Calculus is the mathematics of local change and accumulated quantity. Limits make approximation exact, derivatives convert infinitesimal input changes into first-order output changes, and integrals recover totals from densities or rates. The fundamental theorem of calculus connects differentiation and integration. In several variables, gradients, Jacobians, Hessians, line and surface integrals, and the integral theorems organize geometry, optimization, conservation, and physical fields. A complete calculus argument must state its domain, regularity, limiting process, units, orientation, and error control.

<svg viewBox="0 0 760 470" xmlns="http://www.w3.org/2000/svg" font-family="Arial,sans-serif"><rect x="0" y="0" width="760" height="470" rx="18" fill="#0d1117"/><text x="380" y="34" fill="#e6edf3" font-size="21" font-weight="700" text-anchor="middle">Calculus connects local change and global accumulation</text><text x="380" y="60" fill="#8b949e" font-size="12" text-anchor="middle">Limits turn approximation into a theorem</text><rect x="45" y="115" width="195" height="220" rx="14" fill="#161b22" stroke="#58a6ff" stroke-width="2"/><text x="142" y="150" fill="#79c0ff" font-size="14" font-weight="700" text-anchor="middle">Derivative</text><path d="M75 275Q140 145 210 255" fill="none" stroke="#58a6ff" stroke-width="4"/><line x1="90" y1="255" x2="205" y2="185" stroke="#3fb950" stroke-width="3"/><text x="142" y="310" fill="#c9d1d9" font-size="11" text-anchor="middle">local slope · sensitivity</text><path d="M240 225H300" stroke="#d29922" stroke-width="4"/><polygon points="300,225 287,217 287,233" fill="#d29922"/><rect x="300" y="95" width="160" height="260" rx="14" fill="#161b22" stroke="#d29922" stroke-width="2"/><text x="380" y="135" fill="#e3b341" font-size="14" font-weight="700" text-anchor="middle">Limit</text><text x="380" y="205" fill="#e6edf3" font-size="16" text-anchor="middle">Δx → 0</text><text x="380" y="250" fill="#c9d1d9" font-size="11" text-anchor="middle">sums ↔ integrals</text><text x="380" y="285" fill="#c9d1d9" font-size="11" text-anchor="middle">ratios ↔ derivatives</text><path d="M460 225H520" stroke="#3fb950" stroke-width="4"/><polygon points="520,225 507,217 507,233" fill="#3fb950"/><rect x="520" y="115" width="195" height="220" rx="14" fill="#161b22" stroke="#3fb950" stroke-width="2"/><text x="617" y="150" fill="#7ee787" font-size="14" font-weight="700" text-anchor="middle">Integral</text><path d="M550 280Q610 145 680 255L680 290H550Z" fill="#3fb950" opacity=".3"/><path d="M550 280Q610 145 680 255" fill="none" stroke="#3fb950" stroke-width="4"/><text x="617" y="310" fill="#c9d1d9" font-size="11" text-anchor="middle">total · area · conserved quantity</text><text x="380" y="410" fill="#e6edf3" font-size="12" text-anchor="middle">The fundamental theorem makes the two operations inverse under suitable hypotheses.</text></svg>

A limit describes behavior near a point without requiring evaluation at that point. $\lim_{x\to a}f(x)=L$ means values of $f(x)$ can be made arbitrarily close to $L$ by taking $x$ sufficiently close to $a$ while $x\ne a$. The function may be undefined or differently defined at $a$. Limits distinguish a removable hole from a jump, divergence, or oscillation. Graphical intuition is valuable, but the definition governs ambiguous cases.

The epsilon–delta definition makes closeness quantitative. For every $\epsilon>0$ there must exist $\delta>0$ such that $0<|x-a|<\delta$ implies $|f(x)-L|<\epsilon$. The order of quantifiers matters: $delta$ may depend on $epsilon$ but not on the particular $x$ after it is chosen. A proof constructs or bounds such a $delta$. Numerical evidence samples finitely many points and cannot establish the universal statement.

One-sided limits diagnose boundaries and jumps. The limits $x\to a^-$ and $x\to a^+$ restrict approach direction. A two-sided limit exists only when both exist and agree. At a domain endpoint, the relevant relative-domain limit may be one-sided. Piecewise models, switching laws, contacts, and threshold devices often have meaningful one-sided behavior even when no two-sided derivative exists.

Infinite limits and limits at infinity express different behaviors. $lim_{x\to a}f(x)=\infty$ states unbounded growth near finite $a$ rather than convergence to a real number. $lim_{x\to\infty}f(x)=L$ describes long-range behavior. Horizontal, vertical, and oblique asymptotes summarize particular limits but do not replace them. Rates such as logarithmic, polynomial, exponential, and factorial growth require comparisons beyond a common infinite label.

Limit laws require existence of the component limits. Sums, products, quotients with nonzero denominator limit, and compositions under continuity rules allow algebraic calculation. Canceling a factor can reveal a removable singularity because the simplified expression agrees only away from the point, which is enough for the limit. Substitution is justified by continuity, not by visual habit. Indeterminate forms signal missing information rather than a final answer.

The squeeze theorem controls a difficult function between easier bounds. If $g(x)\le f(x)\le h(x)$ near a point and both outer functions approach the same limit, then $f$ does too. It proves limits with oscillation such as $x\sin(1/x)$ and underlies geometric trigonometric limits. A bound whose upper and lower limits differ proves nothing. Dimensional or sign errors often become obvious when proposed bounds are checked.

Continuity means the function value agrees with its limiting behavior. At $a$, require $f(a)$ defined, $lim_{x\to a}f(x)$ existent, and equality between them. Polynomial, rational away from poles, exponential, logarithmic on its domain, and trigonometric functions inherit continuity through operations and composition. Continuous functions on closed bounded intervals attain maxima and minima and take every intermediate value, results essential to existence arguments.

Uniform continuity controls one input tolerance across an entire domain. Ordinary continuity lets $delta$ depend on the point; uniform continuity does not. Every continuous function on a compact interval is uniformly continuous. This stronger control justifies exchanging limits, controlling numerical steps globally, and proving Riemann integrability. Functions such as $1/x$ on $(0,1)$ are continuous but not uniformly continuous because sensitivity diverges near the excluded endpoint.

