Grade 12 Math

Slope is rise over run — how far a line climbs for each unit it travels sideways. What makes it a useful number rather than an arbitrary one is that you get the same answer whichever two points on the line you choose. That is not a coincidence; it is the defining property of a straight line, and it follows from similar triangles: any two points on the line, dropped to a horizontal and a vertical, form right triangles with the same angles, so their legs stay in the same ratio no matter how far apart you take them. A curve has no slope in this sense, because the ratio would depend on which pair you picked.

A worked instance in units a reader would hold. A wheelchair ramp is limited to a 1:12 slope, or 0.0833. A doorway 400 mm above the walk therefore needs $400/0.0833 = 4.8$ m of run, plus landings — which is why a step that looks trivial so often needs half the front yard. Drainage runs the same arithmetic in the other direction: a sanitary branch falling a quarter inch per foot has a slope of 0.0208, so 6 m of pipe drops 125 mm, and that number decides whether the fixture can reach the stack.

Slope is the ancestor of the derivative. Bring the two points closer and closer together on a curve and the ratio settles on the instantaneous rate of change at that spot, which is the whole of differential calculus in one sentence. It is also worth remembering that a slope carries units in any applied setting — dollars per kilometre, degrees per minute, metres per second. Only on a bare coordinate grid, with both axes in the same units, is it a pure number.

Three things to watch. First, keep the subtractions consistent: $(y_2 - y_1)/(x_2 - x_1)$ is fine and $(y_1 - y_2)/(x_1 - x_2)$ gives the identical answer, because negating both halves leaves the quotient alone — but mixing the two flips the sign. Second, a horizontal line has slope 0 while a vertical line has no slope at all. Its run is zero, and the quotient is undefined rather than infinite; these two get swapped constantly. Third, and most common outside the classroom: a percent grade is not an angle. Grade is $\tan\theta$ expressed as a percentage, so a 6% grade is 3.4°, not 6°, and a 100% grade is 45° rather than a cliff. Below about 10% the two are close enough that the habit forms; above it, the error grows quickly.

Under constant acceleration, velocity changes along a straight line, so the average velocity over the interval is the plain arithmetic mean of the value at the start and the value at the end. Multiply that by the elapsed time and you have the displacement: $d = \tfrac{v_0 + v}{2}\,t$. The reason this works is best seen on a velocity-versus-time graph, where displacement is the area underneath the curve. With constant acceleration the curve is a straight line, so the area is a trapezoid, and the area of a trapezoid is the mean of the two parallel sides times the width. The half in this formula is the half in the trapezoid rule.

A train easing from 30 m/s down to 10 m/s over 20 s covers $d = \tfrac{30 + 10}{2} \times 20 = 400$ m. Notice what you never needed: the acceleration. That is what makes this equation worth having as a separate page rather than treating it as a corollary — when the two speeds and the duration are what you have measured, it answers directly.

It is the bridge between the other kinematic equations rather than an independent fact. Substitute $v = v_0 + at$ into it and the algebra collapses to $d = v_0 t + \tfrac{1}{2}at^2$; substitute $v_0 = v - at$ instead and you recover $d = vt - \tfrac{1}{2}at^2$. All five SUVAT relations are the same two facts — velocity changes at a steady rate, displacement is the area under the velocity curve — rearranged to suit whichever variable is missing.

The failure mode is applying it when the acceleration was not constant, and it fails silently. Consider a car that sits at 10 m/s for 55 s and then accelerates hard to 30 m/s in the last 5 s. The starting velocity is 10, the final is 30, the time is 60 s, and this formula confidently reports 1200 m. The true distance is about 650 m. Nothing in the arithmetic warns you, because the shortcut assumes a straight line between the endpoints and the real motion was nothing of the kind. Whenever the acceleration varies, go back to the definition — total displacement over total time — or split the trip into segments where it genuinely is constant. One further distinction: this is average velocity, not average speed. If the motion reverses within the interval, the two are different numbers, and it is the velocity version that this formula computes.

Under constant acceleration, velocity changes at a steady rate, so the final speed is simply the starting speed plus the acceleration multiplied by the elapsed time. Picture a car merging onto a highway: entering the ramp at 15 m/s and holding a steady 2 m/s² for 5 seconds, it reaches v = 15 + (2)(5) = 25 m/s — right at highway pace. Deceleration works the same way with a negative a, which is how stopping times are estimated from braking data.

This is the first of the SUVAT equations, the toolkit of uniformly accelerated motion that traces back to Galileo's inclined-plane experiments in the early 1600s, where he showed that falling bodies gain equal speed in equal times. Because the relationship is linear in every variable, each of the four rearrangements has exactly one answer — no square roots, no ambiguity — making it the friendliest member of the kinematics family.

When acceleration is constant, displacement has two parts: the distance you would cover at your initial velocity alone, plus the extra distance contributed by speeding up — and that extra grows with the square of time. A jet starting its takeoff roll from rest and holding 2 m/s² covers d = 0 + ½(2)(30²) = 900 m in 30 seconds, which is why runways are measured in kilometres.

The ½ appears because the acceleration term is built from the average of a speed that grows linearly from zero. Galileo uncovered the underlying pattern — distances in successive equal time intervals follow the odd numbers 1, 3, 5, 7 — by rolling bronze balls down inclined planes. Note that solving for t would mean solving a quadratic with potentially two positive roots, so this calculator rearranges only for d, v₀, and a, where the answer is always single-valued.

This is the SUVAT equation with the clock taken out of it. Every other member of the family needs a time; this one relates the starting speed, the finishing speed, the acceleration and the distance directly, which makes it the right tool whenever you know where something ended up but not how long it took getting there. You obtain it by solving $v = v_0 + at$ for $t$ and substituting into $d = v_0 t + \tfrac{1}{2}at^2$; the time cancels and $v^2 = v_0^2 + 2ad$ is what survives.

A car braking from 25 m/s — 90 km/h — at a firm $-8$ m/s² comes to rest in $d = (0 - 25^2)/(2 \times -8) \approx 39$ m. That is the distance the tyres need after the brakes are on, and adding the driver's reaction time at 25 m/s puts another 25 m or so in front of it. Braking-distance tables in road-safety pamphlets are this equation, run once per speed.

Multiply the whole thing by $\tfrac{1}{2}m$ and it turns into something familiar: $\tfrac{1}{2}mv^2 = \tfrac{1}{2}mv_0^2 + (ma)d$, which is the work–energy theorem, final kinetic energy equals initial kinetic energy plus the work done by the net force. That is not a coincidence — it is the same statement written twice, once in the language of kinematics and once in the language of energy. It also explains the squares: energy has always gone as $v^2$, so a relation involving distance and force had to.

Signs are where this equation punishes carelessness. Pick a positive direction, then stay in it: a car braking while travelling in the positive direction has a negative $a$, and entering $+8$ instead of $-8$ returns a stopping distance of $-39$ m, which is the calculator telling you the car would have had to be reversing. The physical consequence of the squares is the one every driving instructor tries to convey and this formula proves: stopping distance goes as the square of speed. The 39 m from 90 km/h becomes about 156 m from 180 km/h. A 20% increase in speed is a 44% increase in the distance you need. Two smaller points: solving for $v$ or $v_0$ takes a square root, and this page returns the principal non-negative branch — if the object actually reversed direction, the negative root is the physical one and you should supply the sign yourself. And $d$ is displacement along the direction of motion, not path length, so it is not the right tool for a curved route.

