18.369, Spring 2016

Mathematical Methods in Nanophotonics

Prof. Steven G. Johnson, Dept. of Mathematics


This is the home page for the 18.369 course at MIT in Spring 2016, where the syllabus, lecture materials, problem sets, and other miscellanea are posted.

You can also download the course announcement flyer, and visit this photonic-crystal tutorial page to find materials for past lectures by SGJ on related subjects. This course was previously offered as 18.325 in Fall 2005 (also on OpenCourseWare) and as 18.369 in Spring 2007, Spring 2008, Spring 2009, Spring 2010, Spring 2012, and Spring 2014.

Tired of doing electromagnetism like it's 1865?

Find out what solid-state physics has brought to 8.02 in the last 20 years, in this new course surveying the physics and mathematics of nanophotonics—electromagnetic waves in media structured on the scale of the wavelength.

In this regime, which is the basis for everything from iridescent butterfly wings to distributed-feedback lasers and integrated optical devices to the next generation of optical fibers, the 140–year-old analytical techniques you learned in 8.02 aren't very useful. Instead, we will cover computational methods combined with high-level algebraic techniques borrowed from solid-state quantum mechanics: linear algebra and eigensystems, group theory, Bloch's theorem and conservation laws, perturbation methods, and coupled-mode theories, to understand surprising optical phenomena from band gaps to slow light to nonlinear filters.

For beginning graduate students and advanced undergraduates.


Lectures: MWF 2–3pm (2-143). Office Hours: Thurs. 4–5pm (2-345). TA/grader: Siddharth Venkatesh.

Probable topics: Methods: linear algebra & eigensystems for Maxwell's equations, symmetry groups and representation theory, Bloch's theorem, numerical eigensolver methods, time and frequency-domain computation, perturbation theory, coupled-mode theories, waveguide theory, adiabatic transitions. Optical phenomena: photonic crystals & band gaps, anomalous diffraction, mechanisms for optical confinement, optical fibers (new & old), nonlinearities, integrated optical devices.

Grading: 33% problem sets (weekly/biweekly). 33% mid-term exam (April 8). 34% final project (proposal due April 15, project due May 11).

Books: Photonic Crystals: Molding the Flow of Light (Second Edition) (readable online). (This book is at an undergraduate level, and 18.369 is somewhat more advanced, but the book should provide a useful foundation.)

Useful (but not required) books (available in the library): Group Theory and Its Applications in Physics by Inui et al. (readable online via MIT), and Group Theory and Quantum Mechanics by Michael Tinkham.

Final projects: A typical project will be to find some interesting nanophotonic structure/phenomenon in the literature (chapter 10 of the book may be a helpful guide to some possibilities), reproduce it (usually in 2d only, so that the simulations are quick), using (e.g.) the numerical software (Meep and/or MPB) introduced in the course/psets, and extend/analyze it in some further way (try some other variation on the geometry, etc.). Then write up the results in a 7-15 page report (in the format of a journal article, ideally Phys. Rev. A style, including figures, a comprehensive review of related work, etcetera)—reports should be written for a target audience of your classmates in 18.369, and should explain what you are doing at that level. Projects should not be a rehash of work you've already done in previous terms for your research (but may be some extension/digression thereof).

Prerequisites: 18.305 or permission of instructor. (Basically, some experience with partial differential equations and linear algebra. e.g. 8.05, 8.07, 6.013, 3.21, 2.062.) This is a graduate-level course aimed at beginning graduate students and suitably advanced undergraduates.

Supplementary lecture notes: Notes on the algebraic structure of wave equations and Notes on Perfectly Matched Layers (PMLs), and several other PDF files that will be made available as the term progresses.

Previous mid-terms: fall 2005 and solutions; spring 2007 and solutions; spring 2008; spring 2009 and solutions; spring 2010 and solutions; spring 2012 and solutions; spring 2014 and solutions.

Lecture Summaries and Handouts

Lecture 1: 3 Feb 2016

Handouts: syllabus (this web page), introductory slides, collaboration policy, pset 1 (due Wed Feb 10).

Motivation and introduction: this class is about electromagnetism where the wavelength is neither very large (quasi-static) nor very small (ray optics), and the analytical and computational methods we can use to understand phenomena in materials that are structured on the wavelength scale. In that situation, there are very few cases that can be solved analytically, but lots of interesting phenomena that we can derive from the structure of the equations.

We start by setting up the source-free Maxwell equations as a linear eigenproblem, which will allow us to bring all of the machinery of linear algebra and (eventually) group theory to bear on this problem without having to solve the PDE explicitly (which is usually impossible to do analytically).

Notational introductions: Hilbert spaces (vector space + inner product), notation for inner products and states (magnetic fields etc.). Defined the adjoint (denoted †) of linear operators. Proved that self-adjoint/Hermitian operators (Â=Â) have real eigenvalues, and orthogonal eigenvectors (for distinct eigenvalues; for other eigenvalues you can orthogonalize by Gram-Schmidt).

Further reading: See chapter 2 of the Photonic Crystals textbook for Maxwell's equations as an eigenproblem etc. For a more sophisticated treatment of Hilbert spaces, adjoints, and other topics in functional analysis, a good text is Basic Classes of Linear Operators by Gohberg et al. The basics of electromagnetism in macroscopic media (ε and μ) are covered in any non-freshman electromagnetism textbook, e.g. Classical Electrodynamics by Jackson or Introduction to Electromagnetism by Griffiths.

Lecture 2: 5 Feb 2016

Showed that the Maxwell eigen-operator ∇×ε-1∇× is Hermitian for real ε (by showing that ∇× is Hermitian). The Maxwell operator is also positive semidefinite, and it follows that the eigenfrequencies are real.

Discussed scale invariance of Maxwell's equations, and the fact that if we scale up the whole system by a factor of s then the solutions are the same, just with wavelengths scaled up by s (frequencies scaled by 1/s). This comes from the multiplicative nature of the Maxwell operator, and is very different for additive operators like the Schrodinger operator in quantum mechanics (see end of chapter 2 in the text).

Simple one-dimensional example of fields in metallic cavity, showed that consequences match predictions from linear algebra.

Further reading: See chapter 2 of the textbook.

Lecture 3: 10 Feb

Handouts: representation theory summary

Discussed consequences of symmetry, and in particular showed that mirror symmetry implies even/odd solutions. Discussed subtleties of mirror symmetries for electromagnetism: although the E and H fields seem to have opposite symmetry, they don't, because H is a pseudovector. Defined general rotation operators ÔR for vector and pseudovector fields.

