Reliability and Risk | Malone Math | Malone Global University

The Central Idea

Suppose a model produces an uncertain outcome represented by a random variable $Y$. Reliability and risk questions begin by identifying which outcomes are acceptable and which count as failure, loss, or harm.

If failure occurs whenever $Y$ exceeds a critical threshold $c$, then the failure probability is

\[ \boxed{P(Y>c)}. \]

If the consequence of each outcome is represented by a loss function $L(Y)$, then one important risk summary is expected loss:

\[ \boxed{E[L(Y)]}. \]

Reliability and risk therefore add a decision layer to uncertainty analysis. The uncertainty model describes what may happen; reliability and risk ask what those possibilities mean for performance and consequences.

What Is Reliability?

Reliability is the probability that a component, process, or system performs its required function under stated conditions for a specified period of time or number of uses.

If $T$ denotes the time until failure, then reliability over a time horizon $t$ is commonly written

\[ \boxed{R(t)=P(T>t)}. \]

This is the probability that the system survives beyond time $t$.

Reliability is always conditional on a model of operating conditions. A device may be highly reliable under normal temperature and load but much less reliable under extreme heat, vibration, overload, or poor maintenance.

Failure-Time Distributions

Let $T$ be a nonnegative random variable representing time to failure. Its cumulative distribution function is

\[ F(t)=P(T\le t). \]

Since the system either fails by time $t$ or survives beyond time $t$,

\[ P(T\le t)+P(T>t)=1. \]

Therefore,

\[ \boxed{R(t)=1-F(t)}. \]

The failure distribution and the reliability function contain the same information viewed from opposite directions.

Example: Reliability from a Failure Distribution

Suppose a component has

\[ P(T\le 1000)=0.08. \]

Then its probability of surviving beyond 1000 hours is

\[ R(1000)=1-0.08=\boxed{0.92}. \]

The model therefore assigns a 92% reliability over the first 1000 hours under the assumed operating conditions.

Hazard Rate

Reliability answers the question, “What is the probability of surviving to a given time?” A related question is, “Given survival up to the present, how intense is the immediate risk of failure?”

For a continuous failure-time model with density $f(t)$, the hazard rate is

\[ \boxed{h(t)=\frac{f(t)}{R(t)}}. \]

The hazard is not itself a probability. It is an instantaneous failure rate conditional on survival up to time $t$.

How to Interpret Hazard

A rising hazard indicates that failure becomes more likely as the system ages or accumulates stress. A roughly constant hazard is associated with a process whose immediate failure tendency does not depend strongly on age. A decreasing hazard can appear when early defects are removed and the surviving population becomes more robust.

Real reliability systems often exhibit different hazard behavior in different life stages. The mathematical form should therefore be chosen from evidence, not from convenience alone.

Constant Hazard and the Exponential Model

A common introductory reliability model assumes a constant hazard rate $\lambda>0$. In that case,

\[ \boxed{R(t)=e^{-\lambda t}}. \]

The corresponding failure-time density is

\[ f(t)=\lambda e^{-\lambda t}. \]

The mean failure time is

\[ E[T]=\frac{1}{\lambda}. \]

This model is useful when a constant hazard assumption is plausible, but it can be misleading for systems dominated by wear, aging, fatigue, or changing operating conditions.

Example: Constant-Hazard Reliability

Suppose a component has constant hazard

\[ \lambda=0.002\text{ per hour}. \]

Its probability of surviving 100 hours is

\[ R(100)=e^{-0.002(100)}=e^{-0.2}\approx\boxed{0.8187}. \]

The model therefore predicts about an 81.9% chance of surviving through the first 100 hours.

Mean Time to Failure

The mean time to failure, or MTTF, is the expected lifetime of a nonrepairable item:

\[ \boxed{\operatorname{MTTF}=E[T]}. \]

For a nonnegative lifetime distribution,

\[ E[T]=\int_0^\infty R(t)\,dt \]

when the integral exists.

MTTF is a useful summary, but it does not describe the full distribution. Two systems can have the same mean lifetime while having very different early- failure probabilities and tail behavior.