Every parabola y = ax² + bx + c is symmetric about a single vertical line, and that line runs through its turning point at x = −b/(2a). The expression is just the quadratic formula with the square root deleted: the two roots sit at −b/(2a) plus and minus √Δ/(2a), so their midpoint — the vertex — is the shared first term. The equation of the axis of symmetry is literally x = h. Worked example: y = 2x² − 8x + 5 turns at h = 8/4 = 2, and its roots 2 ± √(6)/2 straddle that value symmetrically.

This is the workhorse of every optimisation question in a first algebra course, because a parabola's maximum or minimum can only occur at the vertex: fence 40 m of wire into the largest rectangle, and the area A = x(20 − x) = −x² + 20x peaks at h = −20/(2 × −1) = 10 m, a square. Descartes' La Géométrie of 1637 is what made this arithmetic possible at all — before coordinates, a parabola was a slice through a cone and its axis was found by construction, not by dividing two numbers. The trap is dropping the minus sign or forgetting the 2 in the denominator; if a is negative, remember that h can still be positive because two negatives cancel.

The discriminant is the part of the quadratic formula that lives under the square root, and its sign alone answers "how many real solutions?" without solving anything. Positive means two distinct real roots (the parabola cuts the x-axis twice), zero means one repeated root (the parabola just kisses the axis at its vertex), and negative means no real roots at all (the parabola floats entirely above or below). Worked example: 3x² − 4x − 2 = 0 gives Δ = (−4)² − 4(3)(−2) = 16 + 24 = 40 — positive but not a perfect square, so two real irrational roots.

That last test is the practical one for students: Δ a perfect square (and a, b, c whole numbers) means the quadratic factors neatly over the integers, so you can skip the formula entirely. Engineers read the same quantity as a stability check — the discriminant of a system's characteristic equation decides whether a damped oscillator returns smoothly to rest or overshoots and rings. The common trap is sign bookkeeping: with c negative, −4ac becomes an addition, which is why Δ = 40 above rather than −8. Square b first, then subtract, and keep the minus sign attached to c.

Substituting the vertex abscissa x = −b/(2a) back into y = ax² + bx + c collapses to k = c − b²/(4a), the height of the turning point. Notice the discriminant hiding inside: k = −Δ/(4a), so a parabola with Δ = 0 has its vertex exactly on the x-axis. Worked example: y = x² − 6x + 5 has k = 5 − 36/4 = −4, so the vertex is (3, −4) and the curve dips four units below the axis before climbing back through its roots at x = 1 and x = 5.

Because the vertex is the only turning point, k is the answer to "what is the smallest (or largest) value this expression can take?" — a projectile launched at 20 m/s from ground level follows h = 20t − 4.9t², peaking at k = 0 − 400/(4 × −4.9) ≈ 20.4 m. Getting there by hand means completing the square, the same manoeuvre a Babylonian scribe used on tablet BM 13901 nearly four thousand years ago and that al-Khwārizmī later justified with a literal picture of a square with rectangles glued to two sides. The classic trap is sign handling when a is negative: 4a is then negative, so subtracting b²/(4a) raises k, which is exactly right for a downward-opening curve.

The quadratic formula solves every equation of the form ax² + bx + c = 0 in one stroke, and this page returns the plus branch — the root you get by adding the square root of the discriminant. Its twin, the minus branch, is a separate calculator here because a solver returns one number at a time; run both with the same a, b and c to see the complete pair. Worked example: 2x² + 3x − 5 = 0 gives b² − 4ac = 9 + 40 = 49, so x = (−3 + 7)/4 = 1 here and (−3 − 7)/4 = −2.5 on the minus branch, and indeed 2(1)² + 3(1) − 5 = 0.

The method is older than the notation. A Babylonian scribe working the clay tablet BM 13901 around 1800 BCE handled "I added the area and the side of my square and got 0;45" by halving the coefficient, squaring it, and adjusting — completing the square, three and a half millennia before anyone wrote a formula down. Around 820 CE in Baghdad, al-Khwārizmī's al-Kitāb al-mukhtaṣar fī ḥisāb al-jabr wa-l-muqābala gave the word al-jabr ("restoration") to the whole subject, though with no negative numbers available he had to treat six separate cases. The trap to avoid: a must not be zero (the equation is then linear), and if the discriminant comes out negative there is no real root at all — the parabola never touches the x-axis. Watch the signs, too: with b negative, −b is positive.

This is the minus branch of the quadratic formula — the same machine as the plus branch, with the square root subtracted instead of added. Because a solver hands back one number, the two roots live on two pages here; feed identical a, b and c to both and you have the full solution set. Worked example: 2x² + 3x − 5 = 0 has discriminant 9 + 40 = 49, so this branch gives x = (−3 − 7)/4 = −2.5 while the plus branch gives 1. Check it: 2(−2.5)² + 3(−2.5) − 5 = 12.5 − 7.5 − 5 = 0.

Two sanity checks catch most mistakes. The roots must sum to −b/a (here 1 + (−2.5) = −1.5 = −3/2 ✓) and multiply to c/a (1 × −2.5 = −2.5 = −5/2 ✓) — relations named for François Viète, who in 1591 became the first algebraist to use letters for the known coefficients as well as the unknowns, which is why we can write "ax² + bx + c" at all. When a is positive the minus branch is the smaller root; when a is negative the labels swap, because dividing by a negative 2a flips the order. And if the discriminant is exactly zero, both branches return the same repeated root.

You can read the sum of a quadratic's two roots straight off its coefficients: expand a(x − x₁)(x − x₂) and the x term comes out as −a(x₁ + x₂), so x₁ + x₂ = −b/a. Worked example: 2x² − 7x + 3 = 0 has S = 7/2 = 3.5, and the actual roots 3 and ½ do indeed add to 3.5 — no formula, no square roots, no arithmetic beyond one division.

The relation carries François Viète's name; his 1591 In artem analyticem isagoge introduced the habit of writing consonants for known coefficients and vowels for unknowns, and only with symbolic coefficients could such a statement even be phrased. Today it is the fastest way to check a solution — if your two roots do not sum to −b/a you have made an arithmetic slip — and the fastest way to build a quadratic to order: roots summing to 5 and multiplying to 6 give x² − 5x + 6 = 0. Two traps: the minus sign is part of the formula (S = −b/a, not b/a), and it holds even when the roots are complex, where the imaginary parts cancel and the sum stays real.

Multiply out a(x − x₁)(x − x₂) and the constant term lands as a·x₁x₂, so the two roots always multiply to c/a. Worked example: 4x² + 5x − 9 = 0 has P = −9/4 = −2.25, and the roots 1 and −9/4 confirm it. Paired with the sum relation S = −b/a, this pins a quadratic completely: any equation whose roots add to S and multiply to P is x² − Sx + P = 0 up to an overall scale factor.

The sign of P is a free diagnostic. Negative means the roots straddle zero — one positive, one negative — so a projectile's height equation with a negative constant term always has one physically meaningless negative time. Positive P with positive S means both roots are positive; positive P with negative S means both are negative. The trap: P = c/a, with no minus sign, unlike the sum. And remember these relations describe the roots you would get, real or complex — a quadratic with no real roots still has a perfectly real product c/a, because the two complex conjugates multiply to the square of their modulus.

Halve the coefficient of x, square it, and add: that single instruction turns x² + bx into the perfect square (x + b/2)². Worked example: x² + 10x needs k = (10/2)² = 25, giving (x + 5)². Once an expression is a perfect square you can take a square root of both sides and finish the problem in two lines, which is exactly how the quadratic formula itself is derived and how vertex form a(x − h)² + k is reached.