Gave a simple 2d example of fields in a 2d metal box, and showed that the symmetries are more complicated, and may include degeneracies. In order to understand this, we need to understand the relationship of different symmetry operations to one another — this relationship is expressed more precisely by the group of symmetry operators. Defined groups, and we will get to the consequences of group theory next time.

Further reading: Chapter 3 of the photonic-crystals text. See the Inui textbook, or many similar sources, on group theory; the most helpful in this context are the many "group theory in physics" books.

Lecture 4: 10 Feb

Handouts: pset 2 (due Friday, Feb 26), pset 1 solutions

Defined groups, and group representations, irreducibility, and partner functions, conjugacy classes and most of the other things on the handout, with some examples (the square symmetry group and the mirror symmetry group). (Covered everything on the handout except Great Orthogonality Theorem, character tables, projection operators, and product representations.)

We will show next that eigenfunctions are partner functions of representations of the symmetry group. For example, even and odd functions in a mirror-symmetric system correspond precisely to the two irreducible representations of the {E,σ} group.

Proved that all eigenfunctions can be chosen to transform as partner functions of an irreducible representation of the symmetry group (also called a "basis" of the representation), with the dimension of the representation given by the degree of degeneracy of the eigenvalue. Proved that all representations derived from a given eigenvalue are equivalent. Noted that orthonormal eigenfunctions give a unitary representation (outlined proof but did not work it through).

If the representation is irreducible, then the degeneracy comes from the symmetry of the system. If the representation is reducible, then we call it an accidental degeneracy (not coming from symmetry). Accidental degeneracies rarely happen by accident—usually the degeneracy has somehow been forced—so generically we only expect degeneracies if there are >1 dimensional irreps.

Further reading: Chapter 3 of the text, but this doesn't get into representation theory. See e.g. Group Theory and Its Applications in Physics by Inui et al. (especially sections 4.1, 6.1, and 6.2) or Group Theory and Quantum Mechanics by Michael Tinkham (especially sections 3-1 and 3-6), or any book with a similar title.

Lecture 5: 12 Feb

Handout: notes on decomposition of functions into partner functions

Introduce the character table of a group, the table of the traces ("characters") of the irreducible representations (which are constant with a given conjugacy class and representation).

Build the simple character table for the {E,σ} mirror-symmetry group, reprising the previous result that in mirror-symmetric systems we expect even/odd eigenfunctions, and we don't expect (non-accidental) degeneracies (unless there are additional symmetries).

Using the rules from the representation theory handout, we build up the character table for the symmetry group of the square (called C4v). Then, look at the eigenfunction solutions that we previously had for this case, and show how we could classify them into the various irreducible representations.

Looked at the projection operator (see handout) in more detail and gave some graphical examples of how we can use it to decompose a function into partner functions.

Used the projection operator to classify the modes of the square cavity, and in particular found that some of the modes are accidental degeneracies. In this way, we are able to find representatives of all five irreps. Conversely, by looking at the irreps, we can guess some of the types of eigenfunctions that should appear, inferring the sign pattern from the character table.

Showed how we can apply the projection operator to "random" functions to find partners of different irreps, even without an eigenproblem, and to "sketch" the qualitative features that we expect to find in the eigenfunctions. And, once we have partner functions, we can obtain representation matrices for each irrep (useful for 2+ dimensional irreps). As an example, looked at ψ(x,y)=1, x, and x2; found in particular that the 2d irrep transforms like {x,y}, i.e. the ordinary 2d rotation matrices.

Further reading: Character tables for all of the common symmetry groups are tabulated in both textbooks and online, e.g. see this page on the C4v group. See Inui section 6.6 on projection operators.

Lecture 6: 16 Feb

Handout: notes on time evolution and conservation laws

Proved in general that the irreducible representation is conserved over time in a linear system, by showing that the projection operator commutes with the time-evolution operator.

Defined the time-evolution operator explicitly via an exponentiated operator on the 6-component vector-field (E, H). Showed that the time-evolution operator is unitary in an appropriate inner product, and that this leads to conservation of energy.

Derived Poynting's theorem in order to define electromagnetic energy and flux in general, and showed that we got the same quantity as we did from unitarity. For time-harmonic fields, showed that |E|2/2 and |H|2/2 and Re[E*×H]/2 are time averages of the corresponding real oscillating fields Re(E) and Re(H). Showed that the time-average energies in the E and H fields are the same.

Talked a little about a more general formulation of the time-dependent problem, including arbitrary dispersion, and derived that the susceptibility χ needs to be a "passive convolution operator" for energy to be non-increasing. It turns out that this has lot of interesting consequences, which we mostly won't have time to get into.

Further reading: See my Notes on the algebraic structure of wave equations for a general discussion of many wave equations, showing that they share the common form dψ/dt D ψ where D is anti-Hermitian. For Poynting's theorem, see any graduate-level book on electromagnetism, e.g. Jackson's Classical Electrodynamics. The result is summarized in chapter 2 of the textbook. Beware that matters are more complicated for dispersive media (ones in which ε and μ depend on ω), as discussed briefly in Jackson. A much more complete review of passive dispersive media, including the consequences of passivity for causality etcetera, can be found in our 2014 paper Speed-of-light limitations in passive linear media: see in particular the discussion of passivity in section II.B and of "dynamical" energy density in section V.A.

Lecture 7: 17 Feb

A bit more discussion of dispersive media: dynamical energy density vs. dispersive energy density vs. non-dispersive energy density (section 3.3 in the notes).

Translational symmetry: Showed that for continuous translational symmetry, the representations are exponential functions exp(ikx) for some number k (real for unitary representations; in weird cases, k may be a nondiagonalizable matrix with imaginary eigenvalues, but these solutions are not needed in periodic or translationally invariant systems). Concluded that the solutions of Maxwell's equations in empty space are planewaves, and discussed the corresponding dispersion relation.

Explained how conservation of the exp(-ikx) representation, which gives conservation of k, means that planewaves are produced by a line-current source Jz(x,y)=δ(x) e-i(ky-ωt) in 2d, assuming outgoing (radiation) boundary conditions.

Explained how conservation of the exp(-ikx) representation, which gives conservation of k, leads immediately to Snell's law at a flat interface.