A sequence is a function on integers whose limit tests long-run behavior. $a_n\to L$ means all sufficiently large terms lie within every tolerance of $L$. Monotone bounded sequences converge, a key completeness consequence. Recurrences, algorithms, Riemann sums, and series partial sums are sequences. A plotted finite prefix can disguise slow divergence, oscillation, or a long transient.

Completeness supplies the numbers that limiting processes need. The real numbers contain suprema of nonempty bounded-above sets, unlike the rationals. Nested intervals, Cauchy sequences, monotone convergence, and intermediate-value results rely on this property. Calculus is not merely algebra with a very small increment; its exact conclusions rest on the topology and completeness of the number system or function space being used.

<svg viewBox="0 0 760 470" xmlns="http://www.w3.org/2000/svg" font-family="Arial,sans-serif"><rect x="0" y="0" width="760" height="470" rx="18" fill="#0d1117"/><text x="380" y="34" fill="#e6edf3" font-size="21" font-weight="700" text-anchor="middle">A derivative is the limiting slope of secant lines</text><text x="380" y="60" fill="#8b949e" font-size="12" text-anchor="middle">The tangent is a local linear model, not merely a touching line</text><line x1="80" y1="360" x2="690" y2="360" stroke="#8b949e"/><line x1="80" y1="360" x2="80" y2="90" stroke="#8b949e"/><path d="M100 330C210 305 265 225 350 190C440 150 535 180 670 95" fill="none" stroke="#58a6ff" stroke-width="5"/><circle cx="350" cy="190" r="8" fill="#d29922"/><circle cx="565" cy="160" r="8" fill="#d29922"/><line x1="350" y1="190" x2="565" y2="160" stroke="#d29922" stroke-width="3"/><text x="470" y="135" fill="#e3b341" font-size="11">secant for h</text><circle cx="430" cy="165" r="7" fill="#3fb950"/><line x1="260" y1="230" x2="540" y2="115" stroke="#3fb950" stroke-width="4"/><text x="525" y="95" fill="#7ee787" font-size="11">tangent as h → 0</text><rect x="225" y="390" width="310" height="42" rx="8" fill="#161b22" stroke="#a371f7"/><text x="380" y="417" fill="#d2a8ff" font-size="13" text-anchor="middle">f′(a) = limₕ→₀ [f(a+h)−f(a)]/h</text></svg>

The derivative is a local linearization. If $f$ is differentiable at $a$, then $f(a+h)=f(a)+f'(a)h+o(h)$. The remainder is small relative to $h$, which is stronger than saying the tangent looks close on a graph. This formulation generalizes cleanly to vectors and matrices. Slope, instantaneous rate, sensitivity, marginal value, and small-signal gain are interpretations of the same linear approximation with different units.

Differentiability implies continuity but continuity does not imply differentiability. A corner such as $|x|$ is continuous at zero but has unequal one-sided slopes. Cusps, vertical tangents, oscillations, and fractal functions provide other failures. A function can be differentiable once but have a discontinuous derivative. Before applying derivative rules or optimization tests, identify the domain and nonsmooth points rather than assuming a formula is smooth everywhere.

Derivative units are output units divided by input units. Velocity is position per time, current slope can be amperes per volt, and a temperature coefficient can be resistance per kelvin. The differential $df\approx f'(x)dx$ propagates small changes with these units. A dimensionally inconsistent derivative often reveals that a logarithm has a dimensional argument or that percent and fraction were mixed. Nondimensionalization makes sensitivities comparable.

The product rule records simultaneous first-order changes. $(fg)'=f'g+fg'$. The term $f'g'$ would multiply two small increments and is second order, so it disappears from the derivative limit. The quotient rule follows from differentiating $f=g(f/g)$ or an inverse. Memorized rules become easier to audit when derived from linearization and order counting.

The chain rule composes local linear maps. $(f\circ g)'(x)=f'(g(x))g'(x)$ in one variable. It converts rates across intermediate variables and powers backpropagation, coordinate transformations, sensitivity analysis, and automatic differentiation. The order of multiplication matters for vector maps. A missing chain factor is one of the most common errors in nonlinear models.

Implicit differentiation handles relations that do not solve globally for one variable. If $F(x,y)=0$ and $F_y\ne0$, then locally $dy/dx=-F_x/F_y$. The implicit function theorem supplies conditions and higher-dimensional generalization. At points where the denominator vanishes, the chosen representation can fail even though the curve is smooth with a vertical tangent or a different coordinate chart.

Inverse-function derivatives reciprocate local scale under nonzero slope. If $y=f(x)$ is locally one-to-one and $f'(x)\ne0$, then $(f^{-1})'(y)=1/f'(x)$. This derives logarithmic and inverse-trigonometric derivatives. A zero derivative can make the inverse nondifferentiable, as cube root at zero illustrates. Global invertibility additionally requires domain restrictions.

Higher derivatives describe curvature and changing sensitivity. $f''$ measures how slope changes, giving acceleration, convexity, and local quadratic behavior. Mixed and higher derivatives organize response tensors. Smoothness assumptions determine whether mixed partials commute. High-order derivatives can grow quickly and amplify noise; their existence and usefulness are separate questions.

Rolle's and mean-value theorems turn local slopes into global conclusions. A continuous function on $[a,b]$ differentiable inside has some point where $f'(c)=[f(b)-f(a)]/(b-a)$. Consequences include monotonicity from derivative sign, constancy from zero derivative, uniqueness estimates, error bounds, and l'Hôpital-type arguments. The theorem asserts existence, not where the point lies, and its hypotheses matter at endpoints and corners.

Taylor's theorem separates a polynomial approximation from its remainder. Near $a$, $f(x)$ equals a degree-$n$ Taylor polynomial plus a remainder controlled by a higher derivative under suitable smoothness. The polynomial records local derivative data; the remainder determines accuracy. A Taylor series may converge yet not to the function, or may have finite radius. Writing several terms without an error estimate is approximation, not equality.

Linearization propagates uncertainty only within a local regime. For small input perturbation $delta x$, $delta f\approx f'(x)delta x$. Independent random errors lead to variance formulas involving squared sensitivities, while correlated errors require covariance terms. Strong nonlinearity, bounds, discontinuities, or large uncertainty make the linear approximation biased. Compare second-order terms or sample the nonlinear map before trusting a differential error bar.