The name is literal. A Babylonian scribe on tablet BM 13901, around 1800 BCE, solved "I added the area and the side of my square and got 45 (in base 60, i.e. 3/4)" by halving the side coefficient, squaring, and adjusting — arriving at ½ with no algebraic symbols at all. Al-Khwārizmī in ninth-century Baghdad drew the same argument as an actual diagram: a square of side x with two rectangles of width b/2 glued to adjacent sides, leaving a missing corner of area (b/2)² to be filled in. Two traps. First, the leading coefficient must be 1 before you start — for 3x² + 12x, factor out the 3 to get 3(x² + 4x) and complete the square inside. Second, adding k changes the expression, so in an equation you must add it to both sides, and in an expression you must immediately subtract it again to keep the value unchanged.

Fixed costs arrive whether you sell anything or not: rent, insurance, the shop truck payment, the phone. Every unit you do sell contributes its price minus its own variable cost toward that pile, and break-even is simply the point where the contributions finish covering it. With $12,000 of monthly fixed cost and units that sell for $40 while costing $25 in material and labour, each one contributes $15, and $12000/15 = 800$ units clears the month.

The quantity $p - v$ is called the contribution margin, and it deserves the attention rather than the price. Raising the price by $5 lifts the contribution from $15 to $20 and drops break-even from 800 units to 600, a 25% cut in the work needed to survive. Shaving $5 off the variable cost does exactly the same thing. Neither is intuitive until you see that both act on a denominator.

If the price does not exceed the variable cost, no quantity ever breaks even and the solver refuses the question. This is not a technicality. A shop losing money on every job cannot make it up in volume, and the formula is the most honest way to demonstrate that to someone who believes otherwise.

A vector is an arrow: a direction plus a length. Drop it onto axes and it splits into two ordinary numbers, the components v_x and v_y, which form the legs of a right triangle whose hypotenuse is the arrow itself — so the magnitude is √(v_x² + v_y²). A hiker who walks 3 km east and 4 km north ends up 5 km from camp, not 7 km, because displacement is a vector and its parts add head-to-tail rather than arithmetically. That gap between 5 and 7 is the single most common student error in the whole subject.

Run it backwards and a magnitude plus one component recovers the other: a vector of length 13 with v_y = 5 must have v_x = √(169 − 25) = 12. The solver returns the positive root, since a component's sign is a matter of which way the axis points and the arithmetic cannot know it. Note also that no component may exceed the magnitude — a leg longer than the hypotenuse describes a triangle that does not exist.

Components tell you where a vector ends; the direction angle tells you which way it leans. Since tan θ = v_y/v_x, the naive answer is arctan(v_y/v_x) — and that answer is wrong half the time. The ratio for (−3, −4) is identical to the ratio for (3, 4), so a plain arctan reports a vector pointing southwest as though it pointed northeast. The fix is the two-argument function atan2(v_y, v_x), which keeps both signs and returns the true bearing anywhere in the full −180° to 180° sweep. It entered programming through early Fortran and is now in every standard library precisely because the quadrant bug was so pervasive.

Example: (−1, 1) points up and to the left, and atan2(1, −1) = 135° — correct, where arctan(1/−1) = −45° would have been off by exactly half a turn. Read the other way, a vector known to lie at 30° with v_x = 10 must have v_y = 10 tan 30° ≈ 5.774. Straight up or straight down the tangent blows up, which is the arithmetic's honest way of saying a vertical vector has no run to divide by.

Resolving a vector is the reverse of building one: the shadow it casts on the x-axis has length |v| cos θ. Simon Stevin demonstrated the underlying parallelogram rule in 1586 with an inclined-plane thought experiment — a closed loop of beads draped over a wedge that would have to move forever if forces did not combine this way — and Newton restated it as Corollary I of the Principia. Every ramp problem since is an application: a 200 N pull on a sled at 25° above the ground drives the sled forward with 200 × cos 25° ≈ 181 N, while the rest of the effort merely lifts.

The classic trap is measuring the angle from the wrong line. The cosine belongs to the component along the axis the angle is measured from; take the angle from the vertical instead and cosine and sine swap places. Check by sanity: at θ = 0° the whole vector lies on x, and cos 0° = 1 delivers exactly that.

The vertical partner of the x-component: |v| sin θ is how much of the arrow points up. A projectile launched at 40 m/s and 30° leaves the muzzle with 40 × sin 30° = 20 m/s of upward speed, and that number alone — not the 40 — decides how high and how long it flies. Together the pair (|v| cos θ, |v| sin θ) is the machinery that lets you replace one awkward diagonal with two independent one-dimensional problems, which is the reason vectors earn their place in physics at all.

Solving for the angle uses arcsin, whose principal branch runs only from −90° to 90°. A vector at 150° has the same y-component as one at 30°, so formula.expert returns the first-quadrant answer; if you know the vector leans left, its true direction is 180° minus what you get. When the whole vector is horizontal, sin θ = 0 and no y-component can tell you how long it is.

TOA — Tangent is Opposite over Adjacent. It is the ratio to reach for when the hypotenuse is unknown, and in the field the hypotenuse usually is unknown, because it runs along your line of sight to something you cannot put a tape on. Tangent is not independent of the other two: dividing $\sin\theta = o/h$ by $\cos\theta = a/h$ cancels the hypotenuse and leaves $o/a$, which is the formula on this page and the reason the hypotenuse drops out of the problem.

Here is the fact that makes this the most quietly useful ratio of the three: tangent is slope. Rise over run is opposite over adjacent, so $m = \tan\theta$, and every grade, pitch and gradient you have ever seen is a tangent in disguise. A 6% road grade is $\arctan(0.06) = 3.43°$. A 6:12 roof pitch is $\arctan(0.5) = 26.6°$. A 1:12 ramp is 4.76°. If you have ever wondered why a highway sign warning of an 8% grade describes something that looks nowhere near eight degrees, this is why — and it is worth noticing that a 100% grade is 45°, not vertical.

The classic use is an angle of elevation. Stand 50 m out from a tower on level ground and sight the top at 31°: the rise above your eye is $50\tan 31° = 30.0$ m. Backwards, $\theta = \arctan(o/a)$, and the principal branch is exactly right here — two positive legs always land in 0°–90°, so there is no ambiguity to resolve.

Four cautions, in the order they cost people money. First, that tower is not 30 m tall. The formula returns the rise above the instrument, so you must add your eye height, or the tripod height, to get the real total — a systematic error of a metre and a half that stays invisible because the answer looks reasonable. Second, $a$ is the horizontal distance. On sloping ground a taped distance runs along the slope and is longer than the run, which inflates the result; that is why survey instruments reduce slope distance to horizontal before anything else happens. Third, tangent has no ceiling — it runs to infinity at 90° — and grows viciously sensitive as it approaches. At 30° a half-degree error in the sighting shifts the height by about 2%; at 80° the same half degree shifts it by 5%, and at 85° by 10%. Standing farther back and sighting at a shallower angle is nearly always the more accurate measurement. Fourth, as ever, check whether your calculator is in degrees.

Two vectors add head to tail, and the closing side of the parallelogram they form is the resultant. Its length is √(A² + B² + 2AB cos θ), where θ is the angle between the two vectors laid tail to tail. At 90° the cosine term vanishes and the familiar √(A² + B²) returns — 3 and 4 at right angles give exactly 5. At 0° the vectors reinforce completely and R = A + B; at 180° they fight and R = |A − B|. Simon Stevin argued the rule into existence in 1586 with his chain of beads over an inclined plane, and Newton made it Corollary I of the Principia in 1687.