Further reading: Textbook, chapter 3 on continuous translational symmetry.

Lecture 8: 19 Feb

Introduced dielectric waveguides, via the simple 2d example of a high-ε region surrounded by a low-ε region, invariant in the x direction. Explained that the solutions far from the waveguide lead to a continuous region, the light cone, and argued (proof to come later) that the higher-ε region pulls down localized guided modes below the light cone. Since they are localized, they form discrete bands in order to stay orthogonal.

Introduced the "reduced" eigenproblem Θk=e-ikxΘeikx for the modes of a particular wavevector k. This is also Hermitian, its solutions ω(k) yield the dispersion relation (or band structure) of the problem.

Explained how mirror symmetry in z means that z-invariant solutions in "2d" structures ε(x,y) can be segregated into two polarizations: even Hz-polarized (what the book calls "TE") and odd Ez-polarized (what the book calls "TM"). (Note that the literature is split on the terminology here: many authors call the former TM and the latter TE.)

Introduced the variational theorem (or min–max theorem), which arises for any Hermitian eigenproblem. Proved the variational theorem (with the simplifying assumption of a basis of eigenfunctions), and mentioned the derivation (in chapter 2 of the book) that all extrema of the Rayleigh quotient are eigenvalues. We will use this theorem to derive general conditions under which guided modes are guaranteed to arise in dielectric waveguides.

Further reading: See the book, chapter 3, on index guiding and the variational principle. (See e.g. Jackson's Classical Electrodynamics for a more traditional viewpoint on dielectric waveguides, focused on the two cases that can be solved analytically, and Marcuse's Theory of Dielectric Optical Waveguides for an expanded version of this. See e.g. Ramaswami and Sivarajan, Optical Networks for a nice practical overview of dielectric waveguiding in modern telecommunications.) See these notes on localization in a different scalar-wave equation via the same variational method. Bamberget and Bonnet (1990) is a classic paper on the theory of dielectric waveguiding. Lee (2008) is an extension of the variational proof to Maxwell's equations in much more complex periodic waveguides and photonic-crystal fibers.

Lecture 9: 22 Feb

Used the variational theorem to prove the existence of index-guided modes (in two dimensions, for the TE polarization), for any translation-invariant structure where ε is increased "on average" in a localized region, for an appropriate definition of "on average."

Considered related theorems in quantum mechanics: an arbitrary attractive potential will always localize a bound state in 1d or 2d, but not in 3d, and sketched a simple dimensional argument in 1d and 3d (but not 2d, which is a difficult borderline case). Discussed the related theorem for 3d waveguides (2d localization), and the case of substrates where the theorem does not apply and the fundamental modes has a low-ω cutoff.

Discrete translational symmetry:

Started by considering a periodic "waveguide" in two dimensions: a sequence of dielectric rods in air. By analogy with ray-optics and total-internal reflection, it seems that this could not support guided modes. However, it does (chapter 7 of the book), and to understand this we need to understand discrete translational symmetry.

Showed that the representations of the discrete translation group are again exponentials, and thereby proved Bloch's theorem: the eigenfunctions can be chosen in the form of a planewave multipled by a periodic function.

Lecture 10: 24 Feb

Reviewed result from last lecture representations of the discrete translation group are again exponentials, and thereby proved Bloch's theorem: the eigenfunctions can be chosen in the form of a planewave multipled by a periodic function. Defined the primitive lattice vectors. As a corollary, the Bloch wavevector k is conserved, and explained how this relates to a famous mystery from the 19th century: electrons in a pure conductor act almost like a dilute gas, because they scatter only from impurities/imperfections that break the periodicity.

Qualitative description of the resulting band diagrams in 1d-periodic systems.

Derived the periodicity of the Bloch wavevector k in one dimension. Adding 2π/a does not change the irrep, and is only a relabeling of the eigensolution. This means that we only need to look at the "unit cell" in k-space to get the band diagram. Discussed the concept of the (first) Brillouin zone in 1d, although a more general definition will have to wait until we get to 2d periodicity (chapter 5). Defined the reciprocal lattice vectors, and gave examples for 1d periodicity and for a 2d square lattice.

Considered interaction of rotational symmetries with k: showed that rotations R transform a solution at k into solution at Rk, and hence ω(k)=ω(Rk).

By conjugating the eigenequation, for real ε, showed that ω(k)=ω(−k) in general, even for structures without mirror symmetry. Connected this to time-reversal symmetry: the conjugated mode corresponds to running time backwards, which still solves the same Maxwell's equation. One way to break time-reversal symmetry is by introducing absorption loss (complex ε, which time-reverses into gain). Alternatively, briefly mentioned magneto-optic materials (complex-Hermitian ε, neglecting absorption) and why a static magnetic field can (locally) break time-reversal symmetry, and of use for Faraday isolators.

Further reading: Chapter 3 of the textbook. See appendix B for more on the reciprocal lattice, a topic we will return to later.

Lecture 11: 26 Feb

Handouts: pset 2 solutions, pset 3 (due March 11) (for problems 3 and 4, you will need 2dwaveguide.ctl and bandgap1d.ctl

As an application of conservation of k (up to addition of reciprocal lattice vectors) in periodic systems, discussed reflection (specular and diffractive) from a periodic surface, and minimum-frequency/maximum-wavelength cutoffs for various diffracted orders.

Began new topic: photonic band gaps in one dimension. Sketched the form of the dispersion relation (band structure) and explained several qualitative features we can predict without solving:

Next time, we will use perturbation theory to derive the magnitude of the gap and other features by starting with a homogeneous medium and then adding a little bit of periodic ε contrast.

Further reading: Chapter 4 of the textbook.

Lecture 12: 29 Feb

A quantitative estimate of the size of the band gap in 1d, via perturbation theory. In particular, derived first-order perturbation theory for the eigenvalue of any Hermitian operator with some small change, by expanding the eigenvalue and eigenfunction as power series in the change and solving order-by-order. We then write down this perturbative expression for the Maxwell operator, and see that the fractional change in frequency is just the fractional change in index multiplied by the fraction of electric-field energy in the changed material.

Using first-order perturbation theory, computed the size of the band gap for a 1d periodic structure to first order in Δε. Defined the "size" of the gap in a dimensionless way as a fraction of mid-gap.

Discussed how perturbation theory can be used to derive the effect of absorption losses (to lowest order): adding a small imaginary part to ε yields a corresponding imaginary part in ω, giving loss or gain depending on the sign.