Local extrema require examining stationary and nonsmooth candidates. At an interior differentiable extremum, $f'=0$, but the converse is false: stationary points can be maxima, minima, or neither. Endpoints, corners, discontinuities, and domain boundaries also matter. The first-derivative sign test is robust; the second-derivative test classifies nondegenerate stationary points but is inconclusive when $f''=0$.

Convexity turns local information into global optimization guarantees. A differentiable convex function lies above every tangent, and any stationary point is a global minimum. Positive semidefinite second derivative or Hessian is a sufficient smooth criterion on a convex domain. Strict convexity gives uniqueness under conditions. Nonconvex functions can have many local minima and saddles, so a vanishing gradient alone is weak evidence.

Related-rates problems are chain-rule models with a constraint. Write the geometric or physical relation first, differentiate with respect to time while all variables still vary, then substitute the instant's values and units. Substituting constants too early can erase their derivatives incorrectly. The sign of the requested rate should follow the geometry and serve as a sanity check.

Optimization word problems require a feasible domain before differentiation. Translate constraints, eliminate variables or use multipliers, identify endpoints, solve stationary conditions, and compare objective values. A critical point outside the feasible set is irrelevant. Dimensional scaling and limiting cases often expose algebraic errors. Real designs may add discrete choices, uncertainty, and multiple objectives beyond elementary calculus.

<svg viewBox="0 0 760 470" xmlns="http://www.w3.org/2000/svg" font-family="Arial,sans-serif"><rect x="0" y="0" width="760" height="470" rx="18" fill="#0d1117"/><text x="380" y="34" fill="#e6edf3" font-size="21" font-weight="700" text-anchor="middle">A definite integral is a limit of weighted sums</text><text x="380" y="60" fill="#8b949e" font-size="12" text-anchor="middle">Signed accumulation survives nonuniform partitions and changing density</text><line x1="75" y1="350" x2="690" y2="350" stroke="#8b949e"/><line x1="75" y1="350" x2="75" y2="90" stroke="#8b949e"/><path d="M90 300C180 270 230 135 340 165C450 195 520 290 670 120" fill="none" stroke="#58a6ff" stroke-width="5"/><g fill="#3fb950" opacity=".3" stroke="#3fb950"><rect x="100" y="285" width="45" height="65"/><rect x="145" y="250" width="45" height="100"/><rect x="190" y="205" width="45" height="145"/><rect x="235" y="165" width="45" height="185"/><rect x="280" y="155" width="45" height="195"/><rect x="325" y="165" width="45" height="185"/><rect x="370" y="185" width="45" height="165"/><rect x="415" y="220" width="45" height="130"/><rect x="460" y="260" width="45" height="90"/><rect x="505" y="275" width="45" height="75"/><rect x="550" y="240" width="45" height="110"/><rect x="595" y="180" width="45" height="170"/></g><text x="380" y="395" fill="#e6edf3" font-size="13" text-anchor="middle">∫ₐᵇ f(x)dx = lim max Δxᵢ→0 Σ f(xᵢ*)Δxᵢ</text><text x="380" y="430" fill="#c9d1d9" font-size="11" text-anchor="middle">Area is one interpretation; mass, charge, probability, work, and flux use the same accumulation.</text></svg>

The definite integral is a limit of Riemann sums. Partition $[a,b]$, sample each subinterval, sum $f(x_i^*)\Delta x_i$, and let the maximum width approach zero. For continuous functions the limit exists and is independent of sampling. The integral is signed: regions below the axis subtract. Physical totals multiply a density by its measure, so units gain a factor of the integration variable.

Integrability is weaker than continuity. Bounded functions with sufficiently small discontinuity sets can be Riemann integrable. A single jump does not prevent integration, while a function discontinuous at every rational/irrational alternation can fail. The Lebesgue integral generalizes by measuring level sets and handles convergence more flexibly. Elementary calculus usually uses continuous or piecewise-continuous functions where both viewpoints agree.

The fundamental theorem of calculus links accumulation and local rate. If $f$ is continuous and $F(x)=\int_a^x f(t)dt$, then $F'(x)=f(x)$. Conversely, if $G'=f$, then $\int_a^b f(x)dx=G(b)-G(a)$. MIT's calculus materials emphasize these complementary parts. The theorem depends on regularity; distributions, jumps, and improper integrals require generalized statements.

An indefinite integral denotes a family of antiderivatives. If $F'=f$, then $\int f(x)dx=F(x)+C$ on a connected interval. Different connected components can have independent constants. The notation does not by itself specify bounds or a numerical area. Initial or boundary data determine the constant when integration solves a differential equation.

Substitution is the chain rule read backward. If $u=g(x)$, then $\int f(g(x))g'(x)dx=\int f(u)du$. For definite integrals, transform the limits or return to the original variable, but do not do both. Monotonicity simplifies the substitution theorem; more general changes of variables account for multiplicity and orientation.

Integration by parts is the product rule integrated. $\int_a^b u,dv=[uv]_a^b-\int_a^b v,du$. It trades one integrand for another and underlies energy identities, Fourier coefficients, weak formulations, asymptotic estimates, and probability moments. Boundary terms are part of the result, not optional decorations. Their vanishing must follow from limits or boundary conditions.

Partial fractions reduce rational integrals after algebraic preparation. First divide improper rational functions, factor the denominator over the chosen field, and decompose repeated linear and irreducible quadratic factors. Integration then yields logarithmic, rational, and inverse-trigonometric terms. Factorization and coefficient solving are algebra steps; differentiation of the answer provides a decisive check.

Trigonometric integrals exploit identities and parity. Odd powers often reserve one factor for substitution, while even powers use half-angle identities. Trigonometric substitution maps square-root quadratics to identities but introduces domain and sign choices. Hyperbolic substitutions can be cleaner. A computer algebra expression may use different branches yet have the same derivative only on a restricted interval.

Improper integrals are limits, not ordinary endpoint evaluations. Infinite intervals and unbounded integrands are defined by one-sided limits. Each singular endpoint must be treated separately; cancellation across a singularity defines a Cauchy principal value, not the standard improper integral. Comparison, limit comparison, and $p$ tests determine convergence without antiderivatives. Units and positivity remain useful checks.