Failure Probability over a Fixed Mission

Many applications do not require a complete lifetime model. Instead, the decision concerns a fixed mission interval, production run, flight, project, transaction, or operating cycle.

If $F$ denotes the event that the system fails during the mission, then

\[ p_f=P(F) \]

is the probability of failure, and

\[ \boxed{R=1-p_f} \]

is mission reliability.

Series Systems: Every Component Must Work

A series system succeeds only if every required component succeeds. If component reliabilities are $R_1,R_2,\ldots,R_n$ and component failures are independent, then

\[ \boxed{R_{\text{series}}=\prod_{i=1}^{n}R_i}. \]

For three independent components with reliabilities $0.98$, $0.97$, and $0.99$,

\[ R_{\text{series}}=(0.98)(0.97)(0.99)\approx\boxed{0.9411}. \]

Even highly reliable components can produce a noticeably lower system reliability when all of them are required.

Parallel Systems and Redundancy

A simple parallel system succeeds if at least one component works. For independent components with reliabilities $R_1,R_2,\ldots,R_n$, the system fails only if every component fails.

Therefore,

\[ \boxed{R_{\text{parallel}}=1-\prod_{i=1}^{n}(1-R_i)}. \]

If two independent components each have reliability $0.90$, then

\[ R_{\text{parallel}}=1-(0.10)(0.10)=\boxed{0.99}. \]

Redundancy can greatly improve reliability when failures are sufficiently independent.

Redundancy Does Not Eliminate Common-Cause Failure

The parallel-system formula can overstate reliability if components share a common failure mechanism. Two backup generators may both fail during the same flood. Two servers may share the same power supply. Two suppliers may depend on the same upstream factory.

Dependence matters because redundancy protects against independent failures much better than against common-cause failures. A reliability model should represent shared vulnerabilities explicitly whenever they are material.

Systems That Require Some, but Not All, Components

Some systems work as long as at least $k$ of $n$ components succeed. Such systems are called $k$-out-of-$n$ systems.

If all $n$ components are independent and have common reliability $p$, then

\[ \boxed{R=P(X\ge k)=\sum_{j=k}^{n}\binom{n}{j}p^j(1-p)^{n-j}}, \]

where $X\sim\operatorname{Binomial}(n,p)$.

This connects reliability directly to the discrete probability models studied earlier.

Fault Trees and Logical Structure

Complex system failures can often be decomposed into combinations of simpler events. A fault tree represents these logical relationships.

An AND gate means all listed events must occur to produce the higher-level failure.
An OR gate means any listed event can produce the higher-level failure.

The logic resembles set operations in probability. If system failure occurs when either event $A$ or event $B$ occurs, then

\[ P(A\cup B)=P(A)+P(B)-P(A\cap B). \]

The intersection term is essential when events are dependent.

What Is Risk?

Risk concerns uncertain outcomes that carry consequences. Informally, risk is often described as probability times consequence:

\[ \text{risk}\approx\text{probability}\times\text{consequence}. \]

This expression is useful for intuition, but it is not a universal definition. Real risk analysis may need to represent several possible outcomes, nonlinear consequences, asymmetric preferences, tail events, and uncertainty about the probabilities themselves.

Expected Loss

Suppose outcome $i$ occurs with probability $p_i$ and produces loss $L_i$. The expected loss is

\[ \boxed{E[L]=\sum_i p_iL_i}. \]

For a continuous uncertain outcome $Y$ with density $f(y)$,

\[ E[L(Y)]=\int L(y)f(y)\,dy. \]

Expected loss converts a distribution of consequences into an average loss per decision or exposure under repeated comparable conditions.

Example: Probability and Consequence

Suppose a process has a 2% probability of causing a $\$500{,}000$ loss and a 98% probability of no loss. Its expected loss is

\[ E[L]=(0.02)(500{,}000)+(0.98)(0). \]

Therefore,

\[ \boxed{E[L]=\$10{,}000}. \]

This does not mean the process will usually lose $\$10{,}000$. It means that $\$10{,}000$ is the probability-weighted average loss across many comparable exposures under the model.

