Asymptotic regularization 2026-10-05
Asymptotic regularization stops the gradient flow at , starting from zero. With the singular system of a compact operator convention , its spectral filter givesThe scalar coefficient solves with zero initial value. Since for , the operator norm is at most . On the domain of the Moore–Penrose inverse of an operator, dominated convergence theorem of the squared spectral coefficients proves . The noise-bias decomposition for linear regularization then gives noisy-data convergence when .
Compact gradient-flow trapping criterion 2026-10-07
A smooth gradient flow confined to a compact invariant sublevel set has finite total integral of when is bounded below. Compactness supplies uniform continuity of that squared norm along the trajectory, so it tends to zero. Every accumulation point is a critical point. If the trapped component contains exactly one critical point, the trajectory converges to it. This supplies a direct convergence proof in place of an inference solely from an energy-contour picture.
Diffusion from a gradient energy Created 2026-10-05 Updated 2026-10-06
The formal gradient flow of this energy is , with a Neumann boundary condition for no flux. The radial Hessian eigenvalues of the gradient integrand are and . Their signs distinguish convex forward diffusion from formal backward diffusion. The heat equation uses , whereas total variation flow uses with a subgradient interpretation at zero.
Exponential convergence of a Morse gradient flow Created 2026-10-06 Updated 2026-10-14
The Morse lemma gives coordinates . Choose a Riemannian metric Euclidean in these coordinates. A downward gradient flow converging to has eventually, while . Thus its Riemannian distance from decays exponentially. A partition of unity makes this a global metric choice; changing the metric may change the trajectories.
Assume finitely many nonempty routes, positive weights and adjustment rates, and continuous nonnegative nondecreasing resource prices. Each route must meet a resource whose price is positive somewhere. Then the primal congestion potential has compact superlevel sets inside the positive orthant and a unique maximum. Along gradient flow with diagonal mobility, . The substitution makes the dynamics ordinary gradient flow ascent for . This is strictly concave: the logarithmic terms are strictly concave, and each integrated price is convex and nondecreasing, composed with a convex quadratic load. Thus its continuous gradient is monotone decreasing, which proves uniqueness of trajectories even without differentiability of the prices. Compact trapping and the dissipation identity force every limit point to be the unique maximizer. If every price on some route vanishes identically, its rate instead grows linearly and there is no equilibrium.
Gradient flow Created 2026-09-24 Updated 2026-11-05
For a smooth real function on Euclidean space, its downward gradient flow is the ordinary differential equation . The chain rule givesThus is a Lyapunov function, and a nonstationary trajectory cannot return to its starting point. For , the solution is . The explicit Euler method gives , namely gradient descent. The continuous decrease does not guarantee decrease for every discrete step size; even this quadratic example requires for strict contraction.
Gradient-flow dissipation identity 2026-10-06
For the downward gradient flow , the chain rule gives . Integrating converts a bound on the decrease of into a bound on elapsed time wherever its Riemannian gradient is bounded away from zero.
For smooth -periodic , the equation is the negative gradient flow of . Periodic integration by parts gives . Its unique equilibrium has Fourier coefficients , and every transient coefficient decays as . Thus the flow converges to the unique minimizer of the energy.
For the conserved-mean convection amplitude equations before amplitude reduction, the functional generates a gradient flow with increasing . Periodic integration by parts identifies its variational gradient with the evolution equation. Completing the square gives . Hence along a globally smooth solution. Under compactness of the orbit, its limiting states are stationary; a nonstationary periodic orbit is impossible.
Lyapunov functional Created 2026-10-14 Updated 2026-10-31
A Lyapunov functional assigns a scalar to the entire state of an evolution equation and is monotone along its trajectories. For a gradient flow , the identity follows from the variational chain rule, with boundary terms removed by periodicity or decay. A spatial density is not itself a Lyapunov functional unless all remaining coordinates are integrated or averaged.
Morse handle-attachment theorem 2026-10-14
Crossing one nondegenerate critical point of Morse index in a compact sublevel interval attaches a handle . Between critical levels a suitably normalized gradient flow identifies the sublevels. The Morse lemma supplies the local quadratic model for the attachment. Retracting each handle to its core disk gives a CW complex model with one -cell per critical point. Closed manifolds have finite such decompositions; a noncompact manifold requires a proper bounded-below exhaustion for the corresponding statement.
Morse-Smale gradient flow 2026-10-14
For a Morse function, a downward gradient flow is Morse-Smale if every unstable manifold meets every stable manifold transversely. An intersection between points whose indices differ by one has dimension one; quotienting by time translation gives isolated connecting trajectories. Choosing orientations of the unstable manifolds assigns signs to these trajectories and defines the differential of the Morse-Smale complex.