Average value divides accumulated quantity by domain measure. $f_{avg}=(b-a)^{-1}\int_a^b f$. For continuous $f$, the integral mean-value theorem guarantees some point where $f$ equals its average. In weighted averages, normalize by total weight. Confusing an average of a nonlinear function with the function of an average leads to Jensen-inequality errors.

Arc length accumulates local metric stretch. For a parametrized curve $\mathbf r(t)$, $L=\int_a^b\|\mathbf r'(t)\|dt$. The formula is invariant under regular orientation-preserving reparameterization. A graph gives $\int\sqrt{1+(y')^2}dx$. Corners can have finite length even without a derivative at the corner; fractal curves can be continuous with infinite length.

Work is a line integral of force along displacement. $W=\int_C\mathbf F\cdot d\mathbf r$ depends on both field and path. For a conservative field $\mathbf F=\nabla\phi$ on a suitable domain, it equals endpoint potential difference and closed-loop work is zero. Curl-free locally does not guarantee a global potential on a domain with holes. Orientation reversal changes the sign.

Accumulation functions solve rate equations directly. If inventory changes at rate $r(t)$, then $Q(t)=Q(t_0)+\int_{t_0}^t r(s)ds$. The derivative recovers the rate under continuity. This structure underlies charge, mass, energy, probability, cash flow, and population balance. A rate model and its accumulated state must use consistent sign, units, and initial condition.

<svg viewBox="0 0 760 470" xmlns="http://www.w3.org/2000/svg" font-family="Arial,sans-serif"><rect x="0" y="0" width="760" height="470" rx="18" fill="#0d1117"/><text x="380" y="34" fill="#e6edf3" font-size="21" font-weight="700" text-anchor="middle">Series require convergence before algebra</text><text x="380" y="60" fill="#8b949e" font-size="12" text-anchor="middle">Partial sums, tails, and a radius of convergence separate approximation from equality</text><line x1="75" y1="345" x2="690" y2="345" stroke="#8b949e"/><line x1="75" y1="345" x2="75" y2="90" stroke="#8b949e"/><line x1="95" y1="145" x2="670" y2="145" stroke="#3fb950" stroke-width="3" stroke-dasharray="8 7"/><text x="682" y="149" fill="#7ee787" font-size="12">S</text><polyline points="105,300 150,105 195,210 240,125 285,170 330,135 375,157 420,140 465,151 510,143 555,148 600,144 650,146" fill="none" stroke="#58a6ff" stroke-width="4"/><g fill="#d29922"><circle cx="105" cy="300" r="5"/><circle cx="150" cy="105" r="5"/><circle cx="195" cy="210" r="5"/><circle cx="240" cy="125" r="5"/><circle cx="285" cy="170" r="5"/><circle cx="330" cy="135" r="5"/><circle cx="375" cy="157" r="5"/><circle cx="420" cy="140" r="5"/><circle cx="465" cy="151" r="5"/><circle cx="510" cy="143" r="5"/><circle cx="555" cy="148" r="5"/><circle cx="600" cy="144" r="5"/><circle cx="650" cy="146" r="5"/></g><text x="380" y="385" fill="#e6edf3" font-size="13" text-anchor="middle">sₙ = Σₖ₌₀ⁿ aₖ  and  Rₙ = S − sₙ</text><text x="380" y="420" fill="#c9d1d9" font-size="11" text-anchor="middle">Convergence controls the infinite object; a remainder bound controls finite computation.</text></svg>

An infinite series is defined through its sequence of partial sums. The notation $\sum_{n=0}^{\infty}a_n=S$ means $s_N=\sum_{n=0}^{N}a_n$ approaches $S$ as $N$ grows. Terms tending to zero are necessary but not sufficient; the harmonic series is the standard warning. Absolute convergence permits rearrangement, while conditional convergence does not. Every numerical use truncates the series, so the remainder $R_N=S-s_N$ matters as much as convergence itself.

Convergence tests answer different structural questions. Comparison and limit comparison exploit positivity, the integral test links a decreasing sequence to an improper integral, and ratio or root tests detect geometric-scale decay. Alternating-series estimates can bound the tail when magnitudes decrease to zero. No single test is universal. Applying a test at its inconclusive boundary, especially a ratio limit of one, gives no conclusion and requires another argument.

Power series behave like polynomials inside their radius of convergence. For $\sum a_n(x-c)^n$, a radius $R$ separates absolute convergence for $|x-c|R$; endpoints need independent tests. Inside the interval, termwise differentiation and integration preserve the radius. Taylor series are power series whose coefficients are derivative data, but equality to the original function requires a remainder that vanishes. Analyticity is stronger than having derivatives of every order.

Fourier series expand periodic functions in trigonometric modes rather than powers. Smoothness controls coefficient decay and therefore approximation speed. At a jump, ordinary symmetric partial sums converge to the midpoint of the one-sided limits under standard hypotheses and exhibit Gibbs overshoot. Orthogonality computes coefficients by projection, while Parseval identities relate integrated energy to squared coefficients. A Fourier expansion is global: a localized feature influences every coefficient.

Uniform convergence is the key condition behind many exchanges of limits. Pointwise convergence allows the index needed for a tolerance to depend on position; uniform convergence supplies one index for the entire domain. Uniform limits of continuous functions remain continuous, and suitable uniform convergence permits integrating term by term. Differentiating a sequence of functions needs stronger control, usually convergence at a point plus uniform convergence of derivatives. Formal interchange without a theorem can change the answer.

Asymptotic notation compares behavior without claiming exact equality. Writing $f(x)=O(g(x))$ bounds the ratio in magnitude, $f(x)=o(g(x))$ makes that ratio tend to zero, and $f(x)\sim g(x)$ makes it tend to one. The limiting regime must be stated. A Taylor expansion with an $O(h^p)$ remainder communicates order but may hide a large constant; for engineering tolerances, a usable bound or empirical refinement study is still needed.

Parametric curves describe geometry without forcing a single-valued graph. Position $\mathbf r(t)$ has tangent $\mathbf r'(t)$, speed $\|\mathbf r'(t)\|$, and acceleration $\mathbf r''(t)$. Regularity fails where velocity vanishes, even if the traced set looks smooth. Curvature measures tangent rotation per unit arc length. A parameter can represent time, angle, distance, or an artificial coordinate, so its units and orientation must be carried through differentiation and integration.