Further reading: For the same derivation of perturbation theory, see "time-independent perturbation theory" in any quantum-mechanics text, e.g. Cohen-Tannoudji. See also the section on small perturbations in chapter 2 of the book. See chapter 4 of the book on the origin of the 1d gap, and on the special formulas for quarter-wave stacks in 1d (discussed in more detail in Yeh's Optical Waves in Layered Media).

Lecture 13: 2 Mar

Handouts: MPB demo (see also the MPB home page) and example files: 2dwaveguide.ctl and 2dwaveguide-periodic.ctl; IJulia notebook

Gave demo of MPB eigensolver software for 2d dielectric waveguide (available on Athena, e.g. in the clusters or via ssh to athena.dialup.mit.edu).

In the handout, I gave the plotting code in Matlab. Because the output of MPB is either comma-delimited text (ω vs. k) or HDF5 (field data), essentially any modern plotting software should be able to handle it. In class, I used Julia, a Matlab-like language with C-like performance in an IJulia notebook with the PyPlot plotting library (which calls the Python matplotlib under the hood). You can use whatever you want.

Further reading: The MPB web page, and Appendix D of the textbook (on numerical methods).

Lecture 14: 4 Mar

Discussed reflection of light from a semi-infinite 1d crystal, at a frequency in the gap. We have no propagating solutions in the crystal, so by conservation of energy we must have 100% reflection. However, showed by analytical continuation of the band edge that we expect exponentially decaying "evanescent" solutions in the crystal, with a complex wavevector k ≈ sqrt(Δω/α) + π/a, where Δω is how far we are into the gap and α is the band-edge curvature.

Discussion of localization of modes by defects in 1d crystals, discussing how a positive Δε "pulls down" a mode from the upper edge of the gap, and a negative Δε "pushes up" a mode from the lower edge. In general, we expect a discrete set of frequencies for such "bound states", and the number of bound states in a given gap should grow asymptotically proportional to the volume of the defect (the diameter in 1d), but for an arbitrarily weak defect we still generically expect at least one bound state.

Computationally, e.g. in MPB, we often compute localized cavity modes by imposing periodic boundary conditions in a supercell consisting of many unit cells, plus a defect. Explained how, in the absence of a defect, such a supercell leads to the original band structure "folded" into the new Brillouin zone. A defect then pulls one of these foldings into the gap, with a cosine-like dispersion relation that becomes flat exponentially fast as the supercell size is increased.

Defects surrounded by a finite crystal are more tricky. Intuitively, they will "leak out" slowly from the defect via their evanescent tails. Mathematically, the topic of such "leaky modes" is quite tricky. Explained why this cannot be an eigenfunction in any usual sense: both any leaky (complex-ω) mode must grow exponentially far from the defect. Physically, this is because an observer far away is seeing radiation from the leaky mode as it was in the past, when the mode was exponentially larger. Really, we will see that such "modes" are local approximations that take into account only a single pole (= resonance) in the Green's function.

Further reading: For evanescent waves and defect cavities, see chapter 4 of the textbook. For a proof of localization in gaps by arbitrarily weak defects in 1d and 2d, for the Schrodinger equation, see Parzygnat et al. (2010). The leaky modes are subtle things. Completeness of leaky modes is discussed in Leung et al. (1994). A common viewpoint is that leaky modes are a steepest-descent approximation (saddle-point approximation) to a Green's function integral, taking into account the contribution of a single pole; see e.g. Snyder and Love Optical Waveguide Theory (1983).

Lecture 15: 7 Mar

Handouts: Section 4.4 (LDOS) of Electromagnetic Wave Source Conditions.

Discussed the concept of a resonance as a pole in the Green's function. Reviewed Green's functions. By expanding in the eigenfunctions for a Hermitian system with a discrete set of eigenvalues, showed that we have a sequence of poles on the real-ω axis. If we add loss, either by adding absorption to the materials or by opening up the system to let radiation escape to infinity, these poles are pushed down to Im[ω]<0. (As discussed last time, these are poles but are not strictly "eigenvalues" in an open problem, because they would lead to "eigenfunctions" that grow exponentially in space.)

If we have a pole at ω0-iΓ0 that is close to the real axis (Γ0 ≪ ω0), considered a narrow-bandwidth pulse that interacts mainly with this pole. If we can approximate the Green's function just by the contribution of this pole (which is true near ω0 at at positions in or near the resonant cavity), then showed via contour integration (or alternatively one could use saddle-point integration) that the response is approximately an exponentially decaying "leaky mode" proportional to exp(iω0t - Γ0t). However, this is only a local approximation; far away from the cavity, the response will not look like this.

Discussed the "principle of limiting absorption:" the "right" way to define a "lossless" system is to add a little bit of loss everywhere and take the limit as this loss goes to zero from above (Im ωε = 0+). This allows us to rigorously deal with poles on the real-ω axis, and also automatically gives us outgoing ("radiation" or "Sommerfield") boundary conditions.

Alternatively, an approach that lets us talk about the "local" spectrum of finite periodic structures, open resonators, and other lossy cases, is the local density of states. Began discussing section 4.4 of the handout (DOS and LDOS), and claimed without proof that the LDOS is also proportional to the power radiated by a dipole source at a given position and frequency: this latter definition has the advantage of being much easier to generalize, and easier to connect to other physical processes like spontaneous emission or antennas.

Further reading: See the Snyder and Love textbook cited last lecture for leaky modes via saddle-point ("steepest-descent") integration. For the limiting absorption principle, see e.g. Schulenberger and Wilcox (1971). See section 4.4 of the handout and references therein for more information on DOS and LDOS.

Lecture 16: 9 Mar

Went through sections 4.4.1-4.4.3 of the handout from last lecture (handed out a few more pages and a plot from the "Electromagnetic Wave Source Conditions" chapter).

Lecture 17: 11 Mar

Handouts: TE/TM projected band diagram and omnidirectional reflection (from book chapter 4, figure 15), pset 3 solutions, pset 4 (due Monday March 28) [see file defect1d.ctl]

Went through sections 4.4.5-4.4.6 of the handout from last lecture: van Hove singularities and Purcell "Q/V" enhancement.

Off-axis propagation, projected band diagrams for multilayer films, Fabry-Perot defect modes, and surface states. Omnidirectional reflection for the TM polarization.