Neural tangent kernel Created 2026-10-25 Updated 2026-10-31
For a differentiable neural network and initialization law , the expected neural tangent kernel is , assuming finite diagonal expectations. At a fixed parameter value the corresponding inner product defines the empirical neural tangent kernel. Both are positive-semidefinite kernels, since a finite quadratic form is respectively an expected squared norm or a squared norm. In suitable infinite-width scalings, the empirical kernel matrix converges to a deterministic limit and remains constant during training. Squared-error gradient flow then has the same training-prediction dynamics as kernel regression with this limit, provided the initial predictions match.
Under the stated infinite-width scaling and freezing assumptions, the deterministic limiting kernel matrix has entriesFor a sequence of finite-width networks, this expectation is interpreted in the stated width limit. The resulting neural tangent kernel is a positive-semidefinite kernel, since every finite coefficient vector gives .
The prediction gradient flow is linear:The matrix exponential solves the constant-coefficient system because the residual satisfies . Each positive-eigenvalue component decays exponentially; a component in the null space of stays at its initial residual. Thus for every initial prediction if is a positive-definite matrix. The limiting kernel matrix being deterministic does not require to be deterministic.
Let be the Jacobian matrix with row . Differentiating the squared-error loss function givesThe parameter gradient flow is consequently . The chain rule then giveswhere, explicitly,It is a Gram matrix and therefore a positive semidefinite matrix: . These derivatives are understood wherever the network is differentiable; for piecewise differentiable activations the same calculation applies along differentiable portions of the trajectory.
Evaluating the negative functional gradient at yieldsThus the evaluation vector followsThis is exactly the limiting prediction dynamics in part (ii). If , subtracting the two systems gives with zero initial value, so for every .
An infinitely wide neural network in this regime follows the same training-prediction dynamics as kernel regression with its neural tangent kernel and matching initial predictions. For this squared-error loss function, the relevant algorithm is gradient flow in the Reproducing kernel Hilbert space. Matching predictions at arbitrary additional inputs also requires the corresponding cross-input neural tangent kernel to freeze and the initial predictions there to match. A finite-width neural network can change its feature map during training, so the equivalence depends on the assumptions rather than holding for every architecture and parameterization.
Past exam of the mathematics course of the University of Cambridge 2001 iii Paper 54 2 b Solution Created 2026-10-03 Updated 2026-10-14
Let denote the right-hand side of the temperature partial differential equation. For a periodic variation , integration by parts givesAlong the evolution, , so this is a gradient flow with increasing functional:For , completing the square givesThus the increasing-gradient functional for poorly conducting convection converges to a finite limit along any globally smooth solution, andIn particular, there are arbitrarily late times at which is arbitrarily small. A nonstationary time-periodic solution is impossible: it would have a strictly positive increase of over a temporal period. More generally, if the orbit is precompact in a topology strong enough for the equation, its limiting states are stationary. With the conserved mean fixed, the displayed square completion bounds , and periodic elliptic estimates give an bound; the usual smoothing and compactness for a globally regular parabolic solution support this limiting-state conclusion. The inequalities alone do not prove convergence to one uniquely specified equilibrium, and stationary families related by translation symmetry can remain.
Now take a smooth, nonconstant periodic roll . Multiplying its stationary partial differential equation by and using periodic integration by parts yieldsThe primes here are essential: the first and fourth terms involve , not . Write , , , , and . The orthogonal-roll perturbation hasThe cross term in the squared Laplacian averages to zero since . Independence of the cell integrals then gives the exact polynomial differenceIts leading coefficient is half the variance of over the period. That variance is strictly positive: a nonconstant smooth periodic function has a derivative which vanishes somewhere and is nonzero somewhere, so is not constant. Hence arbitrarily small transverse perturbations increase above the roll value.
One can turn this energy comparison into linear instability. The linearized evolution at the roll is the self-adjoint operatorChoose the fixed-mean perturbation . Since , the Rayleigh quotient is positive. The periodic self-adjoint fourth-order operator therefore has a positive eigenvalue. This proves the transverse energy instability of nonconstant temperature rolls: every nonconstant roll is unstable wherever it exists.
The existence qualification matters. Constants are also stationary and are stable on the fixed-mean subspace for . Moreover, the periodic Poincaré inequality gives , so the roll identity implies for a nonconstant -periodic roll. Such rolls require on this specified cell; the conclusion for all positive is conditional on a nonconstant roll being available.
