An unpredictable daily sequence can still have meaningful statistical properties.
By the end: Explain why climate questions differ from exact weather forecasts.
Weather concerns the evolving atmospheric state over short periods. Climate concerns statistical patterns over longer periods, including averages, variability, seasonality, and extremes. A claim about the distribution of hot days is different from a claim about the temperature on a specific distant date.
A chaotic system can have stable statistical structure under fixed conditions even when individual trajectories diverge. This is not a guarantee that every statistic in every system is easy to predict. Some averages converge slowly, and changing forcing can change the distribution.
A temperature distribution assigns weights to possible temperatures. It does not assign dates to those temperatures. Suppose a fictional season has four equally weighted daily temperature categories, 16, 20, 24, and 28°C. Each receives probability one quarter. The mean is the sum of each value multiplied by its weight, which gives 22°C. None of the categories equals that mean. The mean describes the balance of the distribution, not a promise that a typical afternoon must have exactly that temperature. These categories are assumptions, not weather observations.
The same arithmetic applies when weights differ. Give those four temperatures weights 0.1, 0.2, 0.4, and 0.3. They sum to one, so they describe all possibilities in this invented distribution. The weighted contributions to the mean are 1.6, 4, 9.6, and 8.4°C. Their sum is 23.6°C. The probability of at least 24°C is the sum of the last two weights, or 0.7. Averaging the temperatures without their weights would give the wrong mean because it would silently replace the stated distribution with equal weights.
The four temperature categories are 16, 20, 24, 28°C; their weights exhaust this invented daily distribution.
Each term is temperature times its dimensionless weight; μ is the distribution mean.
| Temperature | Weight | Mean contribution |
|---|---|---|
| 16°C | 0.1 | 1.6°C |
| 20°C | 0.2 | 4°C |
| 24°C | 0.4 | 9.6°C |
| 28°C | 0.3 | 8.4°C |
An event definition determines which weights enter a probability. In the equal-weight example, temperatures of at least 24°C have probability one half, whereas temperatures strictly above 24°C have probability one quarter. The category exactly at the boundary explains the difference. With a continuously varying temperature, a single exact value often has zero probability, but this lesson's finite distribution deliberately places weight on individual values. Its inclusive threshold therefore matters. A numerical threshold claim needs both the units and the comparison rule before its numerator can be counted.
A mean also leaves many distributions possible. Consider a second fictional distribution with equal weights on 20, 21, 23, and 24°C. Its mean is again 22°C, but it has no values of at least 26°C. The first equal-weight distribution has one such value out of four. Equal means therefore do not imply equal heat exposure at a threshold. The difference comes from how far the values extend away from the mean. Predicting an average and predicting a rare hot category require evidence about different aspects of the distribution.
Spatial averaging introduces another distinction. Imagine two fictional locations with simultaneous temperatures of 18 and 26°C. Their equally weighted spatial average is 22°C. A person at the warmer location experiences 26°C, not that spatial average. Averaging over locations and averaging over dates answer different questions even when both calculations happen to return 22°C. A climate claim should specify whether it describes a site, a region, or a larger spatial average. The numerical size of a global average change cannot simply be copied into every local daily temperature.
| Equal-weight values | Mean | At least 26°C |
|---|---|---|
| 16, 20, 24, 28°C | 22°C | 1/4 |
| 20, 21, 23, 24°C | 22°C | 0/4 |
| 18, 20, 24, 26°C | 22°C | 1/4 |
The weather and climate distinction is thus partly a distinction between requested outputs. A daily forecast asks about a dated state or event. A climate question might ask for the mean or threshold frequency for a defined season and region under stated conditions. Both depend on atmospheric physics, but they summarize its consequences differently. A failure to identify a distant date's temperature does not, by algebra alone, prevent estimating a distribution. Whether a particular distribution estimate is credible still requires physical reasoning, observations, and tests appropriate to that statistic.
A forcing changes the energy balance or other conditions under which the system evolves. Greenhouse gases, solar changes, and volcanic aerosols can affect climate through physical mechanisms. Internal variability is variation produced within the coupled system. The two can coexist.