The Same Expected Loss Can Hide Very Different Risks

Consider two possibilities:

Model A: a 10% chance of losing $\$100{,}000$;
Model B: a 1% chance of losing $\$1{,}000{,}000$.

Both have expected loss

\[ \$10{,}000. \]

Yet the severity and frequency profiles are very different. For decisions in which catastrophic outcomes matter, expected loss alone may be insufficient.

Risk as a Distribution, Not Just a Single Number

A model can produce a full loss distribution. From that distribution, analysts may examine:

the probability of any loss;
the expected loss;
the probability of exceeding a critical loss level;
high quantiles of the loss distribution;
the expected loss conditional on being in an extreme tail;
the frequency and severity of rare adverse outcomes.

Different summaries answer different decision questions.

Threshold and Exceedance Risk

A common risk question asks whether a quantity exceeds a limit. If $L$ is a random loss and $c$ is a critical threshold, then

\[ \boxed{P(L>c)} \]

is the exceedance probability.

Thresholds may represent structural capacity, regulatory limits, cash reserves, temperature limits, schedule deadlines, inventory shortages, or any other boundary separating acceptable from unacceptable outcomes.

Annual Exceedance Probability and Return Period

In environmental and infrastructure modeling, an event may be described by an annual exceedance probability $p$. A commonly reported return period is

\[ \boxed{T_r=\frac{1}{p}}. \]

If $p=0.01$, the corresponding return period is 100 years.

This does not mean such an event occurs exactly once every 100 years. It means the modeled annual exceedance probability is 1% in each year under the assumptions being used.

Risk Accumulates across Repeated Exposure

If an adverse event has probability $p$ in each independent period, then the probability of avoiding it for $n$ periods is

\[ (1-p)^n. \]

Therefore, the probability of at least one occurrence is

\[ \boxed{1-(1-p)^n}. \]

For a 1% annual event over 30 independent years,

\[ 1-(0.99)^{30}\approx\boxed{0.2603}. \]

A seemingly small annual probability can therefore become substantial over a long exposure horizon.

Quantile-Based Risk Measures

A loss quantile identifies a threshold that the loss distribution exceeds only with a specified probability.

If $q_{0.95}$ is the 95th percentile of loss, then

\[ P(L\le q_{0.95})=0.95 \]

for a continuous model.

In finance and some operational contexts, this type of quantile is often used as a Value-at-Risk-style summary. It gives information about a tail threshold, but not about how severe losses can become beyond that threshold.

Tail-Average Risk

To describe the severity of the tail beyond a high quantile, one may examine a tail-average quantity such as

\[ E[L\mid L\text{ lies in the selected upper tail}]. \]

In some fields this is called expected shortfall or conditional value at risk. The precise definition depends on conventions, especially for discrete loss distributions.

The important modeling point is that two distributions can have the same high quantile while differing greatly in the severity of outcomes beyond that quantile.

Risk Matrices

Organizations sometimes classify risks by placing probability categories on one axis and consequence categories on another. The resulting grid may label risks as low, moderate, high, or critical.

Risk matrices can be useful for communication and screening, but they are coarse. Category boundaries can hide important numerical differences, and the ranking may depend on how the categories are defined. Quantitative analysis should be retained when the decision justifies it.

Risk Tolerance and Decision Thresholds

A model does not determine whether a risk is acceptable. That requires a decision criterion.

A decision rule might require, for example,

\[ P(F)\le p_{\max}, \]

where $p_{\max}$ is a specified maximum tolerable failure probability.

Alternatively, a project may need to satisfy several constraints involving expected loss, worst-case exposure, service availability, safety margins, and cost. The mathematical model provides evidence for the decision; the decision standard comes from engineering requirements, business objectives, contracts, regulation, policy, or organizational risk appetite.

Safety Factors and Reliability

A deterministic design may compare resistance $C$ with demand $D$ using a safety factor such as

\[ \text{SF}=\frac{C}{D}. \]

A safety factor greater than 1 provides margin, but it does not by itself give a probability of failure. If both capacity and demand are uncertain, then a probabilistic formulation considers

\[ \boxed{P(D>C)}. \]

Reliability methods therefore complement deterministic safety factors by quantifying how uncertainty interacts with the available margin.