Watch the sign. The law of cosines carries a minus, this formula a plus, and the difference is not a typo: the law of cosines uses the interior angle of the triangle, while θ here is the angle between the vectors, and the two are supplementary. Mixing them up is the single most common error in force-addition problems. Worked backwards: two forces of 5 N and 8 N producing a 7 N resultant must be separated by arccos((49 − 25 − 64)/80) = arccos(−0.5) = 120°.

The law of cosines is the Pythagorean theorem with a correction term: when angle C is exactly 90°, cos C vanishes and c² = a² + b² reappears. Open the angle wider and the correction adds length; squeeze it and the correction subtracts. Euclid proved both cases geometrically around 300 BCE, and the Persian astronomer al-Kashi gave the modern trigonometric form in the 1400s — in France the result still carries his name. A surveyor's example: from one station, two landmarks lie 8 km and 5 km away with 60° between the sightlines, so their separation is $\sqrt{64 + 25 - 2 \cdot 8 \cdot 5 \cdot \cos 60^\circ} = \sqrt{49} = 7$ km.

Solved for the angle, C = arccos((a² + b² − c²)/(2ab)) uses the principal branch of arccos, spanning 0° to 180° — precisely the range of a triangle's interior angle, so the answer is unique with no ambiguous case. The three sides must obey the triangle inequality, or the arccos argument escapes [−1, 1] and no triangle exists.

The dot product pairs up matching components, multiplies, and adds — and the answer is a plain number, not a vector. That single scalar measures agreement of direction: positive when two vectors broadly point the same way, zero when they are perpendicular, negative when they oppose. Example: (3, 4)·(2, −1) = 6 − 4 = 2, a small positive number, meaning the two arrows lean the same way but only just. Change b to (4, −3) and the product is 12 − 12 = 0: exactly perpendicular, which is far easier to test this way than by measuring any angle.

The notation is younger than the idea. Through the 1880s Hamilton's quaternions were the standard tool for three-dimensional algebra, and their product mixed a scalar part with a vector part in one unwieldy object. Josiah Willard Gibbs at Yale, printing lecture notes privately for his students, and Oliver Heaviside in England, working on Maxwell's equations, independently tore that product in half — keeping the scalar piece as the dot product and the vector piece as the cross product. Quaternion loyalists fought the change bitterly in the pages of Nature; Gibbs and Heaviside won, and every physics course since teaches their version.

This is the dot product seen from the geometry side rather than the component side: length times length times the cosine of the angle between. The two views are the same number, which is exactly why the dot product is useful — you can compute it cheaply from coordinates and then read it as a statement about angles. With |a| = 6, |b| = 5 and 60° between them, a·b = 6 × 5 × 0.5 = 15. Push the angle to 90° and the product collapses to zero; push past it and the product turns negative.

The everyday consequence is physical: work is a dot product, so a porter carrying a suitcase horizontally does no work against gravity at all, the force being straight up and the motion straight along. The trap is assuming that θ is any angle in your diagram; it must be the angle between the two vectors when they are placed tail to tail. Nose-to-tail sketches quietly hand you the supplement, and cos flips sign the moment you cross 90°.

Divide the dot product by both magnitudes and what is left is a pure cosine, so the angle between any two vectors falls out of their coordinates with no drawing required. Take a = (1, 0) and b = (1, 1): the dot product is 1, the magnitudes are 1 and √2, and arccos(1/√2) = 45°, which is obviously right. Take a = (3, 4) and b = (4, 3): the dot product is 24, both magnitudes are 5, and arccos(24/25) = arccos(0.96) ≈ 16.26° — a result no sketch would give you to two decimals.

The Cauchy–Schwarz inequality guarantees the fraction never leaves the range −1 to 1, so an angle always exists. Floating-point arithmetic does not care: for two nearly parallel vectors the computed ratio can come out as 1.0000000000000002, and arccos of that is NaN. This is one of the most reported bugs in graphics and robotics code, and the fix is a one-line clamp before the arccos — which formula.expert applies. Note also that the answer runs from 0° to 180° and carries no sense of rotation; if you need to know whether b sits clockwise or counterclockwise from a, use the cross product's sign instead.

Where the dot product keeps the part of two vectors that agrees, the cross product keeps the part that disagrees: |a||b| sin θ, zero for parallel vectors and largest at a right angle. Numerically it is the area of the parallelogram the two vectors span, which is why torque, angular momentum and magnetic force all wear a cross product. With |a| = 4, |b| = 5 and 30° between them, |a × b| = 4 × 5 × 0.5 = 10.

The operation was born on 16 October 1843, when William Rowan Hamilton — after fifteen years of failing to multiply triples — walked along the Royal Canal in Dublin with his wife, saw the quaternion rule in a flash, and carved i² = j² = k² = ijk = −1 into the stone of Broom Bridge with a penknife. The vector part of a quaternion product is precisely the modern cross product; Gibbs and Heaviside later extracted it as a standalone operation. The trap when solving backwards for the angle is that arcsin returns only 0° to 90°: 30° and 150° give identical cross-product magnitudes, and only the sign of the dot product can tell you which of the two you have.

Two vectors sharing a corner sweep out a parallelogram, and its area is the absolute value of the cross product's z-component: |a_xb_y − a_yb_x|. The result needs no angle, no perpendicular height and no trigonometry — just four coordinates. Sides a = (3, 1) m and b = (1, 4) m enclose |12 − 1| = 11 m². The construction is a special case of Hermann Grassmann's exterior product from his 1844 Ausdehnungslehre, a book so far ahead of its style that almost nobody read it; Grassmann gave up mathematics and became a distinguished Sanskrit scholar instead, and the algebra he invented was only rediscovered decades later.

The solver handles this one for the area only. That is not a limitation of the algebra but of the geometry: infinitely many pairs of vectors enclose the same area, so a single component cannot be recovered from it. The determinant's sign, stripped away by the absolute value, is worth keeping when you compute it by hand — it tells you whether b lies clockwise or counterclockwise from a, and polygon area algorithms depend on exactly that.

The diagonal of a parallelogram cuts it into two congruent triangles, so a triangle built on two vectors has exactly half the cross-product area: ½|a_xb_y − a_yb_x|. Legs of a = (4, 0) m and b = (0, 3) m give ½ × 12 = 6 m², matching the familiar half-base-times-height. But the formula does not care about right angles: a = (6, 2) m and b = (1, 5) m give ½|30 − 2| = 14 m² with no height ever measured.

Written for three arbitrary corner points rather than two vectors, this becomes the shoelace formula — so named because the cross-multiplied terms lace together like a boot — published by A. L. F. Meister in 1769 and later generalized by Gauss to polygons of any number of sides. Surveyors computed parcel areas with it for two centuries before computers, and GIS software still does. Because the absolute value discards the sign, the area alone cannot be run backwards to recover a component; that sign, however, is what lets the polygon version cancel overlapping strips automatically and return the correct area for any simple polygon, convex or not.

Work is the dot product of force and displacement, so in components it is simply F_xd_x + F_yd_y — no angle to measure and no cosine to look up. Each axis contributes independently, and the contributions can cancel: a force of (12, −5) N acting through a displacement of (3, 4) m does 36 − 20 = 16 J of work, the downward pull eating into what the forward push delivered. A component pair that sums to zero means the force was perpendicular to the motion and did no work at all, which is why a satellite in a circular orbit neither gains nor loses energy despite being pulled on the whole way around.

