Reference

Glossary

Every technical word in the course and what it means in plain language, with flashcards for practice.

Every technical term in the course, with a plain-language meaning. On every page these words are marked with a dotted underline: hover or tap one to see its meaning. The “Used in” links point to the sections where each term appears.

Flashcards — term ⇄ meaning
Ch 1 · First-Order Ordinary Differential Equations
TermMeaningUsed in
asymptotically stableAn equilibrium that is stable and attracts every solution starting close enough to it. For $x'=Ax$: iff every eigenvalue has negative real part; in the plane iff $\operatorname{tr}A\lt0$ and $\det A\gt0$.1.3 8.2 8.4
blow-upThe solution becomes unbounded, $|y(x)|\to\infty$, as $x$ approaches a finite point; e.g. $y=1/(1-x)$ for $y'=y^2$, $y(0)=1$.1.1 1.3 2.1 2.2
c-discriminantThe relation obtained by eliminating $c$ from $\varphi(x,y,c)=0$ and $\partial\varphi/\partial c=0$. It contains envelopes, node-loci and cusp-loci of the family.1.2
Clairaut's equationThe equation $y=px+f(p)$, $p=y'$. Its general solution is the family of lines $y=cx+f(c)$; for $f''\ne0$ its singular solution is their envelope $x=-f'(p)$, $y=f(p)-pf'(p)$.1.2
companion matrixThe matrix of the first-order system obtained from $y^{(n)}+a_{n-1}y^{(n-1)}+\dots+a_0y=0$ with $x=(y,y',\dots,y^{(n-1)})$: ones above the diagonal and last row $(-a_0,\dots,-a_{n-1})$.1.3 2.1
cusp-locusThe locus of cusps of the members of a family; it typically appears once in the $p$-discriminant and cubed in the $c$-discriminant, and is not a solution.1.2
envelopeA curve that touches every member of a one-parameter family of curves, each at its own point. It satisfies $\varphi=0$, $\partial\varphi/\partial c=0$; an envelope of solution curves is itself a solution.1.2 2.1 3.2 6.2
exact equationAn equation $M\,dx+N\,dy=0$ with $M=\Phi_x$, $N=\Phi_y$ for some $\Phi$; on a rectangle this holds iff $M_y=N_x$, and the solutions satisfy $\Phi(x,y)=c$.1.1
fundamental matrixA matrix $\Phi(t)$ whose columns form a basis of the solutions of $x'=A(t)x$; equivalently $\Phi'=A\Phi$ and $\det\Phi\ne0$. Every solution is $\Phi(t)c$, and $e^{tA}$ is the one with $\Phi(0)=I$ when $A$ is constant.1.3 2.2
generalised eigenvectorA nonzero $v$ with $(A-\lambda I)^kv=0$ for some $k$; these span $G_\lambda=\ker(A-\lambda I)^m$ ($m$ the algebraic multiplicity), on which $e^{tA}v=e^{\lambda t}\sum_{k\lt m}\frac{t^k}{k!}(A-\lambda I)^kv$.1.3
Gronwall's inequalityIf $u\le C+\int_{x_0}^x k u$ with $k\ge0$, then $u\le C\exp\int_{x_0}^x k$. It gives uniqueness and continuous dependence of solutions on the data.1.1 1.3 4.3 5.5
initial value problemA differential equation $y'=f(x,y)$ together with a starting condition $y(x_0)=y_0$; a solution is a differentiable function satisfying both on an interval containing $x_0$.1.1 1.2 1.3 2.1
integrating factorA nonzero function $\mu$ such that $\mu M\,dx+\mu N\,dy=0$ is exact; for the linear equation $y'+Py=Q$ it is $e^{\int P\,dx}$.1.1 2.1 8.1
Liouville's formulaFor a matrix solution of $x'=A(t)x$: $\det\Phi(t)=\det\Phi(t_0)\exp\int_{t_0}^t\operatorname{tr}A$. So the Wronskian of solutions vanishes nowhere or everywhere; Abel's formula is the scalar special case.1.3 2.1 8.2
Lipschitz condition$|f(x,y_1)-f(x,y_2)|\le L|y_1-y_2|$ for all points of the domain with the same $x$. It holds, for example, when $|\partial f/\partial y|\le L$ on a domain convex in $y$, and it guarantees uniqueness of solutions.1.1 1.3 5.3 7.1
matrix exponential$e^{M}=\sum_{k\ge0}M^k/k!$. The function $e^{tA}$ solves $\Phi'=A\Phi$, $\Phi(0)=I$, so $x(t)=e^{tA}x_0$; $e^{A+B}=e^Ae^B$ holds when $AB=BA$.1.3
maximal interval of existenceThe largest open interval $(\omega_-,\omega_+)$ on which a solution of an IVP can be defined. At a finite end the solution leaves every compact subset of the domain, e.g. by blowing up.1.1
node-locusThe locus of double points (nodes) of the members of a family; it typically appears squared in the $c$-discriminant and is not a solution.1.2
orthogonal trajectoryA curve cutting every member of a family at right angles; found by replacing $p$ by $-1/p$ in the family's differential equation (in polar form, $dr/d\theta$ by $-r^2\,d\theta/dr$).1.2 3.1
p-discriminantThe relation obtained by eliminating $p=y'$ from $F(x,y,p)=0$ and $\partial F/\partial p=0$: the points where two slopes coincide. It contains every singular solution, and also tac-loci and cusp-loci.1.2
phase portraitThe picture of all trajectories of an autonomous system $x'=f(x)$ in the state space, with arrows in the direction of increasing $t$.1.3 8.2
Picard iteratesThe functions $\varphi_0=y_0$, $\varphi_{n+1}(x)=y_0+\int_{x_0}^x f(t,\varphi_n(t))\,dt$. Under a Lipschitz condition they converge uniformly to the unique solution of the IVP.1.1 1.3 5.1 5.5
saddle pointAn equilibrium of a planar system with two trajectories entering and two leaving along different lines; for $x'=Ax$ it occurs exactly when $\det A\lt0$, and it is unstable.1.3 6.2 8.2 8.3
singular solutionA solution of $F(x,y,y')=0$ that is not a member of the general solution for any value of the constant, and along which uniqueness fails; typically the envelope of the general solution, e.g. $y=\pm1$ for $(y')^2+y^2=1$.1.2 3.2
spiral pointAn equilibrium around which trajectories wind infinitely often while approaching it (stable) or leaving it (unstable); for $x'=Ax$ it occurs when $(\operatorname{tr}A)^2\lt4\det A$ and $\operatorname{tr}A\ne0$.1.3
tac-locusThe locus of points where two non-neighbouring members of a family touch; it typically appears squared in the $p$-discriminant and is not a solution.1.2
Ch 2 · Linear Ordinary Differential Equations
TermMeaningUsed in
Abel's formulaFor solutions of $y^{(n)}+p_{n-1}y^{(n-1)}+\dots+p_0y=0$: $W(x)=W(x_0)\exp\left(-\int_{x_0}^xp_{n-1}\right)$; for $y''+py'+qy=0$, $W=W_0e^{-\int p}$.1.3 2.1
annihilatorA constant-coefficient operator $A(D)$ with $A(D)r=0$; e.g. $(D-a)^{k+1}$ annihilates $x^ke^{ax}$ and $D^2+\beta^2$ annihilates $\cos\beta x$, $\sin\beta x$. It explains the trial forms of undetermined coefficients.2.2
Bessel function$J_\nu(x)=\sum_k\frac{(-1)^k}{k!\,\Gamma(k+\nu+1)}\left(\frac x2\right)^{2k+\nu}$, a solution of $x^2y''+xy'+(x^2-\nu^2)y=0$; e.g. $J_{1/2}=\sqrt{2/(\pi x)}\,\sin x$ and $J_0'=-J_1$.2.1 2.2 5.3