Polar coordinates replace $x=r\cos\theta$ and $y=r\sin\theta$. A polar curve can traverse the same geometric point more than once or use negative radius, so bounds require geometric interpretation. Its area contribution is $\tfrac12 r^2d\theta$, and arc length is $\int\sqrt{r^2+(dr/d\theta)^2}\,d\theta$. Sketching the angular range prevents double counting and reveals where the curve changes direction.

The following distinctions prevent several common category errors.

<svg viewBox="0 0 760 470" xmlns="http://www.w3.org/2000/svg" font-family="Arial,sans-serif"><rect x="0" y="0" width="760" height="470" rx="18" fill="#0d1117"/><text x="380" y="34" fill="#e6edf3" font-size="21" font-weight="700" text-anchor="middle">Multivariable derivatives are linear maps</text><text x="380" y="60" fill="#8b949e" font-size="12" text-anchor="middle">Gradient, Jacobian, and Hessian encode first- and second-order response</text><ellipse cx="205" cy="245" rx="135" ry="95" fill="#161b22" stroke="#58a6ff" stroke-width="3"/><ellipse cx="555" cy="225" rx="135" ry="115" fill="#161b22" stroke="#a371f7" stroke-width="3"/><circle cx="205" cy="245" r="8" fill="#d29922"/><line x1="205" y1="245" x2="320" y2="165" stroke="#3fb950" stroke-width="5"/><polygon points="320,165 304,168 313,181" fill="#3fb950"/><text x="272" y="150" fill="#7ee787" font-size="12">direction v</text><path d="M345 245H415" stroke="#d29922" stroke-width="5"/><polygon points="415,245 400,236 400,254" fill="#d29922"/><text x="380" y="225" fill="#e3b341" font-size="13" text-anchor="middle">J</text><line x1="555" y1="225" x2="630" y2="145" stroke="#f778ba" stroke-width="5"/><polygon points="630,145 614,150 625,161" fill="#f778ba"/><text x="610" y="128" fill="#ff7b72" font-size="12">Jv</text><text x="205" y="390" fill="#79c0ff" font-size="13" text-anchor="middle">input tangent space</text><text x="555" y="390" fill="#d2a8ff" font-size="13" text-anchor="middle">output tangent space</text><text x="380" y="432" fill="#e6edf3" font-size="12" text-anchor="middle">f(x+h) = f(x) + J(x)h + o(‖h‖)</text></svg>

A multivariable derivative is one linear map that works in every direction. For $f:\mathbb R^n\to\mathbb R^m$, differentiability at $x$ means $f(x+h)=f(x)+J(x)h+o(\|h\|)$, where the Jacobian $J$ is the best local linear transformation. Existence of every directional derivative alone does not guarantee this uniform linear approximation. Continuous partial derivatives near the point are a common sufficient condition, not the definition.

The gradient converts directions into scalar directional rates. For scalar $f$, $D_{\mathbf v}f=\nabla f\cdot\mathbf v$ when differentiable. Among unit directions, the gradient gives greatest increase, its negative greatest decrease, and directions perpendicular to it are tangent to a regular level surface. Coordinates and metrics matter: in curvilinear coordinates the physical gradient is not obtained by simply listing coordinate partials.

The Hessian captures second-order curvature and variable interaction. Its entries are second partial derivatives, and the quadratic model is $f(x+h)\approx f(x)+\nabla f^Th+\tfrac12h^THh$. A positive-definite Hessian at a stationary point gives a strict local minimum; negative definite gives a maximum; mixed signs indicate a saddle. Semidefinite cases need higher-order or direct analysis. Scaling variables can radically improve the Hessian's condition number.

Constrained extrema satisfy geometric tangency conditions. For equality constraints $g_i(x)=0$, Lagrange multipliers express $\nabla f$ as a combination of active constraint gradients under regularity conditions. They are necessary conditions, not automatic optima. Boundaries, corners, inequality activity, constraint qualifications, and comparison of candidates remain essential. Multiplier values often represent marginal sensitivity of the optimum to relaxing a constraint.

Multiple integrals accumulate density over area or volume. Fubini-type results justify iterated integration when integrability conditions hold. Bounds describe the region, not just an algebraic ritual; reversing an inner bound changes sign. For positive quantities, Tonelli-type reasoning is especially forgiving, while conditionally convergent signed integrals can depend on integration order. A quick region sketch and unit check should precede calculation.

A change of variables replaces a region and density together. If $x=T(u)$ is locally invertible, the volume element gains $|\det DT(u)|$. The absolute determinant measures local volume scaling; orientation matters for differential forms but ordinary volume uses its magnitude. Polar, cylindrical, and spherical factors such as $r$ and $r^2\sin\phi$ are Jacobians, not mnemonic extras. Non-one-to-one maps require restricting domains or counting multiplicity.

The multivariable chain rule multiplies Jacobians in composition order. If $z=f(y)$ and $y=g(x)$, then $D(f\circ g)=Df(g(x))Dg(x)$. For scalar loss functions, reverse-mode automatic differentiation propagates covectors backward and evaluates many input sensitivities efficiently. The computed derivative is exact for the executed elementary operations up to floating-point effects, but it does not validate the model, resolve discontinuous branching, or cure an ill-conditioned problem.

The implicit function theorem turns a system of relations into local functions when the relevant Jacobian block is invertible. Its rank condition explains where a solution branch can fold, bifurcate, or cease to use the chosen coordinates. The inverse function theorem is its square-map counterpart. Both are local results; global uniqueness needs additional topology, monotonicity, or boundary information.

Differentials clarify propagation through many variables. For scalar $f$, $df=\sum_i(\partial f/\partial x_i)dx_i$ is a linear functional on a displacement, not a collection of independent infinitesimals. Total derivatives account for every path-dependent variable. Holding the wrong quantities fixed produces the wrong partial derivative, an especially important distinction in thermodynamics, materials models, and coupled simulations.

Sensitivity to parameters can be computed forward or adjoint. If a state satisfies $F(u,p)=0$, differentiation gives $F_u u_p=-F_p$. Forward sensitivity solves per parameter; an adjoint method can evaluate the gradient of one scalar objective with respect to many parameters at roughly one additional linear solve. Both rely on a consistent derivative of the governing residual and become unreliable near singular Jacobians or discontinuous events.