Past exam of the mathematics course of the University of Cambridge 2003 iii Paper 60 3 i Solution Created 2026-10-03 Updated 2026-10-14
Because , a periodic primitive exists. Fix its additive constant by . Integrating the conserved-field equation once in space gives : the possible integration constant vanishes on taking the spatial mean. Treat the spatial average as the inner product defining functional derivatives. Integration by parts, with no boundary terms under periodic boundary conditions, givesThus the conserved-field real amplitude equation is a gradient flow ascending , or descending , with positive diagonal mobilities. Along a classical solution,For the mixed term put and . The Cauchy-Schwarz inequality and a completed square giveConsequently, for ,The negative leading quartic coefficient proves the upper bound. The stronger pointwise square completionis also useful: a superlevel set bounds the amplitude in , its spatial derivative in , and in , because a quartic polynomial with negative leading coefficient dominates its lower-order terms.
The dynamics dissipates toward the steady-state set rather than sustaining nonconstant periodic motion. Indeed increases to a finite limit, and the time integral of is finite. For global regular trajectories with the usual parabolic precompactness, every limit point is stationary, by the LaSalle invariance principle. A nonconstant periodic or recurrent orbit would force a strictly increasing to return to its original value and is impossible. Boundedness of a monotone functional alone does not prove convergence to one particular equilibrium; isolation of the limiting equilibrium or an additional convergence theorem supplies that stronger claim.
Past exam of the mathematics course of the University of Cambridge 2004 iii Paper 74 4 Solution Created 2026-10-03 Updated 2026-10-14
Let . Incompressibility makes the transverse velocity . In the axial Navier-Stokes equation, transverse viscous stress scales as , longitudinal viscous stress as , and convective inertia as . Thus lubrication theory requiresUnsteady inertia is also negligible if ; this coincides with the preceding condition when . The transverse pressure variation is smaller than this pressure scale by , and viscous normal stress at the free surface is similarly smaller. Consequently pressure is uniform across the film to leading order, while axial viscous shear balances its axial gradient. These estimates justify the lubrication approximation with small surface slopes and a slowly varying rigid wall.
Write the physical cylindrical-film thickness as , so that , and use for dimensional coordinates. For a surface of radius , its outward mean curvature, with the convention giving cylindrical curvature , isThe Young–Laplace equation gives . With local normal coordinate measured from the rigid cylinder, the axial velocity solves , with no-slip boundary condition at and stress-free boundary condition at . ThereforeThe factors of circumference in axisymmetric mass conservation cancel at leading thin-film order, giving . IntroduceThis produces the capillary instability of a thin film on a rigid cylinder governed byThe curvature term destabilizes long waves; the axial-curvature term smooths short waves.
Linearize about using with small . The linear equation is , soThe growth rate is positive for , zero at , and negative for . Its derivative is , givingThe requested sketch is even in :
For the Rayleigh–Plateau instability of an entire liquid cylinder in the purely viscous limit with negligible exterior viscosity, the fastest mode tends to infinite wavelength. The rigid-wall coating instead selects a finite wavelength. The difference comes from the flow resistance: a coating must shear across its thickness, so its long-wave growth vanishes as ; a free liquid cylinder can undergo nearly plug-like axial extension. For example, the long-wave free-cylinder balances and give , which does not vanish as . This comparison is for the creeping-flow, negligible-exterior-viscosity limit; finite inertia or exterior viscosity changes the wavelength selection. The full free-cylinder dispersion relation is discussed in Plateau-Rayleigh Instability of a Cylinder of Viscous Liquid.
The collar is a quasisteady sliding collar on a coated cylinder. It must slowly grow as well as translate: an exact travelling wave would obey , whereas its stationary uniform far films have and would require . Thus the stipulated unequal films require the stated slow evolution of the collar amplitude. In its core, and the volume flux is small, so the leading pressure gradient vanishes: . Integration gives a constant plus sine and cosine terms. Choose the collar centre at its maximum to remove the sine term, and parameterize its edge height as . ThenThe core is a local capillary equilibrium. In the moving frame, swept-up volume minus deposited volume gives the growth rate . Its leading excess volume is , and for constant ,Thus the printed rate is the rate of collar volume, or an order-of-magnitude amplitude rate; interpreted as an exact equation for , it lacks . This slow change is smaller than the translational core terms by the small far-film thickness and does not alter the leading local matching below.
Near either edge , let denote the adjacent film thickness, at the rear and at the front. Across the short transition region the collar's slow growth is higher order, so local mass conservation integrates to . The axial third derivative dominates the first derivative in the capillary flux, givingFor collar-to-film capillary matching, chooseThe two sides become and respectively, and henceThe resulting transition length is small compared with the core length: the matching relations below imply , so for an order-one core amplitude.
