Black Holes, Firewalls, and the Complexity of States and Unitaries Scott Aaronson (MIT) Papers and slides at

My Starting Point PHYSICS COMPUTER SCIENCE QUANTUM COMPUTING BOSONS & FERMIONS (Cf. my talk at IQC yesterday) BLACK HOLES, AdS/CFT (Today’s talk) AND MORE!

Black Holes and Computational Complexity?? YES! Amazing connection made last year by Harlow & Hayden But first, let’s review 40 years of black hole history SZK QSZK BPP BQP AM QAM

Bekenstein, Hawking 1970s: Black holes have entropy and temperature! They emit radiation The Information Loss Problem: Calculations suggest that Hawking radiation is thermal—uncorrelated with whatever fell in. So, is infalling information lost forever? Would seem to violate the unitarity / reversibility of QM OK then, assume the information somehow gets out! The Xeroxing Problem: How could the same qubit |  fall inexorably toward the singularity, and emerge in Hawking radiation? Would violate the No-Cloning Theorem Black Hole Complementarity (Susskind, ‘t Hooft): An external observer can describe everything unitarily without including the interior at all! Interior should be seen as “just a scrambled re-encoding” of the exterior degrees of freedom

Violates monogamy of entanglement! The same qubit can’t be maximally entangled with 2 things The Firewall Paradox (AMPS 2012) H = Interior of “Old” Black Hole R = Faraway Hawking Radiation B = Just-Emitted Hawking Radiation Near-maximal entanglement Also near-maximal entanglement

Harlow-Hayden 2013 (arXiv: ): Striking argument that Alice’s first task, decoding the entanglement between R and B, would take exponential time—by which point, the black hole would’ve long ago evaporated anyway Complexity theory to the rescue of quantum field theory? Are we saying that an inconsistency in the laws of physics is OK, as long as it takes exponential time to discover it? NO! “Inconsistency” is only in low-energy effective field theories; question is in what regimes they break down

Caveats of Complexity Arguments 1.Asymptotic E.g., 8  8 chess takes O(1) time! Only for n  n chess can we give evidence of hardness. But for black holes, n  … 2.(Usually) Conjectural Right now, we can’t even prove P≠NP! To get where we want, we almost always need to make assumptions. Question is, which assumptions? 3.Worst-Case We can argue that a natural formalization of Alice’s decoding task is “generically” hard. We can’t rule out that a future quantum gravity theory would make her task easy, for deep reasons not captured by our formalization.

Quantum Circuits

Given a description of a quantum circuit C, such that Promised that, by acting only on R (the “Hawking radiation part”), it’s possible to distill an EPR pair between R and B Problem: Distill such an EPR pair, by applying a unitary transformation U R to the qubits in R The HH Decoding Problem

Problem: That would require waiting until the black hole was fully evaporated (  no more firewall problem) When the BH is “merely” >50% evaporated, we know from a dimension-counting argument that “generically,” there will exist a U R that distills an EPR pair between R and B But interestingly, this argument doesn’t suggest any efficient procedure to find U R or apply it! Isn’t the Decoding Task Trivial? Just invert C!

Set Equality: Given two efficiently-computable injective functions f,g:{0,1} n  {0,1} p(n). Promised that Range(f) and Range(g) are either equal or disjoint. Decide which. In the “black-box” setting, this problem takes exp(n) time even with a quantum computer (a main result from my 2004 PhD thesis, the “collision lower bound”). Even in non-black- box setting, would let us solve e.g. Graph Isomorphism Theorem (Harlow-Hayden): Suppose there’s a polynomial-time quantum algorithm for HH decoding. Then there’s also a polynomial-time quantum algorithm for Set Equality! The HH Hardness Result

Intuition: If Range(f) and Range(g) are disjoint, then the H register decoheres all entanglement between R and B, leaving only classical correlation If, on the other hand, Range(f)=Range(g), then there’s some permutation of the |x,1  R states that puts the last qubit of R into an EPR pair with B Thus, if we had a reliable way to distill EPR pairs whenever possible, then we could also decide Set Equality The HH Construction (easy to prepare in poly(n) time given f,g)

My strengthening: Harlow-Hayden decoding is as hard as inverting an arbitrary one-way function B is maximally entangled with the last qubit of R. But in order to see that B and R are even classically correlated, one would need to learn x  s (a “hardcore bit” of f), and therefore invert f Is computational intractability the only “armor” protecting the geometry of spacetime inside the black hole? R: “old” Hawking photons / B: photons just coming out / H: still in black hole