Gaspard-Gustave de Coriolis named the quantity travail in 1829 and fixed the modern convention of force times distance; the joule was attached to it later, honouring James Prescott Joule's paddle-wheel measurements of the mechanical equivalent of heat. The component form makes the sign convention obvious in a way the |F||d| cos θ form does not: negative work simply means the force opposed the motion, exactly what friction and braking do.

Shine a light perpendicular to b and the shadow that a casts along it has length (a·b)/|b| — equivalently |a| cos θ. This is how a diagonal quantity gets charged to a chosen direction: the component of a ship's velocity along the intended course, the part of a rope's tension that actually lifts, the useful fraction of an angled push. A dot product of 24 against a vector b of length 6 means a reaches 4 units along b's direction, regardless of how long a itself is or how far it strays sideways.

The trap sits in the denominator. It is |b|, the vector being projected onto, never |a|. Dividing by the wrong magnitude quietly returns the cosine of the angle scaled by the wrong length, and the mistake survives dimensional analysis, so it will not announce itself. A negative result is perfectly legitimate and means a leans backwards along b — its shadow falls on the far side of the origin.

When every outcome is equally likely, probability is just bookkeeping: count the outcomes you want, divide by all the outcomes there are. Gerolamo Cardano wrote the first honest analysis of this in his Liber de Ludo Aleae around 1564 — a gambler's handbook so frank about cheating that it stayed unpublished until 1663 — and Laplace made it the formal definition in 1812. A single die has six faces, three of them even, so P(even) = 3/6 = 0.5. Enter probabilities here as decimals between 0 and 1, or switch the unit to % if you prefer 50 to 0.5.

The trap is the phrase "equally likely". Rolling two dice gives eleven possible totals, 2 through 12, but they are not equally likely: there are 36 equally likely face-pairs, and six of them total 7 while only one totals 12, so P(7) = 6/36 ≈ 0.167 against P(12) = 1/36 ≈ 0.028. Always count the underlying equal outcomes, never the labels people put on them. Run the formula backwards to size a sample space: if 13 of the cards in a deck are hearts and P = 0.25, then n = 13/0.25 = 52.

Probability measures successes against all trials; odds measure successes against failures. A probability of 0.75 is odds of 0.75/0.25 = 3, quoted as "3 to 1 on"; odds of 1:4 against, entered here as 0.25, correspond to P = 0.25/1.25 = 0.2. Enter probabilities as decimals from 0 to 1 (the % unit works too) and odds as the number of units in favour per one unit against.

Bookmakers quote the mirror image — "4/1" means four units against one in favour, so P = 1/5 = 0.2 — and they deliberately publish odds implying probabilities that sum to more than 1 across a field, the overround that pays the shop. Odds also make Bayes' theorem elegant: posterior odds equal prior odds times the likelihood ratio, which is why epidemiologists and machine-learning models work in log-odds. The usual trap is reading "3 to 1" as a 1-in-3 chance; it is 1 in 4.

Every trial ends either in A or in not-A, and those two possibilities carry all the probability there is — so they must add to 1. Probabilities on this site are decimals from 0 to 1 (the % unit is there if you prefer percentages), and the complement rule is the cheapest trick in the subject: if a weather model gives a 0.3 chance of rain, the chance of a dry day is 1 − 0.3 = 0.7, no further modelling required.

Its real power is turning hard problems inside out. "At least one" questions are almost always easier as "not none": the probability of at least one head in five coin tosses means adding five separate cases, while the complement is a single product, 1 − 0.5⁵ = 1 − 0.03125 = 0.96875. Pascal and Fermat leaned on exactly this inversion in their 1654 letters. The classic error is subtracting from the wrong whole — the complement of "at least two" is "zero or one", not "at most two".

Mutually exclusive (disjoint) events have no outcomes in common — a single card cannot be both a king and a queen, one die roll cannot be both a 2 and a 5 — so their chances simply add. Drawing one card, P(king or queen) = 4/52 + 4/52 = 8/52 ≈ 0.154. Enter each probability as a decimal from 0 to 1, or switch the unit to % and type 25 instead of 0.25.

Galileo was asked by his Medici patrons why, with three dice, a total of 10 shows up more often than 9 even though both can be made six ways; his short memo Sopra le Scoperte dei Dadi answered it by adding the probabilities of genuinely distinct, equally likely outcomes rather than of the unordered "ways". That is the standing trap: the rule breaks the moment the events can happen together. "Drawing a king" and "drawing a heart" overlap in the king of hearts, so adding them double-counts one card — use the general addition rule for that. A useful sanity check is built in here: if your two exclusive probabilities sum past 1, one of them is wrong.

Adding P(A) and P(B) counts every outcome that belongs to both events twice, so the general addition rule refunds one copy of the overlap. Drawing one card from a standard deck, P(king or heart) = 4/52 + 13/52 − 1/52 = 16/52 ≈ 0.308 — the king of hearts is a single card and gets counted once. This is the two-event case of inclusion–exclusion, the counting principle Abraham de Moivre put to work in The Doctrine of Chances (1718), the book that taught a generation of gamblers and annuity sellers to reason about risk.

Enter probabilities as decimals from 0 to 1 (or use the % unit). Rearranged for the overlap it becomes a quiet diagnostic tool: if a survey reports that 60% of people drink coffee, 50% drink tea, and 80% drink at least one, then P(both) = 0.6 + 0.5 − 0.8 = 0.3, so 30% drink both. If a set of reported figures forces the overlap negative — say 0.3 + 0.3 − 0.9 — the numbers are simply inconsistent, and the calculator says so instead of pretending.

Independence means learning that A happened tells you nothing about B, and then the joint chance is a plain product: two fair coins both landing heads is 0.5 × 0.5 = 0.25. Because probabilities are fractions, multiplying shrinks them fast — a component that fails 1 time in 100 gives a five-component chain a 0.99⁵ ≈ 0.951 chance of surviving, so reliability engineers watch products, not sums. Enter each probability as a decimal from 0 to 1, or switch the unit to %.

On 18 August 1913 the roulette wheel at Monte Carlo landed on black twenty-six times running, and gamblers lost fortunes betting ever larger sums on red, convinced it was "due". The wheel had no memory: each spin remained an independent 18/37 ≈ 0.486. That is the trap in both directions — assuming independence where it does not exist is just as costly. Two mortgages defaulting are not independent when the same recession hits both borrowers, which is roughly how the models of 2007 went wrong.

Conditioning throws away every outcome in which B did not happen and re-normalises what is left, which is why you divide by P(B). Roll one die: P(it is a 2) = 1/6, but if you are told the roll is even, only three outcomes survive and P(2 | even) = (1/6)/(1/2) = 1/3. Probabilities go in as decimals from 0 to 1 (the % unit is available too).

In September 1990 Marilyn vos Savant told Parade readers that a contestant on Let's Make a Deal should switch doors after Monty Hall opens a losing one, because switching wins 2/3 of the time. Some 10,000 letters arrived, roughly a thousand from people with PhDs, insisting she was wrong; Paul Erdős stayed unconvinced until he saw a simulation. The host's choice is not independent of where the car is, and conditioning on which door he opened is the whole game. The standard trap is confusing P(A|B) with P(B|A): almost every fatal cycling accident involves a bicycle, yet almost no bicycle ride is fatal.