bilinear expansionThe eigenfunction series $G(x,t;\mu)=\sum_n\varphi_n(x)\varphi_n(t)/(\mu-\lambda_n)$ of a Green's function in normalised Sturm–Liouville eigenfunctions; it shows the poles of $G$ at the eigenvalues and gives trace identities such as $\sum n^{-2}=\pi^2/6$.2.4 7.2
boundary value problemA differential equation on $[a,b]$ with conditions imposed at both end points, e.g. $y''+y=f$, $y(0)=y(\pi)=0$. Unlike an initial value problem it may have no solution or infinitely many; it has exactly one for every $f$ iff the homogeneous problem has only the trivial solution.2.1 2.3 2.4 4.3
causal Green's functionThe Green's function of an initial value problem $L[u]=f$, $u(a)=\dots=u^{(n-1)}(a)=0$: equal to the impulse response $K(x,t)$ for $t\le x$ and to $0$ for $t\gt x$. It is never symmetric.2.4
characteristic rootFor $y^{(n)}+a_{n-1}y^{(n-1)}+\dots+a_0y=0$ with constant coefficients, a root of the characteristic polynomial $P(\lambda)=\lambda^n+a_{n-1}\lambda^{n-1}+\dots+a_0$; a root $\lambda$ of multiplicity $m$ contributes the solutions $x^ke^{\lambda x}$, $k\lt m$.2.1 2.2 2.3
complementary functionThe general solution $y_c=c_1y_1+\dots+c_ny_n$ of the homogeneous equation $L[y]=0$; every solution of $L[y]=r$ is $y_c+y_p$.2.2 4.1
eigenfunctionA nonzero solution of a homogeneous boundary value problem containing a parameter $\lambda$, which exists only for special values of $\lambda$ (the eigenvalues); e.g. $\sin nx$ for $y''+\lambda y=0$, $y(0)=y(\pi)=0$, $\lambda=n^2$.2.3 2.4 4.2 4.3
Euler–Cauchy equationA linear equation $x^ny^{(n)}+b_{n-1}x^{n-1}y^{(n-1)}+\dots+b_0y=0$ with constant $b_k$. The trial $y=x^r$ (or $x=e^t$) reduces it to constant coefficients; repeated roots bring in powers of $\log x$.2.1 4.2
Fredholm alternativeEither $L[u]=f$ (with homogeneous boundary conditions) has exactly one solution for every $f$, or the homogeneous problem has a nontrivial solution $\varphi$; in the self-adjoint case the problem is then solvable iff $\int f\varphi=0$, and solutions differ by multiples of $\varphi$.2.3 2.4 6.1 7.1
fundamental set of solutionsA basis of the solution space of a homogeneous linear equation of order $n$: $n$ linearly independent solutions, equivalently $n$ solutions with nonzero Wronskian. Every solution is a unique linear combination of them.1.3 2.1 2.2 2.3
Green's functionThe kernel $G(x,t)$ with $u(x)=\int_a^bG(x,t)f(t)\,dt$ for every right-hand side of a linear boundary value problem: the response at $x$ to a unit point source at $t$. For $-u''=f$, $u(0)=u(1)=0$ it is $\min(x,t)(1-\max(x,t))$. It exists iff the homogeneous problem has only the trivial solution.1.3 2.2 2.4 6.2
indicial equationAt a regular singular point $0$ with $xP=\sum p_kx^k$, $x^2Q=\sum q_kx^k$: $r(r-1)+p_0r+q_0=0$. Its roots are the possible leading exponents $r$ of Frobenius solutions.2.1 2.2 2.4
jump conditionThe requirement that $\partial_xG(x,t)$ jump by $1/p(t)$ across $x=t$ for $L[u]=(pu')'+qu$ (by $1$ in normal form, by $-1$ for $-u''$), while $G$ itself stays continuous; for an order-$n$ equation the $(n-1)$-th derivative jumps.2.4
Legendre polynomialThe polynomial solution $P_n$ of $(1-x^2)y''-2xy'+n(n+1)y=0$ with $P_n(1)=1$; $P_n=\frac{1}{2^nn!}\frac{d^n}{dx^n}(x^2-1)^n$, and $\int_{-1}^1P_mP_n=\frac{2}{2n+1}\delta_{mn}$.2.1 2.2 2.3 5.4
Maxwell–Betti reciprocityFor a linear elastic structure, the deflection at $x$ due to a unit load at $t$ equals the deflection at $t$ due to a unit load at $x$ — the symmetry $G(x,t)=G(t,x)$ of the Green's function of a self-adjoint problem.2.4
modified Green's functionA kernel used when the homogeneous problem has a normalised solution $\varphi$: it satisfies $L_xG_M=\delta(x-t)-\varphi(x)\varphi(t)$, the boundary conditions and $\int G_M\varphi\,dx=0$, and gives the solution orthogonal to $\varphi$ when $\int f\varphi=0$.2.4
particular integralAny one solution $y_p$ of a non-homogeneous linear equation $L[y]=r$, without arbitrary constants. It is unique only up to adding a solution of $L[y]=0$.2.2 3.2 4.1
Prüfer substitutionThe polar coordinates $y=\rho\sin\theta$, $py'=\rho\cos\theta$ for a solution of $(py')'+(q+\lambda r)y=0$; the angle obeys $\theta'=\frac1p\cos^2\theta+(q+\lambda r)\sin^2\theta$, crosses multiples of $\pi$ (zeros of $y$) only upwards, and counts eigenvalues and zeros.2.3
Rayleigh quotient$R[u]=\dfrac{-[puu']_a^b+\int(pu'^2-qu^2)}{\int ru^2}$. It equals $\lambda$ at an eigenfunction, and its minimum over trial functions satisfying the Dirichlet conditions is the first eigenvalue $\lambda_1$; so every trial function gives an upper bound for $\lambda_1$.2.3 4.3 6.1 6.2
reduction of orderFinding a second solution from a known solution $y_1$ by substituting $y=vy_1$; for $y''+py'+qy=0$ it gives $y_2=y_1\int e^{-\int p}/y_1^2\,dx$.1.3 2.1 2.2
regular singular pointA singular point $x_0$ of $y''+Py'+Qy=0$ at which $(x-x_0)P$ and $(x-x_0)^2Q$ are analytic. Frobenius series $x^r\sum a_kx^k$ (possibly with a $\log x$ term) solve the equation there.2.2
resonanceThe forcing frequency coincides with a characteristic root of multiplicity $s$; the particular integral then carries an extra factor $x^s$, e.g. $y''+y=\sin x$ has $y_p=-\tfrac x2\cos x$, which is unbounded.1.3 2.1 2.2 6.1
Robin conditionA boundary condition $\beta_1y(b)+\beta_2y'(b)=0$ with both coefficients nonzero (a mixture of Dirichlet $y=0$ and Neumann $y'=0$). Its sign matters: $y'(1)+hy(1)=0$ with $h\ge0$ keeps eigenvalues positive, $h\lt0$ can create a negative one.2.3 4.3 6.1 6.2
self-adjoint formThe form $(p\,y')'+q\,y+\lambda r\,y=0$ of a second-order linear equation, obtained from $a_2y''+a_1y'+a_0y+\lambda by=0$ by multiplying by $p/a_2$ with $p=e^{\int a_1/a_2}$. In it Green's formula holds and the weight $r$ can be read off.2.3 2.4 6.2
Sturm comparison theoremIf $p_1\ge p_2\gt0$ and $q_2\ge q_1$, then between two zeros of a nontrivial solution of $(p_1u')'+q_1u=0$ lies a zero of every solution of $(p_2v')'+q_2v=0$ (unless $v$ is a multiple of $u$). Consequence: if $m^2\le Q\le M^2$, zeros of $u''+Qu=0$ are between $\pi/M$ and $\pi/m$ apart.1.1 2.1 2.3
Sturm separation theoremThe zeros of two linearly independent solutions of $y''+p(x)y'+q(x)y=0$ (continuous $p,q$) strictly interlace: between consecutive zeros of one lies exactly one zero of the other.2.1
Sturm–Liouville problemThe eigenvalue problem $(py')'+qy+\lambda ry=0$ on $[a,b]$ with separated boundary conditions $\alpha_1y(a)+\alpha_2y'(a)=0$, $\beta_1y(b)+\beta_2y'(b)=0$; regular when $p\gt0$, $r\gt0$ on the closed bounded interval. Its eigenvalues are real, simple and increase to $\infty$.2.3 2.4 4.2 4.3