Further reading: For off-axis propagation in multilayer films, see chapter 4 of the textbook. See also projected TM band diagram from multilayer film (corrected from figure 10 of chapter 4 in the book). See chapter 4 of the book, final section on omnidirectional reflection; see any book on optics or advanced electromagnetism for Brewster's angle (e.g. Jackson or Hecht). For Van Hove singularities, see e.g. Solid State Physics by Ashcroft and Mermin and the other citations in the handout.

Lecture 18: 14 Mar

Omnidirectional reflection: sketched TM/TE projected band diagram for multilayer film and identified the possibility of a range of omnidirectional reflection from air (i.e. a range of 100% reflection for all incident angles and polarizations of incident propagating waves, as long as translational symmetry is not broken). Identified the two key criteria that the index contrast be large enough and that the lower of the two mirror indices be larger than that of the ambient medium (air). Explained how the latter condition, and the odd shape of the TE projected band diagram, arise from Brewster's angle.

Wave propagation velocity: defined phase velocity (along homogeneous directions) and group velocity. Explained why phase velocity is not uniquely defined in a periodic medium (and even in a uniform waveguide it can easily be infinite). Showed that group velocity is the velocity of propagation of wave packets, by considering a narrow-bandwidth packet and Taylor-expanding the dispersion ω(k) to first order.

Another viewpoint is that group velocity is the energy-propagation velocity (in a lossless medium), and explained the general principle that the velocity of any "stuff" can be expressed as the ratio of the flux rate of the stuff to the density of the stuff...we will show that in the electromagnetic case, this ratio gives exactly dω/dk. In particular, we will apply the Hellmann-Feynman theorem to our Θk eigenproblem to show that the group velocity dω/dk is precisely the energy velocity (ratio of energy flux to energy density, averaged over time and the unit cell).

Further reading: See chapter 3 of the book, section on phase and group velocity. See the footnotes in that section, e.g. Jackson, Classical Electrodynamics, for a derivation of group velocity from the Fourier perspective; see also my notes on wave velocity and Fourier transforms from 18.303. A much more complete review of velocity in lossy and dispersive media can be found in our 2014 paper Speed-of-light limitations in passive linear media.

Lecture 19: 16 Mar

Applied the Hellmann-Feynman theorem to our Θk eigenproblem to show that the group velocity dω/dk is precisely the energy velocity (ratio of energy flux to energy density, averaged over time and the unit cell).

From the energy velocity expression, proved that this group velocity is always ≤c for ε≥1. (At a deeper level, it turns out that you can prove this for any passive media.) Also gave a simple proof that the "front velocity" (the rate at which the wave "front" of nonzero fields can move) is bounded by the upper bound of the energy velocity.

Discussed how the group velocity implies that the derivative of phase with respect to frequency is a time delay, but emphasized that this only applies in (approximately) lossless where group velocity is an accurate measure of energy velocity. Gave a simple example of transmission through evanescent waves (through the band gap of a finite 1d-periodic medium), in which a naive application of this "phase delay" formula would imply instantaneous (zero delay) transmission.

Discussed group-velocity dispersion, qualitatively.

Further reading: See chapter 3 of the book, section on phase and group velocity. For a discussion of dispersion (and dispersion compensation) as it applies in optical fibers, see e.g. R. Ramaswami and K. N. Sivarajan, Optical Networks: A Practical Perspective.

Lecture 19: 18 Mar

Handout: BEM lecture notes

Guest lecture by Dr. Homer Reid: surface-integral equations (SIEs) and boundary-element methods (BEMs) in electromagnetism.

Further reading: SCUFF-EM, Homer's free BEM code for electromagnetic scattering problems.

Lecture 20: 28 Mar

Handout: pset 4 solutions, pset 5 (due next Monday), figures 2 and 3 from book, chapter 5, 2d square/triangular-lattice Brillouin zones (from appendix B of the book)

New topic: 2d periodicity

Reviewed Bloch's theorem, the primitive lattice vectors, the Bravais lattice, the primitive reciprocal lattice vectors, and the reciprocal lattice, for 2d periodicity. Reviewed the periodicity in k-space (reciprocal space),

In 1d, we already saw the simplest example of a Brillouin zone, the interval [-π/a,+π/a]. Showed that in the square lattice, things are similarly simple: the natural Brillouin zone is just a square "unit cell" centered on the origin, with diameter 2π/a. Showed how the symmetries of the structure can reduce this to an "irreducible Brillouin zone" (I.B.Z) that is just a triangle, and gave the canonical Γ/X/M names for the corners of this triangle. Pointed out that there are four equivalent M points and two equivalent X points, by periodicity in k space; there is also a Y point that is the 90-degree rotation of the X point, whose solutions are related (in a symmetric structure) but are not the same as at X.

Began more careful discussion of Brillouin zones, by looking at the triangular lattice. Defined lattice vectors, found reciprocal lattice vectors, and showed that the reciprocial lattice is also triangular but rotated 30°. Noted that the "unit cell" of the lattice, however it is chosen, does not have the full symmetry, motivating us to seek a better definition of the first Brillouin zone.

Showed how to construct the first Brillouin zone (and the second Brillouin zone, etc.) via perpendicular bisectors between reciprocal lattice points. (The generalization of this to non-periodic structues is called a Voronoi cell, and in the real lattice it is called a Wigner–Seitz cell.) Showed that B.Z. contains no equivalent k points (not including the B.Z. boundaries), and all inequivalent k points (if you include the B.Z. boundaries). Showed that the B.Z. has the full symmetry of the point group. We can therefore construct the irreducible Brillouin zone (I.B.Z.), which is the B.Z. reduced by all of the symmetries in the point group (+ time reversal), and are the only k we need to consider.

Gave the examples of the square-lattice B.Z. and the triangular-lattice B.Z., constructed in this way, and reduced the latter to the I.B.Z. for a 6-fold symmetrical (C6v) structure.

Considered the TM band diagram of the square lattice of rods (figure 1 of the handout). Discussed the origin of the gap from the variational theorem (explaining the band-edge field patterns in figure 2), and the reason for a minimum index contrast to get a gap (the differing periodicities and hence differing gaps in different directions).

Considered the space group at various k points in the I.B.Z., where k breaks some of the symmetry. Showed that Γ and M have the full symmetry of the lattice, whereas X has a reduced symmetry group. Furthermore, from the symmetry of the points between Γ and M or Γ and X, explained why we have zero group velocity at the X and M points, and why the local maxima (usually) lie along the I.B.Z. boundaries.