This is the honest version of "multiply the probabilities": the second factor is the chance of B after A has already happened. Independence is only the special case where P(B|A) = P(B). Drawing two aces from a shuffled deck without replacement, P = (4/52) × (3/51) = 12/2652 ≈ 0.00452 — one chance in 221 — because the first ace leaves 51 cards holding 3 aces. Enter probabilities as decimals from 0 to 1, or switch the unit to %.

Edward Thorp built Beat the Dealer (1962) on precisely this dependence: in blackjack the cards already dealt change the odds on the next one, so the deck acquires a memory that roulette wheels lack, and casinos answered by shuffling several decks together. The trap is reusing the unconditional figure for the second draw — 4/52 twice gives 0.0059, about 30% too high. Rearranged, the rule extracts the conditional from survey data: if 60% of customers own a car and 15% own both a car and a bike, then P(bike | car) = 0.15/0.6 = 0.25.

n! counts the orderings of n distinct things: the first slot has n candidates, the next n − 1, and so on. Five books on a shelf can be arranged 5! = 5 × 4 × 3 × 2 × 1 = 120 ways; ten books already reach 10! = 3,628,800. The factorial's explosive growth is why brute-force scheduling and travelling-salesman searches collapse so quickly — 52!, the number of orders a shuffled deck can take, is about 8 × 10⁶⁷, so essentially every properly shuffled deck in history has been a first.

The "!" notation is Christian Kramp's, introduced in 1808 mainly because printers hated the older bracket symbols. Two conventions catch people out: 0! = 1, not 0 — there is exactly one way to arrange nothing, and the whole edifice of combinations depends on it — and n! is only defined for whole numbers in counting problems. This calculator quietly extends to fractional inputs through the gamma function (Euler's continuous version, with n! = Γ(n + 1)), so 0.5! returns $\sqrt{\pi}/2 \approx 0.886$, but for counting keep n a whole number. Above 170 the answer exceeds what a double can hold.

A permutation is a selection where order matters — gold, silver and bronze are three different results, not one. With ten sprinters in a final there are 10P3 = 10!/7! = 10 × 9 × 8 = 720 possible podiums. The formula is just the falling product n(n−1)…(n−r+1) written as a ratio of factorials, since dividing by (n − r)! cancels the tail you never reach.

Note that there is no closed-form way to run this backwards, so this calculator solves only for the count. The classic trap is choosing permutations where combinations belong: a five-card poker hand is not 52P5 = 311,875,200, because the same five cards dealt in a different order are the same hand — divide by 5! to get 2,598,960. Ask yourself whether rearranging the chosen items produces a genuinely different outcome. If yes, permutations; if no, combinations. A four-letter code from 26 distinct letters with no repeats allows 26P4 = 26 × 25 × 24 × 23 = 358,800 codes.

A combination counts groups, not orders. Start from the permutations n!/(n−r)! and divide by r!, because each group of r items was counted once for every way of shuffling it. A five-card poker hand from a 52-card deck: 52C5 = 2,598,960. A 6/49 lottery ticket: 49C6 = 13,983,816, which is exactly why the jackpot chance is about 1 in 14 million.

These numbers are the entries of Pascal's triangle, which Pascal himself laid out in the Traité du triangle arithmétique (written 1654, printed 1665) while corresponding with Fermat about how to divide the stakes of an interrupted game — the exchange usually credited with founding probability theory. The triangle was known centuries earlier to Halayudha in India and to Yang Hui in China. Two things to remember: nCr = nC(n−r), so choosing 47 cards to discard is the same job as choosing 5 to keep, and the count has no closed-form inverse, so this calculator solves only for the number of groups.

Three ideas multiply here: $p^k$ for the successes, $(1-p)^{n-k}$ for the failures, and the binomial coefficient for the number of orders in which they can occur. Exactly two heads in five fair tosses: C(5,2) × 0.5² × 0.5³ = 10 × 0.03125 = 0.3125. The distribution is Jacob Bernoulli's, worked out before his death in 1705 and published in Ars Conjectandi in 1713 by his nephew Nicolaus, together with the first law of large numbers.

Enter p as a decimal from 0 to 1, or switch that input to %. Three conditions must hold or the answer is meaningless: fixed n, independent trials, and a constant p — sampling without replacement from a small population violates the last two, and that is the hypergeometric distribution's job instead. There is no closed form for n, k or p given the probability, so this calculator solves only for P. And read the question carefully: "exactly k" is one term, while "at least k" means summing several. Three sixes in ten rolls of a fair die: C(10,3) × (1/6)³ × (5/6)⁷ ≈ 0.155.

Expectation is linear, so the average number of successes is simply the per-trial chance repeated n times: μ = np. Toss a fair coin 100 times and expect 50 heads; screen 100 parts with a 3% defect rate and expect 3 defects. Enter p as a decimal from 0 to 1, or switch that input to %.

Reversed, this is the estimator behind opinion polling and quality control: if 25 heads appeared in an experiment with a fair coin, n = 25/0.5 = 50 tosses; if 12 defects turned up in 400 units, p = 12/400 = 0.03. Two traps deserve naming. First, the mean need not be an achievable outcome — 10 rolls of a die average 10/6 ≈ 1.67 sixes, and nobody ever rolls two-thirds of a six. Second, the mean says nothing about spread on its own; pair it with the variance np(1 − p) before deciding whether an observed count is surprising.

Each trial contributes p(1 − p) of variance, and independent variances add, so n trials give np(1 − p). The factor peaks at p = 0.5, where a coin is at its most unpredictable, and vanishes at p = 0 or 1, where the outcome is certain. One hundred fair coin tosses have σ² = 100 × 0.5 × 0.5 = 25, so σ = 5: seeing 60 heads is two standard deviations out, unusual but hardly impossible. Enter p as a decimal from 0 to 1, or use the % unit.

This is where poll margins of error come from — a sample of 1000 with p ≈ 0.5 has σ ≈ 15.8 respondents, about 1.6 percentage points, and the familiar "±3 points" is roughly two of those. Notice that the standard deviation grows like $\sqrt{n}$ while the mean grows like n, which is why doubling a sample only cuts relative uncertainty by about 30%. Solving for p has no unique answer — a variance of 24 in 100 trials fits both p = 0.4 and p = 0.6 — so that rearrangement is deliberately omitted; use the mean μ = np to pin p down.

"At least one" is the complement of "none at all", and "none at all" is a simple product: all n attempts must fail, which happens with probability (1 − p)^n. At least one six in four rolls of a die is 1 − (5/6)⁴ = 1 − 625/1296 ≈ 0.518, a favourable bet. The Chevalier de Méré made money on exactly that wager, then reasoned that 24 throws of two dice should be equally good since 24/36 matches 4/6 — but 1 − (35/36)²⁴ ≈ 0.491, just under even, and his losses prompted the 1654 Pascal–Fermat correspondence that founded the subject.

Enter p as a decimal from 0 to 1, or switch the unit to %. Solving for n answers the planning question: to reach a 99% chance of at least one success when each attempt works 10% of the time, n = ln(0.01)/ln(0.9) ≈ 43.7, so 44 attempts. The trap is adding probabilities instead — ten 10% attempts do not give 100%, they give 1 − 0.9¹⁰ ≈ 0.651. Note the useful rule of thumb that n = 1/p attempts lands near 1 − e⁻¹ ≈ 63%, never certainty.

Expected value weights each outcome by its probability: win W a fraction p of the time, lose L the rest. Bet $1 on a single number at American roulette and you win $35 with probability 1/38 and lose $1 with probability 37/38, so E = (1/38)(35) − (37/38)(1) = −2/38 ≈ −$0.0526 — the famous 5.26% house edge, identical on every bet the American wheel offers except the five-number line. Enter p as a decimal from 0 to 1, or use the % unit; W and L are plain amounts in whatever currency you like.

