[By prior agreement, this post will be cross-posted on Microsoft’s Q# blog, even though it has nothing to do with the Q# programming language. It does, however, contain many examples that might be fun to implement in Q#!]

Why should Nature have been quantum-mechanical? It’s totally unclear what would count as an answer to such a question, and also totally clear that people will never stop asking it.

Short of an ultimate answer, we can at least try to explain why, if you want this or that *piece* of quantum mechanics, then the rest of the structure is inevitable: why quantum mechanics is an “island in theoryspace,” as I put it in 2003.

In this post, I’d like to focus on a question that any “explanation” for QM at some point needs to address, in a non-question-begging way: **why should amplitudes have been complex numbers? **When I was a grad student, it was his relentless focus on that question, and on others in its vicinity, that made me a lifelong fan of Chris Fuchs (see for example his samizdat), despite my philosophical differences with him.

It’s not that complex numbers are a *bad* choice for the foundation of the deepest known description of the physical universe—far from it! (They’re a field, they’re algebraically closed, they’ve got a norm, how much more could you want?) It’s just that they seem like a *specific* choice, and not the only possible one. There are also the real numbers, for starters, and in the other direction, the quaternions.

Quantum mechanics over the reals or the quaternions still has constructive and destructive interference among amplitudes, and unitary transformations, and probabilities that are absolute squares of amplitudes. Moreover, these variants turn out to lead to precisely the same power for quantum computers—namely, the class BQP—as “standard” quantum mechanics, the one over the complex numbers. So none of those are relevant differences.

Indeed, having just finished teaching an undergrad Intro to Quantum Information course, I can attest that the complex nature of amplitudes is needed only rarely—shockingly rarely, one might say—in quantum computing and information. Real amplitudes typically suffice. Teleportation, superdense coding, the Bell inequality, quantum money, quantum key distribution, the Deutsch-Jozsa and Bernstein-Vazirani and Simon and Grover algorithms, quantum error-correction: all of those and more can be fully explained without using a single *i* that’s not a summation index. (Shor’s factoring algorithm is an exception; it’s much more natural with complex amplitudes. But as the previous paragraph implied, their use is removable even there.)

It’s true that, if you look at even the simplest “real” examples of quantum systems—or as a software engineer might put it, at the application layers built on top of the quantum OS—then complex numbers are everywhere, in a way that seems impossible to remove. The Schrödinger equation, energy eigenstates, the position/momentum commutation relation, the state space of a spin-1/2 particle in 3-dimensional space: none of these make much sense without complex numbers (though it can be fun to try).

But from a sufficiently Olympian remove, it feels circular to use any of this as a “reason” for why quantum mechanics should’ve involved complex amplitudes in the first place. It’s like, once your OS provides a certain core functionality (in this case, complex numbers), it’d be surprising if the application layer *didn’t* exploit that functionality to the hilt—especially if we’re talking about fundamental physics, where we’d like to imagine that nothing is wasted or superfluous (hence Rabi’s famous question about the muon: “who ordered that?”).

But why should the quantum OS have provided complex-number functionality at all? Is it possible to answer that question purely in terms of the OS’s internal logic (i.e., abstract quantum information), making minimal reference to how the OS will eventually get used? Maybe not—but if so, then that itself would seem worthwhile to know.

If we stick to abstract quantum information language, then the most “obvious, elementary” argument for why amplitudes should be complex numbers is one that I spelled out in *Quantum Computing Since Democritus*, as well as my Is quantum mechanics an island in theoryspace? paper. Namely, it seems desirable to be able to implement a “fraction” of any unitary operation U: for example, some V such that V^{2}=U, or V^{3}=U. With complex numbers, this is trivial: we can simply diagonalize U, or use the Hamiltonian picture (i.e., take e^{-iH/2} where U=e^{-iH}), both of which ultimately depend on the complex numbers being algebraically closed. Over the reals, by contrast, a 2×2 orthogonal matrix like $$ U = \left(\begin{array}[c]{cc}1 & 0\\0 & -1\end{array}\right)$$

has no 2×2 orthogonal square root, as follows immediately from its determinant being -1. If we want a square root of U (or rather, of something that acts like U on a subspace) while sticking to real numbers only, then we need to add another dimension, like so: $$ \left(\begin{array}[c]{ccc}1 & 0 & 0\\0 & -1 & 0\\0 & 0&-1\end{array}\right)=\left(\begin{array}[c]{ccc}1 & 0 & 0\\0 & 0 & 1\\0 & -1 & 0\end{array}\right) ^{2} $$

This is directly related to the fact that there’s no way for a Flatlander to “reflect herself” (i.e., switch her left and right sides while leaving everything else unchanged) by any continuous motion, unless she can lift off the plane and rotate herself through the third dimension. Similarly, for *us* to reflect ourselves would require rotating through a fourth dimension.