To count admissible branches, put and linearize. The characteristic equation is , with rootsAt the rear, decay as permits only . For the nontrivial branch growing into a positive collar, ; translation rescales , leaving one orbit. This branch has positive . It exists for all forward , since for , and once , grows at least quadratically. Then , so successive integrations give a quadratic asymptote with a finite constant term; translation removes the linear term. The constant solution is not a rear-to-collar match. On the downward branch, gives while and are initially negative; the slope remains negative and reaches zero, so it cannot match a positive collar. Uniqueness here refers to the positive, nontrivial rear matching branch, modulo translations.
At the front, decay as permits the two real amplitudes inTranslation removes one degree of freedom, leaving one parameter . These oscillatory branches account for the small capillary ripples ahead of the collar. This local stable manifold count gives the family of front branches; admissible global branches must also remain positive and reach a parabolic core asymptote. It does not, by itself, prove the global uniqueness of the matching parameter supplied in the question.
Write the rear and front core asymptotes, after translation removes their linear terms, as and . Expanding the outer collar near either edge giveswhereas the inner expansion givesMatching the constant and quadratic coefficients therefore yieldsEliminate to obtain . Using the specified selected root and writing , ,The physical branch has , as required by . No numerical values of these universal matching coefficients are needed for the requested relations.
Finally, the energy source is decreasing capillary surface energy, not Newtonian gravity. This can be checked through the volume-constrained capillary energy of a cylindrical coating. The exact liquid volume per axial length is , and the free-surface area is . Expanding the latter and eliminating with fixed exact volume gives the nonconstant quadratic partApart from its positive dimensional factor, this is the surface energy . Its variational derivative is , and the thin-film equation is a conserved gradient flow. With periodic or zero-energy-flux boundaries, integration by parts givesFor an isolated collar between unequal infinite films, use the corresponding local energy balance, including far-film fluxes and subtracting background energies; the unsubtracted total energy is infinite. Collecting film into the growing collar lowers surface energy, which supplies the viscous dissipation sustaining its slow motion.
Past exam of the mathematics course of the University of Cambridge 2005 iii Paper 25 4 Solution Created 2026-10-03 Updated 2026-10-14
Choose real constants and define the smooth homogeneous quotientIt is unchanged by nonzero rescaling, so it is well defined on Real projective space. On the unit sphere its differential vanishes on the tangent space precisely when is proportional to , where . The entries are distinct, so the only critical projective lines are . In the affine chart , write for . ThenIts Hessian at zero is diagonal with entries . Exactly are negative, and none vanish. Thus this is a Morse function with one critical point of index for every .
Choose the quotient of the round metric on the unit sphere. The downward gradient flow has sphere equation and projective solutionNormalization to the unit sphere supplies the common positive factor. As , the smallest index with nonzero coordinate dominates; as , the largest dominates. ConsequentlyThese are its unstable manifolds and stable manifolds. If , the tangent directions in allow coordinates , and those in allow , modulo the common projective line. Together they span the full tangent space at every intersection. If the intersection is empty. Thus the flow is a Morse-Smale gradient flow, with no generic perturbation needed.
Orient each unstable manifold using its affine coordinates. The Morse-Smale complex has ; its differential counts oriented unparametrized connecting trajectories of index difference one. A trajectory from to has only coordinates nonzero, and there are exactly two, distinguished by the sign of . Each sign branch is a single orbit modulo time translation.
To calculate their signs rather than merely count two, use the characteristic closed disk of the unstable cell: the closed upper hemisphere of with , followed by projection onto . Its interior is , and its boundary map is the antipodal double cover . After collapsing , its degree on the -cell is the sum of the degrees of the two sheets. Comparing the sheets by the antipodal map on gives relative orientation : it is induced by on , whose determinant is . Thus the two trajectory signs add to , up to the choice of generator orientations. The projective quadratic Morse complex is thereforeIn particular , as required for the connected one-dimensional skeleton. The signed trajectory count agrees with this attaching-map incidence calculation, so it computes integral rather than only mod-two homology.
For , taking kernels modulo images givesFor example, an odd interior-degree generator has zero outgoing differential but twice itself is the image from the next degree; an even positive-degree differential is injective. At the top there is no incoming differential, leaving a copy of exactly when is odd. For the space is one point and only remains. This recovers the cellular homology of real projective space by explicitly constructing its Morse-Smale complex and signs.