Rearranged for p, it gives the break-even chance a wager needs to be worth taking: risking $1 to win $4 requires p = (0 + 1)/(4 + 1) = 0.2, so anything above a 20% chance is profitable in the long run. That framing — comparing your estimated probability against the price on offer — is how professional bettors and insurers actually work. The trap is treating expected value as what will happen: you never lose 5.26 cents on a spin, you lose $1 or gain $35, and it is only across thousands of spins that the average asserts itself. Expected value also ignores risk of ruin, which is why the St Petersburg game, with infinite expectation, is worth only a few dollars to a real player.

To have your first success on trial k you must fail k − 1 times and then succeed, which is $(1-p)^{k-1} \times p$. Rolling a fair die, the chance the first six comes on the third roll is (5/6)² × (1/6) = 25/216 ≈ 0.116. Enter p as a decimal from 0 to 1 or switch that input to %; k is a whole trial number, 1 or more.

The geometric distribution is memoryless — the only discrete distribution that is — so after twenty failed rolls the chance the next roll is a six is still 1/6, exactly as it was at the start. That is the gambler's fallacy in its purest form, and the "hot hand" is its mirror image. Notice that the single most likely value of k is always 1, however small p is: each additional wait multiplies by (1 − p) < 1. Solving for p has no closed form, so this calculator solves for P or for k: with p = 0.2 and P = 0.08192, k = 1 + ln(0.4096)/ln(0.8) = 5.

A z-score strips the units off a measurement and reports it in standard deviations: subtract the mean, divide by the standard deviation, and a 130 IQ (μ = 100, σ = 15) becomes z = (130 − 100)/15 = 2.0, exactly the same standing as a 700 on an SAT section scaled to μ = 500, σ = 100. That common currency is the whole point — it lets you compare a bench-press result to a reading score. The Belgian astronomer Adolphe Quetelet was the first to push this idea hard: in the 1830s he applied the astronomers' error curve to chest circumferences of Scottish soldiers, inventing l'homme moyen and, with it, the habit of measuring people against a mean and a spread.

The trap is forgetting that z-scores only carry probability meaning when the distribution is roughly normal. A z of 2.0 is about the 98th percentile in a bell curve, but in a badly skewed distribution — household income, say — the same z can be nowhere near the top 2%. Run it backwards to recover a raw score: on a test with μ = 72 and σ = 8, the cutoff for the top decile (z ≈ 1.28) is x = 72 + 1.28 × 8 ≈ 82.2.

The interquartile range throws away the top and bottom quarters of the data and measures what is left: Q₃ − Q₁, the width of the middle half. If a class's quartile scores are 48 and 72, the IQR is 24 points. Because it depends only on the ranks of the middle observations, it is immune to outliers in a way the range and standard deviation are not — you could replace the single highest value with a million and the IQR would not move at all. That robustness is why it is the spread reported for skewed data like incomes, house prices and reaction times.

The trap is comparing quartiles computed by different software. There are at least nine defensible definitions of a quartile, and Excel, R and most textbooks do not all agree on where Q₁ falls in a small data set — differences that vanish in large samples but can be embarrassing in a homework answer. John Tukey put the IQR at the centre of exploratory data analysis in 1977, using it both for the box in his box plot and for the fences that flag outliers at 1.5 IQR beyond each quartile.

The range is spread at its crudest: highest minus lowest, no averaging, no squares. Daily weather reports live on it — a high of 98 °F and a low of 62 °F is a range of 36 degrees — and quality engineers still use it on the shop floor because an operator can compute it in their head. It is also the basis of the handy field estimate that a standard deviation is roughly the range divided by four for moderate samples, useful when you need a σ to plug into a sample-size calculation and have nothing else.

The fatal weakness is that it depends on exactly two numbers and ignores everything in between, so a single mis-keyed value destroys it. Worse, the range grows systematically with sample size: draw 10 values from a distribution and you might span 3σ, draw 1,000 from the same distribution and you will span more like 6σ, even though nothing about the population changed. That makes ranges from different-sized samples incomparable — one reason the interquartile range, which trims the extremes away, is preferred whenever you have enough data to compute it.

Variance and standard deviation are the same fact in two costumes. Variance averages the squared deviations from the mean, which makes it mathematically well behaved — variances of independent quantities simply add — but leaves it in squared units, so the variance of a set of heights in centimetres comes out in cm². Taking the square root undoes that and returns the standard deviation, a spread you can actually mark on the axis alongside the data. Ronald Fisher coined the word "variance" in a 1918 paper on Mendelian inheritance, precisely because he needed a quantity that partitioned additively across sources of variation; the whole of analysis of variance grew out of that choice.

The common trap is averaging standard deviations. You cannot: if two independent measurements have σ = 3 and σ = 4, the combined spread is √(9 + 16) = 5, not 3.5. Add the variances, then take the root. Numerically the conversion is trivial in both directions — a standard deviation of 12 means a variance of 144, and a reported variance of 6.25 means a standard deviation of 2.5.

John Tukey's fences turn outlier hunting into arithmetic. Take the interquartile range, multiply by 1.5, and subtract that from Q₁: anything below the result is a candidate outlier and gets plotted as an individual dot beyond the box plot's whisker. With Q₁ = 48 and IQR = 24, the lower fence sits at 48 − 36 = 12, so a score of 9 is flagged while a score of 15 is not. The 1.5 is not derived from anything deep — Tukey chose it because on normal data it flags roughly 0.7% of observations, often enough to be useful and rarely enough not to cry wolf.

The essential trap is treating a flagged point as bad data. A fence says "this value is unusual relative to the middle half", not "this value is wrong". Skewed distributions such as incomes or hospital lengths of stay throw points past the fences constantly and legitimately, which is why some analysts use 3.0 × IQR for a "far out" tier before considering exclusion. Delete points only when you can identify a cause — a transcription error, a broken instrument — never merely because a fence flagged them.

The upper fence is the mirror of the lower one: Q₃ plus one and a half interquartile ranges. With Q₃ = 72 and IQR = 24 it lands at 72 + 36 = 108, so any observation above 108 is drawn as a separate point outside the whisker. Because it is built from quartiles rather than the mean and standard deviation, the rule cannot be sabotaged by the very outliers it is looking for — a mean-and-σ rule inflates σ when an extreme value is present and can end up hiding it, a failure known as masking.

Real-world skew is the thing to watch. Right-skewed data — salaries, insurance claims, download sizes — routinely put several percent of observations above the upper fence with nothing wrong at all, so a box plot of claims that shows a dozen high dots is describing the business, not finding errors. In practice the upper fence earns its keep as a screening tool: in laboratory quality control it flags samples for a re-run rather than for deletion. Solving backwards is handy when reading someone else's plot — a whisker ending at 100 above a box top of 70 implies IQR = 30/1.5 = 20.

Correlation and slope are the same relationship in different clothes. The correlation r is unit-free; multiply it by the ratio of the spreads and you get a slope carrying real units — change in y per unit of x. With r = 0.8, $s_y$ = 5 and $s_x$ = 2, the slope is 0.8 × 5/2 = 2.0. Francis Galton stumbled on this in the 1880s while plotting the heights of children against their parents: the best-fit line was flatter than the 45° line of pure heredity, and he called the effect "regression towards mediocrity" — the origin of the word regression itself.