Limit-State Functions

A convenient reliability formulation defines a function

\[ g(\mathbf{X}) \]

whose sign indicates whether the system is safe or failed. A common convention is

\[ g(\mathbf{X})>0 \quad \text{safe}, \] \[ g(\mathbf{X})=0 \quad \text{boundary}, \] \[ g(\mathbf{X})<0 \quad \text{failure}. \]

Then the probability of failure is

\[ \boxed{p_f=P(g(\mathbf{X})<0)}. \]

This framework is especially useful when failure depends on several uncertain inputs simultaneously.

Estimating Reliability by Monte Carlo Simulation

Monte Carlo simulation provides a direct way to estimate failure probability. Generate $N$ simulated input sets, run the model, and count the number of failures.

If $I_i=1$ when simulation $i$ fails and $I_i=0$ otherwise, then

\[ \boxed{\hat p_f=\frac{1}{N}\sum_{i=1}^{N}I_i}. \]

Estimated reliability is then

\[ \boxed{\hat R=1-\hat p_f}. \]

Sampling Error in Estimated Failure Probability

If failures are independent across Monte Carlo trials and the true failure probability is $p_f$, then the approximate standard error of the estimator is

\[ \operatorname{SE}(\hat p_f)\approx \sqrt{\frac{p_f(1-p_f)}{N}}. \]

In practice, $p_f$ is often replaced by $\hat p_f$ for an estimated standard error.

When failures are rare, ordinary Monte Carlo may require an extremely large number of simulations before enough failures are observed to estimate the probability precisely.

Rare-Event Risk

Rare events are difficult because the outcomes of greatest concern may occur far out in the tail of a distribution. If the true failure probability is approximately

\[ 10^{-6}, \]

then one million ordinary Monte Carlo simulations would produce only about one failure on average.

Specialized methods such as importance sampling, subset simulation, analytical approximations, or carefully designed stress simulations may be needed. The exact technique depends on the model and on how the rare failure region is structured.

Uncertainty about the Risk Estimate Itself

A calculated risk is not necessarily known with certainty. The model may contain uncertain parameters, limited data, uncertain dependence structures, and structural assumptions.

Symbolically, if the failure probability depends on uncertain parameter $\theta$, then

\[ p_f=p_f(\theta). \]

Uncertainty in $\theta$ creates uncertainty in $p_f$. A complete analysis should distinguish the modeled frequency of failure from uncertainty about the estimate of that frequency.

Model Risk

Model risk is the possibility that decisions are harmed by an inadequate model, inappropriate assumptions, incorrect implementation, poor calibration, or use outside the model's valid domain.

A numerical result can be precise while the model is conceptually wrong. Good risk analysis therefore asks not only

\[ \text{What does the model predict?} \]

but also

\[ \boxed{\text{How could the model itself be wrong?}} \]

Dependence Can Dominate Risk

Independent-event assumptions often make calculations convenient, but many real losses become correlated during stress. Market positions can fall together, suppliers can fail under the same regional disruption, and backup systems can share environmental or infrastructure vulnerabilities.

Joint risk therefore depends not only on marginal failure probabilities but also on how adverse events occur together.

For events $A$ and $B$,

\[ P(A\cap B) \]

must be modeled or estimated rather than automatically replaced by

\[ P(A)P(B). \]

Reliability and Resilience Are Different

Reliability concerns avoiding failure. Resilience concerns the ability to absorb disruption, recover, adapt, or continue providing essential function after failure or degradation occurs.

A system can have high reliability but poor recovery capability. Another system may fail more often but recover quickly and maintain acceptable service through redundancy, repair, inventory buffers, rerouting, or other adaptive mechanisms.

Risk analysis may need both concepts when consequences depend on outage duration or recovery speed rather than only on whether an initial failure occurs.

Reliability and Availability

Reliability asks whether a system avoids failure over a period. Availability asks whether the system is operational when needed.