Further reading: beginning of chapter 5 of the book (2d photonic crystals), and appendix B on the reciprocal lattice and Brillouin zone. See this paper for some counterexamples and further discussion regarding the occurrence of band extrema at the edges of the I.B.Z.

Lecture 21: 30 March

Discussed reflection/diffraction/refraction at 2d crystal interfaces, following closely the treatment at the end of chapter 10 in the book. Relationship of isofrequency diagrams, group velocity, and conservation of k||. Briefly discussed negative refraction, flat-lens imaging, supercollimation. Talked a little about metamaterials, in the limit λ>>a, where the crystal can be replaced by a homogenized effective medium; for the most part this course deals with the regime where λ is comparable to a.

Further reading: Chapter 5, chapter 10 (section on reflection, refraction, and diffraction).

Lecture 22: 1 April

Discussed metamaterials and homogenezation theory: at long wavelengths (much larger than the periodicity), one can replace the crystal with an effective ε (and possibly μ) that reproduces essentially all of the unit-cell-averaged behaviors: dispersion relations, scattering at interfaces, etcetera. Discussed why this ε is in general anisotropic, using the example of a 1d-periodic (multilayer-film) metamaterial. However, if the crystal has sufficient symmetry, we will get an isotropic effective medium. In contrast, at longer wavelengths (comparable to the periodicity), one can sometimes define an "effective medium" that captures some aspects of the crystal (e.g. matching the dispersion relation), but in general no single effective medium will capture all of the behaviors (e.g. it will give the wrong reflection coefficient at an interface). (This can sometimes lead to confusing remarks in the literature.) I would tend to use the term "metamaterial" only in the long-wavelength regime where the true homogenization theory is (approximately) valid, but some authors use the term more loosely.

Point-defect states in the square lattice of rods. Either decreasing the radius of a rod to push up a "monopole" state, or increasing the radius of a rod to pull down a "dipole" state. Showed how we can easily predict the qualitative field patterns and symmetries from the corresponding bands of the bulk crystal. Related the defect modes to the 5 irreps of the C4v symmetry group, and showed how we can easily guess the field patterns and degeneracies that we will get.

Further reading: textbook, chapter 5 and 6

Lecture 23: 4 April

Handout: pset 5 solutions

Line-defect states and waveguides in 2d photonic crystals. Projected band diagrams for the line defect, and the guided mode. Emphasize differences from index-guiding (can guide in air) and Fabry-Perot waveguides (even if we break translational symmetry, light can only scatter forwards or back—the waveguide effectively forms a one-dimensional system).

Surface states in 2d crystals.

Lecture 24: 6 April

Handout: computation slides (also ppt format)

New topic: Computational photonics. Began by categorizing computational methods along three axes: what problem is solved, what basis/discretization is used to reduce the problem to finitely many unknowns, and how are the resulting finitely many equations solved? Discussed three categories of problems: full time-dependent Maxwell solvers, responses to time-harmonic currents J(x) e-iωt, and eigenproblems (finding ω from k or vice-versa). Emphasized that there is no "best" method; each method has its strengths and weaknesses, and there are often strong tradeoffs (e.g. between generality/simplicity and efficiency).

Frequency-domain eigensolvers, e.g. MPB.

Explained the Galerkin method to turn linear differential/integral equation, plus a finite-basis approximation, into a finite set of N equations in N unknowns. Showed that Galerkin methods preserve nice properties like positive-definiteness and Hermitian-ness, but generally turn ordinary eigenproblems into generalized ones (unless you happen to have an orthonormal basis).

Talked about solving the frequency-domain eigenproblem in a planewave (spectral) basis, ala MPB. One big motivation for using a planewave basis is that it makes it trivial to enforce the transversality constraint (∇ċH=0), which is diagonal in Fourier space.

In order to solve this equation, we could simply throw it directly at Matlab or LAPACK (LAPACK is the standard free linear-algebra library that everyone uses). With N degrees of freedom, however, this requires O(N2) storage and O(N3) time, and this quickly gets out of hand. Instead, since we only want a few low-frequency eigenvalues (not N!), we use iterative methods, which start with a guess for the solution (e.g. random numbers) and then iteratively improve it to converge to any desired accuracy. Most iterative solvers require only a black-box routine that computes matrix times vector.

For Hermitian eigenproblems, one class of iterative techniques is based on minimizing the Rayleigh quotient: given any starting guess, if we "go downhill" in the Rayleigh quotient then we will end up at the lowest eigenvalue and corresponding eigenvector. We can find subsequent eigenvalues/eigenvectors by deflation: repeating the process in the subspace orthogonal to the previous eigenvectors. A very simple optimization technique is steepest-descent: repeated line searches in the downhill direction given by the gradient of the Rayleigh quotient. In practice, there are better optimization methods for this problem than steepest descent, such as the nonlinear conjugate-gradient method, and preconditioning, but they have a similar flavor. Showed the effect of the different iteration schemes on convergence rate (see handout).

The key to applying iterative methods efficiently for this problem is to use fast Fourier transforms (FFTs) to perform the Θk matrix-vector product in O(N log N) time and O(N) storage.

A planewave basis actually converges exponentially fast if everything is a smooth (analytic) periodic function, but this is not true if ε is discontinuous (as it usually) is: the Fourier series of a discontinuous function converges only at a linear rate (error ~ 1/#terms in 1d). The planewave basis is dual to a uniform grid under a discrete Fourier transform (DFT), so we can equivalently think of "staircasing" of interface, and in general the question is what ε to assign to pixels straddling the boundaries. Intuitively, boundary pixels should be assigned some intermediate ε value, which is equivalent to discretizing a smoothed structure—but then we face the problem that the act of smoothing changed the structure, and itself introduces a 1st-order error in general. Argued (see handout) that the right thing to do is to assign an anisotropic ε to interfaces: one can show that the proper anisotropic ε corresponds to a smoothing that introduces zero 1st-order error, and hence leads to 2nd-order convergence as shown in the handout.

Further reading Textbook, appendix D. See our paper on MPB for more detail on planewave-based eigensolvers for electromagnetism. Spectral methods, Galerkin, etcetera: J. P. Boyd, Chebyshev and Fourier Spectral Methods. Iterative eigensolver methods: Bai et al, Templates for the Solution of Algebraic Eigenvalue Problems; also Numerical Linear Algebra by Trefethen and Bau (readable online with MIT certificates). See e.g. this paper on subpixel-smoothing and perturbation theory.