<svg viewBox="0 0 760 470" xmlns="http://www.w3.org/2000/svg" font-family="Arial,sans-serif"><rect x="0" y="0" width="760" height="470" rx="18" fill="#0d1117"/><text x="380" y="34" fill="#e6edf3" font-size="21" font-weight="700" text-anchor="middle">Vector calculus turns local field structure into boundary measurements</text><text x="380" y="60" fill="#8b949e" font-size="12" text-anchor="middle">Orientation determines every circulation and flux sign</text><path d="M165 125C95 185 100 310 190 350C285 392 360 315 350 215C340 120 245 78 165 125Z" fill="#161b22" stroke="#58a6ff" stroke-width="3"/><g stroke="#3fb950" stroke-width="3"><line x1="150" y1="180" x2="120" y2="145"/><line x1="255" y1="155" x2="275" y2="115"/><line x1="290" y1="265" x2="335" y2="260"/><line x1="195" y1="320" x2="180" y2="365"/></g><g fill="#3fb950"><polygon points="120,145 126,160 138,150"/><polygon points="275,115 261,124 274,132"/><polygon points="335,260 320,251 320,269"/><polygon points="180,365 174,349 190,353"/></g><path d="M450 150C520 95 635 130 660 230C685 325 565 375 475 320C405 277 395 195 450 150Z" fill="#161b22" stroke="#a371f7" stroke-width="3"/><path d="M470 265C485 180 585 160 635 230" fill="none" stroke="#f778ba" stroke-width="4"/><polygon points="635,230 619,223 625,244" fill="#f778ba"/><text x="225" y="225" fill="#79c0ff" font-size="15" text-anchor="middle">flux through ∂V</text><text x="550" y="265" fill="#ff7b72" font-size="15" text-anchor="middle">circulation on ∂S</text><text x="225" y="415" fill="#7ee787" font-size="12" text-anchor="middle">∫ᵥ div F dV = ∮∂ᵥ F·n dS</text><text x="550" y="415" fill="#d2a8ff" font-size="12" text-anchor="middle">∫ₛ curl F·n dS = ∮∂ₛ F·dr</text></svg>

Line integrals distinguish scalar accumulation from vector circulation. Integrating a scalar field along a curve uses $\int_C f\,ds$ and is independent of orientation. Integrating a vector field as $\int_C\mathbf F\cdot d\mathbf r$ measures tangential work or circulation and changes sign when orientation reverses. A parametrization supplies both location and differential displacement; reparametrization preserves the value when it preserves the geometric curve with appropriate orientation.

Surface integrals require a chosen normal orientation. Scalar surface area uses $dS=\|\mathbf r_u\times\mathbf r_v\|dudv$, while flux uses $\mathbf F\cdot(\mathbf r_u\times\mathbf r_v)dudv$. Swapping parameters reverses the oriented cross product. Closed surfaces conventionally use outward normal. An orientable surface admits a continuous normal field; a Möbius strip does not, so a global signed flux integral is not defined in the ordinary way.

The divergence theorem is a multidimensional fundamental theorem. It equates total source density $\int_V\nabla\cdot\mathbf F\,dV$ with outward boundary flux $\int_{\partial V}\mathbf F\cdot\mathbf n\,dS$. Internal face contributions cancel when adjacent cells use opposite normals, which explains finite-volume conservation. Singularities require care: a field can have zero classical divergence away from a point yet nonzero flux enclosing that point, represented by a distributional source.

Stokes' theorem equates surface curl with boundary circulation. With compatible orientations, $\int_S(\nabla\times\mathbf F)\cdot\mathbf n\,dS=\oint_{\partial S}\mathbf F\cdot d\mathbf r$. Green's theorem is the planar case, and the one-dimensional fundamental theorem is part of the same boundary-of-a-domain pattern. Reversing either the normal or boundary direction changes the sign. Smoothness and domain assumptions decide whether holes or singularities invalidate a shortcut.

Gradient, divergence, and curl have different input-output types. The gradient maps a scalar field to a vector, divergence maps a vector field to a scalar, and curl in three dimensions maps a vector field to a vector. Identities such as $\nabla\times\nabla f=0$ and $\nabla\cdot(\nabla\times\mathbf F)=0$ follow from commuting mixed partials under smoothness. Their converses need topological hypotheses; a punctured domain can carry closed but non-exact fields.

Flux expresses transport through a boundary. If density is $\rho$ and velocity is $\mathbf v$, the outward material flux is $\rho\mathbf v\cdot\mathbf n$. Combining a control-volume balance with the divergence theorem produces the local conservation law $\partial_t\rho+\nabla\cdot(\rho\mathbf v)=s$. The sign convention for sources and outward flow must be declared. Dimensional analysis distinguishes flux per area from an already integrated rate.

Coordinate formulas conceal geometric scale factors. In cylindrical and spherical systems, basis directions vary with position and the divergence, curl, and Laplacian include radius and angle terms. Deriving them from a coordinate-invariant theorem or a Jacobian is safer than treating them as Cartesian substitutions. Coordinate singularities at an axis or pole do not necessarily signal a singular physical field.

Differential forms unify these integral theorems by stating that integration of an exterior derivative over a region equals integration of the form over its oriented boundary. Even without formal form notation, the operational lesson is stable: pair the kind of field with the correct geometric element, orient the boundary consistently, and check whether the domain includes singularities.

<svg viewBox="0 0 760 470" xmlns="http://www.w3.org/2000/svg" font-family="Arial,sans-serif"><rect x="0" y="0" width="760" height="470" rx="18" fill="#0d1117"/><text x="380" y="34" fill="#e6edf3" font-size="21" font-weight="700" text-anchor="middle">Numerical calculus balances truncation and roundoff</text><text x="380" y="60" fill="#8b949e" font-size="12" text-anchor="middle">Refinement helps only until finite precision and data noise take over</text><line x1="90" y1="365" x2="680" y2="365" stroke="#8b949e"/><line x1="90" y1="365" x2="90" y2="95" stroke="#8b949e"/><path d="M110 115C210 180 300 270 420 350" fill="none" stroke="#58a6ff" stroke-width="4"/><path d="M110 350C250 345 360 325 455 230C525 160 585 125 660 110" fill="none" stroke="#f778ba" stroke-width="4"/><path d="M110 120C230 180 340 275 435 275C520 275 585 190 660 115" fill="none" stroke="#3fb950" stroke-width="5"/><text x="180" y="155" fill="#79c0ff" font-size="12">truncation</text><text x="560" y="155" fill="#ff7b72" font-size="12">roundoff/noise</text><circle cx="435" cy="275" r="8" fill="#d29922"/><text x="435" y="300" fill="#e3b341" font-size="11" text-anchor="middle">useful step scale</text><text x="385" y="405" fill="#e6edf3" font-size="13" text-anchor="middle">step size h decreases →</text><text x="32" y="235" fill="#e6edf3" font-size="13" text-anchor="middle" transform="rotate(-90 32 235)">error</text><text x="380" y="440" fill="#c9d1d9" font-size="11" text-anchor="middle">A refinement plateau or reversal is diagnostic information, not a request for more digits.</text></svg>