Quantum Circuit Complexity and Wormholes [A.-Susskind, in progress] The AdS/CFT correspondence relates anti- deSitter quantum gravity in D spacetime dimensions to conformal field theories (without gravity) in D-1 dimensions But the mapping is extremely nonlocal! It was recently found that an expanding wormhole, on the AdS side, maps to a collection of qubits on the CFT side that just seems to get more and more “complex”:

But does C(|  t  ) actually increase like this, for natural scrambling dynamics U? Susskind’s Proposal: The quantum circuit complexity C(|  t  )— that is, the number of gates in the smallest circuit that prepares |  t  from |0   n (Not clear if it’s right, but has survived some nontrivial tests) Question: What function of |  t  can we point to on the CFT side, that’s “dual” to wormhole length on the AdS side? Time t C(|  t  ) 2n2n 0 0 2n2n

Theorem: Suppose U implements (say) a computationally- universal, reversible cellular automaton. Then after t=exp(n) iterations, C(|  t  ) is superpolynomial in n, unless something very unlikely happens with complexity classes (PSPACE  PP/poly) Proof Sketch: I proved in 2004 that PP=PostBQP Suppose C(|  t  )=n O(1). Then we could give a description of C as advice to a PostBQP machine, and the machine could efficiently prepare The machine could then measure the first register, postselect on some |x  of interest, then measure the second register to learn U t |x  —thereby solving a PSPACE-complete problem! Also have results for approximate circuit complexity, C(|  t  )  exp(n), and more Note that some complexity assumption must be made to lower-bound C(|  t  )

A Favorite Research Direction Understand, more systematically, the quantum circuit complexity of preparing n-qubit states and applying unitary transformations (“not just for quantum gravity! also for quantum algorithms, quantum money, and so much more”) Example question: For every n-qubit unitary U, is there a Boolean function f such that U can be realized by a polynomial-time quantum algorithm with an oracle for f? (I’m giving you any computational capability f you could possibly want—but it’s still far from obvious how to get the physical capability U!) Easy to show: For every n-qubit state | , there’s a Boolean function f such that |  can be prepared by a polynomial-time quantum algorithm with an oracle for f

A Related Grand Challenge Can we classify all possible sets of quantum gates acting on qubits, in terms of which unitary transformations they approximately generate? “Quantum Computing’s Classification of Finite Simple Groups” A.-Bouland 2014: Every nontrivial two-mode beamsplitter is universal Baby case that already took lots of representation theory… Warmup: Classify all the possible Hamiltonians / Lie algebras. Even just on 1 and 2 qubits!

The Classical Case A.-Grier-Schaefer 2015: Classified all sets of reversible gates in terms of which reversible transformations F:{0,1} n  {0,1} n they generate (assuming swaps and ancilla bits are free) Fredkin Toffoli CNOT

Schaeffer 2014: The first known “physically- universal” cellular automaton (able to implement any transformation in any bounded region, by suitably initializing the complement of that region) Solved open problem of Janzing 2010

Bonus: Rise and Fall of Complexity in Closed Thermodynamic Systems Unlike entropy, “interesting structure” seems to first increase and then decrease as systems mix to equilibrium Sean Carroll’s example: But how to quantify this pattern? “Apparent Complexity”: Entropy (as measured, e.g., by compressed file size) of a coarse-grained version of the image

The Coffee Automaton A., Carroll, Mohan, Ouellette, Werness 2015: A probabilistic n  n reversible system that starts half “coffee” and half “cream.” At each time step, we randomly “shear” half the coffee cup horizontally or vertically (assuming a toroidal cup) We prove that the apparent complexity of this image has a rising-falling pattern, with a maximum of at least ~n 1/6

Summary Quantum computing established a remarkable intellectual bridge between computer science and physics That’s always been why I’ve cared! Actual devices would be a bonus My research agenda: to see just how much weight this bridge can carry Rebuilding physics in the language of computation won’t be nearly as easy as some people (e.g., Wolfram) have thought! Not only does it require engaging our actual understanding of physics (QM, QFT, AdS/CFT…); it requires hard mathematical work, often making new demands on theoretical computer science But I think it’s ultimately possible