One could reasonably ask: is that it? Aren’t there any “deeper” reasons in quantum information for why amplitudes should be complex numbers?

Indeed, there are certain phenomena in quantum information that, slightly mysteriously, work out more elegantly if amplitudes are complex than if they’re real. (By “mysteriously,” I mean not that these phenomena can’t be 100% verified by explicit calculations, but simply that I don’t know of any deep principle by which the results of those calculations could’ve been predicted in advance.)

One famous example of such a phenomenon is due to Bill Wootters: if you take a uniformly random pure state in d dimensions, and then you measure it in an orthonormal basis, what will the probability distribution (p_{1},…,p_{d}) over the d possible measurement outcomes look like? The answer, amazingly, is that you’ll get a *uniformly random probability distribution*: that is, a uniformly random point on the simplex defined by p_{i}≥0 and p_{1}+…+p_{d}=1. This fact, which I’ve used in several papers, is closely related to Archimedes’ Hat-Box Theorem, beloved by friend-of-the-blog Greg Kuperberg. But here’s the kicker: it only works if amplitudes are complex numbers. If amplitudes are real, then the resulting distribution over distributions will be too bunched up near the corners of the probability simplex; if they’re quaternions, it will be too bunched up near the middle.

There’s an even more famous example of such a Goldilocks coincidence—one that’s been elevated, over the past two decades, to exalted titles like “the Axiom of Local Tomography.” Namely: suppose we have an unknown finite-dimensional mixed state ρ, shared by two players Alice and Bob. For example, ρ might be an EPR pair, or a correlated classical bit, or simply two qubits both in the state |0⟩. We imagine that Alice and Bob share many identical copies of ρ, so that they can learn more and more about it by measuring this copy in this basis, that copy in that basis, and so on.

We then ask: can ρ be fully determined from the joint statistics of *product measurements*—that is, measurements that Alice and Bob can apply separately and locally to their respective subsystems, with no communication between them needed? A good example here would be the set of measurements that arise in a Bell experiment—measurements that, despite being local, certify that Alice and Bob must share an entangled state.

If we asked the analogous question for classical probability distributions, the answer is clearly “yes.” That is, once you’ve specified the individual marginals, and you’ve *also* specified all the possible correlations among the players, you’ve fixed your distribution; there’s nothing further to specify.

For quantum mixed states, the answer again turns out to be yes, *but only because amplitudes are complex numbers!* In quantum mechanics over the reals, you could have a 2-qubit state like $$ \rho=\frac{1}{4}\left(\begin{array}[c]{cccc}1 & 0 & 0 & -1\\0 & 1 & 1 & 0\\0 & 1 & 1 & 0\\-1& 0 & 0 & 1\end{array}\right) ,$$

which clearly isn’t the maximally mixed state, yet which is indistinguishable from the maximally mixed state by any local measurement that can be specified using real numbers only. (Proof: exercise!)

In quantum mechanics over the quaternions, something even “worse” happens: namely, the tensor product of two Hermitian matrices need not be Hermitian. Alice’s measurement results might be described by the 2×2 quaternionic density matrix $$ \rho_{A}=\frac{1}{2}\left(\begin{array}[c]{cc}1 & -i\\i & 1\end{array}\right), $$

and Bob’s results might be described by the 2×2 quaternionic density matrix $$ \rho_{B}=\frac{1}{2}\left(\begin{array}[c]{cc}1 & -j\\j & 1\end{array}\right), $$

and yet there might not be (and in this case, isn’t) any 4×4 quaternionic density matrix corresponding to ρ_{A}⊗ρ_{B}, which would explain both results separately.