variation of parametersLagrange's method for a particular integral: replace the constants of $y_c$ by functions. For $y''+py'+qy=r$ (normal form): $y_p=-y_1\int y_2r/W+y_2\int y_1r/W$.1.3 2.1 2.2 2.4
weight functionThe positive coefficient $r$ in $(py')'+qy+\lambda ry=0$. Eigenfunctions of distinct eigenvalues satisfy $\int_a^b r\,\varphi_m\varphi_n\,dx=0$ — orthogonality with respect to $r$, not (in general) with respect to $1$.2.3
WronskianFor functions $y_1,\dots,y_n$, the determinant $W=\det\big(y_j^{(i-1)}\big)$ of their values and first $n-1$ derivatives; for two functions $W=y_1y_2'-y_1'y_2$. For solutions of one linear equation it vanishes either nowhere or identically.1.3 2.1 2.2 2.4
Ch 3 · First-Order Partial Differential Equations
TermMeaningUsed in
Cauchy problemFor a first-order PDE: find the integral surface containing a given curve $\Gamma$, i.e. a solution taking prescribed values on a curve in the plane.3.1 3.2 4.1 4.2
characteristic curveFor $Pp+Qq=R$: an integral curve of $\frac{dx}{P}=\frac{dy}{Q}=\frac{dz}{R}$. Integral surfaces are unions of characteristic curves, and along them the PDE reduces to ODEs.3.1 3.2 4.1
characteristic stripA solution $(x,y,z,p,q)(t)$ of Charpit's equations: a curve together with a tangent plane at each of its points; integral surfaces of nonlinear equations are woven from such strips.3.1 3.2
Charpit's methodA method for a complete integral of $F(x,y,z,p,q)=0$: find a first integral $G=a$ of $\frac{dx}{F_p}=\frac{dy}{F_q}=\frac{dz}{pF_p+qF_q}=\frac{dp}{-(F_x+pF_z)}=\frac{dq}{-(F_y+qF_z)}$, solve for $p,q$ and integrate $dz=p\,dx+q\,dy$.3.1 3.2
Clairaut formThe first-order PDE $z=px+qy+f(p,q)$; its complete integral is the family of planes $z=ax+by+f(a,b)$, and its singular integral is their envelope.1.2 3.1 3.2
compatible equationsTwo first-order PDEs $F=0$, $G=0$ with a common solution through every point; when $\partial(F,G)/\partial(p,q)\ne0$ this holds iff the bracket $[F,G]$ vanishes on $F=G=0$.3.2
complete integralA two-parameter family of solutions $z=\phi(x,y,a,b)$ of a first-order PDE in which both parameters are essential (a rank-2 condition on the derivatives of $\phi$ with respect to $a,b$).3.2
first integralA function with nonzero gradient that is constant along every solution curve of a system of ODEs; for $\frac{dx}{P}=\frac{dy}{Q}=\frac{dz}{R}$ it satisfies $Pu_x+Qu_y+Ru_z=0$.3.1 3.2 4.1 6.1
general integralA solution formula of a first-order PDE containing an arbitrary function: $\varphi(u,v)=0$ for Lagrange's equation; for a nonlinear equation, the envelope of a one-parameter subfamily of a complete integral.3.1 3.2
integral surfaceThe graph of a solution $z=z(x,y)$ of a first-order PDE; for $Pp+Qq=R$ it is a surface to which the vector $(P,Q,R)$ is tangent at every point.3.1 3.2
Jacobi's methodCharpit's idea for $F(x_1,x_2,x_3,p_1,p_2,p_3)=0$ (unknown absent): find $F_1=a_1$, $F_2=a_2$ with all Jacobi brackets zero, then integrate $du=\sum p_i\,dx_i$.3.2
Lagrange's auxiliary equationsThe system $\frac{dx}{P}=\frac{dy}{Q}=\frac{dz}{R}$ attached to the quasi-linear equation $Pp+Qq=R$; its solution curves are the characteristic curves.3.1
method of multipliersFinding a first integral of $\frac{dx}{P}=\frac{dy}{Q}=\frac{dz}{R}$ by functions $\ell,m,n$ with $\ell P+mQ+nR=0$ and $\ell\,dx+m\,dy+n\,dz=dw$ exact; then $w=$ const.3.1
orthogonal surfacesSurfaces cutting every member of a family $f(x,y,z)=c$ at right angles; they solve the Lagrange equation $f_x\,p+f_y\,q=f_z$.3.1
quasi-linear equationA first-order PDE linear in the derivatives, $P(x,y,z)\,p+Q(x,y,z)\,q=R(x,y,z)$, whose coefficients may depend on the unknown $z$ (Lagrange's equation).3.1 3.2 4.1
semilinear equationA first-order PDE $a(x,y)\,p+b(x,y)\,q=c(x,y,z)$: the derivatives enter linearly with coefficients free of $z$, but the right side may be nonlinear in $z$.3.1 4.1
singular integralThe envelope of the whole two-parameter family of a complete integral, when it is a solution; along it $F=F_p=F_q=0$.3.2
transversality conditionFor $Pp+Qq=R$ with data on $(x_0(s),y_0(s),z_0(s))$: $x_0'Q-y_0'P\ne0$, i.e. the data curve is not tangent to the projected characteristics; it guarantees a unique local solution.3.1 3.2 4.1 6.1
Ch 4 · Second-Order Partial Differential Equations
TermMeaningUsed in
canonical formThe simplest form of a second-order equation obtained by a change of variables: $u_{\xi\eta}=\Phi$ or $u_{\alpha\alpha}-u_{\beta\beta}=\Phi$ (hyperbolic), $u_{\eta\eta}=\Phi$ (parabolic), $u_{\alpha\alpha}+u_{\beta\beta}=\Phi$ (elliptic).1.3 4.1 4.3
characteristic coordinatesNew independent variables $\xi,\eta$ whose level curves are the characteristics of a second-order equation; in them the coefficients of $u_{\xi\xi}$ and $u_{\eta\eta}$ vanish.4.1 4.3
d'Alembert's formulaThe solution $u=\frac12[f(x+ct)+f(x-ct)]+\frac1{2c}\int_{x-ct}^{x+ct}g(s)\,ds$ of $u_{tt}=c^2u_{xx}$ on $\mathbb{R}$ with $u(x,0)=f$, $u_t(x,0)=g$.4.3
Dirichlet problemFind $u$ with $\Delta u=F$ in $\Omega$ and prescribed values $u=g$ on $\partial\Omega$. On a bounded domain it has at most one solution, by the maximum principle.2.3 4.2 6.2
domain of dependenceFor the wave equation, the interval $[x_0-ct_0,\,x_0+ct_0]$ of initial data that determine $u(x_0,t_0)$; it is cut out by the two characteristics through $(x_0,t_0)$.4.3
domain of influenceFor the wave equation, the wedge $|x-x_1|\le ct$ of points whose values can be affected by the initial data at $x_1$; outside it a disturbance from $x_1$ is not felt (finite speed of propagation).4.3
elliptic equationA second-order equation with $B^2-4AC\lt0$ at the point considered, i.e. with a definite coefficient matrix. It has no real characteristics; the model is Laplace's equation $u_{xx}+u_{yy}=0$.4.1 4.2
energy methodProving uniqueness or stability by showing that an energy integral — $\int u^2$ for the heat equation, $\frac12\int(u_t^2+c^2u_x^2)$ for the wave equation — cannot increase, so zero data give zero solution.4.2 4.3
Fourier seriesThe expansion $\frac{a_0}2+\sum(a_n\cos\frac{n\pi x}L+b_n\sin\frac{n\pi x}L)$ of a $2L$-periodic function; for piecewise $C^1$ functions it converges to $\frac12(f(x^+)+f(x^-))$, uniformly if the periodic function is continuous.2.3 4.2 4.3 6.2
