4.6.14 · D5Theory of Computation
Question bank — Variants — multi-tape TM, non-deterministic TM, all equivalent
Two words we lean on throughout, defined once so no symbol is unearned:
True or false — justify
A multi-tape TM can recognize languages a single-tape TM cannot.
False. Every -tape TM is simulated step-for-step by a single-tape TM (blocks separated by
#, dotted heads), so they recognize exactly the same languages — only the single-tape version is slower.A non-deterministic TM can recognize languages no deterministic TM can.
False. The breadth-first branch search simulates any NTM on a DTM; acceptance ("some branch accepts") is faithfully reproduced, so the language class is identical — see Non-determinism in Finite Automata (NFA = DFA) for the same phenomenon one level down.
The multi-tape → single-tape simulation runs in the same time as the original.
False. It runs in : each simulated step needs about two sweeps across the -long used region, giving a quadratic (not identical) slowdown.
The NTM → DTM simulation runs in polynomial time.
False (as far as anyone knows). The known simulation is because the computation tree has up to nodes; whether a polynomial simulation exists is exactly the open P vs NP problem.
Because the DTM simulation of an NTM is exponential, the NTM must be more powerful.
False. Exponential cost is about speed, not power. The DTM eventually accepts exactly the same strings, so they are equal in power — the blow-up is a complexity fact, not a computability one.
An NTM "guesses the right answer instantly", so it violates the idea that a machine follows fixed rules.
False. Non-determinism is a mathematical convenience for defining acceptance ("some branch accepts"), not a physical oracle. A DTM realises it mechanically by searching all branches — no magic, just slower.
The robustness of the TM (many variants, one power) is evidence for the Church–Turing Thesis.
True. If a small tweak had added power, "computable" would be fragile and arbitrary; the fact that every reasonable tweak collapses to one model is the strongest empirical support that we captured the right notion of algorithm.
A two-way-infinite tape TM (tape unbounded in both directions) is more powerful than the standard one-way-infinite tape.
False. You fold the negative half onto the positive half using a two-track encoding; the standard model simulates it, so same power (again just bookkeeping).
Adding a second head to a single tape gives a machine strictly stronger than one head.
False. Extra heads are just more "dotted positions" to track in the finite control, exactly like extra tapes — simulated by one head, same power.
Spot the error
"To simulate an NTM we do a depth-first search of the computation tree."
Error: DFS can dive into an infinite non-accepting branch and never return, missing a nearby accepting branch. We must use breadth-first search so every finite-depth accepting branch is reached.
"The address tape stores which state the branch is in at each step."
Error: The address tape stores which choice (a digit in ) to take at each step, where is the maximum number of options in any entry — not the states. States are computed by simulating, not stored as the address.
"In the -tape simulation we only need one left-to-right sweep per simulated step."
Error: You need (at least) two sweeps: the first reads all dotted symbols so the control can compute ; the second writes the new symbols and moves each dot. One sweep can't read-all-then-write-all.
"When a virtual head runs off the end of its block onto a #, we just overwrite the #."
Error: Overwriting
# would destroy the block boundary. Instead you shift everything to the right to insert a fresh blank, growing that virtual tape while keeping all delimiters intact."An NTM accepts only if all branches reach an accept state."
Error: It accepts if at least one branch accepts. Requiring all branches would be a different (universal / co-nondeterministic) acceptance rule.
"We proved multi-tape ≡ single-tape, so we may freely assume a 3-tape DTM when simulating the NTM."
Not an error — this is correct. Because 3-tape ≡ 1-tape is already established, using three tapes (input, working copy, address) is just a convenience that adds no power; it can itself be flattened to one tape afterward.
"Since for a multi-tape TM outputs symbols and moves, it can compute infinitely many things per step."
Error: is a finite table; it maps a finite tuple of read symbols to a fixed output. Finiteness of the control is exactly what lets a single-tape machine record the read symbols and reproduce the step.
Why questions
Why does "power" not include how fast the machine runs?
Because power/computability asks which problems are solvable in principle; a slow-but-correct machine still solves the problem. Timing lives in a separate axis, Time Complexity and Big-O, precisely so we can say "same power, different speed."
Why is the NTM→DTM slowdown at the heart of P vs NP?
The simulation is correct but costs ; P vs NP asks whether every NTM that accepts in polynomial time can be simulated in polynomial time deterministically. If yes, P = NP; nobody knows, so the exponent stays.
Why does the -tape simulation cost specifically?
In steps no head moves past cells, so the used region is long. Each simulated step sweeps that whole region ( work), and there are steps, giving .
Why do we mark virtual heads with a dot instead of a special separate symbol per head?
A dotted symbol carries both pieces of information at once — the tape content and "a head is here" — without consuming an extra cell or breaking the block layout, so the finite control can read position and content together.
Why must tape 1 (the input) stay read-only during the NTM simulation?
Every branch must start from the same original input; if a branch modified tape 1, the next branch would begin from corrupted data. Keeping it read-only lets us copy a fresh clone to the working tape each time.
Why does robustness of the model justify building a universal machine and speaking of "the" algorithm?
Because all reasonable variants compute the same class, there is a single model-independent notion of "computable," so one machine that can simulate any TM captures all of computation, not just one flavour.
Why does non-determinism speed things up in the definition but not necessarily in practice?
The definition lets us "declare success if any branch works," skipping the search in the model. But a real deterministic device must actually do the search, so the apparent speed-up is only realised if a clever deterministic algorithm exists — again the P vs NP problem.
Edge cases
What happens in the NTM simulation if a entry has zero options (a dead configuration)?
That branch simply halts without accepting; BFS moves on to the next address. Zero-option nodes are non-accepting leaves and never block other branches from being explored.
What if an address digit on tape 3 exceeds the number of choices available at that step?
The branch is invalid, so the DTM aborts it and increments to the next address. Skipping invalid addresses costs nothing in correctness because they correspond to no real branch.
If an NTM has no accepting branch on input (and never halts), does the DTM decide or just recognize?
It merely recognizes: the DTM keeps exploring ever-deeper addresses forever, never accepting and never halting, mirroring that no finite accepting branch exists — see Decidability and Recognizability.
Is a -tape TM a special case of a multi-tape TM?
Yes, take ; the multi-tape definition with one tape is the ordinary TM, so the equivalence theorem trivially holds in that degenerate direction.
What is the branching factor when the NTM is actually deterministic?
: every entry has exactly one option, the tree is a single line, and the "search" has one address of each length — the exponential collapses, confirming a DTM simulating a DTM has no blow-up.
What if a multi-tape machine never uses tapes (they stay blank)?
The simulation still lays out blocks separated by
#; the unused blocks just hold blanks with a dotted blank for their idle head. Correctness is unaffected — the extra empty blocks cost only linear space.Does the equivalence still hold if we allow the head to "stay put" (an move) in addition to ?
Yes. A stay move is simulated by moving right then left (or by encoding it in the control), adding no power — one more harmless bookkeeping variant that collapses to the base model.
Recall One-line summary to carry into the exam
All variants same power, possibly different speed: multi-tape costs , non-determinism costs , and none of them recognizes a single new language.