Galton's discovery is also the trap. Because |r| ≤ 1, the slope is always shallower than $s_y/s_x$, so predicted values are pulled toward the mean of y. Tall parents have tall-but-shorter children; a company's best quarter is usually followed by a worse one — not because anything caused it, but because the extreme was partly luck. Note too that r is symmetric while the slope is not: regressing y on x gives 2.0 here, but regressing x on y gives 0.8 × 2/5 = 0.32, and 1/2.0 = 0.5 is not the answer.

Every least-squares line passes through the point of averages (x̄, ȳ) — that is a consequence of minimising squared vertical errors, not a coincidence. So once you know the slope, the intercept is forced: slide back from ȳ by b times x̄. With ȳ = 50, b = 2.5 and x̄ = 12, the intercept is 50 − 30 = 20, and the fitted line is ŷ = 20 + 2.5x. Adrien-Marie Legendre published the least-squares recipe in 1805; Gauss claimed he had been using it since 1795 and proved it by predicting where the lost asteroid Ceres would reappear, which it duly did.

The trap is over-reading the intercept. It is the fitted value at x = 0, which is often far outside the data and physically meaningless — a regression of adult weight on height gives an intercept for a person of zero height. It exists to anchor the line, nothing more, and a "significant" intercept usually tells you nothing interesting. When x = 0 really is nonsense, centring the predictor (using x − x̄) makes the intercept equal ȳ and gives it an interpretation worth having.

Once a line is fitted, prediction is arithmetic: multiply the predictor by the slope and add the intercept. A line of ŷ = 20 + 2.5x evaluated at x = 8 predicts 40. The hat on ŷ matters — it marks a fitted value, an average outcome for that x, not a promise about any individual. Real observations scatter around the line by the residual standard deviation, so a prediction interval for one new case is always much wider than the confidence interval for the line itself.

The classic trap is extrapolation. Within the range of the data a line can be an excellent summary; outside it, it is a guess dressed as arithmetic. Mark Twain skewered this perfectly in Life on the Mississippi: noting that the river had shortened by 242 miles over 176 years, he extrapolated back to conclude that the Lower Mississippi was once 1.3 million miles long and would eventually be a mile and three-quarters. Running the line backwards is legitimate and useful — solving 65 = 5 + 3x gives x = 20 — but it is calibration, and it inherits all the uncertainty of the fit.

Square the correlation and you get the fraction of the variance in y that the regression accounts for. A correlation of 0.9 means R² = 0.81, so 81% of the variation is explained by the line and 19% remains as residual scatter. The squaring is deflating on purpose: a correlation of 0.5, which sounds like a solid relationship, explains only a quarter of the variation. Karl Pearson formalised the correlation coefficient in 1896, building on Galton's earlier work, and the squared version quickly became the standard summary of how well a straight line fits.

Two traps. First, R² says nothing about causation or about whether a line is the right shape — Anscombe's famous 1973 quartet contains four data sets with identical R² of 0.67, one a clean linear trend, one a perfect parabola, one a straight line ruined by a single outlier. Always plot the data. Second, this calculator returns the positive root when you invert it, because squaring destroys the sign: R² = 0.64 implies |r| = 0.8, but only the scatterplot or the slope tells you whether the relationship runs up or down.

Percent error grades a measurement against a known truth: take the size of the miss and divide it by the accepted value. Measuring the boiling point of water as 97.0 °C when the accepted value is 100.0 °C is a 3.0% error; determining a density as 9.7 g/cm³ against a reference 10.0 gives 0.3/10 = 0.03, or 3%. Pick the % display unit to read it directly as a percentage. It is the standard closing line of a school lab report and the standard acceptance criterion in instrument calibration.

The trap is the denominator. Percent error divides by the accepted value, not by your measurement and not by the average of the two — divide by the wrong one and a 3% error becomes 3.09%, which is the sort of thing that costs marks. If there is no accepted value, because you are comparing two equally trustworthy measurements, you want percent difference instead. And note what this number is not: a small percent error means your result is accurate, but it says nothing about precision. A thermometer reading 97.0 every single time is beautifully precise and still 3% wrong.

You cannot average two averages unless the groups are the same size. Twenty students scoring a mean of 80 and thirty scoring 70 do not combine to 75; the correct answer weights each mean by its count: (20 × 80 + 30 × 70)/50 = 3700/50 = 74. The pull is always toward the larger group. This is the scalar form of the general weighted mean, and it covers most of what people actually need — merging two class sections, two shifts, two lab runs — without requiring a full data list.

Ignoring the weights is one of the most common quantitative errors in circulation, and it has a famous escalation: Simpson's paradox, where a treatment can look better in every subgroup yet worse overall once unequal group sizes are folded in. The 1973 Berkeley graduate-admissions case is the textbook example — the university appeared to favour men overall, but almost every individual department favoured women; the aggregate simply reflected that women applied more often to departments that admitted few of anyone. Reversed, the formula does mixture arithmetic: if 20 students averaging 80 must be blended down to a combined 74 with a second group averaging 70, that second group needs 20 × (74 − 80)/(70 − 74) = 30 students.

Simple interest pays only on the original principal, never on past interest: a $2,000 loan at 6% per year for 3 years costs I = 2000 × 0.06 × 3 = $360, exactly $120 each year. Enter the rate as a decimal (6% → 0.06) and keep the rate and the time in the same period — a monthly rate with months, a yearly rate with years.

Short-term personal loans, car loans, and bonds' coupon payments often work this way. Compare with compound interest, where each period's interest joins the principal and the total pulls ahead of the simple-interest straight line.

Percent change measures a difference relative to where you started: a price moving from $80 to $92 changed by (92 − 80)/80 = 0.15, a 15% rise; a fall from $80 to $68 gives −0.15. Pick the % display unit to see it as a percentage directly, and note that the old value is always the reference — which is why a 50% loss needs a 100% gain to break even.

The rearrangements handle both everyday directions: what a $60 jacket costs after a 30% markup (60 × 1.30 = $78), and the reverse — the pre-sale price of an item now $45 after a 25% discount, 45 / 0.75 = $60, not $56.25 as adding 25% back would wrongly suggest.

With periodic compounding, each period's rate r is split into n slices and applied n times, so past interest starts earning interest of its own. $5,000 at 4% per year compounded monthly (n = 12) for 10 years grows to A = 5000 × (1 + 0.04/12)¹²⁰ ≈ $7,454.16 — about $50 more than yearly compounding would give, because 120 small boosts beat 10 large ones.

Solving for r recovers the rate a savings product actually paid between two statements, and solving for t answers "how long until my balance reaches A?" Enter r as a decimal (4% → 0.04), and note that pushing n toward infinity lands on the continuous-compounding formula $A = Pe^{rt}$.

Anything that grows by a fixed percentage each period — money at compound interest, populations, subscriber counts — multiplies by (1 + r) every step, so growth compounds on itself. Enter r as a decimal: 5% is 0.05. Worked example: $1,000 invested at 5% per year for 10 years gives A = 1000 × 1.05¹⁰ ≈ $1,628.89 — the extra $128.89 beyond simple interest is interest earned on earlier interest. Einstein may never have called compound interest the eighth wonder of the world, but bankers behave as if he did.

The same formula runs backward. Solving for r extracts an average per-period growth rate from two snapshots, and solving for t (using logarithms) answers "how long until we reach A?" A handy shortcut, the Rule of 72, falls straight out of the t rearrangement: dividing 72 by the percentage rate approximates the doubling time, so money at 6% doubles in roughly 12 years.