Gibbs phenomenonNear a jump of $f$, the partial sums of its Fourier series overshoot by about $9\%$ of the jump on each side, however many terms are taken; the convergence there is not uniform.2.3 4.2
harmonic functionA function $u\in C^2(\Omega)$ with $\Delta u=u_{xx}+u_{yy}=0$ in $\Omega$; for example $x^2-y^2$, $e^x\cos y$, $r^n\cos n\theta$, and the real and imaginary parts of every analytic function.4.1 4.2
heat equationThe parabolic equation $u_t=k\,u_{xx}$ ($k\gt0$ the diffusivity) for the temperature in a rod; on $[0,L]$ with zero ends its modes decay like $e^{-kn^2\pi^2t/L^2}$.2.3 4.1 4.2 4.3
heat kernel$\Phi(x,t)=(4\pi kt)^{-1/2}e^{-x^2/(4kt)}$, the temperature due to a unit point source of heat; the solution on $\mathbb{R}$ is $u=\int\Phi(x-y,t)f(y)\,dy$.4.3
hyperbolic equationA second-order equation $Au_{xx}+Bu_{xy}+Cu_{yy}+\dots=0$ with $B^2-4AC\gt0$ at the point considered. It has two real families of characteristics; the model is the wave equation $u_{tt}=c^2u_{xx}$.4.1 4.3
maximum principleA non-constant harmonic (or subharmonic) function on a bounded domain attains its maximum only on the boundary: $\max_{\overline\Omega}u=\max_{\partial\Omega}u$. It gives uniqueness and stability for the Dirichlet problem; the heat equation has an analogous principle on the parabolic boundary.2.3 2.4 4.2 4.3
mean value propertyFor a harmonic $u$, the value at the centre of any closed disc in the domain equals the average of $u$ over the boundary circle (and over the disc).4.2 5.3
Neumann problemFind $u$ with $\Delta u=F$ in $\Omega$ and prescribed normal derivative $\partial u/\partial n=g$ on $\partial\Omega$. It is solvable only if $\iint_\Omega F=\oint_{\partial\Omega}g$, and solutions are unique up to a constant.2.3 4.2 4.3 6.2
parabolic boundaryFor the rectangle $[0,L]\times[0,T]$, the bottom $t=0$ together with the two sides $x=0$, $x=L$ (not the top). The maximum principle for the heat equation says a solution attains its maximum there.4.3
parabolic equationA second-order equation with $B^2-4AC=0$ (and $A,B,C$ not all zero) at the point considered. It has one double family of characteristics; the model is the heat equation $u_t=ku_{xx}$.
Poisson kernel$P(r,\psi)=\dfrac{R^2-r^2}{R^2-2Rr\cos\psi+r^2}$; the solution of the Dirichlet problem in a disc is $u(r,\theta)=\frac1{2\pi}\int_{-\pi}^{\pi}P(r,\theta-\varphi)f(\varphi)\,d\varphi$, a positive weighted average of the boundary values.4.2 4.3
principal partThe terms of highest order of a differential equation; for a second-order equation in two variables, $Au_{xx}+Bu_{xy}+Cu_{yy}$. The type and the characteristics depend only on the principal part.4.1 4.3
reducible operatorA constant-coefficient operator $F(D,D')$ that factors into first-order linear factors $\alpha D+\beta D'+\gamma$; otherwise it is irreducible (e.g. the heat operator $D^2-D'$). Each real factor $D-mD'-c$ contributes $e^{cx}\varphi(y+mx)$ to the complementary function.4.1
separation of variablesSeeking solutions of a linear PDE as products such as $X(x)Y(y)$ or $X(x)T(t)$; the factors satisfy ODEs, one of which is a Sturm–Liouville problem, and the general solution is a series of such products.1.1 2.3 4.1 4.2
steady stateA time-independent solution of an evolution equation with its boundary conditions; for the heat equation on a rod with end temperatures $T_1$, $T_2$ it is the straight line $T_1+(T_2-T_1)x/L$.2.2 4.2 4.3
Tricomi equationThe equation $y\,u_{xx}+u_{yy}=0$ of transonic gas flow: elliptic for $y\gt0$, parabolic on $y=0$ and hyperbolic for $y\lt0$, where its characteristics are $x\pm\tfrac23(-y)^{3/2}=\text{const}$.4.1
wave equationThe hyperbolic equation $u_{tt}=c^2u_{xx}$ for small vibrations of a string with speed $c=\sqrt{T/\rho}$; its solutions are superpositions $F(x+ct)+G(x-ct)$ of two travelling waves.2.2 3.1 4.1 4.3
Ch 5 · Numerical Analysis
TermMeaningUsed in
A-stableA method is A-stable if its region of absolute stability contains the whole left half-plane $\operatorname{Re}z\lt0$. Backward Euler and the implicit trapezoidal rule are A-stable; no explicit Runge–Kutta method is.5.5
bisection methodThe bracketing method that halves an interval $[a,b]$ with $f(a)f(b)\lt0$ at every step, keeping the half on which $f$ changes sign. For continuous $f$ the $n$-th midpoint satisfies $|c_n-\alpha|\le(b-a)/2^n$.5.1
Chebyshev nodesThe zeros $x_i=\cos\frac{(2i+1)\pi}{2n+2}$ of the Chebyshev polynomial $T_{n+1}(x)=\cos((n+1)\arccos x)$. They minimise $\max_{[-1,1]}|\prod(x-x_i)|$ (the minimum is $2^{-n}$) and cluster near $\pm1$.5.3
Cholesky factorisationThe factorisation $A=LL^{\mathsf T}$ with $L$ lower triangular with positive diagonal; it exists (uniquely) exactly when $A$ is symmetric positive definite, and costs about $n^3/6$ multiplications.5.2
condition number$\kappa(A)=\|A\|\,\|A^{-1}\|\ge1$, the largest possible factor by which a relative error in $b$ (or $A$) can be amplified in the solution of $Ax=b$. For symmetric $A$ in the 2-norm it is $|\lambda|_{\max}/|\lambda|_{\min}$.5.2
contractionA map $g$ with $|g(x)-g(y)|\le L|x-y|$ for all $x,y$ and a fixed constant $L\lt1$. A contraction of a closed interval into itself has exactly one fixed point, and every fixed-point iteration converges to it.1.1 5.1 7.1
cubic splineA $C^2$ function that is a cubic polynomial between consecutive knots. The interpolating cubic spline is unique once two end conditions are fixed: natural ($s''=0$ at both ends) or clamped (end slopes prescribed). The natural spline minimises $\int(s'')^2$ among all $C^2$ interpolants.5.2 5.3
degree of precisionThe largest $d$ such that a quadrature rule integrates every polynomial of degree $\le d$ exactly. Simpson's 1/3 and 3/8 rules have degree $3$; the $n$-point Gauss–Legendre rule has degree $2n-1$, the maximum possible with $n$ nodes.5.4
diagonally dominant matrixA matrix is strictly (row) diagonally dominant if $|a_{ii}|\gt\sum_{j\ne i}|a_{ij}|$ in every row. Such a matrix is nonsingular, and the Jacobi and Gauss–Seidel iterations converge for it from every starting vector.5.2 5.3
divided differenceDefined by $f[x_i]=f(x_i)$ and $f[x_i,\dots,x_{i+k}]=\dfrac{f[x_{i+1},\dots,x_{i+k}]-f[x_i,\dots,x_{i+k-1}]}{x_{i+k}-x_i}$. It is symmetric in its arguments, equals the leading coefficient of the interpolating polynomial, and equals $f^{(k)}(\xi)/k!$ for some $\xi$ between the nodes.5.1 5.3
efficiency indexFor a method of order $p$ that needs $w$ function evaluations per step, the number $p^{1/w}$: about $1.618$ for the secant method and $\sqrt2\approx1.414$ for Newton's method (which needs $f$ and $f'$).5.1