The lab stipulates ten equally likely daily temperatures from 18 to 27°C, then shifts every value by a chosen amount. A two-degree shift moves the mean from 22.5 to 24.5°C and creates a 20% chance of reaching at least 28°C. This is distribution arithmetic, not an estimate of real local warming or a causal climate simulation.
Adding the same amount to every category has a specific mathematical effect. If a temperature category T receives an added shift s, its new value is T plus s. Its weight stays the same. The new mean is the old mean plus s because the weights sum to one. Every distance between two categories stays unchanged. This translation changes location without changing spread. It is a useful controlled comparison, but it excludes possible changes in variability, seasonal timing, or distribution shape that a physical climate response could involve.
Use a separate invented example with equally weighted values 15, 19, 23, and 27°C. Its mean is 21°C. Add 3°C to obtain 18, 22, 26, and 30°C, with mean 24°C. At the inclusive threshold 25°C, the probability rises from one quarter to one half. At 29°C, it rises from zero to one quarter. The same mean change produces different probability changes at different thresholds because each threshold cuts through a different part of the original distribution. No universal percentage follows from a three-degree shift alone.
P(daily temperature ≥ 25°C)
Added shift, °C
You can calculate the shifted probability without listing the new values. The statement T plus s is at least a threshold H is equivalent to T being at least H minus s. Subtracting the same quantity from both sides preserves the inequality. For a 3°C shift and a 25°C threshold, count original values at least 22°C. The qualifying original values are 23 and 27°C. This change of threshold is an algebraic shortcut for a translation. It would fail if different categories changed by different amounts or if their weights changed.
T is original temperature, s the uniform shift, H the threshold; temperatures and differences are in °C. Weights do not change.
Of 15, 19, 23, 27°C, only 23 and 27°C reach 22°C. The original at-least-25°C fraction was 1/4.
Now hold the mean fixed and change the spread instead. Start with equal weights on 18, 20, 22, and 24°C, whose mean is 21°C. Multiply each departure from 21°C by two. The resulting categories are 15, 19, 23, and 27°C. Their mean remains 21°C, while the largest temperature rises. At least 25°C now has probability one quarter instead of zero. A higher tail probability does not uniquely identify a mean shift. A change in spread can also move weight across a threshold, with different consequences elsewhere.
The lab isolates translation rather than these alternatives. Its shift slider changes all ten values together in whole degrees, while the ten weights remain equal. A higher shift moves more categories across the fixed threshold, but the lab does not calculate how greenhouse gases cause that shift. Greenhouse gases alter radiation and the energy balance; the coupled atmosphere, ocean, and land respond through processes absent from a ten-number list. Treat the shift as a stipulated condition for arithmetic, not as an inferred sensitivity of the real climate system.
| Original category | Departure from 21°C | Doubled-departure value |
|---|---|---|
| 18°C | −3°C | 15°C |
| 20°C | −1°C | 19°C |
| 22°C | +1°C | 23°C |
| 24°C | +3°C | 27°C |
Forcing and internal variability also answer different causal questions. An external forcing changes conditions affecting the energy budget. Internal variability redistributes energy and produces fluctuations within the coupled system. A period can contain both a forced tendency and internal fluctuations, so a short local cooling interval does not alone measure the sign of a longer-term forced response. Conversely, a warm interval alone does not quantify a particular cause. Attribution requires comparisons and physical evidence beyond these category calculations. The toy shift deliberately removes that causal problem so its probability arithmetic stays visible.
Changing the shift replaces the selected distribution. The reference remains visible. Each of the ten values has weight 10%; no time ordering, independence between days, or real climate response is implied.
Reference mean 22.5°C; selected mean 24.5°C. At least 28°C: reference 0%; selected 20%.
| Distribution | Mean | P(temperature ≥ 28°C) |
|---|---|---|
| 18 to 27°C | 22.5°C | 0% |
| 20 to 29°C | 24.5°C | 20% |
| 22 to 31°C | 26.5°C | 40% |
The displayed temperature dots are possible daily values, not positions along a timeline. The reference uses ten values from 18 through 27°C. Each contributes one tenth of the probability. The selected distribution adds the chosen shift to every reference value. Because both rows share a temperature axis, horizontal displacement has a literal meaning in degrees Celsius. The vertical separation only keeps the rows readable. It does not represent a second physical variable, elapsed time, or differences in the probability assigned to individual dots in these rows.