Finite differences approximate derivatives but also amplify noise. Forward difference has truncation error proportional to $h$, while a centered difference is typically second order for smooth data. Yet subtracting nearby nearly equal values loses significant digits, and measurement noise is divided by $h$. Therefore making $h$ smaller eventually worsens the result. Complex-step differentiation avoids subtractive cancellation for analytic code paths, while automatic differentiation avoids truncation but has its own model and implementation caveats.

Numerical quadrature combines sampling with an error model. Trapezoidal and Simpson rules approximate the integrand locally, Gaussian quadrature chooses nodes to integrate high-degree polynomials exactly, and adaptive rules refine where estimated error is large. Smooth periodic functions can make the trapezoidal rule exceptionally accurate. Endpoint singularities, discontinuities, oscillations, and narrow peaks need transformations or specialized schemes; a small estimated error is credible only when estimator assumptions match the integrand.

Conditioning and algorithmic stability are separate sources of reliability. Conditioning asks how much the exact answer changes when input data change; stability asks whether the algorithm adds avoidable error. No algorithm can recover information destroyed by severe ill-conditioning, although reformulation, scaling, regularization, or additional data may help. Comparing formulas algebraically before evaluation can prevent overflow, cancellation, and loss of relative accuracy.

Interpolation passes through data, while approximation need not. High-degree polynomial interpolation at equally spaced nodes can oscillate near endpoints even for smooth functions. Chebyshev-like nodes reduce worst-case polynomial error, and piecewise splines provide local control. Extrapolation is far riskier than interpolation because no surrounding data constrain it. Differentiating an interpolant magnifies its defects; integrating it often smooths them.

Root finding turns an equation $f(x)=0$ into an iteration. Bisection is slow but guaranteed for a continuous sign-changing bracket. Newton's method is locally fast when the derivative is reliable and the root is simple, but it can diverge, cross invalid domains, or converge to an unintended root. Secant methods avoid explicit derivatives. A robust solver commonly combines a bracket with safeguarded interpolation or Newton steps.

Optimization algorithms operationalize derivative information. Gradient descent uses first-order direction, Newton methods solve with the Hessian, and quasi-Newton methods learn curvature from gradients. Line searches or trust regions control whether a local model is trusted. Stopping because a step is small can be misleading under poor scaling; gradient norm, feasibility, objective change, and model agreement should be considered together.

Ordinary differential equations use calculus in both model and solver. An initial-value problem $y'=f(t,y)$ evolves from a specified state; Euler's method replaces the derivative with a finite step, while Runge–Kutta methods combine staged slopes for higher order. Local truncation error accumulates into global error. Stability can impose a step far smaller than accuracy alone, especially for stiff systems with widely separated decay rates.

Adaptive ODE solvers estimate local error, accept or reject a step, and change step size. Relative and absolute tolerances define a scale-dependent norm, so they must match the variables and desired observables. Event detection locates threshold crossings between steps. Conservation, positivity, symplectic structure, or monotonicity may matter more than high formal order for long simulations; structure-preserving methods are chosen accordingly.

Numerical integration of sampled data must distinguish resolution from precision. A dense table does not recover features omitted by the measurement bandwidth, and repeated digits do not imply accuracy. Baseline drift, endpoint handling, missing samples, and correlated noise can dominate a quadrature formula's textbook truncation term. Report the preprocessing, sampling interval, units, and uncertainty with the total.

Richardson extrapolation uses known leading error scaling. If $A(h)=A+c h^p+O(h^{p+1})$, combining results at $h$ and $h/2$ can cancel the leading term and estimate error. The observed order from several refinements tests whether the asymptotic regime has been reached. An irregular order often signals nonsmoothness, coding defects, solver tolerances, roundoff, or an incorrect theoretical model.

Floating-point arithmetic represents a finite nonuniform subset of the reals. Addition is not associative, overflow and underflow exist, and comparisons near a threshold need scale-aware tolerances. Stable primitives such as log1p, expm1, hypot, compensated summation, and scaled norms preserve information in common edge cases. Symbolic identities over real numbers are not automatically equivalent implementations in floating point.

st=>start: State the quantity, domain, units, and assumptions
op1=>operation: Choose limit, derivative, integral, series, or field theorem
cond1=>condition: Is the task symbolic or numerical?
op2=>operation: Derive with hypotheses and preserve constants, bounds, orientation
op3=>operation: Select stable discretization, step scale, tolerance, and estimator
cond2=>condition: Do units, signs, limits, and independent checks agree?
op4=>operation: Refine, compare, and diagnose the failed assumption
e=>end: Report result with domain and error or validity range
st->op1->cond1
cond1(yes)->op2->cond2
cond1(no)->op3->cond2
cond2(yes)->e
cond2(no)->op4->op1

A dependable calculus workflow begins by naming the mathematical object. Decide whether the requested quantity is local slope, accumulated total, extremum, limiting value, approximation, circulation, or flux. State variables, domains, units, regularity, and orientation. Choose a theorem only after checking its hypotheses. Carry an error or remainder when approximating, then verify by differentiation, integration, dimensional analysis, limiting cases, conservation, or independent computation.

Applications reveal why those distinctions matter. Kinematics links position, velocity, and acceleration through derivatives and integrals, but integration constants encode initial state. Mechanics obtains work from force along a path and potential from conservative force. Circuits relate charge and current, while capacitor and inductor laws turn rates and accumulations into differential equations. In every case, sign conventions and units are part of the model.