Euler's methodThe one-step method $y_{n+1}=y_n+hf(x_n,y_n)$: follow the tangent line for one step. Local truncation error $\tfrac{h^2}{2}y''(\xi)$, global error $O(h)$; absolutely stable on $y'=\lambda y$ iff $|1+h\lambda|\lt1$.5.5 8.2
Euler–Maclaurin formulaThe expansion $T(h)-I=\tfrac{h^2}{12}[f'(b)-f'(a)]-\tfrac{h^4}{720}[f'''(b)-f'''(a)]+\cdots$ of the error of the composite trapezoidal rule in even powers of $h$, the basis of Romberg integration.5.4
finite differencesFor equally spaced nodes $x_i=x_0+ih$: $\Delta f_i=f_{i+1}-f_i$ (forward), $\nabla f_i=f_i-f_{i-1}$ (backward), $\delta f_i=f_{i+1/2}-f_{i-1/2}$ (central), with higher differences defined recursively. For a polynomial of degree $n$ with leading coefficient $a$, $\Delta^nf\equiv n!\,a\,h^n$.4.1 5.1 5.3 5.4
fixed-point iterationThe iteration $x_{n+1}=g(x_n)$ for an equation written as $x=g(x)$. Near a fixed point $\alpha$ it converges if $|g'(\alpha)|\lt1$ (linearly with rate $|g'(\alpha)|$) and diverges if $|g'(\alpha)|\gt1$.5.1
Gaussian quadratureThe $n$-node rule whose nodes are the zeros of the Legendre polynomial $P_n$ (on $[-1,1]$) and whose weights are $\int\ell_i=\int\ell_i^2\gt0$. It is exact for all polynomials of degree $\le2n-1$; e.g. nodes $\pm1/\sqrt3$ with weights $1,1$ for $n=2$.5.4
Gauss–Seidel iterationThe iterative method for $Ax=b$ that, like Jacobi's, solves equation $i$ for $x_i$ but uses each new component as soon as it is computed; its iteration matrix is $T_{GS}=-(D+L)^{-1}U$. For tridiagonal matrices $\rho(T_{GS})=\rho(T_J)^2$.4.2 5.2
global errorThe difference $e_n=y(x_n)-y_n$ between the exact and computed solutions after $n$ steps. For a method of order $p$ with a Lipschitz increment function it is $O(h^p)$ on a fixed interval.5.4 5.5
Hermite interpolationInterpolation of values and derivatives: at $n+1$ distinct nodes, prescribing $f(x_i)$ and $f'(x_i)$ determines a unique polynomial of degree $\le2n+1$, with error $\frac{f^{(2n+2)}(\xi)}{(2n+2)!}\prod(x-x_i)^2$.5.3
ill-conditioned systemA linear system (or matrix) with a large condition number: tiny changes in the data cause large changes in the solution, and a small residual does not guarantee a small error. Nearly parallel lines and Hilbert matrices are the standard examples.5.1 5.2
interpolating polynomialThe unique polynomial $p$ of degree $\le n$ with $p(x_i)=y_i$ at $n+1$ distinct nodes; in Lagrange's form $p=\sum y_iL_i$ with $L_i(x)=\prod_{j\ne i}\frac{x-x_j}{x_i-x_j}$. Its error is $\frac{f^{(n+1)}(\xi)}{(n+1)!}\prod(x-x_i)$.5.3 5.4
Jacobi iterationThe iterative method for $Ax=b$ that computes every component of $x^{(k+1)}$ from the equation $\sum_ja_{ij}x_j=b_i$ using only the components of $x^{(k)}$; its iteration matrix is $T_J=-D^{-1}(L+U)$.5.2
Lebesgue constantFor interpolation at given nodes, $\Lambda_n=\max_x\sum_i|L_i(x)|$: the worst factor by which errors in the data are amplified, with $\|f-p_n\|\le(1+\Lambda_n)\min_{q}\|f-q\|$. It grows exponentially for equally spaced nodes and like $\tfrac2\pi\log n$ for Chebyshev nodes.5.3
local truncation errorThe amount $T_{n+1}=y(x_{n+1})-y(x_n)-h\Phi(x_n,y(x_n);h)$ by which the exact solution fails to satisfy one step of a method. A method has order $p$ if $T_{n+1}=O(h^{p+1})$.5.4 5.5
LU factorisationA factorisation $A=LU$ with $L$ lower and $U$ upper triangular; Doolittle's form has $l_{ii}=1$, Crout's has $u_{ii}=1$. A nonsingular $A$ has a Doolittle factorisation without row interchanges iff its leading principal minors $\Delta_1,\dots,\Delta_{n-1}$ are nonzero.5.2
modified Euler methodHeun's predictor–corrector: $y^*=y_n+hf(x_n,y_n)$, then $y_{n+1}=y_n+\tfrac h2[f(x_n,y_n)+f(x_{n+1},y^*)]$. A two-stage Runge–Kutta method of order $2$; for $y'=f(x)$ it is the trapezoidal rule.5.5
multiple rootA root $\alpha$ of multiplicity $m\ge2$: $f(x)=(x-\alpha)^mh(x)$ with $h(\alpha)\ne0$, so $f(\alpha)=f'(\alpha)=0$. Multiple roots are not bracketed when $m$ is even, slow Newton's method down to linear convergence and are sensitive to perturbations.5.1
Newton–Cotes formulaAn interpolatory quadrature rule with equally spaced nodes. The closed rules with $2,3,4,5$ nodes are the trapezoidal, Simpson 1/3, Simpson 3/8 and Boole rules; with an odd number of nodes the degree of precision is one more than expected, by symmetry.5.4
Newton–Raphson methodThe iteration $x_{n+1}=x_n-f(x_n)/f'(x_n)$: the next guess is the zero of the tangent line. It converges quadratically to a simple root from nearby starting values, but only linearly (ratio $1-1/m$) to a root of multiplicity $m$.5.1 5.2
order of convergenceA sequence $x_n\to\alpha$ converges with order $p\ge1$ if $|x_{n+1}-\alpha|/|x_n-\alpha|^p\to C$ with $0\lt C\lt\infty$ (and $C\lt1$ when $p=1$). Order $1$ adds a fixed number of correct digits per step; order $p\gt1$ multiplies the number of correct digits by about $p$.5.1 7.2
partial pivotingIn Gauss elimination, choosing at each step the entry of largest absolute value in the current column (on or below the diagonal) as the pivot and interchanging rows accordingly. It keeps all multipliers at most $1$ in size and gives $PA=LU$ for every nonsingular $A$.5.2
quadrature ruleAn approximation $\int_a^bf\approx\sum_iw_if(x_i)$ with nodes $x_i$ and weights $w_i$. It is interpolatory when $w_i=\int_a^b\ell_i$, i.e. when it integrates the polynomial interpolating $f$ at the nodes.5.4 5.5
region of absolute stabilityFor a method giving $y_{n+1}=R(h\lambda)y_n$ on $y'=\lambda y$, the set of $z=h\lambda$ with $|R(z)|\lt1$. Euler: the disc $|1+z|\lt1$; real intervals $(-2,0)$ for Euler and RK2, $(-2.785,0)$ for RK4.5.5
regula falsiA bracketing method that replaces the midpoint of bisection by the zero of the chord through $(a,f(a))$ and $(b,f(b))$. If $f''$ has constant sign one end point never moves and the convergence is only linear.5.1
Richardson extrapolationIf $A(h)=A+ch^p+O(h^q)$ with $q\gt p$, then $\dfrac{2^pA(h/2)-A(h)}{2^p-1}=A+O(h^q)$: two approximations with steps $h$ and $h/2$ combine to cancel the leading error term.5.4 5.5
Romberg integrationRepeated Richardson extrapolation of the composite trapezoidal rule with $h,h/2,h/4,\dots$, based on the Euler–Maclaurin expansion of its error in even powers of $h$. The first extrapolated column is composite Simpson, the second composite Boole.5.4