For a repairable system under stable long-run conditions, a common simplified availability relationship is

\[ \boxed{A\approx\frac{\operatorname{MTBF}} {\operatorname{MTBF}+\operatorname{MTTR}}}, \]

where MTBF is mean time between failures and MTTR is mean time to repair.

A system can experience failures yet maintain high availability if repairs are rapid. Conversely, rare failures can still create poor availability if recovery takes a very long time.

Risk Reduction Can Act on Probability or Consequence

Risk controls generally operate through one or more pathways:

reduce the probability of the initiating event;
reduce the probability that the initiating event causes failure;
reduce the severity of consequences after failure;
improve detection, warning, containment, or recovery;
reduce uncertainty by collecting better information.

The best intervention is therefore not always the one that produces the largest change in failure probability. Cost, consequence reduction, recovery, and information value may all matter.

Example: Comparing Two Risk-Reduction Options

Suppose a system faces a baseline failure probability of $0.04$, and failure would create a loss of $\$2{,}000{,}000$. The baseline expected loss is

\[ (0.04)(2{,}000{,}000)=\$80{,}000. \]

Option A reduces failure probability to $0.015$. Its expected loss becomes

\[ (0.015)(2{,}000{,}000)=\boxed{\$30{,}000}. \]

Option B leaves probability at $0.04$ but reduces the failure consequence to $\$900{,}000$. Its expected loss becomes

\[ (0.04)(900{,}000)=\boxed{\$36{,}000}. \]

Under expected loss alone, Option A produces the lower residual risk. A final decision would also consider implementation cost and any consequences not captured by the simple loss measure.

Communicating Reliability and Risk

A useful risk statement should identify the event, time horizon, operating conditions, probability measure, consequence measure, and major assumptions. Saying “the risk is 2%” is incomplete unless the reader knows what event has a 2% probability and over what period.

Better statements include forms such as:

“The modeled probability of component failure within 5 years is 2% under the stated load distribution.”
“The probability that annual loss exceeds $\$1$ million is estimated at 0.7%.”
“Mission reliability is estimated at 99.95%, assuming the two backup channels fail independently.”

These statements make the event and modeling assumptions visible.

Validate the Risk Model, Not Just the Arithmetic

Reliability and risk calculations can fail for reasons that are not numerical. Important validation questions include:

Is the failure event defined correctly?
Are the input distributions supported by data or defensible assumptions?
Are dependencies and common-cause mechanisms represented?
Does the time horizon match the decision?
Are consequences modeled at an appropriate level of detail?
Are rare-event estimates numerically stable?
Has the model been compared with observed failures, near misses, or historical outcomes?
Are structural uncertainties and omitted mechanisms acknowledged?

Reliability and risk are only as credible as the model used to generate them.

A Practical Reliability-and-Risk Workflow

Define the system, mission, or decision horizon.
Define success, failure, and relevant loss measures.
Identify uncertain inputs and failure mechanisms.
Specify probability models and dependence assumptions.
Propagate uncertainty through the model.
Estimate reliability, failure probability, loss, and tail measures.
Quantify sampling and parameter uncertainty where practical.
Test sensitivity to assumptions and model structure.
Compare results with decision thresholds or risk criteria.
Communicate both the numerical result and the assumptions that support it.

This workflow treats risk analysis as a model-based decision process rather than as a single formula.

Practice 1: Reliability from Failure Probability

A component has a 6% probability of failing during a mission. What is its mission reliability?

Practice 2: Reliability Function

If $F(500)=0.12$, find $R(500)$.

Practice 3: Exponential Reliability

A component has constant hazard $\lambda=0.0015$ per hour. Find its reliability over 200 hours.

Practice 4: Series System

Three independent components have reliabilities $0.99$, $0.96$, and $0.98$. Find the reliability of a system that requires all three components to work.

Practice 5: Parallel Redundancy

Two independent backup units each have reliability $0.92$. The system works if at least one unit works. Find the system reliability.

Practice 6: Expected Loss

A project has a 4% probability of a $\$750{,}000$ loss and otherwise no loss. Find the expected loss.

Practice 7: Repeated Exposure

An event has independent annual probability $p=0.02$. Find the probability of at least one occurrence during 10 years.