Choose No shift, then move the shift slider to 1°C. This replaces the selected distribution; it does not advance a season by one day. The selected values now run from 19 through 28°C. Their mean is 23.5°C because the reference mean 22.5°C increases by one degree. Only the final value reaches the inclusive 28°C threshold, giving one tenth. The readout combines a distribution average with a threshold count. Those are two summaries of the same ten categories, not two independent predictions generated by separate models.
Mean daily temperature, °C
Stipulated shift, °C
Move the slider to 3°C for a second comparison that is not a preset button. The values become 21 through 30°C, and their mean becomes 25.5°C. Three values reach at least 28°C, so the count is three tenths. Returning to 1°C restores the previous stipulated distribution without memory of the intervening selection. These controls do not assimilate observations, estimate an emissions pathway, or advance differential equations. The retained reference row lets you compare changed assumptions directly rather than trying to remember an earlier plot after it disappears.
A useful extra calculation concerns the distance of the mean from the threshold. At shifts of 0, 1, 2, 3, and 4°C, the mean remains below 28°C. Yet some categories reach the threshold for every positive shift. A mean below a hazard threshold therefore does not imply zero chance of crossing it. For the selected shift of 3°C, the mean is 2.5°C below the threshold while the exceedance probability is 0.3. Neither number replaces the other; the distribution supports both statements without contradiction under its equal-weight assumptions.
The plotted probabilities change by equal amounts across the five available shifts because the original categories have equal spacing and equal weights near this threshold. That pattern is a property of this example. A distribution with unequal weights would have unequal jumps when categories crossed 28°C. A threshold outside the shifted support would instead give zero or one over part of the range. Extending the displayed straight relationship to larger warming values or to a real location would invent assumptions that the control does not establish about either the tail or the physical response.
| Shift | Mean below 28°C by | Weight at least 28°C |
|---|---|---|
| 0°C | 5.5°C | 0/10 |
| 1°C | 4.5°C | 1/10 |
| 3°C | 2.5°C | 3/10 |
| 4°C | 1.5°C | 4/10 |
There is no sampling uncertainty in counting the stipulated ten weights. The program knows the entire distribution it was given. That certainty about arithmetic differs from certainty about whether those weights describe reality. If the weights came from limited observations or model runs, estimation error and model limitations would matter. Ten dots do not by themselves mean ten independent observed days. The display is a finite teaching distribution, so its exact percentages should remain attached to its assumptions rather than presented as verified frequencies for any real season or site.
A climate projection depends on assumptions about future forcing, emissions, and responses. Model uncertainty and internal variability remain relevant. Greater confidence in a long-term global tendency does not imply exact knowledge of each local daily event.
The failure to predict a particular rainy day years ahead does not refute an energy-balance argument or a tested statistical projection. Conversely, a plausible physical mechanism does not validate every claimed regional magnitude. Inspect the variable, scale, scenario, and evidence.
Suppose a different fictional climate distribution assigns probability 0.3 to a defined hot daily event. Across twenty days, the expected number of hot days is six if every day's marginal probability is 0.3. To see this, assign each day an indicator equal to one when hot and zero otherwise. Each indicator has expected value 0.3. Adding the twenty expected values gives six. This calculation uses addition of expectations, not an assumption that days are independent. It also gives an average count over possible seasons, not a guaranteed count in each season.
Independence matters for questions about at least one event. Under the additional assumption of twenty independent days, the probability of no hot days is 0.7 raised to the twentieth power, about 0.000798. Subtracting that from one gives about 0.999202 for at least one hot day. Multiplication here comes from independence, which lets each non-event probability remain 0.7 after the preceding days are known. The daily probability alone did not authorize this multiplication. Weather persistence can make such an independence assumption unsuitable for a real sequence even when its marginal frequencies are correct.