Runge's phenomenonThe divergence of polynomial interpolants at equally spaced nodes as their number grows, for functions such as $1/(1+25x^2)$ on $[-1,1]$: the error tends to $\infty$ near the ends of the interval. Chebyshev nodes or splines avoid it.5.3
Runge–Kutta methodA one-step method $y_{n+1}=y_n+h\sum b_ik_i$ whose stages $k_i=f(x_n+c_ih,\,y_n+h\sum_{j\lt i}a_{ij}k_j)$ are slopes at trial points. The classical RK4 uses four stages with weights $\tfrac16(1,2,2,1)$ and has order $4$.1.1 1.3 2.1 2.2
secant methodThe derivative-free iteration $x_{n+1}=x_n-f(x_n)\,\dfrac{x_n-x_{n-1}}{f(x_n)-f(x_{n-1})}$, which uses the chord through the two latest iterates. Its order at a simple root is the golden ratio $\tfrac{1+\sqrt5}{2}\approx1.618$.5.1
shift operatorThe operator $E$ with $Ef(x)=f(x+h)$. The difference operators are polynomials in $E^{1/2}$: $\Delta=E-1$, $\nabla=1-E^{-1}$, $\delta=E^{1/2}-E^{-1/2}$; on polynomials $E=e^{hD}$.5.3
Simpson's ruleThe 1/3 rule $\int_{x_0}^{x_2}f\approx\tfrac h3(f_0+4f_1+f_2)$ integrates the interpolating parabola; its error is $-\tfrac{h^5}{90}f^{(4)}(\xi)$ per panel and $O(h^4)$ for the composite rule. The 3/8 rule $\tfrac{3h}8(f_0+3f_1+3f_2+f_3)$ has error $-\tfrac{3h^5}{80}f^{(4)}(\xi)$.5.4 5.5
spectral radius$\rho(T)=\max|\lambda|$ over the (complex) eigenvalues of $T$. An iteration $x^{(k+1)}=Tx^{(k)}+c$ converges for every start iff $\rho(T)\lt1$, and the error eventually shrinks by about the factor $\rho(T)$ per step.5.2
stiff equationAn equation or system with rapidly decaying components alongside the slow behaviour of interest (eigenvalues of the Jacobian with large negative real part). Explicit methods then need very small steps for stability, not for accuracy.5.5
successive over-relaxationThe modification $x_i^{\rm new}=(1-\omega)x_i^{\rm old}+\omega\,\tilde x_i$ of Gauss–Seidel ($\tilde x_i$ the Gauss–Seidel value). It can converge for every start only if $0\lt\omega\lt2$ (Kahan); a well-chosen $\omega\gt1$ can speed up slow iterations dramatically.5.2
Taylor series methodThe one-step method $y_{n+1}=y_n+hy'_n+\tfrac{h^2}{2}y''_n+\cdots+\tfrac{h^p}{p!}y^{(p)}_n$, with the derivatives obtained by differentiating $y'=f(x,y)$; it has order $p$.5.5
trapezoidal rule$\int_a^bf\approx\tfrac h2(f_0+f_1)$, error $-\tfrac{h^3}{12}f''(\xi)$; the composite rule $T_n$ has error $-\tfrac{(b-a)h^2}{12}f''(\xi)$. For smooth periodic integrands over a full period it converges faster than any power of $h$.5.4 5.5 7.2
Ch 6 · Calculus of Variations
TermMeaningUsed in
Beltrami identityFor $F=F(y,y')$ not depending on $x$: $F-y'F_{y'}=$ const along every extremal, a first integral of the Euler–Lagrange equation.6.1 8.2
brachistochroneThe curve of fastest frictionless descent between two points under gravity; it is an arc of a cycloid $x=a(\theta-\sin\theta)$, $y=a(1-\cos\theta)$.6.1 6.2
conjugate pointA point $a^*\gt a$ where the solution of the Jacobi equation with $u(a)=0$, $u'(a)=1$ vanishes; nearby extremals from $(a,y(a))$ cross the extremal again there. A minimum cannot contain one in its interior.6.1 6.2 8.2
Dirichlet principleThe solution of $-\Delta u=f$ in $\Omega$, $u=g$ on $\partial\Omega$ minimises $\iint(\frac12|\nabla v|^2-fv)$ among functions with the same boundary values (when a minimiser exists).6.1 6.2
energy functionalFor $-(pu')'+qu=f$, $u(a)=u(b)=0$ with $p\gt0$, $q\ge0$: $I[v]=\int(pv'^2+qv^2-2fv)$, whose unique minimiser is the solution.6.2
Euler–Lagrange equationThe necessary condition $F_y-\frac{d}{dx}F_{y'}=0$ satisfied by every (weak) extremum of $\int F(x,y,y')\,dx$; a second-order ODE when $F_{y'y'}\ne0$.6.1 6.2 8.1 8.2
extremalA solution of the Euler–Lagrange equation — a candidate for an extremum, which need not actually be a minimum or maximum.6.1 6.2 7.2 8.2
first variationThe directional derivative $\delta J[y;\eta]=\frac{d}{d\varepsilon}J[y+\varepsilon\eta]\big|_{\varepsilon=0}=\int(F_y\eta+F_{y'}\eta')\,dx$; it vanishes at every weak extremum.6.1 6.2 8.2
functionalA map that assigns a number to each function of an admissible class, e.g. $J[y]=\int_a^bF(x,y,y')\,dx$ (a length, a time, an area, an energy).5.4 6.1 6.2 8.1
fundamental lemma of the calculus of variationsIf $M$ is continuous and $\int_a^bM\eta\,dx=0$ for every smooth $\eta$ vanishing at $a$ and $b$, then $M\equiv0$. It turns $\delta J=0$ into a differential equation.6.1 8.2
Galerkin methodChoose the approximation in a trial space so that the residual $Lu_n-f$ is orthogonal to that space; equals Rayleigh–Ritz for symmetric problems and also works for non-symmetric ones.6.2
isoperimetric problemExtremising $\int F$ subject to an integral constraint $\int G=\ell$; the extremals satisfy the Euler–Lagrange equation of $F+\lambda G$ with a constant multiplier $\lambda$.6.1
Jacobi conditionNo point conjugate to $a$ in $(a,b)$ (necessary for a weak minimum when $P\gt0$); together with $P\gt0$, no conjugate point in $(a,b]$ is sufficient.6.1 6.2 8.2
Jacobi equationThe Euler–Lagrange equation $-(Pu')'+Qu=0$ of the second variation; its solutions are the variations $\partial y/\partial c$ of families of extremals.6.2
Legendre conditionThe necessary condition $F_{y'y'}(x,y,y')\ge0$ along an extremal that is a weak minimum; the strengthened form $F_{y'y'}\gt0$ is part of the sufficient conditions.6.2
natural boundary conditionThe condition a minimiser must satisfy at an end where its value is not prescribed: $F_{y'}=0$ there (plus $\Gamma'(y)$ if a boundary term $\Gamma(y(b))$ is present).6.1 6.2
Rayleigh–Ritz methodMinimise the energy (or Rayleigh quotient) over a finite-dimensional trial space; gives a linear system $Kc=F$ (or $\det(K-\Lambda M)=0$) and upper bounds for energies and eigenvalues.2.3 6.1 6.2
second variation$\delta^2J[y;\eta]=\frac{d^2}{d\varepsilon^2}J[y+\varepsilon\eta]\big|_0=\int(P\eta'^2+Q\eta^2)\,dx$ with $P=F_{y'y'}$, $Q=F_{yy}-\frac{d}{dx}F_{yy'}$; it is $\ge0$ at a weak minimum.6.1 6.2
weak and strong extremumWeak: $J[y]\ge J[y_0]$ for curves close to $y_0$ together with their slopes ($\|y-y_0\|_1$ small). Strong: for all curves uniformly close ($\|y-y_0\|_0$ small), slopes arbitrary. Strong implies weak.4.2 4.3 6.1 6.2
Weierstrass excess function$E(x,y,p,q)=F(x,y,q)-F(x,y,p)-(q-p)F_{y'}(x,y,p)$; a strong minimum requires $E\ge0$ for every slope $q$.6.2
Ch 7 · Linear Integral Equations
TermMeaningUsed in