Midterm Exam: 8 April

Room 2-146. You have 2 hours: come either 1-3pm or 2-4pm. You can bring in one sheet (2 sides) of notes (though it probably won't help), and you can also bring the representation theory handout. Covers through Friday, April 1.

Previous midterms: 2005 (was way too long) and solutions, 2007 and solutions; 2008; 2009 and solutions; 2010 and solutions; 2012 and solutions; 2014 and solutions.

2016 exam and solutions (see also IJulia notebook for problem 1).

Conflict date: 11 April, 10am-noon, room 2-143.

Lecture 25: 11 April

See slides from lecture 24.

Switched problems to time-domain solvers: find the time-dependent fields in response to an arbitrary time-dependent current, for some initial conditions. This is the most general solution technique, and can handle things like nonlinearities and time-dependent media in which frequency is not conserved (a problem for frequency-domain methods). On the other hand, when a more specialized method (e.g. a frequency-domain eigensolver) is available, often it is easier and more bulletproof than using the most general tool.

In particular, talked about finite-difference time-domain (FDTD) methods, in which space and time are broken up into uniform grids. Started with 1+1 dimensions (1 space + 1 time). Mentioned the second-order accuracy of center-difference approximations, and in order to utilize this in FDTD concluded that we need to store H and E on grids staggered in time and space: a Yee grid. Wrote down the general "leapfrog" scheme for time-stepping the fields.

Derived the CFL stability condition (in 1d) relating Δt to Δx. This arises from the simple fact that, if the scheme is to converge as Δx and Δt go to zero, it must support solutions that propagate at speed c, but the maximum speed at which information propagates in the grid is Δx/Δt, and hence Δx/Δt>c (in 1d). As a consequence, if we make the spatial discretization finer, we must also make the time discretization finer. e.g. in 3d this means that doubling the spatial resolution increases the total simulation time by (at least) a factor of 16=24.

Further reading: See e.g. these notes on finite-difference approximations for the basic ideas of center differences etc. For FDTD in general, see e.g. Allen Taflove and Susan C. Hagness, Computational Electrodynamics: The Finite-Difference Time-Domain Method (Artech, 2005). For the CFL condition in general, see e.g. this book chapter by Trefethen. See also our free FDTD software: Meep, and in particular the introduction and tutorial sections of the Meep manual.

Lecture 26: 13 April

Discussed absorbing boundary conditions and perfectly matched layers (PML); see notes below and slide handout above. Introduced PML as an analytic continuation of the solution and equations into complex coordinates in the direction perpendicular to the boundary. Showed how this transforms oscillating solutions into decaying ones without introducing reflections (in theory). Showed how we transform back to real coordinates, and the entire PML implementation can be summarized by a single equation: ∂/∂x → (1+iσ/ω)−1∂/∂x, where σ(x) is some function that is positive in the PML and zero elsewhere, characterizing the strength of the decay.

Discussed fact (to be proved later, see notes below) that any coordinate transformation (including the complex one for PML) can be represented as merely a change in ε and μ, while keeping Maxwell's equations in Cartesian form. This means that PML can be interpreted simply as an anisotropic absorbing material ("UPML").

Limitations of PML. Discussed fact that PML is no longer reflectionless in discretized equations, but this is compensated for by turning on the absorption (e.g.) quadratically over a wavelength or so. (Any absorption, turned on slowly enough, has negligible reflections; this idea is used e.g. in anechoic chambers.) Note that PML requires Maxwell's equations to be invariant in the direction ⊥ to the PML, which excludes photonic crystals from having any true PML. Briefly discussed (see slides for more detail) on how this fact has sometimes been confused in the literature, since the lack of a true PML can be disguised if you turn on the PML gradually enough (over many periods).

Further reading: Notes on PML; see also e.g. the discussion of PML in Taflove's book. Notes on coordinate transforms in electromagnetism; see also slides on transformation optics

Lecture 27: 15 April

See slides and notes from Lecture 24. Covered implementation of PML in the time domain and auxiliary differential equations.

Also covered the Principle of Equivalence in order to design sources corresponding to specific field patterns.

Further reading: For the principle of equivalence and its application to wave sources in FDTD, see our review article Electromagnetic Wave Source Conditions (excerpted from Advances in FDTD Computational Electrodynamics: Photonics and Nanotechnology, 2013). For a free/open-source BEM code, see SCUFF-EM by Homer Reid.

Lecture 27: 20 April

Demo of Meep FDTD code (installed on Athena/Linux machines: add meep). In particular, went through the tutorial in the Meep manual, and covered the basic techniques to find transmission/reflection spectra.

New topic: temporal coupled-mode theory (TCMT). Started with a canonical device, a waveguide-cavity-waveguide filter, and began to derive how the universal behavior of device in this class can be derived from very general principles such as conservation of energy, parameterized only by the (geometry-dependent) frequency and lifetime of the cavity mode.

Began by parameterizing the unknowns: the amplitude A in the cavity and the incoming/outgoing wave amplitudes s in each channel k, normalized so that |A|2 is energy in the cavity and |s|2 is power. Then wrote down the most general linear time-invariant equation relating A and the incoming wave from a single input port: A(ω)=g(ω) s1+(ω), where g(ω) is some function of frequency in the frequency domain (a type of Green's function or generalized susceptibility). The key assumption of TCMT is resonance: we assume that there is a resonant mode, corresponding to a pole in g(ω) (or the LDOS) at a complex frequency ω0−i/τ, and that 1/τ<<ω0 so that g(ω)≈α/(iω-iω0−1/τ), where α/i is the residue of the pole (i.e. g is dominated near ω0 by the contribution of the pole, and far from ω0 the amplitude A is so small that we will neglect it). In time domain, this corresponds to a simple ODE dA/dt = -i ω0A − A/τ + αs1+, where α is a constant to be determined.

Also wrote down the most general linear time-invariant relation for the outgoing amplitude s1−=βs1++γA for some constants β and γ in the freuqency domain. If we are only interested in the response of the system near resonance, then we can approximate β and γ by their values at ω0, in which case they are constants and the s1−=βs1++γA is valid in the time domain as well.