Probability uses integrals to normalize densities and compute expectations, derivatives to transform likelihoods and optimize estimates, and limits to define convergence. A probability density is not probability at a point; it must be integrated over a region. Changing variables requires a Jacobian. Tail probabilities often demand numerical methods designed to avoid catastrophic cancellation, and differentiation under an integral needs domination or comparable regularity.

Economics interprets derivatives as marginal quantities and constrained multipliers as shadow prices. A marginal cost is local and does not equal the average cost. Elasticity nondimensionalizes response as a fractional output change per fractional input change. Integrating a marginal curve recovers a change only with a baseline constant. Discontinuities, discrete decisions, strategic behavior, and uncertainty can limit a smooth calculus model.

Biology and chemistry use rate laws, compartment balances, gradients, and optimization. Exponential growth assumes a constant per-capita rate; logistic growth introduces a state-dependent limit. Reaction rates may be stiff and temperature sensitive. Dose-response curves can be differentiated locally, but extrapolation outside measured concentration ranges remains a modeling decision rather than a calculus consequence.

Geometry and graphics use parametric curves, surface normals, curvature, Jacobians, and optimization. A transformation's determinant predicts local area or volume scaling and whether orientation reverses. Rendering and inverse problems often differentiate through a pipeline, but visibility changes introduce nonsmooth points. Mesh resolution, coordinate charts, and normal consistency affect the numerical result.

Data science uses gradients to fit models and integrals to average over distributions. A gradient computed perfectly can still optimize the wrong loss, inherit sampling bias, or exploit a data leak. Regularization changes the objective and thus the optimum. Nondifferentiable penalties can be handled with subgradients or proximal methods. Training convergence is an algorithmic observation, not proof of statistical validity.

Calculus also identifies when a local model is insufficient. Near a bifurcation, singular Jacobian, shock, phase transition, contact event, or topology change, small input perturbations can produce non-small qualitative changes. Generalized derivatives, weak solutions, distributions, measure theory, or nonsmooth analysis may be the correct extension. Recognizing that boundary is part of competent calculus, not a failure of it.

A symbolic result should be tested on its domain. Differentiate an antiderivative, substitute an implicit derivative back into the relation, compare an optimization candidate with boundaries, and examine singular points excluded during algebra. Branches of logarithms, roots, and inverse trigonometric functions can make apparently equivalent formulas differ by constants or signs. Computer algebra output inherits these domain issues.

A numerical result should be accompanied by convergence evidence. Repeat with smaller steps or tighter tolerances, compare methods with different failure modes, monitor conserved quantities, and distinguish solver error from uncertain input. Agreement of many digits between closely related algorithms is weaker than agreement between independent formulations. Report only digits supported by conditioning, discretization, and data quality.

The deepest organizing idea is local-to-global reasoning. Derivatives compress local response into a linear map; integrals assemble local density over a domain; the fundamental theorem relates the two; integral theorems move derivatives between interiors and boundaries. Limits certify all of these transitions. Series and numerical methods then replace infinite definitions with finite computations whose remainders can be controlled.

The subject developed from older geometric, astronomical, and mechanical problems rather than from one isolated invention. Ancient exhaustion arguments anticipated limit-based area calculation; seventeenth-century work by Newton and Leibniz organized systematic differential and integral methods; later analysis supplied precise definitions of limit, continuity, and convergence. Modern notation is a compressed language built over that logical foundation. Historical priority disputes do not change the practical point that calculus became powerful through a combination of algorithms, interpretations, and rigor.

Learning calculus is more durable when representations are translated deliberately. A derivative can be a limit, tangent slope, local linear coefficient, physical rate, graph feature, or sensitivity entry; an integral can be a Riemann-sum limit, signed area, net change, expectation, work, or flux. These meanings agree only when their assumptions and geometric elements match. Moving among formula, graph, table, units, and verbal interpretation exposes misconceptions that symbol manipulation can conceal.

A useful diagnostic for a limit asks what varies, what stays fixed, which side or path approaches, and whether the claimed value is finite. For a derivative, identify input and output units, the held-fixed variables, and the domain where linearization is accurate. For an integral, identify density, measure, bounds, orientation, and whether the integral is proper. For a series, identify partial sums, convergence mode, and truncation error. For an optimization, identify feasible points before stationary points.

Exact and approximate answers serve different purposes. An exact expression may reveal symmetry, scaling, and parameter dependence yet be numerically unstable or impossible to evaluate cheaply. A numerical approximation may be operationally superior but needs tolerances and verification. Hybrid practice derives identities symbolically, simplifies with domain awareness, evaluates using stable algorithms, and checks against asymptotics or conservation. More elaborate notation is not automatically more rigorous; explicit assumptions and controlled error are.

Dimensionless groups often reveal the true variables of a problem. Rescaling $x=L\hat x$ and $f=F\hat f$ separates units from shape, improves numerical conditioning, and shows which parameter ratios govern behavior. A derivative rescales by $F/L$, an $n$-dimensional integral by $FL^n$, and a Hessian by $F/L^2$. Limits such as a small dimensionless parameter approaching zero can then be interpreted without taking the limit of a dimensional quantity ambiguously.

Counterexamples define the edges of familiar rules. Continuity need not imply differentiability; existence of partial derivatives need not imply multivariable differentiability; a zero derivative need not identify an extremum; terms approaching zero need not make a series converge; curl-free need not mean globally conservative on a domain with holes; and decreasing numerical step size need not decrease total error. Remembering one concrete counterexample for each implication prevents unjustified theorem reversal.

When communicating a solution, state the result before the algebra and attach its validity conditions. Show the decisive transformation, not every routine manipulation. Include constants of integration, transformed limits, orientation, and uncertainty. If a theorem establishes existence without constructing the object, say so. If a plotted or computed result is evidence rather than proof, label it accordingly. This makes the reasoning auditable by someone who uses a different notation or software system.

NIST's Digital Library of Mathematical Functions collects standard one-variable and multivariable calculus identities with conditions and notation, while MIT calculus materials present the definite integral and both directions of the fundamental theorem as the bridge between rates and totals. These references reinforce a sound practice: formulas should be read together with domains, endpoint behavior, smoothness, and convergence conditions, not detached from them.

Read calculus through a limit-linearization-and-accumulation lens rather than a formula-and-symbol-manipulation lens.