What’s going on here? Why do the local measurement statistics *underdetermine* the global quantum state with real amplitudes, and *overdetermine* it with quaternionic amplitudes, being in one-to-one correspondence with it only when amplitudes are complex?

We can get some insight by looking at the number of independent real parameters needed to specify a d-dimensional Hermitian matrix. Over the complex numbers, the number is exactly d^{2}: we need 1 parameter for each of the d diagonal entries, and 2 (a real part and an imaginary part) for each of the d(d-1)/2 upper off-diagonal entries (the lower off-diagonal entries being determined by the upper ones). Over the real numbers, by contrast, “Hermitian matrices” are just real symmetric matrices, so the number of independent real parameters is only d(d+1)/2. And over the quaternions, the number is d+4[d(d-1)/2] = 2d(d-1).

Now, it turns out that the Goldilocks phenomenon that we saw above—with local measurement statistics determining a unique global quantum state when and only when amplitudes are complex numbers—ultimately boils down to the simple fact that $$ (d_A d_B)^2 = d_A^2 d_B^2, $$

but $$\frac{d_A d_B (d_A d_B + 1)}{2} > \frac{d_A (d_A + 1)}{2} \cdot \frac{d_B (d_B + 1)}{2},$$

and conversely $$ 2 d_A d_B (d_A d_B – 1) < 2 d_A (d_A – 1) \cdot 2 d_B (d_B – 1).$$

In other words, only with complex numbers does the number of real parameters needed to specify a “global” Hermitian operator, exactly match the product of the number of parameters needed to specify an operator on Alice’s subsystem, and the number of parameters needed to specify an operator on Bob’s. With real numbers it overcounts, and with quaternions it undercounts.

A major research goal in quantum foundations, since at least the early 2000s, has been to “derive” the formalism of QM purely from “intuitive-sounding, information-theoretic” postulates—analogous to how, in 1905, some guy whose name I forget derived the otherwise strange-looking Lorentz transformations purely from the assumption that the laws of physics (including a fixed, finite value for the speed of light) take the same form in every inertial frame. There have been some nontrivial successes of this program: most notably, the “axiomatic derivations” of QM due to Lucien Hardy and (more recently) Chiribella et al. Starting from axioms that sound suitably general and nontechnical (if sometimes unmotivated and weird), these derivations perform the impressive magic trick of *deriving* the full mathematical structure of QM: complex amplitudes, unitary transformations, tensor products, the Born rule, everything.

However, in every such derivation that I know of, some axiom needs to get introduced to capture “local tomography”: i.e., the “principle” that composite systems must be uniquely determined by the statistics of local measurements. And while this principle might sound vague and unobjectionable, to those in the business, it’s obvious what it’s going to be used for the second it’s introduced. Namely, it’s going to be used to rule out quantum mechanics over the real numbers, which would otherwise be a model for the axioms, and thus to “explain” why amplitudes have to be complex.

I confess that I was always dissatisfied with this. For I kept asking myself: would I have ever formulated the “Principle of Local Tomography” in the first place—or if someone else had proposed it, would I have ever accepted it as intuitive or natural—if I didn’t *already know* that QM over the complex numbers just happens to satisfy it? And I could never honestly answer “yes.” It always felt to me like a textbook example of drawing the target around where the arrow landed—i.e., of handpicking your axioms so that they yield a predetermined conclusion, which is then no more “explained” than it was at the beginning.

Two months ago, something changed for me: namely, I smacked into the “Principle of Local Tomography,” and its reliance on complex numbers, in my own research, when I hadn’t in any sense set out to look for it. This still doesn’t convince me that the principle is any sort of *a-priori* necessity. But it at least convinces me that it’s, you know, the sort of thing you can smack into when you’re not looking for it.

The aforementioned smacking occurred while I was writing up a small part of a huge paper with Guy Rothblum, about a new connection between so-called “gentle measurements” of quantum states (that is, measurements that don’t damage the states much), and the subfield of classical CS called differential privacy. That connection is a story in itself; here’s our paper and here are some PowerPoint slides.

Anyway, for the paper with Guy, it was of interest to know the following: suppose we have a two-outcome measurement E (let’s say, on n qubits), and suppose it accepts every product state with the same probability p. Must E then accept every entangled state with probability p as well? Or, a closely-related question: suppose we know E’s acceptance probabilities on every product state. Is that enough to determine its acceptance probabilities on *all* n-qubit states?