What remains is to eliminate the unknowns α, β, and γ. Do this by the method in chapter 10 of the book: apply energy conservation to find γ=√(2/τ) (up to an arbitrary phase choice), and then time-reversal symmetry (or reciprocity) to find β=−1 and γ=√(2/τ). Hence, the only geometry- and physics- dependent parameters in the problem are ω0 and τ.

Further reading: chapter 10 of the book.

Lecture 28: 22 April

Handouts: slides on TCMT resonant-mode calculations (also in ppt format)

Showed that the transmission in a waveguide-cavity-waveguide system is always a Lorentzian curve peaked at 100% (for symmetric decay) with a width inversely proportional to the lifetime, and showed that this happens because of a resonant cancellation in the reflected wave.

More examples of coupled-mode theory: A waveguide splitter, a waveguide crossing, external loss (and the need for and absorption Q much larger than the total Q for small loss) and resonant absorption (maximizing absorption, with application to photovoltaics).

Analyzed the time-delay in a resonant filter and showed that: it is given by dφ/dω (the derivative of the phase) if we have a narrow-band pulse near the transmission maximum (so that amplitude is independent of frequency to first order), and that this time delay is precisely τ (the lifetime of the cavity mode).

Furthermore defined the quality factor Q of the cavity, which is simply a dimensionless lifetime Q=ω0τ/2.

Further reading: chapter 10 of the book and references therein. For waveguide splitters, see also Fan et al. (2001) and Manolatou et al. (1999); for crossings, see also Johnson et al. (1998).

Lecture 29: 25 April

Handouts: pages 144–195 of my photonic-crystal tutorial slides

New topic periodic dielectric waveguides (chapter 7).

Reviewed periodic dielectric waveguides, which we've seen once or twice before: periodic replication of the light cone and bands below that which flatten out at the edge of the Brillouin zone. Incomplete gaps: ranges of frequencies where there are no guided modes (but still light-cone modes). Symmetry and polarization.

Photonic-crystal slabs: band gaps, symmetry/polarization, and line-defect waveguides. Microcavities (very similar to analysis in periodic dielectric waveguides).

Partial confinement of light by defects, and intrinsic radiation losses due to coupling to light-line mode. Tradeoff between localization and loss (due to Fourier components inside the light cone). Discussed two mechanisms for large radiation Q despite the incomplete gap: delocalization and cancellation.

Further reading: Chapters 7 and 8 of the book.

Lecture 30: 27 April

Handouts: pages 197–236 of my photonic-crystal tutorial slides

Finished discussing delocalization and cancellation mechanisms for high Q slab cavities.

New topic: photonic-crystal fibers. Discussed the various types from the handouts: photonic-bandgap vs. index guiding, and 2d-periodic vs. Bragg fibers (concentric "1d" crystals). Emphasized the importance of the band gap lying above the light line of air.

Lecture 31: 29 April

The short-wavelength scalar approximation and its consequences for holey fibers.

Discussed consequences of the scalar limit. First for a dielectric waveguide with a square or rectangular cross-section (which maps to the square TM metallic cavity of pset 2), and then for a holey fiber with a solid core (which maps to a 2d metallic photonic crystal). In both cases, applied product representation theory to the relationship between the scalar LP modes and the vector modes. Noted that holey fibers will support only a finite number of guided modes (and can even be "endlessly single mode" for the right parameters).

Discussed the origin of band gaps in the holey-fiber light cone, from the scalar limit, and band-gap guidance in hollow-core fibers.

Further reading: chapter 9 (section on index-guiding holey fibers and the scalar limit). For a rigorous derivation of the scalar limit, see this 1994 paper by Bonnet-Bendhia and Djellouli.

Lecture 32: 2 May

Handouts: notes, first few pages of our 2002 adiabatic-theorem paper

Discussed Bragg fibers (from chapter 9 of the book), emphasing the role of rotational symmetry (in excluding solutions propagating azimuthally at large radii).

New topic (see notes): Going full-circle back to the beginning of the course, we again derive an algebraic (linear operator / eigenproblem) formulation of Maxwell's equations. This time, however, we do so for constant-ω separating out the z derivative and the corresponding k component (for z-periodic structures) kz (denoted β). That is, we write Maxwell's equations in the form:

A ψ = -i B ∂ψ/∂z

where ψ is a four-component vector field consisting of (Ex, Ey, Hx, Hy), and A and B are linear operators. This is the most convenient formulation for considering problems of propagation in the z direction along a waveguide, where perturbations may break translational symmetry but frequency is still conserved. Showed that A andB are Hermitian (but not positive-definite) for real ε and μ.

Lecture 32: 4 May

Continued with notes from previous lecture, discussed orthogonality of modes and unconjugated "inner products," propagating vs. evanescent modes (showing that the latter carry zero power), and the β eigenproblem for z-periodic problems.

Began setting up the expansion for slowly z-varying systems (e.g. tapers).

Lecture 33: 6 May

Derived the coupled-wave equations (see paper from last time) and the adiabatic theorem for slowly z-varying problems (e.g. waveguide tapers, gradual bends, etcetera).

Connected the adiabatic limit to the rate of convergence of the Fourier transform of the rate of change. See section 2.1 of Oskooi et al. (2012).

Briefly discussed the periodic version of the adiabatic theorem (periodic systems where the unit cell is slowly changing from one period to the next), which is similar in spirit but much more complicated in detail.

Lecture 34: 9 May

Handouts: Notes on coordinate transforms in electromagnetism, slides on transformation optics

Discussed how any coordinate transformation (including the complex one for PML) can be represented as merely a change in ε and μ, while keeping Maxwell's equations in Cartesian form. (See handout for proof.) This can be used to derive the "UPML" formulation of PML as anisotropic absorbing materials, and for neat theoretical results such as "invisibility cloaks."

Coordinate transformation allow us to relate waveguide bends to material perturbations Δε, as discussed in section 7.1 of Johnson et al. (2001).

Lecture 35: 11 May

Slides: my slides from a 2014 seminar

Discussed a basic picture of lasers: the Maxwell-Bloch equations, the SALT nonlinear eigenproblem for the steady-state modes, recent SALT computational methods, and laser linewidth limits due to quantum/thermal noise.

Further reading: See e.g. Haken's Laser Theory for the Maxwell-Bloch equations, Ge, Chong, & Stone (2010) on SALT, Esterhazy et al. (2014) on computational SALT, Pick et al. (2014) and references therein on laser linewidth.