Practice 8: Monte Carlo Failure Estimate

In $50{,}000$ simulations, a system fails $425$ times. Estimate the failure probability and reliability.

Practice 9: Reliability versus Availability

Explain why a repairable system can have lower reliability over a year but still have high availability.

Practice 10: Common-Cause Failure

A model treats two backup pumps as independent, but both pumps depend on the same electrical panel. Explain why the independent parallel-system formula may overstate reliability.

Practice Answers

Practice 1

Reliability is one minus failure probability:

\[ R=1-0.06=\boxed{0.94}. \]

Practice 2

Since $R(t)=1-F(t)$,

\[ R(500)=1-0.12=\boxed{0.88}. \]

Practice 3

Under the exponential model,

\[ R(t)=e^{-\lambda t}. \]

Therefore,

\[ R(200)=e^{-0.0015(200)}=e^{-0.3}\approx\boxed{0.7408}. \]

Practice 4

For an independent series system,

\[ R=(0.99)(0.96)(0.98)\approx\boxed{0.9314}. \]

Practice 5

Each unit fails with probability

\[ 1-0.92=0.08. \]

Both fail with probability

\[ (0.08)^2=0.0064. \]

Therefore,

\[ R=1-0.0064=\boxed{0.9936}. \]

Practice 6

The expected loss is

\[ E[L]=(0.04)(750{,}000)=\boxed{\$30{,}000}. \]

Practice 7

The probability of no occurrence in 10 years is

\[ (0.98)^{10}\approx0.8171. \]

Therefore, the probability of at least one occurrence is

\[ 1-(0.98)^{10}\approx\boxed{0.1829}. \]

Practice 8

The estimated failure probability is

\[ \hat p_f=\frac{425}{50{,}000}=\boxed{0.0085}. \]

Therefore, estimated reliability is

\[ \hat R=1-0.0085=\boxed{0.9915}. \]

Practice 9

Reliability decreases when failures occur during the year because reliability measures uninterrupted successful operation over the specified interval. Availability can remain high if failures are repaired quickly, because availability measures the fraction of time the system is operational or the probability that it is operating when needed.

Practice 10

The pumps share a common electrical dependency. A failure of the panel can disable both pumps at once, so their failures are not fully independent. The independent parallel formula assumes simultaneous failure is much less likely than it may actually be and can therefore overstate system reliability.

What to Remember

Reliability is the probability that a system performs successfully over a stated mission or time horizon.
If $T$ is time to failure, then $R(t)=P(T>t)=1-F(t)$.
The hazard rate describes the instantaneous tendency to fail conditional on survival to the present.
A constant hazard produces the exponential reliability model $R(t)=e^{-\lambda t}$.
Series reliability multiplies component reliabilities only when the required independence assumptions are justified.
Parallel redundancy can raise reliability, but common-cause failures can defeat the apparent protection.
Risk combines uncertainty with consequences; expected loss is one useful summary but not the whole risk picture.
Threshold exceedance, tail quantiles, and tail-average measures answer different risk questions.
Small per-period probabilities can accumulate into substantial long-horizon exposure.
Monte Carlo simulation can estimate reliability and failure probabilities directly from a complex model.
Rare-event estimation may require specialized methods because ordinary simulation can observe too few failures.
Risk estimates inherit uncertainty from parameters, dependence assumptions, data limitations, and model structure.
Reliability, availability, and resilience are related but distinct system-performance concepts.
A risk calculation becomes decision-relevant only after the event, horizon, consequence, assumptions, and acceptance criteria are made explicit.

The essential principle is: uncertainty becomes risk when uncertain outcomes are connected to performance requirements, failure boundaries, and consequences.

Where We Go Next

Reliability and risk summarize what can go wrong and how much it may matter. Yet many decisions are not driven by one probability model alone. Decision- makers often need to compare qualitatively different futures: high growth and low growth, normal operations and disruption, favorable weather and extreme weather, stable markets and stressed markets, or alternative policy and technology environments.

The next lesson, Scenario Analysis, develops a structured way to compare these distinct possible futures without pretending that every uncertainty can be reduced to a single precisely known probability distribution.