N counts hot days in twenty days. Every day's marginal hot probability is 0.3; this expectation does not require independence.
This multiplication adds the assumption of independence; the marginal probability alone is insufficient.
| Assumption | Expected hot days | P(at least one hot day) |
|---|---|---|
| Independent, daily p = 0.3 | 6 | 0.999202 approximately |
| Whole block shares outcome | 6 | 0.3 |
| Marginals only specified | 6 | Not determined |
For a contrasting dependence model, stipulate that the entire twenty-day block is either hot every day with probability 0.3 or never hot with probability 0.7. Each individual day still has probability 0.3 of being hot, and the expected count is still six. But the chance of at least one hot day is now 0.3 rather than about 0.999202. These are constructed limiting comparisons, not proposed weather generators. They demonstrate why a daily climate distribution cannot determine the risk of a multi-day spell without information about the sequence's dependence.
Another source of uncertainty concerns the estimated probability itself. If a model assigns 0.3 but a plausible alternative assigns 0.2, the twenty-day expected count changes from six to four. That is a disagreement about the distribution, not ordinary variation of outcomes within one known distribution. More observed days can reduce some sampling uncertainty if they provide relevant information. They do not automatically remove a shared model bias or make old observations representative of changed conditions. A long record needs interpretation alongside the assumptions about what remained comparable during its collection.
Conditioning makes a projection statement precise. A projection may say what a distribution would look like under specified future forcing and a particular model response. Different forcing assumptions can produce different conditional distributions without the calculation contradicting itself. The uncertainty about which conditions will occur is separate from variation within each condition. Combining scenarios into one probability distribution would require justified scenario weights, which this lab never supplies. Equal-weight temperature categories inside one toy distribution do not establish equal probabilities for different future emissions choices or for different models of the climate response.
| Specified quantity | Constructed comparison | Consequence |
|---|---|---|
| Daily event probability | 0.2 versus 0.3 | Expected twenty-day count 4 versus 6 |
| Dependence at p = 0.3 | Independent versus block | Any-event probability changes |
| Scenario weights | Unspecified | No combined scenario probability |
The evidence needed to assess a claim depends on its target. A regional seasonal average calls for suitable spatial and seasonal comparisons. A multi-day heat-risk claim also needs information about persistence and the relevant threshold. A causal warming claim needs physical and attribution evidence rather than only a shifted mean in a diagram. These distinctions leave room for meaningful statistical climate predictions despite unpredictable daily details, without granting equal confidence to every statistic. The ten-dot lab establishes counting and translation identities only; it cannot certify any particular regional projection or observed trend.
If the shifted distribution gives 20% hot days, can you name the first hot day next year?
A frequency does not specify an ordering.
| Ten-day sequence | Hot-day count |
|---|---|
| Hot on days 1 and 2 | 2 |
| Hot on days 4 and 9 | 2 |
| Either ordering | Same count, different persistence |
No. The distribution constrains frequencies, not the sequence of dates. Dependence between days also affects runs of heat; independent draws were not established by the marginal distribution alone.
In the toy distribution, how does a four-degree shift affect the probability of at least 30°C?
List the shifted values, then count inclusively.
| Distribution | Qualifying values | Probability |
|---|---|---|
| 18 to 27°C | None | 0% |
| 22 to 31°C | 30, 31°C | 20% |
The shifted values run from 22 to 31°C. Values 30 and 31 qualify, so the probability is 2/10 = 20%, versus zero in the reference. A mean shift can change a threshold frequency; its magnitude depends on the whole stipulated distribution.
Selected reference passages support the definitions and modeling distinctions below. The examples, explanations, and implementations are original. Terminology and assumptions are stated within each lesson; source review is not an independent audit of this course. Optional links are not runtime dependencies.
Weather versus long-term climate patterns, forcing, and variability. The shifted ten-value temperature distribution is illustrative, not a climate projection.
Atmospheric dynamics, physical processes, model limitations, and probabilistic forecasting. No current operational resolution or skill claim is inferred.