I’m embarrassed to admit that I dithered around with these questions, finding complicated proofs for special cases, before I finally stumbled on the one-paragraph, obvious-in-retrospect “Proof from the Book” that slays them in complete generality.

Here it is: if E accepts every product state with probability p, then clearly it accepts every separable mixed state (i.e., every convex combination of product states) with the same probability p. Now, a well-known result of Braunstein et al., from 1998, states that (surprisingly enough) the separable mixed states have *nonzero density* within the set of all mixed states, in any given finite dimension. Also, the probability that E accepts ρ can be written as f(ρ)=Tr(Eρ), which is linear in the entries of ρ. OK, but a linear function that’s determined on a subset of nonzero density is determined everywhere. And in particular, if f is constant on that subset then it’s constant everywhere, QED.

But what does any of this have to do with why amplitudes are complex numbers? Well, it turns out that the 1998 Braunstein et al. result, which was the linchpin of the above argument, only works in complex QM, not in real QM. We can see its failure in real QM by simply counting parameters, similarly to what we did before. An n-qubit density matrix requires 4^{n} real parameters to specify (OK, 4^{n}-1, if we demand that the trace is 1). Even if we restrict to n-qubit density matrices with real entries only, we still need 2^{n}(2^{n}+1)/2 parameters. By contrast, it’s not hard to show that an n-qubit real *separable* density matrix can be specified using only 3^{n} real parameters—and indeed, that any such density matrix lies in a 3^{n}-dimensional subspace of the full 2^{n}(2^{n}+1)/2-dimensional space of 2^{n}×2^{n} symmetric matrices. (This is simply the subspace spanned by all possible tensor products of n Pauli I, X, and Z matrices—excluding the Y matrix, which is the one that involves imaginary numbers.)

But it’s not only the Braunstein et al. result that fails in real QM: the fact that I wanted for my paper with Guy fails as well. As a counterexample, consider the 2-qubit measurement that accepts the state ρ with probability Tr(Eρ), where $$ E=\frac{1}{2}\left(\begin{array}[c]{cccc}1 & 0 & 0 & -1\\0 & 1 & 1 & 0\\0 & 1 & 1 & 0\\-1 & 0 & 0 & 1\end{array}\right).$$

I invite you to check that this measurement, which we specified using a real matrix, accepts every product state (a|0⟩+b|1⟩)(c|0⟩+d|1⟩), where a,b,c,d are real, with the same probability, namely 1/2—just like the “measurement” that simply returns a coin flip without even looking at the state at all. And yet the measurement can clearly be nontrivial on entangled states: for example, it always rejects $$\frac{\left|00\right\rangle+\left|11\right\rangle}{\sqrt{2}},$$ and it always accepts $$ \frac{\left|00\right\rangle-\left|11\right\rangle}{\sqrt{2}}.$$

Is it a coincidence that we used exactly the same 4×4 matrix (up to scaling) to produce a counterexample to the real-QM version of Local Tomography, and *also* to the real-QM version of the property I wanted for the paper with Guy? Is anything *ever* a coincidence in this sort of discussion?

I claim that, looked at the right way, Local Tomography and the property I wanted are the same property, their truth in complex QM is the same truth, and their falsehood in real QM is the same falsehood. Why? Simply because Tr(Eρ), the probability that the measurement E accepts the mixed state ρ, is a function of two Hermitian matrices E and ρ (both of which can be either “product” or “entangled”), and—crucially—is symmetric under the interchange of E and ρ.

Now it’s time for another confession. We’ve identified an elegant property of quantum mechanics that’s true but only because amplitudes are complex numbers: namely, if you know the probability that your quantum circuit accepts every product state, then you also know the probability that it accepts an arbitrary state. Yet, despite its elegance, this property turns out to be nearly useless for “real-world applications” in quantum information and computing. The reason for the uselessness is that, for the property to kick in, you really do need to know the probabilities on product states almost *exactly*—meaning (say) to 1/exp(n) accuracy for an n-qubit state.