Abel's integral equationThe singular Volterra equation of the first kind $\int_0^x\frac{y(t)}{\sqrt{x-t}}\,dt=f(x)$; for $f\in C^1$, $f(0)=0$, its solution is $y=\frac1\pi\int_0^x\frac{f'(t)}{\sqrt{x-t}}\,dt$. It arose from the tautochrone problem.7.1
characteristic numberA value of $\lambda$ for which the homogeneous equation $y=\lambda\int_a^bK(x,t)y(t)\,dt$ has a continuous solution $y\not\equiv0$ (an eigenfunction). It is the reciprocal of an eigenvalue of the integral operator; for a separable kernel the characteristic numbers are the roots of $\det(I-\lambda A)$.7.1 7.2
convolution kernelA Volterra kernel depending only on $x-t$, $K=k(x-t)$, so that $\int_0^xk(x-t)y(t)\,dt$ is a convolution; the Laplace transform turns $y=f+\lambda k*y$ into $Y=F/(1-\lambda\hat k)$.7.1 7.2
equation of the second kindAn integral equation in which the unknown also appears outside the integral, $y=f+\lambda\int Ky$; in an equation of the first kind it appears only under the integral, $\int Ky=f$, which is typically ill-posed.7.1
Fredholm determinantThe entire function $D(\lambda)$, with $D(0)=1$, whose zeros are the characteristic numbers of a continuous kernel; for a separable kernel $D(\lambda)=\det(I-\lambda A)$, and the resolvent is $R=D(x,t;\lambda)/D(\lambda)$.7.2
Fredholm equationA linear integral equation whose limits of integration are fixed: $\int_a^bK(x,t)y(t)\,dt=f(x)$ (first kind) or $y=f+\lambda\int_a^bK(x,t)y(t)\,dt$ (second kind). Boundary value problems lead to Fredholm equations.2.4 7.1 7.2
Hilbert–Schmidt norm$\|K\|_2=\left(\iint|K(x,t)|^2\,dx\,dt\right)^{1/2}$. Every characteristic number satisfies $|\lambda|\ge1/\|K\|_2$, and for symmetric kernels $\sum1/\lambda_n^2=\|K\|_2^2$.7.2
Hilbert–Schmidt theoremFor a continuous symmetric kernel, every function $g=\int K h$ expands as $\sum(g,\varphi_n)\varphi_n$ in the orthonormal eigenfunctions, uniformly; and $\sum1/\lambda_n^2=\iint K^2$.2.3 7.2
integral equationAn equation in which the unknown function appears under an integral sign, such as $y(x)=f(x)+\lambda\int_a^bK(x,t)y(t)\,dt$. It is linear when the unknown enters linearly; the given function $K(x,t)$ is its kernel.1.1 2.4 7.1 7.2
iterated kernelThe kernels $K_1=K$, $K_{n+1}(x,t)=\int_a^bK(x,s)K_n(s,t)\,ds$ of the powers of the integral operator: $\mathcal{K}^nf=\int K_nf$.7.2
Mercer's theoremIf a continuous symmetric kernel is positive (all characteristic numbers $\gt0$), then $K(x,t)=\sum_n\varphi_n(x)\varphi_n(t)/\lambda_n$ uniformly and $\sum1/\lambda_n=\int_a^bK(x,x)\,dx$.7.2
Neumann seriesThe series $y=f+\lambda\mathcal{K}f+\lambda^2\mathcal{K}^2f+\cdots$ for the solution of a second-kind equation, the analogue of the geometric series for $(I-\lambda\mathcal{K})^{-1}$. It converges if $|\lambda|M(b-a)\lt1$ or $|\lambda|\,\|K\|_2\lt1$; its true radius is the smallest modulus of a characteristic number.7.1 7.2
resolvent kernelThe kernel $R(x,t;\lambda)=\sum_{n\ge1}\lambda^{n-1}K_n(x,t)$ (and its continuation), giving the solution of $y=f+\lambda\int Ky$ as $y=f+\lambda\int R(x,t;\lambda)f(t)\,dt$. Its poles are the characteristic numbers; for a Volterra kernel it is entire in $\lambda$.7.2
separable kernelA kernel of the form $K(x,t)=\sum_{i=1}^na_i(x)b_i(t)$, e.g. $xt$, $x+t$, $\cos(x-t)$. For it, $y=f+\lambda\int Ky$ reduces to the linear system $(I-\lambda A)c=F$ with $A_{ij}=\int b_ia_j$.7.1 7.2
symmetric kernelA real kernel with $K(x,t)=K(t,x)$ (Hermitian: $K(x,t)=\overline{K(t,x)}$). A nonzero continuous symmetric kernel has at least one characteristic number; all are real and eigenfunctions of distinct ones are orthogonal.2.3 2.4 7.1 7.2
transposed kernelThe kernel $K^T(x,t)=K(t,x)$. At a characteristic number, $y=f+\lambda\int Ky$ is solvable iff $\int fz=0$ for every solution $z$ of the transposed homogeneous equation $z=\lambda\int K(t,x)z(t)\,dt$.7.1 7.2
Volterra equationA linear integral equation whose upper limit is the variable: $\int_a^xK(x,t)y(t)\,dt=f(x)$ (first kind) or $y=f+\lambda\int_a^xK(x,t)y(t)\,dt$ (second kind). Initial value problems lead to Volterra equations; with a continuous kernel the second-kind equation is uniquely solvable for every $\lambda$.1.1 7.1 7.2
Ch 8 · Classical Mechanics
TermMeaningUsed in
beatsThe slow periodic swelling and fading of an oscillation made of two nearby frequencies, $\cos\omega_1t+\cos\omega_2t=2\cos\tfrac{(\omega_2-\omega_1)t}2\cos\tfrac{(\omega_1+\omega_2)t}2$. In weakly coupled identical oscillators the energy passes from one to the other in time $\pi/(\omega_2-\omega_1)$.8.4
canonical equationsHamilton's first-order equations of motion $\dot q_j=\partial H/\partial p_j$, $\dot p_j=-\partial H/\partial q_j$ in phase space.8.2
canonical transformationA change of phase-space variables that preserves the form of Hamilton's equations; for one degree of freedom (time-independent) it is characterised by $[Q,P]_{q,p}=1$.8.2
cyclic coordinateA generalised coordinate that does not appear in the Lagrangian; its conjugate momentum $\partial L/\partial\dot q_j$ is conserved.8.1 8.2 8.3
D'Alembert's principleFor ideal (workless) constraints, $\sum_i(\mathbf F_i-m_i\ddot{\mathbf r}_i)\cdot\delta\mathbf r_i=0$ for every virtual displacement; it eliminates the constraint forces and leads to Lagrange's equations.8.1 8.2
degrees of freedomThe number $n=3N-k$ of independent coordinates needed to fix the configuration of $N$ particles subject to $k$ independent holonomic constraints.2.2 8.1 8.2 8.3
Euler's dynamical equationsThe equations $A\dot\omega_1-(B-C)\omega_2\omega_3=N_1$ and the two obtained by cycling the indices, for a rigid body moving about a fixed point or its centre of mass, written in principal axes fixed in the body.8.3
Eulerian anglesThree angles $(\phi,\theta,\psi)$ — precession, nutation (tilt) and spin — describing the orientation of body axes relative to space axes by successive rotations about $\mathbf z$, the line of nodes and $\mathbf e_3$. In them $\omega_3=\dot\psi+\dot\phi\cos\theta$.8.1 8.3
generalised coordinatesIndependent parameters $q_1,\dots,q_n$ (lengths, angles, …) in terms of which every particle position $\mathbf r_i(q,t)$ satisfying the constraints can be written.8.1 8.3 8.4
Hamilton's principleAmong all paths with the same end configurations at the same end times, the actual motion makes the action $S=\int L\,dt$ stationary; its Euler–Lagrange equations are Lagrange's equations.6.1 8.2
HamiltonianThe Legendre transform $H(q,p,t)=\sum p_j\dot q_j-L$ of the Lagrangian with respect to the velocities; it equals $T+V$ for scleronomic systems with velocity-independent potentials.8.2
