safety stock for four percentage points of service. The non-linearity of the normal tail is the entire reason very high service levels are expensive, and seeing the z values side by side makes that concrete in a way a percentage target does not. Manufacturing tolerances, grading curves, and value-at-risk thresholds in finance all follow the same pattern: a chosen probability, converted to a quantile, scaled by a spread.
The Assumption That Does the Damage
Everything above depends on the quantity actually being normally distributed, and the further into the tail you go, the more that assumption matters. Near the middle of the distribution, a mildly skewed or heavy-tailed variable still gives roughly the right answer — the 60th percentile is not very sensitive to shape. Out at the 99th or 99.9th percentile, the shape is doing almost all the work, and a distribution with fatter tails than the normal will produce far more extreme values than this calculation predicts.
Delivery times, incomes, insurance claims and financial returns are all commonly right-skewed or heavy-tailed, and all of them are routinely modelled as normal because it is convenient. The practical safeguard is to compare the calculated quantile against the empirical one: if you have several hundred observations, sort them and look at the actual value at that percentile. If the two disagree materially, the normal model is the problem, not the arithmetic. Penn State's open STAT 500 applied statistics course notes cover the normality checks worth running before trusting a tail quantile. Our percentile calculator does the empirical version directly from a data set, with no distributional assumption at all.
Why p Cannot Be 0 or 1
The normal distribution extends to infinity in both directions, so there is no finite value below which zero percent of the population falls, and none above which one hundred percent falls. Ask for p = 1 and the honest answer is positive infinity. The calculator refuses those endpoints rather than returning a large finite number that would look like a real threshold.
This has a practical edge to it. A "100% service level" is not achievable under a normal demand model at any finite stock level, and a specification limit that promises zero defects on a normally distributed characteristic is promising something the model cannot deliver. What can be quoted is a probability arbitrarily close to one — 99.99% corresponds to z = 3.719, and 99.9999% to z = 4.753. Each additional nine costs more standard deviations than the last, which is the whole reason six-sigma programmes are framed in sigma counts rather than percentages. Our z-score calculator and standard deviation calculator cover the supporting pieces of that arithmetic.
Arb Digital's analysts turn campaign and site data into thresholds you can actually commit to, with the distribution checked first.
Browse All Free Tools Talk To Our TeamCommon Mistakes to Avoid
- Entering p as a percentage — typing 95 instead of 0.95 puts the value outside the valid range, since a probability cannot exceed one.
- Confusing left tail with right tail — the two answers sit on opposite sides of the mean, and a threshold set the wrong way round is worse than no threshold.
- Using the 95th percentile for a 95% central interval — a two-sided interval needs the 97.5th percentile, because 2.5% is left in each tail.
- Trusting extreme quantiles on non-normal data — the further into the tail you go, the more the answer depends on the shape assumption rather than on the mean and spread.
- Feeding in a sample standard deviation as if it were known — with a small sample the spread is itself uncertain, and a t-based interval is the more honest instrument.
Related Free Tools From Arb Digital
Go the forward direction with the normal distribution calculator, find hypothesis-testing cut-offs with the critical value calculator, standardise a single observation with the z-score calculator, read percentiles straight from data with the percentile calculator, or check spread with the standard deviation calculator. The free online tools hub lists every statistics tool we publish.
Frequently Asked Questions
It is the function that takes a cumulative probability and returns the value at that point on a normal distribution. It is also called the percent point function, the quantile function, or the inverse cumulative distribution function.
The forward tool takes a value and returns the probability below it. This one takes a probability and returns the value. They answer opposite questions on the same curve.
A critical value calculator returns hypothesis-testing cut-offs on standard reference distributions such as z, t, chi-square and F. This page handles only the normal, but on any mean and standard deviation, which is what a real measured quantity needs.
Because the normal cumulative distribution function has no elementary closed form, so neither does its inverse. Values are produced by numerical approximation, which is why tables and software exist for it.
1.644854 if the 95 percent is all in one tail, and 1.959964 if it is a central interval with 2.5 percent left in each tail. Choosing the wrong one is a frequent source of error.
Because the normal distribution has no finite endpoints. A probability of exactly one corresponds to positive infinity, so the calculator restricts p to values strictly between zero and one.
Only approximately, and the error grows as you move into the tails. For skewed or heavy-tailed data, compare the calculated quantile against the empirical percentile from your actual observations before relying on it.
This page explains a statistical calculation for educational purposes only. It is not financial, medical or operational advice, and every result assumes the quantity you are modelling really does follow a normal distribution.