Once again a simple example illustrates the point. Suppose n is even, and suppose our measurement simply projects the n-qubit state onto a tensor product of n/2 Bell pairs. Clearly, this measurement accepts every n-qubit product state with exponentially small probability, even as it accepts the entangled state

$$\left(\frac{\left|00\right\rangle+\left|11\right\rangle}{\sqrt{2}}\right)^{\otimes n/2}$$

with probability 1. But this implies that noticing the nontriviality on entangled states, would require knowing the acceptance probabilities on product states to exponential accuracy.

In a sense, then, I come back full circle to my original puzzlement: why *should* Local Tomography, or (alternatively) the-determination-of-a-circuit’s-behavior-on-arbitrary-states-from-its-behavior-on-product-states, have been important principles for Nature’s laws to satisfy? Especially given that, in practice, the exponential accuracy required makes it difficult or impossible to exploit these principles anyway? How could we have known a-priori that these principles would be important—if indeed they *are* important, and are not just mathematical spandrels?

But, while I remain less than 100% satisfied about “why the complex numbers? why not just the reals?,” there’s *one* conclusion that my recent circling-back to these questions has made me fully confident about. Namely: quantum mechanics over the quaternions is a **flaming garbage fire**, which would’ve been rejected at an extremely early stage of God and the angels’ deliberations about how to construct our universe.

In the literature, when the question of “why not quaternionic amplitudes?” is discussed at all, you’ll typically read things about how the parameter-counting doesn’t quite work out (just like it doesn’t for real QM), or how the tensor product of quaternionic Hermitian matrices need not be Hermitian. In this paper by McKague, you’ll read that the CHSH game is winnable with probability 1 in quaternionic QM, while in this paper by Fernandez and Schneeberger, you’ll read that the non-commutativity of the quaternions introduces an order-dependence even for spacelike-separated operations.

But none of that does justice to the enormity of the problem. To put it bluntly: unless something clever is done to fix it, quaternionic QM allows superluminal signaling. This is easy to demonstrate: suppose Alice holds a qubit in the state |1⟩, while Bob holds a qubit in the state |+⟩ (yes, this will work even for unentangled states!) Also, let $$U=\left(\begin{array}[c]{cc}1 & 0\\0 & j\end{array}\right) ,~~~V=\left(\begin{array}[c]{cc}1 & 0\\0& i\end{array}\right).$$

We can calculate that, if Alice applies U to her qubit and then Bob applies V to his qubit, Bob will be left with the state $$ \frac{j \left|0\right\rangle +

k \left|1\right\rangle}{\sqrt{2}}.$$

By contrast, if Alice decided to apply U only *after* Bob applied V, Bob would be left with the state

$$ \frac{j \left|0\right\rangle – k \left|1\right\rangle}{\sqrt{2}}.$$

But Bob can distinguish these two states with certainty, for example by applying the unitary $$ \frac{1}{\sqrt{2}}\left(\begin{array}[c]{cc}j & k\\k & j\end{array}\right). $$

Therefore Alice communicated a bit to Bob.

I’m aware that there’s a whole literature on quaternionic QM, including for example a book by Adler. Would anyone who knows that literature be kind enough to enlighten us on how it proposes to escape the signaling problem? Regardless of the answer, though, it seems worth knowing that the “naïve” version of quaternionic QM—i.e., the version that gets invoked in quantum information discussions like the ones I mentioned above—is just immediately blasted to smithereens by the signaling problem, without the need for any subtle considerations like the ones that differentiate real from complex QM.

**Update (Dec. 20):** In response to this post, Stephen Adler was kind enough to email me with further details about his quaternionic QM proposal, and to allow me to share them here. Briefly, Adler completely agrees that quaternionic QM inevitably leads to superluminal signaling—but in his proposal, the surprising and nontrivial part is that quaternionic QM would reduce to standard, complex QM at large distances. In particular, the strength of a superluminal signal would fall off exponentially with distance, quickly becoming negligible beyond the Planck or grand unification scales. Despite this, Adler says that he eventually abandoned his proposal for quaternionic QM, since he was unable to make specific particle physics ideas work out (but the quaternionic QM proposal then influenced his later work).

**Unrelated Update (Dec. 18):** Probably many of you have already seen it, and/or already know what it covers, but the NYT profile of Donald Knuth (entitled “The Yoda of Silicon Valley”) is enjoyable and nicely written.