holonomic constraintA constraint expressible as an equation $f(\mathbf r_1,\dots,\mathbf r_N,t)=0$ between the coordinates (and possibly time). Inequalities and non-integrable velocity relations are non-holonomic.6.1 8.1 8.3
inertia tensorThe symmetric positive semidefinite matrix $\mathcal I_O=\sum m(|\mathbf r|^2E-\mathbf r\mathbf r^T)$ at a point $O$. It gives the moment about any line through $O$ as $\mathbf n^T\mathcal I_O\mathbf n$, and the angular momentum $\mathbf L_O=\mathcal I_O\vec\omega$ of a body turning about $O$.8.3
instantaneous centreIn plane motion with $\omega\ne0$, the unique point (rigidly attached to the body) whose velocity is zero at that instant; every other point then moves perpendicular to the line joining it to this point, with speed $|\omega|$ times the distance. Its acceleration is generally not zero.8.3
Jacobi integralThe energy function $h=\sum_j\dot q_j\,\partial L/\partial\dot q_j-L$, conserved when $L$ has no explicit time dependence; it equals $T_2-T_0+V$, and $T+V$ only for scleronomic systems.8.1 8.4
Lagrange–Dirichlet theoremIf the potential energy of a natural (conservative, scleronomic) system has a strict local minimum at an equilibrium, that equilibrium is stable. The proof uses only conservation of energy; the converse is false in general.8.4
Legendre transformFor a strictly convex $f(v)$, the function $g(p)=pv-f(v)$ with $p=f'(v)$; it satisfies $g'(p)=v$ and is again convex. It turns a Lagrangian into a Hamiltonian.8.2
Liouville's theoremThe flow of Hamilton's equations preserves phase-space volume, because the Hamiltonian vector field has zero divergence.4.2 8.2
moment of inertiaFor a rigid body and a line $\ell$, $I_\ell=\sum m\,d^2$, the sum of mass times squared distance from $\ell$. It plays the role of mass for rotation about $\ell$: $L=I\dot\theta$ and $T=\tfrac12I\dot\theta^2$ for rotation about a fixed axis.8.1 8.3
momental ellipsoidThe surface $\mathbf x^T\mathcal I_O\mathbf x=1$ at a point $O$: its distance from $O$ in the direction $\mathbf n$ is $1/\sqrt{I_{\mathbf n}}$, and its axes are the principal axes at $O$. It is a sphere for a spherical top.8.3
non-holonomic constraintA constraint that cannot be written as an equation between coordinates: an inequality, or a relation between velocities that does not integrate (a disc rolling on a plane).
normal coordinatesCoordinates $\xi=P^{-1}\eta$, with $P$ the matrix of $\mathsf T$-orthonormal mode vectors, in which the quadratic Lagrangian separates into independent oscillators: $L_2=\tfrac12\sum(\dot\xi_k^2-\omega_k^2\xi_k^2)$.8.4
normal frequencyThe frequency $\omega_k=\sqrt{\lambda_k}$ of a normal mode, $\lambda_k$ a root of the secular equation. The normal frequencies are the square roots of the eigenvalues of $\mathsf T^{-1}\mathsf V$ and do not depend on the choice of coordinates.1.3 8.4
normal modeA small oscillation $\eta(t)=\mathbf a\cos(\omega t+\varphi)$ in which every coordinate moves with the same frequency and phase, the amplitudes in the fixed ratios of the mode vector $\mathbf a$, where $(\mathsf V-\omega^2\mathsf T)\mathbf a=\mathbf 0$. Every small oscillation is a superposition of normal modes.4.3 8.1 8.4
phase spaceThe $2n$-dimensional space of generalised coordinates and momenta $(q,p)$ in which Hamilton's equations define a flow.8.2
Poisson bracket$[f,g]=\sum_j(\partial_{q_j}f\,\partial_{p_j}g-\partial_{p_j}f\,\partial_{q_j}g)$; it gives $\dot f=[f,H]+\partial f/\partial t$, and $[q_j,p_k]=\delta_{jk}$.8.2
principal axes of inertiaThe eigenvector directions of the inertia tensor at a point; in principal axes all products of inertia vanish and the tensor is $\operatorname{diag}(A,B,C)$, the principal moments. Every point has three mutually perpendicular principal axes; axes and planes of symmetry are principal.8.3
principle of least actionMaupertuis–Jacobi principle: among paths of fixed energy $E$ between two configurations, the actual orbit makes $\int2T\,dt=\int\sqrt{2(E-V)}\,ds$ stationary; orbits are geodesics of the metric $2(E-V)\,ds^2$.8.2
product of inertiaWith respect to axes $Oxyz$, $I_{xy}=\sum m\,xy$ (and similarly $I_{yz}$, $I_{zx}$); they appear with a minus sign as off-diagonal entries of the inertia tensor. They depend on the origin and the axes, and all vanish exactly when the axes are principal.8.3
radius of gyrationThe length $k$ with $I=Mk^2$: the distance at which the whole mass, concentrated, would have the same moment of inertia about the given line. For a uniform rod about its centre $k=\ell/(2\sqrt3)$.8.3
rheonomic constraintA constraint that contains time explicitly, such as a bead on a wire rotating at a prescribed rate. A time-independent constraint is scleronomic.8.1 8.2 8.4
rolling without slippingMotion of a body on a surface in which the material point in contact has zero velocity; for a wheel of radius $a$ on a line, $v_G=a\omega$. The friction force is static, does no work and is whatever the motion requires (at most $\mu R$).8.1 8.3
Routhian$R=L-\sum_c p_c\dot q_c$, a Legendre transform in the cyclic velocities only; it is a Lagrangian for the remaining coordinates (with the opposite sign convention in some books).8.1 8.2
scleronomic constraintA constraint that does not contain time explicitly. With scleronomic coordinates the kinetic energy is a quadratic form in the generalised velocities.8.1 8.2 8.4
secular equationThe equation $\det(\mathsf V-\lambda\mathsf T)=0$ for the squared normal frequencies $\lambda=\omega^2$ of small oscillations, where $\mathsf T$ and $\mathsf V$ are the mass and stiffness matrices at the equilibrium.8.4
stable equilibriumAn equilibrium $q^0$ such that motions starting close enough to $q^0$ with small enough velocity stay as close as we please to $q^0$, with small velocity, for all later time. For a natural system a strict minimum of the potential energy is stable (Lagrange–Dirichlet).1.3 8.1 8.2 8.4
virtual displacementAn infinitesimal change of configuration $\delta\mathbf r_i=\sum_j(\partial\mathbf r_i/\partial q_j)\,\delta q_j$ compatible with the constraints at a frozen instant ($\delta t=0$).8.1
zero modeA normal mode with $\omega=0$, i.e. a direction in $\ker\mathsf V$ along which the system can move with no restoring force — for example a rigid translation or rotation of a free molecule. Its normal coordinate moves as $\alpha+\beta t$.8.4